From 154e0d75fccf1784fe9ff6fd76a630b66563da3d Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Sat, 27 Jul 2024 07:55:01 +0200 Subject: Merge mainline llama.cpp (#3) * Merging mainline - WIP * Merging mainline - WIP AVX2 and CUDA appear to work. CUDA performance seems slightly (~1-2%) lower as it is so often the case with llama.cpp/ggml after some "improvements" have been made. * Merging mainline - fix Metal * Remove check --------- Co-authored-by: Iwan Kawrakow --- src/CMakeLists.txt | 33 + src/llama-grammar.cpp | 539 ++ src/llama-grammar.h | 39 + src/llama-impl.h | 26 + src/llama-sampling.cpp | 635 ++ src/llama-sampling.h | 56 + src/llama-vocab.cpp | 1721 +++++ src/llama-vocab.h | 130 + src/llama.cpp | 19336 +++++++++++++++++++++++++++++++++++++++++++++++ src/unicode-data.cpp | 7032 +++++++++++++++++ src/unicode-data.h | 20 + src/unicode.cpp | 818 ++ src/unicode.h | 67 + 13 files changed, 30452 insertions(+) create mode 100644 src/CMakeLists.txt create mode 100644 src/llama-grammar.cpp create mode 100644 src/llama-grammar.h create mode 100644 src/llama-impl.h create mode 100644 src/llama-sampling.cpp create mode 100644 src/llama-sampling.h create mode 100644 src/llama-vocab.cpp create mode 100644 src/llama-vocab.h create mode 100644 src/llama.cpp create mode 100644 src/unicode-data.cpp create mode 100644 src/unicode-data.h create mode 100644 src/unicode.cpp create mode 100644 src/unicode.h (limited to 'src') diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt new file mode 100644 index 00000000..46a6ad56 --- /dev/null +++ b/src/CMakeLists.txt @@ -0,0 +1,33 @@ +# TODO: should not use this +if (WIN32) + if (BUILD_SHARED_LIBS) + set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON) + endif() +endif() + +# +# libraries +# + +# llama + +add_library(llama + ../include/llama.h + llama.cpp + llama-vocab.cpp + llama-grammar.cpp + llama-sampling.cpp + unicode.h + unicode.cpp + unicode-data.cpp + ) + +target_include_directories(llama PUBLIC . ../include) +target_compile_features (llama PUBLIC cxx_std_11) # don't bump + +target_link_libraries(llama PUBLIC ggml) + +if (BUILD_SHARED_LIBS) + set_target_properties(llama PROPERTIES POSITION_INDEPENDENT_CODE ON) + target_compile_definitions(llama PRIVATE LLAMA_SHARED LLAMA_BUILD) +endif() diff --git a/src/llama-grammar.cpp b/src/llama-grammar.cpp new file mode 100644 index 00000000..bd9322e2 --- /dev/null +++ b/src/llama-grammar.cpp @@ -0,0 +1,539 @@ +#include "llama-grammar.h" + +#include "llama-vocab.h" +#include "llama-sampling.h" + +#include + +// Decodes a UTF-8 string which may end in an incomplete sequence. Adds a terminating 0 for use as +// pointer. If an invalid sequence is encountered, returns `llama_partial_utf8.n_remain == -1`. +std::pair, llama_partial_utf8> decode_utf8( + const std::string & src, + llama_partial_utf8 partial_start) { + static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 2, 2, 3, 4 }; + const char * pos = src.c_str(); + std::vector code_points; + + // common english strings have the same number of codepoints and bytes. `+ 1` for the terminating 0. + code_points.reserve(src.size() + 1); + uint32_t value = partial_start.value; + int n_remain = partial_start.n_remain; + + // continue previous decode, if applicable + while (*pos != 0 && n_remain > 0) { + uint8_t next_byte = static_cast(*pos); + if ((next_byte >> 6) != 2) { + // invalid sequence, abort + code_points.push_back(0); + return std::make_pair(std::move(code_points), llama_partial_utf8{ 0, -1 }); + } + value = (value << 6) + (next_byte & 0x3F); + ++pos; + --n_remain; + } + + if (partial_start.n_remain > 0 && n_remain == 0) { + code_points.push_back(value); + } + + // decode any subsequent utf-8 sequences, which may end in an incomplete one + while (*pos != 0) { + uint8_t first_byte = static_cast(*pos); + uint8_t highbits = first_byte >> 4; + n_remain = lookup[highbits] - 1; + + if (n_remain < 0) { + // invalid sequence, abort + code_points.clear(); + code_points.push_back(0); + return std::make_pair(std::move(code_points), llama_partial_utf8{ 0, n_remain }); + } + + uint8_t mask = (1 << (7 - n_remain)) - 1; + value = first_byte & mask; + + ++pos; + while (*pos != 0 && n_remain > 0) { + value = (value << 6) + (static_cast(*pos) & 0x3F); + ++pos; + --n_remain; + } + if (n_remain == 0) { + code_points.push_back(value); + } + } + code_points.push_back(0); + + return std::make_pair(std::move(code_points), llama_partial_utf8{ value, n_remain }); +} + +const llama_grammar_rules & llama_grammar_get_rules(const struct llama_grammar * grammar) { + return grammar->rules; +} + +llama_grammar_stacks & llama_grammar_get_stacks(struct llama_grammar * grammar) { + return grammar->stacks; +} + +// returns true iff pos points to the end of one of the definitions of a rule +static bool llama_grammar_is_end_of_sequence(const llama_grammar_element * pos) { + switch (pos->type) { + case LLAMA_GRETYPE_END: return true; // NOLINT + case LLAMA_GRETYPE_ALT: return true; // NOLINT + default: return false; + } +} + +// returns true iff chr satisfies the char range at pos (regular or inverse range) +// asserts that pos is pointing to a char range element +static std::pair llama_grammar_match_char( + const llama_grammar_element * pos, + const uint32_t chr) { + + bool found = false; + bool is_positive_char = pos->type == LLAMA_GRETYPE_CHAR || pos->type == LLAMA_GRETYPE_CHAR_ANY; + + GGML_ASSERT(is_positive_char || pos->type == LLAMA_GRETYPE_CHAR_NOT); // NOLINT + + do { + if (pos[1].type == LLAMA_GRETYPE_CHAR_RNG_UPPER) { + // inclusive range, e.g. [a-z] + found = found || (pos->value <= chr && chr <= pos[1].value); + pos += 2; + } else if (pos->type == LLAMA_GRETYPE_CHAR_ANY) { + // Any character matches "." + found = true; + pos += 1; + } else { + // exact char match, e.g. [a] or "a" + found = found || pos->value == chr; + pos += 1; + } + } while (pos->type == LLAMA_GRETYPE_CHAR_ALT); + + return std::make_pair(found == is_positive_char, pos); +} + +// returns true iff some continuation of the given partial UTF-8 sequence could satisfy the char +// range at pos (regular or inverse range) +// asserts that pos is pointing to a char range element +static bool llama_grammar_match_partial_char( + const llama_grammar_element * pos, + const llama_partial_utf8 partial_utf8) { + bool is_positive_char = pos->type == LLAMA_GRETYPE_CHAR || pos->type == LLAMA_GRETYPE_CHAR_ANY; + GGML_ASSERT(is_positive_char || pos->type == LLAMA_GRETYPE_CHAR_NOT); + + uint32_t partial_value = partial_utf8.value; + int n_remain = partial_utf8.n_remain; + + // invalid sequence or 7-bit char split across 2 bytes (overlong) + if (n_remain < 0 || (n_remain == 1 && partial_value < 2)) { + return false; + } + + // range of possible code points this partial UTF-8 sequence could complete to + uint32_t low = partial_value << (n_remain * 6); + uint32_t high = low | ((1 << (n_remain * 6)) - 1); + + if (low == 0) { + if (n_remain == 2) { + low = 1 << 11; + } else if (n_remain == 3) { + low = 1 << 16; + } + } + + do { + if (pos[1].type == LLAMA_GRETYPE_CHAR_RNG_UPPER) { + // inclusive range, e.g. [a-z] + if (pos->value <= high && low <= pos[1].value) { + return is_positive_char; + } + pos += 2; + } else if (pos->type == LLAMA_GRETYPE_CHAR_ANY) { + // Any character matches "." + return true; + } else { + // exact char match, e.g. [a] or "a" + if (low <= pos->value && pos->value <= high) { + return is_positive_char; + } + pos += 1; + } + } while (pos->type == LLAMA_GRETYPE_CHAR_ALT); + + return !is_positive_char; +} + +// transforms a grammar pushdown stack into N possible stacks, all ending +// at a character range (terminal element) +static void llama_grammar_advance_stack( + const llama_grammar_rules & rules, + const llama_grammar_stack & stack, + llama_grammar_stacks & new_stacks) { + if (stack.empty()) { + if (std::find(new_stacks.begin(), new_stacks.end(), stack) == new_stacks.end()) { + new_stacks.emplace_back(stack); + } + return; + } + + const llama_grammar_element * pos = stack.back(); + + switch (pos->type) { + case LLAMA_GRETYPE_RULE_REF: { + const size_t rule_id = static_cast(pos->value); + const llama_grammar_element * subpos = rules[rule_id].data(); + do { + // init new stack without the top (pos) + llama_grammar_stack new_stack(stack.begin(), stack.end() - 1); + if (!llama_grammar_is_end_of_sequence(pos + 1)) { + // if this rule ref is followed by another element, add that to stack + new_stack.push_back(pos + 1); + } + if (!llama_grammar_is_end_of_sequence(subpos)) { + // if alternate is nonempty, add to stack + new_stack.push_back(subpos); + } + llama_grammar_advance_stack(rules, new_stack, new_stacks); + while (!llama_grammar_is_end_of_sequence(subpos)) { + // scan to end of alternate def + subpos++; + } + if (subpos->type == LLAMA_GRETYPE_ALT) { + // there's another alternate def of this rule to process + subpos++; + } else { + break; + } + } while (true); + break; + } + case LLAMA_GRETYPE_CHAR: + case LLAMA_GRETYPE_CHAR_NOT: + case LLAMA_GRETYPE_CHAR_ANY: + if (std::find(new_stacks.begin(), new_stacks.end(), stack) == new_stacks.end()) { + // only add the stack if it's not a duplicate of one we already have + new_stacks.emplace_back(stack); + } + break; + default: + // end of alternate (LLAMA_GRETYPE_END, LLAMA_GRETYPE_ALT) or middle of char range + // (LLAMA_GRETYPE_CHAR_ALT, LLAMA_GRETYPE_CHAR_RNG_UPPER); stack should never be left on + // those + GGML_ASSERT(false); + } +} + +// takes a set of possible pushdown stacks on a grammar, which are required to +// be positioned at a character range (see `llama_grammar_advance_stack`), and +// produces the N possible stacks if the given char is accepted at those +// positions +void llama_grammar_accept( + const llama_grammar_rules & rules, + const llama_grammar_stacks & stacks, + const uint32_t chr, + llama_grammar_stacks & new_stacks) { + new_stacks.clear(); + + for (const auto & stack : stacks) { + if (stack.empty()) { + continue; + } + + auto match = llama_grammar_match_char(stack.back(), chr); + if (match.first) { + const llama_grammar_element * pos = match.second; + + // update top of stack to next element, if any + llama_grammar_stack new_stack(stack.begin(), stack.end() - 1); + if (!llama_grammar_is_end_of_sequence(pos)) { + new_stack.push_back(pos); + } + llama_grammar_advance_stack(rules, new_stack, new_stacks); + } + } +} + +static llama_grammar_candidates llama_grammar_reject_candidates( + const llama_grammar_rules & rules, + const llama_grammar_stacks & stacks, + const llama_grammar_candidates & candidates) { + GGML_ASSERT(!stacks.empty()); // REVIEW + + if (candidates.empty()) { + return {}; + } + + auto rejects = llama_grammar_reject_candidates_for_stack(rules, stacks.front(), candidates); + + for (size_t i = 1, size = stacks.size(); i < size; ++i) { + rejects = llama_grammar_reject_candidates_for_stack(rules, stacks[i], rejects); + } + return rejects; +} + +llama_grammar_candidates llama_grammar_reject_candidates_for_stack( + const llama_grammar_rules & rules, + const llama_grammar_stack & stack, + const llama_grammar_candidates & candidates) { + + llama_grammar_candidates rejects; + rejects.reserve(candidates.size()); + + if (stack.empty()) { + for (const auto & tok : candidates) { + if (*tok.code_points != 0 || tok.partial_utf8.n_remain != 0) { + rejects.push_back(tok); + } + } + return rejects; + } + + const llama_grammar_element * stack_pos = stack.back(); + + llama_grammar_candidates next_candidates; + next_candidates.reserve(candidates.size()); + + for (const auto & tok : candidates) { + if (*tok.code_points == 0) { + // reached end of full codepoints in token, reject iff it ended in a partial sequence + // that cannot satisfy this position in grammar + if (tok.partial_utf8.n_remain != 0 && + !llama_grammar_match_partial_char(stack_pos, tok.partial_utf8)) { + rejects.push_back(tok); + } + } else if (llama_grammar_match_char(stack_pos, *tok.code_points).first) { + next_candidates.push_back({ tok.index, tok.code_points + 1, tok.partial_utf8 }); + } else { + rejects.push_back(tok); + } + } + + const auto * stack_pos_after = llama_grammar_match_char(stack_pos, 0).second; + + // update top of stack to next element, if any + llama_grammar_stack stack_after(stack.begin(), stack.end() - 1); + if (!llama_grammar_is_end_of_sequence(stack_pos_after)) { + stack_after.push_back(stack_pos_after); + } + llama_grammar_stacks next_stacks; + llama_grammar_advance_stack(rules, stack_after, next_stacks); + + auto next_rejects = llama_grammar_reject_candidates(rules, next_stacks, next_candidates); + for (const auto & tok : next_rejects) { + rejects.push_back({ tok.index, tok.code_points - 1, tok.partial_utf8 }); + } + + return rejects; +} + +static bool llama_grammar_detect_left_recursion( + const llama_grammar_rules & rules, + size_t rule_index, + std::vector * rules_visited, + std::vector * rules_in_progress, + std::vector * rules_may_be_empty) { + if ((*rules_in_progress)[rule_index]) { + return true; + } + + (*rules_in_progress)[rule_index] = true; + + const llama_grammar_rule & rule = rules[rule_index]; + + // First check if the rule might produce the empty string. This could be done combined with the second + // step but it's more readable as two steps. + bool at_rule_start = true; + for (size_t i = 0; i < rule.size(); i++) { + if (llama_grammar_is_end_of_sequence(&rule[i])) { + if (at_rule_start) { + (*rules_may_be_empty)[rule_index] = true; + break; + } + at_rule_start = true; + } else { + at_rule_start = false; + } + } + + // Second, recurse into leftmost nonterminals (or next-leftmost as long as the previous nonterminal may + // be empty) + bool recurse_into_nonterminal = true; + for (size_t i = 0; i < rule.size(); i++) { + if (rule[i].type == LLAMA_GRETYPE_RULE_REF && recurse_into_nonterminal) { + if (llama_grammar_detect_left_recursion(rules, (size_t)rule[i].value, rules_visited, rules_in_progress, rules_may_be_empty)) { + return true; + } + if (!((*rules_may_be_empty)[(size_t)rule[i].value])) { + recurse_into_nonterminal = false; + } + } else if (llama_grammar_is_end_of_sequence(&rule[i])) { + recurse_into_nonterminal = true; + } else { + recurse_into_nonterminal = false; + } + } + + (*rules_in_progress)[rule_index] = false; + (*rules_visited)[rule_index] = true; + return false; +} + +// +// grammar - external +// + +struct llama_grammar * llama_grammar_init_impl( + const llama_grammar_element ** rules, + size_t n_rules, + size_t start_rule_index) { + const llama_grammar_element * pos; + + // copy rule definitions into vectors + llama_grammar_rules vec_rules(n_rules); + for (size_t i = 0; i < n_rules; i++) { + for (pos = rules[i]; pos->type != LLAMA_GRETYPE_END; pos++) { + vec_rules[i].push_back(*pos); + } + vec_rules[i].push_back({LLAMA_GRETYPE_END, 0}); + } + + // Check for left recursion + std::vector rules_visited(n_rules); + std::vector rules_in_progress(n_rules); + std::vector rules_may_be_empty(n_rules); + for (size_t i = 0; i < n_rules; i++) { + if (rules_visited[i]) { + continue; + } + if (llama_grammar_detect_left_recursion(vec_rules, i, &rules_visited, &rules_in_progress, &rules_may_be_empty)) { + LLAMA_LOG_ERROR("unsupported grammar, left recursion detected for nonterminal at index %zu", i); + return nullptr; + } + } + + // loop over alternates of start rule to build initial stacks + llama_grammar_stacks stacks; + pos = vec_rules[start_rule_index].data(); + do { + llama_grammar_stack stack; + if (!llama_grammar_is_end_of_sequence(pos)) { + // if alternate is nonempty, add to stack + stack.push_back(pos); + } + llama_grammar_advance_stack(vec_rules, stack, stacks); + while (!llama_grammar_is_end_of_sequence(pos)) { + // scan to end of alternate def + pos++; + } + if (pos->type == LLAMA_GRETYPE_ALT) { + // there's another alternate def of this rule to process + pos++; + } else { + break; + } + } while (true); + + // Important: vec_rules has to be moved here, not copied, because stacks contains + // pointers to elements of vec_rules. If vec_rules were copied into llama_grammar + // then the pointers would be invalidated when the local vec_rules goes out of scope. + return new llama_grammar{ std::move(vec_rules), std::move(stacks), {} }; +} + +void llama_grammar_free_impl(struct llama_grammar * grammar) { + delete grammar; +} + +struct llama_grammar * llama_grammar_copy_impl(const struct llama_grammar * grammar) { + llama_grammar * result = new llama_grammar{ grammar->rules, grammar->stacks, grammar->partial_utf8 }; + + // redirect elements in stacks to point to new rules + for (size_t is = 0; is < result->stacks.size(); is++) { + for (size_t ie = 0; ie < result->stacks[is].size(); ie++) { + for (size_t ir0 = 0; ir0 < grammar->rules.size(); ir0++) { + for (size_t ir1 = 0; ir1 < grammar->rules[ir0].size(); ir1++) { + if (grammar->stacks[is][ie] == &grammar->rules[ir0][ir1]) { + result->stacks[is][ie] = &result->rules[ir0][ir1]; + } + } + } + } + } + + return result; +} + +void llama_grammar_sample_impl(const struct llama_grammar * grammar, const struct llama_vocab * vocab, const struct llama_sampling * smpl, llama_token_data_array * candidates) { + GGML_ASSERT(grammar); + GGML_ASSERT(vocab); + + int64_t t_start_sample_us = ggml_time_us(); + + bool allow_eog = false; + for (const auto & stack : grammar->stacks) { + if (stack.empty()) { + allow_eog = true; + break; + } + } + + std::vector, llama_partial_utf8>> candidates_decoded; + candidates_decoded.reserve(candidates->size); + + llama_grammar_candidates candidates_grammar; + candidates_grammar.reserve(candidates->size); + + for (size_t i = 0; i < candidates->size; ++i) { + const llama_token id = candidates->data[i].id; + const std::string & piece = vocab->cache_token_to_piece.at(id); + + if (llama_token_is_eog_impl(*vocab, id)) { + if (!allow_eog) { + candidates->data[i].logit = -INFINITY; + } + } else if (piece.empty() || piece[0] == 0) { + candidates->data[i].logit = -INFINITY; + } else { + candidates_decoded.push_back(decode_utf8(piece, grammar->partial_utf8)); + candidates_grammar.push_back({ i, candidates_decoded.back().first.data(), candidates_decoded.back().second }); + } + } + + const auto rejects = llama_grammar_reject_candidates(grammar->rules, grammar->stacks, candidates_grammar); + for (const auto & reject : rejects) { + candidates->data[reject.index].logit = -INFINITY; + } + + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; +} + +void llama_grammar_accept_token_impl(struct llama_grammar * grammar, const struct llama_vocab * vocab, const struct llama_sampling * smpl, llama_token token) { + const int64_t t_start_sample_us = ggml_time_us(); + + if (llama_token_is_eog_impl(*vocab, token)) { + for (const auto & stack : grammar->stacks) { + if (stack.empty()) { + return; + } + } + GGML_ASSERT(false); + } + + const std::string & piece = vocab->cache_token_to_piece.at(token); + + // Note terminating 0 in decoded string + const auto decoded = decode_utf8(piece, grammar->partial_utf8); + const auto & code_points = decoded.first; + + llama_grammar_stacks tmp_new_stacks; + for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) { + llama_grammar_accept(grammar->rules, grammar->stacks, *it, tmp_new_stacks); + grammar->stacks = tmp_new_stacks; + } + + grammar->partial_utf8 = decoded.second; + GGML_ASSERT(!grammar->stacks.empty()); + + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; +} diff --git a/src/llama-grammar.h b/src/llama-grammar.h new file mode 100644 index 00000000..695ea063 --- /dev/null +++ b/src/llama-grammar.h @@ -0,0 +1,39 @@ +#pragma once + +#include "llama-impl.h" + +struct llama_vocab; +struct llama_sampling; + +struct llama_grammar { + const llama_grammar_rules rules; + llama_grammar_stacks stacks; + + // buffer for partially generated UTF-8 sequence from accepted tokens + llama_partial_utf8 partial_utf8; +}; + +// +// internal API +// + +struct llama_grammar * llama_grammar_init_impl( + const llama_grammar_element ** rules, + size_t n_rules, + size_t start_rule_index); + +void llama_grammar_free_impl(struct llama_grammar * grammar); + +struct llama_grammar * llama_grammar_copy_impl(const struct llama_grammar * grammar); + +void llama_grammar_sample_impl( + const struct llama_grammar * grammar, + const struct llama_vocab * vocab, + const struct llama_sampling * smpl, + llama_token_data_array * candidates); + +void llama_grammar_accept_token_impl( + struct llama_grammar * grammar, + const struct llama_vocab * vocab, + const struct llama_sampling * smpl, + llama_token token); diff --git a/src/llama-impl.h b/src/llama-impl.h new file mode 100644 index 00000000..dcc8c1c1 --- /dev/null +++ b/src/llama-impl.h @@ -0,0 +1,26 @@ +#pragma once + +#define LLAMA_API_INTERNAL +#include "llama.h" + +#ifdef __GNUC__ +#ifdef __MINGW32__ +#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__))) +#else +#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__))) +#endif +#else +#define LLAMA_ATTRIBUTE_FORMAT(...) +#endif + +// +// logging +// + +LLAMA_ATTRIBUTE_FORMAT(2, 3) +void llama_log_internal (ggml_log_level level, const char * format, ...); +void llama_log_callback_default(ggml_log_level level, const char * text, void * user_data); + +#define LLAMA_LOG_INFO(...) llama_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__) +#define LLAMA_LOG_WARN(...) llama_log_internal(GGML_LOG_LEVEL_WARN , __VA_ARGS__) +#define LLAMA_LOG_ERROR(...) llama_log_internal(GGML_LOG_LEVEL_ERROR, __VA_ARGS__) diff --git a/src/llama-sampling.cpp b/src/llama-sampling.cpp new file mode 100644 index 00000000..8910f6d6 --- /dev/null +++ b/src/llama-sampling.cpp @@ -0,0 +1,635 @@ +#include "llama-sampling.h" + +#include +#include +#include +#include +#include +#include + +static void llama_log_softmax(float * array, size_t size) { + float max_l = *std::max_element(array, array + size); + float sum = 0.f; + for (size_t i = 0; i < size; ++i) { + float p = expf(array[i] - max_l); + sum += p; + array[i] = p; + } + + for (size_t i = 0; i < size; ++i) { + array[i] = logf(array[i] / sum); + } +} + +void llama_set_rng_seed_impl(struct llama_sampling * smpl, uint32_t seed) { + if (seed == LLAMA_DEFAULT_SEED) { + seed = time(NULL); + } + + smpl->rng.seed(seed); +} + +void llama_sample_softmax_impl(struct llama_sampling * smpl, llama_token_data_array * candidates) { + GGML_ASSERT(candidates->size > 0); + + const int64_t t_start_sample_us = ggml_time_us(); + + // Sort the logits in descending order + if (!candidates->sorted) { + std::sort(candidates->data, candidates->data + candidates->size, [](const llama_token_data & a, const llama_token_data & b) { + return a.logit > b.logit; + }); + candidates->sorted = true; + } + + float max_l = candidates->data[0].logit; + float cum_sum = 0.0f; + for (size_t i = 0; i < candidates->size; ++i) { + float p = expf(candidates->data[i].logit - max_l); + candidates->data[i].p = p; + cum_sum += p; + } + for (size_t i = 0; i < candidates->size; ++i) { + candidates->data[i].p /= cum_sum; + } + + if (smpl) { + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + +void llama_sample_top_k_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, int32_t k, size_t min_keep) { + // TODO: move bucket sort to separate function so that top_p/tail_free/typical/softmax first is equally fast + // if (k >= (int32_t)candidates->size) { + // return; + // } + + const int64_t t_start_sample_us = ggml_time_us(); + + if (k <= 0) { + k = candidates->size; + } + + k = std::max(k, (int) min_keep); + k = std::min(k, (int) candidates->size); + + // Sort scores in descending order + if (!candidates->sorted) { + auto comp = [](const llama_token_data & a, const llama_token_data & b) { + return a.logit > b.logit; + }; + if (k <= 128) { + std::partial_sort(candidates->data, candidates->data + k, candidates->data + candidates->size, comp); + } else { + constexpr int nbuckets = 128; + constexpr float bucket_low = -10.0f; + constexpr float bucket_high = 10.0f; + constexpr float bucket_scale = nbuckets/(bucket_high - bucket_low); + constexpr float bucker_inter = -bucket_low * bucket_scale; + + std::vector bucket_idx(candidates->size); + std::vector histo(nbuckets, 0); + + for (int i = 0; i < (int)candidates->size; ++i) { + const float val = candidates->data[i].logit; + int ib = int(bucket_scale * val + bucker_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low); + ib = std::max(0, std::min(nbuckets-1, ib)); + bucket_idx[i] = ib; + ++histo[ib]; + } + int nhave = 0; + int ib = nbuckets - 1; + for ( ; ib >= 0; --ib) { + nhave += histo[ib]; + if (nhave >= k) break; + } + std::vector tmp_tokens(nhave); + auto ptr = tmp_tokens.data(); + std::vector bucket_ptrs; + bucket_ptrs.reserve(nbuckets - ib); + for (int j = nbuckets - 1; j >= ib; --j) { + bucket_ptrs.push_back(ptr); + ptr += histo[j]; + } + for (int i = 0; i < (int)candidates->size; ++i) { + int j = bucket_idx[i]; + if (j >= ib) { + *bucket_ptrs[nbuckets-1-j]++ = candidates->data[i]; + } + } + + ptr = tmp_tokens.data(); + int ndone = 0; + for (int j = nbuckets-1; j > ib; --j) { + std::sort(ptr, ptr + histo[j], comp); + ptr += histo[j]; + ndone += histo[j]; + } + std::partial_sort(ptr, ptr + k - ndone, ptr + histo[ib], comp); + + std::memcpy(candidates->data, tmp_tokens.data(), k*sizeof(llama_token_data)); + + } + candidates->sorted = true; + } + candidates->size = k; + + if (smpl) { + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + +void llama_sample_top_p_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float p, size_t min_keep) { + if (p >= 1.0f) { + return; + } + + llama_sample_softmax_impl(smpl, candidates); + + const int64_t t_start_sample_us = ggml_time_us(); + + // Compute the cumulative probabilities + float cum_sum = 0.0f; + size_t last_idx = candidates->size; + + for (size_t i = 0; i < candidates->size; ++i) { + cum_sum += candidates->data[i].p; + + // Check if the running sum is at least p or if we have kept at least min_keep tokens + // we set the last index to i+1 to indicate that the current iterate should be included in the set + if (cum_sum >= p && i + 1 >= min_keep) { + last_idx = i + 1; + break; + } + } + + // Resize the output vector to keep only the top-p tokens + candidates->size = last_idx; + + if (smpl) { + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + +void llama_sample_min_p_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float p, size_t min_keep) { + if (p <= 0.0f || !candidates->size) { + return; + } + + const int64_t t_start_sample_us = ggml_time_us(); + + bool min_p_applied = false; + + // if the candidates aren't sorted, try the unsorted implementation first + if (!candidates->sorted) { + std::vector filtered_tokens; + + float max_logit = -FLT_MAX; + for (size_t i = 0; i < candidates->size; ++i) { + max_logit = std::max(max_logit, candidates->data[i].logit); + } + const float min_logit = max_logit + logf(p); // min logit for p_i >= p * p_max + + for (size_t i = 0; i < candidates->size; ++i) { + if (candidates->data[i].logit >= min_logit) { + filtered_tokens.push_back(candidates->data[i]); + } + } + + // if we have enough values the operation was a success + if (filtered_tokens.size() >= min_keep) { + memcpy(candidates->data, filtered_tokens.data(), filtered_tokens.size()*sizeof(llama_token_data)); + candidates->size = filtered_tokens.size(); + min_p_applied = true; + } + } + + // if the candidates are sorted or the unsorted implementation failed, use this implementation + if (!min_p_applied) { + // Sort the logits in descending order + if (!candidates->sorted) { + std::sort(candidates->data, candidates->data + candidates->size, [](const llama_token_data & a, const llama_token_data & b) { + return a.logit > b.logit; + }); + candidates->sorted = true; + } + + const float min_logit = candidates->data[0].logit + logf(p); // min logit for p_i >= p * p_max + size_t i = 1; // first token always matches + + for (; i < candidates->size; ++i) { + if (candidates->data[i].logit < min_logit && i >= min_keep) { + break; // prob too small + } + } + + // Resize the output vector to keep only the matching tokens + candidates->size = i; + } + + if (smpl) { + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + +void llama_sample_tail_free_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float z, size_t min_keep) { + if (z >= 1.0f || candidates->size <= 2) { + return; + } + + llama_sample_softmax_impl((struct llama_sampling *) nullptr, candidates); + const int64_t t_start_sample_us = ggml_time_us(); + + // Compute the first and second derivatives + std::vector first_derivatives(candidates->size - 1); + std::vector second_derivatives(candidates->size - 2); + + for (size_t i = 0; i < first_derivatives.size(); ++i) { + first_derivatives[i] = candidates->data[i].p - candidates->data[i + 1].p; + } + for (size_t i = 0; i < second_derivatives.size(); ++i) { + second_derivatives[i] = first_derivatives[i] - first_derivatives[i + 1]; + } + + // Calculate absolute value of second derivatives + for (size_t i = 0; i < second_derivatives.size(); ++i) { + second_derivatives[i] = std::abs(second_derivatives[i]); + } + + // Normalize the second derivatives + { + const float second_derivatives_sum = std::accumulate(second_derivatives.begin(), second_derivatives.end(), 0.0f); + + if (second_derivatives_sum > 1e-6f) { + for (float & value : second_derivatives) { + value /= second_derivatives_sum; + } + } else { + for (float & value : second_derivatives) { + value = 1.0f / second_derivatives.size(); + } + } + } + + float cum_sum = 0.0f; + size_t last_idx = candidates->size; + for (size_t i = 0; i < second_derivatives.size(); ++i) { + cum_sum += second_derivatives[i]; + + // Check if the running sum is greater than z or if we have kept at least min_keep tokens + if (cum_sum > z && i >= min_keep) { + last_idx = i; + break; + } + } + + // Resize the output vector to keep only the tokens above the tail location + candidates->size = last_idx; + + if (smpl) { + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + +void llama_sample_typical_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float p, size_t min_keep) { + // Reference implementation: + // https://github.com/huggingface/transformers/compare/main...cimeister:typical-sampling:typical-pr + if (p >= 1.0f) { + return; + } + + // Compute the softmax of logits and calculate entropy + llama_sample_softmax_impl((struct llama_sampling *) nullptr, candidates); + + const int64_t t_start_sample_us = ggml_time_us(); + + float entropy = 0.0f; + for (size_t i = 0; i < candidates->size; ++i) { + entropy += -candidates->data[i].p * logf(candidates->data[i].p); + } + + // Compute the absolute difference between negative log probability and entropy for each candidate + std::vector shifted_scores; + for (size_t i = 0; i < candidates->size; ++i) { + float shifted_score = fabsf(-logf(candidates->data[i].p) - entropy); + shifted_scores.push_back(shifted_score); + } + + // Sort tokens based on the shifted_scores and their corresponding indices + std::vector indices(candidates->size); + std::iota(indices.begin(), indices.end(), 0); + + std::sort(indices.begin(), indices.end(), [&](size_t a, size_t b) { + return shifted_scores[a] < shifted_scores[b]; + }); + + // Compute the cumulative probabilities + float cum_sum = 0.0f; + size_t last_idx = indices.size(); + + for (size_t i = 0; i < indices.size(); ++i) { + size_t idx = indices[i]; + cum_sum += candidates->data[idx].p; + + // Check if the running sum is greater than typical or if we have kept at least min_keep tokens + if (cum_sum > p && i >= min_keep - 1) { + last_idx = i + 1; + break; + } + } + + // Resize the output vector to keep only the locally typical tokens + std::vector new_candidates; + for (size_t i = 0; i < last_idx; ++i) { + size_t idx = indices[i]; + new_candidates.push_back(candidates->data[idx]); + } + + // Replace the data in candidates with the new_candidates data + std::copy(new_candidates.begin(), new_candidates.end(), candidates->data); + candidates->size = new_candidates.size(); + candidates->sorted = false; + + if (smpl) { + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + +void llama_sample_entropy_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float min_temp, float max_temp, float exponent_val) { + const int64_t t_start_sample_us = ggml_time_us(); + + // no need to do anything if there is only one (or zero) candidates + if(candidates->size <= 1) { + return; + } + + // Calculate maximum possible entropy + float max_entropy = -logf(1.0f / candidates->size); + + llama_sample_softmax_impl((struct llama_sampling *) nullptr, candidates); + + // Calculate entropy of the softmax probabilities + float entropy = 0.0f; + for (size_t i = 0; i < candidates->size; ++i) { + float prob = candidates->data[i].p; + if (prob > 0.0f) { // Ensure no log(0) + entropy -= prob * logf(prob); + } + } + + // Normalize the entropy (max_entropy cannot be 0 here because we checked candidates->size != 1 above) + float normalized_entropy = entropy / max_entropy; + + // Map the normalized entropy to the desired temperature range using the power function + float dyn_temp = min_temp + (max_temp - min_temp) * powf(normalized_entropy, exponent_val); + +#ifdef DEBUG + LLAMA_LOG_INFO("Your text maxtemp value is: %f\n", max_temp); + LLAMA_LOG_INFO("Entropy: %f\n", entropy); + LLAMA_LOG_INFO("Max Possible Entropy: %f\n", max_entropy); + LLAMA_LOG_INFO("Normalized Entropy: %f\n", normalized_entropy); + LLAMA_LOG_INFO("Exponent: %f\n", exponent_val); + LLAMA_LOG_INFO("Dynamic Temperature (dyn_temp): %f\n", dyn_temp); +#endif + + // Apply the dynamically calculated temperature scaling + for (size_t i = 0; i < candidates->size; ++i) { + candidates->data[i].logit /= dyn_temp; + } + + // Re-compute softmax probabilities after scaling logits with dynamic temperature + double max_l_double = candidates->data[0].logit; + double cum_sum_double = 0.0; + for (size_t i = 0; i < candidates->size; ++i) { + double p = exp(candidates->data[i].logit - max_l_double); + candidates->data[i].p = p; // Store the scaled probability + cum_sum_double += p; + } + for (size_t i = 0; i < candidates->size; ++i) { + candidates->data[i].p /= cum_sum_double; // Re-normalize the probabilities + } + +#ifdef DEBUG + // Print the updated top 25 probabilities after temperature scaling + LLAMA_LOG_INFO("\nUpdated Top 25 Probabilities After Dynamic Temperature Scaling (in percentages):\n"); + for (size_t i = 0; i < 25 && i < candidates->size; ++i) { + LLAMA_LOG_INFO("Token %zu: %f%%\n", i + 1, candidates->data[i].p * 100.0f); + } +#endif + + if (smpl) { + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + +void llama_sample_temp_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float temp) { + const int64_t t_start_sample_us = ggml_time_us(); + + for (size_t i = 0; i < candidates->size; ++i) { + candidates->data[i].logit /= temp; + } + + if (smpl) { + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + +void llama_sample_repetition_penalties_impl( + struct llama_sampling * smpl, + llama_token_data_array * candidates, + const llama_token * last_tokens, + size_t penalty_last_n, + float penalty_repeat, + float penalty_freq, + float penalty_present) { + if (penalty_last_n == 0 || (penalty_repeat == 1.0f && penalty_freq == 0.0f && penalty_present == 0.0f)) { + return; + } + + const int64_t t_start_sample_us = ggml_time_us(); + + // Create a frequency map to count occurrences of each token in last_tokens + std::unordered_map token_count; + for (size_t i = 0; i < penalty_last_n; ++i) { + token_count[last_tokens[i]]++; + } + + // Apply frequency and presence penalties to the candidates + for (size_t i = 0; i < candidates->size; ++i) { + const auto token_iter = token_count.find(candidates->data[i].id); + if (token_iter == token_count.end()) { + continue; + } + + const int count = token_iter->second; + + // The academic publication that described this technique actually just only divided, but that would cause tokens with negative logits to become more likely, which is obviously wrong. + // This is common fix for this problem, which is to multiply by the penalty instead of dividing. + if (candidates->data[i].logit <= 0) { + candidates->data[i].logit *= penalty_repeat; + } else { + candidates->data[i].logit /= penalty_repeat; + } + + candidates->data[i].logit -= float(count) * penalty_freq + float(count > 0) * penalty_present; + } + + candidates->sorted = false; + + if (smpl) { + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + +void llama_sample_apply_guidance_impl( + struct llama_sampling * smpl, + float * logits, + float * logits_guidance, + float scale) { + GGML_ASSERT(smpl); + + const auto t_start_sample_us = ggml_time_us(); + const auto n_vocab = smpl->n_vocab; + + llama_log_softmax(logits, n_vocab); + llama_log_softmax(logits_guidance, n_vocab); + + for (int i = 0; i < n_vocab; ++i) { + auto & l = logits[i]; + const auto & g = logits_guidance[i]; + + l = scale * (l - g) + g; + } + + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; +} + +llama_token llama_sample_token_mirostat_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float tau, float eta, int32_t m, float * mu) { + GGML_ASSERT(smpl); + + const int32_t n_vocab = float(smpl->n_vocab); + + int64_t t_start_sample_us = ggml_time_us(); + + llama_sample_softmax_impl((struct llama_sampling *) nullptr, candidates); + + // Estimate s_hat using the most probable m tokens + float s_hat = 0.0; + float sum_ti_bi = 0.0; + float sum_ti_sq = 0.0; + for (size_t i = 0; i < size_t(m - 1) && i < candidates->size - 1; ++i) { + float t_i = logf(float(i + 2) / float(i + 1)); + float b_i = logf(candidates->data[i].p / candidates->data[i + 1].p); + sum_ti_bi += t_i * b_i; + sum_ti_sq += t_i * t_i; + } + s_hat = sum_ti_bi / sum_ti_sq; + + // Compute k from the estimated s_hat and target surprise value + float epsilon_hat = s_hat - 1; + float k = powf((epsilon_hat * powf(2, *mu)) / (1 - powf(n_vocab, -epsilon_hat)), 1 / s_hat); + + // Sample the next word X using top-k sampling + llama_sample_top_k_impl((struct llama_sampling *) nullptr, candidates, int(k), 1); + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; + llama_token X = llama_sample_token_impl(smpl, candidates); + t_start_sample_us = ggml_time_us(); + + // Compute error as the difference between observed surprise and target surprise value + size_t X_idx = std::distance(candidates->data, std::find_if(candidates->data, candidates->data + candidates->size, [&](const llama_token_data & candidate) { + return candidate.id == X; + })); + float observed_surprise = -log2f(candidates->data[X_idx].p); + float e = observed_surprise - tau; + + // Update mu using the learning rate and error + *mu = *mu - eta * e; + + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; + return X; +} + +llama_token llama_sample_token_mirostat_v2_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float tau, float eta, float * mu) { + int64_t t_start_sample_us; + t_start_sample_us = ggml_time_us(); + + llama_sample_softmax_impl(smpl, candidates); + + // Truncate the words with surprise values greater than mu + candidates->size = std::distance(candidates->data, std::find_if(candidates->data, candidates->data + candidates->size, [&](const llama_token_data & candidate) { + return -log2f(candidate.p) > *mu; + })); + + if (candidates->size == 0) { + candidates->size = 1; + } + + if (smpl) { + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; + } + + // Normalize the probabilities of the remaining words + llama_sample_softmax_impl(smpl, candidates); + + // Sample the next word X from the remaining words + llama_token X = llama_sample_token_impl(smpl, candidates); + t_start_sample_us = ggml_time_us(); + + // Compute error as the difference between observed surprise and target surprise value + size_t X_idx = std::distance(candidates->data, std::find_if(candidates->data, candidates->data + candidates->size, [&](const llama_token_data & candidate) { + return candidate.id == X; + })); + float observed_surprise = -log2f(candidates->data[X_idx].p); + float e = observed_surprise - tau; + + // Update mu using the learning rate and error + *mu = *mu - eta * e; + + if (smpl) { + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; + } + return X; +} + +llama_token llama_sample_token_greedy_impl(struct llama_sampling * smpl, llama_token_data_array * candidates) { + const int64_t t_start_sample_us = ggml_time_us(); + + // Find max element + auto * max_iter = std::max_element(candidates->data, candidates->data + candidates->size, [](const llama_token_data & a, const llama_token_data & b) { + return a.logit < b.logit; + }); + + llama_token result = max_iter->id; + if (smpl) { + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; + smpl->n_sample++; + } + return result; +} + +llama_token llama_sample_token_with_rng_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, std::mt19937 & rng) { + GGML_ASSERT(smpl); + + const int64_t t_start_sample_us = ggml_time_us(); + llama_sample_softmax_impl((struct llama_sampling *) nullptr, candidates); + + std::vector probs; + probs.reserve(candidates->size); + for (size_t i = 0; i < candidates->size; ++i) { + probs.push_back(candidates->data[i].p); + } + + std::discrete_distribution<> dist(probs.begin(), probs.end()); + int idx = dist(rng); + + llama_token result = candidates->data[idx].id; + + smpl->t_sample_us += ggml_time_us() - t_start_sample_us; + smpl->n_sample++; + + return result; +} + +llama_token llama_sample_token_impl(struct llama_sampling * smpl, llama_token_data_array * candidates) { + return llama_sample_token_with_rng_impl(smpl, candidates, smpl->rng); +} diff --git a/src/llama-sampling.h b/src/llama-sampling.h new file mode 100644 index 00000000..f7f8e3ef --- /dev/null +++ b/src/llama-sampling.h @@ -0,0 +1,56 @@ +#pragma once + +#include "llama-impl.h" + +struct llama_sampling { + llama_sampling(int32_t n_vocab) : n_vocab(n_vocab) {} + + std::mt19937 rng; + + int32_t n_vocab = 0; + + mutable int64_t t_sample_us = 0; + mutable int32_t n_sample = 0; + + void reset_timings() const { + t_sample_us = 0; + n_sample = 0; + } +}; + +// +// internal API +// + +void llama_set_rng_seed_impl(struct llama_sampling * smpl, uint32_t seed); + +void llama_sample_softmax_impl (struct llama_sampling * smpl, llama_token_data_array * candidates); +void llama_sample_top_k_impl (struct llama_sampling * smpl, llama_token_data_array * candidates, int32_t k, size_t min_keep); +void llama_sample_top_p_impl (struct llama_sampling * smpl, llama_token_data_array * candidates, float p, size_t min_keep); +void llama_sample_min_p_impl (struct llama_sampling * smpl, llama_token_data_array * candidates, float p, size_t min_keep); +void llama_sample_tail_free_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float z, size_t min_keep); +void llama_sample_typical_impl (struct llama_sampling * smpl, llama_token_data_array * candidates, float p, size_t min_keep); +void llama_sample_entropy_impl (struct llama_sampling * smpl, llama_token_data_array * candidates, float min_temp, float max_temp, float exponent_val); +void llama_sample_temp_impl (struct llama_sampling * smpl, llama_token_data_array * candidates, float temp); + +void llama_sample_repetition_penalties_impl( + struct llama_sampling * smpl, + llama_token_data_array * candidates, + const llama_token * last_tokens, + size_t penalty_last_n, + float penalty_repeat, + float penalty_freq, + float penalty_present); + +void llama_sample_apply_guidance_impl( + struct llama_sampling * smpl, + float * logits, + float * logits_guidance, + float scale); + +llama_token llama_sample_token_mirostat_impl (struct llama_sampling * smpl, llama_token_data_array * candidates, float tau, float eta, int32_t m, float * mu); +llama_token llama_sample_token_mirostat_v2_impl(struct llama_sampling * smpl, llama_token_data_array * candidates, float tau, float eta, float * mu); +llama_token llama_sample_token_greedy_impl (struct llama_sampling * smpl, llama_token_data_array * candidates); +llama_token llama_sample_token_with_rng_impl (struct llama_sampling * smpl, llama_token_data_array * candidates, std::mt19937 & rng); +llama_token llama_sample_token_impl (struct llama_sampling * smpl, llama_token_data_array * candidates); + diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp new file mode 100644 index 00000000..c482b368 --- /dev/null +++ b/src/llama-vocab.cpp @@ -0,0 +1,1721 @@ +#include "llama-vocab.h" + +#include "unicode.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +// +// helpers +// + +static void replace_all(std::string & s, const std::string & search, const std::string & replace) { + std::string result; + for (size_t pos = 0; ; pos += search.length()) { + auto new_pos = s.find(search, pos); + if (new_pos == std::string::npos) { + result += s.substr(pos, s.size() - pos); + break; + } + result += s.substr(pos, new_pos - pos) + replace; + pos = new_pos; + } + s = std::move(result); +} + +LLAMA_ATTRIBUTE_FORMAT(1, 2) +static std::string format(const char * fmt, ...) { + va_list ap; + va_list ap2; + va_start(ap, fmt); + va_copy(ap2, ap); + int size = vsnprintf(NULL, 0, fmt, ap); + GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT + std::vector buf(size + 1); + int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2); + GGML_ASSERT(size2 == size); + va_end(ap2); + va_end(ap); + return std::string(buf.data(), size); +} + +struct naive_trie { + naive_trie() : has_value(false), value(0) { + } + void insert(const char * key, size_t len, int32_t value = 0) { + if (len == 0) { + this->has_value = true; + this->value = value; + return; + } + char c = key[0]; + auto res = children.find(c); + if (res != children.end()) { + res->second.insert(key + 1, len - 1, value); + } else { + auto res = children.insert(std::make_pair(c, naive_trie())); + res.first->second.insert(key + 1, len - 1, value); + } + } + std::pair get_longest_prefix(const char * key, size_t len, size_t offset = 0) { + if (len == 0 || offset == len) { + return std::make_pair(key, offset); + } + char c = key[offset]; + auto res = children.find(c); + if (res != children.end()) { + return res->second.get_longest_prefix(key, len, offset + 1); + } else { + return std::make_pair(key, offset); + } + } + struct naive_trie * traverse(const char c) { + auto res = children.find(c); + if (res != children.end()) { + return &res->second; + } else { + return NULL; + } + } + std::map children; + bool has_value; + llama_token value; +}; + +// +// impl +// + +int llama_vocab::find_bpe_rank(const std::string & token_left, const std::string & token_right) const { + GGML_ASSERT(token_left.find(' ') == std::string::npos); + GGML_ASSERT(token_left.find('\n') == std::string::npos); + GGML_ASSERT(token_right.find(' ') == std::string::npos); + GGML_ASSERT(token_right.find('\n') == std::string::npos); + + auto it = bpe_ranks.find(std::make_pair(token_left, token_right)); + if (it == bpe_ranks.end()) { + return -1; + } + + return it->second; +} + +static enum llama_vocab_type llama_vocab_get_type(const llama_vocab & vocab) { + return vocab.type; +} + +static bool llama_is_normal_token(const llama_vocab & vocab, llama_token id) { + GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE); + return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_NORMAL; +} + +static bool llama_is_unknown_token(const llama_vocab & vocab, llama_token id) { + GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE); + return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_UNKNOWN; +} + +static bool llama_is_control_token(const llama_vocab & vocab, llama_token id) { + GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE); + return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_CONTROL; +} + +static bool llama_is_byte_token(const llama_vocab & vocab, llama_token id) { + GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE); + return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_BYTE; +} + +static bool llama_is_user_defined_token(const llama_vocab & vocab, llama_token id) { + GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE); + return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_USER_DEFINED; +} + +static bool llama_is_unused_token(const llama_vocab & vocab, llama_token id) { + GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE); + return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_UNUSED; +} + +static uint8_t llama_token_to_byte(const llama_vocab & vocab, llama_token id) { + GGML_ASSERT(llama_vocab_get_type(vocab) != LLAMA_VOCAB_TYPE_NONE); + GGML_ASSERT(llama_is_byte_token(vocab, id)); + const auto & token_data = vocab.id_to_token.at(id); + switch (llama_vocab_get_type(vocab)) { + case LLAMA_VOCAB_TYPE_SPM: + case LLAMA_VOCAB_TYPE_UGM: { + auto buf = token_data.text.substr(3, 2); + return strtol(buf.c_str(), NULL, 16); + } + case LLAMA_VOCAB_TYPE_BPE: { + GGML_ASSERT(false); + return unicode_utf8_to_byte(token_data.text); // TODO: why is this here after GGML_ASSERT? + } + case LLAMA_VOCAB_TYPE_WPM: { + GGML_ASSERT(false); + } + default: + GGML_ASSERT(false); + } +} + +static void llama_escape_whitespace(std::string & text) { + replace_all(text, " ", "\xe2\x96\x81"); +} + +static void llama_unescape_whitespace(std::string & word) { + replace_all(word, "\xe2\x96\x81", " "); +} + +struct llm_symbol { + using index = int; + index prev; + index next; + const char * text; + size_t n; +}; + +static_assert(std::is_trivially_copyable::value, "llm_symbol is not trivially copyable"); + +// +// SPM tokenizer +// original implementation: +// https://github.com/ggerganov/llama.cpp/commit/074bea2eb1f1349a0118239c4152914aecaa1be4 +// + +struct llm_bigram_spm { + struct comparator { + bool operator()(llm_bigram_spm & l, llm_bigram_spm & r) { + return (l.score < r.score) || (l.score == r.score && l.left > r.left); + } + }; + using queue_storage = std::vector; + using queue = std::priority_queue; + llm_symbol::index left; + llm_symbol::index right; + float score; + size_t size; +}; + +struct llm_tokenizer_spm { + llm_tokenizer_spm(const llama_vocab & vocab) : vocab(vocab) {} + + void tokenize(const std::string & text, std::vector & output) { + // split string into utf8 chars + int index = 0; + size_t offs = 0; + while (offs < text.size()) { + llm_symbol sym; + size_t len = unicode_len_utf8(text[offs]); + sym.text = text.c_str() + offs; + sym.n = std::min(len, text.size() - offs); + offs += sym.n; + sym.prev = index - 1; + sym.next = offs == text.size() ? -1 : index + 1; + index++; + symbols.emplace_back(sym); + } + + // seed the work queue with all possible 2-character tokens. + for (size_t i = 1; i < symbols.size(); ++i) { + try_add_bigram(i - 1, i); + } + + // keep substituting the highest frequency pairs for as long as we can. + while (!work_queue.empty()) { + auto bigram = work_queue.top(); + work_queue.pop(); + + auto & left_sym = symbols[bigram.left]; + auto & right_sym = symbols[bigram.right]; + + // if one of the symbols already got merged, skip it. + if (left_sym.n == 0 || right_sym.n == 0 || + left_sym.n + right_sym.n != bigram.size) { + continue; + } + + // merge the right sym into the left one + left_sym.n += right_sym.n; + right_sym.n = 0; + + //LLAMA_LOG_INFO("left = '%*s' size = %zu\n", (int) left_sym.n, left_sym.text, bigram.size); + + // remove the right sym from the chain + left_sym.next = right_sym.next; + if (right_sym.next >= 0) { + symbols[right_sym.next].prev = bigram.left; + } + + // find more substitutions + try_add_bigram(left_sym.prev, bigram.left); + try_add_bigram(bigram.left, left_sym.next); + } + + for (int i = 0; i != -1; i = symbols[i].next) { + auto & symbol = symbols[i]; + resegment(symbol, output); + } + } + +private: + void resegment(llm_symbol & symbol, std::vector & output) { + auto text = std::string(symbol.text, symbol.n); + auto token = vocab.token_to_id.find(text); + + // Do we need to support is_unused? + if (token != vocab.token_to_id.end()) { + output.push_back((*token).second); + return; + } + + const auto p = rev_merge.find(text); + + if (p == rev_merge.end()) { + // output any symbols that did not form tokens as bytes. + output.reserve(output.size() + symbol.n); + for (int j = 0; j < (int)symbol.n; ++j) { + llama_vocab::id token_id = llama_byte_to_token_impl(vocab, symbol.text[j]); + output.push_back(token_id); + } + return; + } + + resegment(symbols[p->second.first], output); + resegment(symbols[p->second.second], output); + } + + void try_add_bigram(int left, int right) { + if (left == -1 || right == -1) { + return; + } + + const std::string text = std::string(symbols[left].text, symbols[left].n + symbols[right].n); + auto token = vocab.token_to_id.find(text); + + if (token == vocab.token_to_id.end()) { + return; + } + + if (static_cast((*token).second) >= vocab.id_to_token.size()) { + return; + } + + const auto & tok_data = vocab.id_to_token[(*token).second]; + + llm_bigram_spm bigram; + bigram.left = left; + bigram.right = right; + bigram.score = tok_data.score; + bigram.size = text.size(); + + work_queue.push(bigram); + + // Do we need to support is_unused? + rev_merge[text] = std::make_pair(left, right); + } + + const llama_vocab & vocab; + + std::vector symbols; + llm_bigram_spm::queue work_queue; + + std::map> rev_merge; +}; + +// +// BPE tokenizer +// adapted from https://github.com/cmp-nct/ggllm.cpp [MIT License] +// tried to simplify unicode stuff, so most likely does not work 100% correctly! +// + +// TODO: there are a lot of common parts between spm and bpe tokenizers, should be refactored and reused + +struct llm_bigram_bpe { + struct comparator { + bool operator()(const llm_bigram_bpe & l, const llm_bigram_bpe & r) const { + return l.rank > r.rank || (l.rank == r.rank && l.left > r.left); + } + }; + + using queue_storage = std::vector; + using queue = std::priority_queue; + llm_symbol::index left; + llm_symbol::index right; + std::string text; + int rank; + size_t size; +}; + +struct llm_tokenizer_bpe { + llm_tokenizer_bpe(const llama_vocab & vocab): vocab(vocab) { + GGML_ASSERT(vocab.type == LLAMA_VOCAB_TYPE_BPE); + switch (vocab.type_pre) { + case LLAMA_VOCAB_PRE_TYPE_LLAMA3: + regex_exprs = { + // original regex from tokenizer.json + //"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + + // adapted: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2080233989 + "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_DBRX: + case LLAMA_VOCAB_PRE_TYPE_SMAUG: + regex_exprs = { + // same as llama3 + "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM: + regex_exprs = { + "[\r\n]", + "\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿDŽ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿff-stﬓ-ﬗＡ-Ｚａ-ｚ𐐀-𐑏𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+", + "\\s?[!-/:-~！-／：-～‘-‟　-。]+", + "\\s+$", + "[一-龥ࠀ-一가-퟿]+", + "\\p{N}+", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER: + regex_exprs = { + "[\r\n]", + "\\s?\\p{L}+", + "\\s?\\p{P}+", + "[一-龥ࠀ-一가-퟿]+", + "\\p{N}", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_FALCON: + regex_exprs = { + "[\\p{P}\\$\\+<=>\\^~\\|`]+", + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + "[0-9][0-9][0-9]", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_STARCODER: + case LLAMA_VOCAB_PRE_TYPE_REFACT: + case LLAMA_VOCAB_PRE_TYPE_COMMAND_R: + case LLAMA_VOCAB_PRE_TYPE_SMOLLM: + case LLAMA_VOCAB_PRE_TYPE_CODESHELL: + regex_exprs = { + "\\p{N}", + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_GPT2: + case LLAMA_VOCAB_PRE_TYPE_MPT: + case LLAMA_VOCAB_PRE_TYPE_OLMO: + case LLAMA_VOCAB_PRE_TYPE_JAIS: + regex_exprs = { + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_STABLELM2: + case LLAMA_VOCAB_PRE_TYPE_QWEN2: + regex_exprs = { + // original regex from tokenizer.json + // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+" + "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_PORO: + regex_exprs = { + " ?[^(\\s|.,!?…。，、।۔،)]+", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_CHATGLM4: + regex_exprs = { + "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_VIKING: + regex_exprs = { + " ?[^(\\s|.,!?…。，、।۔،)]+", + "\\p{N}", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_TEKKEN: + // original regex from tokenizer.json + // "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+" + regex_exprs = { + "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }; + break; + default: + // default regex for BPE tokenization pre-processing + regex_exprs = { + "[\\p{P}\\$\\+<=>\\^~\\|]+", + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + "\\p{N}+", + "[0-9][0-9][0-9]", + }; + break; + } + } + + void append(const llama_vocab::id token_id, std::vector & output) const { + output.push_back(token_id); + } + + bool append_bos(std::vector & output) const { + if (vocab.tokenizer_add_bos) { + GGML_ASSERT(vocab.special_bos_id != -1); + output.push_back(vocab.special_bos_id); + return true; + } + return false; + } + + bool append_eos(std::vector & output) const { + if (vocab.tokenizer_add_eos) { + GGML_ASSERT(vocab.special_eos_id != -1); + output.push_back(vocab.special_eos_id); + return true; + } + return false; + } + + void check_double_bos_eos(const std::vector & output) const { + if (vocab.tokenizer_add_bos && output.size() >= 2 && output[1] == vocab.special_bos_id) { + LLAMA_LOG_WARN( + "%s: Added a BOS token to the prompt as specified by the model but the prompt " + "also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. " + "Are you sure this is what you want?\n", __FUNCTION__); + } + if (vocab.tokenizer_add_eos && output.size() >= 2 && *(output.end()-2) == vocab.special_eos_id) { + LLAMA_LOG_WARN( + "%s: Added a EOS token to the prompt as specified by the model but the prompt " + "also ends with a EOS token. So now the final prompt ends with 2 EOS tokens. " + "Are you sure this is what you want?\n", __FUNCTION__); + } + } + + void tokenize(const std::string & text, std::vector & output) { + int final_prev_index = -1; + + const auto word_collection = unicode_regex_split(text, regex_exprs); + + symbols_final.clear(); + + for (auto & word : word_collection) { + work_queue = llm_bigram_bpe::queue(); + symbols.clear(); + + int index = 0; + size_t offset = 0; + + if (vocab.tokenizer_ignore_merges && vocab.token_to_id.find(word) != vocab.token_to_id.end()) { + symbols.emplace_back(llm_symbol{-1, -1, word.c_str(), word.size()}); + offset = word.size(); + } + + while (offset < word.size()) { + llm_symbol sym; + size_t char_len = std::min(word.size() - offset, (size_t) unicode_len_utf8(word[offset])); + sym.text = word.c_str() + offset; + sym.n = char_len; + offset += sym.n; + sym.prev = index - 1; + sym.next = offset == word.size() ? -1 : index + 1; + index++; + symbols.emplace_back(sym); + } + for (size_t i = 1; i < symbols.size(); ++i) { + add_new_bigram(i - 1, i); + } + + // build token(s) + while (!work_queue.empty()) { + auto bigram = work_queue.top(); + work_queue.pop(); + + auto & left_symbol = symbols[bigram.left]; + auto & right_symbol = symbols[bigram.right]; + + if (left_symbol.n == 0 || right_symbol.n == 0) { + continue; + } + std::string left_token = std::string(left_symbol.text, left_symbol.n); + std::string right_token = std::string(right_symbol.text, right_symbol.n); + if (left_token + right_token != bigram.text) { + continue; // Skip this bigram if it's outdated + } + + // merge the right sym into the left one + left_symbol.n += right_symbol.n; + right_symbol.n = 0; + + // remove the right sym from the chain + left_symbol.next = right_symbol.next; + if (right_symbol.next >= 0) { + symbols[right_symbol.next].prev = bigram.left; + } + + add_new_bigram(left_symbol.prev, bigram.left); // left side of current symbol + add_new_bigram(bigram.left, left_symbol.next); // right side of current symbol + } + + // add the finished tokens to the final list keeping correct order for next and prev + for (auto & sym : symbols) { + if (sym.n > 0) { + sym.prev = final_prev_index; + sym.next = -1; + if (final_prev_index != -1) { + symbols_final[final_prev_index].next = symbols_final.size(); + } + symbols_final.emplace_back(sym); + final_prev_index = symbols_final.size() - 1; + } + } + } + + symbols = symbols_final; + + if (!symbols.empty()) { + for (int i = 0; i != -1; i = symbols[i].next) { + auto & symbol = symbols[i]; + if (symbol.n == 0) { + continue; + } + + const std::string str = std::string(symbol.text, symbol.n); + const auto token = vocab.token_to_id.find(str); + + if (token == vocab.token_to_id.end()) { + for (auto j = str.begin(); j != str.end(); ++j) { + std::string byte_str(1, *j); + auto token_multibyte = vocab.token_to_id.find(byte_str); + if (token_multibyte != vocab.token_to_id.end()) { + output.push_back(token_multibyte->second); + } + } + } else { + output.push_back((*token).second); + } + } + } + } + +private: + void add_new_bigram(int left, int right) { + if (left == -1 || right == -1) { + return; + } + + std::string left_token = std::string(symbols[left].text, symbols[left].n); + std::string right_token = std::string(symbols[right].text, symbols[right].n); + + int rank_found = -1; + + rank_found = vocab.find_bpe_rank(left_token, right_token); + + if (rank_found < 0) { + return; + } + + llm_bigram_bpe bigram; + + bigram.left = left; + bigram.right = right; + bigram.text = left_token + right_token; + bigram.size = left_token.size() + right_token.size(); + bigram.rank = rank_found; + + work_queue.push(bigram); + } + + const llama_vocab & vocab; + + std::vector regex_exprs; + + std::vector symbols; + std::vector symbols_final; + + llm_bigram_bpe::queue work_queue; +}; + +// +// WPM tokenizer +// + +struct llm_tokenizer_wpm { + llm_tokenizer_wpm(const llama_vocab & vocab): vocab(vocab) {} + + void tokenize(const std::string & text, std::vector & output) const { + const auto & token_map = vocab.token_to_id; + + // normalize and split by whitespace + std::vector words = preprocess(text); + + // bos token prepended already + + // find the longest tokens that form the words + for (const std::string & word : words) { + // skip empty words + if (word.size() == 0) { + continue; + } + + // prepend phantom space + const std::string word1 = "\xe2\x96\x81" + word; + const int n = word1.size(); + + const size_t current_tokens = output.size(); + + // we're at the start of a new word + // move through character position in word + for (int i = 0; i < n; ++i) { + // loop through possible match length + bool match = false; + for (int j = std::min(n, i + vocab.max_token_len + 1); j > i; j--) { + auto it = token_map.find(word1.substr(i, j - i)); + if (it != token_map.end()) { + output.push_back(it->second); + match = true; + i = j - 1; + break; + } + } + + if (!match) { // discard all + output.resize(current_tokens); + break; // and discard next tokens + } + } + + // we didn't find any matches for this word + if (current_tokens == output.size()) { + output.push_back(vocab.special_unk_id); + } + } + } + + // TODO: reduce string copies by using cpts_offs array + std::vector preprocess(const std::string & text) const { + const std::vector cpts_nfd = unicode_cpts_normalize_nfd(unicode_cpts_from_utf8(text)); + std::vector words(1, ""); + + for (const uint32_t cpt : cpts_nfd) { + const auto flags = unicode_cpt_flags(cpt); + + if (flags.is_whitespace) { + if (words.back().size()) { // finish previous word if any + words.emplace_back(); + } + continue; + } + + assert (!flags.is_separator); + if (cpt == 0 || cpt == 0xFFFD || flags.is_control) { + continue; + } + + const std::string s = unicode_cpt_to_utf8(unicode_tolower(cpt)); + if (flags.is_punctuation || ( cpt < 0x7F && flags.is_symbol ) || is_chinese_char(cpt)) { + if (words.back().size()) { // finish previous word if any + words.emplace_back(); + } + words.back() = s; // single char word + words.emplace_back(); // start a new word + } else { + words.back() += s; // append char to word + } + } + + if (!words.back().size()) { + words.pop_back(); + } + + return words; + } + + static bool is_chinese_char(uint32_t cpt) { + return + (cpt >= 0x04E00 && cpt <= 0x09FFF) || + (cpt >= 0x03400 && cpt <= 0x04DBF) || + (cpt >= 0x20000 && cpt <= 0x2A6DF) || + (cpt >= 0x2A700 && cpt <= 0x2B73F) || + (cpt >= 0x2B740 && cpt <= 0x2B81F) || + (cpt >= 0x2B920 && cpt <= 0x2CEAF) || // this should be 0x2B820 but in hf rust code it is 0x2B920 + (cpt >= 0x0F900 && cpt <= 0x0FAFF) || + (cpt >= 0x2F800 && cpt <= 0x2FA1F); + //(cpt >= 0x3000 && cpt <= 0x303F) || + //(cpt >= 0xFF00 && cpt <= 0xFFEF); + } + + const llama_vocab & vocab; +}; + +// +// UGM tokenizer +// + +struct llm_tokenizer_ugm { + llm_tokenizer_ugm(const llama_vocab & vocab) : vocab(vocab) { + if (vocab.precompiled_charsmap.size() > 0) { + size_t charsmap_offset = 0; + + // First four bytes of precompiled_charsmap contains length of binary + // blob containing XOR-compressed compact double array (XCDA) entries + uint32_t xcda_blob_size = *(const uint32_t *) &vocab.precompiled_charsmap[0]; + charsmap_offset += sizeof(xcda_blob_size); + if (xcda_blob_size + charsmap_offset >= vocab.precompiled_charsmap.size()) { + throw std::runtime_error("Index out of array bounds in precompiled charsmap!"); + } + + // Next xcda_blob_size bytes contain entries of XOR-compressed compact + // double array (XCDA). Each entry is bit-packed into a 32-bit integer. + xcda_array = (const uint32_t *) &vocab.precompiled_charsmap[charsmap_offset]; + xcda_array_size = xcda_blob_size / sizeof(uint32_t); + charsmap_offset += xcda_blob_size; + + // Remaining bytes of precompiled charsmap contain null-terminated + // replacement strings for prefixes matched by the XCDA. + prefix_replacements = &vocab.precompiled_charsmap[charsmap_offset]; + prefix_replacements_size = vocab.precompiled_charsmap.size() - charsmap_offset; + } + + for (unsigned int id = 0; id < vocab.id_to_token.size(); ++id) { + const auto &token_data = vocab.id_to_token[id]; + + if (llama_is_normal_token(vocab, id)) { + min_score = std::min(min_score, token_data.score); + max_score = std::max(max_score, token_data.score); + } + + if (llama_is_normal_token(vocab, id) || + llama_is_user_defined_token(vocab, id) || + llama_is_unused_token(vocab, id)) { + token_matcher.insert(token_data.text.data(), token_data.text.size(), id); + } + + if (llama_is_user_defined_token(vocab, id)) { + user_defined_token_matcher.insert(token_data.text.data(), token_data.text.size()); + } + } + + unknown_token_score = min_score - unknown_token_score_penalty; + } + + /* This implementation is based on SentencePiece optimized Viterbi algorithm for + * unigram language models. The general idea is to: + * - move along the input sequence in steps of one UTF code point, + * - at each step find all possible tokenizations of the prefix by + * traversing the tokens trie, + * - for each tokenization store the best one so far (by higher score) + * - use the position in sequence after given token as an index to store + * results + * - if there was no valid tokenization of the current UTF code point + * then use unknown token with additional score penalty + * After processing the whole sequence we backtrack from the end to get + * the best tokenization. + */ + void tokenize(const std::string & text, std::vector & output) { + // normalize the input first + std::string normalized; + normalize(text, &normalized); + size_t input_len = normalized.size(); + if (input_len == 0) { + return; + } + + // initialize score_sum to -FLT_MAX so it will be always lower than sums of token scores + std::vector tokenization_results(input_len + 1, {vocab.special_unk_id, 0, -FLT_MAX}); + // at the beginning tokenization score is zero + tokenization_results[0] = { vocab.special_unk_id, 0, 0 }; + + for (size_t input_offset = 0; input_offset < input_len;) { + size_t prefix_offset = input_offset; + // calculate how many code units are in the currently processed UTF code point + size_t n_utf8_code_units = std::min(unicode_len_utf8(normalized[input_offset]), input_len - input_offset); + + // traverse the token matcher trie to find a matching token + bool single_codepoint_token_found = false; + const struct best_tokenization & current_best = tokenization_results[input_offset]; + struct naive_trie * node = token_matcher.traverse(normalized[prefix_offset++]); + + while (prefix_offset <= input_len && node != NULL) { + // check if we found valid token in prefix + if (node->has_value) { + // check if it corresponds to the whole UTF code point + if (prefix_offset - input_offset == n_utf8_code_units) { + single_codepoint_token_found = true; + } + llama_token token_id = node->value; + const auto & token_data = vocab.id_to_token[token_id]; + + // we set the user-defined token scores to 0 to make them more likely to be selected + // (normal token scores are log probabilities, so they are negative) + // score type is double here to make tokenization results exactly + // the same as in the HF tokenizer using SentencePiece + const double token_score = llama_is_user_defined_token(vocab, token_id) ? 0.0 : token_data.score; + const double challenger_score = current_best.score_sum + token_score; + struct best_tokenization & current_champ = tokenization_results[prefix_offset]; + if (challenger_score > current_champ.score_sum) { + struct best_tokenization challenger = { token_id, input_offset, (float) challenger_score }; + current_champ = challenger; + } + } + node = node->traverse(normalized[prefix_offset++]); + } + + // if we didn't find a valid token corresponding to the whole UTF code point + // then use unknown token as the tokenization of this UTF code point + if (!single_codepoint_token_found) { + const double challenger_score = current_best.score_sum + unknown_token_score; + prefix_offset = input_offset + n_utf8_code_units; + struct best_tokenization & current_champ = tokenization_results[prefix_offset]; + if (challenger_score > current_champ.score_sum) { + struct best_tokenization challenger = { vocab.special_unk_id, input_offset, (float) challenger_score }; + current_champ = challenger; + } + } + + // move to the next UTF code point + input_offset += n_utf8_code_units; + } + + // now backtrack from the end to gather token ids of the best tokenization + // merge sequences of consecutive unknown tokens into single unknown tokens + bool is_prev_unknown = false; + for (struct best_tokenization & tokenization = tokenization_results[input_len]; ; tokenization = tokenization_results[tokenization.input_offset]) { + bool is_unknown = tokenization.token_id == vocab.special_unk_id; + if (!(is_prev_unknown && is_unknown)) { + output.push_back(tokenization.token_id); + } + if (tokenization.input_offset == 0) { + break; + } + is_prev_unknown = is_unknown; + } + + // reverse the output since we added tokens starting from the end of the input + std::reverse(output.begin(), output.end()); + } + +private: + const llama_vocab & vocab; + + // helper structure for returning normalization results + struct normalization_result { + const char * normalized; + size_t normalized_len; + size_t consumed_input; + }; + + void normalize(const std::string& input, std::string * normalized) { + normalized->clear(); + normalized->reserve(input.size() * 3); + + const std::string space = vocab.tokenizer_escape_whitespaces ? escaped_space : " "; + + bool shall_prepend_space = !vocab.tokenizer_treat_whitespace_as_suffix && vocab.tokenizer_add_space_prefix; + bool shall_append_space = vocab.tokenizer_treat_whitespace_as_suffix && vocab.tokenizer_add_space_prefix; + bool shall_merge_spaces = vocab.tokenizer_remove_extra_whitespaces; + + bool is_space_prepended = false; + bool processing_non_ws = false; + + size_t input_len = input.size(); + + for (size_t input_offset = 0; input_offset < input_len; ) { + auto norm_res = normalize_prefix(input, input_offset); + for (size_t i = 0; i < norm_res.normalized_len; i++) { + char c = norm_res.normalized[i]; + if (c != ' ') { + if (!processing_non_ws) { + processing_non_ws = true; + if ((shall_prepend_space && !is_space_prepended) || shall_merge_spaces) { + normalized->append(space); + is_space_prepended = true; + } + } + normalized->push_back(c); + } else { + if (processing_non_ws) { + processing_non_ws = false; + } + if (!shall_merge_spaces) { + normalized->append(space); + } + } + } + + input_offset += norm_res.consumed_input; + } + + if (shall_append_space) { + normalized->append(space); + } + } + + /* + * This structure is a view wrapper for XOR-compressed double array (XCDA) + * See Shunsuke Kanda (2018). Space- and Time-Efficient String Dictionaries. + * Eeach bit-packed entry contains: + * - BASE array value in bits 10-30 + * - LCHECK array value in bits 0-7 + * - LEAF array value in bit 9 + * Entries containing indexes of replacement sequences have set bit 31 + */ + struct xcda_array_view { + public: + xcda_array_view(const uint32_t * xcda_array, size_t xcda_array_size) : xcda_array(xcda_array), xcda_array_size(xcda_array_size) { + } + uint32_t get_base(size_t index) { + uint32_t packed_node = get_node(index); + return (packed_node >> 10) << ((packed_node & (1U << 9)) >> 6); + } + uint32_t get_lcheck(size_t index) { + uint32_t packed_node = get_node(index); + return packed_node & ((1U << 31) | 0xff); + } + bool get_leaf(size_t index) { + uint32_t packed_node = get_node(index); + return (packed_node >> 8) & 1; + } + uint32_t get_value(size_t index) { + uint32_t packed_node = get_node(index); + return packed_node & ((1U << 31) - 1); + } + private: + uint32_t get_node(size_t index) { + if (index > xcda_array_size) { + throw std::runtime_error("Index out of array bounds in XCDA array!"); + } + return xcda_array[index]; + } + const uint32_t * xcda_array; + size_t xcda_array_size; + }; + + struct normalization_result normalize_prefix(const std::string & input, size_t input_offset) { + if (input_offset == input.size()) { + return { &input[input_offset], 0, 0 }; + } + + // if input prefix matches some user-defined token return this token as normalization result + auto user_defined_token_match = user_defined_token_matcher.get_longest_prefix(&input[input_offset], input.size() - input_offset); + if (user_defined_token_match.second > 0) { + return { &input[input_offset], user_defined_token_match.second, user_defined_token_match.second }; + } + + size_t longest_prefix_length = 0; + size_t longest_prefix_offset = 0; + + if (xcda_array_size > 0) { + struct xcda_array_view xcda_view(xcda_array, xcda_array_size); + + // Find the longest normalized sequence matching the input prefix by walking + // the XOR-compressed compact double array (XCDA) starting from the root node + // We find the index of the next node by calculating BASE[s] ^ c where s is + // the index of the previous node and c is a numerical character value + uint32_t node_index = 0; + // get BASE of the root node + node_index = xcda_view.get_base(node_index); + for (size_t prefix_offset = input_offset; prefix_offset < input.size(); prefix_offset++) { + unsigned char c = input[prefix_offset]; + if (c == 0) { + break; + } + node_index ^= c; + // if value of LCHECK is not c it means that this is not a child of + // the previous node, so we stop matching + if (xcda_view.get_lcheck(node_index) != c) { + break; + } + bool is_leaf = xcda_view.get_leaf(node_index); + // get BASE of the current node + node_index ^= xcda_view.get_base(node_index); + // if LEAF of the current node is true, it means that its BASE points to the node + // containing index of replacement sequence for currently matched input prefix + if (is_leaf) + { + longest_prefix_length = prefix_offset - input_offset + 1; + // get index of replacement sequence for currently matched input prefix + longest_prefix_offset = xcda_view.get_value(node_index); + } + } + } + + if (longest_prefix_length > 0) { + // we have a match, so return the replacement sequence + if (longest_prefix_offset >= prefix_replacements_size) { + throw std::runtime_error("Index out of array bounds in precompiled charsmap!"); + } + const char * prefix_replacement = &prefix_replacements[longest_prefix_offset]; + return { prefix_replacement, strlen(prefix_replacement), longest_prefix_length }; + } else { + // check if the input prefix contains a valid sequence of UTF-8 code units + try { + // if yes, return this sequence unmodified + size_t prefix_offset = input_offset; + unicode_cpt_from_utf8(input, prefix_offset); + return { &input[input_offset], prefix_offset - input_offset, prefix_offset - input_offset }; + } catch (std::invalid_argument & /*ex*/) { + // if no, consume 1 byte and return U+FFFD - REPLACEMENT CHARACTER + return { "\xEF\xBF\xBD", 3, 1 }; + } + } + } + + // escaped space symbol - U+2581 (Lower One Eighth Block) + const std::string escaped_space = "\xE2\x96\x81"; + + const char * prefix_replacements = NULL; + size_t prefix_replacements_size = 0; + + const uint32_t * xcda_array = NULL; + size_t xcda_array_size = 0; + + struct naive_trie user_defined_token_matcher; + + // this structure stores the best tokenization so far at input_offset + struct best_tokenization { + llama_token token_id; + size_t input_offset; + float score_sum; + }; + + float min_score = FLT_MAX; + float max_score = -FLT_MAX; + + float unknown_token_score_penalty = 10.0; + float unknown_token_score; + + struct naive_trie token_matcher; +}; + +// +// (de-) tokenize +// + +typedef enum FRAGMENT_BUFFER_VARIANT_TYPE { + FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN, + FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT +} FRAGMENT_BUFFER_VARIANT_TYPE; + +struct fragment_buffer_variant { + fragment_buffer_variant(llama_vocab::id _token) + : + type(FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN), + token(_token), + raw_text(_dummy), + offset(0), + length(0) {} + + fragment_buffer_variant(const std::string & _raw_text, int64_t _offset, int64_t _length) + : + type(FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT), + token((llama_vocab::id) - 1), + raw_text(_raw_text), + offset(_offset), + length(_length){ + GGML_ASSERT(_offset >= 0); + GGML_ASSERT(_length >= 1); + GGML_ASSERT(offset + length <= raw_text.length()); + } + + const FRAGMENT_BUFFER_VARIANT_TYPE type; + const llama_vocab::id token; + const std::string _dummy; + const std::string & raw_text; + const uint64_t offset; + const uint64_t length; +}; + +// #define PRETOKENIZERDEBUG + +static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list & buffer, bool parse_special) { + // for each special token + for (const llama_vocab::id special_id : vocab.cache_special_tokens) { + const auto & data = vocab.id_to_token[special_id]; + const auto & special_token = data.text; + + if (!parse_special && (data.attr & (LLAMA_TOKEN_ATTR_CONTROL | LLAMA_TOKEN_ATTR_UNKNOWN))) { + // Ignore control and unknown tokens when parse_special == false + continue; + // User-defined tokens are still pre-tokenized before everything else + // ref: https://github.com/huggingface/tokenizers/blob/fdd26ba9a3f0c133427aab0423888cbde91362d7/tokenizers/src/tokenizer/mod.rs#L726 + // This is mostly relevant for neox-style tokenizers (mpt, olmo, stablelm, etc.) + } + + // for each text fragment + std::forward_list::iterator it = buffer.begin(); + while (it != buffer.end()) { + auto & fragment = (*it); + + // if a fragment is text ( not yet processed ) + if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) { + auto & raw_text = fragment.raw_text; + + auto raw_text_base_offset = fragment.offset; + auto raw_text_base_length = fragment.length; + + // loop over the text + while (true) { + // find the first occurrence of a given special token in this fragment + // passing offset argument only limit the "search area" but match coordinates + // are still relative to the source full raw_text + auto match = raw_text.find(special_token, raw_text_base_offset); + + // no occurrences found, stop processing this fragment for a given special token + if (match == std::string::npos) break; + + // check if match is within bounds of offset <-> length + if (match + special_token.length() > raw_text_base_offset + raw_text_base_length) break; + +#ifdef PRETOKENIZERDEBUG + LLAMA_LOG_WARN("FF: (%ld %ld %ld) '%s'\n", raw_text->length(), raw_text_base_offset, raw_text_base_length, raw_text->substr(raw_text_base_offset, raw_text_base_length).c_str()); +#endif + auto source = std::distance(buffer.begin(), it); + + // if match is further than base offset + // then we have some text to the left of it + if (match > raw_text_base_offset) { + // left + const int64_t left_reminder_offset = raw_text_base_offset + 0; + int64_t left_reminder_length = match - raw_text_base_offset; + + if (data.attr & LLAMA_TOKEN_ATTR_LSTRIP) { + while (left_reminder_length > 0 && isspace(raw_text[left_reminder_offset + left_reminder_length - 1])) { + left_reminder_length--; + } + } + + if (left_reminder_length > 0) { + buffer.emplace_after(it, raw_text, left_reminder_offset, left_reminder_length); + it++; + } + +#ifdef PRETOKENIZERDEBUG + LLAMA_LOG_WARN("FL: (%ld %ld) '%s'\n", left_reminder_offset, left_reminder_length, raw_text->substr(left_reminder_offset, left_reminder_length).c_str()); +#endif + } + + // special token + buffer.emplace_after(it, special_id); + it++; + + // right + if (match + special_token.length() < raw_text_base_offset + raw_text_base_length) { + int64_t right_reminder_offset = match + special_token.length(); + int64_t right_reminder_length = raw_text_base_length - ((match - raw_text_base_offset) + special_token.length()); + + if (data.attr & LLAMA_TOKEN_ATTR_RSTRIP) { + while (right_reminder_length > 0 && isspace(raw_text[right_reminder_offset])) { + right_reminder_offset++; + right_reminder_length--; + } + } + + if (right_reminder_length > 0) { + buffer.emplace_after(it, raw_text, right_reminder_offset, right_reminder_length); + it++; + } + +#ifdef PRETOKENIZERDEBUG + LLAMA_LOG_WARN("FR: (%ld %ld) '%s'\n", right_reminder_offset, right_reminder_length, raw_text->substr(right_reminder_offset, right_reminder_length).c_str()); +#endif + + if (source == 0) { + buffer.erase_after(buffer.before_begin()); + } else { + buffer.erase_after(std::next(buffer.begin(), (source-1))); + } + + // repeat for the right side + raw_text_base_offset = right_reminder_offset; + raw_text_base_length = right_reminder_length; + +#ifdef PRETOKENIZERDEBUG + LLAMA_LOG_WARN("RR: (%ld %ld) '%s'\n", raw_text_base_offset, raw_text_base_length, raw_text->substr(raw_text_base_offset, raw_text_base_length).c_str()); +#endif + } else { + if (source == 0) { + buffer.erase_after(buffer.before_begin()); + } else { + buffer.erase_after(std::next(buffer.begin(), (source-1))); + } + break; + } + } + } + it++; + } + } +} + +std::vector llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool add_special, bool parse_special) { + std::vector output; + std::forward_list fragment_buffer; + + if (!raw_text.empty()) { + fragment_buffer.emplace_front(raw_text, 0, raw_text.length()); + tokenizer_st_partition(vocab, fragment_buffer, parse_special); + } + + switch (vocab.type) { + case LLAMA_VOCAB_TYPE_SPM: + { + // OG tokenizer behavior: + // + // tokenizer.encode('', add_special_tokens=True) returns [1] + // tokenizer.encode('', add_special_tokens=False) returns [] + + bool is_prev_special = true; // prefix with space if first token + + if (add_special && vocab.tokenizer_add_bos) { + GGML_ASSERT(vocab.special_bos_id != -1); + output.push_back(vocab.special_bos_id); + is_prev_special = true; + } + + for (const auto & fragment : fragment_buffer) { + if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) { + auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length); + + // prefix with space if previous is special + if (vocab.tokenizer_add_space_prefix && is_prev_special) { + raw_text = " " + raw_text; + } + +#ifdef PRETOKENIZERDEBUG + LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str()); +#endif + llm_tokenizer_spm tokenizer(vocab); + llama_escape_whitespace(raw_text); + tokenizer.tokenize(raw_text, output); + is_prev_special = false; + } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN) + output.push_back(fragment.token); + is_prev_special = true; + } + } + + if (add_special && vocab.tokenizer_add_bos && output.size() >= 2 && output[1] == vocab.special_bos_id) { + LLAMA_LOG_WARN( + "%s: Added a BOS token to the prompt as specified by the model but the prompt " + "also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. " + "Are you sure this is what you want?\n", __FUNCTION__); + } + + if (add_special && vocab.tokenizer_add_eos) { + GGML_ASSERT(vocab.special_eos_id != -1); + output.push_back(vocab.special_eos_id); + } + } break; + case LLAMA_VOCAB_TYPE_BPE: + { + llm_tokenizer_bpe tokenizer(vocab); + + if (add_special) { + tokenizer.append_bos(output); + } + for (const auto & fragment : fragment_buffer) { + if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) { + auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length); + +#ifdef PRETOKENIZERDEBUG + LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str()); +#endif + tokenizer.tokenize(raw_text, output); + } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN) + tokenizer.append(fragment.token, output); + } + } + + if (add_special) { + tokenizer.append_eos(output); + tokenizer.check_double_bos_eos(output); + } + } break; + case LLAMA_VOCAB_TYPE_WPM: + { + if (add_special) { + GGML_ASSERT(vocab.special_cls_id != -1); + output.push_back(vocab.special_cls_id); + } + + llm_tokenizer_wpm tokenizer(vocab); + + for (const auto & fragment : fragment_buffer) { + if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) { + auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length); + +#ifdef PRETOKENIZERDEBUG + LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str()); +#endif + tokenizer.tokenize(raw_text, output); + } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN) + output.push_back(fragment.token); + } + } + + if (add_special) { + GGML_ASSERT(vocab.special_sep_id != -1); + output.push_back(vocab.special_sep_id); + } + } break; + case LLAMA_VOCAB_TYPE_UGM: + { + llm_tokenizer_ugm tokenizer(vocab); + + if (add_special && vocab.tokenizer_add_bos != 0) { + GGML_ASSERT(vocab.special_bos_id != -1); + output.push_back(vocab.special_bos_id); + } + + for (const auto & fragment : fragment_buffer) { + if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) { + auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length); +#ifdef PRETOKENIZERDEBUG + LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str()); +#endif + tokenizer.tokenize(raw_text, output); + } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN) + output.push_back(fragment.token); + } + } + + if (add_special && vocab.tokenizer_add_bos != 0 && output.size() >= 2 && output[1] == vocab.special_bos_id) { + LLAMA_LOG_WARN( + "%s: Added a BOS token to the prompt as specified by the model but the prompt " + "also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. " + "Are you sure this is what you want?\n", __FUNCTION__); + } + + if (add_special && vocab.tokenizer_add_eos == 1) { + GGML_ASSERT(vocab.special_eos_id != -1); + output.push_back(vocab.special_eos_id); + } + } break; + case LLAMA_VOCAB_TYPE_NONE: + GGML_ASSERT(false); + } + + return output; +} + +llama_token llama_byte_to_token_impl(const llama_vocab & vocab, uint8_t ch) { + GGML_ASSERT(llama_vocab_get_type(vocab) != LLAMA_VOCAB_TYPE_NONE); + static const char * hex = "0123456789ABCDEF"; + switch (llama_vocab_get_type(vocab)) { + case LLAMA_VOCAB_TYPE_SPM: + case LLAMA_VOCAB_TYPE_UGM: { + const char buf[7] = { '<', '0', 'x', hex[ch >> 4], hex[ch & 15], '>', 0 }; + auto token = vocab.token_to_id.find(buf); + if (token != vocab.token_to_id.end()) { + return (*token).second; + } + // Try to fall back to just the byte as a string + const char buf2[2] = { (char)ch, 0 }; + return vocab.token_to_id.at(buf2); + } + case LLAMA_VOCAB_TYPE_WPM: + case LLAMA_VOCAB_TYPE_BPE: { + return vocab.token_to_id.at(unicode_byte_to_utf8(ch)); + } + default: + GGML_ASSERT(false); + } +} + +const char * llama_token_get_text_impl(const struct llama_vocab & vocab, llama_token token) { + GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE); + return vocab.id_to_token[token].text.c_str(); +} + +float llama_token_get_score_impl(const struct llama_vocab & vocab, llama_token token) { + GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE); + return vocab.id_to_token[token].score; +} + +llama_token_attr llama_token_get_attr_impl(const struct llama_vocab & vocab, llama_token token) { + GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE); + return vocab.id_to_token[token].attr; +} + +bool llama_token_is_eog_impl(const struct llama_vocab & vocab, llama_token token) { + return token != -1 && ( + token == llama_token_eos_impl(vocab) || + token == llama_token_eot_impl(vocab) + ); +} + +bool llama_token_is_control_impl(const struct llama_vocab & vocab, llama_token token) { + return llama_is_control_token(vocab, token); +} + +llama_token llama_token_bos_impl(const struct llama_vocab & vocab) { + return vocab.special_bos_id; +} + +llama_token llama_token_eos_impl(const struct llama_vocab & vocab) { + return vocab.special_eos_id; +} + +llama_token llama_token_cls_impl(const struct llama_vocab & vocab) { + return vocab.special_cls_id; +} + +llama_token llama_token_sep_impl(const struct llama_vocab & vocab) { + return vocab.special_sep_id; +} + +llama_token llama_token_nl_impl(const struct llama_vocab & vocab) { + return vocab.linefeed_id; +} + +llama_token llama_token_pad_impl(const struct llama_vocab & vocab) { + return vocab.special_pad_id; +} + +int32_t llama_add_bos_token_impl(const struct llama_vocab & vocab) { + return vocab.tokenizer_add_bos; +} + +int32_t llama_add_eos_token_impl(const struct llama_vocab & vocab) { + return vocab.tokenizer_add_eos; +} + +llama_token llama_token_prefix_impl(const struct llama_vocab & vocab) { + return vocab.special_prefix_id; +} + +llama_token llama_token_middle_impl(const struct llama_vocab & vocab) { + return vocab.special_middle_id; +} + +llama_token llama_token_suffix_impl(const struct llama_vocab & vocab) { + return vocab.special_suffix_id; +} + +llama_token llama_token_eot_impl(const struct llama_vocab & vocab) { + return vocab.special_eot_id; +} + +int32_t llama_tokenize_impl( + const struct llama_vocab & vocab, + const char * text, + int32_t text_len, + llama_token * tokens, + int32_t n_tokens_max, + bool add_special, + bool parse_special) { + auto res = llama_tokenize_internal(vocab, std::string(text, text_len), add_special, parse_special); + if (n_tokens_max < (int) res.size()) { + // LLAMA_LOG_ERROR("%s: too many tokens\n", __func__); + return -((int) res.size()); + } + + for (size_t i = 0; i < res.size(); i++) { + tokens[i] = res[i]; + } + + return res.size(); +} + +static std::string llama_decode_text(const std::string & text) { + std::string decoded_text; + + const auto cpts = unicode_cpts_from_utf8(text); + for (const auto cpt : cpts) { + const auto utf8 = unicode_cpt_to_utf8(cpt); + try { + decoded_text += unicode_utf8_to_byte(utf8); + } catch (const std::out_of_range & /*e*/) { + decoded_text += "[UNK_BYTE_0x"; + for (const auto c : utf8) { + decoded_text += format("%02x", (uint8_t) c); + } + decoded_text += text + "]"; + } + } + + return decoded_text; +} + +// does not write null-terminator to buf +int32_t llama_token_to_piece_impl(const struct llama_vocab & vocab, llama_token token, char * buf, int32_t length, int32_t lstrip, bool special) { + // ref: https://github.com/ggerganov/llama.cpp/pull/7587#discussion_r1620983843 + static const int attr_special = LLAMA_TOKEN_ATTR_UNKNOWN | LLAMA_TOKEN_ATTR_CONTROL; + const llama_token_attr attr = llama_token_get_attr_impl(vocab, token); + if (!special && (attr & attr_special)) { + return 0; + } + + // copy piece chars to output text buffer + // skip up to 'lstrip' leading spaces before copying + auto _try_copy = [=] (const char * token, size_t size) -> int32_t { + for (int32_t i = 0; i < lstrip && size && *token == ' '; ++i) { + token++; + size--; + } + if (length < (int32_t)size) { + return -(int32_t) size; + } + memcpy(buf, token, size); + return (int32_t) size; + }; + + // if we have a cache - use it + { + const auto & cache = vocab.cache_token_to_piece; + + if (!cache.empty()) { + const auto & result = cache.at(token); + return _try_copy(result.data(), result.size()); + } + } + + if (0 <= token && token < (int32_t) vocab.id_to_token.size()) { + const std::string & token_text = vocab.id_to_token[token].text; + switch (llama_vocab_get_type(vocab)) { + case LLAMA_VOCAB_TYPE_WPM: + case LLAMA_VOCAB_TYPE_SPM: + case LLAMA_VOCAB_TYPE_UGM: { + // NOTE: we accept all unsupported token types, + // suppressing them like CONTROL tokens. + if (attr & (attr_special | LLAMA_TOKEN_ATTR_USER_DEFINED)) { + return _try_copy(token_text.data(), token_text.size()); + } else if (attr & LLAMA_TOKEN_ATTR_NORMAL) { + std::string result = token_text; + llama_unescape_whitespace(result); + return _try_copy(result.data(), result.size()); + } else if (attr & LLAMA_TOKEN_ATTR_BYTE) { + char byte = (char) llama_token_to_byte(vocab, token); + return _try_copy((char*) &byte, 1); + } + break; + } + case LLAMA_VOCAB_TYPE_BPE: { + // NOTE: we accept all unsupported token types, + // suppressing them like CONTROL tokens. + if (attr & (attr_special | LLAMA_TOKEN_ATTR_USER_DEFINED)) { + return _try_copy(token_text.data(), token_text.size()); + } else if (attr & LLAMA_TOKEN_ATTR_NORMAL) { + std::string result = llama_decode_text(token_text); + return _try_copy(result.data(), result.size()); + } + break; + } + default: + GGML_ASSERT(false); + } + } + + return 0; +} + +int32_t llama_detokenize_impl( + const struct llama_vocab & vocab, + const llama_token * tokens, + int32_t n_tokens, + char * text, + int32_t text_len_max, + bool remove_special, + bool unparse_special) { + int32_t avail = text_len_max; + int32_t total = 0; + + // remove the leading space + bool remove_space = vocab.tokenizer_add_space_prefix; + + if (remove_special && vocab.tokenizer_add_bos) { + if (n_tokens > 0 && tokens[0] == vocab.special_bos_id) { + remove_space = false; + n_tokens--; + tokens++; + } + } + + if (remove_special && vocab.tokenizer_add_eos) { + if (n_tokens > 0 && tokens[n_tokens-1] == vocab.special_eos_id) { + n_tokens--; + } + } + + for (int32_t i = 0; i < n_tokens; ++i) { + GGML_ASSERT(avail >= 0); + int32_t n_chars = llama_token_to_piece_impl(vocab, tokens[i], text, avail, remove_space, unparse_special); + remove_space = false; + if (n_chars < 0) { + avail = 0; + total -= n_chars; + } else if (n_chars > 0) { + avail -= n_chars; + text += n_chars; + total += n_chars; + } + } + + if (total > text_len_max) { + return -total; + } + + if (vocab.tokenizer_clean_spaces) { + text -= total; // restart text + + // first pass: characters ?!., //TODO: where do these characters come from? + const int32_t total1 = total; + total = total ? 1 : 0; + for (int32_t i = 1; i < total1; ++i) { + const char x = text[i]; + if (text[i - 1] == ' ') { + if (x == '?' || x == '!' || x == '.' || x == ',') { // " ?", " !", " .", " ," + total--; // remove space + } + } + text[total++] = x; + } + + // second pass: strip single apostrophe between spaces + const int32_t total2 = total; + total = total ? 1 : 0; + for (int32_t i = 1; i < total2; ++i) { + const char x = text[i]; + if (x == '\'' && i + 1 < total2 && text[i - 1] == ' ' && text[i + 1] == ' ') { // " ' " + total--; // remove prev space + text[++i] = '\0'; // remove next space + } + text[total++] = x; + } + + // third pass: apostrophe contractions //NOTE: this makes sense? + const int32_t total3 = total; + total = total ? 1 : 0; + for (int32_t i = 1; i < total3; ++i) { + const char x = text[i]; + if (text[i - 1] == ' ') { + if (x == '\'' && i + 1 < total3) { + const char x1 = text[i + 1]; + if (x1 == 't' || x1 == 'd') { // " 't", " 'd" + //total--; // remove space + } else if (x1 == 's' || x1 == 'm') { // " 's", " 'm" + total--; // remove space + } else if (i + 2 < total3) { + const char x2 = text[i + 2]; + if ((x1 == 'l' && x2 == 'l')) { // " 'll" + //total--; // remove space + } else if ((x1 == 'r' && x2 == 'e') || (x1 == 'v' && x2 == 'e')) { // " 're", " 've" + total--; // remove space + } else { + //total--; // remove space + } + } else { + //total--; // remove space + } + } + } + text[total++] = x; + } + } + + return total <= text_len_max ? total : -total; +} diff --git a/src/llama-vocab.h b/src/llama-vocab.h new file mode 100644 index 00000000..30b565d5 --- /dev/null +++ b/src/llama-vocab.h @@ -0,0 +1,130 @@ +#pragma once + +#include "llama-impl.h" + +#include +#include +#include +#include + +struct llama_vocab { + using id = llama_token; + using token = std::string; + using tattr = llama_token_attr; + + struct token_data { + token text; + float score; + tattr attr; + }; + + enum llama_vocab_type type = LLAMA_VOCAB_TYPE_SPM; + enum llama_vocab_pre_type type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; + + int max_token_len = 0; // used for optimizing longest token search + + std::unordered_map token_to_id; + std::vector id_to_token; + + std::vector cache_special_tokens; + std::vector cache_token_to_piece; // llama_token_to_piece(special = true); + + std::map, int> bpe_ranks; + + // default LLaMA special tokens + id special_bos_id = 1; + id special_eos_id = 2; + id special_unk_id = 0; + id special_sep_id = -1; + id special_pad_id = -1; + id special_cls_id = -1; + id special_mask_id = -1; + + id linefeed_id = 13; + id special_prefix_id = -1; + id special_suffix_id = -1; + id special_middle_id = -1; + id special_eot_id = -1; // TODO: move above after "eos_id", and here add "file separator" token + + // tokenizer flags + bool tokenizer_add_space_prefix = false; + bool tokenizer_add_bos = false; + bool tokenizer_add_eos = false; + bool tokenizer_ignore_merges = false; + bool tokenizer_clean_spaces = false; // clean_up_tokenization_spaces + bool tokenizer_remove_extra_whitespaces = false; + bool tokenizer_escape_whitespaces = true; + bool tokenizer_treat_whitespace_as_suffix = false; + + std::vector precompiled_charsmap; + + int find_bpe_rank(const std::string & token_left, const std::string & token_right) const; +}; + +const struct llama_vocab * llama_get_vocab(const struct llama_context * ctx); + +// +// internal API +// + +// TODO: rename to llama_tokenize_impl +// TODO: This should probably be in llama.h +std::vector llama_tokenize_internal( + const llama_vocab & vocab, + std::string raw_text, + bool add_special, + bool parse_special = false); + +llama_token llama_byte_to_token_impl(const llama_vocab & vocab, uint8_t ch); + +const char * llama_token_get_text_impl(const struct llama_vocab & vocab, llama_token token); + +float llama_token_get_score_impl(const struct llama_vocab & vocab, llama_token token); + +llama_token_attr llama_token_get_attr_impl(const struct llama_vocab & vocab, llama_token token); + +bool llama_token_is_eog_impl(const struct llama_vocab & vocab, llama_token token); + +bool llama_token_is_control_impl(const struct llama_vocab & vocab, llama_token token); + +llama_token llama_token_bos_impl(const struct llama_vocab & vocab); +llama_token llama_token_eos_impl(const struct llama_vocab & vocab); +llama_token llama_token_cls_impl(const struct llama_vocab & vocab); +llama_token llama_token_sep_impl(const struct llama_vocab & vocab); +llama_token llama_token_nl_impl (const struct llama_vocab & vocab); +llama_token llama_token_pad_impl(const struct llama_vocab & vocab); + +int32_t llama_add_bos_token_impl(const struct llama_vocab & vocab); +int32_t llama_add_eos_token_impl(const struct llama_vocab & vocab); + +llama_token llama_token_prefix_impl(const struct llama_vocab & vocab); +llama_token llama_token_middle_impl(const struct llama_vocab & vocab); +llama_token llama_token_suffix_impl(const struct llama_vocab & vocab); +llama_token llama_token_eot_impl (const struct llama_vocab & vocab); + +int32_t llama_tokenize_impl( + const struct llama_vocab & vocab, + const char * text, + int32_t text_len, + llama_token * tokens, + int32_t n_tokens_max, + bool add_special, + bool parse_special); + +// does not write null-terminator to buf +int32_t llama_token_to_piece_impl( + const struct llama_vocab & vocab, + llama_token token, + char * buf, + int32_t length, + int32_t lstrip, + bool special); + +int32_t llama_detokenize_impl( + const struct llama_vocab & vocab, + const llama_token * tokens, + int32_t n_tokens, + char * text, + int32_t text_len_max, + bool remove_special, + bool unparse_special); diff --git a/src/llama.cpp b/src/llama.cpp new file mode 100644 index 00000000..eecfccbd --- /dev/null +++ b/src/llama.cpp @@ -0,0 +1,19336 @@ +#include "llama-impl.h" +#include "llama-vocab.h" +#include "llama-grammar.h" +#include "llama-sampling.h" + +#include "unicode.h" + +#include "ggml.h" +#include "ggml-alloc.h" +#include "ggml-backend.h" + +#ifdef GGML_USE_RPC +# include "ggml-rpc.h" +#endif + +#ifdef GGML_USE_CUDA +# include "ggml-cuda.h" +#elif defined(GGML_USE_VULKAN) +# include "ggml-vulkan.h" +#elif defined(GGML_USE_SYCL) +# include "ggml-sycl.h" +#elif defined(GGML_USE_KOMPUTE) +# include "ggml-kompute.h" +#elif defined(GGML_USE_CANN) +# include "ggml-cann.h" +#endif + +#ifdef GGML_USE_BLAS +# include "ggml-blas.h" +#endif + +#ifdef GGML_USE_METAL +# include "ggml-metal.h" +#endif + +// TODO: replace with ggml API call +#define QK_K 256 +#define QK_IQ1BN 64 + +#ifdef __has_include + #if __has_include() + #include + #if defined(_POSIX_MAPPED_FILES) + #include + #include + #endif + #if defined(_POSIX_MEMLOCK_RANGE) + #include + #endif + #endif +#endif + +#if defined(_WIN32) + #define WIN32_LEAN_AND_MEAN + #ifndef NOMINMAX + #define NOMINMAX + #endif + #include + #ifndef PATH_MAX + #define PATH_MAX MAX_PATH + #endif + #include +#endif + +#if __cplusplus >= 202000L + #define LU8(x) (const char*)(u8##x) +#else + #define LU8(x) u8##x +#endif + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#if defined(_MSC_VER) +#pragma warning(disable: 4244 4267) // possible loss of data +#endif + +// bump if necessary +#define LLAMA_MAX_NODES 8192 +#define LLAMA_MAX_LAYERS 512 +#define LLAMA_MAX_EXPERTS 160 // DeepSeekV2 + +// +// helpers +// + +// trim whitespace from the beginning and end of a string +static std::string trim(const std::string & str) { + size_t start = 0; + size_t end = str.size(); + while (start < end && isspace(str[start])) { + start += 1; + } + while (end > start && isspace(str[end - 1])) { + end -= 1; + } + return str.substr(start, end - start); +} + +static void replace_all(std::string & s, const std::string & search, const std::string & replace) { + std::string result; + for (size_t pos = 0; ; pos += search.length()) { + auto new_pos = s.find(search, pos); + if (new_pos == std::string::npos) { + result += s.substr(pos, s.size() - pos); + break; + } + result += s.substr(pos, new_pos - pos) + replace; + pos = new_pos; + } + s = std::move(result); +} + +static bool is_float_close(float a, float b, float abs_tol) { + // Check for non-negative tolerance + if (abs_tol < 0.0) { + throw std::invalid_argument("Tolerance must be non-negative"); + } + + // Exact equality check + if (a == b) { + return true; + } + + // Check for infinities + if (std::isinf(a) || std::isinf(b)) { + return false; + } + + // Regular comparison using the provided absolute tolerance + return std::fabs(b - a) <= abs_tol; +} + +static void zeros(std::ofstream & file, size_t n) { + char zero = 0; + for (size_t i = 0; i < n; ++i) { + file.write(&zero, 1); + } +} + +LLAMA_ATTRIBUTE_FORMAT(1, 2) +static std::string format(const char * fmt, ...) { + va_list ap; + va_list ap2; + va_start(ap, fmt); + va_copy(ap2, ap); + int size = vsnprintf(NULL, 0, fmt, ap); + GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT + std::vector buf(size + 1); + int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2); + GGML_ASSERT(size2 == size); + va_end(ap2); + va_end(ap); + return std::string(buf.data(), size); +} + +// +// gguf constants (sync with gguf.py) +// + +enum llm_arch { + LLM_ARCH_LLAMA, + LLM_ARCH_FALCON, + LLM_ARCH_BAICHUAN, + LLM_ARCH_GROK, + LLM_ARCH_GPT2, + LLM_ARCH_GPTJ, + LLM_ARCH_GPTNEOX, + LLM_ARCH_MPT, + LLM_ARCH_STARCODER, + LLM_ARCH_REFACT, + LLM_ARCH_BERT, + LLM_ARCH_NOMIC_BERT, + LLM_ARCH_JINA_BERT_V2, + LLM_ARCH_BLOOM, + LLM_ARCH_STABLELM, + LLM_ARCH_QWEN, + LLM_ARCH_QWEN2, + LLM_ARCH_QWEN2MOE, + LLM_ARCH_PHI2, + LLM_ARCH_PHI3, + LLM_ARCH_PLAMO, + LLM_ARCH_CODESHELL, + LLM_ARCH_ORION, + LLM_ARCH_INTERNLM2, + LLM_ARCH_MINICPM, + LLM_ARCH_GEMMA, + LLM_ARCH_GEMMA2, + LLM_ARCH_STARCODER2, + LLM_ARCH_MAMBA, + LLM_ARCH_XVERSE, + LLM_ARCH_COMMAND_R, + LLM_ARCH_DBRX, + LLM_ARCH_OLMO, + LLM_ARCH_OPENELM, + LLM_ARCH_ARCTIC, + LLM_ARCH_DEEPSEEK2, + LLM_ARCH_CHATGLM, + LLM_ARCH_BITNET, + LLM_ARCH_T5, + LLM_ARCH_JAIS, + LLM_ARCH_UNKNOWN, +}; + +static const std::map LLM_ARCH_NAMES = { + { LLM_ARCH_LLAMA, "llama" }, + { LLM_ARCH_FALCON, "falcon" }, + { LLM_ARCH_GROK, "grok" }, + { LLM_ARCH_GPT2, "gpt2" }, + { LLM_ARCH_GPTJ, "gptj" }, + { LLM_ARCH_GPTNEOX, "gptneox" }, + { LLM_ARCH_MPT, "mpt" }, + { LLM_ARCH_BAICHUAN, "baichuan" }, + { LLM_ARCH_STARCODER, "starcoder" }, + { LLM_ARCH_REFACT, "refact" }, + { LLM_ARCH_BERT, "bert" }, + { LLM_ARCH_NOMIC_BERT, "nomic-bert" }, + { LLM_ARCH_JINA_BERT_V2, "jina-bert-v2" }, + { LLM_ARCH_BLOOM, "bloom" }, + { LLM_ARCH_STABLELM, "stablelm" }, + { LLM_ARCH_QWEN, "qwen" }, + { LLM_ARCH_QWEN2, "qwen2" }, + { LLM_ARCH_QWEN2MOE, "qwen2moe" }, + { LLM_ARCH_PHI2, "phi2" }, + { LLM_ARCH_PHI3, "phi3" }, + { LLM_ARCH_PLAMO, "plamo" }, + { LLM_ARCH_CODESHELL, "codeshell" }, + { LLM_ARCH_ORION, "orion" }, + { LLM_ARCH_INTERNLM2, "internlm2" }, + { LLM_ARCH_MINICPM, "minicpm" }, + { LLM_ARCH_GEMMA, "gemma" }, + { LLM_ARCH_GEMMA2, "gemma2" }, + { LLM_ARCH_STARCODER2, "starcoder2" }, + { LLM_ARCH_MAMBA, "mamba" }, + { LLM_ARCH_XVERSE, "xverse" }, + { LLM_ARCH_COMMAND_R, "command-r" }, + { LLM_ARCH_DBRX, "dbrx" }, + { LLM_ARCH_OLMO, "olmo" }, + { LLM_ARCH_OPENELM, "openelm" }, + { LLM_ARCH_ARCTIC, "arctic" }, + { LLM_ARCH_DEEPSEEK2, "deepseek2" }, + { LLM_ARCH_CHATGLM, "chatglm" }, + { LLM_ARCH_BITNET, "bitnet" }, + { LLM_ARCH_T5, "t5" }, + { LLM_ARCH_JAIS, "jais" }, + { LLM_ARCH_UNKNOWN, "(unknown)" }, +}; + +enum llm_kv { + LLM_KV_GENERAL_TYPE, + LLM_KV_GENERAL_ARCHITECTURE, + LLM_KV_GENERAL_QUANTIZATION_VERSION, + LLM_KV_GENERAL_ALIGNMENT, + LLM_KV_GENERAL_NAME, + LLM_KV_GENERAL_AUTHOR, + LLM_KV_GENERAL_VERSION, + LLM_KV_GENERAL_URL, + LLM_KV_GENERAL_DESCRIPTION, + LLM_KV_GENERAL_LICENSE, + LLM_KV_GENERAL_SOURCE_URL, + LLM_KV_GENERAL_SOURCE_HF_REPO, + + LLM_KV_VOCAB_SIZE, + LLM_KV_CONTEXT_LENGTH, + LLM_KV_EMBEDDING_LENGTH, + LLM_KV_BLOCK_COUNT, + LLM_KV_LEADING_DENSE_BLOCK_COUNT, + LLM_KV_FEED_FORWARD_LENGTH, + LLM_KV_EXPERT_FEED_FORWARD_LENGTH, + LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, + LLM_KV_USE_PARALLEL_RESIDUAL, + LLM_KV_TENSOR_DATA_LAYOUT, + LLM_KV_EXPERT_COUNT, + LLM_KV_EXPERT_USED_COUNT, + LLM_KV_EXPERT_SHARED_COUNT, + LLM_KV_EXPERT_WEIGHTS_SCALE, + LLM_KV_POOLING_TYPE, + LLM_KV_LOGIT_SCALE, + LLM_KV_DECODER_START_TOKEN_ID, + LLM_KV_ATTN_LOGIT_SOFTCAPPING, + LLM_KV_FINAL_LOGIT_SOFTCAPPING, + + LLM_KV_ATTENTION_HEAD_COUNT, + LLM_KV_ATTENTION_HEAD_COUNT_KV, + LLM_KV_ATTENTION_MAX_ALIBI_BIAS, + LLM_KV_ATTENTION_CLAMP_KQV, + LLM_KV_ATTENTION_KEY_LENGTH, + LLM_KV_ATTENTION_VALUE_LENGTH, + LLM_KV_ATTENTION_LAYERNORM_EPS, + LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, + LLM_KV_ATTENTION_CAUSAL, + LLM_KV_ATTENTION_Q_LORA_RANK, + LLM_KV_ATTENTION_KV_LORA_RANK, + LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, + LLM_KV_ATTENTION_SLIDING_WINDOW, + + LLM_KV_ROPE_DIMENSION_COUNT, + LLM_KV_ROPE_FREQ_BASE, + LLM_KV_ROPE_SCALE_LINEAR, + LLM_KV_ROPE_SCALING_TYPE, + LLM_KV_ROPE_SCALING_FACTOR, + LLM_KV_ROPE_SCALING_ATTN_FACTOR, + LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, + LLM_KV_ROPE_SCALING_FINETUNED, + LLM_KV_ROPE_SCALING_YARN_LOG_MUL, + + LLM_KV_SPLIT_NO, + LLM_KV_SPLIT_COUNT, + LLM_KV_SPLIT_TENSORS_COUNT, + + LLM_KV_SSM_INNER_SIZE, + LLM_KV_SSM_CONV_KERNEL, + LLM_KV_SSM_STATE_SIZE, + LLM_KV_SSM_TIME_STEP_RANK, + + LLM_KV_TOKENIZER_MODEL, + LLM_KV_TOKENIZER_PRE, + LLM_KV_TOKENIZER_LIST, + LLM_KV_TOKENIZER_TOKEN_TYPE, + LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, + LLM_KV_TOKENIZER_SCORES, + LLM_KV_TOKENIZER_MERGES, + LLM_KV_TOKENIZER_BOS_ID, + LLM_KV_TOKENIZER_EOS_ID, + LLM_KV_TOKENIZER_UNK_ID, + LLM_KV_TOKENIZER_SEP_ID, + LLM_KV_TOKENIZER_PAD_ID, + LLM_KV_TOKENIZER_CLS_ID, + LLM_KV_TOKENIZER_MASK_ID, + LLM_KV_TOKENIZER_ADD_BOS, + LLM_KV_TOKENIZER_ADD_EOS, + LLM_KV_TOKENIZER_ADD_PREFIX, + LLM_KV_TOKENIZER_REMOVE_EXTRA_WS, + LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, + LLM_KV_TOKENIZER_HF_JSON, + LLM_KV_TOKENIZER_RWKV, + LLM_KV_TOKENIZER_PREFIX_ID, + LLM_KV_TOKENIZER_SUFFIX_ID, + LLM_KV_TOKENIZER_MIDDLE_ID, + LLM_KV_TOKENIZER_EOT_ID, + + LLM_KV_ADAPTER_TYPE, + LLM_KV_ADAPTER_LORA_ALPHA, +}; + +static const std::map LLM_KV_NAMES = { + { LLM_KV_GENERAL_TYPE, "general.type" }, + { LLM_KV_GENERAL_ARCHITECTURE, "general.architecture" }, + { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version" }, + { LLM_KV_GENERAL_ALIGNMENT, "general.alignment" }, + { LLM_KV_GENERAL_NAME, "general.name" }, + { LLM_KV_GENERAL_AUTHOR, "general.author" }, + { LLM_KV_GENERAL_VERSION, "general.version" }, + { LLM_KV_GENERAL_URL, "general.url" }, + { LLM_KV_GENERAL_DESCRIPTION, "general.description" }, + { LLM_KV_GENERAL_LICENSE, "general.license" }, + { LLM_KV_GENERAL_SOURCE_URL, "general.source.url" }, + { LLM_KV_GENERAL_SOURCE_HF_REPO, "general.source.huggingface.repository" }, + + { LLM_KV_VOCAB_SIZE, "%s.vocab_size" }, + { LLM_KV_CONTEXT_LENGTH, "%s.context_length" }, + { LLM_KV_EMBEDDING_LENGTH, "%s.embedding_length" }, + { LLM_KV_BLOCK_COUNT, "%s.block_count" }, + { LLM_KV_LEADING_DENSE_BLOCK_COUNT, "%s.leading_dense_block_count" }, + { LLM_KV_FEED_FORWARD_LENGTH, "%s.feed_forward_length" }, + { LLM_KV_EXPERT_FEED_FORWARD_LENGTH, "%s.expert_feed_forward_length" }, + { LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, "%s.expert_shared_feed_forward_length" }, + { LLM_KV_USE_PARALLEL_RESIDUAL, "%s.use_parallel_residual" }, + { LLM_KV_TENSOR_DATA_LAYOUT, "%s.tensor_data_layout" }, + { LLM_KV_EXPERT_COUNT, "%s.expert_count" }, + { LLM_KV_EXPERT_USED_COUNT, "%s.expert_used_count" }, + { LLM_KV_EXPERT_SHARED_COUNT, "%s.expert_shared_count" }, + { LLM_KV_EXPERT_WEIGHTS_SCALE, "%s.expert_weights_scale" }, + { LLM_KV_POOLING_TYPE , "%s.pooling_type" }, + { LLM_KV_LOGIT_SCALE, "%s.logit_scale" }, + { LLM_KV_DECODER_START_TOKEN_ID, "%s.decoder_start_token_id" }, + { LLM_KV_ATTN_LOGIT_SOFTCAPPING, "%s.attn_logit_softcapping" }, + { LLM_KV_FINAL_LOGIT_SOFTCAPPING, "%s.final_logit_softcapping" }, + + { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" }, + { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" }, + { LLM_KV_ATTENTION_MAX_ALIBI_BIAS, "%s.attention.max_alibi_bias" }, + { LLM_KV_ATTENTION_CLAMP_KQV, "%s.attention.clamp_kqv" }, + { LLM_KV_ATTENTION_KEY_LENGTH, "%s.attention.key_length" }, + { LLM_KV_ATTENTION_VALUE_LENGTH, "%s.attention.value_length" }, + { LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" }, + { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" }, + { LLM_KV_ATTENTION_CAUSAL, "%s.attention.causal" }, + { LLM_KV_ATTENTION_Q_LORA_RANK, "%s.attention.q_lora_rank" }, + { LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" }, + { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" }, + { LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" }, + + { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, + { LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" }, + { LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" }, + { LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" }, + { LLM_KV_ROPE_SCALING_FACTOR, "%s.rope.scaling.factor" }, + { LLM_KV_ROPE_SCALING_ATTN_FACTOR, "%s.rope.scaling.attn_factor" }, + { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" }, + { LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" }, + { LLM_KV_ROPE_SCALING_YARN_LOG_MUL, "%s.rope.scaling.yarn_log_multiplier" }, + + { LLM_KV_SPLIT_NO, "split.no" }, + { LLM_KV_SPLIT_COUNT, "split.count" }, + { LLM_KV_SPLIT_TENSORS_COUNT, "split.tensors.count" }, + + { LLM_KV_SSM_CONV_KERNEL, "%s.ssm.conv_kernel" }, + { LLM_KV_SSM_INNER_SIZE, "%s.ssm.inner_size" }, + { LLM_KV_SSM_STATE_SIZE, "%s.ssm.state_size" }, + { LLM_KV_SSM_TIME_STEP_RANK, "%s.ssm.time_step_rank" }, + + { LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" }, + { LLM_KV_TOKENIZER_PRE, "tokenizer.ggml.pre" }, + { LLM_KV_TOKENIZER_LIST, "tokenizer.ggml.tokens" }, + { LLM_KV_TOKENIZER_TOKEN_TYPE, "tokenizer.ggml.token_type" }, + { LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, "tokenizer.ggml.token_type_count" }, + { LLM_KV_TOKENIZER_SCORES, "tokenizer.ggml.scores" }, + { LLM_KV_TOKENIZER_MERGES, "tokenizer.ggml.merges" }, + { LLM_KV_TOKENIZER_BOS_ID, "tokenizer.ggml.bos_token_id" }, + { LLM_KV_TOKENIZER_EOS_ID, "tokenizer.ggml.eos_token_id" }, + { LLM_KV_TOKENIZER_UNK_ID, "tokenizer.ggml.unknown_token_id" }, + { LLM_KV_TOKENIZER_SEP_ID, "tokenizer.ggml.seperator_token_id" }, + { LLM_KV_TOKENIZER_PAD_ID, "tokenizer.ggml.padding_token_id" }, + { LLM_KV_TOKENIZER_CLS_ID, "tokenizer.ggml.cls_token_id" }, + { LLM_KV_TOKENIZER_MASK_ID, "tokenizer.ggml.mask_token_id" }, + { LLM_KV_TOKENIZER_ADD_BOS, "tokenizer.ggml.add_bos_token" }, + { LLM_KV_TOKENIZER_ADD_EOS, "tokenizer.ggml.add_eos_token" }, + { LLM_KV_TOKENIZER_ADD_PREFIX, "tokenizer.ggml.add_space_prefix" }, + { LLM_KV_TOKENIZER_REMOVE_EXTRA_WS, "tokenizer.ggml.remove_extra_whitespaces" }, + { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, "tokenizer.ggml.precompiled_charsmap" }, + { LLM_KV_TOKENIZER_HF_JSON, "tokenizer.huggingface.json" }, + { LLM_KV_TOKENIZER_RWKV, "tokenizer.rwkv.world" }, + { LLM_KV_TOKENIZER_PREFIX_ID, "tokenizer.ggml.prefix_token_id" }, + { LLM_KV_TOKENIZER_SUFFIX_ID, "tokenizer.ggml.suffix_token_id" }, + { LLM_KV_TOKENIZER_MIDDLE_ID, "tokenizer.ggml.middle_token_id" }, + { LLM_KV_TOKENIZER_EOT_ID, "tokenizer.ggml.eot_token_id" }, + + { LLM_KV_ADAPTER_TYPE, "adapter.type" }, + { LLM_KV_ADAPTER_LORA_ALPHA, "adapter.lora.alpha" }, +}; + +struct LLM_KV { + LLM_KV(llm_arch arch) : arch(arch) {} + + llm_arch arch; + + std::string operator()(llm_kv kv) const { + return ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch)); + } +}; + +enum llm_tensor { + LLM_TENSOR_TOKEN_EMBD, + LLM_TENSOR_TOKEN_EMBD_NORM, + LLM_TENSOR_TOKEN_TYPES, + LLM_TENSOR_POS_EMBD, + LLM_TENSOR_OUTPUT, + LLM_TENSOR_OUTPUT_NORM, + LLM_TENSOR_ROPE_FREQS, + LLM_TENSOR_ROPE_FACTORS_LONG, + LLM_TENSOR_ROPE_FACTORS_SHORT, + LLM_TENSOR_ATTN_Q, + LLM_TENSOR_ATTN_K, + LLM_TENSOR_ATTN_V, + LLM_TENSOR_ATTN_QKV, + LLM_TENSOR_ATTN_OUT, + LLM_TENSOR_ATTN_NORM, + LLM_TENSOR_ATTN_NORM_2, + LLM_TENSOR_ATTN_OUT_NORM, + LLM_TENSOR_ATTN_POST_NORM, + LLM_TENSOR_ATTN_ROT_EMBD, + LLM_TENSOR_FFN_GATE_INP, + LLM_TENSOR_FFN_GATE_INP_SHEXP, + LLM_TENSOR_FFN_NORM, + LLM_TENSOR_FFN_POST_NORM, + LLM_TENSOR_FFN_GATE, + LLM_TENSOR_FFN_DOWN, + LLM_TENSOR_FFN_UP, + LLM_TENSOR_FFN_ACT, + LLM_TENSOR_FFN_DOWN_EXP, // split experts for backward compatibility + LLM_TENSOR_FFN_GATE_EXP, + LLM_TENSOR_FFN_UP_EXP, + LLM_TENSOR_FFN_NORM_EXPS, + LLM_TENSOR_FFN_DOWN_EXPS, // merged experts + LLM_TENSOR_FFN_GATE_EXPS, + LLM_TENSOR_FFN_UP_EXPS, + LLM_TENSOR_FFN_DOWN_SHEXP, + LLM_TENSOR_FFN_GATE_SHEXP, + LLM_TENSOR_FFN_UP_SHEXP, + LLM_TENSOR_ATTN_Q_NORM, + LLM_TENSOR_ATTN_K_NORM, + LLM_TENSOR_LAYER_OUT_NORM, + LLM_TENSOR_SSM_IN, + LLM_TENSOR_SSM_CONV1D, + LLM_TENSOR_SSM_X, + LLM_TENSOR_SSM_DT, + LLM_TENSOR_SSM_A, + LLM_TENSOR_SSM_D, + LLM_TENSOR_SSM_OUT, + LLM_TENSOR_ATTN_Q_A, + LLM_TENSOR_ATTN_Q_B, + LLM_TENSOR_ATTN_KV_A_MQA, + LLM_TENSOR_ATTN_KV_B, + LLM_TENSOR_ATTN_Q_A_NORM, + LLM_TENSOR_ATTN_KV_A_NORM, + LLM_TENSOR_ATTN_SUB_NORM, + LLM_TENSOR_FFN_SUB_NORM, + LLM_TENSOR_DEC_ATTN_NORM, + LLM_TENSOR_DEC_ATTN_Q, + LLM_TENSOR_DEC_ATTN_K, + LLM_TENSOR_DEC_ATTN_V, + LLM_TENSOR_DEC_ATTN_OUT, + LLM_TENSOR_DEC_ATTN_REL_B, + LLM_TENSOR_DEC_CROSS_ATTN_NORM, + LLM_TENSOR_DEC_CROSS_ATTN_Q, + LLM_TENSOR_DEC_CROSS_ATTN_K, + LLM_TENSOR_DEC_CROSS_ATTN_V, + LLM_TENSOR_DEC_CROSS_ATTN_OUT, + LLM_TENSOR_DEC_CROSS_ATTN_REL_B, + LLM_TENSOR_DEC_FFN_NORM, + LLM_TENSOR_DEC_FFN_GATE, + LLM_TENSOR_DEC_FFN_DOWN, + LLM_TENSOR_DEC_FFN_UP, + LLM_TENSOR_DEC_OUTPUT_NORM, + LLM_TENSOR_ENC_ATTN_NORM, + LLM_TENSOR_ENC_ATTN_Q, + LLM_TENSOR_ENC_ATTN_K, + LLM_TENSOR_ENC_ATTN_V, + LLM_TENSOR_ENC_ATTN_OUT, + LLM_TENSOR_ENC_ATTN_REL_B, + LLM_TENSOR_ENC_FFN_NORM, + LLM_TENSOR_ENC_FFN_GATE, + LLM_TENSOR_ENC_FFN_DOWN, + LLM_TENSOR_ENC_FFN_UP, + LLM_TENSOR_ENC_OUTPUT_NORM, +}; + +static const std::map> LLM_TENSOR_NAMES = { + { + LLM_ARCH_LLAMA, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_GATE_EXP, "blk.%d.ffn_gate.%d" }, + { LLM_TENSOR_FFN_DOWN_EXP, "blk.%d.ffn_down.%d" }, + { LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + }, + }, + { + LLM_ARCH_BAICHUAN, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_FALCON, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_NORM_2, "blk.%d.attn_norm_2" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_GROK, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE_EXP, "blk.%d.ffn_gate.%d" }, + { LLM_TENSOR_FFN_DOWN_EXP, "blk.%d.ffn_down.%d" }, + { LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + { LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" }, + { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" }, + }, + }, + { + LLM_ARCH_GPT2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_POS_EMBD, "position_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + }, + }, + { + LLM_ARCH_GPTJ, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + }, + }, + { + LLM_ARCH_GPTNEOX, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_MPT, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output"}, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_ACT, "blk.%d.ffn.act" }, + { LLM_TENSOR_POS_EMBD, "position_embd" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm"}, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm"}, + }, + }, + { + LLM_ARCH_STARCODER, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_POS_EMBD, "position_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + }, + }, + { + LLM_ARCH_REFACT, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_BERT, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, + { LLM_TENSOR_TOKEN_TYPES, "token_types" }, + { LLM_TENSOR_POS_EMBD, "position_embd" }, + { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_NOMIC_BERT, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, + { LLM_TENSOR_TOKEN_TYPES, "token_types" }, + { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_JINA_BERT_V2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, + { LLM_TENSOR_TOKEN_TYPES, "token_types" }, + { LLM_TENSOR_ATTN_NORM_2, "blk.%d.attn_norm_2" }, + { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_BLOOM, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + }, + }, + { + LLM_ARCH_STABLELM, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + }, + }, + { + LLM_ARCH_QWEN, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_QWEN2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_QWEN2MOE, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" }, + { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" }, + { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" }, + { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" }, + }, + }, + { + LLM_ARCH_PHI2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_PHI3, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FACTORS_LONG, "rope_factors_long" }, + { LLM_TENSOR_ROPE_FACTORS_SHORT, "rope_factors_short" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_PLAMO, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_CODESHELL, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_ORION, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_INTERNLM2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_MINICPM, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_GATE_EXP, "blk.%d.ffn_gate.%d" }, + { LLM_TENSOR_FFN_DOWN_EXP, "blk.%d.ffn_down.%d" }, + { LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" }, + }, + }, + { + LLM_ARCH_GEMMA, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_GEMMA2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" }, + }, + }, + { + LLM_ARCH_STARCODER2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_MAMBA, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" }, + { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" }, + { LLM_TENSOR_SSM_X, "blk.%d.ssm_x" }, + { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" }, + { LLM_TENSOR_SSM_A, "blk.%d.ssm_a" }, + { LLM_TENSOR_SSM_D, "blk.%d.ssm_d" }, + { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" }, + }, + }, + { + LLM_ARCH_XVERSE, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_COMMAND_R, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + }, + }, + { + LLM_ARCH_DBRX, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + }, + }, + { + LLM_ARCH_OLMO, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_OPENELM, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_ARCTIC, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_NORM_EXPS, "blk.%d.ffn_norm_exps" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + }, + }, + { + LLM_ARCH_DEEPSEEK2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q_A_NORM, "blk.%d.attn_q_a_norm" }, + { LLM_TENSOR_ATTN_KV_A_NORM, "blk.%d.attn_kv_a_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_Q_A, "blk.%d.attn_q_a" }, + { LLM_TENSOR_ATTN_Q_B, "blk.%d.attn_q_b" }, + { LLM_TENSOR_ATTN_KV_A_MQA, "blk.%d.attn_kv_a_mqa" }, + { LLM_TENSOR_ATTN_KV_B, "blk.%d.attn_kv_b" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" }, + { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" }, + { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" }, + { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" }, + }, + }, + { + LLM_ARCH_BITNET, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_SUB_NORM, "blk.%d.attn_sub_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_SUB_NORM, "blk.%d.ffn_sub_norm" }, + }, + }, + { + LLM_ARCH_T5, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_DEC_OUTPUT_NORM, "dec.output_norm" }, + { LLM_TENSOR_DEC_ATTN_NORM, "dec.blk.%d.attn_norm" }, + { LLM_TENSOR_DEC_ATTN_Q, "dec.blk.%d.attn_q" }, + { LLM_TENSOR_DEC_ATTN_K, "dec.blk.%d.attn_k" }, + { LLM_TENSOR_DEC_ATTN_V, "dec.blk.%d.attn_v" }, + { LLM_TENSOR_DEC_ATTN_OUT, "dec.blk.%d.attn_o" }, + { LLM_TENSOR_DEC_ATTN_REL_B, "dec.blk.%d.attn_rel_b" }, + { LLM_TENSOR_DEC_CROSS_ATTN_NORM, "dec.blk.%d.cross_attn_norm" }, + { LLM_TENSOR_DEC_CROSS_ATTN_Q, "dec.blk.%d.cross_attn_q" }, + { LLM_TENSOR_DEC_CROSS_ATTN_K, "dec.blk.%d.cross_attn_k" }, + { LLM_TENSOR_DEC_CROSS_ATTN_V, "dec.blk.%d.cross_attn_v" }, + { LLM_TENSOR_DEC_CROSS_ATTN_OUT, "dec.blk.%d.cross_attn_o" }, + { LLM_TENSOR_DEC_CROSS_ATTN_REL_B, "dec.blk.%d.cross_attn_rel_b" }, + { LLM_TENSOR_DEC_FFN_NORM, "dec.blk.%d.ffn_norm" }, + { LLM_TENSOR_DEC_FFN_GATE, "dec.blk.%d.ffn_gate" }, + { LLM_TENSOR_DEC_FFN_DOWN, "dec.blk.%d.ffn_down" }, + { LLM_TENSOR_DEC_FFN_UP, "dec.blk.%d.ffn_up" }, + { LLM_TENSOR_ENC_OUTPUT_NORM, "enc.output_norm" }, + { LLM_TENSOR_ENC_ATTN_NORM, "enc.blk.%d.attn_norm" }, + { LLM_TENSOR_ENC_ATTN_Q, "enc.blk.%d.attn_q" }, + { LLM_TENSOR_ENC_ATTN_K, "enc.blk.%d.attn_k" }, + { LLM_TENSOR_ENC_ATTN_V, "enc.blk.%d.attn_v" }, + { LLM_TENSOR_ENC_ATTN_OUT, "enc.blk.%d.attn_o" }, + { LLM_TENSOR_ENC_ATTN_REL_B, "enc.blk.%d.attn_rel_b" }, + { LLM_TENSOR_ENC_FFN_NORM, "enc.blk.%d.ffn_norm" }, + { LLM_TENSOR_ENC_FFN_GATE, "enc.blk.%d.ffn_gate" }, + { LLM_TENSOR_ENC_FFN_DOWN, "enc.blk.%d.ffn_down" }, + { LLM_TENSOR_ENC_FFN_UP, "enc.blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_JAIS, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + }, + }, + { + LLM_ARCH_UNKNOWN, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + }, + }, +}; + +static llm_arch llm_arch_from_string(const std::string & name) { + for (const auto & kv : LLM_ARCH_NAMES) { // NOLINT + if (kv.second == name) { + return kv.first; + } + } + + return LLM_ARCH_UNKNOWN; +} + +// helper to handle gguf constants +// usage: +// +// const auto tn = LLM_TN(LLM_ARCH_LLAMA); +// +// std::string name = tn(LLM_TENSOR_OUTPUT); -> "output" +// std::string name = tn(LLM_TENSOR_TOKEN_EMBD, "bias"); -> "token_embd.bias" +// std::string name = tn(LLM_TENSOR_ATTN_NORM, "weight", 3); -> "blk.3.attn_norm.weight" +// +struct LLM_TN { + LLM_TN(llm_arch arch) : arch(arch) {} + + llm_arch arch; + + std::string operator()(llm_tensor tensor) const { + if (LLM_TENSOR_NAMES.at(arch).find(tensor) == LLM_TENSOR_NAMES.at(arch).end()) { + return "__missing__"; + } + return LLM_TENSOR_NAMES.at(arch).at(tensor); + } + + std::string operator()(llm_tensor tensor, const std::string & suffix) const { + if (LLM_TENSOR_NAMES.at(arch).find(tensor) == LLM_TENSOR_NAMES.at(arch).end()) { + return "__missing__"; + } + return LLM_TENSOR_NAMES.at(arch).at(tensor) + "." + suffix; + } + + std::string operator()(llm_tensor tensor, int bid) const { + if (LLM_TENSOR_NAMES.at(arch).find(tensor) == LLM_TENSOR_NAMES.at(arch).end()) { + return "__missing__"; + } + return ::format(LLM_TENSOR_NAMES.at(arch).at(tensor).c_str(), bid); + } + + std::string operator()(llm_tensor tensor, const std::string & suffix, int bid) const { + if (LLM_TENSOR_NAMES.at(arch).find(tensor) == LLM_TENSOR_NAMES.at(arch).end()) { + return "__missing__"; + } + return ::format(LLM_TENSOR_NAMES.at(arch).at(tensor).c_str(), bid) + "." + suffix; + } + + std::string operator()(llm_tensor tensor, const std::string & suffix, int bid, int xid) const { + if (LLM_TENSOR_NAMES.at(arch).find(tensor) == LLM_TENSOR_NAMES.at(arch).end()) { + return "__missing__"; + } + return ::format(LLM_TENSOR_NAMES.at(arch).at(tensor).c_str(), bid, xid) + "." + suffix; + } +}; + +// +// gguf helpers +// + +static const std::map LLAMA_ROPE_SCALING_TYPES = { + { LLAMA_ROPE_SCALING_TYPE_NONE, "none" }, + { LLAMA_ROPE_SCALING_TYPE_LINEAR, "linear" }, + { LLAMA_ROPE_SCALING_TYPE_YARN, "yarn" }, +}; + +static llama_rope_scaling_type llama_rope_scaling_type_from_string(const std::string & name) { + for (const auto & kv : LLAMA_ROPE_SCALING_TYPES) { + if (kv.second == name) { + return (llama_rope_scaling_type) kv.first; + } + } + + return LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED; +} + +static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) { + switch (type) { + case GGUF_TYPE_UINT8: return std::to_string(((const uint8_t *)data)[i]); + case GGUF_TYPE_INT8: return std::to_string(((const int8_t *)data)[i]); + case GGUF_TYPE_UINT16: return std::to_string(((const uint16_t *)data)[i]); + case GGUF_TYPE_INT16: return std::to_string(((const int16_t *)data)[i]); + case GGUF_TYPE_UINT32: return std::to_string(((const uint32_t *)data)[i]); + case GGUF_TYPE_INT32: return std::to_string(((const int32_t *)data)[i]); + case GGUF_TYPE_UINT64: return std::to_string(((const uint64_t *)data)[i]); + case GGUF_TYPE_INT64: return std::to_string(((const int64_t *)data)[i]); + case GGUF_TYPE_FLOAT32: return std::to_string(((const float *)data)[i]); + case GGUF_TYPE_FLOAT64: return std::to_string(((const double *)data)[i]); + case GGUF_TYPE_BOOL: return ((const bool *)data)[i] ? "true" : "false"; + default: return format("unknown type %d", type); + } +} + +static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) { + const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i); + + switch (type) { + case GGUF_TYPE_STRING: + return gguf_get_val_str(ctx_gguf, i); + case GGUF_TYPE_ARRAY: + { + const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i); + int arr_n = gguf_get_arr_n(ctx_gguf, i); + const void * data = gguf_get_arr_data(ctx_gguf, i); + std::stringstream ss; + ss << "["; + for (int j = 0; j < arr_n; j++) { + if (arr_type == GGUF_TYPE_STRING) { + std::string val = gguf_get_arr_str(ctx_gguf, i, j); + // escape quotes + replace_all(val, "\\", "\\\\"); + replace_all(val, "\"", "\\\""); + ss << '"' << val << '"'; + } else if (arr_type == GGUF_TYPE_ARRAY) { + ss << "???"; + } else { + ss << gguf_data_to_str(arr_type, data, j); + } + if (j < arr_n - 1) { + ss << ", "; + } + } + ss << "]"; + return ss.str(); + } + default: + return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0); + } +} + +// +// llama helpers +// + +#if defined(_WIN32) +static std::string llama_format_win_err(DWORD err) { + LPSTR buf; + size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS, + NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL); + if (!size) { + return "FormatMessageA failed"; + } + std::string ret(buf, size); + LocalFree(buf); + return ret; +} +#endif + +template +struct no_init { + T value; + no_init() { /* do nothing */ } +}; + +struct llama_file { + +#if defined(_WIN32) + // use FILE * so we don't have to re-open the file to mmap + FILE * fp; + HANDLE fp_win32; + size_t size; + +private: + std::string GetErrorMessageWin32(DWORD error_code) const { + std::string ret; + LPSTR lpMsgBuf = NULL; + DWORD bufLen = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS, + NULL, error_code, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&lpMsgBuf, 0, NULL); + if (!bufLen) { + ret = format("Win32 error code: %s", error_code); + } else { + ret = lpMsgBuf; + LocalFree(lpMsgBuf); + } + + return ret; + } + +public: + + llama_file(const char * fname, const char * mode) { + fp = ggml_fopen(fname, mode); + if (fp == NULL) { + throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno))); + } + fp_win32 = (HANDLE) _get_osfhandle(_fileno(fp)); + seek(0, SEEK_END); + size = tell(); + seek(0, SEEK_SET); + } + + size_t tell() const { + // SetFilePointerEx returns the current position when seeking relative 0 bytes + LARGE_INTEGER li; + li.QuadPart = 0; + BOOL ret = SetFilePointerEx(fp_win32, li, &li, FILE_CURRENT); + if (!ret) { + throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str())); + } + + return li.QuadPart; + } + + void seek(size_t offset, int whence) const { + // no need to convert SEEK_* to FILE_*. The enums are the same. + // Still, keep static asserts to avoid failures in the future. + static_assert(SEEK_SET == FILE_BEGIN, "SEEK_SET != FILE_BEGIN"); + static_assert(SEEK_CUR == FILE_CURRENT, "SEEK_CUR != FILE_CURRENT"); + static_assert(SEEK_END == FILE_END, "SEEK_END != FILE_END"); + + LARGE_INTEGER li; + li.QuadPart = offset; + BOOL ret = SetFilePointerEx(fp_win32, li, NULL, whence); + if (!ret) { + throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str())); + } + } + + void read_raw(void * ptr, size_t len) const { + // On Win32 ReadFile is significant faster than fread which is again significant faster than std::fstream. Thus + // use the Win32 API to do file io instead of the C/C++ library functions. + + // There are conditions under which ReadFile cannot read chunks >64MB. + // Thus split the operation into smaller chunks if len exceeds this limit. + size_t bytes_read = 0; + while (bytes_read < len) { + size_t chunk_size = std::min(len - bytes_read, 64*1024*1024); + DWORD chunk_read = 0; + BOOL result = ReadFile(fp_win32, reinterpret_cast(ptr) + bytes_read, chunk_size, &chunk_read, NULL); + if (!result) { + throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str())); + } + if (chunk_read < chunk_size || chunk_read == 0) { + throw std::runtime_error("unexpectedly reached end of file"); + } + + bytes_read += chunk_read; + } ; + } + + uint32_t read_u32() const { + uint32_t val; + read_raw(&val, sizeof(val)); + return val; + } + + void write_raw(const void * ptr, size_t len) const { + // There are conditions under which WriteFile cannot write chunks >64MB. + // Thus split the operation into smaller chunks if len exceeds this limit. + size_t bytes_written = 0; + while (bytes_written < len) { + size_t chunk_size = std::min(len - bytes_written, 64*1024*1024); + DWORD chunk_written = 0; + BOOL result = WriteFile(fp_win32, reinterpret_cast(ptr) + bytes_written, chunk_size, &chunk_written, NULL); + if (!result) { + throw std::runtime_error(format("write error: %s", GetErrorMessageWin32(GetLastError()).c_str())); + } + if (chunk_written < chunk_size || chunk_written == 0) { + throw std::runtime_error("unexpectedly failed to write bytes"); + } + + bytes_written += chunk_written; + } + } + + void write_u32(std::uint32_t val) const { + write_raw(&val, sizeof(val)); + } + + ~llama_file() { + if (fp) { + std::fclose(fp); + } + } +#else + // use FILE * so we don't have to re-open the file to mmap + FILE * fp; + size_t size; + + llama_file(const char * fname, const char * mode) { + fp = ggml_fopen(fname, mode); + if (fp == NULL) { + throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno))); + } + seek(0, SEEK_END); + size = tell(); + seek(0, SEEK_SET); + } + + size_t tell() const { +#ifdef _WIN32 + __int64 ret = _ftelli64(fp); +#else + long ret = std::ftell(fp); +#endif + if (ret == -1) { + throw std::runtime_error(format("ftell error: %s", strerror(errno))); + } + + return (size_t) ret; + } + + void seek(size_t offset, int whence) const { +#ifdef _WIN32 + int ret = _fseeki64(fp, (__int64) offset, whence); +#else + int ret = std::fseek(fp, (long) offset, whence); +#endif + if (ret != 0) { + throw std::runtime_error(format("seek error: %s", strerror(errno))); + } + } + + void read_raw(void * ptr, size_t len) const { + if (len == 0) { + return; + } + errno = 0; + std::size_t ret = std::fread(ptr, len, 1, fp); + if (ferror(fp)) { + throw std::runtime_error(format("read error: %s", strerror(errno))); + } + if (ret != 1) { + throw std::runtime_error("unexpectedly reached end of file"); + } + } + + uint32_t read_u32() const { + uint32_t ret; + read_raw(&ret, sizeof(ret)); + return ret; + } + + void write_raw(const void * ptr, size_t len) const { + if (len == 0) { + return; + } + errno = 0; + size_t ret = std::fwrite(ptr, len, 1, fp); + if (ret != 1) { + throw std::runtime_error(format("write error: %s", strerror(errno))); + } + } + + void write_u32(std::uint32_t val) const { + write_raw(&val, sizeof(val)); + } + + ~llama_file() { + if (fp) { + std::fclose(fp); + } + } +#endif +}; +using llama_files = std::vector>; + +struct llama_mmap { + void * addr; + size_t size; + + llama_mmap(const llama_mmap &) = delete; + +#ifdef _POSIX_MAPPED_FILES + static constexpr bool SUPPORTED = true; + + // list of mapped fragments (first_offset, last_offset) + std::vector> mapped_fragments; + + llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */, bool numa = false) { + size = file->size; + int fd = fileno(file->fp); + int flags = MAP_SHARED; + // prefetch/readahead impairs performance on NUMA systems + if (numa) { prefetch = 0; } +#ifdef __linux__ + // advise the kernel to read the file sequentially (increases readahead) + if (posix_fadvise(fd, 0, 0, POSIX_FADV_SEQUENTIAL)) { + LLAMA_LOG_WARN("warning: posix_fadvise(.., POSIX_FADV_SEQUENTIAL) failed: %s\n", + strerror(errno)); + } + if (prefetch) { flags |= MAP_POPULATE; } +#endif + addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0); + if (addr == MAP_FAILED) { // NOLINT + throw std::runtime_error(format("mmap failed: %s", strerror(errno))); + } + + if (prefetch > 0) { + // advise the kernel to preload the mapped memory + if (posix_madvise(addr, std::min(file->size, prefetch), POSIX_MADV_WILLNEED)) { + LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n", + strerror(errno)); + } + } + if (numa) { + // advise the kernel not to use readahead + // (because the next page might not belong on the same node) + if (posix_madvise(addr, file->size, POSIX_MADV_RANDOM)) { + LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_RANDOM) failed: %s\n", + strerror(errno)); + } + } + + // initialize list of mapped_fragments + mapped_fragments.emplace_back(0, file->size); + } + + static void align_range(size_t * first, size_t * last, size_t page_size) { + // align first to the next page + size_t offset_in_page = *first & (page_size - 1); + size_t offset_to_page = offset_in_page == 0 ? 0 : page_size - offset_in_page; + *first += offset_to_page; + + // align last to the previous page + *last = *last & ~(page_size - 1); + + if (*last <= *first) { + *last = *first; + } + } + + // partially unmap the file in the range [first, last) + void unmap_fragment(size_t first, size_t last) { + // note: this function must not be called multiple times with overlapping ranges + // otherwise, there is a risk of invalidating addresses that have been repurposed for other mappings + int page_size = sysconf(_SC_PAGESIZE); + align_range(&first, &last, page_size); + size_t len = last - first; + + if (len == 0) { + return; + } + + GGML_ASSERT(first % page_size == 0); + GGML_ASSERT(last % page_size == 0); + GGML_ASSERT(last > first); + + void * next_page_start = (uint8_t *) addr + first; + + // unmap the range + if (munmap(next_page_start, len)) { + LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno)); + } + + // update the list of mapped fragments to avoid unmapping the same range again in the destructor + std::vector> new_mapped_fragments; + for (const auto & frag : mapped_fragments) { + if (frag.first < first && frag.second > last) { + // the range is in the middle of the fragment, split it + new_mapped_fragments.emplace_back(frag.first, first); + new_mapped_fragments.emplace_back(last, frag.second); + } else if (frag.first < first && frag.second > first) { + // the range starts in the middle of the fragment + new_mapped_fragments.emplace_back(frag.first, first); + } else if (frag.first < last && frag.second > last) { + // the range ends in the middle of the fragment + new_mapped_fragments.emplace_back(last, frag.second); + } else if (frag.first >= first && frag.second <= last) { + // the range covers the entire fragment + } else { + // the range is outside the fragment + new_mapped_fragments.push_back(frag); + } + } + mapped_fragments = std::move(new_mapped_fragments); + } + + ~llama_mmap() { + for (const auto & frag : mapped_fragments) { + if (munmap((char *) addr + frag.first, frag.second - frag.first)) { + LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno)); + } + } + } +#elif defined(_WIN32) + static constexpr bool SUPPORTED = true; + + llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1, bool numa = false) { + GGML_UNUSED(numa); + + size = file->size; + + HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp)); + + HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL); + + if (hMapping == NULL) { + DWORD error = GetLastError(); + throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str())); + } + + addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0); + DWORD error = GetLastError(); + CloseHandle(hMapping); + + if (addr == NULL) { + throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str())); + } + + if (prefetch > 0) { +#if _WIN32_WINNT >= 0x602 + // PrefetchVirtualMemory is only present on Windows 8 and above, so we dynamically load it + BOOL (WINAPI *pPrefetchVirtualMemory) (HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG); + HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll"); + + // may fail on pre-Windows 8 systems + pPrefetchVirtualMemory = reinterpret_cast (GetProcAddress(hKernel32, "PrefetchVirtualMemory")); + + if (pPrefetchVirtualMemory) { + // advise the kernel to preload the mapped memory + WIN32_MEMORY_RANGE_ENTRY range; + range.VirtualAddress = addr; + range.NumberOfBytes = (SIZE_T) std::min(size, prefetch); + if (!pPrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) { + LLAMA_LOG_WARN("warning: PrefetchVirtualMemory failed: %s\n", + llama_format_win_err(GetLastError()).c_str()); + } + } +#else + throw std::runtime_error("PrefetchVirtualMemory unavailable"); +#endif + } + } + + void unmap_fragment(size_t first, size_t last) { + // not supported + GGML_UNUSED(first); + GGML_UNUSED(last); + } + + ~llama_mmap() { + if (!UnmapViewOfFile(addr)) { + LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n", + llama_format_win_err(GetLastError()).c_str()); + } + } +#else + static constexpr bool SUPPORTED = false; + + llama_mmap(struct llama_file * file, size_t prefetch = -1, bool numa = false) { + GGML_UNUSED(file); + GGML_UNUSED(prefetch); + GGML_UNUSED(numa); + + throw std::runtime_error("mmap not supported"); + } + + void unmap_fragment(size_t first, size_t last) { + GGML_UNUSED(first); + GGML_UNUSED(last); + + throw std::runtime_error("mmap not supported"); + } +#endif +}; +using llama_mmaps = std::vector>; + +// Represents some region of memory being locked using mlock or VirtualLock; +// will automatically unlock on destruction. +struct llama_mlock { + void * addr = NULL; + size_t size = 0; + + bool failed_already = false; + + llama_mlock() {} + llama_mlock(const llama_mlock &) = delete; + + ~llama_mlock() { + if (size) { + raw_unlock(addr, size); + } + } + + void init(void * ptr) { + GGML_ASSERT(addr == NULL && size == 0); // NOLINT + addr = ptr; + } + + void grow_to(size_t target_size) { + GGML_ASSERT(addr); + if (failed_already) { + return; + } + size_t granularity = lock_granularity(); + target_size = (target_size + granularity - 1) & ~(granularity - 1); + if (target_size > size) { + if (raw_lock((uint8_t *) addr + size, target_size - size)) { + size = target_size; + } else { + failed_already = true; + } + } + } + +#ifdef _POSIX_MEMLOCK_RANGE + static constexpr bool SUPPORTED = true; + + static size_t lock_granularity() { + return (size_t) sysconf(_SC_PAGESIZE); + } + + #ifdef __APPLE__ + #define MLOCK_SUGGESTION \ + "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \ + "decreasing 'vm.global_no_user_wire_amount'. Also try increasing RLIMIT_MEMLOCK (ulimit -l).\n" + #else + #define MLOCK_SUGGESTION \ + "Try increasing RLIMIT_MEMLOCK ('ulimit -l' as root).\n" + #endif + + bool raw_lock(const void * addr, size_t size) const { + if (!mlock(addr, size)) { + return true; + } + + char* errmsg = std::strerror(errno); + bool suggest = (errno == ENOMEM); + + // Check if the resource limit is fine after all + struct rlimit lock_limit; + if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit)) { + suggest = false; + } + if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size)) { + suggest = false; + } + + LLAMA_LOG_WARN("warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s", + size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : ""); + return false; + } + + #undef MLOCK_SUGGESTION + + static void raw_unlock(void * addr, size_t size) { + if (munlock(addr, size)) { + LLAMA_LOG_WARN("warning: failed to munlock buffer: %s\n", std::strerror(errno)); + } + } +#elif defined(_WIN32) + static constexpr bool SUPPORTED = true; + + static size_t lock_granularity() { + SYSTEM_INFO si; + GetSystemInfo(&si); + return (size_t) si.dwPageSize; + } + + bool raw_lock(void * ptr, size_t len) const { + for (int tries = 1; ; tries++) { + if (VirtualLock(ptr, len)) { + return true; + } + if (tries == 2) { + LLAMA_LOG_WARN("warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n", + len, size, llama_format_win_err(GetLastError()).c_str()); + return false; + } + + // It failed but this was only the first try; increase the working + // set size and try again. + SIZE_T min_ws_size, max_ws_size; + if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) { + LLAMA_LOG_WARN("warning: GetProcessWorkingSetSize failed: %s\n", + llama_format_win_err(GetLastError()).c_str()); + return false; + } + // Per MSDN: "The maximum number of pages that a process can lock + // is equal to the number of pages in its minimum working set minus + // a small overhead." + // Hopefully a megabyte is enough overhead: + size_t increment = len + 1048576; + // The minimum must be <= the maximum, so we need to increase both: + min_ws_size += increment; + max_ws_size += increment; + if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) { + LLAMA_LOG_WARN("warning: SetProcessWorkingSetSize failed: %s\n", + llama_format_win_err(GetLastError()).c_str()); + return false; + } + } + } + + static void raw_unlock(void * ptr, size_t len) { + if (!VirtualUnlock(ptr, len)) { + LLAMA_LOG_WARN("warning: failed to VirtualUnlock buffer: %s\n", + llama_format_win_err(GetLastError()).c_str()); + } + } +#else + static constexpr bool SUPPORTED = false; + + static size_t lock_granularity() { + return (size_t) 65536; + } + + bool raw_lock(const void * addr, size_t len) const { + LLAMA_LOG_WARN("warning: mlock not supported on this system\n"); + return false; + } + + static void raw_unlock(const void * addr, size_t len) {} +#endif +}; +using llama_mlocks = std::vector>; + +// NOTE: avoid ever using this except for building the token_to_piece caches +static std::string llama_token_to_piece(const struct llama_model * model, llama_token token, bool special) { + std::string piece; + piece.resize(piece.capacity()); // using string internal cache + const int n_chars = llama_token_to_piece(model, token, &piece[0], piece.size(), 0, special); + if (n_chars < 0) { + piece.resize(-n_chars); + int check = llama_token_to_piece(model, token, &piece[0], piece.size(), 0, special); + GGML_ASSERT(check == -n_chars); + } + else { + piece.resize(n_chars); + } + + return piece; +} + +static ggml_backend_buffer_type_t llama_default_buffer_type_cpu(bool host_buffer) { + ggml_backend_buffer_type_t buft = nullptr; + +#if defined(GGML_USE_CUDA) + // host buffers should only be used when data is expected to be copied to/from the GPU + if (host_buffer) { + buft = ggml_backend_cuda_host_buffer_type(); + } +#elif defined(GGML_USE_SYCL) + if (host_buffer) { + buft = ggml_backend_sycl_host_buffer_type(); + } +#elif defined(GGML_USE_CPU_HBM) + buft = ggml_backend_cpu_hbm_buffer_type(); +#elif defined(GGML_USE_VULKAN) + if (host_buffer) { + buft = ggml_backend_vk_host_buffer_type(); + } +#endif + + if (buft == nullptr) { + buft = ggml_backend_cpu_buffer_type(); + } + return buft; + + GGML_UNUSED(host_buffer); +} + +// +// globals +// + +struct llama_state { + llama_state() { +#ifdef GGML_USE_METAL + ggml_backend_metal_log_set_callback(log_callback, log_callback_user_data); +#elif defined(GGML_USE_CUDA) + ggml_backend_cuda_log_set_callback(log_callback, log_callback_user_data); +#elif defined(GGML_USE_CANN) + ggml_backend_cann_log_set_callback(log_callback, log_callback_user_data); +#endif + } + + // We save the log callback globally + ggml_log_callback log_callback = llama_log_callback_default; + void * log_callback_user_data = nullptr; +}; + +static llama_state g_state; + +// available llama models +enum e_model { + MODEL_UNKNOWN, + MODEL_14M, + MODEL_17M, + MODEL_22M, + MODEL_33M, + MODEL_60M, + MODEL_70M, + MODEL_80M, + MODEL_109M, + MODEL_137M, + MODEL_160M, + MODEL_220M, + MODEL_250M, + MODEL_270M, + MODEL_335M, + MODEL_410M, + MODEL_450M, + MODEL_770M, + MODEL_780M, + MODEL_0_5B, + MODEL_1B, + MODEL_1_3B, + MODEL_1_4B, + MODEL_2B, + MODEL_2_8B, + MODEL_3B, + MODEL_4B, + MODEL_6B, + MODEL_6_9B, + MODEL_7B, + MODEL_8B, + MODEL_9B, + MODEL_11B, + MODEL_12B, + MODEL_13B, + MODEL_14B, + MODEL_15B, + MODEL_16B, + MODEL_20B, + MODEL_30B, + MODEL_34B, + MODEL_35B, + MODEL_40B, + MODEL_65B, + MODEL_70B, + MODEL_236B, + MODEL_314B, + MODEL_SMALL, + MODEL_MEDIUM, + MODEL_LARGE, + MODEL_XL, + MODEL_A2_7B, + MODEL_8x7B, + MODEL_8x22B, + MODEL_16x12B, + MODEL_10B_128x3_66B, + MODEL_57B_A14B, + MODEL_27B, +}; + +static const size_t kiB = 1024; +static const size_t MiB = 1024*kiB; +static const size_t GiB = 1024*MiB; + +struct llama_hparams { + bool vocab_only; + bool rope_finetuned; + bool use_par_res; + + uint32_t n_vocab; + uint32_t n_ctx_train; // context size the model was trained on + uint32_t n_embd; + uint32_t n_layer; + uint32_t n_rot; + uint32_t n_swa = 0; // sliding window attention (SWA) + uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads + uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head + uint32_t n_expert = 0; + uint32_t n_expert_used = 0; + uint32_t n_vocab_type = 0; // for BERT-style token types + uint32_t n_rel_attn_bkts = 0; + + std::array n_head_arr; + std::array n_head_kv_arr; + std::array n_ff_arr; + + uint32_t n_layer_dense_lead = 0; + uint32_t n_lora_q = 0; + uint32_t n_lora_kv = 0; + uint32_t n_ff_exp = 0; + uint32_t n_ff_shexp = 0; + uint32_t n_expert_shared = 0; + float expert_weights_scale = 0.0; + + float f_norm_eps; + float f_norm_rms_eps; + + float f_attn_logit_softcapping = 50.0f; + float f_final_logit_softcapping = 30.0f; + + float rope_attn_factor = 1.0f; + float rope_freq_base_train; + float rope_freq_scale_train; + uint32_t n_ctx_orig_yarn; + float rope_yarn_log_mul; + + // for State Space Models + uint32_t ssm_d_conv = 0; + uint32_t ssm_d_inner = 0; + uint32_t ssm_d_state = 0; + uint32_t ssm_dt_rank = 0; + + float f_clamp_kqv = 0.0f; + float f_max_alibi_bias = 0.0f; + float f_logit_scale = 0.0f; + + bool causal_attn = true; + bool use_alibi = false; + bool attn_soft_cap = false; + + // needed by encoder-decoder models (e.g. T5, FLAN-T5) + // ref: https://github.com/ggerganov/llama.cpp/pull/8141 + llama_token dec_start_token_id = -1; + + enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_NONE; + enum llama_rope_type rope_type = LLAMA_ROPE_TYPE_NONE; + enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE; + + bool operator!=(const llama_hparams & other) const { + if (this->vocab_only != other.vocab_only) return true; + if (this->n_vocab != other.n_vocab) return true; + if (this->n_ctx_train != other.n_ctx_train) return true; + if (this->n_embd != other.n_embd) return true; + if (this->n_layer != other.n_layer) return true; + if (this->n_rot != other.n_rot) return true; + if (this->n_swa != other.n_swa) return true; + if (this->n_embd_head_k != other.n_embd_head_k) return true; + if (this->n_embd_head_v != other.n_embd_head_v) return true; + if (this->n_expert != other.n_expert) return true; + if (this->n_expert_used != other.n_expert_used) return true; + + if (this->n_head_arr != other.n_head_arr) return true; + if (this->n_head_kv_arr != other.n_head_kv_arr) return true; + if (this->n_ff_arr != other.n_ff_arr) return true; + + if (this->n_rel_attn_bkts != other.n_rel_attn_bkts) return true; + if (this->n_layer_dense_lead != other.n_layer_dense_lead) return true; + if (this->n_lora_q != other.n_lora_q) return true; + if (this->n_lora_kv != other.n_lora_kv) return true; + if (this->n_ff_exp != other.n_ff_exp) return true; + if (this->n_ff_shexp != other.n_ff_shexp) return true; + if (this->n_expert_shared != other.n_expert_shared) return true; + + if (this->rope_finetuned != other.rope_finetuned) return true; + if (this->n_ctx_orig_yarn != other.n_ctx_orig_yarn) return true; + + if (this->ssm_d_conv != other.ssm_d_conv) return true; + if (this->ssm_d_inner != other.ssm_d_inner) return true; + if (this->ssm_d_state != other.ssm_d_state) return true; + if (this->ssm_dt_rank != other.ssm_dt_rank) return true; + + if (this->dec_start_token_id != other.dec_start_token_id) return true; + + const float EPSILON = 1e-9f; + + if (!is_float_close(this->f_norm_eps, other.f_norm_eps, EPSILON)) return true; + if (!is_float_close(this->f_norm_rms_eps, other.f_norm_rms_eps, EPSILON)) return true; + if (!is_float_close(this->rope_attn_factor, other.rope_attn_factor, EPSILON)) return true; + if (!is_float_close(this->rope_freq_base_train, other.rope_freq_base_train, EPSILON)) return true; + if (!is_float_close(this->rope_freq_scale_train, other.rope_freq_scale_train, EPSILON)) return true; + if (!is_float_close(this->expert_weights_scale, other.expert_weights_scale, EPSILON)) return true; + if (!is_float_close(this->rope_yarn_log_mul, other.rope_yarn_log_mul, EPSILON)) return true; + + return false; + } + + uint32_t n_head(uint32_t il = 0) const { + if (il < n_layer) { + return n_head_arr[il]; + } + + GGML_ASSERT(false); + return 0; + } + + uint32_t n_head_kv(uint32_t il = 0) const { + if (il < n_layer) { + return n_head_kv_arr[il]; + } + + GGML_ASSERT(false); + return 0; + } + + uint32_t n_ff(uint32_t il = 0) const { + if (il < n_layer) { + return n_ff_arr[il]; + } + + GGML_ASSERT(false); + return 0; + } + + uint32_t n_gqa(uint32_t il = 0) const { + const uint32_t n_head = this->n_head(il); + const uint32_t n_head_kv = this->n_head_kv(il); + + if (n_head_kv == 0) { + return 0; + } + + return n_head/n_head_kv; + } + + uint32_t n_embd_k_gqa(uint32_t il = 0) const { // dimension of key embeddings across all k-v heads + const uint32_t n_head_kv = this->n_head_kv(il); + + return n_embd_head_k * n_head_kv; + } + + uint32_t n_embd_v_gqa(uint32_t il = 0) const { // dimension of value embeddings across all k-v heads + const uint32_t n_head_kv = this->n_head_kv(il); + + return n_embd_head_v * n_head_kv; + } + + uint32_t n_embd_k_s() const { // dimension of the rolling state embeddings + // corresponds to Mamba's conv_states size + // TODO: maybe support other convolution strides than 1 + // NOTE: since the first column of the conv_state is shifted out each time, it's not actually needed + return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * ssm_d_inner; + } + + uint32_t n_embd_v_s() const { // dimension of the recurrent state embeddings + // corresponds to Mamba's ssm_states size + return ssm_d_state * ssm_d_inner; + } +}; + +static_assert(std::is_trivially_copyable::value, "llama_hparams must be trivially copyable"); + +struct llama_cparams { + uint32_t n_ctx; // context size used during inference + uint32_t n_batch; + uint32_t n_ubatch; + uint32_t n_seq_max; + uint32_t n_threads; // number of threads to use for generation + uint32_t n_threads_batch; // number of threads to use for batch processing + + float rope_freq_base; + float rope_freq_scale; + + uint32_t n_ctx_orig_yarn; + // These hyperparameters are not exposed in GGUF, because all + // existing YaRN models use the same values for them. + float yarn_ext_factor; + float yarn_attn_factor; + float yarn_beta_fast; + float yarn_beta_slow; + float defrag_thold; + + bool embeddings; + bool causal_attn; + bool offload_kqv; + bool flash_attn; + + enum llama_pooling_type pooling_type; + + ggml_backend_sched_eval_callback cb_eval; + void * cb_eval_user_data; +}; + +// TODO: separate into "llama_layer_enc" and "llama_layer_dec" +struct llama_layer { + // normalization + struct ggml_tensor * attn_norm; + struct ggml_tensor * attn_norm_b; + struct ggml_tensor * attn_norm_2; + struct ggml_tensor * attn_norm_2_b; + struct ggml_tensor * attn_q_norm; + struct ggml_tensor * attn_q_norm_b; + struct ggml_tensor * attn_k_norm; + struct ggml_tensor * attn_k_norm_b; + struct ggml_tensor * attn_out_norm; + struct ggml_tensor * attn_out_norm_b; + struct ggml_tensor * attn_q_a_norm; + struct ggml_tensor * attn_kv_a_norm; + struct ggml_tensor * attn_sub_norm; + struct ggml_tensor * attn_post_norm; + struct ggml_tensor * ffn_sub_norm; + struct ggml_tensor * attn_norm_cross; + struct ggml_tensor * attn_norm_enc; + + // attention + struct ggml_tensor * wq; + struct ggml_tensor * wk; + struct ggml_tensor * wv; + struct ggml_tensor * wo; + struct ggml_tensor * wqkv; + struct ggml_tensor * wq_a; + struct ggml_tensor * wq_b; + struct ggml_tensor * wkv_a_mqa; + struct ggml_tensor * wkv_b; + struct ggml_tensor * wq_cross; + struct ggml_tensor * wk_cross; + struct ggml_tensor * wv_cross; + struct ggml_tensor * wo_cross; + struct ggml_tensor * wq_enc; + struct ggml_tensor * wk_enc; + struct ggml_tensor * wv_enc; + struct ggml_tensor * wo_enc; + + // attention bias + struct ggml_tensor * bq; + struct ggml_tensor * bk; + struct ggml_tensor * bv; + struct ggml_tensor * bo; + struct ggml_tensor * bqkv; + + // relative position bias + struct ggml_tensor * attn_rel_b; + struct ggml_tensor * attn_rel_b_enc; + struct ggml_tensor * attn_rel_b_cross; + + // normalization + struct ggml_tensor * ffn_norm; + struct ggml_tensor * ffn_norm_b; + struct ggml_tensor * ffn_post_norm; + struct ggml_tensor * layer_out_norm; + struct ggml_tensor * layer_out_norm_b; + struct ggml_tensor * ffn_norm_exps; + struct ggml_tensor * ffn_norm_enc; + + // ff + struct ggml_tensor * ffn_gate; // w1 + struct ggml_tensor * ffn_down; // w2 + struct ggml_tensor * ffn_up; // w3 + struct ggml_tensor * ffn_gate_enc; + struct ggml_tensor * ffn_down_enc; + struct ggml_tensor * ffn_up_enc; + + // ff MoE + struct ggml_tensor * ffn_gate_inp; + struct ggml_tensor * ffn_gate_exps; + struct ggml_tensor * ffn_down_exps; + struct ggml_tensor * ffn_up_exps ; + + // ff shared expert (shexp) + struct ggml_tensor * ffn_gate_inp_shexp; + struct ggml_tensor * ffn_gate_shexp; + struct ggml_tensor * ffn_down_shexp; + struct ggml_tensor * ffn_up_shexp; + + // ff bias + struct ggml_tensor * ffn_gate_b = nullptr; + struct ggml_tensor * ffn_down_b = nullptr; // b2 + struct ggml_tensor * ffn_up_b = nullptr; // b3 + struct ggml_tensor * ffn_act; + + // mamba proj + struct ggml_tensor * ssm_in; + struct ggml_tensor * ssm_x; + struct ggml_tensor * ssm_dt; + struct ggml_tensor * ssm_out; + + // mamba + struct ggml_tensor * ssm_conv1d; + struct ggml_tensor * ssm_a; + struct ggml_tensor * ssm_d; + + // mamba bias + struct ggml_tensor * ssm_conv1d_b; + struct ggml_tensor * ssm_dt_b; + + // long rope factors + struct ggml_tensor * rope_long = nullptr; + struct ggml_tensor * rope_short = nullptr; + + // bitnet scale + struct ggml_tensor * wq_scale; + struct ggml_tensor * wk_scale; + struct ggml_tensor * wv_scale; + struct ggml_tensor * wo_scale; + struct ggml_tensor * ffn_gate_scale; + struct ggml_tensor * ffn_up_scale; + struct ggml_tensor * ffn_down_scale; +}; + +struct llama_kv_cell { + llama_pos pos = -1; + llama_pos delta = 0; + int32_t src = 0; // used by recurrent state models to copy states + + std::set seq_id; + + bool has_seq_id(const llama_seq_id & id) const { + return seq_id.find(id) != seq_id.end(); + } + + bool is_empty() const { + return seq_id.empty(); + } + + bool is_same_seq(const llama_kv_cell & other) const { + return seq_id == other.seq_id; + } +}; + +// ring-buffer of cached KV data +struct llama_kv_cache { + bool has_shift = false; + bool do_defrag = false; + bool do_copy = false; + bool recurrent = false; // with recurrent state models, a cell can hold the state for more than one past token + bool v_trans = true; // the value tensor is transposed + + // Note: The value of head isn't only used to optimize searching + // for a free KV slot. llama_decode_internal also uses it, so it + // cannot be freely changed after a slot has been allocated. + uint32_t head = 0; + uint32_t size = 0; + uint32_t used = 0; // used cells (i.e. at least one seq_id) + + // computed before each graph build + uint32_t n = 0; + + ggml_type type_k = GGML_TYPE_F16; + ggml_type type_v = GGML_TYPE_F16; + + std::vector cells; + + std::vector k_l; // per layer + std::vector v_l; + + std::vector ctxs; + std::vector bufs; + + size_t total_size() const { + size_t size = 0; + for (ggml_backend_buffer_t buf : bufs) { + size += ggml_backend_buffer_get_size(buf); + } + return size; + } + + ~llama_kv_cache() { + for (struct ggml_context * ctx : ctxs) { + ggml_free(ctx); + } + for (ggml_backend_buffer_t buf : bufs) { + ggml_backend_buffer_free(buf); + } + } +}; + +struct llama_control_vector { + std::vector tensors; // per layer + std::vector ctxs; + std::vector bufs; + + int32_t layer_start = -1; + int32_t layer_end = -1; + + struct ggml_tensor * tensor_for(int il) const { + if (il < 0 || il < layer_start || il > layer_end || (size_t) il >= tensors.size()) { + return nullptr; + } + return tensors[il]; + } + + struct ggml_tensor * apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int il) const { + ggml_tensor * layer_dir = tensor_for(il); + if (layer_dir != nullptr) { + cur = ggml_add(ctx, cur, layer_dir); + } + return cur; + } + + ~llama_control_vector() { + for (struct ggml_context * ctx : ctxs) { + ggml_free(ctx); + } + for (ggml_backend_buffer_t buf : bufs) { + ggml_backend_buffer_free(buf); + } + } +}; + +struct llama_model { + e_model type = MODEL_UNKNOWN; + llm_arch arch = LLM_ARCH_UNKNOWN; + llama_ftype ftype = LLAMA_FTYPE_ALL_F32; + + std::string name = "n/a"; + + llama_hparams hparams = {}; + llama_vocab vocab; + + struct ggml_tensor * tok_embd; + struct ggml_tensor * type_embd; + struct ggml_tensor * pos_embd; + struct ggml_tensor * tok_norm; + struct ggml_tensor * tok_norm_b; + + struct ggml_tensor * output_norm; + struct ggml_tensor * output_norm_b; + struct ggml_tensor * output; + struct ggml_tensor * output_b; + struct ggml_tensor * output_norm_enc; + + std::vector layers; + + llama_split_mode split_mode; + int main_gpu; + int n_gpu_layers; + + std::vector rpc_servers; + + // gguf metadata + std::unordered_map gguf_kv; + + // layer -> buffer type mapping + struct layer_buft { + layer_buft() : buft_matrix(nullptr), buft(nullptr) {} + layer_buft(ggml_backend_buffer_type_t matrix) : buft_matrix(matrix), buft(matrix) {} + layer_buft(ggml_backend_buffer_type_t matrix, ggml_backend_buffer_type_t other) : buft_matrix(matrix), buft(other) {} + + ggml_backend_buffer_type_t buft_matrix; // matrices only - used by split buffers and backends that support only matrix multiplication + ggml_backend_buffer_type_t buft; // everything else + }; + + layer_buft buft_input; + layer_buft buft_output; + std::vector buft_layer; + + // contexts where the model tensors metadata is stored + std::vector ctxs; + + // the model memory buffers for the tensor data + std::vector bufs; + + // model memory mapped files + llama_mmaps mappings; + + // objects representing data potentially being locked in memory + llama_mlocks mlock_bufs; + llama_mlocks mlock_mmaps; + + // for quantize-stats only + std::vector> tensors_by_name; + + int64_t t_load_us = 0; + int64_t t_start_us = 0; + + // keep track of loaded lora adapters + std::set lora_adapters; + + ~llama_model() { + for (struct ggml_context * ctx : ctxs) { + ggml_free(ctx); + } + for (ggml_backend_buffer_t buf : bufs) { +#ifdef GGML_USE_CUDA + if (ggml_backend_buffer_get_type(buf) == ggml_backend_cpu_buffer_type()) { + ggml_backend_cuda_unregister_host_buffer(ggml_backend_buffer_get_base(buf)); + } +#endif + ggml_backend_buffer_free(buf); + } + while (!lora_adapters.empty()) { + llama_lora_adapter_free(*lora_adapters.begin()); + } + } +}; + +struct llama_context { + llama_context(const llama_model & model) + : model(model) + , sampling(llama_n_vocab(&model)) + , t_start_us(model.t_start_us) + , t_load_us(model.t_load_us) {} + + ~llama_context() { + ggml_backend_sched_free(sched); + + for (ggml_backend_t backend : backends) { + ggml_backend_free(backend); + } + + ggml_backend_buffer_free(buf_output); + } + + const struct llama_model & model; + + struct llama_cparams cparams; + struct llama_sampling sampling; + struct llama_kv_cache kv_self; + struct llama_control_vector cvec; + + std::unordered_map lora_adapters; + + std::vector backends; +#ifdef GGML_USE_METAL + ggml_backend_t backend_metal = nullptr; +#endif +#ifdef GGML_USE_BLAS + ggml_backend_t backend_blas = nullptr; +#endif + ggml_backend_t backend_cpu = nullptr; + + bool has_evaluated_once = false; + + int64_t t_start_us; + int64_t t_load_us; + int64_t t_p_eval_us = 0; + int64_t t_eval_us = 0; + + int64_t t_compute_start_us = 0; + int64_t n_queued_tokens = 0; + + int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1) + int32_t n_eval = 0; // number of eval calls + + // host buffer for the model output (logits and embeddings) + ggml_backend_buffer_t buf_output = nullptr; + + // decode output (2-dimensional array: [n_outputs][n_vocab]) + size_t logits_size = 0; // capacity (of floats) for logits + float * logits = nullptr; + + std::vector output_ids; // map batch token positions to ids of the logits and embd buffers + size_t output_size = 0; // capacity (of tokens positions) for the output buffers + int32_t n_outputs = 0; // number of actually-used outputs in the current ubatch or last logical batch + + bool logits_all = false; + + // embeddings output (2-dimensional array: [n_outputs][n_embd]) + // populated only when pooling_type == LLAMA_POOLING_TYPE_NONE + size_t embd_size = 0; // capacity (of floats) for embeddings + float * embd = nullptr; + + // sequence embeddings output (map of [n_embd] vectors) + // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE + std::map> embd_seq; + + // whether we are computing encoder output or decoder output + bool is_encoding = false; + + // output of the encoder part of the encoder-decoder models + std::vector embd_enc; + std::vector> seq_ids_enc; + + // memory buffers used to evaluate the model + std::vector buf_compute_meta; + ggml_backend_sched_t sched = nullptr; + + ggml_abort_callback abort_callback = nullptr; + void * abort_callback_data = nullptr; + + // input tensors + struct ggml_tensor * inp_tokens; // I32 [n_batch] + struct ggml_tensor * inp_embd; // F32 [n_embd, n_batch] + struct ggml_tensor * inp_pos; // I32 [n_batch] + struct ggml_tensor * inp_out_ids; // I32 [n_outputs] + struct ggml_tensor * inp_KQ_mask; // F32 [kv_size, n_batch] + struct ggml_tensor * inp_KQ_mask_swa; // F32 [kv_size, n_batch] + struct ggml_tensor * inp_K_shift; // I32 [kv_size] + struct ggml_tensor * inp_mean; // F32 [n_batch, n_batch] + struct ggml_tensor * inp_cls; // I32 [n_batch] + struct ggml_tensor * inp_s_copy; // I32 [kv_size] + struct ggml_tensor * inp_s_mask; // F32 [1, n_kv] + struct ggml_tensor * inp_s_seq; // I32 [n_kv, n_batch] + struct ggml_tensor * inp_pos_bucket; // I32 [n_batch|n_kv, n_batch] + struct ggml_tensor * inp_embd_enc; // F32 [n_embd, n_outputs_enc] + struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch] +}; + +struct llama_lora_weight { + struct ggml_tensor * a = nullptr; + struct ggml_tensor * b = nullptr; + llama_lora_weight() = default; + llama_lora_weight(struct ggml_tensor * a, struct ggml_tensor * b): a(a), b(b) {} +}; + +struct llama_lora_adapter { + struct llama_model * base_model; + // map tensor name to lora_a_b + std::unordered_map ab_map; + std::vector ctxs; + std::vector bufs; + + float alpha; + + llama_lora_adapter(struct llama_model * base_model): base_model(base_model) { + base_model->lora_adapters.insert(this); + } + + llama_lora_weight * get_weight(struct ggml_tensor * w) { + std::string name(w->name); + auto pos = ab_map.find(name); + if (ab_map.find(name) != ab_map.end()) { + return &pos->second; + } + return nullptr; + } + + ~llama_lora_adapter() { + for (struct ggml_context * ctx : ctxs) { + ggml_free(ctx); + } + for (ggml_backend_buffer_t buf : bufs) { + ggml_backend_buffer_free(buf); + } + auto pos = base_model->lora_adapters.find(this); + if (pos != base_model->lora_adapters.end()) { + base_model->lora_adapters.erase(pos); + } + } +}; + +static size_t llama_get_device_count(const llama_model & model) { + size_t count = 1; +#if defined(GGML_USE_CUDA) + count = ggml_backend_cuda_get_device_count(); +#elif defined(GGML_USE_SYCL) + count = ggml_backend_sycl_get_device_count(); +#elif defined(GGML_USE_VULKAN) + count = ggml_backend_vk_get_device_count(); +#elif defined(GGML_USE_CANN) + return ggml_backend_cann_get_device_count(); +#endif +#if defined(GGML_USE_RPC) + count += model.rpc_servers.size(); +#endif + return count; + GGML_UNUSED(model); +} + +static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_model & model, int gpu) { + ggml_backend_buffer_type_t buft = nullptr; + +#if defined(GGML_USE_RPC) + int dev_count = (int)llama_get_device_count(model); + int rpc_count = (int)model.rpc_servers.size(); + if (gpu >= dev_count - rpc_count) { + const char * endpoint = model.rpc_servers[gpu - dev_count + rpc_count].c_str(); + return ggml_backend_rpc_buffer_type(endpoint); + } +#endif +#if defined(GGML_USE_METAL) + buft = ggml_backend_metal_buffer_type(); +#elif defined(GGML_USE_CUDA) + buft = ggml_backend_cuda_buffer_type(gpu); +#elif defined(GGML_USE_VULKAN) + buft = ggml_backend_vk_buffer_type(gpu); +#elif defined(GGML_USE_SYCL) + buft = ggml_backend_sycl_buffer_type(gpu); +#elif defined(GGML_USE_KOMPUTE) + buft = ggml_backend_kompute_buffer_type(gpu); + if (buft == nullptr) { + LLAMA_LOG_WARN("%s: cannot use GPU %d, check `vulkaninfo --summary`\n", __func__, gpu); + } +#elif defined(GGML_USE_CANN) + buft = ggml_backend_cann_buffer_type(gpu); +#endif + + if (buft == nullptr) { + buft = llama_default_buffer_type_cpu(true); + } + return buft; + GGML_UNUSED(model); + GGML_UNUSED(gpu); +} + +static ggml_backend_buffer_type_t llama_default_buffer_type_split(const llama_model & model, int fallback_gpu, const float * tensor_split) { + ggml_backend_buffer_type_t buft = nullptr; + +#ifdef GGML_USE_CUDA + if (ggml_backend_cuda_get_device_count() > 1) { + buft = ggml_backend_cuda_split_buffer_type(tensor_split); + } +#endif + +#ifdef GGML_USE_SYCL + if (ggml_backend_sycl_get_device_count() > 1) { + buft = ggml_backend_sycl_split_buffer_type(tensor_split); + } +#endif + + if (buft == nullptr) { + buft = llama_default_buffer_type_offload(model, fallback_gpu); + } + return buft; + + GGML_UNUSED(tensor_split); +} + +static size_t llama_get_device_memory(const llama_model & model, int device) { +#if defined(GGML_USE_RPC) + int dev_count = (int)llama_get_device_count(model); + int rpc_count = (int)model.rpc_servers.size(); + if (device >= dev_count - rpc_count) { + size_t total; + size_t free; + const char * endpoint = model.rpc_servers[device - dev_count + rpc_count].c_str(); + ggml_backend_rpc_get_device_memory(endpoint, &free, &total); + return free; + } +#endif +#if defined(GGML_USE_CUDA) + size_t total; + size_t free; + ggml_backend_cuda_get_device_memory(device, &free, &total); + return free; +#elif defined(GGML_USE_SYCL) + size_t total; + size_t free; + ggml_backend_sycl_get_device_memory(device, &free, &total); + return free; +#elif defined(GGML_USE_VULKAN) + size_t total; + size_t free; + ggml_backend_vk_get_device_memory(device, &free, &total); + return free; +#elif defined(GGML_USE_CANN) + size_t total; + size_t free; + ggml_backend_cann_get_device_memory(device, &free, &total); + return free; +#else + return 1; +#endif + GGML_UNUSED(model); + GGML_UNUSED(device); +} + +// +// kv cache helpers +// + +static bool llama_kv_cache_init( + struct llama_kv_cache & cache, + const llama_context * ctx, + ggml_type type_k, + ggml_type type_v, + uint32_t kv_size, + bool offload) { + const llama_model & model = ctx->model; + const llama_cparams & cparams = ctx->cparams; + + const struct llama_hparams & hparams = model.hparams; + + const int64_t n_layer = hparams.n_layer; + + cache.has_shift = false; + + // TODO: find a nicer way to add other recurrent model architectures + cache.recurrent = model.arch == LLM_ARCH_MAMBA; + cache.v_trans = !cparams.flash_attn; + + cache.head = 0; + cache.size = kv_size; + cache.used = 0; + + cache.type_k = type_k; + cache.type_v = type_v; + + cache.cells.clear(); + cache.cells.resize(kv_size); + + if (cache.recurrent) { + // init state copy sources + for (uint32_t i = 0; i < cache.size; ++i) { + cache.cells[i].src = i; + } + } + + // count used buffer types + std::map buft_layer_count; + if (offload) { + for (int64_t i = 0; i < n_layer; ++i) { + buft_layer_count[model.buft_layer[i].buft]++; + } + } else { + buft_layer_count[llama_default_buffer_type_cpu(true)] = n_layer; + } + + // create a context for each buffer type + std::map ctx_map; + for (auto & it : buft_layer_count) { + int n_layers = it.second; + struct ggml_init_params params = { + /*.mem_size =*/ 2u*n_layers*ggml_tensor_overhead(), + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ true, + }; + ggml_context * ctx = ggml_init(params); + if (!ctx) { + LLAMA_LOG_ERROR("%s: failed to allocate context for kv cache\n", __func__); + return false; + } + ctx_map[it.first] = ctx; + cache.ctxs.push_back(ctx); + } + + cache.k_l.reserve(n_layer); + cache.v_l.reserve(n_layer); + + for (int i = 0; i < (int) n_layer; i++) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s(); + + struct ggml_context * ctx = offload ? ctx_map.at(model.buft_layer[i].buft) : cache.ctxs.front(); + ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size); + ggml_tensor * v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size); + ggml_format_name(k, "cache_k_l%d", i); + ggml_format_name(v, "cache_v_l%d", i); + cache.k_l.push_back(k); + cache.v_l.push_back(v); + } + + // allocate tensors and initialize the buffers to avoid NaNs in the padding + for (auto it : ctx_map) { + ggml_backend_buffer_type_t buft = it.first; + ggml_context * ctx = it.second; + ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); + if (!buf) { + LLAMA_LOG_ERROR("%s: failed to allocate buffer for kv cache\n", __func__); + return false; + } + ggml_backend_buffer_clear(buf, 0); + LLAMA_LOG_INFO("%s: %10s KV buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0); + cache.bufs.push_back(buf); + } + + return true; +} + +// find an empty slot of size "n_tokens" in the cache +// updates the cache head +// Note: On success, it's important that cache.head points +// to the first cell of the slot. +static bool llama_kv_cache_find_slot( + struct llama_kv_cache & cache, + const struct llama_batch & batch) { + const uint32_t n_tokens = batch.n_tokens; + + if (cache.recurrent) { + // For recurrent state architectures (like Mamba), + // each KV cache cell can store the state for a whole sequence. + + llama_seq_id min = cache.size - 1; + llama_seq_id max = 0; + + for (uint32_t i = 0; i < n_tokens; ++i) { + for (int32_t j = 0; j < batch.n_seq_id[i]; ++j) { + llama_seq_id seq_id = batch.seq_id[i][j]; + // make sure it's a valid seq_id + if ((uint32_t) seq_id < cache.size) { + if (seq_id > max) { + max = seq_id; + } + if (seq_id < min) { + min = seq_id; + } + // Assuming the tokens are in-order + if (batch.pos[i] != cache.cells[seq_id].pos + 1) { + // What should happen when the pos backtracks or skips a value? + // Clearing the state mid-batch would require special-casing which isn't done. + LLAMA_LOG_WARN("%s: non-consecutive token position %d after %d for sequence %d\n", + __func__, batch.pos[i], cache.cells[seq_id].pos, seq_id); + } + if (cache.cells[seq_id].pos < 0 && 0 <= batch.pos[i]) { + cache.used += 1; + } + cache.cells[seq_id].pos = batch.pos[i]; + // NOTE: seq_ids are not inserted here; they are handled when the input tensors are set + } else { + // too big seq_id + // TODO: would it be possible to resize the KV cache size instead? + LLAMA_LOG_ERROR("%s: seq_id=%d >= kv_size=%d Try using a bigger --parallel value\n", __func__, seq_id, cache.size); + return false; + } + } + } + + // allow getting the range of used cells, from head to head + n + cache.head = min; + cache.n = max - min + 1; + + // sanity check + return max >= min; + } + // otherwise, one cell per token. + + if (n_tokens > cache.size) { + LLAMA_LOG_ERROR("%s: n_tokens=%d > cache.size=%d\n", __func__, n_tokens, cache.size); + return false; + } + + uint32_t n_tested = 0; + + while (true) { + if (cache.head + n_tokens > cache.size) { + n_tested += cache.size - cache.head; + cache.head = 0; + continue; + } + + bool found = true; + for (uint32_t i = 0; i < n_tokens; i++) { + if (cache.cells[cache.head + i].pos >= 0) { + found = false; + cache.head += i + 1; + n_tested += i + 1; + break; + } + } + + if (found) { + break; + } + + if (n_tested >= cache.size) { + //LLAMA_LOG_ERROR("%s: failed to find a slot for %d tokens\n", __func__, n_tokens); + return false; + } + } + + for (uint32_t i = 0; i < n_tokens; i++) { + cache.cells[cache.head + i].pos = batch.pos[i]; + + for (int32_t j = 0; j < batch.n_seq_id[i]; j++) { + cache.cells[cache.head + i].seq_id.insert(batch.seq_id[i][j]); + } + } + + cache.used += n_tokens; + + return true; +} + +// find how many cells are currently in use +static uint32_t llama_kv_cache_cell_max(const struct llama_kv_cache & cache) { + for (uint32_t i = cache.size; i > 0; --i) { + const llama_kv_cell & cell = cache.cells[i - 1]; + + if (cell.pos >= 0 && !cell.is_empty()) { + return i; + } + } + + return 0; +} + +static void llama_kv_cache_clear(struct llama_kv_cache & cache) { + for (int32_t i = 0; i < (int32_t) cache.size; ++i) { + cache.cells[i].pos = -1; + cache.cells[i].seq_id.clear(); + } + cache.head = 0; + cache.used = 0; + + for (auto & buf : cache.bufs) { + ggml_backend_buffer_clear(buf, 0); + } +} + +static bool llama_kv_cache_seq_rm( + struct llama_kv_cache & cache, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1) { + uint32_t new_head = cache.size; + + if (p0 < 0) p0 = 0; + if (p1 < 0) p1 = std::numeric_limits::max(); + + // models like Mamba can't have a state partially erased + if (cache.recurrent) { + if (seq_id >= (int64_t) cache.size) { + // could be fatal + return false; + } + if (0 <= seq_id) { + // partial intersection is invalid + if ((0 < p0 && p0 <= cache.cells[seq_id].pos) || (0 < p1 && p1 <= cache.cells[seq_id].pos)) { + return false; + } + } else { + // seq_id is negative, then the range should include everything or nothing + if (p0 != p1 && (p0 != 0 || p1 != std::numeric_limits::max())) { + return false; + } + } + } + + for (uint32_t i = 0; i < cache.size; ++i) { + if (cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { + if (seq_id < 0) { + cache.cells[i].seq_id.clear(); + } else if (cache.cells[i].has_seq_id(seq_id)) { + cache.cells[i].seq_id.erase(seq_id); + } else { + continue; + } + if (cache.cells[i].is_empty()) { + // keep count of the number of used cells + if (cache.cells[i].pos >= 0) cache.used--; + + cache.cells[i].pos = -1; + if (new_head == cache.size) new_head = i; + } + } + } + + // If we freed up a slot, set head to it so searching can start there. + if (new_head != cache.size && new_head < cache.head) cache.head = new_head; + + return true; +} + +static void llama_kv_cache_seq_cp( + struct llama_kv_cache & cache, + llama_seq_id seq_id_src, + llama_seq_id seq_id_dst, + llama_pos p0, + llama_pos p1) { + if (p0 < 0) p0 = 0; + if (p1 < 0) p1 = std::numeric_limits::max(); + + if (cache.recurrent) { + if ((uint32_t) seq_id_dst < cache.size && (uint32_t) seq_id_src < cache.size) { + seq_id_src = cache.cells[seq_id_src].src; + GGML_ASSERT((uint32_t) seq_id_src < cache.size); + // intent to "copy from" + // supports copy chains thanks to taking the source of the source + cache.cells[seq_id_dst].src = seq_id_src; + + // preserve the "keep or clear" status of the copied sequence + if (cache.cells[seq_id_src].has_seq_id(seq_id_src)) { + cache.cells[seq_id_dst].seq_id.insert(seq_id_dst); + } else { + cache.cells[seq_id_dst].seq_id.erase(seq_id_dst); + } + + cache.do_copy = true; + + cache.cells[seq_id_dst].pos = cache.cells[seq_id_src].pos; + } + return; + } + // otherwise, this is the KV cache of a Transformer-like model + + cache.head = 0; + + for (uint32_t i = 0; i < cache.size; ++i) { + if (cache.cells[i].has_seq_id(seq_id_src) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { + cache.cells[i].seq_id.insert(seq_id_dst); + } + } +} + +static void llama_kv_cache_seq_keep(struct llama_kv_cache & cache, llama_seq_id seq_id) { + uint32_t new_head = cache.size; + + for (uint32_t i = 0; i < cache.size; ++i) { + if (!cache.cells[i].has_seq_id(seq_id)) { + if (cache.cells[i].pos >= 0) cache.used--; + cache.cells[i].pos = -1; + cache.cells[i].seq_id.clear(); + if (new_head == cache.size) new_head = i; + } else { + cache.cells[i].seq_id.clear(); + cache.cells[i].seq_id.insert(seq_id); + } + } + + // If we freed up a slot, set head to it so searching can start there. + if (new_head != cache.size && new_head < cache.head) cache.head = new_head; +} + +static void llama_kv_cache_seq_add( + struct llama_kv_cache & cache, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + llama_pos delta) { + uint32_t new_head = cache.size; + + if (p0 < 0) p0 = 0; + if (p1 < 0) p1 = std::numeric_limits::max(); + // If there is no range then return early to avoid looping over the cache. + if (p0 == p1) return; + + if (cache.recurrent) { + // for Mamba-like models, only the pos needs to be shifted + if (0 <= seq_id && seq_id < (int64_t) cache.size) { + llama_kv_cell & cell = cache.cells[seq_id]; + if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) { + cell.pos += delta; + } + } + return; + } + + for (uint32_t i = 0; i < cache.size; ++i) { + if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { + cache.has_shift = true; + cache.cells[i].pos += delta; + cache.cells[i].delta += delta; + + if (cache.cells[i].pos < 0) { + if (!cache.cells[i].is_empty()) { + cache.used--; + } + cache.cells[i].pos = -1; + cache.cells[i].seq_id.clear(); + if (new_head == cache.size) { + new_head = i; + } + } + } + } + + // If we freed up a slot, set head to it so searching can start there. + // Otherwise we just start the next search from the beginning. + cache.head = new_head != cache.size ? new_head : 0; +} + +static void llama_kv_cache_seq_div( + struct llama_kv_cache & cache, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + int d) { + if (p0 < 0) p0 = 0; + if (p1 < 0) p1 = std::numeric_limits::max(); + // If there is no range then return early to avoid looping over the cache. + if (p0 == p1) return; + + if (cache.recurrent) { + // for Mamba-like models, only the pos needs to be changed + if (0 <= seq_id && seq_id < (int64_t) cache.size) { + llama_kv_cell & cell = cache.cells[seq_id]; + if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) { + cell.pos /= d; + } + } + return; + } + + for (uint32_t i = 0; i < cache.size; ++i) { + if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { + cache.has_shift = true; + + { + llama_pos p_old = cache.cells[i].pos; + cache.cells[i].pos /= d; + cache.cells[i].delta += cache.cells[i].pos - p_old; + } + } + } +} + +static llama_pos llama_kv_cache_seq_pos_max(struct llama_kv_cache & cache, llama_seq_id seq_id) { + llama_pos result = 0; + + for (uint32_t i = 0; i < cache.size; ++i) { + if (cache.cells[i].has_seq_id(seq_id)) { + result = std::max(result, cache.cells[i].pos); + } + } + + return result; +} + +static void llama_kv_cache_defrag(struct llama_kv_cache & cache) { + cache.do_defrag = true; +} + +static uint32_t llama_kv_cache_get_padding(const struct llama_cparams & cparams) { + // the FA kernels require padding to avoid extra runtime boundary checks + return cparams.flash_attn ? 256u : 32u; +} + +// +// model loading and saving +// + +enum llama_fver { + GGUF_FILE_VERSION_V1 = 1, + GGUF_FILE_VERSION_V2 = 2, + GGUF_FILE_VERSION_V3 = 3, +}; + +static const char * llama_file_version_name(llama_fver version) { + switch (version) { + case GGUF_FILE_VERSION_V1: return "GGUF V1 (support until nov 2023)"; + case GGUF_FILE_VERSION_V2: return "GGUF V2"; + case GGUF_FILE_VERSION_V3: return "GGUF V3 (latest)"; + } + + return "unknown"; +} + +static std::string llama_format_tensor_shape(const std::vector & ne) { + char buf[256]; + snprintf(buf, sizeof(buf), "%5" PRId64, ne.at(0)); + for (size_t i = 1; i < ne.size(); i++) { + snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, ne.at(i)); + } + return buf; +} + +static std::string llama_format_tensor_shape(const struct ggml_tensor * t) { + char buf[256]; + snprintf(buf, sizeof(buf), "%5" PRId64, t->ne[0]); + for (int i = 1; i < GGML_MAX_DIMS; i++) { + snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, t->ne[i]); + } + return buf; +} + +namespace GGUFMeta { + template + struct GKV_Base_Type { + static constexpr gguf_type gt = gt_; + + static T getter(const gguf_context * ctx, const int kid) { + return gfun(ctx, kid); + } + }; + + template struct GKV_Base; + + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + template<> struct GKV_Base: GKV_Base_Type {}; + + template<> struct GKV_Base { + static constexpr gguf_type gt = GGUF_TYPE_STRING; + + static std::string getter(const gguf_context * ctx, const int kid) { + return gguf_get_val_str(ctx, kid); + } + }; + + struct ArrayInfo { + const gguf_type gt; + const size_t length; + const void * data; + }; + + template<> struct GKV_Base { + public: + static constexpr gguf_type gt = GGUF_TYPE_ARRAY; + static ArrayInfo getter(const gguf_context *ctx, const int k) { + return ArrayInfo { + gguf_get_arr_type(ctx, k), + size_t(gguf_get_arr_n(ctx, k)), + gguf_get_arr_data(ctx, k), + }; + } + }; + + template + class GKV : public GKV_Base { + GKV() = delete; + + public: + static T get_kv(const gguf_context * ctx, const int k) { + const enum gguf_type kt = gguf_get_kv_type(ctx, k); + + if (kt != GKV::gt) { + throw std::runtime_error(format("key %s has wrong type %s but expected type %s", + gguf_get_key(ctx, k), gguf_type_name(kt), gguf_type_name(GKV::gt))); + } + return GKV::getter(ctx, k); + } + + static const char * override_type_to_str(const llama_model_kv_override_type ty) { + switch (ty) { + case LLAMA_KV_OVERRIDE_TYPE_BOOL: return "bool"; + case LLAMA_KV_OVERRIDE_TYPE_INT: return "int"; + case LLAMA_KV_OVERRIDE_TYPE_FLOAT: return "float"; + case LLAMA_KV_OVERRIDE_TYPE_STR: return "str"; + } + return "unknown"; + } + + static bool validate_override(const llama_model_kv_override_type expected_type, const struct llama_model_kv_override * ovrd) { + if (!ovrd) { return false; } + if (ovrd->tag == expected_type) { + LLAMA_LOG_INFO("%s: Using metadata override (%5s) '%s' = ", + __func__, override_type_to_str(ovrd->tag), ovrd->key); + switch (ovrd->tag) { + case LLAMA_KV_OVERRIDE_TYPE_BOOL: { + LLAMA_LOG_INFO("%s\n", ovrd->val_bool ? "true" : "false"); + } break; + case LLAMA_KV_OVERRIDE_TYPE_INT: { + LLAMA_LOG_INFO("%" PRId64 "\n", ovrd->val_i64); + } break; + case LLAMA_KV_OVERRIDE_TYPE_FLOAT: { + LLAMA_LOG_INFO("%.6f\n", ovrd->val_f64); + } break; + case LLAMA_KV_OVERRIDE_TYPE_STR: { + LLAMA_LOG_INFO("%s\n", ovrd->val_str); + } break; + default: + // Shouldn't be possible to end up here, but just in case... + throw std::runtime_error( + format("Unsupported attempt to override %s type for metadata key %s\n", + override_type_to_str(ovrd->tag), ovrd->key)); + } + return true; + } + LLAMA_LOG_WARN("%s: Warning: Bad metadata override type for key '%s', expected %s but got %s\n", + __func__, ovrd->key, override_type_to_str(expected_type), override_type_to_str(ovrd->tag)); + return false; + } + + template + static typename std::enable_if::value, bool>::type + try_override(OT & target, const struct llama_model_kv_override * ovrd) { + if (validate_override(LLAMA_KV_OVERRIDE_TYPE_BOOL, ovrd)) { + target = ovrd->val_bool; + return true; + } + return false; + } + + template + static typename std::enable_if::value && std::is_integral::value, bool>::type + try_override(OT & target, const struct llama_model_kv_override * ovrd) { + if (validate_override(LLAMA_KV_OVERRIDE_TYPE_INT, ovrd)) { + target = ovrd->val_i64; + return true; + } + return false; + } + + template + static typename std::enable_if::value, bool>::type + try_override(T & target, const struct llama_model_kv_override * ovrd) { + if (validate_override(LLAMA_KV_OVERRIDE_TYPE_FLOAT, ovrd)) { + target = ovrd->val_f64; + return true; + } + return false; + } + + template + static typename std::enable_if::value, bool>::type + try_override(T & target, const struct llama_model_kv_override * ovrd) { + if (validate_override(LLAMA_KV_OVERRIDE_TYPE_STR, ovrd)) { + target = ovrd->val_str; + return true; + } + return false; + } + + static bool set(const gguf_context * ctx, const int k, T & target, const struct llama_model_kv_override * ovrd = nullptr) { + if (try_override(target, ovrd)) { + return true; + } + if (k < 0) { return false; } + target = get_kv(ctx, k); + return true; + } + + static bool set(const gguf_context * ctx, const char * key, T & target, const struct llama_model_kv_override * ovrd = nullptr) { + return set(ctx, gguf_find_key(ctx, key), target, ovrd); + } + + static bool set(const gguf_context * ctx, const std::string & key, T & target, const struct llama_model_kv_override * ovrd = nullptr) { + return set(ctx, key.c_str(), target, ovrd); + } + }; +} + +using llama_buf_map = std::unordered_map; + +struct llama_model_loader { + int n_kv = 0; + int n_tensors = 0; + int n_created = 0; + + int64_t n_elements = 0; + size_t n_bytes = 0; + + bool use_mmap = false; + bool check_tensors; + + llama_files files; + llama_ftype ftype; + llama_fver fver; + + llama_mmaps mappings; + + // Holds information on a model weight + struct llama_tensor_weight { + uint16_t idx; // source file index + size_t offs; // tensor data offset in the original file + + ggml_tensor * tensor; + + llama_tensor_weight(const llama_file * file, uint16_t idx, const char * name, const struct gguf_context * gguf_ctx, ggml_tensor * tensor) : idx(idx), tensor(tensor) { + const int tensor_idx = gguf_find_tensor(gguf_ctx, name); + offs = gguf_get_data_offset(gguf_ctx) + gguf_get_tensor_offset(gguf_ctx, tensor_idx); + + if (offs + ggml_nbytes(tensor) < offs || offs + ggml_nbytes(tensor) > file->size) { + throw std::runtime_error(format("tensor '%s' data is not within the file bounds, model is corrupted or incomplete", name)); + } + } + }; + std::vector weights; + + std::unordered_map kv_overrides; + + struct gguf_context * meta = NULL; + std::vector contexts; + + std::string arch_name; + LLM_KV llm_kv = LLM_KV(LLM_ARCH_UNKNOWN); + + llama_model_loader(const std::string & fname, bool use_mmap, bool check_tensors, const struct llama_model_kv_override * param_overrides_p) { + int trace = 0; + if (getenv("LLAMA_TRACE")) { + trace = atoi(getenv("LLAMA_TRACE")); + } + + if (param_overrides_p != nullptr) { + for (const struct llama_model_kv_override * p = param_overrides_p; p->key[0] != 0; p++) { + kv_overrides.insert({std::string(p->key), *p}); + } + } + + struct ggml_context * ctx = NULL; + struct gguf_init_params params = { + /*.no_alloc = */ true, + /*.ctx = */ &ctx, + }; + + meta = gguf_init_from_file(fname.c_str(), params); + if (!meta) { + throw std::runtime_error(format("%s: failed to load model from %s\n", __func__, fname.c_str())); + } + + get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false); + llm_kv = LLM_KV(llm_arch_from_string(arch_name)); + + files.emplace_back(new llama_file(fname.c_str(), "rb")); + contexts.emplace_back(ctx); + + // Save tensors data offset of the main file. + // For subsidiary files, `meta` tensor data offset must not be used, + // so we build a unified tensors index for weights. + for (ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) { + weights.emplace_back(files.back().get(), 0, cur->name, meta, cur); + } + uint16_t n_split = 0; + get_key(llm_kv(LLM_KV_SPLIT_COUNT), n_split, false); + + // Load additional GGML contexts + if (n_split > 1) { + uint16_t idx = 0; + get_key(llm_kv(LLM_KV_SPLIT_NO), idx); + if (idx != 0) { + throw std::runtime_error(format("illegal split file: %d, model must be loaded with the first split", idx)); + } + + char split_prefix[PATH_MAX] = {0}; + if (!llama_split_prefix(split_prefix, sizeof(split_prefix), fname.c_str(), idx, n_split)) { + throw std::runtime_error(format("invalid split file: %s", fname.c_str())); + } + + if (trace > 0) { + LLAMA_LOG_INFO("%s: loading additional %d GGUFs\n", __func__, n_split); + } + + char split_path[PATH_MAX] = {0}; + for (idx = 1; idx < n_split; idx++) { + llama_split_path(split_path, sizeof(split_path), split_prefix, idx, n_split); + + struct gguf_init_params split_params = { + /*.no_alloc = */ true, + /*.ctx = */ &ctx, + }; + struct gguf_context * ctx_gguf = gguf_init_from_file(split_path, split_params); + if (!ctx_gguf) { + throw std::runtime_error(format("%s: failed to load GGUF split from %s\n", __func__, split_path)); + } + + files.emplace_back(new llama_file(split_path, "rb")); + contexts.emplace_back(ctx); + + // Save tensors data offset info of the shard. + for (ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) { + weights.emplace_back(files.back().get(), idx, cur->name, ctx_gguf, cur); + } + + gguf_free(ctx_gguf); + } + + get_key(llm_kv(LLM_KV_SPLIT_TENSORS_COUNT), n_tensors); + + // sanity check + { + const int n_tensors_loaded = (int) weights.size(); + if (n_tensors != n_tensors_loaded) { + throw std::runtime_error(format("corrupted model: %d tensors expected but %d found", n_tensors, n_tensors_loaded)); + } + } + + LLAMA_LOG_INFO("%s: additional %d GGUFs metadata loaded.\n", __func__, n_split - 1); + } + + n_kv = gguf_get_n_kv(meta); + n_tensors = weights.size(); + + fver = (enum llama_fver) gguf_get_version(meta); + + std::set tensor_names; + for (auto & w : weights) { + n_elements += ggml_nelements(w.tensor); + n_bytes += ggml_nbytes(w.tensor); + // make sure there is no duplicated tensor names + const std::string name(w.tensor->name); + auto found = tensor_names.find(name); + if (found != tensor_names.end()) { + throw std::runtime_error(format("invalid model: tensor '%s' is duplicated", w.tensor->name)); + } + tensor_names.insert(name); + } + + LLAMA_LOG_INFO("%s: loaded meta data with %d key-value pairs and %d tensors from %s (version %s)\n", + __func__, n_kv, n_tensors, fname.c_str(), llama_file_version_name(fver)); + + // determine file type based on the number of tensors for each quantization and print meta data + // TODO: make optional + { + std::map n_type; + + uint32_t n_type_max = 0; + enum ggml_type type_max = GGML_TYPE_F32; + + for (int i = 0; i < n_tensors; i++) { + const ggml_tensor * tensor = weights.at(i).tensor; + enum ggml_type type = tensor->type; + + n_type[type]++; + + if (n_type_max < n_type[type]) { + n_type_max = n_type[type]; + type_max = type; + } + + if (trace > 0) { + const uint16_t sid = weights.at(i).idx; + LLAMA_LOG_INFO("%s: - tensor %4d, split %2d: %32s %-8s [ %s ]\n", __func__, i, sid, ggml_get_name(tensor), ggml_type_name(type), llama_format_tensor_shape(tensor).c_str()); + } + } + + switch (type_max) { + case GGML_TYPE_F32: ftype = LLAMA_FTYPE_ALL_F32; break; + case GGML_TYPE_F16: ftype = LLAMA_FTYPE_MOSTLY_F16; break; + case GGML_TYPE_BF16: ftype = LLAMA_FTYPE_MOSTLY_BF16; break; + case GGML_TYPE_Q4_0: ftype = LLAMA_FTYPE_MOSTLY_Q4_0; break; + case GGML_TYPE_Q4_1: ftype = LLAMA_FTYPE_MOSTLY_Q4_1; break; + case GGML_TYPE_Q5_0: ftype = LLAMA_FTYPE_MOSTLY_Q5_0; break; + case GGML_TYPE_Q5_1: ftype = LLAMA_FTYPE_MOSTLY_Q5_1; break; + case GGML_TYPE_Q8_0: ftype = LLAMA_FTYPE_MOSTLY_Q8_0; break; + case GGML_TYPE_Q2_K: ftype = LLAMA_FTYPE_MOSTLY_Q2_K; break; + case GGML_TYPE_Q3_K: ftype = LLAMA_FTYPE_MOSTLY_Q3_K_M; break; + case GGML_TYPE_Q4_K: ftype = LLAMA_FTYPE_MOSTLY_Q4_K_M; break; + case GGML_TYPE_Q5_K: ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M; break; + case GGML_TYPE_Q6_K: ftype = LLAMA_FTYPE_MOSTLY_Q6_K; break; + case GGML_TYPE_IQ2_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS; break; + case GGML_TYPE_IQ2_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XS; break; + case GGML_TYPE_IQ2_S: ftype = LLAMA_FTYPE_MOSTLY_IQ2_S; break; + case GGML_TYPE_IQ3_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ3_XXS; break; + case GGML_TYPE_IQ1_S: ftype = LLAMA_FTYPE_MOSTLY_IQ1_S; break; + case GGML_TYPE_IQ1_M: ftype = LLAMA_FTYPE_MOSTLY_IQ1_M; break; + case GGML_TYPE_IQ1_BN: ftype = LLAMA_FTYPE_MOSTLY_IQ1_BN; break; + case GGML_TYPE_IQ2_BN: ftype = LLAMA_FTYPE_MOSTLY_IQ2_BN; break; + case GGML_TYPE_IQ4_NL: ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL; break; + case GGML_TYPE_IQ4_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS; break; + case GGML_TYPE_IQ3_S: ftype = LLAMA_FTYPE_MOSTLY_IQ3_S; break; + case GGML_TYPE_Q4_0_4_4: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_4; break; + case GGML_TYPE_Q4_0_4_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_8; break; + case GGML_TYPE_Q4_0_8_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_8_8; break; + default: + { + LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max)); + ftype = LLAMA_FTYPE_ALL_F32; + } break; + } + + // this is a way to mark that we have "guessed" the file type + ftype = (llama_ftype) (ftype | LLAMA_FTYPE_GUESSED); + + { + const int kid = gguf_find_key(meta, "general.file_type"); // TODO: use LLM_KV + if (kid >= 0) { + ftype = (llama_ftype) gguf_get_val_u32(meta, kid); + } + } + + LLAMA_LOG_INFO("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__); + + for (int i = 0; i < n_kv; i++) { + const char * name = gguf_get_key(meta, i); + const enum gguf_type type = gguf_get_kv_type(meta, i); + const std::string type_name = + type == GGUF_TYPE_ARRAY + ? format("%s[%s,%d]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(meta, i)), gguf_get_arr_n(meta, i)) + : gguf_type_name(type); + + std::string value = gguf_kv_to_str(meta, i); + const size_t MAX_VALUE_LEN = 40; + if (value.size() > MAX_VALUE_LEN) { + value = format("%s...", value.substr(0, MAX_VALUE_LEN - 3).c_str()); + } + replace_all(value, "\n", "\\n"); + + LLAMA_LOG_INFO("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str()); + } + + // print type counts + for (auto & kv : n_type) { + if (kv.second == 0) { + continue; + } + + LLAMA_LOG_INFO("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second); + } + } + + if (!llama_mmap::SUPPORTED) { + LLAMA_LOG_WARN("%s: mmap is not supported on this platform\n", __func__); + use_mmap = false; + } + + this->use_mmap = use_mmap; + this->check_tensors = check_tensors; + } + + ~llama_model_loader() { + if (meta) { + gguf_free(meta); + } + for (auto * ctx : contexts) { + ggml_free(ctx); + } + } + + template + typename std::enable_if::value, bool>::type + get_arr_n(const std::string & key, T & result, const bool required = true) { + const int kid = gguf_find_key(meta, key.c_str()); + + if (kid < 0) { + if (required) { + throw std::runtime_error(format("key not found in model: %s", key.c_str())); + } + return false; + } + + struct GGUFMeta::ArrayInfo arr_info = + GGUFMeta::GKV::get_kv(meta, kid); + + + result = arr_info.length; + return true; + } + + template + typename std::enable_if::value, bool>::type + get_arr_n(const enum llm_kv kid, T & result, const bool required = true) { + return get_arr_n(llm_kv(kid), result, required); + } + + template + bool get_arr(const std::string & key, std::vector & result, const bool required = true) { + const int kid = gguf_find_key(meta, key.c_str()); + + if (kid < 0 || gguf_get_kv_type(meta, kid) != GGUF_TYPE_ARRAY) { + if (required) { + throw std::runtime_error(format("array key not found in model: %s", key.c_str())); + } + return false; + } + + struct GGUFMeta::ArrayInfo arr_info = + GGUFMeta::GKV::get_kv(meta, kid); + + switch (arr_info.gt) { + case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same::value)); break; + case GGUF_TYPE_INT32: GGML_ASSERT( + (std::is_same::value) || + (std::is_same::value)); break; + default: + throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str())); + } + + result.resize(arr_info.length); + result.assign((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length); + + return true; + } + + template + bool get_arr(const std::string & key, std::array & result, const bool required = true) { + const int kid = gguf_find_key(meta, key.c_str()); + + if (kid < 0 || gguf_get_kv_type(meta, kid) != GGUF_TYPE_ARRAY) { + if (required) { + throw std::runtime_error(format("array key not found in model: %s", key.c_str())); + } + return false; + } + + struct GGUFMeta::ArrayInfo arr_info = + GGUFMeta::GKV::get_kv(meta, kid); + + switch (arr_info.gt) { + case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same::value)); break; + case GGUF_TYPE_INT32: GGML_ASSERT( + (std::is_same::value) || + (std::is_same::value)); break; + default: + throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str())); + } + + if (arr_info.length > N_MAX) { + throw std::runtime_error(format("array length %u for key %s exceeds max %u", (uint32_t) arr_info.length, key.c_str(), (uint32_t) N_MAX)); + } + + std::copy((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length, result.begin()); + + return true; + } + + template + bool get_arr(const enum llm_kv kid, T & result, const bool required = true) { + return get_arr(llm_kv(kid), result, required); + } + + template + bool get_key(const std::string & key, T & result, const bool required = true) { + auto it = kv_overrides.find(key); + + const struct llama_model_kv_override * override = + it != kv_overrides.end() ? &it->second : nullptr; + + const bool found = GGUFMeta::GKV::set(meta, key, result, override); + + if (required && !found) { + throw std::runtime_error(format("key not found in model: %s", key.c_str())); + } + + return found; + } + + template + bool get_key(const enum llm_kv kid, T & result, const bool required = true) { + return get_key(llm_kv(kid), result, required); + } + + // get array of n <= N_MAX elements, or a single element repeated n times + template + bool get_key_or_arr(const std::string & key, std::array & result, uint32_t n, const bool required = true) { + const int kid = gguf_find_key(meta, key.c_str()); + + if (kid < 0) { + if (required) { + throw std::runtime_error(format("key not found in model: %s", key.c_str())); + } + return false; + } + + if (n > N_MAX) { + throw std::runtime_error(format("n > N_MAX: %u > %u for key %s", (uint32_t) n, (uint32_t) N_MAX, key.c_str())); + } + + if (gguf_get_kv_type(meta, kid) == GGUF_TYPE_ARRAY) { + struct GGUFMeta::ArrayInfo arr_info = + GGUFMeta::GKV::get_kv(meta, kid); + + if (n != arr_info.length) { + throw std::runtime_error(format("key %s has wrong array length; expected %u, got %u", key.c_str(), n, (uint32_t) arr_info.length)); + } + + return get_arr(key, result, required); + } else { + T value; + + bool ok = get_key(key, value, required); + if (!ok) { + return false; + } + + for (uint32_t i = 0; i < n; i++) { + result[i] = value; + } + + return true; + } + } + + template + bool get_key_or_arr(const enum llm_kv kid, T & result, uint32_t n, const bool required = true) { + return get_key_or_arr(llm_kv(kid), result, n, required); + } + + std::string get_arch_name() const { + return arch_name; + } + + enum llm_arch get_arch() const { + return llm_kv.arch; + } + + const char * get_tensor_name(int i) const { + return weights.at(i).tensor->name; + } + + const llama_tensor_weight * get_weight(const char * name) const { + for (const auto & weight : weights) { + if (strcmp(name, weight.tensor->name) == 0) { + return &weight; + } + } + return nullptr; + } + + const llama_tensor_weight * get_weight(int i) const { + return get_weight(get_tensor_name(i)); + } + + const llama_tensor_weight & require_weight(const char * name) const { + const llama_tensor_weight * weight = get_weight(name); + if (!weight) { + throw std::runtime_error(format("%s: tensor '%s' not found", __func__, name)); + } + return *weight; + } + + struct ggml_tensor * get_tensor_meta(const char * name) const { + const auto * weight = get_weight(name); + if (!weight) { + return nullptr; + } + return weight->tensor; + } + + struct ggml_tensor * require_tensor_meta(const char * name) const { + struct ggml_tensor * tensor = get_tensor_meta(name); + if (!tensor) { + throw std::runtime_error(format("%s: tensor '%s' not found", __func__, name)); + } + return tensor; + } + + struct ggml_tensor * get_tensor_meta(int i) const { + return get_tensor_meta(get_tensor_name(i)); + } + + struct ggml_tensor * create_tensor_for(struct ggml_context * ctx, const struct ggml_tensor * cur, bool duplicated) { + struct ggml_tensor * tensor = ggml_dup_tensor(ctx, cur); + ggml_set_name(tensor, ggml_get_name(cur)); + + if (duplicated) { + size_data += ggml_nbytes(cur); + } else { + n_created++; + } + + return tensor; + } + + const struct ggml_tensor * check_tensor_dims(const std::string & name, const std::vector & ne, bool required) const { + const struct ggml_tensor * cur = get_tensor_meta(name.c_str()); + + if (cur == NULL) { + if (!required) { + return NULL; + } + throw std::runtime_error(format("%s: tensor '%s' not found", __func__, name.c_str())); + } + + { + bool is_ok = true; + for (size_t i = 0; i < GGML_MAX_DIMS; ++i) { + if ((i < ne.size() && ne[i] != cur->ne[i]) || (i >= ne.size() && cur->ne[i] != 1)) { + is_ok = false; + break; + } + } + if (!is_ok) { + throw std::runtime_error( + format("%s: tensor '%s' has wrong shape; expected %s, got %s", + __func__, name.c_str(), + llama_format_tensor_shape(ne).c_str(), + llama_format_tensor_shape(cur).c_str())); + } + } + + return cur; + } + + static const int TENSOR_NOT_REQUIRED = 1; + static const int TENSOR_DUPLICATED = 2; + + struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector & ne, int flags = 0) { + const struct ggml_tensor * cur = check_tensor_dims(name, ne, !(flags & TENSOR_NOT_REQUIRED)); + + if (cur == NULL) { + return NULL; + } + + return create_tensor_for(ctx, cur, flags & TENSOR_DUPLICATED); + } + + struct ggml_tensor * create_tensor_as_view(struct ggml_context * ctx, struct ggml_tensor * base, const std::string & name, const std::vector & ne, size_t offset, bool required = true) { + const struct ggml_tensor * cur = check_tensor_dims(name, ne, required); + + if (cur == NULL) { + return NULL; + } + + if (cur->type != base->type) { + throw std::runtime_error(format("%s: tensor '%s' has wrong type; expected %s, got %s", __func__, name.c_str(), ggml_type_name(base->type), ggml_type_name(cur->type))); + } + + std::array dims; + for (size_t i = 0; i < GGML_MAX_DIMS; ++i) { + dims[i] = i < ne.size() ? ne[i] : 1; + } + + struct ggml_tensor * tensor = ggml_view_4d(ctx, base, + dims[0], dims[1], dims[2], dims[3], + cur->nb[1], cur->nb[2], cur->nb[3], + offset); + + ggml_set_name(tensor, name.c_str()); + + n_created++; + + return tensor; + } + + void done_getting_tensors() const { + if (n_created != n_tensors) { + throw std::runtime_error(format("%s: wrong number of tensors; expected %d, got %d", __func__, n_tensors, n_created)); + } + } + + void init_mappings(bool prefetch = true, llama_mlocks * mlock_mmaps = nullptr) { + if (use_mmap) { + mappings.reserve(files.size()); + mmaps_used.reserve(files.size()); + for (const auto & file : files) { + std::unique_ptr mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, ggml_is_numa())); + mmaps_used.emplace_back(mapping->size, 0); + if (mlock_mmaps) { + std::unique_ptr mlock_mmap(new llama_mlock()); + mlock_mmap->init(mapping->addr); + mlock_mmaps->emplace_back(std::move(mlock_mmap)); + } + mappings.emplace_back(std::move(mapping)); + } + } + + // compute the total size of all tensors for progress reporting + for (auto & w : weights) { + size_data += ggml_nbytes(w.tensor); + } + } + + void get_mapping_range(size_t * first, size_t * last, void ** addr, int idx, ggml_context * ctx) const { + GGML_ASSERT(!mappings.empty()); + const auto & mapping = mappings.at(idx); + + *first = mapping->size; + *last = 0; + *addr = mapping->addr; + for (ggml_tensor * tensor = ggml_get_first_tensor(ctx); tensor; tensor = ggml_get_next_tensor(ctx, tensor)) { + try { + const auto * weight = get_weight(ggml_get_name(tensor)); + if (!weight) { + continue; + } + if (weight->idx != idx) { + continue; + } + *first = std::min(*first, weight->offs); + *last = std::max(*last, weight->offs + ggml_nbytes(tensor)); + } catch(...) { + // the tensor is not in the model + } + } + } + + // for backwards compatibility, does not support ggml-backend + void load_data_for(struct ggml_tensor * cur) const { + const auto & w = require_weight(ggml_get_name(cur)); + + if (use_mmap) { + const auto & mapping = mappings.at(w.idx); + if (cur->data == nullptr) { + cur->data = (uint8_t *)mapping->addr + w.offs; + } else { + memcpy(cur->data, (uint8_t *)mapping->addr + w.offs, ggml_nbytes(cur)); + } + } else { + GGML_ASSERT(cur->data != nullptr); + GGML_ASSERT(w.idx < files.size()); + const auto & file = files.at(w.idx); + file->seek(w.offs, SEEK_SET); + file->read_raw(cur->data, ggml_nbytes(cur)); + } + + if (check_tensors && !ggml_validate_row_data(cur->type, cur->data, ggml_nbytes(cur))) { + throw std::runtime_error(format("tensor '%s' has invalid data", ggml_get_name(cur))); + } + } + + size_t size_done = 0; + size_t size_data = 0; + std::vector> mmaps_used; + + // Returns false if cancelled by progress_callback + bool load_all_data( + struct ggml_context * ctx, + llama_buf_map & bufs_mmap, + llama_mlocks * lmlocks, + llama_progress_callback progress_callback, + void * progress_callback_user_data) { + GGML_ASSERT(size_data != 0 && "call init_mappings() first"); + + std::vector> read_buf; + std::vector>> validation_result; + +#if defined(GGML_USE_CUDA) + // 4 staging buffers for async uploads, each sized 1MB seems to be a good default for single NVMe drives. + // NVMe raid configurations might require more / larger buffers. + constexpr size_t n_buffers = 4; + constexpr size_t buffer_size = 1 * 1024 * 1024; // 1MB + + std::vector host_buffers; + std::vector host_ptrs; + std::vector events; + size_t buffer_idx = 0; // buffer to use for async loads + + ggml_backend_t cuda_backend = nullptr; + if (!use_mmap && !check_tensors) { + // When not using mmaped io use async uploads from pinned memory to GPU memory. + // First determine if the CUDA backend is active, and if so, determine the device ID. + ggml_backend_buffer_t buf = bufs_mmap.count(0) ? bufs_mmap.at(0) : nullptr; + if (buf) { + ggml_backend_buffer_type_t buffer_type = ggml_backend_buffer_get_type(buf); + for (int i = 0; i < ggml_backend_cuda_get_device_count(); ++i) { + auto * cuda_buffer_type = ggml_backend_cuda_buffer_type(i); + if (buffer_type == cuda_buffer_type) { + cuda_backend = ggml_backend_cuda_init(i); + break; + } + } + } + + // If the cuda backend is active create pinned memory buffers and events for synchronisation. + if (cuda_backend) { + for (size_t idx = 0; idx < n_buffers; ++idx) { + host_buffers.emplace_back(ggml_backend_buft_alloc_buffer(llama_default_buffer_type_cpu(true), buffer_size)); + host_ptrs.emplace_back(ggml_backend_buffer_get_base(host_buffers[idx])); + events.emplace_back(ggml_backend_event_new(cuda_backend)); + } + } + } +#endif + + for (struct ggml_tensor * cur = ggml_get_first_tensor(ctx); cur != NULL; cur = ggml_get_next_tensor(ctx, cur)) { + const auto * weight = get_weight(ggml_get_name(cur)); + if (weight == nullptr) { + // this can happen with split experts models + continue; + } + + if (progress_callback) { + if (!progress_callback((float) size_done / size_data, progress_callback_user_data)) { + return false; + } + } + + size_t n_size = ggml_nbytes(cur); + + if (use_mmap) { + const auto & mapping = mappings.at(weight->idx); + ggml_backend_buffer_t buf_mmap = nullptr; + if (bufs_mmap.count(weight->idx)) { + buf_mmap = bufs_mmap.at(weight->idx); + } + uint8_t * data = (uint8_t *) mapping->addr + weight->offs; + + if (check_tensors) { + validation_result.emplace_back(std::async(std::launch::async, [cur, data, n_size] { + return std::make_pair(cur, ggml_validate_row_data(cur->type, data, n_size)); + })); + } + + GGML_ASSERT(buf_mmap || cur->data); // either we have a buffer to allocate the tensor in, or it is already allocated + if (buf_mmap && cur->data == nullptr) { + ggml_backend_tensor_alloc(buf_mmap, cur, data); + if (lmlocks) { + const auto & lmlock = lmlocks->at(weight->idx); + lmlock->grow_to(weight->offs + n_size); + } + + auto & mmap_used = mmaps_used[weight->idx]; + mmap_used.first = std::min(mmap_used.first, weight->offs); + mmap_used.second = std::max(mmap_used.second, weight->offs + n_size); + } else { + ggml_backend_tensor_set(cur, data, 0, n_size); + } + } else { + GGML_ASSERT(weight->idx < files.size()); + const auto & file = files.at(weight->idx); + if (ggml_backend_buffer_is_host(cur->buffer)) { + file->seek(weight->offs, SEEK_SET); + file->read_raw(cur->data, n_size); + if (check_tensors) { + validation_result.emplace_back(std::async(std::launch::async, [cur, n_size] { + return std::make_pair(cur, ggml_validate_row_data(cur->type, cur->data, n_size)); + })); + } + } else { +#if defined(GGML_USE_CUDA) + // If cuda_backend is valid load the tensor in chunks to pinned memory and upload the buffers asynchronously to the GPU. + if (cuda_backend) { + file->seek(weight->offs, SEEK_SET); + + size_t bytes_read = 0; + + while (bytes_read < n_size) { + size_t read_iteration = std::min(buffer_size, n_size - bytes_read); + + ggml_backend_event_synchronize(events[buffer_idx]); + file->read_raw(host_ptrs[buffer_idx], read_iteration); + ggml_backend_tensor_set_async(cuda_backend, cur, host_ptrs[buffer_idx], bytes_read, read_iteration); + ggml_backend_event_record(events[buffer_idx]); + + bytes_read += read_iteration; + ++buffer_idx; + buffer_idx %= n_buffers; + } + } + else +#endif + { + read_buf.resize(n_size); + file->seek(weight->offs, SEEK_SET); + file->read_raw(read_buf.data(), n_size); + ggml_backend_tensor_set(cur, read_buf.data(), 0, n_size); + if (check_tensors && !ggml_validate_row_data(cur->type, read_buf.data(), n_size)) { + throw std::runtime_error(format("tensor '%s' has invalid data", ggml_get_name(cur))); + } + } + } + } + + size_done += n_size; + } + +#if defined(GGML_USE_CUDA) + // free temporary resources used for async cuda uploads + if (cuda_backend) { + for (size_t idx = 0; idx < n_buffers;++idx) { + ggml_backend_event_synchronize(events[idx]); + ggml_backend_event_free(events[idx]); + ggml_backend_buffer_free(host_buffers[idx]); + } + ggml_backend_free(cuda_backend); + } +#endif + + // check validation results + bool validation_failed = false; + for (auto & future : validation_result) { + auto result = future.get(); + if (!result.second) { + LLAMA_LOG_ERROR("%s: tensor '%s' has invalid data\n", __func__, ggml_get_name(result.first)); + validation_failed = true; + } + } + if (validation_failed) { + throw std::runtime_error("found tensors with invalid data"); + } + + // check if this is the last call and do final cleanup + if (size_done >= size_data) { + // unmap offloaded tensors and metadata + if (use_mmap) { + for (uint32_t idx = 0; idx < mappings.size(); idx++) { + const auto & mmap_used = mmaps_used.at(idx); + auto & mapping = mappings.at(idx); + mapping->unmap_fragment(0, mmap_used.first); + if (mmap_used.second != 0) { + mapping->unmap_fragment(mmap_used.second, mapping->size); + } + } + } + if (progress_callback) { + // Even though the model is done loading, we still honor + // cancellation since we need to free allocations. + return progress_callback(1.0f, progress_callback_user_data); + } + } + + return true; + } +}; + +template<> +bool llama_model_loader::get_key(const enum llm_kv kid, enum llama_pooling_type & result, const bool required) { + uint32_t tmp; + const bool found = get_key(kid, tmp, required); + if (found) { + result = (enum llama_pooling_type) tmp; + } else { + result = LLAMA_POOLING_TYPE_UNSPECIFIED; + } + return found; +} + + +// +// load LLaMA models +// + +static const char * llama_model_arch_name(llm_arch arch) { + auto it = LLM_ARCH_NAMES.find(arch); + if (it == LLM_ARCH_NAMES.end()) { + return "unknown"; + } + return it->second; +} + +static std::string llama_model_ftype_name(llama_ftype ftype) { + if (ftype & LLAMA_FTYPE_GUESSED) { + return llama_model_ftype_name((enum llama_ftype) (ftype & ~LLAMA_FTYPE_GUESSED)) + " (guessed)"; + } + + switch (ftype) { + case LLAMA_FTYPE_ALL_F32: return "all F32"; + case LLAMA_FTYPE_MOSTLY_F16: return "F16"; + case LLAMA_FTYPE_MOSTLY_BF16: return "BF16"; + case LLAMA_FTYPE_MOSTLY_Q4_0: return "Q4_0"; + case LLAMA_FTYPE_MOSTLY_Q4_1: return "Q4_1"; + case LLAMA_FTYPE_MOSTLY_Q5_0: return "Q5_0"; + case LLAMA_FTYPE_MOSTLY_Q5_1: return "Q5_1"; + case LLAMA_FTYPE_MOSTLY_Q8_0: return "Q8_0"; + case LLAMA_FTYPE_MOSTLY_Q2_K: return "Q2_K - Medium"; + case LLAMA_FTYPE_MOSTLY_Q2_K_S: return "Q2_K - Small"; + case LLAMA_FTYPE_MOSTLY_Q3_K_S: return "Q3_K - Small"; + case LLAMA_FTYPE_MOSTLY_Q3_K_M: return "Q3_K - Medium"; + case LLAMA_FTYPE_MOSTLY_Q3_K_L: return "Q3_K - Large"; + case LLAMA_FTYPE_MOSTLY_Q4_K_S: return "Q4_K - Small"; + case LLAMA_FTYPE_MOSTLY_Q4_K_M: return "Q4_K - Medium"; + case LLAMA_FTYPE_MOSTLY_Q5_K_S: return "Q5_K - Small"; + case LLAMA_FTYPE_MOSTLY_Q5_K_M: return "Q5_K - Medium"; + case LLAMA_FTYPE_MOSTLY_Q6_K: return "Q6_K"; + case LLAMA_FTYPE_MOSTLY_IQ2_XXS: return "IQ2_XXS - 2.0625 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ2_XS: return "IQ2_XS - 2.3125 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ2_S: return "IQ2_S - 2.5 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ2_M: return "IQ2_M - 2.7 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ3_XS: return "IQ3_XS - 3.3 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ3_XXS: return "IQ3_XXS - 3.0625 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ1_S: return "IQ1_S - 1.5625 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ1_M: return "IQ1_M - 1.75 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ4_NL: return "IQ4_NL - 4.5 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ4_XS: return "IQ4_XS - 4.25 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ3_S: return "IQ3_S - 3.4375 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ3_M: return "IQ3_S mix - 3.66 bpw"; + case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: return "Q4_0_4_4"; + case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: return "Q4_0_4_8"; + case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: return "Q4_0_8_8"; + case LLAMA_FTYPE_MOSTLY_IQ1_BN: return "IQ1_BN - 1.625 bpw Bitnet"; + case LLAMA_FTYPE_MOSTLY_IQ2_BN: return "IQ2_BN - 2.00 bpw Bitnet"; + + default: return "unknown, may not work"; + } +} + +static const char * llama_model_type_name(e_model type) { + switch (type) { + case MODEL_14M: return "14M"; + case MODEL_17M: return "17M"; + case MODEL_22M: return "22M"; + case MODEL_33M: return "33M"; + case MODEL_60M: return "60M"; + case MODEL_70M: return "70M"; + case MODEL_80M: return "80M"; + case MODEL_109M: return "109M"; + case MODEL_137M: return "137M"; + case MODEL_160M: return "160M"; + case MODEL_220M: return "220M"; + case MODEL_250M: return "250M"; + case MODEL_270M: return "270M"; + case MODEL_335M: return "335M"; + case MODEL_410M: return "410M"; + case MODEL_450M: return "450M"; + case MODEL_770M: return "770M"; + case MODEL_780M: return "780M"; + case MODEL_0_5B: return "0.5B"; + case MODEL_1B: return "1B"; + case MODEL_1_3B: return "1.3B"; + case MODEL_1_4B: return "1.4B"; + case MODEL_2B: return "2B"; + case MODEL_2_8B: return "2.8B"; + case MODEL_3B: return "3B"; + case MODEL_4B: return "4B"; + case MODEL_6B: return "6B"; + case MODEL_6_9B: return "6.9B"; + case MODEL_7B: return "7B"; + case MODEL_8B: return "8B"; + case MODEL_9B: return "9B"; + case MODEL_11B: return "11B"; + case MODEL_12B: return "12B"; + case MODEL_13B: return "13B"; + case MODEL_14B: return "14B"; + case MODEL_15B: return "15B"; + case MODEL_16B: return "16B"; + case MODEL_20B: return "20B"; + case MODEL_30B: return "30B"; + case MODEL_34B: return "34B"; + case MODEL_35B: return "35B"; + case MODEL_40B: return "40B"; + case MODEL_65B: return "65B"; + case MODEL_70B: return "70B"; + case MODEL_236B: return "236B"; + case MODEL_314B: return "314B"; + case MODEL_SMALL: return "0.1B"; + case MODEL_MEDIUM: return "0.4B"; + case MODEL_LARGE: return "0.8B"; + case MODEL_XL: return "1.5B"; + case MODEL_A2_7B: return "A2.7B"; + case MODEL_8x7B: return "8x7B"; + case MODEL_8x22B: return "8x22B"; + case MODEL_16x12B: return "16x12B"; + case MODEL_10B_128x3_66B: return "10B+128x3.66B"; + case MODEL_57B_A14B: return "57B.A14B"; + case MODEL_27B: return "27B"; + default: return "?B"; + } +} + +static const char * llama_model_vocab_type_name(enum llama_vocab_type type){ + switch (type) { + case LLAMA_VOCAB_TYPE_NONE: return "no vocab"; + case LLAMA_VOCAB_TYPE_SPM: return "SPM"; + case LLAMA_VOCAB_TYPE_BPE: return "BPE"; + case LLAMA_VOCAB_TYPE_WPM: return "WPM"; + case LLAMA_VOCAB_TYPE_UGM: return "UGM"; + default: return "unknown"; + } +} + +static void llm_load_arch(llama_model_loader & ml, llama_model & model) { + model.arch = ml.get_arch(); + if (model.arch == LLM_ARCH_UNKNOWN) { + throw std::runtime_error("unknown model architecture: '" + ml.get_arch_name() + "'"); + } +} + +static void llm_load_hparams( + llama_model_loader & ml, + llama_model & model) { + auto & hparams = model.hparams; + const gguf_context * ctx = ml.meta; + + // get metadata as string + for (int i = 0; i < gguf_get_n_kv(ctx); i++) { + enum gguf_type type = gguf_get_kv_type(ctx, i); + if (type == GGUF_TYPE_ARRAY) { + continue; + } + const char * name = gguf_get_key(ctx, i); + const std::string value = gguf_kv_to_str(ctx, i); + model.gguf_kv.emplace(name, value); + } + + // get general kv + ml.get_key(LLM_KV_GENERAL_NAME, model.name, false); + + // get hparams kv + ml.get_key(LLM_KV_VOCAB_SIZE, hparams.n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, hparams.n_vocab); + + // everything past this point is not vocab-related + if (hparams.vocab_only) { + return; + } + + ml.get_key(LLM_KV_CONTEXT_LENGTH, hparams.n_ctx_train); + ml.get_key(LLM_KV_EMBEDDING_LENGTH, hparams.n_embd); + ml.get_key(LLM_KV_BLOCK_COUNT, hparams.n_layer); + ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert, false); + ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used, false); + + GGML_ASSERT(hparams.n_expert <= LLAMA_MAX_EXPERTS); + GGML_ASSERT(hparams.n_expert_used <= hparams.n_expert); + if (hparams.n_expert > 0) { + GGML_ASSERT(hparams.n_expert_used > 0); + } else { + GGML_ASSERT(hparams.n_expert_used == 0); + } + + // zero-out the per-layer hparams + std::fill(hparams.n_head_arr.begin(), hparams.n_head_arr.end(), 0); + std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0); + std::fill(hparams.n_ff_arr.begin(), hparams.n_ff_arr.end(), 0); + + ml.get_key_or_arr(LLM_KV_FEED_FORWARD_LENGTH, hparams.n_ff_arr, hparams.n_layer); + ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, hparams.n_layer); + + // n_head_kv is optional, default to n_head + hparams.n_head_kv_arr = hparams.n_head_arr; + + ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT_KV, hparams.n_head_kv_arr, hparams.n_layer, false); + + bool rope_finetuned = false; + ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false); + hparams.rope_finetuned = rope_finetuned; + + hparams.n_ctx_orig_yarn = hparams.n_ctx_train; + ml.get_key(LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, hparams.n_ctx_orig_yarn, false); + + // rope_freq_base (optional) + hparams.rope_freq_base_train = 10000.0f; + ml.get_key(LLM_KV_ROPE_FREQ_BASE, hparams.rope_freq_base_train, false); + + std::string rope_scaling("linear"); + ml.get_key(LLM_KV_ROPE_SCALING_TYPE, rope_scaling, false); + hparams.rope_scaling_type_train = llama_rope_scaling_type_from_string(rope_scaling); + GGML_ASSERT(hparams.rope_scaling_type_train != LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED); + + // rope_freq_scale (inverse of the kv) is optional + float ropescale = 0.0f; + if (!ml.get_key(LLM_KV_ROPE_SCALING_FACTOR, ropescale, false)) { + // try the old key name + ml.get_key(LLM_KV_ROPE_SCALE_LINEAR, ropescale, false); + } + hparams.rope_freq_scale_train = ropescale == 0.0f ? 1.0f : 1.0f/ropescale; + + ml.get_key(LLM_KV_ROPE_SCALING_ATTN_FACTOR, hparams.rope_attn_factor, false); + + // non-transformer models do not have attention heads + if (hparams.n_head() > 0) { + // gpt-neox n_rot = rotary_pct * (n_embd / n_head) + // gpt-j n_rot = rotary_dim + + hparams.n_embd_head_k = hparams.n_embd / hparams.n_head(); + ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k, false); + + hparams.n_embd_head_v = hparams.n_embd / hparams.n_head(); + ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v, false); + + // sanity check for n_rot (optional) + hparams.n_rot = hparams.n_embd_head_k; + + ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false); + + if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) { + if (hparams.n_rot != hparams.n_embd_head_k) { + throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k)); + } + } + } else { + hparams.n_rot = 0; + hparams.n_embd_head_k = 0; + hparams.n_embd_head_v = 0; + } + + // arch-specific KVs + switch (model.arch) { + case LLM_ARCH_LLAMA: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + if (hparams.n_expert == 8) { + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_8x7B; break; + case 56: model.type = e_model::MODEL_8x22B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } else { + switch (hparams.n_layer) { + case 22: model.type = e_model::MODEL_1B; break; + case 26: model.type = e_model::MODEL_3B; break; + // granite uses a vocab with len 49152 + case 32: model.type = hparams.n_vocab == 49152 ? e_model::MODEL_3B : (hparams.n_vocab < 40000 ? e_model::MODEL_7B : e_model::MODEL_8B); break; + case 36: model.type = e_model::MODEL_8B; break; // granite + case 40: model.type = e_model::MODEL_13B; break; + case 48: model.type = e_model::MODEL_34B; break; + case 60: model.type = e_model::MODEL_30B; break; + case 80: model.type = hparams.n_head() == hparams.n_head_kv() ? e_model::MODEL_65B : e_model::MODEL_70B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } + } break; + case LLM_ARCH_MINICPM: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 40: model.type = e_model::MODEL_2B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_GROK: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 64: model.type = e_model::MODEL_314B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_FALCON: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_7B; break; + case 60: model.type = e_model::MODEL_40B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_BAICHUAN: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_7B; break; + case 40: model.type = e_model::MODEL_13B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + + if (model.type == e_model::MODEL_13B) { + // TODO: become GGUF KV parameter + hparams.f_max_alibi_bias = 8.0f; + } + } break; + case LLM_ARCH_STARCODER: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1B; break; + case 36: model.type = e_model::MODEL_3B; break; + case 42: model.type = e_model::MODEL_7B; break; + case 40: model.type = e_model::MODEL_15B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_REFACT: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_1B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + + // TODO: become GGUF KV parameter + hparams.f_max_alibi_bias = 8.0f; + } break; + case LLM_ARCH_BERT: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_ATTENTION_CAUSAL, hparams.causal_attn); + ml.get_key(LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, hparams.n_vocab_type); + ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type, false); + + switch (hparams.n_layer) { + case 3: + model.type = e_model::MODEL_17M; break; // bge-micro + case 6: + model.type = e_model::MODEL_22M; break; // MiniLM-L6 + case 12: + switch (hparams.n_embd) { + case 384: model.type = e_model::MODEL_33M; break; // MiniLM-L12, bge-small + case 768: model.type = e_model::MODEL_109M; break; // bge-base + } break; + case 24: + model.type = e_model::MODEL_335M; break; // bge-large + } + } break; + case LLM_ARCH_JINA_BERT_V2: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_ATTENTION_CAUSAL, hparams.causal_attn); + ml.get_key(LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, hparams.n_vocab_type); + ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type); + hparams.f_max_alibi_bias = 8.0f; + + switch (hparams.n_layer) { + case 4: model.type = e_model::MODEL_33M; break; // jina-embeddings-small + case 12: model.type = e_model::MODEL_137M; break; // jina-embeddings-base + } + } break; + case LLM_ARCH_NOMIC_BERT: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_ATTENTION_CAUSAL, hparams.causal_attn); + ml.get_key(LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, hparams.n_vocab_type); + ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type); + + if (hparams.n_layer == 12 && hparams.n_embd == 768) { + model.type = e_model::MODEL_137M; + } + } break; + case LLM_ARCH_BLOOM: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + + switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1B; break; + case 30: + switch (hparams.n_embd) { + case 2560: model.type = e_model::MODEL_3B; break; + case 4096: model.type = e_model::MODEL_7B; break; + } break; + } + + // TODO: become GGUF KV parameter + hparams.f_max_alibi_bias = 8.0f; + } break; + case LLM_ARCH_MPT: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_ATTENTION_CLAMP_KQV, hparams.f_clamp_kqv, false); + ml.get_key(LLM_KV_ATTENTION_MAX_ALIBI_BIAS, hparams.f_max_alibi_bias); + + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_7B; break; + case 48: model.type = e_model::MODEL_30B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_STABLELM: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + + switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1B; break; + case 32: model.type = e_model::MODEL_3B; break; + case 40: model.type = e_model::MODEL_12B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_QWEN: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_7B; break; + case 40: model.type = e_model::MODEL_13B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_QWEN2: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + case 24: model.type = hparams.n_embd == 1024 ? e_model::MODEL_0_5B : e_model::MODEL_1B; break; + case 32: model.type = e_model::MODEL_7B; break; + case 40: model.type = hparams.n_head() == 20 ? e_model::MODEL_4B : e_model::MODEL_13B; break; + case 80: model.type = e_model::MODEL_70B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_QWEN2MOE: + { + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false); + ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false); + + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_A2_7B; break; + case 28: model.type = e_model::MODEL_57B_A14B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_PHI2: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + + switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1B; break; + case 32: model.type = e_model::MODEL_3B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_PHI3: + { + ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1B; break; + case 32: model.type = e_model::MODEL_3B; break; + case 40: model.type = e_model::MODEL_14B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_PLAMO: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 40: model.type = e_model::MODEL_13B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_GPT2: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + switch (hparams.n_layer) { + case 12: model.type = e_model::MODEL_SMALL; break; + case 24: model.type = e_model::MODEL_MEDIUM; break; + case 36: model.type = e_model::MODEL_LARGE; break; + case 48: model.type = e_model::MODEL_XL; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_CODESHELL: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + switch (hparams.n_layer) { + case 42: model.type = e_model::MODEL_7B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_ORION: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + + switch (hparams.n_layer) { + case 40: model.type = e_model::MODEL_14B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_INTERNLM2: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_7B; break; + case 48: model.type = e_model::MODEL_20B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_GEMMA: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 18: model.type = e_model::MODEL_2B; break; + case 28: model.type = e_model::MODEL_7B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_GEMMA2: + { + hparams.n_swa = 4096; // default value of gemma 2 + ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_ATTN_LOGIT_SOFTCAPPING, hparams.f_attn_logit_softcapping, false); + ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING, hparams.f_final_logit_softcapping, false); + hparams.attn_soft_cap = true; + + switch (hparams.n_layer) { + case 42: model.type = e_model::MODEL_9B; break; + case 46: model.type = e_model::MODEL_27B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_STARCODER2: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + switch (hparams.n_layer) { + case 30: model.type = e_model::MODEL_3B; break; + case 32: model.type = e_model::MODEL_7B; break; + case 40: model.type = e_model::MODEL_15B; break; + case 52: model.type = e_model::MODEL_20B; break; // granite + case 88: model.type = e_model::MODEL_34B; break; // granite + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_MAMBA: + { + ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv); + ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner); + ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state); + ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank); + + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 24: + switch (hparams.n_embd) { + case 768: model.type = e_model::MODEL_SMALL; break; + default: model.type = e_model::MODEL_UNKNOWN; + } break; + case 48: + switch (hparams.n_embd) { + case 1024: model.type = e_model::MODEL_MEDIUM; break; + case 1536: model.type = e_model::MODEL_LARGE; break; + case 2048: model.type = e_model::MODEL_XL; break; + default: model.type = e_model::MODEL_UNKNOWN; + } break; + case 64: + switch (hparams.n_embd) { + case 2560: model.type = e_model::MODEL_3B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_XVERSE: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_7B; break; + case 40: model.type = e_model::MODEL_13B; break; + case 80: model.type = e_model::MODEL_65B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_COMMAND_R: + { + ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + switch (hparams.n_layer) { + case 40: model.type = e_model::MODEL_35B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_DBRX: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_ATTENTION_CLAMP_KQV, hparams.f_clamp_kqv); + + switch (hparams.n_layer) { + case 40: model.type = e_model::MODEL_16x12B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_OLMO: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_ATTENTION_CLAMP_KQV, hparams.f_clamp_kqv, false); + + switch (hparams.n_layer) { + case 22: model.type = e_model::MODEL_1B; break; + case 32: model.type = e_model::MODEL_7B; break; + case 80: model.type = e_model::MODEL_70B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_OPENELM: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 16: model.type = e_model::MODEL_270M; break; + case 20: model.type = e_model::MODEL_450M; break; + case 28: model.type = e_model::MODEL_1B; break; + case 36: model.type = e_model::MODEL_3B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_GPTNEOX: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_USE_PARALLEL_RESIDUAL, hparams.use_par_res); + switch (hparams.n_layer) { + case 6: + switch (hparams.n_ff()) { + case 512: model.type = e_model::MODEL_14M; break; + case 2048: model.type = e_model::MODEL_70M; break; + default: model.type = e_model::MODEL_UNKNOWN; + } break; + case 12: + switch (hparams.n_ff()) { + case 3072: model.type = e_model::MODEL_160M; break; + default: model.type = e_model::MODEL_UNKNOWN; + } break; + case 16: + switch (hparams.n_ff()) { + case 8192: model.type = e_model::MODEL_1B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } break; + case 24: + switch (hparams.n_ff()) { + case 4096: model.type = e_model::MODEL_410M; break; + case 8192: model.type = e_model::MODEL_1_4B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } break; + case 32: + switch (hparams.n_ff()) { + case 10240: model.type = e_model::MODEL_2_8B; break; + case 16384: model.type = e_model::MODEL_6_9B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } break; + case 36: + switch (hparams.n_ff()) { + case 20480: model.type = e_model::MODEL_12B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } break; + case 44: + switch (hparams.n_ff()) { + case 24576: model.type = e_model::MODEL_20B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_ARCTIC: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + if (hparams.n_expert == 128) { + switch (hparams.n_layer) { + case 35: model.type = e_model::MODEL_10B_128x3_66B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } else { + model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_DEEPSEEK2: + { + bool is_lite = (hparams.n_layer == 27); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead); + if (!is_lite) { + ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q); + } + ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv); + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); + ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared); + ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale); + ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul); + + switch (hparams.n_layer) { + case 27: model.type = e_model::MODEL_16B; break; + case 60: model.type = e_model::MODEL_236B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_CHATGLM: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + case 28: model.type = e_model::MODEL_6B; break; + case 40: model.type = e_model::MODEL_9B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_BITNET: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 26: model.type = e_model::MODEL_3B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_T5: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, hparams.n_rel_attn_bkts); + + uint32_t dec_start_token_id; + if (ml.get_key(LLM_KV_DECODER_START_TOKEN_ID, dec_start_token_id, false)) { + hparams.dec_start_token_id = dec_start_token_id; + } + + switch (hparams.n_layer) { + case 6: model.type = e_model::MODEL_60M; break; // t5-small + case 8: model.type = e_model::MODEL_80M; break; // flan-t5-small + case 12: + switch (hparams.n_ff()) { + case 3072: model.type = e_model::MODEL_220M; break; // t5-base + case 2048: model.type = e_model::MODEL_250M; break; // flan-t5-base + default: model.type = e_model::MODEL_UNKNOWN; + } break; + case 24: + switch (hparams.n_ff()) { + case 4096: model.type = e_model::MODEL_770M; break; // t5-large + case 2816: model.type = e_model::MODEL_780M; break; // flan-t5-large + case 16384: model.type = e_model::MODEL_3B; break; // t5-3b + case 5120: model.type = e_model::MODEL_3B; break; // flan-t5-xl + case 65536: model.type = e_model::MODEL_11B; break; // t5-11b + case 10240: model.type = e_model::MODEL_11B; break; // flan-t5-xxl + default: model.type = e_model::MODEL_UNKNOWN; + } break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_JAIS: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_ATTENTION_MAX_ALIBI_BIAS, hparams.f_max_alibi_bias); + + switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1_3B; break; + case 40: model.type = e_model::MODEL_13B; break; + /* TODO: add variants */ + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + default: (void)0; + } + + model.ftype = ml.ftype; + + if (hparams.f_max_alibi_bias > 0.0f) { + hparams.use_alibi = true; + } + + hparams.rope_type = llama_rope_type(&model); +} + +static void llm_load_vocab( + llama_model_loader & ml, + llama_model & model) { + auto & vocab = model.vocab; + + struct gguf_context * ctx = ml.meta; + + const auto kv = LLM_KV(model.arch); + + // determine vocab type + { + std::string tokenizer_model; + std::string tokenizer_pre; + + ml.get_key(LLM_KV_TOKENIZER_MODEL, tokenizer_model); + ml.get_key(LLM_KV_TOKENIZER_PRE, tokenizer_pre, false); + + if (tokenizer_model == "no_vocab") { + vocab.type = LLAMA_VOCAB_TYPE_NONE; + + // default special tokens + vocab.special_bos_id = -1; + vocab.special_eos_id = -1; + vocab.special_unk_id = -1; + vocab.special_sep_id = -1; + vocab.special_pad_id = -1; + vocab.special_cls_id = -1; + vocab.special_mask_id = -1; + vocab.linefeed_id = -1; + + return; + } else if (tokenizer_model == "llama") { + vocab.type = LLAMA_VOCAB_TYPE_SPM; + + // default special tokens + vocab.special_bos_id = 1; + vocab.special_eos_id = 2; + vocab.special_unk_id = 0; + vocab.special_sep_id = -1; + vocab.special_pad_id = -1; + vocab.special_cls_id = -1; + vocab.special_mask_id = -1; + } else if (tokenizer_model == "bert") { + vocab.type = LLAMA_VOCAB_TYPE_WPM; + + // default special tokens + vocab.special_bos_id = -1; + vocab.special_eos_id = -1; + vocab.special_unk_id = 100; + vocab.special_sep_id = 102; + vocab.special_pad_id = 0; + vocab.special_cls_id = 101; + vocab.special_mask_id = 103; + } else if (tokenizer_model == "gpt2") { + vocab.type = LLAMA_VOCAB_TYPE_BPE; + + // read bpe merges and populate bpe ranks + const int merges_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_MERGES).c_str()); + if (merges_keyidx == -1) { + throw std::runtime_error("cannot find tokenizer merges in model file\n"); + } + + const int n_merges = gguf_get_arr_n(ctx, merges_keyidx); + for (int i = 0; i < n_merges; i++) { + const std::string word = gguf_get_arr_str(ctx, merges_keyidx, i); + GGML_ASSERT(unicode_cpts_from_utf8(word).size() > 0); + + std::string first; + std::string second; + + const size_t pos = word.find(' ', 1); + + if (pos != std::string::npos) { + first = word.substr(0, pos); + second = word.substr(pos + 1); + } + + vocab.bpe_ranks.emplace(std::make_pair(first, second), i); + } + + // default special tokens + vocab.special_bos_id = 11; + vocab.special_eos_id = 11; + vocab.special_unk_id = -1; + vocab.special_sep_id = -1; + vocab.special_pad_id = -1; + vocab.special_cls_id = -1; + vocab.special_mask_id = -1; + } else if (tokenizer_model == "t5") { + vocab.type = LLAMA_VOCAB_TYPE_UGM; + + // default special tokens + vocab.special_bos_id = -1; + vocab.special_eos_id = 1; + vocab.special_unk_id = 2; + vocab.special_sep_id = -1; + vocab.special_pad_id = 0; + vocab.special_cls_id = -1; + vocab.special_mask_id = -1; + + const int precompiled_charsmap_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP).c_str()); + if (precompiled_charsmap_keyidx != -1) { + size_t n_precompiled_charsmap = gguf_get_arr_n(ctx, precompiled_charsmap_keyidx); + const char * precompiled_charsmap = (const char *) gguf_get_arr_data(ctx, precompiled_charsmap_keyidx); + vocab.precompiled_charsmap.assign(precompiled_charsmap, precompiled_charsmap + n_precompiled_charsmap); +#ifdef IS_BIG_ENDIAN + // correct endiannes of data in precompiled_charsmap binary blob + uint32_t * xcda_blob_size = (uint32_t *) &vocab.precompiled_charsmap[0]; + *xcda_blob_size = __builtin_bswap32(*xcda_blob_size); + assert(*xcda_blob_size + sizeof(uint32_t) < n_precompiled_charsmap); + size_t xcda_array_size = *xcda_blob_size / sizeof(uint32_t); + uint32_t * xcda_array = (uint32_t *) &vocab.precompiled_charsmap[sizeof(uint32_t)]; + for (size_t i = 0; i < xcda_array_size; ++i) { + xcda_array[i] = __builtin_bswap32(xcda_array[i]); + } +#endif + } + } else { + throw std::runtime_error(format("unknown tokenizer: '%s'", tokenizer_model.c_str())); + } + + // for now, only BPE models have pre-tokenizers + if (vocab.type == LLAMA_VOCAB_TYPE_BPE) { + vocab.tokenizer_add_space_prefix = false; + vocab.tokenizer_clean_spaces = true; + if (tokenizer_pre.empty()) { + //!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! + // OK - I don't feel like recreati8ng the LLaMA-v3 models. Considering that, at least for now, + // LLaMA-v3 is the only model wehere we end up here, let's just force the pre-tokanizer to be + // llama3. + tokenizer_pre = "llama3"; + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_LLAMA3; + LLAMA_LOG_WARN("%s: missing pre-tokenizer type, using: 'llama3'\n", __func__); + LLAMA_LOG_WARN("%s: \n", __func__); + LLAMA_LOG_WARN("%s: ************************************ \n", __func__); + LLAMA_LOG_WARN("%s: GENERATION QUALITY MAY BE DEGRADED! \n", __func__); + LLAMA_LOG_WARN("%s: CONSIDER REGENERATING THE MODEL \n", __func__); + LLAMA_LOG_WARN("%s: ************************************ \n", __func__); + LLAMA_LOG_WARN("%s: \n", __func__); + //vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; + //!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! + } else if (tokenizer_pre == "default") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; + } else if ( + tokenizer_pre == "llama3" || + tokenizer_pre == "llama-v3" || + tokenizer_pre == "llama-bpe") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_LLAMA3; + vocab.tokenizer_ignore_merges = true; + vocab.tokenizer_add_bos = true; + } else if ( + tokenizer_pre == "deepseek-llm") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM; + vocab.tokenizer_clean_spaces = false; + } else if ( + tokenizer_pre == "deepseek-coder") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER; + vocab.tokenizer_clean_spaces = false; + } else if ( + tokenizer_pre == "falcon") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_FALCON; + } else if ( + tokenizer_pre == "mpt") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_MPT; + } else if ( + tokenizer_pre == "starcoder") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_STARCODER; + } else if ( + tokenizer_pre == "gpt-2" || + tokenizer_pre == "phi-2" || + tokenizer_pre == "jina-es" || + tokenizer_pre == "jina-de" || + tokenizer_pre == "jina-v2-es" || + tokenizer_pre == "jina-v2-de" || + tokenizer_pre == "jina-v2-code") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_GPT2; + } else if ( + tokenizer_pre == "refact") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_REFACT; + } else if ( + tokenizer_pre == "command-r") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_COMMAND_R; + vocab.tokenizer_clean_spaces = false; + } else if ( + tokenizer_pre == "qwen2") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_QWEN2; + vocab.tokenizer_clean_spaces = false; + } else if ( + tokenizer_pre == "stablelm2") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_STABLELM2; + } else if ( + tokenizer_pre == "olmo") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_OLMO; + } else if ( + tokenizer_pre == "dbrx") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DBRX; + } else if ( + tokenizer_pre == "smaug-bpe") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_SMAUG; + } else if ( + tokenizer_pre == "poro-chat") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_PORO; + vocab.tokenizer_clean_spaces = false; + } else if ( + tokenizer_pre == "chatglm-bpe") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CHATGLM4; + vocab.special_bos_id = -1; + } else if ( + tokenizer_pre == "viking") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_VIKING; + vocab.tokenizer_clean_spaces = false; + } else if ( + tokenizer_pre == "jais") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_JAIS; + } else if ( + tokenizer_pre == "tekken") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_TEKKEN; + vocab.tokenizer_clean_spaces = false; + vocab.tokenizer_ignore_merges = true; + vocab.tokenizer_add_bos = true; + } else if ( + tokenizer_pre == "smollm") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_SMOLLM; + vocab.tokenizer_clean_spaces = false; + } else if ( + tokenizer_pre == "codeshell") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CODESHELL; + } else { + throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str())); + } + } else if (vocab.type == LLAMA_VOCAB_TYPE_SPM) { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; + vocab.tokenizer_add_space_prefix = true; + vocab.tokenizer_clean_spaces = false; + vocab.tokenizer_add_bos = true; + vocab.tokenizer_add_eos = false; + } else if (vocab.type == LLAMA_VOCAB_TYPE_WPM) { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; + vocab.tokenizer_add_space_prefix = false; + vocab.tokenizer_clean_spaces = true; + vocab.tokenizer_add_bos = true; + vocab.tokenizer_add_eos = false; + } else if (vocab.type == LLAMA_VOCAB_TYPE_UGM) { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; + vocab.tokenizer_add_bos = false; + vocab.tokenizer_add_eos = true; + } else { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; + } + + ml.get_key(LLM_KV_TOKENIZER_ADD_PREFIX, vocab.tokenizer_add_space_prefix, false); + ml.get_key(LLM_KV_TOKENIZER_REMOVE_EXTRA_WS, vocab.tokenizer_remove_extra_whitespaces, false); + } + + const int token_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_LIST).c_str()); + if (token_idx == -1) { + throw std::runtime_error("cannot find tokenizer vocab in model file\n"); + } + + const float * scores = nullptr; + const int score_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_SCORES).c_str()); + if (score_idx != -1) { + scores = (const float * ) gguf_get_arr_data(ctx, score_idx); + } + + const int * toktypes = nullptr; + const int toktype_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_TOKEN_TYPE).c_str()); + if (toktype_idx != -1) { + toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx); + } + + const uint32_t n_vocab = gguf_get_arr_n(ctx, token_idx); + + vocab.id_to_token.resize(n_vocab); + + for (uint32_t i = 0; i < n_vocab; i++) { + std::string word = gguf_get_arr_str(ctx, token_idx, i); + GGML_ASSERT(unicode_cpts_from_utf8(word).size() > 0); + + vocab.token_to_id[word] = i; + vocab.max_token_len = std::max(vocab.max_token_len, (int) word.size()); + + auto & token_data = vocab.id_to_token[i]; + token_data.text = std::move(word); + token_data.score = scores ? scores[i] : 0.0f; + token_data.attr = LLAMA_TOKEN_ATTR_NORMAL; + + if (toktypes) { //TODO: remove, required until per token attributes are available from GGUF file + switch(toktypes[i]) { + case LLAMA_TOKEN_TYPE_UNKNOWN: token_data.attr = LLAMA_TOKEN_ATTR_UNKNOWN; break; + case LLAMA_TOKEN_TYPE_UNUSED: token_data.attr = LLAMA_TOKEN_ATTR_UNUSED; break; + case LLAMA_TOKEN_TYPE_NORMAL: token_data.attr = LLAMA_TOKEN_ATTR_NORMAL; break; + case LLAMA_TOKEN_TYPE_CONTROL: token_data.attr = LLAMA_TOKEN_ATTR_CONTROL; break; + case LLAMA_TOKEN_TYPE_USER_DEFINED: token_data.attr = LLAMA_TOKEN_ATTR_USER_DEFINED; break; + case LLAMA_TOKEN_TYPE_BYTE: token_data.attr = LLAMA_TOKEN_ATTR_BYTE; break; + case LLAMA_TOKEN_TYPE_UNDEFINED: token_data.attr = LLAMA_TOKEN_ATTR_UNDEFINED; break; + default: token_data.attr = LLAMA_TOKEN_ATTR_UNDEFINED; break; + } + } + } + GGML_ASSERT(vocab.id_to_token.size() == vocab.token_to_id.size()); + + // determine the newline token: LLaMA "<0x0A>" == 10 == '\n', Falcon 193 == '\n' + if (vocab.type == LLAMA_VOCAB_TYPE_SPM) { + // For Fill-In-the-Middle (FIM)/infill models which where converted + // prior to support of FIM special tokens in GGUF, the following + // will allow those models to continue to work. The general names + // of the known models are currently CodeLlama (LLM_ARCH_LLAMA) and + // CodeGemma (LLM_ARCH_GEMMA). This can potentially be removed once + // new versions of these models have been published. + std::string gen_name; + ml.get_key(LLM_KV_GENERAL_NAME, gen_name, false); + + std::transform(gen_name.begin(), gen_name.end(), gen_name.begin(), + [](unsigned char c){ return std::tolower(c); }); + + if (gen_name.find("code") != std::string::npos) { + if (model.arch == LLM_ARCH_LLAMA + && 32010 < vocab.id_to_token.size() + && vocab.id_to_token[32007].text.find("
") != std::string::npos
+              && vocab.id_to_token[32008].text.find("") != std::string::npos
+              && vocab.id_to_token[32009].text.find("") != std::string::npos
+              && vocab.id_to_token[32010].text.find("") != std::string::npos) {
+                vocab.special_prefix_id = 32007;
+                vocab.special_suffix_id = 32008;
+                vocab.special_middle_id = 32009;
+                vocab.special_eot_id    = 32010;
+            } else if (model.arch == LLM_ARCH_GEMMA
+              && 107 < vocab.id_to_token.size()
+              && vocab.id_to_token[67].text == "<|fim_prefix|>"
+              && vocab.id_to_token[69].text == "<|fim_suffix|>"
+              && vocab.id_to_token[68].text == "<|fim_middle|>"
+              && vocab.id_to_token[107].text == "") {
+                vocab.special_prefix_id = 67;
+                vocab.special_suffix_id = 69;
+                vocab.special_middle_id = 68;
+                // TODO: this is not EOT, it is "file separator" token, needs fix
+                //       https://huggingface.co/google/codegemma-7b-it/blob/9b1d9231388358c04d90bd003458f5070d97db44/tokenizer_config.json#L565-L572
+                //vocab.special_eot_id    = 70;
+                vocab.special_eot_id    = 107;
+            }
+        }
+        try {
+            vocab.linefeed_id = llama_byte_to_token_impl(vocab, '\n');
+        } catch (const std::exception & e) {
+            LLAMA_LOG_WARN("%s: SPM vocabulary, but newline token not found: %s! Using special_pad_id instead.", __func__, e.what());
+            vocab.linefeed_id = vocab.special_pad_id;
+        }
+    } else if (vocab.type == LLAMA_VOCAB_TYPE_WPM) {
+        vocab.linefeed_id = vocab.special_pad_id;
+    } else {
+        const std::vector ids = llama_tokenize_internal(vocab, "\xC4\x8A", false); // U+010A
+        GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
+        vocab.linefeed_id = ids[0];
+    }
+
+    // special tokens
+    {
+        const std::vector> special_token_types = {
+            { LLM_KV_TOKENIZER_BOS_ID,    vocab.special_bos_id    },
+            { LLM_KV_TOKENIZER_EOS_ID,    vocab.special_eos_id    },
+            { LLM_KV_TOKENIZER_UNK_ID,    vocab.special_unk_id    },
+            { LLM_KV_TOKENIZER_SEP_ID,    vocab.special_sep_id    },
+            { LLM_KV_TOKENIZER_PAD_ID,    vocab.special_pad_id    },
+            { LLM_KV_TOKENIZER_CLS_ID,    vocab.special_cls_id    },
+            { LLM_KV_TOKENIZER_MASK_ID,   vocab.special_mask_id   },
+            { LLM_KV_TOKENIZER_PREFIX_ID, vocab.special_prefix_id },
+            { LLM_KV_TOKENIZER_SUFFIX_ID, vocab.special_suffix_id },
+            { LLM_KV_TOKENIZER_MIDDLE_ID, vocab.special_middle_id },
+            { LLM_KV_TOKENIZER_EOT_ID,    vocab.special_eot_id    },
+        };
+
+        for (const auto & it : special_token_types) {
+            const std::string & key = kv(std::get<0>(it));
+            int32_t & id = std::get<1>(it);
+
+            uint32_t new_id;
+            if (!ml.get_key(std::get<0>(it), new_id, false)) {
+                continue;
+            }
+            if (new_id >= vocab.id_to_token.size()) {
+                LLAMA_LOG_WARN("%s: bad special token: '%s' = %ud, using default id %d\n",
+                    __func__, key.c_str(), new_id, id);
+            } else {
+                id = new_id;
+            }
+        }
+
+        // Handle add_bos_token and add_eos_token
+        {
+            bool temp = true;
+
+            if (ml.get_key(LLM_KV_TOKENIZER_ADD_BOS, temp, false)) {
+                vocab.tokenizer_add_bos = temp;
+            }
+            if (ml.get_key(LLM_KV_TOKENIZER_ADD_EOS, temp, false)) {
+                vocab.tokenizer_add_eos = temp;
+            }
+        }
+
+        // find EOT token: "<|eot_id|>", "<|im_end|>", "", etc.
+        //
+        // TODO: convert scripts should provide this token through the KV metadata LLAMA_KV_TOKENIZER_EOT_ID
+        //       for now, we apply this workaround to find the EOT token based on its text
+        if (vocab.special_eot_id == -1) {
+            for (const auto & t : vocab.token_to_id) {
+                if (
+                        // TODO: gemma "" is exported as a normal token, so the following check does not work
+                        //       need to fix convert script
+                        //vocab.id_to_token[t.second].type == LLAMA_TOKEN_TYPE_CONTROL &&
+                        (t.first == "<|eot_id|>" ||
+                         t.first == "<|im_end|>" ||
+                         t.first == "<|end|>" ||
+                         t.first == "" ||
+                         t.first == "<|endoftext|>"
+                        )
+                   ) {
+                    vocab.special_eot_id = t.second;
+                    break;
+                }
+            }
+        }
+    }
+
+    // build special tokens cache
+    {
+        for (llama_vocab::id id = 0; id < (llama_vocab::id)n_vocab; ++id) {
+            if (vocab.id_to_token[id].attr & (LLAMA_TOKEN_ATTR_CONTROL | LLAMA_TOKEN_ATTR_USER_DEFINED | LLAMA_TOKEN_ATTR_UNKNOWN)) {
+                vocab.cache_special_tokens.push_back(id);
+            }
+        }
+
+        std::sort(vocab.cache_special_tokens.begin(), vocab.cache_special_tokens.end(),
+            [&] (const llama_vocab::id a, const llama_vocab::id b) {
+                return vocab.id_to_token[a].text.size() > vocab.id_to_token[b].text.size();
+            }
+        );
+
+        LLAMA_LOG_INFO("%s: special tokens cache size = %u\n", __func__, (uint32_t)vocab.cache_special_tokens.size());
+    }
+
+    // build token to piece cache
+    {
+        size_t size_cache = 0;
+
+        std::vector cache_token_to_piece(n_vocab);
+
+        for (uint32_t id = 0; id < n_vocab; ++id) {
+            cache_token_to_piece[id] = llama_token_to_piece(&model, id, true);
+
+            size_cache += cache_token_to_piece[id].size();
+        }
+
+        std::swap(vocab.cache_token_to_piece, cache_token_to_piece);
+
+        LLAMA_LOG_INFO("%s: token to piece cache size = %.4f MB\n", __func__, size_cache / 1024.0 / 1024.0);
+    }
+
+    // Handle per token attributes
+    //NOTE: Each model customizes per token attributes.
+    //NOTE: Per token attributes are missing from the GGUF file.
+    //TODO: Extract attributes from GGUF file.
+    {
+        auto _contains_any = [] (const std::string &str, const std::vector &substrs) -> bool {
+            for (auto substr : substrs) {
+                if (str.find(substr) < std::string::npos) {
+                    return true;
+                }
+            }
+            return false;
+        };
+
+        auto _set_tokenid_attr = [&] (const llama_vocab::id id, llama_token_attr attr, bool value) {
+            uint32_t current = vocab.id_to_token.at(id).attr;
+            current = value ? (current | attr) : (current & ~attr);
+            vocab.id_to_token[id].attr = (llama_token_attr) current;
+        };
+
+        auto _set_token_attr = [&] (const std::string & token, llama_token_attr attr, bool value) {
+            _set_tokenid_attr(vocab.token_to_id.at(token), attr, value);
+        };
+
+        std::string model_name;
+        std::string tokenizer_pre;
+
+        ml.get_key(LLM_KV_GENERAL_NAME, model_name, false);
+        ml.get_key(LLM_KV_TOKENIZER_PRE, tokenizer_pre, false);
+
+        // model name to lowercase
+        std::transform(model_name.begin(), model_name.end(), model_name.begin(),
+            [] (const std::string::value_type x) {
+                return std::tolower(x);
+            }
+        );
+
+        // set attributes by model/tokenizer name
+        if (_contains_any(tokenizer_pre, {"jina-v2-de", "jina-v2-es", "jina-v2-code"})) {
+            _set_token_attr("", LLAMA_TOKEN_ATTR_LSTRIP, true);
+        } else if (_contains_any(model_name, {"phi-3", "phi3"})) {
+            for (auto id : vocab.cache_special_tokens) {
+                _set_tokenid_attr(id, LLAMA_TOKEN_ATTR_RSTRIP, true);
+            }
+            for (auto token : {""}) {
+                _set_token_attr(token, LLAMA_TOKEN_ATTR_RSTRIP, true);
+            }
+            for (auto token : {"", "", "<|endoftext|>"}) {
+                _set_token_attr(token, LLAMA_TOKEN_ATTR_RSTRIP, false);
+            }
+        }
+    }
+}
+
+static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
+    const auto & hparams = model.hparams;
+    const auto & vocab   = model.vocab;
+
+    const char * rope_scaling_type = LLAMA_ROPE_SCALING_TYPES.at(hparams.rope_scaling_type_train);
+
+    auto print_f = [](const std::function & f, uint32_t n) {
+        bool is_var = false;
+
+        std::vector v;
+        for (uint32_t i = 0; i < n; ++i) {
+            v.push_back(f(i));
+            if (v[i] != v[0]) {
+                is_var = true;
+            }
+        }
+
+        std::stringstream ss;
+
+        if (is_var) {
+            ss << "[";
+            for (uint32_t i = 0; i < n; ++i) {
+                ss << v[i];
+                if (i < n - 1) {
+                    ss << ", ";
+                }
+            }
+            ss << "]";
+        } else {
+            ss << v[0];
+        }
+
+        return ss.str();
+    };
+
+    // hparams
+    LLAMA_LOG_INFO("%s: format           = %s\n",     __func__, llama_file_version_name(ml.fver));
+    LLAMA_LOG_INFO("%s: arch             = %s\n",     __func__, LLM_ARCH_NAMES.at(model.arch));
+    LLAMA_LOG_INFO("%s: vocab type       = %s\n",     __func__, llama_model_vocab_type_name(vocab.type));
+    LLAMA_LOG_INFO("%s: n_vocab          = %u\n",     __func__, hparams.n_vocab);
+    LLAMA_LOG_INFO("%s: n_merges         = %u\n",     __func__, (int) vocab.bpe_ranks.size());
+    LLAMA_LOG_INFO("%s: vocab_only       = %d\n",     __func__, hparams.vocab_only);
+
+    if (!hparams.vocab_only) {
+        LLAMA_LOG_INFO("%s: n_ctx_train      = %u\n",     __func__, hparams.n_ctx_train);
+        LLAMA_LOG_INFO("%s: n_embd           = %u\n",     __func__, hparams.n_embd);
+        LLAMA_LOG_INFO("%s: n_layer          = %u\n",     __func__, hparams.n_layer);
+        LLAMA_LOG_INFO("%s: n_head           = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_head(il);    }, hparams.n_layer).c_str());
+        LLAMA_LOG_INFO("%s: n_head_kv        = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer).c_str());
+        LLAMA_LOG_INFO("%s: n_rot            = %u\n",     __func__, hparams.n_rot);
+        LLAMA_LOG_INFO("%s: n_swa            = %u\n",     __func__, hparams.n_swa);
+        LLAMA_LOG_INFO("%s: n_embd_head_k    = %u\n",     __func__, hparams.n_embd_head_k);
+        LLAMA_LOG_INFO("%s: n_embd_head_v    = %u\n",     __func__, hparams.n_embd_head_v);
+        LLAMA_LOG_INFO("%s: n_gqa            = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_gqa(il);        }, hparams.n_layer).c_str());
+        LLAMA_LOG_INFO("%s: n_embd_k_gqa     = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_embd_k_gqa(il); }, hparams.n_layer).c_str());
+        LLAMA_LOG_INFO("%s: n_embd_v_gqa     = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_embd_v_gqa(il); }, hparams.n_layer).c_str());
+        LLAMA_LOG_INFO("%s: f_norm_eps       = %.1e\n",   __func__, hparams.f_norm_eps);
+        LLAMA_LOG_INFO("%s: f_norm_rms_eps   = %.1e\n",   __func__, hparams.f_norm_rms_eps);
+        LLAMA_LOG_INFO("%s: f_clamp_kqv      = %.1e\n",   __func__, hparams.f_clamp_kqv);
+        LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n",   __func__, hparams.f_max_alibi_bias);
+        LLAMA_LOG_INFO("%s: f_logit_scale    = %.1e\n",   __func__, hparams.f_logit_scale);
+        LLAMA_LOG_INFO("%s: n_ff             = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_ff(il); }, hparams.n_layer).c_str());
+        LLAMA_LOG_INFO("%s: n_expert         = %u\n",     __func__, hparams.n_expert);
+        LLAMA_LOG_INFO("%s: n_expert_used    = %u\n",     __func__, hparams.n_expert_used);
+        LLAMA_LOG_INFO("%s: causal attn      = %d\n",     __func__, hparams.causal_attn);
+        LLAMA_LOG_INFO("%s: pooling type     = %d\n",     __func__, hparams.pooling_type);
+        LLAMA_LOG_INFO("%s: rope type        = %d\n",     __func__, hparams.rope_type);
+        LLAMA_LOG_INFO("%s: rope scaling     = %s\n",     __func__, rope_scaling_type);
+        LLAMA_LOG_INFO("%s: freq_base_train  = %.1f\n",   __func__, hparams.rope_freq_base_train);
+        LLAMA_LOG_INFO("%s: freq_scale_train = %g\n",     __func__, hparams.rope_freq_scale_train);
+        LLAMA_LOG_INFO("%s: n_ctx_orig_yarn  = %u\n",     __func__, hparams.n_ctx_orig_yarn);
+        LLAMA_LOG_INFO("%s: rope_finetuned   = %s\n",     __func__, hparams.rope_finetuned ? "yes" : "unknown");
+        LLAMA_LOG_INFO("%s: ssm_d_conv       = %u\n",     __func__, hparams.ssm_d_conv);
+        LLAMA_LOG_INFO("%s: ssm_d_inner      = %u\n",     __func__, hparams.ssm_d_inner);
+        LLAMA_LOG_INFO("%s: ssm_d_state      = %u\n",     __func__, hparams.ssm_d_state);
+        LLAMA_LOG_INFO("%s: ssm_dt_rank      = %u\n",     __func__, hparams.ssm_dt_rank);
+    }
+
+    LLAMA_LOG_INFO("%s: model type       = %s\n",     __func__, llama_model_type_name(model.type));
+    LLAMA_LOG_INFO("%s: model ftype      = %s\n",     __func__, llama_model_ftype_name(model.ftype).c_str());
+    if (ml.n_elements >= 1e12) {
+        LLAMA_LOG_INFO("%s: model params     = %.2f T\n", __func__, ml.n_elements*1e-12);
+    } else if (ml.n_elements >= 1e9) {
+        LLAMA_LOG_INFO("%s: model params     = %.2f B\n", __func__, ml.n_elements*1e-9);
+    } else if (ml.n_elements >= 1e6) {
+        LLAMA_LOG_INFO("%s: model params     = %.2f M\n", __func__, ml.n_elements*1e-6);
+    } else {
+        LLAMA_LOG_INFO("%s: model params     = %.2f K\n", __func__, ml.n_elements*1e-3);
+    }
+    if (ml.n_bytes < GiB) {
+        LLAMA_LOG_INFO("%s: model size       = %.2f MiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0,        ml.n_bytes*8.0/ml.n_elements);
+    } else {
+        LLAMA_LOG_INFO("%s: model size       = %.2f GiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0/1024.0, ml.n_bytes*8.0/ml.n_elements);
+    }
+
+    // general kv
+    LLAMA_LOG_INFO("%s: general.name     = %s\n",    __func__, model.name.c_str());
+
+    // special tokens
+    if (vocab.special_bos_id    != -1) { LLAMA_LOG_INFO( "%s: BOS token        = %d '%s'\n", __func__, vocab.special_bos_id,  vocab.id_to_token[vocab.special_bos_id].text.c_str() );  }
+    if (vocab.special_eos_id    != -1) { LLAMA_LOG_INFO( "%s: EOS token        = %d '%s'\n", __func__, vocab.special_eos_id,  vocab.id_to_token[vocab.special_eos_id].text.c_str() );  }
+    if (vocab.special_unk_id    != -1) { LLAMA_LOG_INFO( "%s: UNK token        = %d '%s'\n", __func__, vocab.special_unk_id,  vocab.id_to_token[vocab.special_unk_id].text.c_str() );  }
+    if (vocab.special_sep_id    != -1) { LLAMA_LOG_INFO( "%s: SEP token        = %d '%s'\n", __func__, vocab.special_sep_id,  vocab.id_to_token[vocab.special_sep_id].text.c_str() );  }
+    if (vocab.special_pad_id    != -1) { LLAMA_LOG_INFO( "%s: PAD token        = %d '%s'\n", __func__, vocab.special_pad_id,  vocab.id_to_token[vocab.special_pad_id].text.c_str() );  }
+    if (vocab.special_cls_id    != -1) { LLAMA_LOG_INFO( "%s: CLS token        = %d '%s'\n", __func__, vocab.special_cls_id,  vocab.id_to_token[vocab.special_cls_id].text.c_str() );  }
+    if (vocab.special_mask_id   != -1) { LLAMA_LOG_INFO( "%s: MASK token       = %d '%s'\n", __func__, vocab.special_mask_id, vocab.id_to_token[vocab.special_mask_id].text.c_str() ); }
+
+    if (vocab.linefeed_id       != -1) { LLAMA_LOG_INFO( "%s: LF token         = %d '%s'\n", __func__, vocab.linefeed_id,       vocab.id_to_token[vocab.linefeed_id].text.c_str() );       }
+    if (vocab.special_prefix_id != -1) { LLAMA_LOG_INFO( "%s: PRE token        = %d '%s'\n", __func__, vocab.special_prefix_id, vocab.id_to_token[vocab.special_prefix_id].text.c_str() ); }
+    if (vocab.special_suffix_id != -1) { LLAMA_LOG_INFO( "%s: SUF token        = %d '%s'\n", __func__, vocab.special_suffix_id, vocab.id_to_token[vocab.special_suffix_id].text.c_str() ); }
+    if (vocab.special_middle_id != -1) { LLAMA_LOG_INFO( "%s: MID token        = %d '%s'\n", __func__, vocab.special_middle_id, vocab.id_to_token[vocab.special_middle_id].text.c_str() ); }
+    if (vocab.special_eot_id    != -1) { LLAMA_LOG_INFO( "%s: EOT token        = %d '%s'\n", __func__, vocab.special_eot_id,    vocab.id_to_token[vocab.special_eot_id].text.c_str() );    }
+
+    LLAMA_LOG_INFO("%s: max token length = %d\n", __func__, vocab.max_token_len);
+
+    if (model.arch == LLM_ARCH_DEEPSEEK2) {
+        LLAMA_LOG_INFO("%s: n_layer_dense_lead   = %d\n",     __func__, hparams.n_layer_dense_lead);
+        LLAMA_LOG_INFO("%s: n_lora_q             = %d\n",     __func__, hparams.n_lora_q);
+        LLAMA_LOG_INFO("%s: n_lora_kv            = %d\n",     __func__, hparams.n_lora_kv);
+        LLAMA_LOG_INFO("%s: n_ff_exp             = %d\n",     __func__, hparams.n_ff_exp);
+        LLAMA_LOG_INFO("%s: n_expert_shared      = %d\n",     __func__, hparams.n_expert_shared);
+        LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n",   __func__, hparams.expert_weights_scale);
+        LLAMA_LOG_INFO("%s: rope_yarn_log_mul    = %.4f\n",   __func__, hparams.rope_yarn_log_mul);
+    }
+
+    if (model.arch == LLM_ARCH_QWEN2MOE) {
+        LLAMA_LOG_INFO("%s: n_ff_exp         = %d\n",     __func__, hparams.n_ff_exp);
+        LLAMA_LOG_INFO("%s: n_ff_shexp       = %d\n",     __func__, hparams.n_ff_shexp);
+    }
+}
+
+// Returns false if cancelled by progress_callback
+static bool llm_load_tensors(
+        llama_model_loader & ml,
+        llama_model & model,
+        int n_gpu_layers,
+        enum llama_split_mode split_mode,
+        int main_gpu,
+        const float * tensor_split,
+        bool use_mlock,
+        llama_progress_callback progress_callback,
+        void * progress_callback_user_data) {
+    model.t_start_us = ggml_time_us();
+
+    auto & hparams = model.hparams;
+
+    model.split_mode   = split_mode;
+    model.main_gpu     = main_gpu;
+    model.n_gpu_layers = n_gpu_layers;
+
+    const int n_layer     = hparams.n_layer;
+    const int i_gpu_start = std::max((int) hparams.n_layer - n_gpu_layers, (int) 0);
+    bool use_mmap_buffer = true;
+
+    // there is very little benefit to offloading the input layer, so always keep it on the CPU
+    model.buft_input = llama_default_buffer_type_cpu(true);
+
+    model.buft_layer.resize(n_layer);
+
+    // assign cpu layers
+    for (int i = 0; i < i_gpu_start; ++i) {
+        model.buft_layer[i] = llama_default_buffer_type_cpu(true);
+    }
+
+    if (split_mode == LLAMA_SPLIT_MODE_LAYER) {
+        // calculate the split points
+        int device_count = llama_get_device_count(model);
+        bool all_zero = tensor_split == nullptr || std::all_of(tensor_split, tensor_split + device_count, [](float x) { return x == 0.0f; });
+        std::vector splits(device_count);
+        if (all_zero) {
+            // default split, by free memory
+            for (int i = 0; i < device_count; ++i) {
+                splits[i] = llama_get_device_memory(model, i);
+            }
+        } else {
+            std::copy(tensor_split, tensor_split + device_count, splits.begin());
+        }
+
+        // sum and normalize the splits to get the split points
+        float split_sum = 0.0f;
+        for (int i = 0; i < device_count; ++i) {
+            split_sum += splits[i];
+            splits[i] = split_sum;
+        }
+        for (int i = 0; i < device_count; ++i) {
+            splits[i] /= split_sum;
+        }
+
+        // assign the repeating layers to the devices according to the splits
+        int act_gpu_layers = std::min(n_gpu_layers, (int)n_layer + 1);
+        for (int i = i_gpu_start; i < n_layer; ++i) {
+            int layer_gpu = std::upper_bound(splits.begin(), splits.begin() + device_count, float(i - i_gpu_start)/act_gpu_layers) - splits.begin();
+            model.buft_layer[i] = llama_default_buffer_type_offload(model, layer_gpu);
+        }
+        // assign the output layer
+        if (n_gpu_layers > n_layer) {
+            int layer_gpu = std::upper_bound(splits.begin(), splits.begin() + device_count, float(act_gpu_layers - 1)/act_gpu_layers) - splits.begin();
+            model.buft_output = llama_default_buffer_type_offload(model, layer_gpu);
+        } else {
+            model.buft_output = llama_default_buffer_type_cpu(true);
+        }
+    } else {
+        ggml_backend_buffer_type_t split_buft;
+        if (split_mode == LLAMA_SPLIT_MODE_ROW) {
+            split_buft = llama_default_buffer_type_split(model, main_gpu, tensor_split);
+        } else {
+            // LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_LAYER in backends where it is not supported
+            split_buft = llama_default_buffer_type_offload(model, main_gpu);
+        }
+        // assign the repeating layers
+        for (int i = i_gpu_start; i < n_layer; ++i) {
+            model.buft_layer[i] = {
+                split_buft,
+                llama_default_buffer_type_offload(model, main_gpu)
+            };
+        }
+        // assign the output layer
+        if (n_gpu_layers > n_layer) {
+            model.buft_output = {
+                split_buft,
+                llama_default_buffer_type_offload(model, main_gpu)
+            };
+        } else {
+            model.buft_output = llama_default_buffer_type_cpu(true);
+        }
+    }
+
+    // count used buffer types
+    std::map buft_layer_count;
+    buft_layer_count[model.buft_input.buft]++;
+    buft_layer_count[model.buft_input.buft_matrix]++;
+    buft_layer_count[model.buft_output.buft]++;
+    buft_layer_count[model.buft_output.buft_matrix]++;
+    for (int i = 0; i < n_layer; ++i) {
+        buft_layer_count[model.buft_layer[i].buft]++;
+        buft_layer_count[model.buft_layer[i].buft_matrix]++;
+    }
+
+    // create one context per buffer type
+    size_t ctx_size = ggml_tensor_overhead()*(ml.n_tensors + 1); // +1 for models where tok_embd is duplicated as output
+
+    // for moe merged tensors
+    ctx_size += ggml_tensor_overhead()*n_layer*3;
+
+    std::map ctx_map;
+    for (auto & it : buft_layer_count) {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ true,
+        };
+        ggml_context * ctx = ggml_init(params);
+        if (!ctx) {
+            throw std::runtime_error(format("failed to create context"));
+        }
+        ctx_map[it.first] = ctx;
+        model.ctxs.push_back(ctx);
+    }
+
+    LLAMA_LOG_INFO("%s: ggml ctx size = %7.2f MiB\n", __func__, model.ctxs.size()*ctx_size/1024.0/1024.0);
+
+    // create tensors for the weights
+    {
+        // note: cast to int64_t since we will use these for the tensor dimensions
+        const int64_t n_head        = hparams.n_head();
+        const int64_t n_head_kv     = hparams.n_head_kv();
+        const int64_t n_embd        = hparams.n_embd;
+        const int64_t n_embd_k_gqa  = hparams.n_embd_k_gqa();
+        const int64_t n_embd_v_gqa  = hparams.n_embd_v_gqa();
+        const int64_t n_embd_head_k = hparams.n_embd_head_k;
+        const int64_t n_embd_head_v = hparams.n_embd_head_v;
+        const int64_t n_ff          = hparams.n_ff();
+        const int64_t n_embd_gqa    = n_embd_v_gqa;
+        const int64_t n_vocab       = hparams.n_vocab;
+        const int64_t n_vocab_type  = hparams.n_vocab_type;
+        const int64_t n_expert      = hparams.n_expert;
+        const int64_t n_expert_used = hparams.n_expert_used;
+        const int64_t n_ctx_train   = hparams.n_ctx_train;
+
+        if (n_expert > 0 && hparams.n_expert_used == 0) {
+            throw std::runtime_error("model has expert layers but no expert layers are used");
+        }
+
+        ggml_context * ctx_input        = ctx_map.at(model.buft_input.buft);
+        ggml_context * ctx_output       = ctx_map.at(model.buft_output.buft);
+        ggml_context * ctx_output_split = ctx_map.at(model.buft_output.buft_matrix);
+
+        auto ctx_for_layer       = [&](int i) { return ctx_map.at(model.buft_layer[i].buft); };
+        auto ctx_for_layer_split = [&](int i) { return ctx_map.at(model.buft_layer[i].buft_matrix); };
+
+        model.layers.resize(n_layer);
+
+        const auto tn = LLM_TN(model.arch);
+        switch (model.arch) {
+            case LLM_ARCH_LLAMA:
+            case LLM_ARCH_REFACT:
+            case LLM_ARCH_MINICPM:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
+
+                        // optional bias tensors
+                        layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        if (n_expert == 0) {
+                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                            layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                            layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+
+                            // optional MLP bias
+                            layer.ffn_gate_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                            layer.ffn_down_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                            layer.ffn_up_b   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        } else {
+                            layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert});
+
+                            layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                            if (layer.ffn_gate_exps) {
+                                layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert});
+                                layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert});
+                            } else {
+                                // merge split expert into a single tensor for compatibility with older models
+                                // requires disabling mmap
+                                use_mmap_buffer = false;
+
+                                ggml_type type_gate = ml.require_tensor_meta(tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, 0).c_str())->type;
+                                ggml_type type_down = ml.require_tensor_meta(tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, 0).c_str())->type;
+                                ggml_type type_up   = ml.require_tensor_meta(tn(LLM_TENSOR_FFN_UP_EXP,   "weight", i, 0).c_str())->type;
+
+                                layer.ffn_gate_exps = ggml_new_tensor_3d(ctx_split, type_gate, n_embd,   n_ff, n_expert);
+                                layer.ffn_down_exps = ggml_new_tensor_3d(ctx_split, type_down,   n_ff, n_embd, n_expert);
+                                layer.ffn_up_exps   = ggml_new_tensor_3d(ctx_split, type_up,   n_embd,   n_ff, n_expert);
+
+                                ggml_set_name(layer.ffn_gate_exps, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i).c_str());
+                                ggml_set_name(layer.ffn_down_exps, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i).c_str());
+                                ggml_set_name(layer.ffn_up_exps,   tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i).c_str());
+
+                                for (uint32_t x = 0; x < n_expert; ++x) {
+                                    // the individual experts are loaded into a view of the merged tensor
+                                    ml.create_tensor_as_view(ctx_split, layer.ffn_gate_exps, tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, x), { n_embd, n_ff }, layer.ffn_gate_exps->nb[2]*x);
+                                    ml.create_tensor_as_view(ctx_split, layer.ffn_down_exps, tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, x), { n_ff, n_embd }, layer.ffn_down_exps->nb[2]*x);
+                                    ml.create_tensor_as_view(ctx_split, layer.ffn_up_exps,   tn(LLM_TENSOR_FFN_UP_EXP,   "weight", i, x), { n_embd, n_ff }, layer.ffn_up_exps->nb[2]*x);
+                                }
+                            }
+                        }
+                    }
+                } break;
+            case LLM_ARCH_GROK:
+                {
+                    if (n_expert == 0) {
+                        throw std::runtime_error("Grok model cannot have zero experts");
+                    }
+
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.attn_out_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_gate_inp  = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert});
+                        layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        if (layer.ffn_gate_exps) {
+                            layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert});
+                            layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert});
+                        } else {
+                            // merge split expert into a single tensor for compatibility with older models
+                            // requires disabling mmap
+                            use_mmap_buffer = false;
+
+                            ggml_type type_gate = ml.require_tensor_meta(tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, 0).c_str())->type;
+                            ggml_type type_down = ml.require_tensor_meta(tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, 0).c_str())->type;
+                            ggml_type type_up   = ml.require_tensor_meta(tn(LLM_TENSOR_FFN_UP_EXP,   "weight", i, 0).c_str())->type;
+
+                            layer.ffn_gate_exps = ggml_new_tensor_3d(ctx_split, type_gate, n_embd,   n_ff, n_expert);
+                            layer.ffn_down_exps = ggml_new_tensor_3d(ctx_split, type_down,   n_ff, n_embd, n_expert);
+                            layer.ffn_up_exps   = ggml_new_tensor_3d(ctx_split, type_up,   n_embd,   n_ff, n_expert);
+
+                            ggml_set_name(layer.ffn_gate_exps, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i).c_str());
+                            ggml_set_name(layer.ffn_down_exps, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i).c_str());
+                            ggml_set_name(layer.ffn_up_exps,   tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i).c_str());
+
+                            for (uint32_t x = 0; x < n_expert; ++x) {
+                                // the individual experts are loaded into a view of the merged tensor
+                                ml.create_tensor_as_view(ctx_split, layer.ffn_gate_exps, tn(LLM_TENSOR_FFN_GATE_EXP, "weight", i, x), { n_embd, n_ff }, layer.ffn_gate_exps->nb[2]*x);
+                                ml.create_tensor_as_view(ctx_split, layer.ffn_down_exps, tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, x), { n_ff, n_embd }, layer.ffn_down_exps->nb[2]*x);
+                                ml.create_tensor_as_view(ctx_split, layer.ffn_up_exps,   tn(LLM_TENSOR_FFN_UP_EXP,   "weight", i, x), { n_embd, n_ff }, layer.ffn_up_exps->nb[2]*x);
+                            }
+                        }
+
+                        layer.layer_out_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd});
+                    }
+                } break;
+            case LLM_ARCH_DBRX:
+            {
+                if (n_expert == 0) {
+                    throw std::runtime_error("DBRX model cannot have zero experts");
+                }
+
+                model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                // output
+                {
+                    model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                    model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                }
+
+                for (int i = 0; i < n_layer; ++i) {
+                    ggml_context * ctx_layer = ctx_for_layer(i);
+                    ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                    auto & layer = model.layers[i];
+
+                    layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                    layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
+                    layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                    layer.attn_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd});
+
+                    layer.ffn_gate_inp  = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert});
+                    layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff,   n_expert});
+                    layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff,   n_embd, n_expert});
+                    layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff,   n_expert});
+                }
+            } break;
+            case LLM_ARCH_BAICHUAN:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
+            case LLM_ARCH_FALCON:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm   = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
+
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        if (!model.output) {
+                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // needs to be on GPU
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd});
+
+                        layer.attn_norm_2   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_norm_2_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
+                        layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
+            case LLM_ARCH_STARCODER:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+                    model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD,   "weight"), {n_embd, n_ctx_train});
+
+                    // output
+                    {
+                        model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        if (!model.output) {
+                            // needs to be on GPU
+                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd});
+
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa});
+
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+                        layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd});
+
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd});
+
+                        layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd});
+
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i),   {n_embd, n_ff});
+                        layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i),     {n_ff});
+                    }
+                } break;
+            case LLM_ARCH_BERT:
+            case LLM_ARCH_NOMIC_BERT:
+                {
+                    model.tok_embd     = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab});
+                    model.type_embd    = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_vocab_type});
+
+                    if (model.arch == LLM_ARCH_BERT) {
+                        model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD,    "weight"), {n_embd, n_ctx_train});
+                    }
+
+                    model.tok_norm   = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd});
+                    model.tok_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"),   {n_embd});
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        if (model.arch == LLM_ARCH_BERT) {
+                            layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                            layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i),   {n_embd});
+
+                            layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                            layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i),   {n_embd_gqa});
+
+                            layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                            layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i),   {n_embd_gqa});
+                        } else {
+                            layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
+                        }
+
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT,      "weight", i), {n_embd, n_embd});
+
+                        layer.attn_out_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd});
+                        layer.attn_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i),   {n_embd});
+
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,        "weight", i), {n_embd, n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN,      "weight", i), {n_ff, n_embd});
+
+                        if (model.arch == LLM_ARCH_BERT) {
+                            layer.bo         = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd});
+                            layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff});
+                            layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd});
+                        } else {
+                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
+                        }
+
+                        layer.layer_out_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd});
+                        layer.layer_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i),   {n_embd});
+                    }
+                } break;
+            case LLM_ARCH_JINA_BERT_V2:
+                {
+                    model.tok_embd  = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}); // word_embeddings
+                    model.type_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_vocab_type}); // token_type_embeddings
+
+                    model.tok_norm   = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}); // LayerNorm
+                    model.tok_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"),   {n_embd}); //LayerNorm bias
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i]; // JinaBertLayer
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd});
+                        layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i),   {n_embd});
+
+                        layer.attn_q_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias",   i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa});
+                        layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias",   i), {n_embd_gqa});
+
+                        layer.attn_k_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias",   i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa});
+                        layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias",   i), {n_embd_gqa});
+
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); //output_dens
+                        layer.bo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "bias",   i), {n_embd}); //output_dens
+
+                        layer.attn_out_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}); //output_norm
+                        layer.attn_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "bias",   i), {n_embd});
+
+                        layer.attn_norm_2   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_norm_2_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "bias",   i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff});
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
+
+                        layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd});
+
+                        layer.layer_out_norm   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd});
+                        layer.layer_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "bias",   i), {n_embd});
+                    }
+                } break;
+            case LLM_ARCH_BLOOM:
+                {
+                    model.tok_embd   = ml.create_tensor(ctx_input,  tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab});
+                    model.tok_norm   = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd});
+                    model.tok_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"),   {n_embd});
+
+                    // output
+                    {
+                        model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias",   i), {n_embd});
+
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias",   i), {n_embd + 2*n_embd_gqa});
+
+                        layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+                        layer.bo   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias",   i), {n_embd});
+
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias",   i), {n_embd});
+
+                        layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd});
+
+                        layer.ffn_up     = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
+                        layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias",   i), {n_ff});
+                    }
+                } break;
+            case LLM_ARCH_MPT:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+                    model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD,   "weight"), {n_embd, n_ctx_train}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                    // output
+                    {
+                        model.output_norm   = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        if (!model.output) {
+                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // needs to be on GPU
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+                        layer.bo   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.ffn_up     = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                        layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.attn_q_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias",   i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.attn_k_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias",   i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        // AWQ ScaleActivation layer
+                        layer.ffn_act = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_ACT, "scales", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                    }
+                } break;
+            case LLM_ARCH_STABLELM:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
+                        model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm =   ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        // optional bias tensors, present in Stable LM 2 1.6B
+                        layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        // optional q and k layernorms, present in StableLM 2 12B
+                        layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k, n_head},    llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k, n_head_kv}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        // optional FFN norm, not present in StableLM 2 12B which uses parallel residual
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
+            case LLM_ARCH_QWEN:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd*3});
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd*3});
+                        layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff/2});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff/2, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff/2});
+                    }
+                } break;
+            case LLM_ARCH_QWEN2:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        // optional bias tensors
+                        layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd});
+                        layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa});
+                        layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
+            case LLM_ARCH_QWEN2MOE:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        // optional bias tensors
+                        layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd});
+                        layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa});
+                        layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert});
+
+                        GGML_ASSERT(n_expert      > 0);
+                        GGML_ASSERT(n_expert_used > 0);
+
+                        // MoE branch
+                        const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+
+                        layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert});
+                        layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert});
+                        layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert});
+
+                        // Shared expert branch
+                        const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff;
+
+                        layer.ffn_gate_inp_shexp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), {n_embd});
+                        layer.ffn_gate_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {    n_embd, n_ff_shexp});
+                        layer.ffn_down_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp,     n_embd});
+                        layer.ffn_up_shexp   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {    n_embd, n_ff_shexp});
+                    }
+                } break;
+            case LLM_ARCH_PHI2:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                        model.output_b      = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT,      "bias"),   {n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd});
+
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        if (layer.wqkv == nullptr) {
+                            layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd});
+                            layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i),   {n_embd});
+
+                            layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa});
+                            layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i),   {n_embd_gqa});
+
+                            layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa});
+                            layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i),   {n_embd_gqa});
+                        }
+
+                        layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+                        layer.bo   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd});
+
+                        layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd});
+
+                        layer.ffn_up     = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff});
+                        layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff});
+                    }
+                } break;
+            case LLM_ARCH_PHI3:
+                {
+                    const int64_t n_embd_head = n_embd / n_head;
+
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab });
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd });
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab });
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd });
+
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd });
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd });
+
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd });
+                        layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, 2 * n_ff });
+
+                        layer.rope_long  = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight"), { n_embd_head/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                        layer.rope_short = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight"), { n_embd_head/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                    }
+                } break;
+            case LLM_ARCH_PLAMO:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
+            case LLM_ARCH_GPT2:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+                    model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD,   "weight"), {n_embd, n_ctx_train});
+
+                    // output
+                    {
+                        model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM,   "weight", i), {n_embd});
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM,   "bias", i),   {n_embd});
+
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa});
+
+                        layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+                        layer.bo   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd});
+
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd});
+
+                        layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd});
+
+                        layer.ffn_up     = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff});
+                        layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff});
+                    }
+                } break;
+            case LLM_ARCH_CODESHELL:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd});
+
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa});
+
+                        layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+                        layer.bo   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd});
+
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd});
+
+                        layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd});
+
+                        layer.ffn_up     = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i),   {n_embd, n_ff});
+                        layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i),     {n_ff});
+                    }
+                } break;
+            case LLM_ARCH_ORION:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+                    {
+                        model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
+            case LLM_ARCH_INTERNLM2:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        // layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
+            case LLM_ARCH_GEMMA:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                    model.output      = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                    }
+                } break;
+            case LLM_ARCH_GEMMA2:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                    model.output      = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
+                        layer.attn_post_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_post_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd});
+                    }
+                } break;
+            case LLM_ARCH_STARCODER2:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm   = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
+
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        // optional bias tensors
+                        layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd});
+                        layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa});
+                        layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa});
+                        layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd});
+
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd});
+
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+
+                        // optional bias tensors
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd});
+                        layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP ,  "bias", i), {  n_ff});
+                    }
+                } break;
+            case LLM_ARCH_MAMBA:
+                {
+                    const int64_t d_conv  = hparams.ssm_d_conv;
+                    const int64_t d_inner = hparams.ssm_d_inner;
+                    const int64_t d_state = hparams.ssm_d_state;
+                    const int64_t dt_rank = hparams.ssm_dt_rank;
+
+                    // only an expansion factor of 2 is supported for now
+                    GGML_ASSERT(2 * n_embd == d_inner);
+
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        // if output is NULL, init from the input tok embed, duplicated to allow offloading
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        // norm
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.ssm_in = ml.create_tensor(ctx_split, tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, 2*d_inner});
+
+                        layer.ssm_conv1d = ml.create_tensor(ctx_split, tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner});
+                        layer.ssm_conv1d_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner});
+
+                        layer.ssm_x = ml.create_tensor(ctx_split, tn(LLM_TENSOR_SSM_X, "weight", i), {d_inner, dt_rank + 2*d_state});
+
+                        layer.ssm_dt = ml.create_tensor(ctx_split, tn(LLM_TENSOR_SSM_DT, "weight", i), {dt_rank, d_inner});
+                        layer.ssm_dt_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_SSM_DT, "bias", i), {d_inner});
+
+                        // no "weight" suffix for these
+                        layer.ssm_a = ml.create_tensor(ctx_split, tn(LLM_TENSOR_SSM_A, i), {d_state, d_inner});
+                        layer.ssm_d = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_SSM_D, i), {d_inner});
+
+                        // out_proj
+                        layer.ssm_out = ml.create_tensor(ctx_split, tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd});
+                    }
+                } break;
+            case LLM_ARCH_XVERSE:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
+            case LLM_ARCH_COMMAND_R:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        // init output from the input tok embed
+                        model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        if (n_layer >= 64){
+                            layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k, n_head});
+                            layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k, n_head_kv});
+                        }
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
+            case LLM_ARCH_OLMO:  // adapted from LLM_ARCH_LLAMA with norm params removed
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
+            case LLM_ARCH_OPENELM:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        // init output from the input tok embed
+                        model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        const int64_t n_head      =   hparams.n_head(i);
+                        const int64_t n_head_qkv  = 2*hparams.n_head_kv(i) + n_head;
+                        const int64_t n_ff        =   hparams.n_ff(i);
+
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_head_qkv*n_embd_head_k});
+                        layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k});
+                        layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head*n_embd_head_k, n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff});
+                    }
+                } break;
+            case LLM_ARCH_GPTNEOX:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd});
+
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa});
+
+                        layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+                        layer.bo   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd});
+
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd});
+
+                        layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd});
+
+                        layer.ffn_up     = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff});
+                        layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff});
+                    }
+                } break;
+            case LLM_ARCH_ARCTIC:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_embd});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_embd, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert});
+                        layer.ffn_norm_exps = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM_EXPS, "weight", i), {n_embd});
+                        layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, false);
+                        layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert});
+                        layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert});
+                    }
+                } break;
+            case LLM_ARCH_DEEPSEEK2:
+                {
+                    const bool is_lite = (hparams.n_layer == 27);
+
+                    const int64_t n_embd_head_qk_rope = hparams.n_rot;
+                    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+
+                    const int64_t q_lora_rank  = hparams.n_lora_q;
+                    const int64_t kv_lora_rank = hparams.n_lora_kv;
+
+                    const int64_t n_ff_exp        = hparams.n_ff_exp;
+                    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        if (!is_lite) {
+                            layer.attn_q_a_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank});
+                        }
+
+                        layer.attn_kv_a_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank});
+
+                        if (!is_lite) {
+                            layer.wq_a = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank});
+                            layer.wq_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k});
+                        } else {
+                            layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa});
+                        }
+
+                        layer.wkv_a_mqa = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + (n_embd_head_qk_rope)});
+                        layer.wkv_b     = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_KV_B,     "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)});
+                        layer.wo        = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT,      "weight", i), {              n_head * (                      n_embd_head_v), n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        if (i < (int) hparams.n_layer_dense_lead) {
+                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                            layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                            layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                        } else {
+                            layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert});
+
+                            GGML_ASSERT(n_expert      > 0);
+                            GGML_ASSERT(n_expert_used > 0);
+
+                            // MoE branch
+                            layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert});
+                            layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert});
+                            layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert});
+
+                            // Shared expert branch
+                            layer.ffn_gate_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared});
+                            layer.ffn_down_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd});
+                            layer.ffn_up_shexp   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared});
+                        }
+                    }
+                } break;
+            case LLM_ARCH_BITNET:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
+                    }
+
+                    const uint32_t n_ff = hparams.n_ff();
+                    model.layers.resize(n_layer);
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_sub_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_SUB_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd});
+                        layer.wq_scale = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "scale", i), {1});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa});
+                        layer.wk_scale = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "scale", i), {1});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv_scale = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "scale", i), {1});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+                        layer.wo_scale = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "scale", i), {1});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_sub_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_SUB_NORM, "weight", i), {n_ff});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
+                        layer.ffn_gate_scale = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "scale", i), {1});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                        layer.ffn_down_scale = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "scale", i), {1});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff});
+                        layer.ffn_up_scale = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "scale", i), {1});
+                    }
+                } break;
+            case LLM_ARCH_T5:
+                {
+                    const auto n_rel_attn_bkts = hparams.n_rel_attn_bkts;
+
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm_enc = ml.create_tensor(ctx_output, tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm     = ml.create_tensor(ctx_output, tn(LLM_TENSOR_DEC_OUTPUT_NORM, "weight"), {n_embd});
+
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm_enc  = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_ATTN_NORM,  "weight", i), {n_embd});
+                        layer.attn_rel_b_enc = ml.create_tensor(ctx_input, tn(LLM_TENSOR_ENC_ATTN_REL_B, "weight", i), {n_head, n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.wq_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_Q,   "weight", i), {n_embd, n_embd_k_gqa});
+                        layer.wk_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
+                        layer.wv_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
+                        layer.wo_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_OUT, "weight", i), {n_embd_v_gqa, n_embd});
+
+                        layer.ffn_norm_enc = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_gate_enc = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_FFN_GATE, "weight", i), {n_embd,   n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.ffn_down_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up_enc   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_FFN_UP,   "weight", i), {n_embd,   n_ff});
+
+                        layer.attn_norm  = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_DEC_ATTN_NORM,  "weight", i), {n_embd});
+                        layer.attn_rel_b = ml.create_tensor(ctx_input, tn(LLM_TENSOR_DEC_ATTN_REL_B, "weight", i), {n_head, n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_ATTN_Q,   "weight", i), {n_embd, n_embd_k_gqa});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_ATTN_OUT, "weight", i), {n_embd_v_gqa, n_embd});
+
+                        layer.attn_norm_cross  = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_DEC_CROSS_ATTN_NORM,  "weight", i), {n_embd});
+                        // this tensor seems to be unused in HF transformers implementation
+                        layer.attn_rel_b_cross = ml.create_tensor(ctx_input, tn(LLM_TENSOR_DEC_CROSS_ATTN_REL_B, "weight", i), {n_head, n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.wq_cross = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_CROSS_ATTN_Q,   "weight", i), {n_embd, n_embd_k_gqa});
+                        layer.wk_cross = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_CROSS_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
+                        layer.wv_cross = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_CROSS_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
+                        layer.wo_cross = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_CROSS_ATTN_OUT, "weight", i), {n_embd_v_gqa, n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_DEC_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_gate = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_DEC_FFN_GATE, "weight", i), {n_embd,   n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
+            case LLM_ARCH_JAIS:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // Output
+                    {
+                        model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM,   "weight", i), {n_embd});
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM,   "bias", i),   {n_embd});
+
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa});
+
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+                        layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd});
+
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd});
+
+                        layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd});
+
+                        layer.ffn_gate   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE,   "weight", i), {n_embd, n_ff});
+                        layer.ffn_gate_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE,   "bias", i),   {n_ff});
+
+                        layer.ffn_up     = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff});
+                        layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff});
+                    }
+                } break;
+            case LLM_ARCH_CHATGLM:
+                {
+                    model.tok_embd   = ml.create_tensor(ctx_input,  tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + (hparams.n_embd_head_k << 2)});
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + (hparams.n_embd_head_k << 2)});
+
+                        layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_up     = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff * 2});
+
+                        layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                    }
+                } break;
+            default:
+                throw std::runtime_error("unknown architecture");
+        }
+    }
+
+    ml.done_getting_tensors();
+
+    ml.init_mappings(true, use_mlock ? &model.mlock_mmaps : nullptr);
+    model.mappings.reserve(ml.mappings.size());
+
+    // create the backend buffers
+    std::vector> ctx_bufs;
+    ctx_bufs.reserve(ctx_map.size());
+
+    // Ensure we have enough capacity for the maximum backend buffer we will potentially create
+    size_t n_max_backend_buffer = ctx_map.size() * ml.files.size();
+    model.bufs.reserve(n_max_backend_buffer);
+
+    for (auto & it : ctx_map) {
+        ggml_backend_buffer_type_t buft = it.first;
+        ggml_context * ctx              = it.second;
+
+        llama_buf_map bufs;
+        bufs.reserve(n_max_backend_buffer);
+
+        // only the mmap region containing the tensors in the model is mapped to the backend buffer
+        // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers
+        // this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size
+        if (ml.use_mmap && use_mmap_buffer && buft == llama_default_buffer_type_cpu(true)) {
+            for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
+                void * addr = nullptr;
+                size_t first, last;
+                ml.get_mapping_range(&first, &last, &addr, idx, ctx);
+                if (first >= last) {
+                    continue;
+                }
+                ggml_backend_buffer_t buf = ggml_backend_cpu_buffer_from_ptr((char *) addr + first, last - first);
+                if (buf == nullptr) {
+                    throw std::runtime_error("unable to allocate backend CPU buffer");
+                }
+                model.bufs.push_back(buf);
+                bufs.emplace(idx, buf);
+#ifdef GGML_USE_CUDA
+                if (n_layer >= n_gpu_layers) {
+                    ggml_backend_cuda_register_host_buffer(
+                        ggml_backend_buffer_get_base(buf),
+                        ggml_backend_buffer_get_size(buf));
+                }
+#endif
+            }
+        }
+#ifdef GGML_USE_METAL
+        else if (ml.use_mmap && use_mmap_buffer && buft == ggml_backend_metal_buffer_type()) {
+            for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
+                const size_t max_size = ggml_get_max_tensor_size(ctx);
+                void * addr = nullptr;
+                size_t first, last;
+                ml.get_mapping_range(&first, &last, &addr, idx, ctx);
+                if (first >= last) {
+                    continue;
+                }
+                ggml_backend_buffer_t buf = ggml_backend_metal_buffer_from_ptr((char *) addr + first, last - first, max_size);
+                if (buf == nullptr) {
+                    throw std::runtime_error("unable to allocate backend metal buffer");
+                }
+                model.bufs.push_back(buf);
+                bufs.emplace(idx, buf);
+            }
+        }
+#endif
+        else {
+            ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+            if (buf == nullptr) {
+                throw std::runtime_error("unable to allocate backend buffer");
+            }
+            model.bufs.push_back(buf);
+            if (use_mlock && ggml_backend_buffer_is_host(buf)) {
+                model.mlock_bufs.emplace_back(new llama_mlock);
+                auto & mlock_buf = model.mlock_bufs.back();
+                mlock_buf->init   (ggml_backend_buffer_get_base(buf));
+                mlock_buf->grow_to(ggml_backend_buffer_get_size(buf));
+            }
+            for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
+                bufs.emplace(idx, buf);
+            }
+        }
+
+        if (bufs.empty()) {
+            throw std::runtime_error("failed to allocate buffer");
+        }
+
+        for (auto & buf : bufs) {
+            // indicate that this buffer contains weights
+            // this is used by ggml_backend_sched to improve op scheduling -> ops that use a weight are preferably scheduled to the backend that contains the weight
+            ggml_backend_buffer_set_usage(buf.second, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
+        }
+
+        ctx_bufs.emplace_back(ctx, bufs);
+    }
+
+    if (llama_supports_gpu_offload()) {
+        const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
+
+        LLAMA_LOG_INFO("%s: offloading %d repeating layers to GPU\n", __func__, n_gpu);
+        if (n_gpu_layers > (int) hparams.n_layer) {
+            LLAMA_LOG_INFO("%s: offloading non-repeating layers to GPU\n", __func__);
+        }
+
+        const int max_backend_supported_layers = hparams.n_layer + 1;
+        const int max_offloadable_layers       = hparams.n_layer + 1;
+
+        LLAMA_LOG_INFO("%s: offloaded %d/%d layers to GPU\n", __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers);
+    }
+
+    // print memory requirements
+    for (ggml_backend_buffer_t buf : model.bufs) {
+        LLAMA_LOG_INFO("%s: %10s buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf) / 1024.0 / 1024.0);
+    }
+
+    // populate tensors_by_name
+    for (ggml_context * ctx : model.ctxs) {
+        for (auto * cur = ggml_get_first_tensor(ctx); cur != NULL; cur = ggml_get_next_tensor(ctx, cur)) {
+            model.tensors_by_name.emplace_back(ggml_get_name(cur), cur);
+        }
+    }
+
+    // load tensor data
+    for (auto & it : ctx_bufs) {
+        ggml_context * ctx = it.first;
+        auto & bufs = it.second;
+        if (!ml.load_all_data(ctx, bufs, use_mlock ? &model.mlock_mmaps : NULL, progress_callback, progress_callback_user_data)) {
+            return false;
+        }
+    }
+
+    if (use_mmap_buffer) {
+        for (auto & mapping : ml.mappings) {
+            model.mappings.emplace_back(std::move(mapping));
+        }
+    }
+
+    if (model.arch == LLM_ARCH_BITNET) {
+        auto set_scale = [] (ggml_tensor * w, ggml_tensor * s) {
+            float scale = 1;
+            if (ggml_backend_buffer_is_host(s->buffer)) {
+                scale = *(const float *)s->data;
+            } else {
+                ggml_backend_tensor_get(s, &scale, 0, sizeof(float));
+            }
+            std::memcpy(w->op_params, &scale, sizeof(scale));
+        };
+        for (auto& l : model.layers) {
+            set_scale(l.ffn_up, l.ffn_up_scale);
+            set_scale(l.ffn_gate, l.ffn_gate_scale);
+            set_scale(l.ffn_down, l.ffn_down_scale);
+            set_scale(l.wq, l.wq_scale);
+            set_scale(l.wk, l.wk_scale);
+            set_scale(l.wv, l.wv_scale);
+            set_scale(l.wo, l.wo_scale);
+        }
+    }
+
+    // loading time will be recalculate after the first eval, so
+    // we take page faults deferred by mmap() into consideration
+    model.t_load_us = ggml_time_us() - model.t_start_us;
+    return true;
+}
+
+// Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback
+static int llama_model_load(const std::string & fname, llama_model & model, llama_model_params & params) {
+    try {
+        llama_model_loader ml(fname, params.use_mmap, params.check_tensors, params.kv_overrides);
+
+        model.hparams.vocab_only = params.vocab_only;
+
+        try {
+            llm_load_arch(ml, model);
+        } catch(const std::exception & e) {
+            throw std::runtime_error("error loading model architecture: " + std::string(e.what()));
+        }
+        try {
+            llm_load_hparams(ml, model);
+        } catch(const std::exception & e) {
+            throw std::runtime_error("error loading model hyperparameters: " + std::string(e.what()));
+        }
+        try {
+            llm_load_vocab(ml, model);
+        } catch(const std::exception & e) {
+            throw std::runtime_error("error loading model vocabulary: " + std::string(e.what()));
+        }
+
+        llm_load_print_meta(ml, model);
+
+        if (model.vocab.type != LLAMA_VOCAB_TYPE_NONE &&
+            model.hparams.n_vocab != model.vocab.id_to_token.size()) {
+            throw std::runtime_error("vocab size mismatch");
+        }
+
+        if (params.vocab_only) {
+            LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__);
+            return 0;
+        }
+
+#ifdef GGML_USE_KOMPUTE
+        if (params.n_gpu_layers > 0 && (
+            !(model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON)
+            || !(
+                model.ftype == LLAMA_FTYPE_ALL_F32 ||
+                model.ftype == LLAMA_FTYPE_MOSTLY_F16 ||
+                model.ftype == LLAMA_FTYPE_MOSTLY_BF16 ||
+                model.ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ||
+                model.ftype == LLAMA_FTYPE_MOSTLY_Q4_1
+            )
+        )) {
+            // TODO(cebtenzzre): propagate this error outside of llama_load_model_from_file
+            LLAMA_LOG_WARN("%s: disabling Kompute due to unsupported model arch or quantization\n", __func__);
+            params.n_gpu_layers = 0;
+        }
+#endif
+
+        if (!llm_load_tensors(
+            ml, model, params.n_gpu_layers, params.split_mode,  params.main_gpu, params.tensor_split, params.use_mlock,
+            params.progress_callback, params.progress_callback_user_data
+        )) {
+            return -2;
+        }
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: error loading model: %s\n", __func__, err.what());
+        return -1;
+    }
+
+    return 0;
+}
+
+//
+// llm_build
+//
+
+using llm_build_cb = std::function;
+
+enum llm_ffn_op_type {
+    LLM_FFN_SILU,
+    LLM_FFN_GELU,
+    LLM_FFN_RELU,
+    LLM_FFN_RELU_SQR,
+    LLM_FFN_SWIGLU,
+};
+
+enum llm_ffn_gate_type {
+    LLM_FFN_SEQ,
+    LLM_FFN_PAR, // ffn_gate is parallel to ffn_up
+};
+
+enum llm_norm_type {
+    LLM_NORM,
+    LLM_NORM_RMS,
+};
+
+static struct ggml_tensor * llm_build_inp_embd(
+        struct ggml_context * ctx,
+       struct llama_context & lctx,
+        const llama_hparams & hparams,
+          const llama_batch & batch,
+         struct ggml_tensor * tok_embd,
+         const llm_build_cb & cb) {
+    const int64_t n_embd = hparams.n_embd;
+
+    struct ggml_tensor * inpL;
+
+    if (batch.token) {
+        lctx.inp_tokens = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, batch.n_tokens);
+        cb(lctx.inp_tokens, "inp_tokens", -1);
+        ggml_set_input(lctx.inp_tokens);
+
+        inpL = ggml_get_rows(ctx, tok_embd, lctx.inp_tokens);
+    } else {
+       lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens);
+        inpL = lctx.inp_embd;
+        ggml_set_input(lctx.inp_embd);
+    }
+
+    cb(inpL, "inp_embd", -1);
+
+    return inpL;
+}
+
+static void llm_build_kv_store(
+        struct ggml_context * ctx,
+        const llama_hparams & hparams,
+        const llama_cparams & cparams,
+       const llama_kv_cache & kv,
+         struct ggml_cgraph * graph,
+         struct ggml_tensor * k_cur,
+         struct ggml_tensor * v_cur,
+                    int32_t   n_tokens,
+                    int32_t   kv_head,
+         const llm_build_cb & cb,
+                    int64_t   il) {
+    const int64_t n_ctx = cparams.n_ctx;
+
+    const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
+    const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
+
+    GGML_ASSERT(kv.size == n_ctx);
+
+    struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k_l[il], n_tokens*n_embd_k_gqa,
+            (ggml_row_size(kv.k_l[il]->type, n_embd_k_gqa))*kv_head);
+    cb(k_cache_view, "k_cache_view", il);
+
+    // note: storing RoPE-ed version of K in the KV cache
+    ggml_build_forward_expand(graph, ggml_cpy(ctx, k_cur, k_cache_view));
+
+    assert(v_cur->ne[0] == n_embd_v_gqa && v_cur->ne[1] == n_tokens);
+
+    struct ggml_tensor * v_cache_view = nullptr;
+
+    if (cparams.flash_attn) {
+        v_cache_view = ggml_view_1d(ctx, kv.v_l[il], n_tokens*n_embd_v_gqa,
+                (kv_head)*ggml_row_size(kv.v_l[il]->type, n_embd_v_gqa));
+    } else {
+        // note: the V cache is transposed when not using flash attention
+        v_cache_view = ggml_view_2d(ctx, kv.v_l[il], n_tokens, n_embd_v_gqa,
+                (  n_ctx)*ggml_element_size(kv.v_l[il]),
+                (kv_head)*ggml_element_size(kv.v_l[il]));
+
+        v_cur = ggml_transpose(ctx, v_cur);
+    }
+    cb(v_cache_view, "v_cache_view", il);
+
+    ggml_build_forward_expand(graph, ggml_cpy(ctx, v_cur, v_cache_view));
+}
+
+// do mat_mul, while optionally apply lora
+static struct ggml_tensor * llm_build_lora_mm(
+        struct llama_context & lctx,
+         struct ggml_context * ctx0,
+          struct ggml_tensor * w,
+          struct ggml_tensor * cur) {
+    struct ggml_tensor * res = ggml_mul_mat(ctx0, w, cur);
+    for (auto & it : lctx.lora_adapters) {
+        struct llama_lora_weight * lora = it.first->get_weight(w);
+        if (lora == nullptr) {
+            continue;
+        }
+        const float alpha = it.first->alpha;
+        const float rank  = (float) lora->b->ne[0];
+        const float scale = alpha ? it.second * alpha / rank : it.second;
+        struct ggml_tensor * ab_cur = ggml_mul_mat(
+            ctx0, lora->b,
+            ggml_mul_mat(ctx0, lora->a, cur)
+        );
+        ab_cur = ggml_scale(ctx0, ab_cur, scale);
+        res = ggml_add(ctx0, res, ab_cur);
+    }
+    return res;
+}
+
+// do mat_mul_id, while optionally apply lora
+static struct ggml_tensor * llm_build_lora_mm_id(
+        struct llama_context & lctx,
+         struct ggml_context * ctx0,
+          struct ggml_tensor * w,   // struct ggml_tensor * as
+          struct ggml_tensor * cur, // struct ggml_tensor * b
+          struct ggml_tensor * ids) {
+    struct ggml_tensor * res = ggml_mul_mat_id(ctx0, w, cur, ids);
+    for (auto & it : lctx.lora_adapters) {
+        struct llama_lora_weight * lora = it.first->get_weight(w);
+        if (lora == nullptr) {
+            continue;
+        }
+        const float alpha = it.first->alpha;
+        const float rank  = (float) lora->b->ne[0];
+        const float scale = alpha ? it.second * alpha / rank : it.second;
+        struct ggml_tensor * ab_cur = ggml_mul_mat_id(
+            ctx0, lora->b,
+            ggml_mul_mat_id(ctx0, lora->a, cur, ids),
+            ids
+        );
+        ab_cur = ggml_scale(ctx0, ab_cur, scale);
+        res = ggml_add(ctx0, res, ab_cur);
+    }
+    return res;
+}
+
+static struct ggml_tensor * llm_build_norm(
+        struct ggml_context * ctx,
+         struct ggml_tensor * cur,
+        const llama_hparams & hparams,
+         struct ggml_tensor * mw,
+         struct ggml_tensor * mb,
+              llm_norm_type   type,
+         const llm_build_cb & cb,
+                        int   il, float scale_eps = 1) {
+    switch (type) {
+        case LLM_NORM:     cur = ggml_norm    (ctx, cur, hparams.f_norm_eps);     break;
+        case LLM_NORM_RMS: cur = ggml_rms_norm(ctx, cur, scale_eps * hparams.f_norm_rms_eps); break;
+    }
+
+    if (mw || mb) {
+        cb(cur, "norm", il);
+    }
+
+    if (mw) {
+        cur = ggml_mul(ctx, cur, mw);
+        if (mb) {
+            cb(cur, "norm_w", il);
+        }
+    }
+
+    if (mb) {
+        cur = ggml_add(ctx, cur, mb);
+    }
+
+    return cur;
+}
+
+static struct ggml_tensor * llm_build_ffn(
+        struct ggml_context * ctx,
+       struct llama_context & lctx,
+         struct ggml_tensor * cur,
+         struct ggml_tensor * up,
+         struct ggml_tensor * up_b,
+         struct ggml_tensor * up_s,
+         struct ggml_tensor * gate,
+         struct ggml_tensor * gate_b,
+         struct ggml_tensor * gate_s,
+         struct ggml_tensor * down,
+         struct ggml_tensor * down_b,
+         struct ggml_tensor * down_s,
+         struct ggml_tensor * act_scales,
+            llm_ffn_op_type   type_op,
+          llm_ffn_gate_type   type_gate,
+         const llm_build_cb & cb,
+                        int   il) {
+    struct ggml_tensor * tmp = up ? llm_build_lora_mm(lctx, ctx, up, cur) : cur;
+    cb(tmp, "ffn_up", il);
+
+    if (up_b) {
+        tmp = ggml_add(ctx, tmp, up_b);
+        cb(tmp, "ffn_up_b", il);
+    }
+
+    if (up_s) {
+        tmp = ggml_mul(ctx, tmp, up_s);
+        cb(tmp, "ffn_up_s", il);
+    }
+
+    if (gate) {
+        switch (type_gate) {
+            case LLM_FFN_SEQ:
+                {
+                    cur = llm_build_lora_mm(lctx, ctx, gate, tmp);
+                    cb(cur, "ffn_gate", il);
+                } break;
+            case LLM_FFN_PAR:
+                {
+                    cur = llm_build_lora_mm(lctx, ctx, gate, cur);
+                    cb(cur, "ffn_gate", il);
+                } break;
+        }
+
+        if (gate_b) {
+            cur = ggml_add(ctx, cur, gate_b);
+            cb(cur, "ffn_gate_b", il);
+        }
+
+        if (gate_s) {
+            cur = ggml_mul(ctx, cur, gate_s);
+            cb(cur, "ffn_gate_s", il);
+        }
+
+    } else {
+        cur = tmp;
+    }
+
+    switch (type_op) {
+        case LLM_FFN_SILU:
+            {
+                cur = ggml_silu(ctx, cur);
+                cb(cur, "ffn_silu", il);
+            } break;
+        case LLM_FFN_GELU:
+            {
+                cur = ggml_gelu(ctx, cur);
+                cb(cur, "ffn_gelu", il);
+                if (act_scales != NULL) {
+                    cur = ggml_div(ctx, cur, act_scales);
+                    cb(cur, "ffn_act", il);
+                }
+            } break;
+        case LLM_FFN_RELU:
+            {
+                cur = ggml_relu(ctx, cur);
+                cb(cur, "ffn_relu", il);
+            } break;
+        case LLM_FFN_RELU_SQR:
+            {
+                cur = ggml_relu(ctx, cur);
+                cb(cur, "ffn_relu", il);
+
+                cur = ggml_sqr(ctx, cur);
+                cb(cur, "ffn_sqr(relu)", il);
+            } break;
+        case LLM_FFN_SWIGLU:
+            {
+                // Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
+                int64_t split_point = cur->ne[0] / 2;
+                struct ggml_tensor * x0 = ggml_cont(ctx, ggml_view_2d(ctx, cur, split_point, cur->ne[1], cur->nb[1], 0));
+                struct ggml_tensor * x1 = ggml_cont(ctx, ggml_view_2d(ctx, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur)));
+
+                x0 = ggml_silu(ctx, x0);
+                cb(cur, "ffn_silu", il);
+
+                cur = ggml_mul(ctx, x0, x1);
+                cb(cur, "ffn_mul", il);
+            } break;
+    }
+
+    if (type_gate == LLM_FFN_PAR) {
+        cur = ggml_mul(ctx, cur, tmp);
+        cb(cur, "ffn_gate_par", il);
+    }
+
+    if (down) {
+        cur = llm_build_lora_mm(lctx, ctx, down, cur);
+    }
+
+    if (down_b) {
+        cb(cur, "ffn_down", il);
+    }
+
+    if (down_b) {
+        cur = ggml_add(ctx, cur, down_b);
+    }
+
+    if (down_s) {
+        cur = ggml_mul(ctx, cur, down_s);
+        cb(cur, "ffn_down_s", il);
+    }
+
+    return cur;
+}
+
+static struct ggml_tensor * llm_build_moe_ffn(
+        struct ggml_context * ctx,
+       struct llama_context & lctx,
+         struct ggml_tensor * cur,
+         struct ggml_tensor * gate_inp,
+         struct ggml_tensor * up_exps,
+         struct ggml_tensor * gate_exps,
+         struct ggml_tensor * down_exps,
+                    int64_t   n_expert,
+                    int64_t   n_expert_used,
+            llm_ffn_op_type   type_op,
+                       bool   norm_w,
+                       bool   scale_w,
+                      float   w_scale,
+         const llm_build_cb & cb,
+                        int   il) {
+    int64_t n_embd = cur->ne[0];
+    int64_t n_tokens = cur->ne[1];
+
+    ggml_tensor * logits = llm_build_lora_mm(lctx, ctx, gate_inp, cur); // [n_expert, n_tokens]
+    cb(logits, "ffn_moe_logits", il);
+
+    ggml_tensor * probs = ggml_soft_max(ctx, logits); // [n_expert, n_tokens]
+    cb(probs, "ffn_moe_probs", il);
+
+    // select experts
+    ggml_tensor * selected_experts = ggml_top_k(ctx, probs, n_expert_used); // [n_expert_used, n_tokens]
+    cb(selected_experts->src[0], "ffn_moe_argsort", il);
+    cb(selected_experts, "ffn_moe_topk", il);
+
+    ggml_tensor * weights = ggml_get_rows(ctx,
+            ggml_reshape_3d(ctx, probs, 1, n_expert, n_tokens), selected_experts); // [1, n_expert_used, n_tokens]
+    cb(weights, "ffn_moe_weights", il);
+
+    if (norm_w) {
+        weights = ggml_reshape_2d(ctx, weights, n_expert_used, n_tokens);
+
+        ggml_tensor * weights_sum = ggml_sum_rows(ctx, weights); // [1, n_tokens]
+        cb(weights_sum, "ffn_moe_weights_sum", il);
+
+        weights = ggml_div(ctx, weights, weights_sum); // [n_expert_used, n_tokens]
+        cb(weights, "ffn_moe_weights_norm", il);
+
+        weights = ggml_reshape_3d(ctx, weights, 1, n_expert_used, n_tokens);
+    }
+    if (scale_w) {
+        weights = ggml_scale(ctx, weights, w_scale);
+        cb(weights, "ffn_moe_weights_scaled", il);
+    }
+
+    cur = ggml_reshape_3d(ctx, cur, n_embd, 1, n_tokens);
+    ggml_tensor * up = llm_build_lora_mm_id(lctx, ctx, up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
+    cb(up, "ffn_moe_up", il);
+
+    ggml_tensor * gate = llm_build_lora_mm_id(lctx, ctx, gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
+    cb(gate, "ffn_moe_gate", il);
+
+    switch (type_op) {
+        case LLM_FFN_SILU:
+            {
+                gate = ggml_silu(ctx, gate);
+                cb(gate, "ffn_moe_silu", il);
+            } break;
+        case LLM_FFN_GELU:
+            {
+                gate = ggml_gelu(ctx, gate);
+                cb(gate, "ffn_moe_gelu", il);
+            } break;
+        default:
+            GGML_ASSERT(false);
+    }
+
+    ggml_tensor * par = ggml_mul(ctx, up, gate); // [n_ff, n_expert_used, n_tokens]
+    cb(par, "ffn_moe_gate_par", il);
+
+    ggml_tensor * experts = llm_build_lora_mm_id(lctx, ctx, down_exps, par, selected_experts); // [n_embd, n_expert_used, n_tokens]
+    cb(experts, "ffn_moe_down", il);
+
+    experts = ggml_mul(ctx, experts, weights);
+
+    // aggregate experts
+    ggml_tensor * moe_out = nullptr;
+    for (int i = 0; i < n_expert_used; ++i) {
+        ggml_tensor * cur_expert = ggml_view_2d(ctx, experts, n_embd, n_tokens,
+                experts->nb[2], i*experts->nb[1]);
+
+        if (i == 0) {
+            moe_out = cur_expert;
+        } else {
+            moe_out = ggml_add(ctx, moe_out, cur_expert);
+        }
+    }
+
+    if (n_expert_used == 1) {
+        // avoid returning a non-contiguous tensor
+        moe_out = ggml_cont(ctx, moe_out);
+    }
+
+    return moe_out;
+}
+
+static struct ggml_tensor * llm_build_kqv(
+        struct ggml_context * ctx,
+       struct llama_context & lctx,
+       const llama_kv_cache & kv,
+         struct ggml_cgraph * graph,
+         struct ggml_tensor * wo,
+         struct ggml_tensor * wo_b,
+         struct ggml_tensor * q_cur,
+         struct ggml_tensor * kq_mask,
+                    int32_t   n_tokens,
+                    int32_t   n_kv,
+                    float     kq_scale,
+         const llm_build_cb & cb,
+                    int       il) {
+    const llama_model   & model   = lctx.model;
+    const llama_hparams & hparams = lctx.model.hparams;
+    const llama_cparams & cparams = lctx.cparams;
+
+    const int64_t n_ctx         = cparams.n_ctx;
+    const int64_t n_head        = hparams.n_head(il);
+    const int64_t n_head_kv     = hparams.n_head_kv(il);
+    const int64_t n_embd_head_k = hparams.n_embd_head_k;
+    const int64_t n_embd_k_gqa  = hparams.n_embd_k_gqa(il);
+    const int64_t n_embd_head_v = hparams.n_embd_head_v;
+    const int64_t n_embd_v_gqa  = hparams.n_embd_v_gqa(il);
+
+    struct ggml_tensor * q = ggml_permute(ctx, q_cur, 0, 2, 1, 3);
+    cb(q, "q", il);
+
+    struct ggml_tensor * k =
+        ggml_view_3d(ctx, kv.k_l[il],
+                n_embd_head_k, n_kv, n_head_kv,
+                ggml_row_size(kv.k_l[il]->type, n_embd_k_gqa),
+                ggml_row_size(kv.k_l[il]->type, n_embd_head_k),
+                0);
+    cb(k, "k", il);
+
+    struct ggml_tensor * cur;
+
+    if (cparams.flash_attn) {
+        GGML_UNUSED(model);
+        GGML_UNUSED(n_ctx);
+
+        // split cached v into n_head heads (not transposed)
+        struct ggml_tensor * v =
+            ggml_view_3d(ctx, kv.v_l[il],
+                    n_embd_head_v, n_kv, n_head_kv,
+                    ggml_row_size(kv.v_l[il]->type, n_embd_v_gqa),
+                    ggml_row_size(kv.v_l[il]->type, n_embd_head_v),
+                    0);
+        cb(v, "v", il);
+
+        cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias);
+
+        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX) {
+            ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
+        }
+
+        cur = ggml_reshape_2d(ctx, cur, n_embd_head_v*n_head, n_tokens);
+    } else {
+        struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
+        cb(kq, "kq", il);
+
+        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2) {
+            // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
+            // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
+            ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
+        }
+
+        if (model.arch == LLM_ARCH_GROK) {
+            // need to do the following:
+            // multiply by attn_output_multiplyer of 0.08838834764831845
+            // and then :
+            // kq = 30 * tanh(kq / 30)
+            // before the softmax below
+
+            //try from phi2
+            //ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
+
+            kq = ggml_tanh(ctx, ggml_scale(ctx, kq, 0.08838834764831845f/30.0f));
+            kq = ggml_scale(ctx, kq, 30);
+        }
+
+        if (hparams.attn_soft_cap) {
+            kq = ggml_scale(ctx, kq, 1.0f / hparams.f_attn_logit_softcapping);
+            kq = ggml_tanh(ctx, kq);
+            kq = ggml_scale(ctx, kq, hparams.f_attn_logit_softcapping);
+        }
+
+        kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias);
+        cb(kq, "kq_soft_max_ext", il);
+
+        GGML_ASSERT(kv.size == n_ctx);
+
+        // split cached v into n_head heads
+        struct ggml_tensor * v =
+            ggml_view_3d(ctx, kv.v_l[il],
+                    n_kv, n_embd_head_v, n_head_kv,
+                    ggml_element_size(kv.v_l[il])*n_ctx,
+                    ggml_element_size(kv.v_l[il])*n_ctx*n_embd_head_v,
+                    0);
+        cb(v, "v", il);
+
+        struct ggml_tensor * kqv = ggml_mul_mat(ctx, v, kq);
+        cb(kqv, "kqv", il);
+
+        struct ggml_tensor * kqv_merged = ggml_permute(ctx, kqv, 0, 2, 1, 3);
+        cb(kqv_merged, "kqv_merged", il);
+
+        cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_v*n_head, n_tokens);
+        cb(cur, "kqv_merged_cont", il);
+    }
+
+    ggml_build_forward_expand(graph, cur);
+
+    if (wo) {
+        cur = llm_build_lora_mm(lctx, ctx, wo, cur);
+    }
+
+    if (wo_b) {
+        cb(cur, "kqv_wo", il);
+    }
+
+    if (wo_b) {
+        cur = ggml_add(ctx, cur, wo_b);
+    }
+
+    return cur;
+}
+
+static struct ggml_tensor * llm_build_kv(
+        struct ggml_context * ctx,
+       struct llama_context & lctx,
+       const llama_kv_cache & kv,
+         struct ggml_cgraph * graph,
+         struct ggml_tensor * wo,
+         struct ggml_tensor * wo_b,
+         struct ggml_tensor * k_cur,
+         struct ggml_tensor * v_cur,
+         struct ggml_tensor * q_cur,
+         struct ggml_tensor * kq_mask,
+                    int32_t   n_tokens,
+                    int32_t   kv_head,
+                    int32_t   n_kv,
+                    float     kq_scale,
+         const llm_build_cb & cb,
+                    int       il) {
+    const llama_hparams & hparams = lctx.model.hparams;
+    const llama_cparams & cparams = lctx.cparams;
+
+    // these nodes are added to the graph together so that they are not reordered
+    // by doing so, the number of splits in the graph is reduced
+    ggml_build_forward_expand(graph, q_cur);
+    ggml_build_forward_expand(graph, k_cur);
+    ggml_build_forward_expand(graph, v_cur);
+
+    llm_build_kv_store(ctx, hparams, cparams, kv, graph, k_cur, v_cur, n_tokens, kv_head, cb, il);
+
+    struct ggml_tensor * cur;
+
+    cur  = llm_build_kqv(ctx, lctx, kv, graph, wo, wo_b,
+            q_cur, kq_mask, n_tokens, n_kv, kq_scale, cb, il);
+    cb(cur, "kqv_out", il);
+
+    return cur;
+}
+
+struct llm_build_context {
+    const llama_model    & model;
+          llama_context  & lctx;
+    const llama_hparams  & hparams;
+    const llama_cparams  & cparams;
+    const llama_batch    & batch;
+    const llama_kv_cache & kv_self;
+
+    const int64_t n_embd;
+    const int64_t n_layer;
+    const int64_t n_rot;
+    const int64_t n_ctx;       // user-specified context size (can be different from n_ctx_train)
+    const int64_t n_head;
+    const int64_t n_head_kv;
+    const int64_t n_embd_head_k;
+    const int64_t n_embd_k_gqa;
+    const int64_t n_embd_head_v;
+    const int64_t n_embd_v_gqa;
+    const int64_t n_expert;
+    const int64_t n_expert_used;
+
+    const float freq_base;
+    const float freq_scale;
+    const float ext_factor;
+    const float attn_factor;
+    const float beta_fast;
+    const float beta_slow;
+    const float norm_eps;
+    const float norm_rms_eps;
+
+    const int32_t n_tokens;
+    const int32_t n_kv;     // size of KV cache to consider (n_kv <= kv_self.size)
+    const int32_t n_outputs;
+    const int32_t n_outputs_enc;
+    const int32_t kv_head;  // index of where we store new KV data in the cache
+    const int32_t n_ctx_orig;
+
+    const bool flash_attn;
+
+    const enum llama_pooling_type pooling_type;
+    const enum llama_rope_type    rope_type;
+
+    const llm_build_cb & cb;
+
+    std::vector & buf_compute_meta;
+
+    struct ggml_context * ctx0 = nullptr;
+
+    // TODO: consider making the entire interface noexcept
+    llm_build_context(
+        llama_context  & lctx,
+    const llama_batch  & batch,
+    const llm_build_cb & cb,
+                  bool   worst_case) :
+        model            (lctx.model),
+        lctx             (lctx),
+        hparams          (model.hparams),
+        cparams          (lctx.cparams),
+        batch            (batch),
+        kv_self          (lctx.kv_self),
+        n_embd           (hparams.n_embd),
+        n_layer          (hparams.n_layer),
+        n_rot            (hparams.n_rot),
+        n_ctx            (cparams.n_ctx),
+        n_head           (hparams.n_head()),
+        n_head_kv        (hparams.n_head_kv()),
+        n_embd_head_k    (hparams.n_embd_head_k),
+        n_embd_k_gqa     (hparams.n_embd_k_gqa()),
+        n_embd_head_v    (hparams.n_embd_head_v),
+        n_embd_v_gqa     (hparams.n_embd_v_gqa()),
+        n_expert         (hparams.n_expert),
+        n_expert_used    (hparams.n_expert_used),
+        freq_base        (cparams.rope_freq_base),
+        freq_scale       (cparams.rope_freq_scale),
+        ext_factor       (cparams.yarn_ext_factor),
+        attn_factor      (cparams.yarn_attn_factor),
+        beta_fast        (cparams.yarn_beta_fast),
+        beta_slow        (cparams.yarn_beta_slow),
+        norm_eps         (hparams.f_norm_eps),
+        norm_rms_eps     (hparams.f_norm_rms_eps),
+        n_tokens         (batch.n_tokens),
+        n_kv             (worst_case ? kv_self.size : kv_self.n),
+        n_outputs        (worst_case ? n_tokens : lctx.n_outputs),
+        n_outputs_enc    (worst_case ? n_tokens : lctx.embd_enc.size() / hparams.n_embd),
+        kv_head          (worst_case ? (kv_self.recurrent ? 0 : kv_self.size - n_tokens) : kv_self.head),
+        n_ctx_orig       (cparams.n_ctx_orig_yarn),
+        flash_attn       (cparams.flash_attn),
+        pooling_type     (cparams.pooling_type),
+        rope_type        (hparams.rope_type),
+        cb               (cb),
+        buf_compute_meta (lctx.buf_compute_meta) {
+            // all initializations should be done in init()
+        }
+
+    void init() {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ buf_compute_meta.size(),
+            /*.mem_buffer =*/ buf_compute_meta.data(),
+            /*.no_alloc   =*/ true,
+        };
+
+        ctx0 = ggml_init(params);
+
+        lctx.inp_tokens      = nullptr;
+        lctx.inp_embd        = nullptr;
+        lctx.inp_pos         = nullptr;
+        lctx.inp_out_ids     = nullptr;
+        lctx.inp_KQ_mask     = nullptr;
+        lctx.inp_KQ_mask_swa = nullptr;
+        lctx.inp_K_shift     = nullptr;
+        lctx.inp_mean        = nullptr;
+        lctx.inp_cls         = nullptr;
+        lctx.inp_s_copy      = nullptr;
+        lctx.inp_s_mask      = nullptr;
+        lctx.inp_s_seq       = nullptr;
+        lctx.inp_pos_bucket    = nullptr;
+        lctx.inp_embd_enc      = nullptr;
+        lctx.inp_KQ_mask_cross = nullptr;
+    }
+
+    void free() {
+        if (ctx0) {
+            ggml_free(ctx0);
+            ctx0 = nullptr;
+        }
+    }
+
+    struct ggml_cgraph * build_k_shift() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        GGML_ASSERT(kv_self.size == n_ctx);
+
+        lctx.inp_K_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_ctx);
+        cb(lctx.inp_K_shift, "K_shift", -1);
+        ggml_set_input(lctx.inp_K_shift);
+
+        for (int il = 0; il < n_layer; ++il) {
+            const int64_t n_head_kv = hparams.n_head_kv(il);
+            const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
+            struct ggml_tensor * rope_factors = build_rope_factors(il);
+            struct ggml_tensor * tmp =
+                // we rotate only the first n_rot dimensions
+                ggml_rope_ext_inplace(ctx0,
+                        ggml_view_3d(ctx0, kv_self.k_l[il],
+                            n_embd_head_k, n_head_kv, n_ctx,
+                            ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k),
+                            ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa),
+                            0),
+                        lctx.inp_K_shift, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+
+            cb(tmp, "K_shifted", il);
+            ggml_build_forward_expand(gf, tmp);
+        }
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_s_copy() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        GGML_ASSERT(kv_self.recurrent);
+
+        struct ggml_tensor * state_copy = build_inp_s_copy();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * conv_states = ggml_reshape_2d(ctx0, kv_self.k_l[il], hparams.n_embd_k_s(), kv_self.size);
+            struct ggml_tensor * ssm_states  = ggml_reshape_2d(ctx0, kv_self.v_l[il], hparams.n_embd_v_s(), kv_self.size);
+
+            conv_states = ggml_get_rows(ctx0, conv_states, state_copy);
+            ssm_states  = ggml_get_rows(ctx0,  ssm_states, state_copy);
+
+            // TODO: name the intermediate tensors with cb()
+
+            ggml_build_forward_expand(gf, ggml_cpy(ctx0, conv_states, kv_self.k_l[il]));
+            ggml_build_forward_expand(gf, ggml_cpy(ctx0,  ssm_states, kv_self.v_l[il]));
+        }
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_defrag(const std::vector & ids) {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        for (uint32_t i = 0; i < ids.size(); ++i) {
+            const uint32_t id = ids[i];
+
+            if (i == id || id == ids.size()) {
+                continue;
+            }
+
+            uint32_t nm = 1;
+
+            while (i + nm < ids.size() && ids[i + nm] == id + nm) {
+                nm++;
+            }
+
+            for (int il = 0; il < n_layer; ++il) {
+                const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
+                const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
+
+                ggml_tensor * view_k_src = ggml_view_2d(ctx0, kv_self.k_l[il],
+                        n_embd_k_gqa, nm,
+                        ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa),
+                        ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*i));
+
+                ggml_tensor * view_k_dst = ggml_view_2d(ctx0, kv_self.k_l[il],
+                        n_embd_k_gqa, nm,
+                        ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa),
+                        ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*id));
+
+                ggml_tensor * view_v_src;
+                ggml_tensor * view_v_dst;
+
+                if (flash_attn) {
+                    // NOTE: the V cache is not transposed when using flash attention
+                    view_v_src = ggml_view_2d(ctx0, kv_self.v_l[il],
+                            n_embd_v_gqa, nm,
+                            ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa),
+                            ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*i));
+
+                    view_v_dst = ggml_view_2d(ctx0, kv_self.v_l[il],
+                            n_embd_v_gqa, nm,
+                            ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa),
+                            ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*id));
+                } else {
+                    view_v_src = ggml_view_2d(ctx0, kv_self.v_l[il],
+                            nm, n_embd_v_gqa,
+                            ggml_row_size(kv_self.v_l[il]->type, kv_self.size),
+                            ggml_row_size(kv_self.v_l[il]->type, i));
+
+                    view_v_dst = ggml_view_2d(ctx0, kv_self.v_l[il],
+                            nm, n_embd_v_gqa,
+                            ggml_row_size(kv_self.v_l[il]->type, kv_self.size),
+                            ggml_row_size(kv_self.v_l[il]->type, id));
+                }
+
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_k_src, view_k_dst));
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_v_src, view_v_dst));
+            }
+
+            i += nm - 1;
+        }
+
+        //LLAMA_LOG_INFO("gf->n_nodes = %d\n", gf->n_nodes);
+
+        return gf;
+    }
+
+    struct ggml_tensor * build_inp_pos() {
+        lctx.inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+        cb(lctx.inp_pos, "inp_pos", -1);
+        ggml_set_input(lctx.inp_pos);
+        return lctx.inp_pos;
+    }
+
+    struct ggml_tensor * build_rope_factors(int il) {
+        // choose long/short freq factors based on the context size
+        const auto n_ctx_pre_seq = cparams.n_ctx / cparams.n_seq_max;
+
+        if (n_ctx_pre_seq > hparams.n_ctx_orig_yarn) {
+            return model.layers[il].rope_long;
+        }
+
+        return model.layers[il].rope_short;
+    }
+
+    struct ggml_tensor * build_inp_out_ids() {
+        lctx.inp_out_ids = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_outputs);
+        cb(lctx.inp_out_ids, "inp_out_ids", -1);
+        ggml_set_input(lctx.inp_out_ids);
+        return lctx.inp_out_ids;
+    }
+
+    struct ggml_tensor * build_inp_KQ_mask(bool causal = true) {
+        lctx.inp_KQ_mask = causal
+            ? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv,     GGML_PAD(n_tokens, GGML_KQ_MASK_PAD))
+            : ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        cb(lctx.inp_KQ_mask, "KQ_mask", -1);
+        ggml_set_input(lctx.inp_KQ_mask);
+
+        return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_mask, GGML_TYPE_F16) : lctx.inp_KQ_mask;
+    }
+
+    struct ggml_tensor * build_inp_KQ_mask_swa(bool causal = true) {
+        GGML_ASSERT(hparams.n_swa > 0);
+
+        lctx.inp_KQ_mask_swa = causal
+            ? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv,     GGML_PAD(n_tokens, GGML_KQ_MASK_PAD))
+            : ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        cb(lctx.inp_KQ_mask_swa, "KQ_mask_swa", -1);
+        ggml_set_input(lctx.inp_KQ_mask_swa);
+
+        return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_mask_swa, GGML_TYPE_F16) : lctx.inp_KQ_mask_swa;
+    }
+
+    struct ggml_tensor * build_inp_mean() {
+        lctx.inp_mean = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens);
+        cb(lctx.inp_mean, "inp_mean", -1);
+        ggml_set_input(lctx.inp_mean);
+        return lctx.inp_mean;
+    }
+
+    struct ggml_tensor * build_inp_cls() {
+        lctx.inp_cls = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+        cb(lctx.inp_cls, "inp_cls", -1);
+        ggml_set_input(lctx.inp_cls);
+        return lctx.inp_cls;
+    }
+
+    struct ggml_tensor * build_inp_s_copy() {
+        lctx.inp_s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, kv_self.size);
+        cb(lctx.inp_s_copy, "inp_s_copy", -1);
+        ggml_set_input(lctx.inp_s_copy);
+        return lctx.inp_s_copy;
+    }
+
+    struct ggml_tensor * build_inp_s_mask() {
+        lctx.inp_s_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, n_kv);
+        cb(lctx.inp_s_mask, "inp_s_mask", -1);
+        ggml_set_input(lctx.inp_s_mask);
+        return lctx.inp_s_mask;
+    }
+
+    struct ggml_tensor * build_inp_s_seq() {
+        lctx.inp_s_seq = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_kv, n_tokens);
+        cb(lctx.inp_s_seq, "inp_s_seq", -1);
+        ggml_set_input(lctx.inp_s_seq);
+        return lctx.inp_s_seq;
+    }
+
+    struct ggml_cgraph * append_pooling(struct ggml_cgraph * gf) {
+        // find result_norm tensor for input
+        struct ggml_tensor * inp = nullptr;
+        for (int i = gf->n_nodes - 1; i >= 0; --i) {
+            inp = gf->nodes[i];
+            if (strcmp(inp->name, "result_norm") == 0 || strcmp(inp->name, "result_embd") == 0) {
+                break;
+            } else {
+                inp = nullptr;
+            }
+        }
+        GGML_ASSERT(inp != nullptr && "missing result_norm/result_embd tensor");
+
+        struct ggml_tensor * cur;
+
+        switch (pooling_type) {
+            case LLAMA_POOLING_TYPE_MEAN:
+                {
+                    struct ggml_tensor * inp_mean = build_inp_mean();
+                    cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, inp)), inp_mean);
+                } break;
+            case LLAMA_POOLING_TYPE_CLS:
+            case LLAMA_POOLING_TYPE_LAST:
+                {
+                    struct ggml_tensor * inp_cls = build_inp_cls();
+                    cur = ggml_get_rows(ctx0, inp, inp_cls);
+                } break;
+            case LLAMA_POOLING_TYPE_NONE:
+                {
+                    cur = inp;
+                } break;
+            default:
+                {
+                    GGML_ASSERT(false && "unknown pooling type");
+                } break;
+        }
+
+        cb(cur, "result_embd_pooled", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_tensor * llm_build_pos_bucket(bool causal) {
+        if (causal) {
+            lctx.inp_pos_bucket = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_kv,     n_tokens);
+        } else {
+            lctx.inp_pos_bucket = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_tokens);
+        }
+
+        ggml_set_input(lctx.inp_pos_bucket);
+        cb(lctx.inp_pos_bucket, "pos_bucket", -1);
+
+        return lctx.inp_pos_bucket;
+    }
+
+    struct ggml_tensor * llm_build_pos_bias(struct ggml_tensor * pos_bucket, struct ggml_tensor * attn_rel_b) {
+        struct ggml_tensor * pos_bucket_1d = ggml_view_1d(ctx0, pos_bucket, pos_bucket->ne[0] * pos_bucket->ne[1], 0);
+        cb(pos_bucket_1d, "pos_bucket_1d", -1);
+
+        struct ggml_tensor * pos_bias = ggml_get_rows(ctx0, attn_rel_b, pos_bucket_1d);
+        cb(pos_bias, "pos_bias", -1);
+
+        pos_bias = ggml_view_3d(ctx0, pos_bias, pos_bias->ne[0], lctx.inp_pos_bucket->ne[0], lctx.inp_pos_bucket->ne[1], ggml_element_size(pos_bias) * pos_bias->ne[0], ggml_element_size(pos_bias) * pos_bias->ne[0] * lctx.inp_pos_bucket->ne[0],  0);
+        cb(pos_bias, "pos_bias", -1);
+
+        pos_bias = ggml_permute(ctx0, pos_bias, 2, 0, 1, 3);
+        cb(pos_bias, "pos_bias", -1);
+
+        pos_bias = ggml_cont(ctx0, pos_bias);
+        cb(pos_bias, "pos_bias", -1);
+
+        return pos_bias;
+    }
+
+    struct ggml_tensor * llm_build_inp_embd_enc() {
+        const int64_t n_embd = hparams.n_embd;
+        lctx.inp_embd_enc = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_outputs_enc);
+        ggml_set_input(lctx.inp_embd_enc);
+        cb(lctx.inp_embd_enc, "embd_enc", -1);
+        return lctx.inp_embd_enc;
+    }
+
+    struct ggml_tensor * llm_build_inp_KQ_mask_cross() {
+        lctx.inp_KQ_mask_cross = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_outputs_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+        ggml_set_input(lctx.inp_KQ_mask_cross);
+        cb(lctx.inp_KQ_mask_cross, "KQ_mask_cross", -1);
+        return lctx.inp_KQ_mask_cross;
+    }
+
+    struct ggml_cgraph * build_llama() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            if (model.layers[il].ffn_gate_inp == nullptr) {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            } else {
+                // MoE branch
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_moe_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_gate_inp,
+                        model.layers[il].ffn_up_exps,
+                        model.layers[il].ffn_gate_exps,
+                        model.layers[il].ffn_down_exps,
+                        n_expert, n_expert_used,
+                        LLM_FFN_SILU, true,
+                        false, 0.0,
+                        cb, il);
+                cb(cur, "ffn_moe_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_baichuan() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = model.type == MODEL_7B ? build_inp_pos() : nullptr;
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                switch (model.type) {
+                    case MODEL_7B:
+                        Qcur = ggml_rope_ext(
+                            ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+                        Kcur = ggml_rope_ext(
+                            ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+                        break;
+                    case MODEL_13B:
+                        Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd/n_head, n_head, n_tokens);
+                        Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd/n_head, n_head, n_tokens);
+                        break;
+                    default:
+                        GGML_ASSERT(false);
+                }
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_xverse() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,      cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, NULL, LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_falcon() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * attn_norm;
+
+            attn_norm = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm,
+                    model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(attn_norm, "attn_norm", il);
+
+            // self-attention
+            {
+                if (model.layers[il].attn_norm_2) {
+                    // Falcon-40B
+                    cur = llm_build_norm(ctx0, inpL, hparams,
+                            model.layers[il].attn_norm_2,
+                            model.layers[il].attn_norm_2_b,
+                            LLM_NORM, cb, il);
+                    cb(cur, "attn_norm_2", il);
+                } else {
+                    cur = attn_norm;
+                }
+
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+                // using mode = 2 for neox mode
+                Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig,
+                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig,
+                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur       = ggml_get_rows(ctx0,       cur, inp_out_ids);
+                inpL      = ggml_get_rows(ctx0,      inpL, inp_out_ids);
+                attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = cur;
+
+            // feed forward
+            {
+                cur = llm_build_ffn(ctx0, lctx, attn_norm, // !! use the attn norm, not the result
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        NULL,                      NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = ggml_add(ctx0, cur, inpL);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        // norm
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm,
+                model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_grok() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // multiply by embedding_multiplier_scale of 78.38367176906169
+        inpL = ggml_scale(ctx0, inpL, 78.38367176906169f);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            // Grok
+            // if attn_out_norm is present then apply it before adding the input
+            if (model.layers[il].attn_out_norm) {
+                cur = llm_build_norm(ctx0, cur, hparams,
+                        model.layers[il].attn_out_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "attn_out_norm", il);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            // MoE branch
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = llm_build_moe_ffn(ctx0, lctx, cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    n_expert, n_expert_used,
+                    LLM_FFN_GELU, true,
+                    false, 0.0,
+                    cb, il);
+            cb(cur, "ffn_moe_out", il);
+
+            // Grok
+            // if layer_out_norm is present then apply it before adding the input
+            // Idea: maybe ffn_out_norm is a better name
+            if (model.layers[il].layer_out_norm) {
+                cur = llm_build_norm(ctx0, cur, hparams,
+                        model.layers[il].layer_out_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "layer_out_norm", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+
+        // Grok
+        // multiply logits by output_multiplier_scale of 0.5773502691896257
+
+        cur = ggml_scale(ctx0, cur, 0.5773502691896257f);
+
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_dbrx() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                                 model.layers[il].attn_norm, NULL,
+                                 LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                struct ggml_tensor * Qcur = nullptr;
+                struct ggml_tensor * Kcur = nullptr;
+                struct ggml_tensor * Vcur = nullptr;
+
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
+                cb(cur, "wqkv_clamped", il);
+
+                Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            // MoE branch
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                                 model.layers[il].attn_out_norm, NULL,
+                                 LLM_NORM, cb, il);
+            cb(cur, "attn_out_norm", il);
+
+            cur = llm_build_moe_ffn(ctx0, lctx, cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    cb, il);
+            cb(cur, "ffn_moe_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                             model.output_norm, NULL,
+                             LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_starcoder() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        struct ggml_tensor * pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
+        cb(pos, "pos_embd", -1);
+
+        inpL = ggml_add(ctx0, inpL, pos);
+        cb(inpL, "inpL", -1);
+
+        for (int il = 0; il < n_layer; ++il) {
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm,
+                    model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                cb(cur, "bqkv", il);
+
+                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            // add the input
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // FF
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm,
+                        model.layers[il].ffn_norm_b,
+                        LLM_NORM, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        NULL,                      NULL,                        NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = llm_build_norm(ctx0, inpL, hparams,
+                model.output_norm,
+                model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_refact() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                cb(Kcur, "Kcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                cb(Qcur, "Qcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_bert() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+        struct ggml_tensor * inp_pos = nullptr;
+
+        if (model.arch != LLM_ARCH_JINA_BERT_V2) {
+            inp_pos = build_inp_pos();
+        }
+
+        // construct input embeddings (token, type, position)
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // token types are hardcoded to zero ("Sentence A")
+        struct ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0);
+        inpL = ggml_add(ctx0, inpL, type_row0);
+        if (model.arch == LLM_ARCH_BERT) {
+            inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL);
+        }
+        cb(inpL, "inp_embd", -1);
+
+        // embed layer norm
+        inpL = llm_build_norm(ctx0, inpL, hparams, model.tok_norm, model.tok_norm_b, LLM_NORM, cb, -1);
+        cb(inpL, "inp_norm", -1);
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask(false);
+
+        // iterate layers
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * cur = inpL;
+
+            struct ggml_tensor * Qcur;
+            struct ggml_tensor * Kcur;
+            struct ggml_tensor * Vcur;
+
+            // self-attention
+            if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_JINA_BERT_V2) {
+                Qcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur), model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+
+                if (model.layers[il].attn_q_norm) {
+                    Qcur = llm_build_norm(ctx0, Qcur, hparams,
+                            model.layers[il].attn_q_norm,
+                            model.layers[il].attn_q_norm_b,
+                            LLM_NORM, cb, il);
+                }
+
+                Kcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur), model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+
+                if (model.layers[il].attn_k_norm) {
+                    Kcur = llm_build_norm(ctx0, Kcur, hparams,
+                            model.layers[il].attn_k_norm,
+                            model.layers[il].attn_k_norm_b,
+                            LLM_NORM, cb, il);
+                }
+                Vcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur), model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            } else {
+                // compute Q and K and RoPE them
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+            }
+
+            struct ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+            struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
+
+            struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+            cb(kq, "kq", il);
+
+            kq = ggml_soft_max_ext(ctx0, kq, KQ_mask, 1.0f/sqrtf(float(n_embd_head)), hparams.f_max_alibi_bias);
+            cb(kq, "kq_soft_max_ext", il);
+
+            struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens)));
+            cb(v, "v", il);
+
+            struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_tokens, n_embd_head, n_head_kv), kq);
+            cb(kqv, "kqv", il);
+
+            struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+            cb(kqv_merged, "kqv_merged", il);
+
+            cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
+            cb(cur, "kqv_merged_cont", il);
+
+            ggml_build_forward_expand(gf, cur);
+
+            cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo, cur);
+            if (model.layers[il].bo) {
+                cb(cur, "kqv_wo", il);
+            }
+
+            if (model.layers[il].bo) {
+                cur = ggml_add(ctx0, cur, model.layers[il].bo);
+            }
+            cb(cur, "kqv_out", il);
+
+            if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            // re-add the layer input
+            cur = ggml_add(ctx0, cur, inpL);
+
+            // attention layer norm
+            cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, cb, il);
+
+            if (model.layers[il].attn_norm_2 != nullptr) {
+                cur = ggml_add(ctx0, cur, inpL); // re-add the layer input
+                cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, cb, il);
+            }
+
+            struct ggml_tensor * ffn_inp = cur;
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            if (model.arch == LLM_ARCH_BERT) {
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        NULL,                      NULL,                        NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+            } else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL,                        NULL,
+                        model.layers[il].ffn_gate, NULL,                        NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
+            } else {
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            }
+            cb(cur, "ffn_out", il);
+
+            // attentions bypass the intermediate layer
+            cur = ggml_add(ctx0, cur, ffn_inp);
+
+            // output layer norm
+            cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, cb, il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        // final output
+        cur = inpL;
+        cb(cur, "result_embd", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_bloom() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        inpL = llm_build_norm(ctx0, inpL, hparams,
+                model.tok_norm,
+                model.tok_norm_b,
+                LLM_NORM, cb, -1);
+        cb(inpL, "inp_norm", -1);
+
+        for (int il = 0; il < n_layer; ++il) {
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm,
+                    model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                cb(cur, "bqkv", il);
+
+                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            // Add the input
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // FF
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm,
+                        model.layers[il].ffn_norm_b,
+                        LLM_NORM, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        NULL,                      NULL,                        NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = llm_build_norm(ctx0, inpL, hparams,
+                model.output_norm,
+                model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_mpt() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * pos;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        if (model.pos_embd) {
+            // inp_pos - contains the positions
+            struct ggml_tensor * inp_pos = build_inp_pos();
+            pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
+            cb(pos, "pos_embd", -1);
+
+            inpL = ggml_add(ctx0, inpL, pos);
+            cb(inpL, "inpL", -1);
+        }
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * attn_norm;
+
+            attn_norm = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm,
+                    model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(attn_norm, "attn_norm", il);
+
+            // self-attention
+            {
+                cur = attn_norm;
+
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                if (model.layers[il].bqkv){
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+
+                if (hparams.f_clamp_kqv > 0.0f) {
+                    cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
+                    cb(cur, "wqkv_clamped", il);
+                }
+
+                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                // Q/K Layernorm
+                if (model.layers[il].attn_q_norm) {
+                    Qcur = llm_build_norm(ctx0, Qcur, hparams,
+                            model.layers[il].attn_q_norm,
+                            model.layers[il].attn_q_norm_b,
+                            LLM_NORM, cb, il);
+                    cb(Qcur, "Qcur", il);
+
+                    Kcur = llm_build_norm(ctx0, Kcur, hparams,
+                            model.layers[il].attn_k_norm,
+                            model.layers[il].attn_k_norm_b,
+                            LLM_NORM, cb, il);
+                    cb(Kcur, "Kcur", il);
+
+                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+                    cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                            model.layers[il].wo, model.layers[il].bo,
+                            Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                } else {
+                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+
+                    cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                            model.layers[il].wo, model.layers[il].bo,
+                            Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                }
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            // Add the input
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed forward
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm,
+                        model.layers[il].ffn_norm_b,
+                        LLM_NORM, cb, il);
+                cb(cur, "ffn_norm", il);
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        NULL,                      NULL,                        NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        model.layers[il].ffn_act,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm,
+                model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_stablelm() {
+        struct ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm,
+                    model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+
+            struct ggml_tensor * inpSA = cur;
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                cb(Qcur, "Qcur", il);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                cb(Kcur, "Kcur", il);
+
+                if (model.layers[il].attn_q_norm) {
+                    Qcur = llm_build_norm(ctx0, Qcur, hparams,
+                            model.layers[il].attn_q_norm,
+                            NULL,
+                            LLM_NORM, cb, il);
+                    cb(Qcur, "Qcur", il);
+                }
+                if (model.layers[il].attn_k_norm) {
+                    Kcur = llm_build_norm(ctx0, Kcur, hparams,
+                            model.layers[il].attn_k_norm,
+                            NULL,
+                            LLM_NORM, cb, il);
+                    cb(Kcur, "Kcur", il);
+                }
+
+
+                Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpL  = ggml_get_rows(ctx0,  inpL, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            {
+                if (model.layers[il].ffn_norm) {
+                    cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                            model.layers[il].ffn_norm,
+                            model.layers[il].ffn_norm_b,
+                            LLM_NORM, cb, il);
+                    cb(cur, "ffn_norm", il);
+                } else {
+                    // parallel residual
+                    cur = inpSA;
+                }
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm,
+                model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_qwen() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                cb(cur, "bqkv", il);
+
+                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 2*sizeof(float)*(n_embd)));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+                // using mode = 2 for neox mode
+                Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig,
+                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig,
+                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward forward
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_qwen2() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = llm_build_ffn(ctx0, lctx, cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_qwen2moe() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self_attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // MoE branch
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            ggml_tensor * moe_out =
+                    llm_build_moe_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_gate_inp,
+                        model.layers[il].ffn_up_exps,
+                        model.layers[il].ffn_gate_exps,
+                        model.layers[il].ffn_down_exps,
+                        n_expert, n_expert_used,
+                        LLM_FFN_SILU, false,
+                        false, 0.0,
+                        cb, il);
+            cb(cur, "ffn_moe_out", il);
+
+            // FFN shared expert
+            {
+                ggml_tensor * cur_gate_inp = llm_build_lora_mm(lctx, ctx0, model.layers[il].ffn_gate_inp_shexp, cur);
+                cb(cur_gate_inp, "ffn_shexp_gate_inp", il);
+
+                // sigmoid
+                ggml_tensor * cur_gate = ggml_div(ctx0, ggml_silu(ctx0, cur_gate_inp), cur_gate_inp);
+                cb(cur_gate, "ffn_shexp_gate", il);
+
+                ggml_tensor * cur_ffn = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up_shexp,   NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur_ffn, "ffn_shexp", il);
+
+                ggml_tensor * ffn_shexp_out = ggml_mul(ctx0, cur_ffn, cur_gate);
+                cb(ffn_shexp_out, "ffn_shexp_out", il);
+
+                moe_out = ggml_add(ctx0, moe_out, ffn_shexp_out);
+                cb(moe_out, "ffn_out", il);
+
+                cur = moe_out;
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_phi2() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * attn_norm_output;
+        struct ggml_tensor * ffn_output;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            attn_norm_output = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm,
+                    model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(attn_norm_output, "attn_norm", il);
+
+            // self-attention
+            {
+                struct ggml_tensor * Qcur = nullptr;
+                struct ggml_tensor * Kcur = nullptr;
+                struct ggml_tensor * Vcur = nullptr;
+
+                if (model.layers[il].wqkv) {
+                    cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, attn_norm_output);
+                    cb(cur, "wqkv", il);
+
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+
+                    Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                    Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                } else {
+                    Qcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, attn_norm_output), model.layers[il].bq);
+                    Kcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, attn_norm_output), model.layers[il].bk);
+                    Vcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, attn_norm_output), model.layers[il].bv);
+                }
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig,
+                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                // with phi2, we scale the Q to avoid precision issues
+                // ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L66
+                Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head)));
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig,
+                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur              = ggml_get_rows(ctx0,              cur, inp_out_ids);
+                inpL             = ggml_get_rows(ctx0,             inpL, inp_out_ids);
+                attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids);
+            }
+
+            // FF
+            {
+                ffn_output = llm_build_ffn(ctx0, lctx, attn_norm_output,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        NULL,                      NULL,                        NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+                cb(ffn_output, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_output);
+            cur = ggml_add(ctx0, cur, inpL);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = llm_build_norm(ctx0, inpL, hparams,
+                model.output_norm,
+                model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output_no_bias", -1);
+
+        cur = ggml_add(ctx0, cur, model.output_b);
+        cb(cur, "result_output", -1);
+        ggml_build_forward_expand(gf, cur);
+        return gf;
+    }
+
+    struct ggml_cgraph * build_phi3() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa();
+
+        for (int il = 0; il < n_layer; ++il) {
+            auto residual = inpL;
+
+            // self-attention
+            {
+                // rope freq factors for 128k context
+                struct ggml_tensor * rope_factors = build_rope_factors(il);
+
+                struct ggml_tensor* attn_norm_output = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm,
+                    NULL,
+                    LLM_NORM_RMS, cb, il);
+                cb(attn_norm_output, "attn_norm", il);
+
+                struct ggml_tensor * Qcur = nullptr;
+                struct ggml_tensor * Kcur = nullptr;
+                struct ggml_tensor * Vcur = nullptr;
+
+                if (model.layers[il].wqkv) {
+                    cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, attn_norm_output);
+                    cb(cur, "wqkv", il);
+
+                    Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0 * sizeof(float) * (n_embd)));
+                    Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd)));
+                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa)));
+                }
+                else {
+                    Qcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, attn_norm_output), model.layers[il].bq);
+                    Kcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, attn_norm_output), model.layers[il].bk);
+                    Vcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, attn_norm_output), model.layers[il].bv);
+                }
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig,
+                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head)));
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig,
+                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask_swa, n_tokens, kv_head, n_kv, 1.0f, cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor* inp_out_ids = build_inp_out_ids();
+                cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+                residual = ggml_get_rows(ctx0, residual, inp_out_ids);
+            }
+
+            cur = ggml_add(ctx0, cur, residual);
+            residual = cur;
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            // FF
+            // special-case: the up and gate tensors are merged into a single tensor
+            // TOOD: support into llm_build_ffn
+            {
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        NULL,                      NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SWIGLU, LLM_FFN_SEQ, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, residual, cur);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = llm_build_norm(ctx0, inpL, hparams,
+            model.output_norm,
+            NULL,
+            LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+
+    struct ggml_cgraph * build_plamo() {
+        struct ggml_cgraph * gf = ggml_new_graph(ctx0);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            struct ggml_tensor * attention_norm = cur;
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_rot, n_head,    n_tokens), inp_pos, nullptr,
+                        n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens), inp_pos, nullptr,
+                        n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+            struct ggml_tensor * sa_out = cur;
+
+            cur = attention_norm;
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur    = ggml_get_rows(ctx0,    cur, inp_out_ids);
+                sa_out = ggml_get_rows(ctx0, sa_out, inp_out_ids);
+                inpL   = ggml_get_rows(ctx0,   inpL, inp_out_ids);
+            }
+
+            // feed-forward network
+            {
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, sa_out);
+            cur = ggml_add(ctx0, cur, inpL);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_gpt2() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * pos;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
+        cb(pos, "pos_embd", -1);
+
+        inpL = ggml_add(ctx0, inpL, pos);
+        cb(inpL, "inpL", -1);
+
+        for (int il = 0; il < n_layer; ++il) {
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm,
+                    model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                cb(cur, "bqkv", il);
+
+                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            // add the input
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // FF
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm,
+                        model.layers[il].ffn_norm_b,
+                        LLM_NORM, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        NULL,                      NULL,                        NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = llm_build_norm(ctx0, inpL, hparams,
+                model.output_norm,
+                model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_codeshell() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm,
+                    model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                cb(cur, "bqkv", il);
+
+                struct ggml_tensor * tmpq = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                struct ggml_tensor * tmpk = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+                cb(tmpq, "tmpq", il);
+                cb(tmpk, "tmpk", il);
+                cb(Vcur, "Vcur", il);
+
+                struct ggml_tensor * Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head,    n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            // add the input
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // FF
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm,
+                        model.layers[il].ffn_norm_b,
+                        LLM_NORM, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        NULL,                      NULL,                        NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = llm_build_norm(ctx0, inpL, hparams,
+                model.output_norm,
+                model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_orion() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                // if (model.layers[il].bq) {
+                //     Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                //     cb(Qcur, "Qcur", il);
+                // }
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                // if (model.layers[il].bk) {
+                //     Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                //     cb(Kcur, "Kcur", il);
+                // }
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                // if (model.layers[il].bv) {
+                //     Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                //     cb(Vcur, "Vcur", il);
+                // }
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, model.layers[il].ffn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = llm_build_ffn(ctx0, lctx, cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_internlm2() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = llm_build_ffn(ctx0, lctx, cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    // ref: https://arxiv.org/abs/2203.03466
+    //      https://github.com/ggerganov/llama.cpp/issues/5276#issuecomment-1925774738
+    // based on the original build_llama() function
+    struct ggml_cgraph * build_minicpm() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        const int64_t n_embd = hparams.n_embd;
+        //TODO: if the model varies, these parameters need to be read from the model
+        const int64_t n_embd_base = 256;
+        const float scale_embd  = 12.0f;
+        const float scale_depth = 1.4f;
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // scale the input embeddings
+        inpL = ggml_scale(ctx0, inpL, scale_embd);
+        cb(inpL, "inp_scaled", -1);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            // scale_res - scale the hidden states for residual connection
+            const float scale_res = scale_depth/sqrtf(float(n_layer));
+            cur = ggml_scale(ctx0, cur, scale_res);
+            cb(cur, "hidden_scaled", -1);
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            // scale the hidden states for residual connection
+            cur = ggml_scale(ctx0, cur, scale_res);
+            cb(cur, "hidden_scaled_ffn", -1);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head scaling
+        const float scale_lmhead = float(n_embd_base)/float(n_embd);
+        cur = ggml_scale(ctx0, cur, scale_lmhead);
+        cb(cur, "lmhead_scaling", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_gemma() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head_k = hparams.n_embd_head_k;
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+        cb(inpL, "inp_scaled", -1);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head,    n_tokens), inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Qcur, "Qcur", il);
+
+                Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
+                cb(Qcur, "Qcur_scaled", il);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+            cb(sa_out, "sa_out", il);
+
+            cur = llm_build_norm(ctx0, sa_out, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            // feed-forward network
+            {
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, sa_out);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_gemma2() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head_k = hparams.n_embd_head_k;
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
+        cb(inpL, "inp_scaled", -1);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        // gemma 2 requires different mask for layers using sliding window (SWA)
+        struct ggml_tensor * KQ_mask     = build_inp_KQ_mask(true);
+        struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa(true);
+
+        for (int il = 0; il < n_layer; ++il) {
+            // (il % 2) layers use SWA
+            struct ggml_tensor * KQ_mask_l = (il % 2 == 0) ? KQ_mask_swa : KQ_mask;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head,    n_tokens), inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Qcur, "Qcur", il);
+
+                // ref: https://github.com/google/gemma_pytorch/commit/03e657582d17cb5a8617ebf333c1c16f3694670e
+                switch (model.type) {
+                    case e_model::MODEL_9B:  Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));   break;
+                    case e_model::MODEL_27B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd / n_head))); break;
+                    default: GGML_ASSERT(false);
+                };
+                cb(Qcur, "Qcur_scaled", il);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow);
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask_l, n_tokens, kv_head, n_kv, 1.0f, cb, il);
+            }
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                    model.layers[il].attn_post_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_post_norm", il);
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
+            cb(sa_out, "sa_out", il);
+
+            cur = llm_build_norm(ctx0, sa_out, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            // feed-forward network
+            {
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+            cb(cur, "ffn_post_norm", -1);
+
+            cur = ggml_add(ctx0, cur, sa_out);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+
+        // final logit soft-capping
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
+        cur = ggml_tanh(ctx0, cur);
+        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
+
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+
+    struct ggml_cgraph * build_starcoder2() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, model.layers[il].ffn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        NULL,                      NULL,                        NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_mamba() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t d_model = n_embd;
+        const int64_t d_conv  = hparams.ssm_d_conv;
+        const int64_t d_inner = hparams.ssm_d_inner;
+        GGML_ASSERT(2 * d_model == d_inner);
+        const int64_t d_state = hparams.ssm_d_state;
+        const int64_t dt_rank = hparams.ssm_dt_rank;
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        // {n_embd, n_tokens}
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        struct ggml_tensor * state_mask = build_inp_s_mask();
+        struct ggml_tensor * state_seq  = build_inp_s_seq();
+
+        for (int il = 0; il < n_layer; ++il) {
+            // (ab)using the KV cache to store the states
+            struct ggml_tensor * conv_states = ggml_reshape_2d(ctx0, kv_self.k_l[il], hparams.n_embd_k_s(), kv_self.size);
+            struct ggml_tensor * ssm_states  = ggml_reshape_2d(ctx0, kv_self.v_l[il], hparams.n_embd_v_s(), kv_self.size);
+
+            // clear states of sequences which are starting at the beginning of this batch
+            {
+                conv_states = ggml_mul(ctx0,
+                    ggml_view_2d(ctx0, conv_states, conv_states->ne[0], n_kv, conv_states->nb[1], kv_head*conv_states->nb[1]),
+                    state_mask);
+                ssm_states  = ggml_mul(ctx0,
+                    ggml_view_2d(ctx0, ssm_states, ssm_states->ne[0], n_kv, ssm_states->nb[1], kv_head*ssm_states->nb[1]),
+                    state_mask);
+            }
+
+            conv_states = ggml_reshape_3d(ctx0, conv_states, d_conv - 1, d_inner, n_kv);
+            ssm_states  = ggml_reshape_3d(ctx0,  ssm_states,    d_state, d_inner, n_kv);
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // {n_embd, 2*d_inner} * {n_embd, n_tokens} => {2*d_inner, n_tokens}
+            struct ggml_tensor * xz = llm_build_lora_mm(lctx, ctx0, model.layers[il].ssm_in, cur);
+            // split the above in two
+            // => {d_inner, n_tokens}
+            struct ggml_tensor * x = ggml_view_2d(ctx0, xz, d_inner, xz->ne[1], xz->nb[1], 0);
+            struct ggml_tensor * z = ggml_view_2d(ctx0, xz, d_inner, xz->ne[1], xz->nb[1], ggml_element_size(xz)*d_inner);
+
+            // conv
+            {
+                // Custom operator which is needed only to ease simultaneous sequence processing.
+                // For a single sequence, the equivalent is to concatenate the columns of conv_states and x,
+                // then make a self-overlapping view of that over d_conv columns at each stride in the 3rd dimension,
+                // then element-wise multiply that with the conv1d weigth,
+                // then sum the elements of each row,
+                // (the last two steps are a dot product over rows (also doable with mul_mat))
+                // then permute away the ne[0] dimension,
+                // and then you're left with the resulting x tensor.
+                // The new conv_states is the last (d_conv - 1) columns
+                // of the last 3rd dimensional "layer" of the self-overlapping view.
+                // For simultaneous sequences, it's more complicated.
+                struct ggml_tensor * x_conv = ggml_ssm_conv(ctx0, conv_states, x, model.layers[il].ssm_conv1d, state_seq);
+
+                // store last (d_conv - 1) columns of the conv_state part of x_conv back into the KV cache
+                ggml_build_forward_expand(gf,
+                    ggml_cpy(ctx0,
+                        ggml_view_2d(ctx0, x_conv, d_conv - 1, d_inner*n_kv, d_conv*ggml_element_size(x_conv), (1+d_inner*n_tokens)*ggml_element_size(x_conv)),
+                        ggml_view_1d(ctx0, kv_self.k_l[il], (d_conv - 1)*(d_inner)*(n_kv), kv_head*(d_conv - 1)*(d_inner)*ggml_element_size(x_conv))));
+
+                // extract x from x_conv
+                x = ggml_view_2d(ctx0, x_conv, d_inner, n_tokens, d_inner*ggml_element_size(x_conv), 0);
+
+                // bias
+                x = ggml_add(ctx0, x, model.layers[il].ssm_conv1d_b);
+
+                x = ggml_silu(ctx0, x);
+            }
+
+            // ssm
+            {
+                // {d_inner, dt_rank + 2*d_state} * {d_inner, n_tokens} => {dt_rank + 2*d_state, n_tokens}
+                struct ggml_tensor * x_db = llm_build_lora_mm(lctx, ctx0, model.layers[il].ssm_x, x);
+                // split
+                struct ggml_tensor * dt = ggml_view_2d(ctx0, x_db, dt_rank, n_tokens, x_db->nb[1], 0);
+                struct ggml_tensor * B  = ggml_view_2d(ctx0, x_db, d_state, n_tokens, x_db->nb[1], ggml_element_size(x_db)*dt_rank);
+                struct ggml_tensor * C  = ggml_view_2d(ctx0, x_db, d_state, n_tokens, x_db->nb[1], ggml_element_size(x_db)*(dt_rank+d_state));
+
+                // {dt_rank, d_inner} * {dt_rank, n_tokens} => {d_inner, n_tokens}
+                dt = llm_build_lora_mm(lctx, ctx0, model.layers[il].ssm_dt, dt);
+                dt = ggml_add(ctx0, dt, model.layers[il].ssm_dt_b);
+
+                // Custom operator to optimize the parallel associative scan
+                // as described in the Annex D of the Mamba paper.
+                // => {d_inner, n_tokens} and {d_state, d_inner, n_kv} combined,
+                // because only a single tensor can be returned.
+                struct ggml_tensor * y_ssm_states = ggml_ssm_scan(ctx0, ssm_states, x, dt, model.layers[il].ssm_a, B, C, state_seq);
+
+                // store last states (the second part of y_ssm_states)
+                ggml_build_forward_expand(gf,
+                    ggml_cpy(ctx0,
+                        ggml_view_1d(ctx0, y_ssm_states, d_state*d_inner*n_kv, d_inner*n_tokens*ggml_element_size(y_ssm_states)),
+                        ggml_view_1d(ctx0, kv_self.v_l[il], d_state*d_inner*n_kv, kv_head*d_state*d_inner*ggml_element_size(ssm_states))));
+
+                struct ggml_tensor * y = ggml_view_2d(ctx0, y_ssm_states, d_inner, n_tokens, d_inner*ggml_element_size(y_ssm_states), 0);
+
+                if (il == n_layer - 1) {
+                    // skip computing output for unused tokens
+                    struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                    x    = ggml_get_rows(ctx0,    x, inp_out_ids);
+                    y    = ggml_get_rows(ctx0,    y, inp_out_ids);
+                    z    = ggml_get_rows(ctx0,    z, inp_out_ids);
+                    inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+                }
+
+                // {d_inner, n_tokens} * {d_inner} => {d_inner, n_tokens}
+                y = ggml_add(ctx0, y, ggml_mul(ctx0, x, model.layers[il].ssm_d));
+                y = ggml_mul(ctx0, y, ggml_silu(ctx0, z));
+
+                // {d_inner, n_embd} * {d_inner, n_tokens} => {n_embd, n_tokens}
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].ssm_out, y);
+            }
+
+            // residual
+            cur = ggml_add(ctx0, cur, inpL);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        // final rmsnorm
+        cur = llm_build_norm(ctx0, inpL, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_command_r() {
+
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        const float f_logit_scale = hparams.f_logit_scale;
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+            struct ggml_tensor * ffn_inp = cur;
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                if (model.layers[il].attn_q_norm) {
+                    Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens,
+                                ggml_element_size(Qcur) * n_embd_head,
+                                ggml_element_size(Qcur) * n_embd_head * n_head,
+                                0);
+                    cb(Qcur, "Qcur", il);
+                    Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens,
+                                ggml_element_size(Kcur) * n_embd_head,
+                                ggml_element_size(Kcur) * n_embd_head * n_head_kv,
+                                0);
+                    cb(Kcur, "Kcur", il);
+
+                    Qcur = llm_build_norm(ctx0, Qcur, hparams,
+                                model.layers[il].attn_q_norm,
+                                NULL,
+                                LLM_NORM, cb, il);
+                    cb(Qcur, "Qcur", il);
+
+                    Kcur = llm_build_norm(ctx0, Kcur, hparams,
+                            model.layers[il].attn_k_norm,
+                            NULL,
+                            LLM_NORM, cb, il);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur     = ggml_get_rows(ctx0,     cur, inp_out_ids);
+                inpL    = ggml_get_rows(ctx0,    inpL, inp_out_ids);
+                ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
+            }
+
+            struct ggml_tensor * attn_out = cur;
+
+            // feed-forward network
+            {
+                cur = llm_build_ffn(ctx0, lctx, ffn_inp,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            // add together residual + FFN + self-attention
+            cur = ggml_add(ctx0, cur, inpL);
+            cur = ggml_add(ctx0, cur, attn_out);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+
+        if (f_logit_scale) {
+            cur = ggml_scale(ctx0, cur, f_logit_scale);
+        }
+
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+
+    }
+
+    // ref: https://allenai.org/olmo
+    // based on the original build_llama() function, changes:
+    //   * non-parametric layer norm
+    //   * clamp qkv
+    //   * removed bias
+    //   * removed MoE
+    struct ggml_cgraph * build_olmo() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    NULL, NULL,
+                    LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (hparams.f_clamp_kqv > 0.0f) {
+                    Qcur = ggml_clamp(ctx0, Qcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (hparams.f_clamp_kqv > 0.0f) {
+                    Kcur = ggml_clamp(ctx0, Kcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (hparams.f_clamp_kqv > 0.0f) {
+                    Vcur = ggml_clamp(ctx0, Vcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, nullptr,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    NULL, NULL,
+                    LLM_NORM, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = llm_build_ffn(ctx0, lctx, cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                NULL, NULL,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_openelm() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            const int64_t n_head    = hparams.n_head(il);
+            const int64_t n_head_kv = hparams.n_head_kv(il);
+            const int64_t n_head_qkv = 2*n_head_kv + n_head;
+
+            cur = inpL;
+            struct ggml_tensor * residual = cur;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                cur = ggml_reshape_3d(ctx0, cur, n_embd_head_k, n_head_qkv, n_tokens);
+
+                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, cur->nb[1], cur->nb[2], 0));
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*n_head));
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*(n_head+n_head_kv)));
+                cb(Vcur, "Vcur", il);
+
+                Qcur = llm_build_norm(ctx0, Qcur, hparams,
+                        model.layers[il].attn_q_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(Qcur, "Qcur", il);
+
+                Kcur = llm_build_norm(ctx0, Kcur, hparams,
+                        model.layers[il].attn_k_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(Kcur, "Kcur", il);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, NULL, n_rot, rope_type, n_ctx_orig,
+                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, NULL, n_rot, rope_type, n_ctx_orig,
+                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                Vcur = ggml_reshape_2d(ctx0, Vcur, n_embd_head * n_head_kv, n_tokens);
+                cb(Qcur, "Vcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                residual = ggml_get_rows(ctx0, residual, inp_out_ids);
+                cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        // norm
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_gptneox() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm,
+                    model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                cb(cur, "bqkv", il);
+
+                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            // ffn
+            if (hparams.use_par_res) {
+                // attention and ffn are computed in parallel
+                // x = x + attn(ln1(x)) + ffn(ln2(x))
+
+                struct ggml_tensor * attn_out = cur;
+
+                cur = llm_build_norm(ctx0, inpL, hparams,
+                        model.layers[il].ffn_norm,
+                        model.layers[il].ffn_norm_b,
+                        LLM_NORM, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        NULL,                      NULL,                        NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+                cb(cur, "ffn_out", il);
+
+                cur = ggml_add(ctx0, cur, inpL);
+                cb(cur, "ffn_out", il);
+
+                cur = ggml_add(ctx0, cur, attn_out);
+                cur = lctx.cvec.apply_to(ctx0, cur, il);
+                cb(cur, "l_out", il);
+
+                // input for next layer
+                inpL = cur;
+            } else {
+                // attention and ffn are computed sequentially
+                // x = x + attn(ln1(x))
+                // x = x + ffn(ln2(x))
+
+                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+                cb(ffn_inp, "ffn_inp", il);
+
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm,
+                        model.layers[il].ffn_norm_b,
+                        LLM_NORM, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        NULL,                      NULL,                        NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+                cb(cur, "ffn_out", il);
+
+                cur = ggml_add(ctx0, cur, ffn_inp);
+                cur = lctx.cvec.apply_to(ctx0, cur, il);
+                cb(cur, "l_out", il);
+
+                // input for next layer
+                inpL = cur;
+            }
+        }
+
+        cur = llm_build_norm(ctx0, inpL, hparams,
+                model.output_norm,
+                model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_arctic() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = llm_build_ffn(ctx0, lctx, cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            struct ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp);
+            cb(ffn_out, "ffn_out", il);
+
+            // MoE
+            cur = llm_build_norm(ctx0, inpSA, hparams,
+                    model.layers[il].ffn_norm_exps, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm_exps", il);
+
+            cur = llm_build_moe_ffn(ctx0, lctx, cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    cb, il);
+            cb(cur, "ffn_moe_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_out);
+            cb(cur, "ffn_out", il);
+
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_deepseek2() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        bool is_lite = (hparams.n_layer == 27);
+
+        // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
+        // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
+        const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
+        const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(hparams.n_embd_head_k));
+        const float attn_factor_scaled = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
+
+        const uint32_t n_embd_head_qk_rope = hparams.n_rot;
+        const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+        const uint32_t kv_lora_rank = hparams.n_lora_kv;
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        // {n_embd, n_tokens}
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self_attention
+            {
+                struct ggml_tensor * q = NULL;
+                if (!is_lite) {
+                    // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens}
+                    q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
+                    cb(q, "q", il);
+
+                    q = llm_build_norm(ctx0, q, hparams,
+                            model.layers[il].attn_q_a_norm, NULL,
+                            LLM_NORM_RMS, cb, il);
+                    cb(q, "q", il);
+
+                    // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens}
+                    q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
+                    cb(q, "q", il);
+                } else {
+                    q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                    cb(q, "q", il);
+                }
+
+                // split into {n_head * n_embd_head_qk_nope, n_tokens}
+                struct ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
+                        ggml_row_size(q->type, hparams.n_embd_head_k),
+                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                        0);
+                cb(q_nope, "q_nope", il);
+
+                // and {n_head * n_embd_head_qk_rope, n_tokens}
+                struct ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
+                        ggml_row_size(q->type, hparams.n_embd_head_k),
+                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                        ggml_row_size(q->type, n_embd_head_qk_nope));
+                cb(q_pe, "q_pe", il);
+
+                // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
+                struct ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+                cb(kv_pe_compresseed, "kv_pe_compresseed", il);
+
+                // split into {kv_lora_rank, n_tokens}
+                struct ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
+                        kv_pe_compresseed->nb[1],
+                        0);
+                cb(kv_compressed, "kv_compressed", il);
+
+                // and {n_embd_head_qk_rope, n_tokens}
+                struct ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
+                        kv_pe_compresseed->nb[1],
+                        kv_pe_compresseed->nb[1],
+                        ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
+                cb(k_pe, "k_pe", il);
+
+                kv_compressed = ggml_cont(ctx0, kv_compressed); // TODO: the CUDA backend does not support non-contiguous norm
+                kv_compressed = llm_build_norm(ctx0, kv_compressed, hparams,
+                        model.layers[il].attn_kv_a_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(kv_compressed, "kv_compressed", il);
+
+                // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
+                struct ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
+                cb(kv, "kv", il);
+
+                // split into {n_head * n_embd_head_qk_nope, n_tokens}
+                struct ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
+                        ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
+                        ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                        0);
+                cb(k_nope, "k_nope", il);
+
+                // and {n_head * n_embd_head_v, n_tokens}
+                struct ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
+                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
+                        ggml_row_size(kv->type, (n_embd_head_qk_nope)));
+                cb(v_states, "v_states", il);
+
+                v_states = ggml_cont(ctx0, v_states);
+                cb(v_states, "v_states", il);
+
+                v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens,
+                    ggml_row_size(kv->type, hparams.n_embd_head_v * n_head),
+                    0);
+                cb(v_states, "v_states", il);
+
+                q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend does not support non-contiguous RoPE
+                q_pe = ggml_rope_ext(
+                    ctx0, q_pe, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor_scaled, beta_fast, beta_slow
+                );
+                cb(q_pe, "q_pe", il);
+
+                // shared RoPE key
+                k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend does not support non-contiguous RoPE
+                k_pe = ggml_rope_ext(
+                    ctx0, k_pe, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor_scaled, beta_fast, beta_slow
+                );
+                cb(k_pe, "k_pe", il);
+
+                struct ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
+                cb(q_states, "q_states", il);
+
+                struct ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
+                cb(k_states, "k_states", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        k_states, v_states, q_states, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            if ((uint32_t) il < hparams.n_layer_dense_lead) {
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            } else {
+                // MoE branch
+                ggml_tensor * moe_out =
+                        llm_build_moe_ffn(ctx0, lctx, cur,
+                            model.layers[il].ffn_gate_inp,
+                            model.layers[il].ffn_up_exps,
+                            model.layers[il].ffn_gate_exps,
+                            model.layers[il].ffn_down_exps,
+                            n_expert, n_expert_used,
+                            LLM_FFN_SILU, false,
+                            true, hparams.expert_weights_scale,
+                            cb, il);
+                cb(moe_out, "ffn_moe_out", il);
+
+                // FFN shared expert
+                {
+                    ggml_tensor * ffn_shexp = llm_build_ffn(ctx0, lctx, cur,
+                            model.layers[il].ffn_up_shexp,   NULL, NULL,
+                            model.layers[il].ffn_gate_shexp, NULL, NULL,
+                            model.layers[il].ffn_down_shexp, NULL, NULL,
+                            NULL,
+                            LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                    cb(ffn_shexp, "ffn_shexp", il);
+
+                    cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                    cb(cur, "ffn_out", il);
+                }
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = ggml_mul_mat(ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_bitnet() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                float q_scale; std::memcpy(&q_scale, model.layers[il].wq->op_params, sizeof(float));
+                // Note: we could save this scale operation by applying the Q scale on the K * Q product further down
+                // (which also uses a scale). This works on the CPU and Metal backends, but produces NaNs on CUDA.
+                Qcur = ggml_scale(ctx0, Qcur, q_scale);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                // B1.K
+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+                float k_scale; std::memcpy(&k_scale, model.layers[il].wk->op_params, sizeof(float));
+                Kcur = ggml_scale(ctx0, Kcur, k_scale);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                // B1.V
+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                float v_scale; std::memcpy(&v_scale, model.layers[il].wv->op_params, sizeof(float));
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_scale(ctx0, Vcur, v_scale);
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                    v_scale = 1;
+                }
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                llm_build_kv_store(ctx0, hparams, cparams, kv_self, gf, Kcur, Vcur, n_tokens, kv_head, cb, il);
+
+                const int64_t n_ctx                 = cparams.n_ctx;
+                const int64_t n_head                = hparams.n_head();
+                const int64_t n_head_kv             = hparams.n_head_kv();
+                const int64_t n_embd_head_k         = hparams.n_embd_head_k;
+                const int64_t n_embd_k_gqa          = hparams.n_embd_k_gqa();
+                const int64_t n_embd_head_v         = hparams.n_embd_head_v;
+                const int64_t n_embd_v_gqa          = hparams.n_embd_v_gqa();
+
+                float                      kq_scale = 1.0f/sqrtf(float(n_embd_head));
+                // We would use this if we did not apply the Q scale above. Sadly, this fails on CUDA.
+                //float                      kq_scale = q_scale/sqrtf(float(n_embd_head));
+                struct ggml_tensor *       cur_attn;
+                struct ggml_tensor *              q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+                cb(q, "q", il);
+
+                struct ggml_tensor * k =
+                    ggml_view_3d(ctx0, kv_self.k_l[il],
+                            n_embd_head_k, n_kv, n_head_kv,
+                            ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa),
+                            ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k),
+                            0);
+                cb(k, "k", il);
+
+                if (cparams.flash_attn) {
+
+                    // split cached v into n_head heads (not transposed)
+                    struct ggml_tensor * v =
+                        ggml_view_3d(ctx0, kv_self.v_l[il],
+                                n_embd_head_v, n_kv, n_head_kv,
+                                ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa),
+                                ggml_row_size(kv_self.v_l[il]->type, n_embd_head_v),
+                                0);
+                    cb(v, "v", il);
+
+                    cur_attn = ggml_flash_attn_ext(ctx0, q, k, v, KQ_mask, kq_scale, hparams.f_max_alibi_bias);
+
+                    cur_attn = ggml_reshape_2d(ctx0, cur, n_embd_head_v*n_head, n_tokens);
+                } else {
+                    struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+                    cb(kq, "kq", il);
+
+                    kq = ggml_soft_max_ext(ctx0, kq, KQ_mask, kq_scale, hparams.f_max_alibi_bias);
+                    cb(kq, "kq_soft_max_ext", il);
+
+                    GGML_ASSERT(kv_self.size == n_ctx);
+
+                    // split cached v into n_head heads
+                    struct ggml_tensor * v =
+                        ggml_view_3d(ctx0, kv_self.v_l[il],
+                                n_kv, n_embd_head_v, n_head_kv,
+                                ggml_element_size(kv_self.v_l[il])*n_ctx,
+                                ggml_element_size(kv_self.v_l[il])*n_ctx*n_embd_head_v,
+                                0);
+                    cb(v, "v", il);
+
+                    struct ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq);
+                    cb(kqv, "kqv", il);
+
+                    struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+                    cb(kqv_merged, "kqv_merged", il);
+
+                    cur_attn = ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens);
+                    cb(cur_attn, "kqv_merged_cont", il);
+                }
+
+                cur_attn = llm_build_norm(ctx0, cur_attn, hparams,
+                        model.layers[il].attn_sub_norm, NULL,
+                        LLM_NORM_RMS, cb, il, 1/(v_scale*v_scale));
+                cb(cur_attn, "attn_sub_norm", il);
+
+                ggml_build_forward_expand(gf, cur_attn);
+
+                cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur_attn);
+                float wo_scale; std::memcpy(&wo_scale, model.layers[il].wo->op_params, sizeof(float));
+                cur = ggml_scale(ctx0, cur, wo_scale);
+
+                cb(cur, "kqv_out", il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward forward
+            if (model.layers[il].ffn_gate_inp == nullptr) {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                struct ggml_tensor *tmp = ggml_mul_mat(ctx0, model.layers[il].ffn_up, cur);
+                float ffn_up_scale; std::memcpy(&ffn_up_scale, model.layers[il].ffn_up->op_params, sizeof(float));
+
+                cb(tmp, "ffn_up", il);
+
+                cur = ggml_mul_mat(ctx0, model.layers[il].ffn_gate, cur);
+                float ffn_gate_scale; std::memcpy(&ffn_gate_scale, model.layers[il].ffn_gate->op_params, sizeof(float));
+                cur = ggml_scale(ctx0, cur, ffn_gate_scale);
+
+                cb(cur, "ffn_gate", il);
+
+
+                // combine this with the above scale into ggml_scaled_silu
+                cur = ggml_silu(ctx0, cur);
+                cb(cur, "ffn_silu", il);
+
+                cur = ggml_mul(ctx0, cur, tmp);
+                cb(cur, "ffn_gate_par", il);
+
+                cur = llm_build_norm(ctx0, cur, hparams,
+                                model.layers[il].ffn_sub_norm, NULL,
+                                LLM_NORM_RMS, cb, il, 1/(ffn_up_scale*ffn_up_scale));
+                cb(cur, "ffn_sub_norm", il);
+
+                cur = ggml_mul_mat(ctx0, model.layers[il].ffn_down, cur);
+                float ffn_down_scale; std::memcpy(&ffn_down_scale, model.layers[il].ffn_down->op_params, sizeof(float));
+                cur = ggml_scale(ctx0, cur, ffn_down_scale);
+                cb(cur, "ffn_down", il);
+            }
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+        return gf;
+    }
+
+    struct ggml_cgraph * build_t5() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        if (lctx.is_encoding) {
+            struct ggml_tensor * pos_bucket_enc = llm_build_pos_bucket(false);
+
+            // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+            struct ggml_tensor * KQ_mask_enc = build_inp_KQ_mask(false);
+
+            for (int il = 0; il < n_layer; ++il) {
+                struct ggml_tensor * inpSA = inpL;
+
+                // norm
+                cur = llm_build_norm(ctx0, inpL, hparams,
+                        model.layers[il].attn_norm_enc, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "attn_norm", il);
+
+                // self-attention
+                {
+                    struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq_enc, cur);
+                    cb(Qcur, "Qcur", il);
+
+                    struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk_enc, cur);
+                    cb(Kcur, "Kcur", il);
+
+                    struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv_enc, cur);
+                    cb(Vcur, "Vcur", il);
+
+                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+                    struct ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+                    struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
+
+                    struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+                    cb(kq, "kq", il);
+
+                    struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc;
+                    struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_enc, attn_rel_b);
+                    struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias);
+                    cb(kq_b, "kq_b", il);
+
+                    kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_enc, 1.0f, hparams.f_max_alibi_bias);
+                    cb(kq, "kq_soft_max_ext", il);
+
+                    struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens)));
+                    cb(v, "v", il);
+
+                    struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_tokens, n_embd_head, n_head_kv), kq);
+                    cb(kqv, "kqv", il);
+
+                    struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+                    cb(kqv_merged, "kqv_merged", il);
+
+                    cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
+                    cb(cur, "kqv_merged_cont", il);
+
+                    ggml_build_forward_expand(gf, cur);
+
+                    cur = ggml_mul_mat(ctx0, model.layers[il].wo_enc, cur);
+                    cb(cur, "kqv_out", il);
+                }
+
+                if (il == n_layer - 1) {
+                    // skip computing output for unused tokens
+                    struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                    n_tokens = n_outputs;
+                    cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                    inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+                }
+
+                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+                cb(ffn_inp, "ffn_inp", il);
+
+                // feed-forward network
+                {
+                    cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                            model.layers[il].ffn_norm_enc, NULL,
+                            LLM_NORM_RMS, cb, il);
+                    cb(cur, "ffn_norm", il);
+
+                    // T5 uses relu, flan-T5 uses gelu-gated
+                    cur = llm_build_ffn(ctx0, lctx, cur,
+                            model.layers[il].ffn_up_enc,   NULL, NULL,
+                            model.layers[il].ffn_gate_enc, NULL, NULL,
+                            model.layers[il].ffn_down_enc, NULL, NULL,
+                            NULL,
+                            model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU,
+                            model.layers[il].ffn_gate_enc ? LLM_FFN_PAR  : LLM_FFN_SEQ,
+                            cb, il);
+                    cb(cur, "ffn_out", il);
+                }
+
+                cur = ggml_add(ctx0, cur, ffn_inp);
+                cb(cur, "ffn_out", il);
+
+                ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
+                if (layer_dir != nullptr) {
+                    cur = ggml_add(ctx0, cur, layer_dir);
+                }
+                cb(cur, "l_out", il);
+
+                // input for next layer
+                inpL = cur;
+            }
+
+            cur = inpL;
+            cb(cur, "result_embd", -1);
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                    model.output_norm_enc, NULL,
+                    LLM_NORM_RMS, cb, -1);
+            cb(cur, "result_norm", -1);
+        } else {
+            GGML_ASSERT(n_outputs_enc > 0 && "call llama_encode() first");
+
+            struct ggml_tensor * embd_enc       = llm_build_inp_embd_enc();
+            struct ggml_tensor * pos_bucket_dec = llm_build_pos_bucket(true);
+
+            struct ggml_tensor * KQ_mask_dec   = build_inp_KQ_mask();
+            struct ggml_tensor * KQ_mask_cross = llm_build_inp_KQ_mask_cross();
+
+            for (int il = 0; il < n_layer; ++il) {
+                struct ggml_tensor * inpSA = inpL;
+
+                // norm
+                cur = llm_build_norm(ctx0, inpL, hparams,
+                        model.layers[il].attn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "attn_norm", il);
+
+                // self-attention
+                {
+                    struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                    cb(Qcur, "Qcur", il);
+
+                    struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+                    cb(Kcur, "Kcur", il);
+
+                    struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                    cb(Vcur, "Vcur", il);
+
+                    llm_build_kv_store(ctx0, hparams, cparams, kv_self, gf, Kcur, Vcur, n_tokens, kv_head, cb, il);
+
+                    struct ggml_tensor * k =
+                        ggml_view_3d(ctx0, kv_self.k_l[il],
+                                n_embd_head_k, n_kv, n_head_kv,
+                                ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa),
+                                ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k),
+                                0);
+                    cb(k, "k", il);
+
+                    struct ggml_tensor * v =
+                        ggml_view_3d(ctx0, kv_self.v_l[il],
+                                n_kv, n_embd_head_v, n_head_kv,
+                                ggml_element_size(kv_self.v_l[il])*n_ctx,
+                                ggml_element_size(kv_self.v_l[il])*n_ctx*n_embd_head_v,
+                                0);
+                    cb(v, "v", il);
+
+                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+
+                    struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+
+                    struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+                    cb(kq, "kq", il);
+
+                    struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b;
+                    struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_dec, attn_rel_b);
+                    struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias);
+                    cb(kq_b, "kq_b", il);
+
+                    kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_dec, 1.0f, hparams.f_max_alibi_bias);
+                    cb(kq, "kq_soft_max_ext", il);
+
+                    struct ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq);
+                    cb(kqv, "kqv", il);
+
+                    struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+                    cb(kqv_merged, "kqv_merged", il);
+
+                    cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
+                    cb(cur, "kqv_merged_cont", il);
+
+                    ggml_build_forward_expand(gf, cur);
+
+                    cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur);
+                    cb(cur, "kqv_out", il);
+                }
+
+                cur = ggml_add(ctx0, cur, inpSA);
+                cb(cur, "cross_inp", il);
+
+                struct ggml_tensor * inpCA = cur;
+
+                // norm
+                cur = llm_build_norm(ctx0, cur, hparams,
+                        model.layers[il].attn_norm_cross, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "attn_norm_cross", il);
+
+                // cross-attention
+                {
+                    struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq_cross, cur);
+                    cb(Qcur, "Qcur", il);
+
+                    struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk_cross, embd_enc);
+                    cb(Kcur, "Kcur", il);
+
+                    struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv_cross, embd_enc);
+                    cb(Vcur, "Vcur", il);
+
+                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc);
+
+                    struct ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
+                    struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
+
+                    struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
+                    cb(kq, "kq", il);
+
+                    kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias);
+                    cb(kq, "kq_soft_max_ext", il);
+
+                    struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc)));
+                    cb(v, "v", il);
+
+                    struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq);
+                    cb(kqv, "kqv", il);
+
+                    struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+                    cb(kqv_merged, "kqv_merged", il);
+
+                    cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
+                    cb(cur, "kqv_merged_cont", il);
+
+                    ggml_build_forward_expand(gf, cur);
+
+                    cur = ggml_mul_mat(ctx0, model.layers[il].wo_cross, cur);
+                    cb(cur, "kqv_out", il);
+                }
+
+                if (il == n_layer - 1) {
+                    // skip computing output for unused tokens
+                    struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                    n_tokens = n_outputs;
+                    cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                    inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+                    inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids);
+                }
+
+                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA);
+                cb(ffn_inp, "ffn_inp", il);
+
+                // feed-forward network
+                {
+                    cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                            model.layers[il].ffn_norm, NULL,
+                            LLM_NORM_RMS, cb, il);
+                    cb(cur, "ffn_norm", il);
+
+                    // T5 uses relu, flan-T5 uses gelu-gated
+                    cur = llm_build_ffn(ctx0, lctx, cur,
+                            model.layers[il].ffn_up,   NULL, NULL,
+                            model.layers[il].ffn_gate, NULL, NULL,
+                            model.layers[il].ffn_down, NULL, NULL,
+                            NULL,
+                            model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU,
+                            model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ,
+                            cb, il);
+                    cb(cur, "ffn_out", il);
+                }
+
+                cur = ggml_add(ctx0, cur, ffn_inp);
+                cb(cur, "ffn_out", il);
+
+                ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
+                if (layer_dir != nullptr) {
+                    cur = ggml_add(ctx0, cur, layer_dir);
+                }
+                cb(cur, "l_out", il);
+
+                // input for next layer
+                inpL = cur;
+            }
+
+            cur = inpL;
+            cb(cur, "result_embd", -1);
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                    model.output_norm, NULL,
+                    LLM_NORM_RMS, cb, -1);
+            cb(cur, "result_norm", -1);
+
+            // lm_head
+            cur = ggml_mul_mat(ctx0, model.output, cur);
+            cb(cur, "result_output", -1);
+        }
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_jais() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm,
+                    model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                cb(cur, "bqkv", il);
+
+                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*cur->nb[0]*(n_embd)));
+                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd)));
+                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd + n_embd_gqa)));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/float(n_embd_head), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            // add the input
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // FF
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm,
+                        model.layers[il].ffn_norm_b,
+                        LLM_NORM, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            }
+
+            inpL = ggml_add(ctx0, cur, ffn_inp);
+            cb(inpL, "l_out", il);
+        }
+
+        cur = llm_build_norm(ctx0, inpL, hparams,
+                model.output_norm,
+                model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_chatglm() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm,
+                    NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                struct ggml_tensor * Qcur = nullptr;
+                struct ggml_tensor * Kcur = nullptr;
+                struct ggml_tensor * Vcur = nullptr;
+
+                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                cb(cur, "bqkv", il);
+
+                Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+                //printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor);
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur_rope", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur_rope", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            // Add the input
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // FF
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm,
+                        NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        NULL,                      NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SWIGLU, LLM_FFN_SEQ, cb, il);
+                cb(cur, "ffn_out", il);
+
+            }
+
+            inpL = ggml_add(ctx0, cur, ffn_inp);
+            cb(inpL, "l_out", il);
+        }
+
+        cur = llm_build_norm(ctx0, inpL, hparams,
+                model.output_norm,
+                NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+};
+
+static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector & ids) {
+    llama_batch dummy;
+    dummy.n_tokens = 0;
+
+    llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { };
+
+    struct llm_build_context llm(lctx, dummy, cb, false);
+
+    llm.init();
+
+    struct ggml_cgraph * result = llm.build_defrag(ids);
+
+    llm.free();
+
+    return result;
+}
+
+static struct ggml_cgraph * llama_build_graph_k_shift(llama_context & lctx) {
+    llama_batch dummy;
+    dummy.n_tokens = 0;
+
+    llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { };
+
+    struct llm_build_context llm(lctx, dummy, cb, false);
+
+    llm.init();
+
+    struct ggml_cgraph * result = llm.build_k_shift();
+
+    llm.free();
+
+    return result;
+}
+
+static struct ggml_cgraph * llama_build_graph_s_copy(llama_context & lctx) {
+    llama_batch dummy;
+    dummy.n_tokens = 0;
+
+    llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { };
+
+    struct llm_build_context llm(lctx, dummy, cb, false);
+
+    llm.init();
+
+    struct ggml_cgraph * result = llm.build_s_copy();
+
+    llm.free();
+
+    return result;
+}
+
+static struct ggml_cgraph * llama_build_graph(
+         llama_context & lctx,
+     const llama_batch & batch,
+                  bool   worst_case) {
+    const auto & model = lctx.model;
+
+    // this callback allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.)
+    llm_build_cb cb = [&](struct ggml_tensor * cur, const char * name, int il) {
+        if (il >= 0) {
+            ggml_format_name(cur, "%s-%d", name, il);
+        } else {
+            ggml_set_name(cur, name);
+        }
+
+        if (!lctx.cparams.offload_kqv) {
+            if (strcmp(name, "kqv_merged_cont") == 0) {
+                // all nodes between the KV store and the attention output are run on the CPU
+                ggml_backend_sched_set_tensor_backend(lctx.sched, cur, lctx.backend_cpu);
+            }
+        }
+
+        // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
+        // FIXME: fix in ggml_backend_sched
+        const bool full_offload = lctx.model.n_gpu_layers > (int)lctx.model.hparams.n_layer;
+        if (batch.n_tokens < 32 || full_offload) {
+            if (il != -1 && strcmp(name, "norm") == 0) {
+                for (auto * backend : lctx.backends) {
+                    if (ggml_backend_supports_buft(backend, lctx.model.buft_layer[il].buft) &&
+                        (ggml_backend_supports_op(backend, cur) || ggml_backend_offload_op(backend, cur))) {
+                        ggml_backend_sched_set_tensor_backend(lctx.sched, cur, backend);
+                        break;
+                    }
+                }
+            }
+        }
+    };
+
+    struct ggml_cgraph * result = NULL;
+
+    struct llm_build_context llm(lctx, batch, cb, worst_case);
+
+    llm.init();
+
+    switch (model.arch) {
+        case LLM_ARCH_LLAMA:
+            {
+                result = llm.build_llama();
+            } break;
+        case LLM_ARCH_BAICHUAN:
+            {
+                result = llm.build_baichuan();
+            } break;
+        case LLM_ARCH_FALCON:
+            {
+                result = llm.build_falcon();
+            } break;
+        case LLM_ARCH_GROK:
+            {
+                result = llm.build_grok();
+            } break;
+        case LLM_ARCH_STARCODER:
+            {
+                result = llm.build_starcoder();
+            } break;
+        case LLM_ARCH_REFACT:
+            {
+                result = llm.build_refact();
+            } break;
+        case LLM_ARCH_BERT:
+        case LLM_ARCH_JINA_BERT_V2:
+        case LLM_ARCH_NOMIC_BERT:
+            {
+                result = llm.build_bert();
+            } break;
+        case LLM_ARCH_BLOOM:
+            {
+                result = llm.build_bloom();
+            } break;
+        case LLM_ARCH_MPT:
+            {
+                result = llm.build_mpt();
+            } break;
+         case LLM_ARCH_STABLELM:
+            {
+                result = llm.build_stablelm();
+            } break;
+        case LLM_ARCH_QWEN:
+            {
+                result = llm.build_qwen();
+            } break;
+        case LLM_ARCH_QWEN2:
+            {
+                result = llm.build_qwen2();
+            } break;
+        case LLM_ARCH_QWEN2MOE:
+            {
+                result = llm.build_qwen2moe();
+            } break;
+        case LLM_ARCH_PHI2:
+            {
+                result = llm.build_phi2();
+            } break;
+        case LLM_ARCH_PHI3:
+            {
+                result = llm.build_phi3();
+            } break;
+        case LLM_ARCH_PLAMO:
+            {
+                result = llm.build_plamo();
+            } break;
+        case LLM_ARCH_GPT2:
+            {
+                result = llm.build_gpt2();
+            } break;
+        case LLM_ARCH_CODESHELL:
+            {
+                result = llm.build_codeshell();
+            } break;
+        case LLM_ARCH_ORION:
+            {
+                result = llm.build_orion();
+            } break;
+        case LLM_ARCH_INTERNLM2:
+            {
+                result = llm.build_internlm2();
+            } break;
+        case LLM_ARCH_MINICPM:
+            {
+                result = llm.build_minicpm();
+            } break;
+        case LLM_ARCH_GEMMA:
+            {
+                result = llm.build_gemma();
+            } break;
+        case LLM_ARCH_GEMMA2:
+            {
+                result = llm.build_gemma2();
+            } break;
+        case LLM_ARCH_STARCODER2:
+            {
+                result = llm.build_starcoder2();
+            } break;
+        case LLM_ARCH_MAMBA:
+            {
+                result = llm.build_mamba();
+            } break;
+        case LLM_ARCH_XVERSE:
+            {
+                result = llm.build_xverse();
+            } break;
+        case LLM_ARCH_COMMAND_R:
+            {
+                result = llm.build_command_r();
+            } break;
+        case LLM_ARCH_DBRX:
+            {
+                result = llm.build_dbrx();
+            } break;
+        case LLM_ARCH_OLMO:
+            {
+                result = llm.build_olmo();
+            } break;
+        case LLM_ARCH_OPENELM:
+            {
+                result = llm.build_openelm();
+            } break;
+        case LLM_ARCH_GPTNEOX:
+            {
+                result = llm.build_gptneox();
+            } break;
+        case LLM_ARCH_ARCTIC:
+            {
+                result = llm.build_arctic();
+            } break;
+        case LLM_ARCH_DEEPSEEK2:
+            {
+                result = llm.build_deepseek2();
+            } break;
+        case LLM_ARCH_CHATGLM:
+            {
+                result = llm.build_chatglm();
+            } break;
+        case LLM_ARCH_BITNET:
+            {
+                result = llm.build_bitnet();
+            } break;
+        case LLM_ARCH_T5:
+            {
+                result = llm.build_t5();
+            } break;
+        case LLM_ARCH_JAIS:
+            {
+                result = llm.build_jais();
+            } break;
+        default:
+            GGML_ASSERT(false);
+    }
+
+    // add on pooling layer
+    if (lctx.cparams.embeddings) {
+        result = llm.append_pooling(result);
+    }
+
+    llm.free();
+
+    return result;
+}
+
+static void llama_set_k_shift(llama_context & lctx) {
+    const int64_t kv_size = lctx.kv_self.size;
+
+    assert(ggml_backend_buffer_is_host(lctx.inp_K_shift->buffer));
+
+    int32_t * data = (int32_t *) lctx.inp_K_shift->data;
+
+    for (int i = 0; i < kv_size; ++i) {
+        data[i] = lctx.kv_self.cells[i].delta;
+    }
+}
+
+static void llama_set_s_copy(llama_context & lctx) {
+    const int64_t kv_size = lctx.kv_self.size;
+
+    assert(ggml_backend_buffer_is_host(lctx.inp_s_copy->buffer));
+
+    int32_t * data = (int32_t *) lctx.inp_s_copy->data;
+
+    for (int i = 0; i < kv_size; ++i) {
+        data[i] = lctx.kv_self.cells[i].src;
+    }
+}
+
+static int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional) {
+    // TODO move to hparams if a T5 variant appears that uses a different value
+    const int64_t max_distance = 128;
+
+    if (bidirectional) {
+        n_buckets >>= 1;
+    }
+
+    const int64_t max_exact = n_buckets >> 1;
+
+    int32_t relative_position = x - y;
+    int32_t relative_bucket = 0;
+    if (bidirectional) {
+        relative_bucket += (relative_position > 0) * n_buckets;
+        relative_position = abs(relative_position);
+    } else {
+        relative_position = -std::min(relative_position, 0);
+    }
+    int32_t relative_position_if_large = floorf(max_exact + logf(1.0 * relative_position / max_exact) * (n_buckets - max_exact) / log(1.0 * max_distance / max_exact));
+    relative_position_if_large = std::min(relative_position_if_large, n_buckets - 1);
+    relative_bucket += (relative_position < max_exact ? relative_position : relative_position_if_large);
+    return relative_bucket;
+}
+
+static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
+    //
+    // set input data
+    //
+
+    const auto & hparams = lctx.model.hparams;
+    const auto & cparams = lctx.cparams;
+    const auto & kv_self = lctx.kv_self;
+
+    if (batch.token) {
+        const int64_t n_tokens = batch.n_tokens;
+
+        ggml_backend_tensor_set(lctx.inp_tokens, batch.token, 0, n_tokens*ggml_element_size(lctx.inp_tokens));
+    }
+
+    if (batch.embd) {
+        const int64_t n_embd   = hparams.n_embd;
+        const int64_t n_tokens = batch.n_tokens;
+
+        ggml_backend_tensor_set(lctx.inp_embd, batch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd));
+    }
+
+    if (batch.pos && lctx.inp_pos) {
+        const int64_t n_tokens = batch.n_tokens;
+
+        ggml_backend_tensor_set(lctx.inp_pos, batch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos));
+    }
+
+    if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) {
+        GGML_ASSERT(lctx.inp_out_ids && "every model that can must skip unused outputs");
+        const int64_t n_tokens = batch.n_tokens;
+
+        GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_out_ids->buffer));
+        int32_t * data = (int32_t *) lctx.inp_out_ids->data;
+
+        if (lctx.n_outputs == n_tokens) {
+            for (int i = 0; i < n_tokens; ++i) {
+                data[i] = i;
+            }
+        } else if (batch.logits) {
+            int32_t n_outputs = 0;
+            for (int i = 0; i < n_tokens; ++i) {
+                if (batch.logits[i]) {
+                    data[n_outputs++] = i;
+                }
+            }
+            // the graph needs to have been passed the correct number of outputs
+            GGML_ASSERT(lctx.n_outputs == n_outputs);
+        } else if (lctx.n_outputs == 1) {
+            // only keep last output
+            data[0] = n_tokens - 1;
+        } else {
+            GGML_ASSERT(lctx.n_outputs == 0);
+        }
+    }
+
+    GGML_ASSERT(
+        // (!a || b) is a logical implication (a -> b)
+        // !hparams.causal_attn -> !cparams.causal_attn
+        (hparams.causal_attn || !cparams.causal_attn) &&
+        "causal attention is not supported by this model"
+    );
+
+    if (lctx.inp_KQ_mask || lctx.inp_KQ_mask_swa) {
+        // NOTE: hparams.causal_attn indicates the model is capable of generation and uses the kv cache.
+        if (cparams.causal_attn && !lctx.is_encoding) {
+            const int64_t n_kv     = kv_self.n;
+            const int64_t n_tokens = batch.n_tokens;
+
+
+            float * data     = nullptr;
+            float * data_swa = nullptr;
+
+            if (lctx.inp_KQ_mask) {
+                GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask->buffer));
+                data = (float *) lctx.inp_KQ_mask->data;
+            }
+
+            if (lctx.inp_KQ_mask_swa) {
+                GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask_swa->buffer));
+                data_swa = (float *) lctx.inp_KQ_mask_swa->data;
+            }
+
+            // For causal attention, use only the previous KV cells
+            // of the correct sequence for each token of the batch.
+            // It's assumed that if a token in the batch has multiple sequences, they are equivalent.
+            for (int h = 0; h < 1; ++h) {
+                for (int j = 0; j < n_tokens; ++j) {
+                    const llama_pos    pos    = batch.pos[j];
+                    const llama_seq_id seq_id = batch.seq_id[j][0];
+
+                    for (int i = 0; i < n_kv; ++i) {
+                        float f;
+                        if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) {
+                            f = -INFINITY;
+                        } else {
+                            if (hparams.use_alibi) {
+                                f = -std::abs(lctx.kv_self.cells[i].pos - pos);
+                            } else {
+                                f = 0.0f;
+                            }
+                        }
+
+                        if (data) {
+                            data[h*(n_kv*n_tokens) + j*n_kv + i] = f;
+                        }
+
+                        // may need to cut off old tokens for sliding window
+                        if (data_swa) {
+                            if (pos - lctx.kv_self.cells[i].pos >= (int32_t)hparams.n_swa) {
+                                f = -INFINITY;
+                            }
+                            data_swa[h*(n_kv*n_tokens) + j*n_kv + i] = f;
+                        }
+                    }
+                }
+
+                if (data) {
+                    for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
+                        for (int j = 0; j < n_kv; ++j) {
+                            data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY;
+                        }
+                    }
+                }
+
+                if (data_swa) {
+                    for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
+                        for (int j = 0; j < n_kv; ++j) {
+                            data_swa[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY;
+                        }
+                    }
+                }
+            }
+        } else {
+            // when using kv cache, the mask needs to match the kv cache size
+            const int64_t n_tokens = batch.n_tokens;
+            const int64_t n_stride = hparams.causal_attn && !lctx.is_encoding ? kv_self.n : n_tokens;
+
+            GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask->buffer));
+
+            float * data = (float *) lctx.inp_KQ_mask->data;
+
+            for (int h = 0; h < 1; ++h) {
+                for (int j = 0; j < n_tokens; ++j) {
+                    const llama_seq_id seq_id = batch.seq_id[j][0];
+
+                    for (int i = 0; i < n_tokens; ++i) {
+                        float f = -INFINITY;
+                        for (int s = 0; s < batch.n_seq_id[i]; ++s) {
+                            if (batch.seq_id[i][s] == seq_id) {
+                                if (hparams.use_alibi) {
+                                    f = -std::abs(batch.pos[i] - batch.pos[j]);
+                                } else {
+                                    f = 0.0f;
+                                }
+                                break;
+                            }
+                        }
+
+                        data[h*(n_tokens*n_tokens) + j*n_stride + i] = f;
+                    }
+
+                    for (int i = n_tokens; i < n_stride; ++i) {
+                        data[h*(n_tokens*n_tokens) + j*n_stride + i] = -INFINITY;
+                    }
+                }
+            }
+        }
+    }
+
+    if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) {
+        const int64_t n_tokens = batch.n_tokens;
+
+        GGML_ASSERT(lctx.inp_mean);
+        GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_mean->buffer));
+
+        float * data = (float *) lctx.inp_mean->data;
+        memset(lctx.inp_mean->data, 0, n_tokens * n_tokens * ggml_element_size(lctx.inp_mean));
+
+        std::vector sum(n_tokens, 0);
+        for (int i = 0; i < n_tokens; ++i) {
+            const llama_seq_id seq_id = batch.seq_id[i][0];
+
+            GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN");
+
+            sum[seq_id] += 1;
+        }
+
+        std::vector div(n_tokens, 0.0f);
+        for (int i = 0; i < n_tokens; ++i) {
+            const uint64_t s = sum[i];
+            if (s > 0) {
+                div[i] = 1.0f/float(s);
+            }
+        }
+
+        for (int i = 0; i < n_tokens; ++i) {
+            const llama_seq_id seq_id = batch.seq_id[i][0];
+            data[seq_id*n_tokens + i] = div[seq_id];
+        }
+    }
+
+    if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_CLS) {
+        const int64_t n_tokens = batch.n_tokens;
+
+        GGML_ASSERT(lctx.inp_cls);
+        GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer));
+
+        uint32_t * data = (uint32_t *) lctx.inp_cls->data;
+        memset(lctx.inp_cls->data, 0, n_tokens * ggml_element_size(lctx.inp_cls));
+
+        for (int i = 0; i < n_tokens; ++i) {
+            const llama_seq_id seq_id = batch.seq_id[i][0];
+            const llama_pos    pos    = batch.pos[i];
+
+            GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS");
+
+            if (pos == 0) {
+                data[seq_id] = i;
+            }
+        }
+    }
+
+    if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_LAST) {
+        const int64_t n_tokens = batch.n_tokens;
+
+        GGML_ASSERT(lctx.inp_cls);
+        GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer));
+
+        uint32_t * data = (uint32_t *) lctx.inp_cls->data;
+        memset(lctx.inp_cls->data, 0, n_tokens * ggml_element_size(lctx.inp_cls));
+
+        std::vector last_pos(n_tokens, -1);
+        std::vector last_row(n_tokens, -1);
+
+        for (int i = 0; i < n_tokens; ++i) {
+            const llama_seq_id seq_id = batch.seq_id[i][0];
+            const llama_pos    pos    = batch.pos[i];
+
+            GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == LAST");
+
+            if (pos >= last_pos[seq_id]) {
+                last_pos[seq_id] = pos;
+                last_row[seq_id] = i;
+            }
+        }
+
+        for (int i = 0; i < n_tokens; ++i) {
+            if (last_row[i] >= 0) {
+                data[i] = last_row[i];
+            }
+        }
+    }
+
+    if (kv_self.recurrent) {
+        const int64_t n_kv = kv_self.n;
+
+        if (lctx.inp_s_mask) {
+            GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_s_mask->buffer));
+            float * data = (float *) lctx.inp_s_mask->data;
+
+            // states which are not affected by the current batch are left untouched
+            for (int i = 0; i < n_kv; ++i) {
+                llama_seq_id    seq_id       = i + lctx.kv_self.head;
+                llama_kv_cell & kv_cell      = lctx.kv_self.cells[seq_id];
+                bool            has_self_seq = kv_cell.has_seq_id(seq_id);
+
+                data[i] = (float) has_self_seq;
+
+                // ensure current sequences will be kept
+                if (!has_self_seq && kv_cell.pos >= 0) {
+                    kv_cell.seq_id.insert(seq_id);
+                }
+            }
+        }
+        // For Mamba (and other recurrent architectures),
+        // update the correct state(s)/sequence(s) for each token of the batch.
+        // Like with the KQ_mask, if a token in the batch has multiple sequences,
+        // they are assumed to be equivalent (not here, but in ggml_ssm_scan and ggml_ssm_conv).
+        if (lctx.inp_s_seq) {
+            const int64_t n_tokens = batch.n_tokens;
+
+            GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_s_seq->buffer));
+            int32_t * data = (int32_t *) lctx.inp_s_seq->data;
+
+            for (int j = 0; j < n_tokens; ++j) {
+                const int32_t n_seq = batch.n_seq_id[j];
+                GGML_ASSERT(0 < n_seq); // a token should be part of at least 1 sequence
+
+                for (int i = 0; i < n_kv; ++i) {
+                    if (i < n_seq) {
+                        // for this type of model, the head is the minimum seq_id of the batch
+                        data[j*n_kv + i] = batch.seq_id[j][i] - kv_self.head;
+                    } else {
+                        data[j*n_kv + i] = -1;
+                    }
+                }
+            }
+        }
+    }
+
+    if (lctx.inp_pos_bucket) {
+        const int64_t n_tokens = batch.n_tokens;
+
+        GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_pos_bucket->buffer));
+
+        int32_t * data = (int32_t *) lctx.inp_pos_bucket->data;
+
+        if (!lctx.is_encoding) {
+            const int64_t n_kv = kv_self.n;
+            for (int h = 0; h < 1; ++h) {
+                for (int j = 0; j < n_tokens; ++j) {
+                    for (int i = 0; i < n_kv; ++i) {
+                        data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(lctx.kv_self.cells[i].pos, batch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding);
+                    }
+                }
+            }
+        } else {
+            for (int h = 0; h < 1; ++h) {
+                for (int j = 0; j < n_tokens; ++j) {
+                    for (int i = 0; i < n_tokens; ++i) {
+                        data[h*(n_tokens*n_tokens) + j*n_tokens + i] = llama_relative_position_bucket(batch.pos[i], batch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding);
+                    }
+                }
+            }
+        }
+    }
+
+    if (!lctx.is_encoding && lctx.inp_embd_enc) {
+        assert(lctx.inp_embd_enc->type == GGML_TYPE_F32);
+        assert((size_t) ggml_nelements(lctx.inp_embd_enc) == lctx.embd_enc.size());
+
+        ggml_backend_tensor_set(lctx.inp_embd_enc, lctx.embd_enc.data(), 0, ggml_nbytes(lctx.inp_embd_enc));
+    }
+
+    if (!lctx.is_encoding && lctx.inp_KQ_mask_cross) {
+        const int64_t n_output_enc = lctx.embd_enc.size() / hparams.n_embd;
+        const int64_t n_tokens = batch.n_tokens;
+
+        GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask_cross->buffer));
+
+        float * data = (float *) lctx.inp_KQ_mask_cross->data;
+
+        for (int h = 0; h < 1; ++h) {
+            for (int j = 0; j < n_tokens; ++j) {
+                for (int i = 0; i < n_output_enc; ++i) {
+                    float f = -INFINITY;
+                    for (int s = 0; s < batch.n_seq_id[j]; ++s) {
+                        const llama_seq_id seq_id = batch.seq_id[j][s];
+                        if (lctx.seq_ids_enc[i].find(seq_id) != lctx.seq_ids_enc[i].end()) {
+                            f = 0.0f;
+                        }
+                    }
+                    data[h*(n_output_enc*n_tokens) + j*n_output_enc + i] = f;
+                }
+            }
+
+            for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
+                for (int j = 0; j < n_output_enc; ++j) {
+                    data[h*(n_output_enc*n_tokens) + i*n_output_enc + j] = -INFINITY;
+                }
+            }
+        }
+    }
+}
+
+// Make sure enough space is available for outputs.
+// Returns max number of outputs for which space was reserved.
+static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
+    const auto & cparams = lctx.cparams;
+    const auto & hparams = lctx.model.hparams;
+
+    const size_t n_outputs_max = std::max(n_outputs, (size_t) cparams.n_seq_max);
+
+    const auto n_batch = cparams.n_batch;
+    const auto n_vocab = hparams.n_vocab;
+    const auto n_embd  = hparams.n_embd;
+
+    // TODO: use a per-batch flag for logits presence instead
+    const bool has_logits = !cparams.embeddings;
+    const bool has_embd   =  lctx.is_encoding || (cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE));
+
+    const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0;
+    const size_t embd_size   = has_embd   ?  n_embd*n_outputs_max : 0;
+
+    if (lctx.output_ids.empty()) {
+        // init, never resized afterwards
+        lctx.output_ids.resize(n_batch);
+    }
+
+    const size_t prev_size = lctx.buf_output ? ggml_backend_buffer_get_size(lctx.buf_output) : 0;
+    const size_t new_size  = (logits_size + embd_size) * sizeof(float);
+
+    // alloc only when more than the current capacity is required
+    // TODO: also consider shrinking the buffer
+    if (!lctx.buf_output || prev_size < new_size) {
+        if (lctx.buf_output) {
+#ifndef NDEBUG
+            // This doesn't happen often, but may be annoying in some cases (like the HellaSwag benchmark)
+            LLAMA_LOG_INFO("%s: reallocating output buffer from size %.02f MiB to %.02f MiB\n", __func__, prev_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+#endif
+            ggml_backend_buffer_free(lctx.buf_output);
+            lctx.buf_output = nullptr;
+            lctx.logits = nullptr;
+            lctx.embd = nullptr;
+        }
+
+        lctx.buf_output = ggml_backend_buft_alloc_buffer(llama_default_buffer_type_cpu(true), new_size);
+        if (lctx.buf_output == nullptr) {
+            LLAMA_LOG_ERROR("%s: failed to allocate output buffer of size %.2f MiB\n", __func__, new_size / (1024.0 * 1024.0));
+            return 0;
+        }
+    }
+
+    float * output_base = (float *) ggml_backend_buffer_get_base(lctx.buf_output);
+
+    lctx.logits = has_logits ? output_base               : nullptr;
+    lctx.embd   = has_embd   ? output_base + logits_size : nullptr;
+
+    lctx.output_size = n_outputs_max;
+    lctx.logits_size = logits_size;
+    lctx.embd_size   = embd_size;
+
+    // set all ids as invalid (negative)
+    std::fill(lctx.output_ids.begin(), lctx.output_ids.end(), -1);
+
+    ggml_backend_buffer_clear(lctx.buf_output, 0);
+
+    lctx.n_outputs = 0;
+
+    return n_outputs_max;
+}
+
+
+static void llama_graph_compute(
+        llama_context & lctx,
+          ggml_cgraph * gf,
+                  int   n_threads) {
+#ifdef GGML_USE_METAL
+    if (ggml_backend_is_metal(lctx.backend_metal)) {
+        ggml_backend_metal_set_n_cb(lctx.backend_metal, n_threads);
+    }
+#endif
+
+    if (lctx.backend_cpu != nullptr) {
+        ggml_backend_cpu_set_n_threads(lctx.backend_cpu, n_threads);
+        ggml_backend_cpu_set_abort_callback(lctx.backend_cpu, lctx.abort_callback, lctx.abort_callback_data);
+    }
+#ifdef GGML_USE_BLAS
+    if (lctx.backend_blas != nullptr) {
+        ggml_backend_blas_set_n_threads(lctx.backend_blas, n_threads);
+    }
+#endif
+
+    ggml_backend_sched_graph_compute_async(lctx.sched, gf);
+
+    // fprintf(stderr, "splits: %d\n", ggml_backend_sched_get_n_splits(lctx.sched));
+}
+
+// decode a batch of tokens by evaluating the transformer
+//
+//   - lctx:      llama context
+//   - batch:     batch to evaluate
+//
+// return 0 on success
+// return positive int on warning
+// return negative int on error
+//
+static int llama_decode_internal(
+         llama_context & lctx,
+           llama_batch   batch_all) { // TODO: rename back to batch
+
+    lctx.is_encoding = false;
+    const uint32_t n_tokens_all = batch_all.n_tokens;
+
+    if (n_tokens_all == 0) {
+        LLAMA_LOG_ERROR("%s: n_tokens == 0", __func__);
+        return -1;
+    }
+
+    const auto & model   = lctx.model;
+    const auto & hparams = model.hparams;
+    const auto & cparams = lctx.cparams;
+
+    GGML_ASSERT((!batch_all.token && batch_all.embd) || (batch_all.token && !batch_all.embd)); // NOLINT
+
+    GGML_ASSERT(n_tokens_all <= cparams.n_batch);
+
+    GGML_ASSERT((cparams.causal_attn || cparams.n_ubatch >= n_tokens_all) && "non-causal attention requires n_ubatch >= n_tokens");
+
+    if (lctx.t_compute_start_us == 0) {
+        lctx.t_compute_start_us = ggml_time_us();
+    }
+    lctx.n_queued_tokens += n_tokens_all;
+
+    auto & kv_self = lctx.kv_self;
+
+    const int64_t n_embd  = hparams.n_embd;
+    const int64_t n_vocab = hparams.n_vocab;
+
+    uint32_t n_outputs = 0;
+    uint32_t n_outputs_prev = 0;
+
+    const auto n_ubatch = cparams.n_ubatch;
+
+    // TODO: simplify or deprecate
+    std::vector pos;
+    std::vector                   n_seq_id;
+    std::vector            seq_id_arr;
+    std::vector> seq_id;
+
+    // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens
+    const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
+
+    // count outputs
+    if (batch_all.logits && !embd_pooled) {
+        for (uint32_t i = 0; i < n_tokens_all; ++i) {
+            n_outputs += batch_all.logits[i] != 0;
+        }
+    } else if (lctx.logits_all || embd_pooled) {
+        n_outputs = n_tokens_all;
+    } else {
+        // keep last output only
+        n_outputs = 1;
+    }
+
+    // reserve output buffer
+    if (llama_output_reserve(lctx, n_outputs) < n_outputs) {
+        LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_outputs);
+        return -2;
+    };
+
+    // set output mappings
+    if (batch_all.logits) {
+        int32_t i_logits = 0;
+        for (uint32_t i = 0; i < n_tokens_all; ++i) {
+            if (batch_all.logits[i]) {
+                lctx.output_ids[i] = i_logits++;
+            }
+        }
+    } else {
+        for (uint32_t i = 0; i < n_outputs; ++i) {
+            lctx.output_ids[i] = i;
+        }
+    }
+
+    for (uint32_t cur_token = 0; cur_token < n_tokens_all; cur_token += n_ubatch) {
+        const uint32_t n_tokens = std::min(n_ubatch, n_tokens_all - cur_token);
+        llama_batch u_batch = {
+            /* .n_tokens   = */ (int32_t) n_tokens,
+            /* .token      = */ batch_all.token     ? batch_all.token    + cur_token        : nullptr,
+            /* .embd       = */ batch_all.embd      ? batch_all.embd     + cur_token*n_embd : nullptr,
+            /* .pos        = */ batch_all.pos       ? batch_all.pos      + cur_token        : nullptr,
+            /* .n_seq_id   = */ batch_all.n_seq_id  ? batch_all.n_seq_id + cur_token        : nullptr,
+            /* .seq_id     = */ batch_all.seq_id    ? batch_all.seq_id   + cur_token        : nullptr,
+            /* .logits     = */ batch_all.logits    ? batch_all.logits   + cur_token        : nullptr,
+            /* .all_pos_0  = */ batch_all.all_pos_0 + (llama_pos) cur_token*batch_all.all_pos_1,
+            /* .all_pos_1  = */ batch_all.all_pos_1,
+            /* .all_seq_id = */ batch_all.all_seq_id,
+        };
+
+        // count the outputs in this u_batch
+        {
+            int32_t n_outputs_new = 0;
+
+            if (u_batch.logits && !embd_pooled) {
+                for (uint32_t i = 0; i < n_tokens; i++) {
+                    n_outputs_new += u_batch.logits[i] != 0;
+                }
+            } else if (n_outputs == n_tokens_all) {
+                n_outputs_new = n_tokens;
+            } else {
+                // keep last output only
+                if (cur_token + n_tokens >= n_tokens_all) {
+                    n_outputs_new = 1;
+                }
+            }
+
+            // needs to happen before the graph is built
+            lctx.n_outputs = n_outputs_new;
+        }
+
+        int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch;
+        GGML_ASSERT(n_threads > 0);
+
+        // helpers for smoother batch API transition
+        // after deprecating the llama_eval calls, these will be removed
+        if (u_batch.pos == nullptr) {
+            pos.resize(n_tokens);
+            for (uint32_t i = 0; i < n_tokens; i++) {
+                pos[i] = u_batch.all_pos_0 + i*u_batch.all_pos_1;
+            }
+
+            u_batch.pos = pos.data();
+        }
+
+        if (u_batch.seq_id == nullptr) {
+            n_seq_id.resize(n_tokens);
+            seq_id.resize(n_tokens);
+            seq_id_arr.resize(n_tokens);
+            for (uint32_t i = 0; i < n_tokens; i++) {
+                n_seq_id[i] = 1;
+                seq_id[i].resize(1);
+                seq_id[i][0] = u_batch.all_seq_id;
+                seq_id_arr[i] = seq_id[i].data();
+            }
+
+            u_batch.n_seq_id = n_seq_id.data();
+            u_batch.seq_id = seq_id_arr.data();
+        }
+
+        // non-causal masks do not use the KV cache
+        if (hparams.causal_attn) {
+            llama_kv_cache_update(&lctx);
+
+            // if we have enough unused cells before the current head ->
+            //   better to start searching from the beginning of the cache, hoping to fill it
+            if (kv_self.head > kv_self.used + 2*n_tokens) {
+                kv_self.head = 0;
+            }
+
+            if (!llama_kv_cache_find_slot(kv_self, u_batch)) {
+                return 1;
+            }
+
+            if (!kv_self.recurrent) {
+                // a heuristic, to avoid attending the full cache if it is not yet utilized
+                // after enough generations, the benefit from this heuristic disappears
+                // if we start defragmenting the cache, the benefit from this will be more important
+                const uint32_t pad = llama_kv_cache_get_padding(cparams);
+                kv_self.n = std::min(kv_self.size, std::max(pad, GGML_PAD(llama_kv_cache_cell_max(kv_self), pad)));
+                //kv_self.n = llama_kv_cache_cell_max(kv_self);
+            }
+        }
+
+        //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head);
+
+        ggml_backend_sched_reset(lctx.sched);
+        ggml_backend_sched_set_eval_callback(lctx.sched, lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data);
+
+        ggml_cgraph * gf = llama_build_graph(lctx, u_batch, false);
+
+        // the output is always the last tensor in the graph
+        struct ggml_tensor * res  = gf->nodes[gf->n_nodes - 1];
+        struct ggml_tensor * embd = gf->nodes[gf->n_nodes - 2];
+
+        if (lctx.n_outputs == 0) {
+            // no output
+            res  = nullptr;
+            embd = nullptr;
+        } else if (cparams.embeddings) {
+            res = nullptr; // do not extract logits for embedding case
+            embd = gf->nodes[gf->n_nodes - 1];
+            if (strcmp(embd->name, "result_embd_pooled") != 0) {
+                embd = gf->nodes[gf->n_nodes - 2];
+            }
+            GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0 && "missing embeddings tensor");
+        } else {
+            embd = nullptr; // do not extract embeddings when not needed
+            GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor");
+        }
+        // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
+
+        ggml_backend_sched_alloc_graph(lctx.sched, gf);
+
+        llama_set_inputs(lctx, u_batch);
+
+        llama_graph_compute(lctx, gf, n_threads);
+
+        // update the kv ring buffer
+        {
+            kv_self.head += n_tokens;
+
+            // Ensure kv cache head points to a valid index.
+            if (kv_self.head >= kv_self.size) {
+                kv_self.head = 0;
+            }
+        }
+
+        // plot the computation graph in dot format (for debugging purposes)
+        //if (n_past%100 == 0) {
+        //    ggml_graph_dump_dot(gf, NULL, "llama.dot");
+        //}
+
+        // extract logits
+        if (res) {
+            ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(lctx.sched, res);
+            GGML_ASSERT(backend_res != nullptr);
+            GGML_ASSERT(lctx.logits != nullptr);
+
+            float * logits_out = lctx.logits + n_outputs_prev*n_vocab;
+            const int32_t n_outputs_new = lctx.n_outputs;
+
+            if (n_outputs_new) {
+                GGML_ASSERT( n_outputs_prev + n_outputs_new <= n_outputs);
+                GGML_ASSERT((n_outputs_prev + n_outputs_new)*n_vocab <= (int64_t) lctx.logits_size);
+                ggml_backend_tensor_get_async(backend_res, res, logits_out, 0, n_outputs_new*n_vocab*sizeof(float));
+            }
+        }
+
+        // extract embeddings
+        if (embd) {
+            ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched, embd);
+            GGML_ASSERT(backend_embd != nullptr);
+
+            switch (cparams.pooling_type) {
+                case LLAMA_POOLING_TYPE_NONE:
+                    {
+                        // extract token embeddings
+                        GGML_ASSERT(lctx.embd != nullptr);
+                        float * embd_out = lctx.embd + n_outputs_prev*n_embd;
+                        const int32_t n_outputs_new = lctx.n_outputs;
+
+                        if (n_outputs_new) {
+                            GGML_ASSERT( n_outputs_prev + n_outputs_new <= n_outputs);
+                            GGML_ASSERT((n_outputs_prev + n_outputs_new)*n_embd <= (int64_t) lctx.embd_size);
+                            ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_outputs_new*n_embd*sizeof(float));
+                        }
+                    } break;
+                case LLAMA_POOLING_TYPE_MEAN:
+                case LLAMA_POOLING_TYPE_CLS:
+                case LLAMA_POOLING_TYPE_LAST:
+                    {
+                        // extract sequence embeddings
+                        auto & embd_seq_out = lctx.embd_seq;
+                        embd_seq_out.clear();
+
+                        for (uint32_t i = 0; i < n_tokens; i++) {
+                            const llama_seq_id seq_id = u_batch.seq_id[i][0];
+                            if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
+                                continue;
+                            }
+                            embd_seq_out[seq_id].resize(n_embd);
+                            ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float));
+                        }
+                    } break;
+                case LLAMA_POOLING_TYPE_UNSPECIFIED:
+                    {
+                        GGML_ASSERT(false && "unknown pooling type");
+                    } break;
+            }
+        }
+        n_outputs_prev += lctx.n_outputs;
+    }
+
+    // set to total number of outputs in the batch, for use in llama_get_logits_ith
+    lctx.n_outputs = n_outputs;
+
+    // wait for the computation to finish (automatically done when obtaining the model output)
+    //llama_synchronize(&lctx);
+
+    // decide if we need to defrag the kv cache
+    if (cparams.causal_attn && cparams.defrag_thold >= 0.0f) {
+        const float fragmentation = kv_self.n >= 128 ? 1.0f - float(kv_self.used)/float(kv_self.n) : 0.0f;
+
+        // queue defragmentation for next llama_kv_cache_update
+        if (fragmentation > cparams.defrag_thold) {
+            //LLAMA_LOG_INFO("fragmentation: %.2f\n", fragmentation);
+
+            llama_kv_cache_defrag(kv_self);
+        }
+    }
+
+    // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
+    // overlap with device computation.
+    ggml_backend_sched_reset(lctx.sched);
+
+    return 0;
+}
+
+// encode a batch of tokens by evaluating the encoder part of the transformer
+//
+//   - lctx:      llama context
+//   - batch:     batch to evaluate
+//
+// return 0 on success
+// return positive int on warning
+// return negative int on error
+//
+static int llama_encode_internal(
+         llama_context & lctx,
+           llama_batch   batch) {
+
+    lctx.is_encoding = true;
+
+    const uint32_t n_tokens = batch.n_tokens;
+
+    if (n_tokens == 0) {
+        LLAMA_LOG_ERROR("%s: n_tokens == 0", __func__);
+        return -1;
+    }
+
+    const auto & model   = lctx.model;
+    const auto & hparams = model.hparams;
+    const auto & cparams = lctx.cparams;
+
+    GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
+
+    // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot
+    GGML_ASSERT(cparams.n_ubatch >= n_tokens && "encoder requires n_ubatch >= n_tokens");
+
+    if (lctx.t_compute_start_us == 0) {
+        lctx.t_compute_start_us = ggml_time_us();
+    }
+
+    lctx.n_queued_tokens += n_tokens;
+
+    const int64_t n_embd = hparams.n_embd;
+
+    // TODO: simplify or deprecate
+    std::vector pos;
+    std::vector                   n_seq_id;
+    std::vector            seq_id_arr;
+    std::vector> seq_id;
+
+    // reserve output buffer
+    if (llama_output_reserve(lctx, n_tokens) < n_tokens) {
+        LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_tokens);
+        return -2;
+    };
+
+    for (uint32_t i = 0; i < n_tokens; ++i) {
+        lctx.output_ids[i] = i;
+    }
+
+    lctx.inp_embd_enc = NULL;
+    lctx.n_outputs = n_tokens;
+
+    const int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch;
+    GGML_ASSERT(n_threads > 0);
+
+    // helpers for smoother batch API transition
+    // after deprecating the llama_eval calls, these will be removed
+    if (batch.pos == nullptr) {
+        pos.resize(n_tokens);
+        for (uint32_t i = 0; i < n_tokens; i++) {
+            pos[i] = batch.all_pos_0 + i*batch.all_pos_1;
+        }
+
+        batch.pos = pos.data();
+    }
+
+    if (batch.seq_id == nullptr) {
+        n_seq_id.resize(n_tokens);
+        seq_id.resize(n_tokens);
+        seq_id_arr.resize(n_tokens);
+        for (uint32_t i = 0; i < n_tokens; i++) {
+            n_seq_id[i] = 1;
+            seq_id[i].resize(1);
+            seq_id[i][0] = batch.all_seq_id;
+            seq_id_arr[i] = seq_id[i].data();
+        }
+
+        batch.n_seq_id = n_seq_id.data();
+        batch.seq_id = seq_id_arr.data();
+    }
+
+    ggml_backend_sched_reset(lctx.sched);
+    ggml_backend_sched_set_eval_callback(lctx.sched, lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data);
+
+    ggml_cgraph * gf = llama_build_graph(lctx, batch, false);
+
+    // the output embeddings after the final encoder normalization
+    struct ggml_tensor * embd = gf->nodes[gf->n_nodes - 1];
+
+    GGML_ASSERT(strcmp(embd->name, "result_norm") == 0);
+
+    ggml_backend_sched_alloc_graph(lctx.sched, gf);
+
+    llama_set_inputs(lctx, batch);
+
+    llama_graph_compute(lctx, gf, n_threads);
+
+    // extract embeddings
+    if (embd) {
+        ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched, embd);
+        GGML_ASSERT(backend_embd != nullptr);
+
+        // extract token embeddings
+        GGML_ASSERT(lctx.embd != nullptr);
+
+        lctx.embd_enc.resize(n_tokens*n_embd);
+        float * embd_out = lctx.embd_enc.data();
+
+        ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_tokens*n_embd*sizeof(float));
+
+        // remember the sequence ids used during the encoding - needed for cross attention later
+        lctx.seq_ids_enc.resize(n_tokens);
+        for (uint32_t i = 0; i < n_tokens; i++) {
+            for (int s = 0; s < batch.n_seq_id[i]; s++) {
+                llama_seq_id seq_id = batch.seq_id[i][s];
+                lctx.seq_ids_enc[i].insert(seq_id);
+            }
+        }
+    }
+
+    // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
+    // overlap with device computation.
+    ggml_backend_sched_reset(lctx.sched);
+
+    return 0;
+}
+
+// find holes from the beginning of the KV cache and fill them by moving data from the end of the cache
+static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
+    auto & kv_self = lctx.kv_self;
+
+    const auto & hparams = lctx.model.hparams;
+
+    const uint32_t n_layer = hparams.n_layer;
+
+    const uint32_t n_kv   = llama_kv_cache_cell_max(kv_self);
+    const uint32_t n_used = kv_self.used;
+
+    assert(n_used <= n_kv);
+
+    //const int64_t t_start = ggml_time_us();
+
+    // number of cells moved
+    uint32_t n_moves = 0;
+
+    // each move requires 6*n_layer tensors (see build_defrag)
+    //   - source view, destination view, copy operation
+    //   - x2 for keys and values
+    //const uint32_t max_moves = LLAMA_MAX_NODES/(6*n_layer);
+    // TODO: tmp fix https://github.com/ggerganov/llama.cpp/issues/6685#issuecomment-2057579516
+    const uint32_t max_moves = (LLAMA_MAX_NODES - 2*n_layer)/(6*n_layer);
+
+    // determine which KV cells to move where
+    //
+    //  cell i moves to ids[i]
+    //
+    //  if ids[i] == i || ids[i] == n_kv, then cell i is not moved
+    //
+    std::vector ids(n_kv, n_kv);
+
+    for (uint32_t i0 = 0; i0 < n_used; ++i0) {
+        const auto & cell0 = kv_self.cells[i0];
+
+        if (!cell0.is_empty()) {
+            ids[i0] = i0;
+
+            continue;
+        }
+
+        // found a hole - fill it with data from the end of the cache
+
+        uint32_t nh = 1;
+
+        // determine the size of the hole
+        while (i0 + nh < n_used && kv_self.cells[i0 + nh].is_empty()) {
+            nh++;
+        }
+
+        uint32_t nf = 0;
+        uint32_t is = n_kv - 1;
+
+        // starting from the end, find nh non-empty cells
+        for (; is > i0; --is) {
+            const auto & cell1 = kv_self.cells[is];
+
+            if (cell1.is_empty() || ids[is] != n_kv) {
+                continue;
+            }
+
+            // non-empty cell which is not yet moved
+            nf++;
+
+            if (nf == nh) {
+                break;
+            }
+        }
+
+        // this can only happen if `n_used` is not accurate, which would be a bug
+        GGML_ASSERT(nf == nh && "KV defrag bug: nf != nh");
+
+        nf = 0;
+
+        uint32_t i1 = is;
+
+        // are we moving a continuous block of memory?
+        bool cont = false;
+
+        // should we stop searching for the next move?
+        bool stop = false;
+
+        // go back and move the nf cells to the hole
+        for (; i1 < n_kv; ++i1) {
+            auto & cell1 = kv_self.cells[i1];
+
+            if (cell1.is_empty() || ids[i1] != n_kv) {
+                if (n_moves == max_moves) {
+                    stop = true;
+                    break;
+                }
+
+                cont = false;
+                continue;
+            }
+
+            // this cell goes to (i0 + nf)
+            ids[i1] = i0 + nf;
+
+            // move the cell meta data
+            kv_self.cells[i0 + nf] = cell1;
+
+            // clear the old cell and move the head there
+            cell1 = llama_kv_cell();
+            kv_self.head = n_used;
+
+            if (!cont) {
+                n_moves++;
+                cont = true;
+            }
+
+            nf++;
+
+            if (nf == nh) {
+                break;
+            }
+        }
+
+        if (stop || n_moves == max_moves) {
+            break;
+        }
+
+        //LLAMA_LOG_INFO("(tmp log) KV defrag: move [%u, %u) to [%u, %u)\n", is, i1 + 1, i0, i0 + nh);
+
+        i0 += nh - 1;
+    }
+
+    if (n_moves == 0) {
+        return;
+    }
+
+    //LLAMA_LOG_INFO("(tmp log) KV defrag cell moves: %u\n", n_moves);
+
+    //LLAMA_LOG_INFO("expected gf nodes: %u\n", 6*n_moves*n_layer);
+
+#if 0
+    // CPU defrag
+    //
+    // TODO: optimizations are possible:
+    //       - multiple threads
+    //       - avoid copying to the host memory when already there
+    //
+    // likely not worth the effort, as we have ggml_graph based defrag
+    //
+
+    const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa();
+    const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa();
+
+    const uint32_t kv_size = kv_self.size;
+
+    std::vector buf_k;
+    std::vector buf_v;
+
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        const size_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa);
+        const size_t k_size     = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_size);
+
+        const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type);
+        const size_t v_size    = ggml_row_size (kv_self.v_l[il]->type, n_embd_v_gqa*kv_size);
+
+        buf_k.resize(k_size);
+        buf_v.resize(v_size);
+
+        ggml_backend_tensor_get(kv_self.k_l[il], buf_k.data(), 0, buf_k.size());
+        ggml_backend_tensor_get(kv_self.v_l[il], buf_v.data(), 0, buf_v.size());
+
+        // batch move [i, i+nm) to [id, id+nm)
+        // note: cells can move only to a lower index
+        for (uint32_t i = 0; i < n_kv; ++i) {
+            const uint32_t id = ids[i];
+
+            if (i == id || id == n_kv) {
+                continue;
+            }
+
+            uint32_t nm = 1;
+
+            while (i + nm < n_kv && ids[i + nm] == id + nm) {
+                nm++;
+            }
+
+            // move keys
+            {
+                const int64_t os =  i*k_size_row;
+                const int64_t od = id*k_size_row;
+
+                memcpy(buf_k.data() + od, buf_k.data() + os, nm*k_size_row);
+            }
+
+            // move values (note: they are transposed)
+            {
+                const int64_t os =  i;
+                const int64_t od = id;
+
+                for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
+                    memcpy(buf_v.data() + (od + j*kv_size)*v_size_el, buf_v.data() + (os + j*kv_size)*v_size_el, nm*v_size_el);
+                }
+            }
+
+            i += nm - 1;
+        }
+
+        ggml_backend_tensor_set(kv_self.k_l[il], buf_k.data(), 0, buf_k.size());
+        ggml_backend_tensor_set(kv_self.v_l[il], buf_v.data(), 0, buf_v.size());
+    }
+#else
+    // ggml_graph defrag
+
+    ggml_backend_sched_reset(lctx.sched);
+
+    ggml_cgraph * gf = llama_build_graph_defrag(lctx, ids);
+
+    llama_graph_compute(lctx, gf, lctx.cparams.n_threads);
+#endif
+
+    //const int64_t t_end = ggml_time_us();
+
+    //LLAMA_LOG_INFO("(tmp log) KV defrag time: %.3f ms\n", (t_end - t_start)/1000.0);
+}
+
+static void llama_kv_cache_update_internal(struct llama_context & lctx) {
+    bool need_reserve = false;
+
+    // apply K-shift if needed
+    if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE && lctx.kv_self.has_shift) {
+        if (lctx.model.arch == LLM_ARCH_DEEPSEEK2) { // not supported due to MLA
+            GGML_ASSERT(false && "Deepseek2 does not support K-shift");
+        }
+
+        {
+            ggml_backend_sched_reset(lctx.sched);
+
+            ggml_cgraph * gf = llama_build_graph_k_shift(lctx);
+
+            ggml_backend_sched_alloc_graph(lctx.sched, gf);
+
+            llama_set_k_shift(lctx);
+
+            llama_graph_compute(lctx, gf, lctx.cparams.n_threads);
+
+            need_reserve = true;
+        }
+
+        {
+            auto & kv_self = lctx.kv_self;
+
+            kv_self.has_shift = false;
+
+            for (uint32_t i = 0; i < kv_self.size; ++i) {
+                kv_self.cells[i].delta = 0;
+            }
+        }
+    }
+
+    if (lctx.kv_self.recurrent && lctx.kv_self.do_copy) {
+        {
+            ggml_backend_sched_reset(lctx.sched);
+
+            ggml_cgraph * gf = llama_build_graph_s_copy(lctx);
+
+            ggml_backend_sched_alloc_graph(lctx.sched, gf);
+
+            llama_set_s_copy(lctx);
+
+            llama_graph_compute(lctx, gf, lctx.cparams.n_threads);
+
+            need_reserve = true;
+        }
+
+        {
+            auto & kv_self = lctx.kv_self;
+
+            kv_self.do_copy = false;
+
+            for (uint32_t i = 0; i < kv_self.size; ++i) {
+                kv_self.cells[i].src = i;
+            }
+        }
+    }
+
+    // defragment the KV cache if needed
+    if (lctx.kv_self.do_defrag) {
+        llama_kv_cache_defrag_internal(lctx);
+
+        need_reserve = true;
+
+        lctx.kv_self.do_defrag = false;
+    }
+
+    // reserve a worst case graph again
+    if (need_reserve) {
+        // TODO: extract to a function
+        // build worst-case graph
+        int n_tokens = (int)std::min(lctx.cparams.n_ctx, lctx.cparams.n_ubatch);
+        int n_past = lctx.cparams.n_ctx - n_tokens;
+        llama_token token = llama_token_bos(&lctx.model); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
+        ggml_cgraph * gf = llama_build_graph(lctx, llama_batch_get_one(&token, n_tokens, n_past, 0), true);
+
+        // initialize scheduler with the worst-case graph
+        ggml_backend_sched_reset(lctx.sched);
+        if (!ggml_backend_sched_reserve(lctx.sched, gf)) {
+            LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
+        }
+    }
+}
+
+//
+// quantization
+//
+
+struct quantize_state_internal {
+    const llama_model                 & model;
+    const llama_model_quantize_params * params;
+
+    int n_attention_wv    = 0;
+    int n_ffn_down        = 0;
+    int n_ffn_gate        = 0;
+    int n_ffn_up          = 0;
+    int i_attention_wv    = 0;
+    int i_ffn_down        = 0;
+    int i_ffn_gate        = 0;
+    int i_ffn_up          = 0;
+
+    int n_k_quantized     = 0;
+    int n_fallback        = 0;
+
+    bool has_imatrix      = false;
+
+    // used to figure out if a model shares tok_embd with the output weight
+    bool has_output       = false;
+
+    quantize_state_internal(const llama_model & model, const llama_model_quantize_params * params)
+        : model(model)
+        , params(params)
+        {}
+};
+
+static void llama_tensor_dequantize_internal(
+    struct ggml_tensor * tensor, std::vector> & output, std::vector & workers,
+    const size_t nelements, const int nthread
+) {
+    if (output.size() < nelements) {
+        output.resize(nelements);
+    }
+    float * f32_output = (float *) output.data();
+
+    ggml_type_traits_t qtype;
+    if (ggml_is_quantized(tensor->type)) {
+        qtype = ggml_internal_get_type_traits(tensor->type);
+        if (qtype.to_float == NULL) {
+            throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available", ggml_type_name(tensor->type)));
+        }
+    } else if (tensor->type != GGML_TYPE_F16 &&
+               tensor->type != GGML_TYPE_BF16) {
+        throw std::runtime_error(format("cannot dequantize/convert tensor type %s", ggml_type_name(tensor->type)));
+    }
+
+    if (nthread < 2) {
+        if (tensor->type == GGML_TYPE_F16) {
+            ggml_fp16_to_fp32_row((ggml_fp16_t *)tensor->data, f32_output, nelements);
+        } else if (tensor->type == GGML_TYPE_BF16) {
+            ggml_bf16_to_fp32_row((ggml_bf16_t *)tensor->data, f32_output, nelements);
+        } else if (ggml_is_quantized(tensor->type)) {
+            qtype.to_float(tensor->data, f32_output, nelements);
+        } else {
+            GGML_ASSERT(false); // unreachable
+        }
+        return;
+    }
+
+    size_t block_size;
+    if (tensor->type == GGML_TYPE_F16 ||
+        tensor->type == GGML_TYPE_BF16) {
+        block_size = 1;
+    } else {
+        block_size = (size_t)ggml_blck_size(tensor->type);
+    }
+
+    size_t block_size_bytes = ggml_type_size(tensor->type);
+
+    GGML_ASSERT(nelements % block_size == 0);
+    size_t nblocks = nelements / block_size;
+    size_t blocks_per_thread = nblocks / nthread;
+    size_t spare_blocks = nblocks - (blocks_per_thread * nthread); // if blocks aren't divisible by thread count
+
+    size_t in_buff_offs = 0;
+    size_t out_buff_offs = 0;
+
+    for (int tnum = 0; tnum < nthread; tnum++) {
+        size_t thr_blocks = blocks_per_thread + (tnum == nthread - 1 ? spare_blocks : 0); // num blocks for this thread
+        size_t thr_elems = thr_blocks * block_size; // number of elements for this thread
+        size_t thr_block_bytes = thr_blocks * block_size_bytes; // number of input bytes for this thread
+
+        auto compute = [qtype] (ggml_type typ, uint8_t * inbuf, float * outbuf, int nels) {
+            if (typ == GGML_TYPE_F16) {
+                ggml_fp16_to_fp32_row((ggml_fp16_t *)inbuf, outbuf, nels);
+            } else if (typ == GGML_TYPE_BF16) {
+                ggml_bf16_to_fp32_row((ggml_bf16_t *)inbuf, outbuf, nels);
+            } else {
+                qtype.to_float(inbuf, outbuf, nels);
+            }
+        };
+        workers.emplace_back(compute, tensor->type, (uint8_t *) tensor->data + in_buff_offs, f32_output + out_buff_offs, thr_elems);
+        in_buff_offs += thr_block_bytes;
+        out_buff_offs += thr_elems;
+    }
+    for (auto & w : workers) { w.join(); }
+    workers.clear();
+}
+
+static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype) {
+    const std::string name = ggml_get_name(tensor);
+
+    // TODO: avoid hardcoded tensor names - use the TN_* constants
+    const llm_arch arch = qs.model.arch;
+    const auto       tn = LLM_TN(arch);
+
+    auto use_more_bits = [](int i_layer, int n_layers) -> bool {
+        return i_layer < n_layers/8 || i_layer >= 7*n_layers/8 || (i_layer - n_layers/8)%3 == 2;
+    };
+    const int n_expert = std::max(1, (int)qs.model.hparams.n_expert);
+    auto layer_info = [n_expert] (int i_layer, int n_layer, const char * name) {
+        if (n_expert > 1) {
+            // Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but iccasionally randomly
+            // sprinkled in the model. Hence, simply dividing i_ffn_down by n_expert does not work
+            // for getting the current layer as I initially thought, and we need to resort to parsing the
+            // tensor name.
+            if (sscanf(name, "blk.%d.", &i_layer) != 1) {
+                throw std::runtime_error(format("Failed to determine layer for tensor %s", name));
+            }
+            if (i_layer < 0 || i_layer >= n_layer) {
+                throw std::runtime_error(format("Bad layer %d for tensor %s. Must be in [0, %d)", i_layer, name, n_layer));
+            }
+        }
+        return std::make_pair(i_layer, n_layer);
+    };
+
+    // for arches that share the same tensor between the token embeddings and the output, we quantize the token embeddings
+    // with the quantization of the output tensor
+    if (name == tn(LLM_TENSOR_OUTPUT, "weight") || (!qs.has_output && name == tn(LLM_TENSOR_TOKEN_EMBD, "weight"))) {
+        if (qs.params->output_tensor_type < GGML_TYPE_COUNT) {
+            new_type = qs.params->output_tensor_type;
+        } else {
+            int nx = tensor->ne[0];
+            if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
+                new_type = GGML_TYPE_Q8_0;
+            }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
+                     ftype == LLAMA_FTYPE_MOSTLY_IQ1_S   || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S  || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M   ||
+                     ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
+                new_type = GGML_TYPE_Q5_K;
+            }
+            else if (new_type != GGML_TYPE_Q8_0) {
+                new_type = GGML_TYPE_Q6_K;
+            }
+        }
+    } else if (name == "token_embd.weight") {
+        if (qs.params->token_embedding_type < GGML_TYPE_COUNT) {
+            new_type = qs.params->token_embedding_type;
+        } else {
+            if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS ||
+                ftype == LLAMA_FTYPE_MOSTLY_IQ1_S   || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
+                new_type = GGML_TYPE_Q2_K;
+            }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
+                new_type = GGML_TYPE_IQ3_S;
+            }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
+                new_type = GGML_TYPE_IQ3_S;
+            }
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_BN || ftype == LLAMA_FTYPE_MOSTLY_IQ2_BN) {
+                new_type = GGML_TYPE_IQ4_NL;
+            }
+            else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 ||
+                     new_type == GGML_TYPE_Q4_0_8_8) {
+                new_type = GGML_TYPE_Q4_0;
+            }
+        }
+    } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S ||
+               ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M    || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
+        if (name.find("attn_v.weight") != std::string::npos) {
+            if (qs.model.hparams.n_gqa() >= 4 || qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q4_K;
+            else new_type = ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ? GGML_TYPE_IQ3_S : GGML_TYPE_Q2_K;
+            ++qs.i_attention_wv;
+        }
+        else if (qs.model.hparams.n_expert == 8 && name.find("attn_k.weight") != std::string::npos) {
+            new_type = GGML_TYPE_Q4_K;
+        }
+        else if (name.find("ffn_down") != std::string::npos) {
+            if (qs.i_ffn_down < qs.n_ffn_down/8) {
+                new_type = ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ? GGML_TYPE_IQ3_S : GGML_TYPE_Q2_K;
+            }
+            ++qs.i_ffn_down;
+        }
+        else if (name.find("attn_output.weight") != std::string::npos) {
+            if (qs.model.hparams.n_expert == 8) {
+                new_type = GGML_TYPE_Q5_K;
+            } else {
+                if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) new_type = GGML_TYPE_IQ2_XXS;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) new_type = GGML_TYPE_IQ3_S;
+            }
+        }
+    } else if (name.find("attn_v.weight") != std::string::npos) {
+        if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) {
+            new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && qs.model.hparams.n_gqa() >= 4) {
+            new_type = GGML_TYPE_Q4_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
+            new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : !qs.has_imatrix ? GGML_TYPE_IQ3_S : GGML_TYPE_IQ3_XXS;
+        }
+        else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S) && qs.model.hparams.n_gqa() >= 4) {
+            new_type = GGML_TYPE_Q4_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) {
+            new_type = GGML_TYPE_Q4_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
+            new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+        else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS) && qs.model.hparams.n_gqa() >= 4) {
+            new_type = GGML_TYPE_Q5_K;
+        }
+        else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
+                use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
+        if (qs.model.type == MODEL_70B) {
+            // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is
+            // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
+            // nearly negligible increase in model size by quantizing this tensor with more bits:
+            if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K;
+        }
+        if (qs.model.hparams.n_expert == 8) {
+            // for the 8-expert model, bumping this to Q8_0 trades just ~128MB
+            // TODO: explore better strategies
+            new_type = GGML_TYPE_Q8_0;
+        }
+        ++qs.i_attention_wv;
+    } else if (name.find("attn_k.weight") != std::string::npos) {
+        if (qs.model.hparams.n_expert == 8) {
+            // for the 8-expert model, bumping this to Q8_0 trades just ~128MB
+            // TODO: explore better strategies
+            new_type = GGML_TYPE_Q8_0;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) {
+            new_type = GGML_TYPE_IQ3_XXS;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
+            new_type = GGML_TYPE_IQ2_S;
+        }
+    } else if (name.find("attn_q.weight") != std::string::npos) {
+        if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) {
+            new_type = GGML_TYPE_IQ3_XXS;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
+            new_type = GGML_TYPE_IQ2_S;
+        }
+    } else if (name.find("ffn_down") != std::string::npos) {
+        auto info = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str());
+        int i_layer = info.first, n_layer = info.second;
+        if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) {
+            if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS && !qs.has_imatrix) {
+            new_type = i_layer < n_layer/8 ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
+            new_type = i_layer < n_layer/16 ? GGML_TYPE_Q5_K
+                     : arch != LLM_ARCH_FALCON || use_more_bits(i_layer, n_layer) ? GGML_TYPE_Q4_K
+                     : GGML_TYPE_Q3_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M && (i_layer < n_layer/8 ||
+                    (qs.model.hparams.n_expert == 8 && use_more_bits(i_layer, n_layer)))) {
+            new_type = GGML_TYPE_Q4_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) {
+            new_type = arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
+            if (arch == LLM_ARCH_FALCON) {
+                new_type = i_layer < n_layer/16 ? GGML_TYPE_Q6_K :
+                           use_more_bits(i_layer, n_layer) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
+            } else {
+                if (use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q6_K;
+            }
+        }
+        else if (i_layer < n_layer/8 && (ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS) && !qs.has_imatrix) {
+            new_type = GGML_TYPE_Q5_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q6_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && i_layer < n_layer/8) {
+            new_type = GGML_TYPE_Q5_K;
+        }
+        else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_0 || ftype == LLAMA_FTYPE_MOSTLY_Q5_0)
+                && qs.has_imatrix && i_layer < n_layer/8) {
+            // Guard against craziness in the first few ffn_down layers that can happen even with imatrix for Q4_0/Q5_0.
+            // We only do it when an imatrix is provided because a) we want to make sure that one can always get the
+            // same quantization as before imatrix stuff, and b) Q4_1/Q5_1 do go crazy on ffn_down without an imatrix.
+            new_type = ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ? GGML_TYPE_Q4_1 : GGML_TYPE_Q5_1;
+        }
+        ++qs.i_ffn_down;
+    } else if (name.find("attn_output.weight") != std::string::npos) {
+        if (arch != LLM_ARCH_FALCON) {
+            if (qs.model.hparams.n_expert == 8) {
+                if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K   || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
+                    ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M  || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL  ||
+                    ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M  || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S  ||
+                    ftype == LLAMA_FTYPE_MOSTLY_IQ3_M  || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS) {
+                    new_type = GGML_TYPE_Q5_K;
+                }
+            } else {
+                if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K   ) new_type = GGML_TYPE_Q3_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) new_type = GGML_TYPE_IQ3_S;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M ) new_type = GGML_TYPE_Q4_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L ) new_type = GGML_TYPE_Q5_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M  ) new_type = GGML_TYPE_Q4_K;
+            }
+        } else {
+            if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K;
+        }
+    }
+    else if (name.find("attn_qkv.weight") != std::string::npos) {
+        if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) {
+            new_type = GGML_TYPE_Q4_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
+    }
+    else if (name.find("ffn_gate") != std::string::npos) {
+        auto info = layer_info(qs.i_ffn_gate, qs.n_ffn_gate, name.c_str());
+        int i_layer = info.first, n_layer = info.second;
+        if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (i_layer >= n_layer/8 && i_layer < 7*n_layer/8)) {
+            new_type = GGML_TYPE_IQ3_XXS;
+        }
+        ++qs.i_ffn_gate;
+    }
+    else if (name.find("ffn_up") != std::string::npos) {
+        auto info = layer_info(qs.i_ffn_up, qs.n_ffn_up, name.c_str());
+        int i_layer = info.first, n_layer = info.second;
+        if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (i_layer >= n_layer/8 && i_layer < 7*n_layer/8)) {
+            new_type = GGML_TYPE_IQ3_XXS;
+        }
+        ++qs.i_ffn_up;
+    }
+
+    //    if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
+    //}
+    // IK: let's remove this, else Q2_K is almost the same as Q3_K_S
+    //else if (name.find("ffn_gate") != std::string::npos || name.find("ffn_up") != std::string::npos) {
+    //    if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
+    //}
+    // This can be used to reduce the size of the Q5_K_S model.
+    // The associated PPL increase is fully in line with the size reduction
+    //else {
+    //    if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q4_K;
+    //}
+    bool convert_incompatible_tensor = false;
+    if (new_type == GGML_TYPE_Q2_K    || new_type == GGML_TYPE_Q3_K    || new_type == GGML_TYPE_Q4_K   ||
+        new_type == GGML_TYPE_Q5_K    || new_type == GGML_TYPE_Q6_K    || new_type == GGML_TYPE_IQ4_XS ||
+        new_type == GGML_TYPE_IQ2_XS  || new_type == GGML_TYPE_IQ2_XXS || new_type == GGML_TYPE_IQ2_S  ||
+        new_type == GGML_TYPE_IQ3_XXS || new_type == GGML_TYPE_IQ1_S   || new_type == GGML_TYPE_IQ3_S  ||
+        new_type == GGML_TYPE_IQ1_M) {
+        int nx = tensor->ne[0];
+        int ny = tensor->ne[1];
+        if (nx % QK_K != 0) {
+            LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for %s", __func__, nx, ny, QK_K, ggml_type_name(new_type));
+            convert_incompatible_tensor = true;
+        } else {
+            ++qs.n_k_quantized;
+        }
+    }
+    if (new_type == GGML_TYPE_IQ1_BN || new_type == GGML_TYPE_IQ2_BN) {
+        int nx = tensor->ne[0];
+        if (nx % QK_IQ1BN != 0) {
+            convert_incompatible_tensor = true;
+        }
+    }
+    if (convert_incompatible_tensor) {
+        switch (new_type) {
+            case GGML_TYPE_IQ2_XXS:
+            case GGML_TYPE_IQ2_XS:
+            case GGML_TYPE_IQ2_S:
+            case GGML_TYPE_IQ3_XXS:
+            case GGML_TYPE_IQ3_S:
+            case GGML_TYPE_IQ1_S:
+            case GGML_TYPE_IQ1_M:
+            case GGML_TYPE_Q2_K:
+            case GGML_TYPE_Q3_K:
+            case GGML_TYPE_IQ4_XS: new_type = GGML_TYPE_IQ4_NL; break;
+            case GGML_TYPE_Q4_K:   new_type = GGML_TYPE_Q5_0;   break;
+            case GGML_TYPE_Q5_K:   new_type = GGML_TYPE_Q5_1;   break;
+            case GGML_TYPE_Q6_K:   new_type = GGML_TYPE_Q8_0;   break;
+            default: throw std::runtime_error("\nUnsupported tensor size encountered\n");
+        }
+        LLAMA_LOG_WARN(" - using fallback quantization %s\n", ggml_type_name(new_type));
+        ++qs.n_fallback;
+    }
+
+    return new_type;
+}
+
+static size_t llama_tensor_quantize_internal(enum ggml_type new_type, const float * f32_data, void * new_data, const int64_t chunk_size, int64_t nrows, int64_t n_per_row, const float * imatrix, std::vector & workers, const int nthread) {
+    if (nthread < 2) {
+        // single-thread
+        size_t new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, nrows, n_per_row, imatrix);
+        if (!ggml_validate_row_data(new_type, new_data, new_size)) {
+            throw std::runtime_error("quantized data validation failed");
+        }
+        return new_size;
+    }
+
+    std::mutex mutex;
+    int64_t counter = 0;
+    size_t new_size = 0;
+    bool valid = true;
+    auto compute = [&mutex, &counter, &new_size, &valid, new_type, f32_data, new_data, chunk_size,
+            nrows, n_per_row, imatrix]() {
+        const int64_t nrows_per_chunk = chunk_size / n_per_row;
+        size_t local_size = 0;
+        while (true) {
+            std::unique_lock lock(mutex);
+            int64_t first_row = counter; counter += nrows_per_chunk;
+            if (first_row >= nrows) {
+                if (local_size > 0) {
+                    new_size += local_size;
+                }
+                break;
+            }
+            lock.unlock();
+            const int64_t this_nrow = std::min(nrows - first_row, nrows_per_chunk);
+            size_t this_size = ggml_quantize_chunk(new_type, f32_data, new_data, first_row * n_per_row, this_nrow, n_per_row, imatrix);
+            local_size += this_size;
+
+            // validate the quantized data
+            const size_t row_size  = ggml_row_size(new_type, n_per_row);
+            void * this_data = (char *) new_data + first_row * row_size;
+            if (!ggml_validate_row_data(new_type, this_data, this_size)) {
+                std::unique_lock lock(mutex);
+                valid = false;
+                break;
+            }
+        }
+    };
+    for (int it = 0; it < nthread - 1; ++it) {
+        workers.emplace_back(compute);
+    }
+    compute();
+    for (auto & w : workers) { w.join(); }
+    workers.clear();
+    if (!valid) {
+        throw std::runtime_error("quantized data validation failed");
+    }
+    return new_size;
+}
+
+static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, const llama_model_quantize_params * params) {
+    ggml_type default_type;
+    llama_ftype ftype = params->ftype;
+
+    switch (params->ftype) {
+        case LLAMA_FTYPE_MOSTLY_Q4_0: default_type = GGML_TYPE_Q4_0; break;
+        case LLAMA_FTYPE_MOSTLY_Q4_1: default_type = GGML_TYPE_Q4_1; break;
+        case LLAMA_FTYPE_MOSTLY_Q5_0: default_type = GGML_TYPE_Q5_0; break;
+        case LLAMA_FTYPE_MOSTLY_Q5_1: default_type = GGML_TYPE_Q5_1; break;
+        case LLAMA_FTYPE_MOSTLY_Q8_0: default_type = GGML_TYPE_Q8_0; break;
+        case LLAMA_FTYPE_MOSTLY_F16:  default_type = GGML_TYPE_F16;  break;
+        case LLAMA_FTYPE_MOSTLY_BF16: default_type = GGML_TYPE_BF16; break;
+        case LLAMA_FTYPE_ALL_F32:     default_type = GGML_TYPE_F32;  break;
+
+        // K-quants
+        case LLAMA_FTYPE_MOSTLY_Q2_K_S:
+        case LLAMA_FTYPE_MOSTLY_Q2_K:    default_type = GGML_TYPE_Q2_K;    break;
+        case LLAMA_FTYPE_MOSTLY_IQ3_XS:  default_type = GGML_TYPE_IQ3_S;   break;
+        case LLAMA_FTYPE_MOSTLY_Q3_K_S:
+        case LLAMA_FTYPE_MOSTLY_Q3_K_M:
+        case LLAMA_FTYPE_MOSTLY_Q3_K_L:  default_type = GGML_TYPE_Q3_K;    break;
+        case LLAMA_FTYPE_MOSTLY_Q4_K_S:
+        case LLAMA_FTYPE_MOSTLY_Q4_K_M:  default_type = GGML_TYPE_Q4_K;    break;
+        case LLAMA_FTYPE_MOSTLY_Q5_K_S:
+        case LLAMA_FTYPE_MOSTLY_Q5_K_M:  default_type = GGML_TYPE_Q5_K;    break;
+        case LLAMA_FTYPE_MOSTLY_Q6_K:    default_type = GGML_TYPE_Q6_K;    break;
+        case LLAMA_FTYPE_MOSTLY_IQ2_XXS: default_type = GGML_TYPE_IQ2_XXS; break;
+        case LLAMA_FTYPE_MOSTLY_IQ2_XS:  default_type = GGML_TYPE_IQ2_XS;  break;
+        case LLAMA_FTYPE_MOSTLY_IQ2_S:   default_type = GGML_TYPE_IQ2_XS;  break;
+        case LLAMA_FTYPE_MOSTLY_IQ2_M:   default_type = GGML_TYPE_IQ2_S;   break;
+        case LLAMA_FTYPE_MOSTLY_IQ3_XXS: default_type = GGML_TYPE_IQ3_XXS; break;
+        case LLAMA_FTYPE_MOSTLY_IQ1_S:   default_type = GGML_TYPE_IQ1_S;   break;
+        case LLAMA_FTYPE_MOSTLY_IQ1_M:   default_type = GGML_TYPE_IQ1_M;   break;
+        case LLAMA_FTYPE_MOSTLY_IQ1_BN:  default_type = GGML_TYPE_IQ1_BN;  break;
+        case LLAMA_FTYPE_MOSTLY_IQ2_BN:  default_type = GGML_TYPE_IQ2_BN;  break;
+        case LLAMA_FTYPE_MOSTLY_IQ4_NL:  default_type = GGML_TYPE_IQ4_NL;  break;
+        case LLAMA_FTYPE_MOSTLY_IQ4_XS:  default_type = GGML_TYPE_IQ4_XS;  break;
+        case LLAMA_FTYPE_MOSTLY_IQ3_S:   default_type = GGML_TYPE_IQ3_S;   break;
+        case LLAMA_FTYPE_MOSTLY_IQ3_M:   default_type = GGML_TYPE_IQ3_S;   break;
+        case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: default_type = GGML_TYPE_Q4_0_4_4; break;
+        case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: default_type = GGML_TYPE_Q4_0_4_8; break;
+        case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: default_type = GGML_TYPE_Q4_0_8_8; break;
+
+        default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
+    }
+
+    int nthread = params->nthread;
+
+    if (nthread <= 0) {
+        nthread = std::thread::hardware_concurrency();
+    }
+
+    // mmap consistently increases speed Linux, and also increases speed on Windows with
+    // hot cache. It may cause a slowdown on macOS, possibly related to free memory.
+#if defined(__linux__) || defined(_WIN32)
+    constexpr bool use_mmap = true;
+#else
+    constexpr bool use_mmap = false;
+#endif
+
+    llama_model_kv_override * kv_overrides = nullptr;
+    if (params->kv_overrides) {
+        auto v = (std::vector*)params->kv_overrides;
+        kv_overrides = v->data();
+    }
+    llama_model_loader ml(fname_inp, use_mmap, /*check_tensors*/ true, kv_overrides);
+    ml.init_mappings(false); // no prefetching
+
+    llama_model model;
+    llm_load_arch(ml, model);
+    llm_load_hparams(ml, model);
+
+    struct quantize_state_internal qs(model, params);
+
+    if (params->only_copy) {
+        ftype = model.ftype;
+    }
+    const std::unordered_map> * imatrix_data = nullptr;
+    if (params->imatrix) {
+        imatrix_data = static_cast>*>(params->imatrix);
+        if (imatrix_data) {
+            LLAMA_LOG_INFO("================================ Have weights data with %d entries\n",int(imatrix_data->size()));
+            qs.has_imatrix = true;
+            // check imatrix for nans or infs
+            for (const auto & kv : *imatrix_data) {
+                for (float f : kv.second) {
+                    if (!std::isfinite(f)) {
+                        throw std::runtime_error(format("imatrix contains non-finite value %f\n", f));
+                    }
+                }
+            }
+        }
+    }
+
+    const size_t align = GGUF_DEFAULT_ALIGNMENT;
+    struct gguf_context * ctx_out = gguf_init_empty();
+
+    // copy the KV pairs from the input file
+    gguf_set_kv     (ctx_out, ml.meta);
+    gguf_set_val_u32(ctx_out, "general.quantization_version", GGML_QNT_VERSION); // TODO: use LLM_KV
+    gguf_set_val_u32(ctx_out, "general.file_type", ftype); // TODO: use LLM_KV
+
+    // Remove split metadata
+    gguf_remove_key(ctx_out, ml.llm_kv(LLM_KV_SPLIT_NO).c_str());
+    gguf_remove_key(ctx_out, ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str());
+    gguf_remove_key(ctx_out, ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str());
+
+    if (params->kv_overrides) {
+        const std::vector & overrides = *(const std::vector *)params->kv_overrides;
+        for (auto & o : overrides) {
+            if (o.key[0] == 0) break;
+            if (o.tag == LLAMA_KV_OVERRIDE_TYPE_FLOAT) {
+                gguf_set_val_f32(ctx_out, o.key, o.val_f64);
+            } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_INT) {
+                gguf_set_val_i32(ctx_out, o.key, o.val_i64);
+            } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_BOOL) {
+                gguf_set_val_bool(ctx_out, o.key, o.val_bool);
+            } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_STR) {
+                gguf_set_val_str(ctx_out, o.key, o.val_str);
+            } else {
+                LLAMA_LOG_WARN("%s: unknown KV override type for key %s\n", __func__, o.key);
+            }
+        }
+    }
+
+    for (int i = 0; i < ml.n_tensors; ++i) {
+        const struct ggml_tensor * meta = ml.get_tensor_meta(i);
+
+        const std::string name = ggml_get_name(meta);
+
+        // TODO: avoid hardcoded tensor names - use the TN_* constants
+        if (name.find("attn_v.weight")   != std::string::npos ||
+            name.find("attn_qkv.weight") != std::string::npos) {
+            ++qs.n_attention_wv;
+        } else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) {
+            qs.has_output = true;
+        }
+    }
+
+    qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer;
+
+    // sanity checks
+    //
+    //  - qs.n_attention_wv == 0                         for Mamba           models
+    //  - qs.n_attention_wv == model.hparams.n_layer     for Transformer     models
+    //  - qs.n_attention_wv == 3 * model.hparams.n_layer for Encoder-Decoder models
+    //
+    GGML_ASSERT((qs.n_attention_wv == 0 || qs.n_attention_wv == (int)model.hparams.n_layer || qs.n_attention_wv == 3 * (int)model.hparams.n_layer) && "n_attention_wv is unexpected");
+
+    size_t total_size_org = 0;
+    size_t total_size_new = 0;
+
+    std::vector workers;
+    workers.reserve(nthread);
+
+    int idx = 0;
+
+    std::vector> read_data;
+    std::vector> work;
+    std::vector> f32_conv_buf;
+
+    uint16_t n_split = 1;
+    // Assume split index is continuous
+    if (params->keep_split) {
+        for (int i = 0; i < ml.n_tensors; ++i) {
+            n_split = std::max(uint16_t(ml.get_weight(i)->idx+1), n_split);
+        }
+    }
+    std::vector ctx_outs(n_split, NULL);
+    ctx_outs[0] = ctx_out;
+
+    // populate the original tensors so we get an initial meta data
+    for (int i = 0; i < ml.n_tensors; ++i) {
+        auto weight = ml.get_weight(i);
+        uint16_t i_split = params->keep_split ? weight->idx : 0;
+        struct ggml_tensor * tensor = weight->tensor;
+        if (ctx_outs[i_split] == NULL) {
+            ctx_outs[i_split] = gguf_init_empty();
+        }
+        gguf_add_tensor(ctx_outs[i_split], tensor);
+    }
+
+    // Set split info if needed
+    if (n_split > 1) {
+        for (size_t i = 0; i < ctx_outs.size(); ++i) {
+            gguf_set_val_u16(ctx_outs[i], ml.llm_kv(LLM_KV_SPLIT_NO).c_str(), i);
+            gguf_set_val_u16(ctx_outs[i], ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str(), n_split);
+            gguf_set_val_i32(ctx_outs[i], ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str(), ml.n_tensors);
+        }
+    }
+
+    int cur_split = -1;
+    std::ofstream fout;
+    auto close_ofstream = [&]() {
+        // Write metadata and close file handler
+        if (fout.is_open()) {
+            fout.seekp(0);
+            std::vector data(gguf_get_meta_size(ctx_outs[cur_split]));
+            gguf_get_meta_data(ctx_outs[cur_split], data.data());
+            fout.write((const char *) data.data(), data.size());
+            fout.close();
+        }
+    };
+    auto new_ofstream = [&](int index) {
+        cur_split = index;
+        GGML_ASSERT(ctx_outs[cur_split] && "Find uninitialized gguf_context");
+        std::string fname = fname_out;
+        if (params->keep_split) {
+            char split_path[PATH_MAX] = {0};
+            llama_split_path(split_path, sizeof(split_path), fname_out.c_str(), cur_split, n_split);
+            fname = std::string(split_path);
+        }
+
+        fout = std::ofstream(fname, std::ios::binary);
+        fout.exceptions(std::ofstream::failbit); // fail fast on write errors
+        const size_t meta_size = gguf_get_meta_size(ctx_outs[cur_split]);
+        // placeholder for the meta data
+        ::zeros(fout, meta_size);
+    };
+
+    const auto tn = LLM_TN(model.arch);
+    new_ofstream(0);
+    for (int i = 0; i < ml.n_tensors; ++i) {
+        auto weight = ml.get_weight(i);
+        struct ggml_tensor * tensor = weight->tensor;
+        if (weight->idx != cur_split && params->keep_split) {
+            close_ofstream();
+            new_ofstream(weight->idx);
+        }
+
+        const std::string name = ggml_get_name(tensor);
+
+        if (!ml.use_mmap) {
+            if (read_data.size() < ggml_nbytes(tensor)) {
+                read_data.resize(ggml_nbytes(tensor));
+            }
+            tensor->data = read_data.data();
+        }
+        ml.load_data_for(tensor);
+
+        LLAMA_LOG_INFO("[%4d/%4d] %36s - [%s], type = %6s, ",
+               ++idx, ml.n_tensors,
+               ggml_get_name(tensor),
+               llama_format_tensor_shape(tensor).c_str(),
+               ggml_type_name(tensor->type));
+
+        // This used to be a regex, but  has an extreme cost to compile times.
+        bool quantize = name.rfind("weight") == name.size() - 6; // ends with 'weight'?
+
+        // quantize only 2D and 3D tensors (experts)
+        quantize &= (ggml_n_dims(tensor) >= 2);
+
+        // do not quantize norm tensors
+        quantize &= name.find("_norm.weight") == std::string::npos;
+
+        quantize &= params->quantize_output_tensor || name != "output.weight";
+        quantize &= !params->only_copy;
+
+        // do not quantize expert gating tensors
+        // NOTE: can't use LLM_TN here because the layer number is not known
+        quantize &= name.find("ffn_gate_inp.weight") == std::string::npos;
+
+        // do not quantize positional embeddings and token types (BERT)
+        quantize &= name != LLM_TN(model.arch)(LLM_TENSOR_POS_EMBD,    "weight");
+        quantize &= name != LLM_TN(model.arch)(LLM_TENSOR_TOKEN_TYPES, "weight");
+
+        // do not quantize Mamba's small yet 2D weights
+        // NOTE: can't use LLM_TN here because the layer number is not known
+        quantize &= name.find("ssm_conv1d.weight") == std::string::npos;
+        quantize &= name.find("ssm_x.weight")      == std::string::npos;
+        quantize &= name.find("ssm_dt.weight")     == std::string::npos;
+
+        // do not quantize relative position bias (T5)
+        quantize &= name.find("attn_rel_b.weight") == std::string::npos;
+
+        enum ggml_type new_type;
+        void * new_data;
+        size_t new_size;
+
+        if (quantize) {
+            new_type = default_type;
+
+            // get more optimal quantization type based on the tensor shape, layer, etc.
+            if (!params->pure && ggml_is_quantized(default_type)) {
+                new_type = llama_tensor_get_type(qs, new_type, tensor, ftype);
+            }
+            if (params->token_embedding_type < GGML_TYPE_COUNT && strcmp(tensor->name, "token_embd.weight") == 0) {
+                new_type = params->token_embedding_type;
+            }
+            if (params->output_tensor_type < GGML_TYPE_COUNT && strcmp(tensor->name, "output.weight") == 0) {
+                new_type = params->output_tensor_type;
+            }
+
+            // If we've decided to quantize to the same type the tensor is already
+            // in then there's nothing to do.
+            quantize = tensor->type != new_type;
+        }
+
+        if (!quantize) {
+            new_type = tensor->type;
+            new_data = tensor->data;
+            new_size = ggml_nbytes(tensor);
+            LLAMA_LOG_INFO("size = %8.3f MB\n", ggml_nbytes(tensor)/1024.0/1024.0);
+        } else {
+            const int64_t nelements = ggml_nelements(tensor);
+
+            const float * imatrix = nullptr;
+            if (imatrix_data) {
+                auto it = imatrix_data->find(tensor->name);
+                if (it == imatrix_data->end()) {
+                    LLAMA_LOG_INFO("\n====== %s: did not find weights for %s\n", __func__, tensor->name);
+                } else {
+                    if (it->second.size() == (size_t)tensor->ne[0]*tensor->ne[2]) {
+                        imatrix = it->second.data();
+                    } else {
+                        LLAMA_LOG_INFO("\n====== %s: imatrix size %d is different from tensor size %d for %s\n", __func__,
+                                int(it->second.size()), int(tensor->ne[0]*tensor->ne[2]), tensor->name);
+
+                        // this can happen when quantizing an old mixtral model with split tensors with a new incompatible imatrix
+                        // this is a significant error and it may be good idea to abort the process if this happens,
+                        // since many people will miss the error and not realize that most of the model is being quantized without an imatrix
+                        // tok_embd should be ignored in this case, since it always causes this warning
+                        if (name != tn(LLM_TENSOR_TOKEN_EMBD, "weight")) {
+                            throw std::runtime_error(format("imatrix size %d is different from tensor size %d for %s",
+                                    int(it->second.size()), int(tensor->ne[0]*tensor->ne[2]), tensor->name));
+                        }
+                    }
+                }
+            }
+            if ((new_type == GGML_TYPE_IQ2_XXS ||
+                 new_type == GGML_TYPE_IQ2_XS  ||
+                 new_type == GGML_TYPE_IQ2_S   ||
+                 new_type == GGML_TYPE_IQ1_S   ||
+                (new_type == GGML_TYPE_IQ1_M && strcmp(tensor->name, "token_embd.weight") && strcmp(tensor->name, "output.weight"))  ||
+                (new_type == GGML_TYPE_Q2_K && params->ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && strcmp(tensor->name, "token_embd.weight") != 0)) && !imatrix) {
+                LLAMA_LOG_ERROR("\n\n============================================================\n");
+                LLAMA_LOG_ERROR("Missing importance matrix for tensor %s in a very low-bit quantization\n", tensor->name);
+                LLAMA_LOG_ERROR("The result will be garbage, so bailing out\n");
+                LLAMA_LOG_ERROR("============================================================\n\n");
+                throw std::runtime_error(format("Missing importance matrix for tensor %s in a very low-bit quantization", tensor->name));
+            }
+
+            float * f32_data;
+
+            if (tensor->type == GGML_TYPE_F32) {
+                f32_data = (float *) tensor->data;
+            } else if (ggml_is_quantized(tensor->type) && !params->allow_requantize) {
+                throw std::runtime_error(format("requantizing from type %s is disabled", ggml_type_name(tensor->type)));
+            } else {
+                llama_tensor_dequantize_internal(tensor, f32_conv_buf, workers, nelements, nthread);
+                f32_data = (float *) f32_conv_buf.data();
+            }
+
+            int chunk_size_multiplier = 1;
+            if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 || new_type == GGML_TYPE_Q4_0_8_8) {
+                if ((new_type == GGML_TYPE_Q4_0_8_8) && (tensor->ne[1] % 8 != 0)) new_type = GGML_TYPE_Q4_0;
+                else if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_0;
+                if (new_type == GGML_TYPE_Q4_0_8_8) chunk_size_multiplier = 8;
+                else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8) chunk_size_multiplier = 4;
+            }
+
+            LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
+            fflush(stdout);
+
+            if (work.size() < (size_t)nelements * 4) {
+                work.resize(nelements * 4); // upper bound on size
+            }
+            new_data = work.data();
+
+            const int64_t n_per_row = tensor->ne[0];
+            const int64_t nrows = tensor->ne[1];
+
+            static const int64_t min_chunk_size = 32 * 512;
+            const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row)) *
+                                       chunk_size_multiplier;
+
+            const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1];
+            const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;
+            const int64_t nthread_use = nthread > 1 ? std::max((int64_t)1, std::min((int64_t)nthread, nchunk)) : 1;
+
+            // quantize each expert separately since they have different importance matrices
+            new_size = 0;
+            for (int64_t i03 = 0; i03 < tensor->ne[2]; ++i03) {
+                const float * f32_data_03 = f32_data + i03 * nelements_matrix;
+                void * new_data_03 = (char *)new_data + ggml_row_size(new_type, n_per_row) * i03 * nrows;
+                const float * imatrix_03 = imatrix ? imatrix + i03 * n_per_row : nullptr;
+
+                new_size += llama_tensor_quantize_internal(new_type, f32_data_03, new_data_03, chunk_size, nrows, n_per_row, imatrix_03, workers, nthread_use);
+            }
+            LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB\n", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0);
+        }
+        total_size_org += ggml_nbytes(tensor);
+        total_size_new += new_size;
+
+        // update the gguf meta data as we go
+        gguf_set_tensor_type(ctx_outs[cur_split], name.c_str(), new_type);
+        gguf_set_tensor_data(ctx_outs[cur_split], name.c_str(), new_data, new_size);
+
+        // write tensor data + padding
+        fout.write((const char *) new_data, new_size);
+        zeros(fout, GGML_PAD(new_size, align) - new_size);
+    }
+    close_ofstream();
+    for (auto & c:ctx_outs) {
+        gguf_free(c);
+    }
+
+    LLAMA_LOG_INFO("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
+    LLAMA_LOG_INFO("%s: quant size  = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0);
+
+    if (qs.n_fallback > 0) {
+        LLAMA_LOG_WARN("%s: WARNING: %d of %d tensor(s) required fallback quantization\n",
+                __func__, qs.n_fallback, qs.n_k_quantized + qs.n_fallback);
+    }
+}
+
+static void llama_lora_adapter_init_internal(struct llama_model * model, const char * path_lora, struct llama_lora_adapter & adapter) {
+    LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora);
+
+    ggml_context * ctx = nullptr;
+    struct gguf_init_params meta_gguf_params = {
+        /* .no_alloc = */ true,
+        /* .ctx      = */ &ctx,
+    };
+    struct gguf_context * ctx_gguf = gguf_init_from_file(path_lora, meta_gguf_params);
+    if (!ctx_gguf) {
+        throw std::runtime_error("failed to load lora adapter file from " + std::string(path_lora));
+    }
+
+    // check metadata
+    {
+        auto get_kv_str = [&](const std::string & key) -> std::string {
+            int id = gguf_find_key(ctx_gguf, key.c_str());
+            return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf, id));
+        };
+        auto get_kv_f32 = [&](const std::string & key) -> float {
+            int id = gguf_find_key(ctx_gguf, key.c_str());
+            return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf, id);
+        };
+        LLM_KV llm_kv = LLM_KV(LLM_ARCH_UNKNOWN);
+
+        auto general_type = get_kv_str(llm_kv(LLM_KV_GENERAL_TYPE));
+        if (general_type != "adapter") {
+            gguf_free(ctx_gguf);
+            throw std::runtime_error("expect general.type to be 'adapter', but got: " + general_type);
+        }
+
+        auto general_arch_str = get_kv_str(llm_kv(LLM_KV_GENERAL_ARCHITECTURE));
+        auto general_arch = llm_arch_from_string(general_arch_str);
+        if (general_arch != model->arch) {
+            gguf_free(ctx_gguf);
+            throw std::runtime_error("model arch and LoRA arch mismatch");
+        }
+
+        auto adapter_type = get_kv_str(llm_kv(LLM_KV_ADAPTER_TYPE));
+        if (adapter_type != "lora") {
+            gguf_free(ctx_gguf);
+            throw std::runtime_error("expect adapter.type to be 'lora', but got: " + adapter_type);
+        }
+
+        adapter.alpha = get_kv_f32(llm_kv(LLM_KV_ADAPTER_LORA_ALPHA));
+    }
+
+    int n_tensors = gguf_get_n_tensors(ctx_gguf);
+
+    // contexts for each buffer type
+    std::map ctx_map;
+    auto get_ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
+        auto it = ctx_map.find(buft);
+        if (it == ctx_map.end()) {
+            // add a new context
+            struct ggml_init_params params = {
+                /*.mem_size   =*/ n_tensors*ggml_tensor_overhead(),
+                /*.mem_buffer =*/ NULL,
+                /*.no_alloc   =*/ true,
+            };
+            ggml_context * buft_ctx = ggml_init(params);
+            ctx_map[buft] = buft_ctx;
+            return buft_ctx;
+        };
+        return it->second;
+    };
+
+    // bundle lora_a and lora_b into pairs
+    std::map ab_map;
+    auto str_endswith = [](const std::string & str, const std::string & suffix) {
+        return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
+    };
+    for (ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) {
+        std::string name(cur->name);
+        if (str_endswith(name, ".lora_a")) {
+            replace_all(name, ".lora_a", "");
+            if (ab_map.find(name) == ab_map.end()) {
+                ab_map[name] = llama_lora_weight(cur, nullptr);
+            } else {
+                ab_map[name].a = cur;
+            }
+        } else if (str_endswith(name, ".lora_b")) {
+            replace_all(name, ".lora_b", "");
+            if (ab_map.find(name) == ab_map.end()) {
+                ab_map[name] = llama_lora_weight(nullptr, cur);
+            } else {
+                ab_map[name].b = cur;
+            }
+        } else {
+            gguf_free(ctx_gguf);
+            ggml_free(ctx);
+            throw std::runtime_error("LoRA tensor '" + name + "' has unexpected suffix");
+        }
+    }
+
+    // add tensors
+    for (auto & it : ab_map) {
+        const std::string & name = it.first;
+        llama_lora_weight & w = it.second;
+
+        if (!w.a || !w.b) {
+            gguf_free(ctx_gguf);
+            ggml_free(ctx);
+            throw std::runtime_error("LoRA tensor pair for '" + name + "' is missing one component");
+        }
+
+        // device buft and device ctx
+        auto * model_tensor = llama_get_model_tensor(model, name.c_str());
+        if (!model_tensor) {
+            gguf_free(ctx_gguf);
+            ggml_free(ctx);
+            throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model");
+        }
+        struct ggml_context * dev_ctx = get_ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer));
+        // validate tensor shape
+        if (model_tensor->ne[0] != w.a->ne[0] || model_tensor->ne[1] != w.b->ne[1]) {
+            gguf_free(ctx_gguf);
+            ggml_free(ctx);
+            throw std::runtime_error("tensor '" + name + "' has incorrect shape");
+        }
+        if (w.a->ne[1] != w.b->ne[0]) {
+            gguf_free(ctx_gguf);
+            ggml_free(ctx);
+            throw std::runtime_error("lora_a tensor is not transposed (hint: adapter from \"finetune\" example is no longer supported)");
+        }
+        // save tensor to adapter
+        struct ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a);
+        struct ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b);
+        ggml_set_name(tensor_a, w.a->name);
+        ggml_set_name(tensor_b, w.b->name);
+        adapter.ab_map[name] = llama_lora_weight(tensor_a, tensor_b);
+    }
+
+    // allocate tensors / buffers and zero
+    {
+        adapter.ctxs.reserve(ctx_map.size());
+        adapter.bufs.reserve(ctx_map.size());
+        for (auto it : ctx_map) {
+            ggml_backend_buffer_type_t buft = it.first;
+            ggml_context * ctx_dev = it.second;
+            ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx_dev, buft);
+            if (!buf) {
+                gguf_free(ctx_gguf);
+                ggml_free(ctx);
+                throw std::runtime_error("failed to allocate buffer for lora adapter\n");
+            }
+            LLAMA_LOG_INFO("%s: %10s LoRA buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0);
+            adapter.ctxs.push_back(ctx_dev);
+            adapter.bufs.push_back(buf);
+        }
+    }
+
+    // set tensor data
+    {
+        llama_file gguf_file(path_lora, "rb");
+        std::vector read_buf;
+        auto set_tensor = [&](struct ggml_tensor * orig, struct ggml_tensor * dev) {
+            size_t offs = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, gguf_find_tensor(ctx_gguf, orig->name));
+            size_t size = ggml_nbytes(orig);
+            read_buf.resize(size);
+            gguf_file.seek(offs, SEEK_SET);
+            gguf_file.read_raw(read_buf.data(), size);
+            ggml_backend_tensor_set(dev, read_buf.data(), 0, size);
+        };
+        for (auto & it : adapter.ab_map) {
+            auto orig = ab_map[it.first];
+            auto dev  = it.second;
+            set_tensor(orig.a, dev.a);
+            set_tensor(orig.b, dev.b);
+        }
+    }
+
+    LLAMA_LOG_INFO("%s: loaded %ld tensors from lora file\n", __func__, adapter.ab_map.size()*2);
+
+    // free ctx for reading gguf
+    gguf_free(ctx_gguf);
+    ggml_free(ctx);
+}
+
+int32_t llama_lora_adapter_set(
+            struct llama_context * ctx,
+            struct llama_lora_adapter * adapter,
+            float scale) {
+    if (ctx->cparams.flash_attn) {
+        LLAMA_LOG_ERROR("%s: flash_attn is not compatible with LoRA\n", __func__);
+        return -1;
+    }
+    ctx->lora_adapters[adapter] = scale;
+    return 0;
+}
+
+int32_t llama_lora_adapter_remove(
+            struct llama_context * ctx,
+            struct llama_lora_adapter * adapter) {
+    auto pos = ctx->lora_adapters.find(adapter);
+    if (pos != ctx->lora_adapters.end()) {
+        ctx->lora_adapters.erase(pos);
+        return 0;
+    }
+    return -1;
+}
+
+void llama_lora_adapter_clear(struct llama_context * ctx) {
+    ctx->lora_adapters.clear();
+}
+
+void llama_lora_adapter_free(struct llama_lora_adapter * adapter) {
+    delete adapter;
+}
+
+//
+// interface implementation
+//
+struct llama_model_params llama_model_default_params() {
+    struct llama_model_params result = {
+        /*.n_gpu_layers                =*/ 0,
+        /*.split_mode                  =*/ LLAMA_SPLIT_MODE_LAYER,
+        /*.main_gpu                    =*/ 0,
+        /*.tensor_split                =*/ nullptr,
+        /*.rpc_servers                 =*/ nullptr,
+        /*.progress_callback           =*/ nullptr,
+        /*.progress_callback_user_data =*/ nullptr,
+        /*.kv_overrides                =*/ nullptr,
+        /*.vocab_only                  =*/ false,
+        /*.use_mmap                    =*/ true,
+        /*.use_mlock                   =*/ false,
+        /*.check_tensors               =*/ false,
+    };
+
+#ifdef GGML_USE_METAL
+    // note: we usually have plenty of VRAM, so by default offload all layers to the GPU
+    result.n_gpu_layers = 999;
+#endif
+
+    return result;
+}
+
+struct llama_context_params llama_context_default_params() {
+    struct llama_context_params result = {
+        /*.seed                        =*/ LLAMA_DEFAULT_SEED,
+        /*.n_ctx                       =*/ 512,
+        /*.n_batch                     =*/ 2048,
+        /*.n_ubatch                    =*/ 512,
+        /*.n_seq_max                   =*/ 1,
+        /*.n_threads                   =*/ GGML_DEFAULT_N_THREADS, // TODO: better default
+        /*.n_threads_batch             =*/ GGML_DEFAULT_N_THREADS,
+        /*.rope_scaling_type           =*/ LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED,
+        /*.pooling_type                =*/ LLAMA_POOLING_TYPE_UNSPECIFIED,
+        /*.attention_type              =*/ LLAMA_ATTENTION_TYPE_UNSPECIFIED,
+        /*.rope_freq_base              =*/ 0.0f,
+        /*.rope_freq_scale             =*/ 0.0f,
+        /*.yarn_ext_factor             =*/ -1.0f,
+        /*.yarn_attn_factor            =*/ 1.0f,
+        /*.yarn_beta_fast              =*/ 32.0f,
+        /*.yarn_beta_slow              =*/ 1.0f,
+        /*.yarn_orig_ctx               =*/ 0,
+        /*.defrag_thold                =*/ -1.0f,
+        /*.cb_eval                     =*/ nullptr,
+        /*.cb_eval_user_data           =*/ nullptr,
+        /*.type_k                      =*/ GGML_TYPE_F16,
+        /*.type_v                      =*/ GGML_TYPE_F16,
+        /*.logits_all                  =*/ false,
+        /*.embeddings                  =*/ false,
+        /*.offload_kqv                 =*/ true,
+        /*.flash_attn                  =*/ false,
+        /*.abort_callback              =*/ nullptr,
+        /*.abort_callback_data         =*/ nullptr,
+    };
+
+    return result;
+}
+
+struct llama_model_quantize_params llama_model_quantize_default_params() {
+    struct llama_model_quantize_params result = {
+        /*.nthread                     =*/ 0,
+        /*.ftype                       =*/ LLAMA_FTYPE_MOSTLY_Q5_1,
+        /*.output_tensor_type          =*/ GGML_TYPE_COUNT,
+        /*.token_embedding_type        =*/ GGML_TYPE_COUNT,
+        /*.allow_requantize            =*/ false,
+        /*.quantize_output_tensor      =*/ true,
+        /*.only_copy                   =*/ false,
+        /*.pure                        =*/ false,
+        /*.keep_split                  =*/ false,
+        /*.imatrix                     =*/ nullptr,
+        /*.kv_overrides                =*/ nullptr,
+    };
+
+    return result;
+}
+
+size_t llama_max_devices(void) {
+#if defined(GGML_USE_RPC)
+    return GGML_RPC_MAX_SERVERS;
+#elif defined(GGML_USE_METAL)
+    return 1;
+#elif defined(GGML_USE_CUDA)
+    return GGML_CUDA_MAX_DEVICES;
+#elif defined(GGML_USE_SYCL)
+    return GGML_SYCL_MAX_DEVICES;
+#elif defined(GGML_USE_VULKAN)
+    return GGML_VK_MAX_DEVICES;
+#elif defined(GGML_USE_CANN)
+    return GGML_CANN_MAX_DEVICES;
+#else
+    return 1;
+#endif
+}
+
+bool llama_supports_mmap(void) {
+    return llama_mmap::SUPPORTED;
+}
+
+bool llama_supports_mlock(void) {
+    return llama_mlock::SUPPORTED;
+}
+
+bool llama_supports_gpu_offload(void) {
+#if defined(GGML_USE_CUDA) || defined(GGML_USE_METAL)   || defined(GGML_USE_VULKAN) || \
+    defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE) || defined(GGML_USE_RPC)
+    // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
+    return true;
+#else
+    return false;
+#endif
+}
+
+void llama_backend_init(void) {
+    ggml_time_init();
+
+    // needed to initialize f16 tables
+    {
+        struct ggml_init_params params = { 0, NULL, false };
+        struct ggml_context * ctx = ggml_init(params);
+        ggml_free(ctx);
+    }
+}
+
+void llama_numa_init(enum ggml_numa_strategy numa) {
+    if (numa != GGML_NUMA_STRATEGY_DISABLED) {
+        ggml_numa_init(numa);
+    }
+}
+
+void llama_backend_free(void) {
+    ggml_quantize_free();
+}
+
+int64_t llama_time_us(void) {
+    return ggml_time_us();
+}
+
+struct llama_model * llama_load_model_from_file(
+        const char * path_model,
+        struct llama_model_params   params) {
+    ggml_time_init();
+
+    llama_model * model = new llama_model;
+
+    unsigned cur_percentage = 0;
+    if (params.progress_callback == NULL) {
+        params.progress_callback_user_data = &cur_percentage;
+        params.progress_callback = [](float progress, void * ctx) {
+            unsigned * cur_percentage_p = (unsigned *) ctx;
+            unsigned percentage = (unsigned) (100 * progress);
+            while (percentage > *cur_percentage_p) {
+                *cur_percentage_p = percentage;
+                LLAMA_LOG_INFO(".");
+                if (percentage >= 100) {
+                    LLAMA_LOG_INFO("\n");
+                }
+            }
+            return true;
+        };
+    }
+    if (params.rpc_servers != nullptr && params.rpc_servers[0] != '\0') {
+        // split the servers set them into model->rpc_servers
+        std::string servers(params.rpc_servers);
+        size_t pos = 0;
+        while ((pos = servers.find(",")) != std::string::npos) {
+            std::string server = servers.substr(0, pos);
+            model->rpc_servers.push_back(server);
+            servers.erase(0, pos + 1);
+        }
+        model->rpc_servers.push_back(servers);
+    }
+    int status = llama_model_load(path_model, *model, params);
+    GGML_ASSERT(status <= 0);
+    if (status < 0) {
+        if (status == -1) {
+            LLAMA_LOG_ERROR("%s: failed to load model\n", __func__);
+        } else if (status == -2) {
+            LLAMA_LOG_INFO("%s: cancelled model load\n", __func__);
+        }
+        delete model;
+        return nullptr;
+    }
+
+    return model;
+}
+
+void llama_free_model(struct llama_model * model) {
+    delete model;
+}
+
+struct llama_context * llama_new_context_with_model(
+                 struct llama_model * model,
+        struct llama_context_params   params) {
+
+    if (!model) {
+        LLAMA_LOG_ERROR("%s: model cannot be NULL\n", __func__);
+        return nullptr;
+    }
+
+    if (params.n_batch == 0 && params.n_ubatch == 0) {
+        LLAMA_LOG_ERROR("%s: n_batch and n_ubatch cannot both be zero\n", __func__);
+        return nullptr;
+    }
+
+    if (params.n_ctx == 0 && model->hparams.n_ctx_train == 0) {
+        LLAMA_LOG_ERROR("%s: n_ctx and model->hparams.n_ctx_train cannot both be zero\n", __func__);
+        return nullptr;
+    }
+
+    if (params.flash_attn && model->arch == LLM_ARCH_GROK) {
+        LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__);
+        params.flash_attn = false;
+    }
+
+    if (params.flash_attn && model->hparams.attn_soft_cap) {
+        LLAMA_LOG_WARN("%s: flash_attn is not compatible with attn_soft_cap - forcing off\n", __func__);
+        params.flash_attn = false;
+    }
+
+
+    if (params.flash_attn && model->hparams.n_embd_head_k != model->hparams.n_embd_head_v) {
+        LLAMA_LOG_WARN("%s: flash_attn requires n_embd_head_k == n_embd_head_v - forcing off\n", __func__);
+        params.flash_attn = false;
+    }
+
+    if (params.type_v != GGML_TYPE_F16 && !params.flash_attn) {
+        LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__);
+        return nullptr;
+    }
+
+    llama_context * ctx = new llama_context(*model);
+
+    const auto & hparams = model->hparams;
+    auto       & cparams = ctx->cparams;
+
+    cparams.n_seq_max        = std::max(1u, params.n_seq_max);
+    cparams.n_threads        = params.n_threads;
+    cparams.n_threads_batch  = params.n_threads_batch;
+    cparams.yarn_ext_factor  = params.yarn_ext_factor;
+    cparams.yarn_attn_factor = params.yarn_attn_factor;
+    cparams.yarn_beta_fast   = params.yarn_beta_fast;
+    cparams.yarn_beta_slow   = params.yarn_beta_slow;
+    cparams.defrag_thold     = params.defrag_thold;
+    cparams.embeddings       = params.embeddings;
+    cparams.offload_kqv      = params.offload_kqv;
+    cparams.flash_attn       = params.flash_attn;
+    cparams.pooling_type     = params.pooling_type;
+
+    cparams.n_ctx            = params.n_ctx           == 0    ? hparams.n_ctx_train           : params.n_ctx;
+    cparams.rope_freq_base   = params.rope_freq_base  == 0.0f ? hparams.rope_freq_base_train  : params.rope_freq_base;
+    cparams.rope_freq_scale  = params.rope_freq_scale == 0.0f ? hparams.rope_freq_scale_train : params.rope_freq_scale;
+
+    // this is necessary due to kv_self.n being padded later during inference
+    cparams.n_ctx            = GGML_PAD(cparams.n_ctx, llama_kv_cache_get_padding(cparams));
+
+    // with causal attention, the batch size is limited by the context size
+    cparams.n_batch          = hparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;
+
+    // the batch has to be at least GGML_KQ_MASK_PAD because we will be padding the KQ_mask
+    // this is required by GPU kernels in order to avoid out-of-bounds accesses (e.g. ggml_flash_attn_ext)
+    // ref: https://github.com/ggerganov/llama.cpp/pull/5021
+    if (cparams.n_batch < GGML_KQ_MASK_PAD) {
+        LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD);
+        cparams.n_batch = GGML_KQ_MASK_PAD;
+    }
+
+    cparams.n_ubatch         = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
+
+    cparams.n_ctx_orig_yarn  = params.yarn_orig_ctx    != 0 ? params.yarn_orig_ctx    :
+                               hparams.n_ctx_orig_yarn != 0 ? hparams.n_ctx_orig_yarn :
+                                                              hparams.n_ctx_train;
+
+    cparams.cb_eval           = params.cb_eval;
+    cparams.cb_eval_user_data = params.cb_eval_user_data;
+
+    auto rope_scaling_type = params.rope_scaling_type;
+    if (rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED) {
+        rope_scaling_type = hparams.rope_scaling_type_train;
+    }
+
+    if (rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_NONE) {
+        cparams.rope_freq_scale = 1.0f; // never scale if scaling type is none
+    }
+
+    if (cparams.yarn_ext_factor < 0.0f) { // negative indicates 'not set'
+        cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_YARN ? 1.0f : 0.0f;
+    }
+
+    cparams.yarn_attn_factor *= hparams.rope_attn_factor;
+
+    if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) {
+        if (hparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) {
+            cparams.pooling_type = LLAMA_POOLING_TYPE_NONE;
+        } else {
+            cparams.pooling_type = hparams.pooling_type;
+        }
+    }
+
+    if (params.attention_type == LLAMA_ATTENTION_TYPE_UNSPECIFIED) {
+        cparams.causal_attn = hparams.causal_attn;
+    } else {
+        cparams.causal_attn = params.attention_type == LLAMA_ATTENTION_TYPE_CAUSAL;
+    }
+
+    if (params.seed == LLAMA_DEFAULT_SEED) {
+        params.seed = time(NULL);
+    }
+
+    LLAMA_LOG_INFO("%s: n_ctx      = %u\n",     __func__, cparams.n_ctx);
+    LLAMA_LOG_INFO("%s: n_batch    = %u\n",     __func__, cparams.n_batch);
+    LLAMA_LOG_INFO("%s: n_ubatch   = %u\n",     __func__, cparams.n_ubatch);
+    LLAMA_LOG_INFO("%s: flash_attn = %d\n",     __func__, cparams.flash_attn);
+    LLAMA_LOG_INFO("%s: freq_base  = %.1f\n",   __func__, cparams.rope_freq_base);
+    LLAMA_LOG_INFO("%s: freq_scale = %g\n",     __func__, cparams.rope_freq_scale);
+
+    ctx->abort_callback      = params.abort_callback;
+    ctx->abort_callback_data = params.abort_callback_data;
+
+    ctx->sampling.rng = std::mt19937(params.seed);
+    ctx->logits_all   = params.logits_all;
+
+    uint32_t kv_size = cparams.n_ctx;
+    ggml_type type_k = params.type_k;
+    ggml_type type_v = params.type_v;
+
+    // Mamba only needs a constant number of KV cache cells per sequence
+    if (model->arch == LLM_ARCH_MAMBA) {
+        // Mamba needs at least as many KV cells as there are sequences kept at any time
+        kv_size = std::max((uint32_t) 1, params.n_seq_max);
+        // it's probably best to keep as much precision as possible for the states
+        type_k = GGML_TYPE_F32; // required by ggml_ssm_conv for Mamba's conv_states
+        type_v = GGML_TYPE_F32; // required by ggml_ssm_scan for Mamba's ssm_states
+    }
+
+    GGML_ASSERT(hparams.n_embd_head_k % ggml_blck_size(type_k) == 0);
+    GGML_ASSERT(hparams.n_embd_head_v % ggml_blck_size(type_v) == 0);
+
+    if (!hparams.vocab_only) {
+        // initialize backends
+#if defined(GGML_USE_METAL)
+        if (model->n_gpu_layers > 0) {
+            ctx->backend_metal = ggml_backend_metal_init();
+            if (ctx->backend_metal == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize Metal backend\n", __func__);
+                llama_free(ctx);
+                return nullptr;
+            }
+            ctx->backends.push_back(ctx->backend_metal);
+        }
+#elif defined(GGML_USE_CUDA)
+        if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) {
+            // with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used
+            ggml_backend_t backend = ggml_backend_cuda_init(model->main_gpu);
+            if (backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize CUDA%d backend\n", __func__, model->main_gpu);
+                llama_free(ctx);
+                return nullptr;
+            }
+            ctx->backends.push_back(backend);
+        } else {
+            // LLAMA_SPLIT_MODE_LAYER requires a backend for each GPU
+            for (int device = 0; device < ggml_backend_cuda_get_device_count(); ++device) {
+                ggml_backend_t backend = ggml_backend_cuda_init(device);
+                if (backend == nullptr) {
+                    LLAMA_LOG_ERROR("%s: failed to initialize CUDA%d backend\n", __func__, device);
+                    llama_free(ctx);
+                    return nullptr;
+                }
+                ctx->backends.push_back(backend);
+            }
+        }
+#elif defined(GGML_USE_VULKAN)
+        if (model->split_mode == LLAMA_SPLIT_MODE_ROW) {
+            LLAMA_LOG_ERROR("%s: Row split not supported. Failed to initialize Vulkan backend\n", __func__);
+            llama_free(ctx);
+            return nullptr;
+        }
+        if (model->split_mode == LLAMA_SPLIT_MODE_NONE) {
+            ggml_backend_t backend = ggml_backend_vk_init(model->main_gpu);
+            if (backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize Vulkan backend\n", __func__);
+                llama_free(ctx);
+                return nullptr;
+            }
+            ctx->backends.push_back(backend);
+        } else {
+            for (int device = 0; device < ggml_backend_vk_get_device_count(); ++device) {
+                ggml_backend_t backend = ggml_backend_vk_init(device);
+                if (backend == nullptr) {
+                    LLAMA_LOG_ERROR("%s: failed to initialize Vulkan%d backend\n", __func__, device);
+                    llama_free(ctx);
+                    return nullptr;
+                }
+                ctx->backends.push_back(backend);
+            }
+        }
+#elif defined(GGML_USE_SYCL)
+        // with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used
+        if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) {
+            ggml_backend_t backend = ggml_backend_sycl_init(model->main_gpu);
+            if (backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d backend\n", __func__, model->main_gpu);
+                llama_free(ctx);
+                return nullptr;
+            }
+            ctx->backends.push_back(backend);
+        } else {
+            // LLAMA_SPLIT_LAYER requires a backend for each GPU
+            for (int i = 0; i < ggml_backend_sycl_get_device_count(); ++i) {
+                ggml_backend_t backend = ggml_backend_sycl_init(i);
+                if (backend == nullptr) {
+                    LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d for No.%d backend\n", __func__, i, i);
+                    llama_free(ctx);
+                    return nullptr;
+                }
+                ctx->backends.push_back(backend);
+            }
+        }
+#elif defined(GGML_USE_KOMPUTE)
+        if (model->n_gpu_layers > 0) {
+            auto * backend = ggml_backend_kompute_init(model->main_gpu);
+            if (backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize Kompute backend\n", __func__);
+                llama_free(ctx);
+                return nullptr;
+            }
+            ctx->backends.push_back(backend);
+        }
+#elif defined(GGML_USE_CANN)
+    // with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used
+    // TODO: ggml_backend_cann is not support split tensor now, just leave code here.
+    if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) {
+        ggml_backend_t backend = ggml_backend_cann_init(model->main_gpu);
+        if (backend == nullptr) {
+            LLAMA_LOG_ERROR("%s: failed to initialize CANN%d backend\n", __func__, model->main_gpu);
+            llama_free(ctx);
+            return nullptr;
+        }
+        ctx->backends.push_back(backend);
+    } else {
+        // LLAMA_SPLIT_MODE_LAYER requires a backend for each GPU
+        // TODO: currently, CANN can't use multi-gpus, just leave code here for further cann version.
+        for (int32_t device = 0; device < ggml_backend_cann_get_device_count(); ++device) {
+            ggml_backend_t backend = ggml_backend_cann_init(device);
+            if (backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize CANN%d backend\n", __func__, device);
+                llama_free(ctx);
+                return nullptr;
+            }
+            ctx->backends.push_back(backend);
+        }
+    }
+#endif
+
+#ifdef GGML_USE_BLAS
+        ctx->backend_blas = ggml_backend_blas_init();
+        if (ctx->backend_blas == nullptr) {
+            LLAMA_LOG_WARN("%s: failed to initialize BLAS backend\n", __func__);
+        } else {
+            ctx->backends.push_back(ctx->backend_blas);
+        }
+#endif
+
+#if defined(GGML_USE_RPC)
+        if (model->n_gpu_layers > 0) {
+            for (const auto & endpoint : model->rpc_servers) {
+                ggml_backend_t backend = ggml_backend_rpc_init(endpoint.c_str());
+                if (backend == nullptr) {
+                    LLAMA_LOG_ERROR("%s: failed to initialize RPC to '%s'\n", __func__, endpoint.c_str());
+                    llama_free(ctx);
+                    return nullptr;
+                }
+                ctx->backends.push_back(backend);
+            }
+        }
+#endif
+        ctx->backend_cpu = ggml_backend_cpu_init();
+        if (ctx->backend_cpu == nullptr) {
+            LLAMA_LOG_ERROR("%s: failed to initialize CPU backend\n", __func__);
+            llama_free(ctx);
+            return nullptr;
+        }
+        ctx->backends.push_back(ctx->backend_cpu);
+
+        if (!llama_kv_cache_init(ctx->kv_self, ctx, type_k, type_v, kv_size, cparams.offload_kqv)) {
+            LLAMA_LOG_ERROR("%s: llama_kv_cache_init() failed for self-attention cache\n", __func__);
+            llama_free(ctx);
+            return nullptr;
+        }
+
+        {
+            size_t memory_size_k = 0;
+            size_t memory_size_v = 0;
+
+            for (auto & k : ctx->kv_self.k_l) {
+                memory_size_k += ggml_nbytes(k);
+            }
+
+            for (auto & v : ctx->kv_self.v_l) {
+                memory_size_v += ggml_nbytes(v);
+            }
+
+            LLAMA_LOG_INFO("%s: KV self size  = %7.2f MiB, K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__,
+                (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f),
+                ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f),
+                ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f));
+        }
+
+        // graph outputs buffer
+        {
+            // resized during inference when a batch uses more outputs
+            if (llama_output_reserve(*ctx, params.n_seq_max) < params.n_seq_max) {
+                LLAMA_LOG_ERROR("%s: failed to reserve initial output buffer\n", __func__);
+                llama_free(ctx);
+                return nullptr;
+            }
+
+            LLAMA_LOG_INFO("%s: %10s  output buffer size = %8.2f MiB\n", __func__,
+                    ggml_backend_buffer_name(ctx->buf_output),
+                    ggml_backend_buffer_get_size(ctx->buf_output) / 1024.0 / 1024.0);
+        }
+
+        // scheduler and compute buffers
+        {
+            // buffer types used for the compute buffer of each backend
+            std::vector backend_buft;
+            for (auto * backend : ctx->backends) {
+                if (ggml_backend_is_cpu(backend)) {
+                    // use host buffers for the CPU backend compute buffer
+                    backend_buft.push_back(llama_default_buffer_type_cpu(true));
+                } else {
+                    backend_buft.push_back(ggml_backend_get_default_buffer_type(backend));
+                }
+            }
+
+            // buffer used to store the computation graph and the tensor meta data
+            ctx->buf_compute_meta.resize(ggml_tensor_overhead()*LLAMA_MAX_NODES + ggml_graph_overhead_custom(LLAMA_MAX_NODES, false));
+
+            // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary
+            bool pipeline_parallel =
+                llama_get_device_count(*model) > 1 &&
+                model->n_gpu_layers > (int)model->hparams.n_layer &&
+                model->split_mode == LLAMA_SPLIT_MODE_LAYER &&
+                params.offload_kqv;
+#ifndef GGML_USE_CUDA
+            // pipeline parallelism requires support for async compute and events
+            // currently this is only implemented in the CUDA backend
+            pipeline_parallel = false;
+#endif
+            ctx->sched = ggml_backend_sched_new(ctx->backends.data(), backend_buft.data(), ctx->backends.size(), LLAMA_MAX_NODES, pipeline_parallel);
+
+            if (pipeline_parallel) {
+                LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(ctx->sched));
+            }
+
+            // build worst-case graph
+            int n_tokens = (int)std::min(cparams.n_ctx, cparams.n_ubatch);
+            int n_past = cparams.n_ctx - n_tokens;
+            llama_token token = llama_token_bos(&ctx->model); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
+            ggml_cgraph * gf = llama_build_graph(*ctx, llama_batch_get_one(&token, n_tokens, n_past, 0), true);
+
+            // initialize scheduler with the worst-case graph
+            if (!ggml_backend_sched_reserve(ctx->sched, gf)) {
+                LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
+                llama_free(ctx);
+                return nullptr;
+            }
+
+            for (size_t i = 0; i < ctx->backends.size(); i++) {
+                ggml_backend_t backend = ctx->backends[i];
+                ggml_backend_buffer_type_t buft = backend_buft[i];
+                size_t size = ggml_backend_sched_get_buffer_size(ctx->sched, backend);
+                if (size > 1) {
+                    LLAMA_LOG_INFO("%s: %10s compute buffer size = %8.2f MiB\n", __func__,
+                            ggml_backend_buft_name(buft),
+                            size / 1024.0 / 1024.0);
+                }
+            }
+
+            // note: the number of splits during measure is higher than during inference due to the kv shift
+            int n_splits = ggml_backend_sched_get_n_splits(ctx->sched);
+            LLAMA_LOG_INFO("%s: graph nodes  = %d\n", __func__, gf->n_nodes);
+            LLAMA_LOG_INFO("%s: graph splits = %d\n", __func__, n_splits);
+        }
+    }
+
+    return ctx;
+}
+
+void llama_free(struct llama_context * ctx) {
+    delete ctx;
+}
+
+const struct llama_model * llama_get_model(const struct llama_context * ctx) {
+    return &ctx->model;
+}
+
+const struct llama_vocab * llama_get_vocab(const struct llama_context * ctx) {
+    return &ctx->model.vocab;
+}
+
+uint32_t llama_n_ctx(const struct llama_context * ctx) {
+    return ctx->cparams.n_ctx;
+}
+
+uint32_t llama_n_batch(const struct llama_context * ctx) {
+    return ctx->cparams.n_batch;
+}
+
+uint32_t llama_n_ubatch(const struct llama_context * ctx) {
+    return ctx->cparams.n_ubatch;
+}
+
+uint32_t llama_n_seq_max(const struct llama_context * ctx) {
+    return ctx->kv_self.size;
+}
+
+enum llama_vocab_type llama_vocab_type(const struct llama_model * model) {
+    return model->vocab.type;
+}
+
+enum llama_rope_type llama_rope_type(const struct llama_model * model) {
+    switch (model->arch) {
+        // these models do not use RoPE
+        case LLM_ARCH_GPT2:
+        case LLM_ARCH_GPTJ:
+        case LLM_ARCH_MPT:
+        case LLM_ARCH_REFACT:
+        case LLM_ARCH_BLOOM:
+        case LLM_ARCH_MAMBA:
+        case LLM_ARCH_JINA_BERT_V2:
+        case LLM_ARCH_T5:
+        case LLM_ARCH_JAIS:
+            return LLAMA_ROPE_TYPE_NONE;
+
+        // use what we call a normal RoPE, operating on pairs of consecutive head values
+        case LLM_ARCH_LLAMA:
+        case LLM_ARCH_BAICHUAN:
+        case LLM_ARCH_STARCODER:
+        case LLM_ARCH_PLAMO:
+        case LLM_ARCH_ORION:
+        case LLM_ARCH_INTERNLM2:
+        case LLM_ARCH_MINICPM:
+        case LLM_ARCH_XVERSE:
+        case LLM_ARCH_COMMAND_R:
+        case LLM_ARCH_OLMO:
+        case LLM_ARCH_ARCTIC:
+        case LLM_ARCH_DEEPSEEK2:
+        case LLM_ARCH_CHATGLM:
+            return LLAMA_ROPE_TYPE_NORM;
+
+        // the pairs of head values are offset by n_rot/2
+        case LLM_ARCH_FALCON:
+        case LLM_ARCH_GROK:
+        case LLM_ARCH_DBRX:
+        case LLM_ARCH_BERT:
+        case LLM_ARCH_NOMIC_BERT:
+        case LLM_ARCH_STABLELM:
+        case LLM_ARCH_BITNET:
+        case LLM_ARCH_QWEN:
+        case LLM_ARCH_QWEN2:
+        case LLM_ARCH_QWEN2MOE:
+        case LLM_ARCH_PHI2:
+        case LLM_ARCH_PHI3:
+        case LLM_ARCH_GEMMA:
+        case LLM_ARCH_GEMMA2:
+        case LLM_ARCH_STARCODER2:
+        case LLM_ARCH_OPENELM:
+        case LLM_ARCH_GPTNEOX:
+        case LLM_ARCH_CODESHELL:
+            return LLAMA_ROPE_TYPE_NEOX;
+
+        // all model arches should be listed explicitly here
+        case LLM_ARCH_UNKNOWN:
+            GGML_ASSERT(false && "unknown architecture");
+            break;
+    }
+
+    return LLAMA_ROPE_TYPE_NONE;
+}
+
+enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx) {
+    return ctx->cparams.pooling_type;
+}
+
+int32_t llama_n_vocab(const struct llama_model * model) {
+    return model->hparams.n_vocab;
+}
+
+int32_t llama_n_ctx_train(const struct llama_model * model) {
+    return model->hparams.n_ctx_train;
+}
+
+int32_t llama_n_embd(const struct llama_model * model) {
+    return model->hparams.n_embd;
+}
+
+int32_t llama_n_layer(const struct llama_model * model) {
+    return model->hparams.n_layer;
+}
+
+float llama_rope_freq_scale_train(const struct llama_model * model) {
+    return model->hparams.rope_freq_scale_train;
+}
+
+int32_t llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size) {
+    const auto & it = model->gguf_kv.find(key);
+    if (it == model->gguf_kv.end()) {
+        if (buf_size > 0) {
+            buf[0] = '\0';
+        }
+        return -1;
+    }
+    return snprintf(buf, buf_size, "%s", it->second.c_str());
+}
+
+int32_t llama_model_meta_count(const struct llama_model * model) {
+    return (int)model->gguf_kv.size();
+}
+
+int32_t llama_model_meta_key_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size) {
+    if (i < 0 || i >= (int)model->gguf_kv.size()) {
+        if (buf_size > 0) {
+            buf[0] = '\0';
+        }
+        return -1;
+    }
+    auto it = model->gguf_kv.begin();
+    std::advance(it, i);
+    return snprintf(buf, buf_size, "%s", it->first.c_str());
+}
+
+int32_t llama_model_meta_val_str_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size) {
+    if (i < 0 || i >= (int)model->gguf_kv.size()) {
+        if (buf_size > 0) {
+            buf[0] = '\0';
+        }
+        return -1;
+    }
+    auto it = model->gguf_kv.begin();
+    std::advance(it, i);
+    return snprintf(buf, buf_size, "%s", it->second.c_str());
+}
+
+int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size) {
+    return snprintf(buf, buf_size, "%s %s %s",
+            llama_model_arch_name(model->arch),
+            llama_model_type_name(model->type),
+            llama_model_ftype_name(model->ftype).c_str());
+}
+
+uint64_t llama_model_size(const struct llama_model * model) {
+    uint64_t size = 0;
+    for (const auto & it : model->tensors_by_name) {
+        size += ggml_nbytes(it.second);
+    }
+    return size;
+}
+
+uint64_t llama_model_n_params(const struct llama_model * model) {
+    uint64_t nparams = 0;
+    for (const auto & it : model->tensors_by_name) {
+        nparams += ggml_nelements(it.second);
+    }
+    return nparams;
+}
+
+struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name) {
+    auto it = std::find_if(model->tensors_by_name.begin(), model->tensors_by_name.end(),
+            [name](const std::pair & it) {
+                return it.first == name;
+            });
+    if (it == model->tensors_by_name.end()) {
+        return nullptr;
+    }
+    return it->second;
+}
+
+bool llama_model_has_encoder(const struct llama_model * model) {
+    switch (model->arch) {
+        case LLM_ARCH_T5: return true;
+        default:          return false;
+    }
+}
+
+llama_token llama_model_decoder_start_token(const struct llama_model * model) {
+    return model->hparams.dec_start_token_id;
+}
+
+uint32_t llama_model_quantize(
+        const char * fname_inp,
+        const char * fname_out,
+        const llama_model_quantize_params * params) {
+    try {
+        llama_model_quantize_internal(fname_inp, fname_out, params);
+        return 0;
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: failed to quantize: %s\n", __func__, err.what());
+        return 1;
+    }
+}
+
+struct llama_lora_adapter * llama_lora_adapter_init(struct llama_model * model, const char * path_lora) {
+    try {
+        struct llama_lora_adapter * adapter = new llama_lora_adapter(model);
+        llama_lora_adapter_init_internal(model, path_lora, *adapter);
+        return adapter;
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
+        return nullptr;
+    }
+}
+
+static bool llama_control_vector_init(struct llama_control_vector & cvec, const llama_model & model) {
+    GGML_ASSERT(cvec.tensors.empty());
+    GGML_ASSERT(cvec.ctxs.empty());
+    GGML_ASSERT(cvec.bufs.empty());
+
+    // count layer buffer types
+    std::map buft_layer_count;
+    for (int64_t i = 0; i < model.hparams.n_layer; i++) {
+        buft_layer_count[model.buft_layer[i].buft]++;
+    }
+
+    // allocate contexts
+    std::map ctx_map;
+    for (auto & it : buft_layer_count) {
+        int n_layers = it.second;
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ n_layers * ggml_tensor_overhead(),
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ true,
+        };
+        ggml_context * ctx = ggml_init(params);
+        if (!ctx) {
+            LLAMA_LOG_ERROR("%s: failed to allocate context for control vector\n", __func__);
+            return 1;
+        }
+        ctx_map[it.first] = ctx;
+    }
+
+    // make tensors
+    cvec.tensors.reserve(model.hparams.n_layer);
+    cvec.tensors.push_back(nullptr); // there's never a tensor for layer 0
+    for (size_t il = 1; il < model.hparams.n_layer; il++) {
+        struct ggml_context * ctx = ctx_map.at(model.buft_layer[il].buft);
+        ggml_tensor * tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, model.hparams.n_embd);
+        cvec.tensors.push_back(tensor);
+    }
+
+    // allocate tensors / buffers and zero
+    cvec.ctxs.reserve(ctx_map.size());
+    cvec.bufs.reserve(ctx_map.size());
+    for (auto it : ctx_map) {
+        ggml_backend_buffer_type_t buft = it.first;
+        ggml_context * ctx = it.second;
+        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+        if (!buf) {
+            LLAMA_LOG_ERROR("%s: failed to allocate buffer for control vector\n", __func__);
+            return false;
+        }
+        ggml_backend_buffer_clear(buf, 0);
+        cvec.ctxs.push_back(ctx);
+        cvec.bufs.push_back(buf);
+    }
+
+    return true;
+}
+
+int32_t llama_control_vector_apply(struct llama_context * lctx, const float * data, size_t len, int32_t n_embd, int32_t il_start, int32_t il_end) {
+    const llama_model & model = lctx->model;
+    llama_control_vector & cvec = lctx->cvec;
+
+    if (data == nullptr) {
+        // disable the current control vector (but leave allocated for later)
+        cvec.layer_start = -1;
+        cvec.layer_end   = -1;
+        return 0;
+    }
+
+    if (n_embd != (int) model.hparams.n_embd) {
+        LLAMA_LOG_ERROR("%s: control vector n_embd does not match model\n", __func__);
+        return 1;
+    }
+
+    if (cvec.tensors.empty()) {
+        if (!llama_control_vector_init(cvec, model)) {
+            return 1;
+        }
+    }
+
+    cvec.layer_start = il_start;
+    cvec.layer_end   = il_end;
+
+    for (size_t il = 1; il < model.hparams.n_layer; il++) {
+        assert(cvec.tensors[il] != nullptr);
+
+        const size_t off = n_embd * (il - 1); // buffer doesn't have data for layer 0, since it's never present
+        if (off + n_embd <= len) {
+            ggml_backend_tensor_set(cvec.tensors[il], data + off, 0, n_embd * ggml_element_size(cvec.tensors[il]));
+        }
+    }
+
+    return 0;
+}
+
+struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_seq_max) {
+    struct llama_kv_cache_view result = {
+        /*.n_cells            = */ 0,
+        /*.n_seq_max          = */ n_seq_max,
+        /*.token_count        = */ 0,
+        /*.used_cells         = */ llama_get_kv_cache_used_cells(ctx),
+        /*.max_contiguous     = */ 0,
+        /*.max_contiguous_idx = */ -1,
+        /*.cells              = */ nullptr,
+        /*.cells_sequences    = */ nullptr,
+    };
+    return result;
+}
+
+void llama_kv_cache_view_free(struct llama_kv_cache_view * view) {
+    if (view->cells != nullptr) {
+        free(view->cells);
+        view->cells = nullptr;
+    }
+    if (view->cells_sequences != nullptr) {
+        free(view->cells_sequences);
+        view->cells_sequences = nullptr;
+    }
+}
+
+void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view) {
+    if (uint32_t(view->n_cells) < ctx->kv_self.size || view->cells == nullptr) {
+        view->n_cells = int32_t(ctx->kv_self.size);
+        void * p = realloc(view->cells, sizeof(struct llama_kv_cache_view_cell) * view->n_cells);
+        GGML_ASSERT(p != nullptr && "Failed to alloc kv_cache_view cells");
+        view->cells = (struct llama_kv_cache_view_cell *)p;
+        p = realloc(view->cells_sequences, sizeof(llama_seq_id) * view->n_seq_max * view->n_cells);
+        GGML_ASSERT(p != nullptr && "Failed to alloc kv_cache_view cells sequences");
+        view->cells_sequences = (llama_seq_id *)p;
+    }
+
+    const std::vector & kv_cells = ctx->kv_self.cells;
+    llama_kv_cache_view_cell * c_curr = view->cells;
+    llama_seq_id * cs_curr = view->cells_sequences;
+    int32_t used_cells = 0;
+    int32_t token_count = 0;
+    int32_t curr_contig_idx = -1;
+    uint32_t max_contig = 0;
+    int32_t max_contig_idx = -1;
+
+    for (int32_t i = 0; i < int32_t(ctx->kv_self.size); i++, c_curr++, cs_curr += view->n_seq_max) {
+        const size_t curr_size = kv_cells[i].seq_id.size();
+        token_count += curr_size;
+        c_curr->pos = kv_cells[i].pos + kv_cells[i].delta;
+
+        if (curr_size > 0) {
+            if (curr_contig_idx >= 0 && uint32_t(i - curr_contig_idx) > max_contig) {
+                max_contig = i - curr_contig_idx;
+                max_contig_idx = curr_contig_idx;
+            }
+            curr_contig_idx = -1;
+        } else if (curr_contig_idx < 0) {
+            curr_contig_idx = i;
+        }
+
+        int seq_idx = 0;
+        for (const llama_seq_id it : kv_cells[i].seq_id) {
+            if (seq_idx >= view->n_seq_max) {
+                break;
+            }
+            cs_curr[seq_idx] = it;
+            seq_idx++;
+        }
+        if (seq_idx != 0) {
+            used_cells++;
+        }
+        for (; seq_idx < view->n_seq_max; seq_idx++) {
+            cs_curr[seq_idx] = -1;
+        }
+    }
+    if (curr_contig_idx >= 0 && kv_cells.size() - curr_contig_idx > max_contig) {
+        max_contig_idx = curr_contig_idx;
+        max_contig = kv_cells.size() - curr_contig_idx;
+    }
+    view->max_contiguous = max_contig;
+    view->max_contiguous_idx = max_contig_idx;
+    view->token_count = token_count;
+    view->used_cells = used_cells;
+    if (uint32_t(used_cells) != ctx->kv_self.used) {
+        LLAMA_LOG_ERROR("%s: used cells mismatch. kv_cache says %d but we calculated %d\n",
+            __func__, ctx->kv_self.used, used_cells);
+    }
+}
+
+int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx) {
+    int result = 0;
+
+    for (uint32_t i = 0; i < ctx->kv_self.size; i++) {
+        result += ctx->kv_self.cells[i].seq_id.size();
+    }
+
+    return result;
+}
+
+int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx) {
+    return ctx->kv_self.used;
+}
+
+void llama_kv_cache_clear(struct llama_context * ctx) {
+    llama_kv_cache_clear(ctx->kv_self);
+}
+
+bool llama_kv_cache_seq_rm(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+    return llama_kv_cache_seq_rm(ctx->kv_self, seq_id, p0, p1);
+}
+
+void llama_kv_cache_seq_cp(struct llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+    if (seq_id_src == seq_id_dst) {
+        return;
+    }
+    llama_kv_cache_seq_cp(ctx->kv_self, seq_id_src, seq_id_dst, p0, p1);
+}
+
+void llama_kv_cache_seq_keep(struct llama_context * ctx, llama_seq_id seq_id) {
+    llama_kv_cache_seq_keep(ctx->kv_self, seq_id);
+}
+
+void llama_kv_cache_seq_add(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) {
+    if (delta == 0) {
+        return;
+    }
+
+    llama_kv_cache_seq_add(ctx->kv_self, seq_id, p0, p1, delta);
+}
+
+void llama_kv_cache_seq_div(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+    if (d == 1) {
+        return;
+    }
+
+    llama_kv_cache_seq_div(ctx->kv_self, seq_id, p0, p1, d);
+}
+
+llama_pos llama_kv_cache_seq_pos_max(struct llama_context * ctx, llama_seq_id seq_id) {
+    return llama_kv_cache_seq_pos_max(ctx->kv_self, seq_id);
+}
+
+void llama_kv_cache_defrag(struct llama_context * ctx) {
+    llama_kv_cache_defrag(ctx->kv_self);
+}
+
+void llama_kv_cache_update(struct llama_context * ctx) {
+    llama_kv_cache_update_internal(*ctx);
+}
+
+// deprecated
+size_t llama_get_state_size(const struct llama_context * ctx) {
+    return llama_state_get_size(ctx);
+}
+
+// deprecated
+size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
+    return llama_state_get_data(ctx, dst);
+}
+
+// deprecated
+size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) {
+    return llama_state_set_data(ctx, src);
+}
+
+// deprecated
+bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
+    return llama_state_load_file(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out);
+}
+
+// deprecated
+bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) {
+    return llama_state_save_file(ctx, path_session, tokens, n_token_count);
+}
+
+// Returns the *maximum* size of the state
+size_t llama_state_get_size(const struct llama_context * ctx) {
+    const auto & cparams = ctx->cparams;
+    const auto & hparams = ctx->model.hparams;
+
+    // we don't know size of rng until we actually serialize it. so reserve more than enough memory for its serialized state.
+    // for reference, std::mt19937(1337) serializes to 6701 bytes.
+    const size_t s_rng_size        = sizeof(size_t);
+    const size_t s_rng             = LLAMA_MAX_RNG_STATE;
+    const size_t s_n_outputs       = sizeof(size_t);
+    // assume worst case for outputs although only currently set ones are serialized
+    const size_t s_output_pos      = ctx->cparams.n_batch * sizeof(int32_t);
+    const size_t s_logits_size     = sizeof(size_t);
+    const size_t s_logits          = ctx->logits_size ? cparams.n_batch * hparams.n_vocab * sizeof(float) : 0;
+    const size_t s_embedding_size  = sizeof(size_t);
+    const size_t s_embedding       = ctx->embd_size   ? cparams.n_batch * hparams.n_embd  * sizeof(float) : 0;
+    const size_t s_kv_buf_size     = sizeof(size_t);
+    const size_t s_kv_head         = sizeof(uint32_t);
+    const size_t s_kv_size         = sizeof(uint32_t);
+    const size_t s_kv_used         = sizeof(uint32_t);
+    const size_t s_v_trans         = sizeof(uint32_t);
+    const size_t s_kv              = ctx->kv_self.total_size();
+    const size_t s_kv_cell         = sizeof(llama_pos) + sizeof(size_t) + cparams.n_seq_max*sizeof(llama_seq_id);
+    const size_t s_kv_cells        = ctx->kv_self.size * s_kv_cell;
+
+    const size_t s_total = (
+        + s_rng_size
+        + s_rng
+        + s_n_outputs
+        + s_output_pos
+        + s_logits_size
+        + s_logits
+        + s_embedding_size
+        + s_embedding
+        + s_kv_buf_size
+        + s_kv_head
+        + s_kv_size
+        + s_kv_used
+        + s_v_trans
+        + s_kv
+        + s_kv_cells
+    );
+
+    // on session change it is very likely that the state size has changed - so we need to update this function
+    static_assert(LLAMA_SESSION_VERSION == 7, "So you just bumped the session version - good. But did you remember to update llama_state_get_size?");
+
+    return s_total;
+}
+
+// llama_context_data
+struct llama_data_context {
+    virtual void write(const void * src, size_t size) = 0;
+    virtual size_t get_size_written() = 0;
+    virtual ~llama_data_context() = default;
+};
+
+struct llama_data_buffer_context : llama_data_context {
+    uint8_t * ptr;
+    size_t size_written = 0;
+
+    llama_data_buffer_context(uint8_t * p) : ptr(p) {}
+
+    void write(const void * src, size_t size) override {
+        memcpy(ptr, src, size);
+        ptr += size;
+        size_written += size;
+    }
+
+    size_t get_size_written() override {
+        return size_written;
+    }
+};
+
+struct llama_data_file_context : llama_data_context {
+    llama_file * file;
+    size_t size_written = 0;
+
+    llama_data_file_context(llama_file * f) : file(f) {}
+
+    void write(const void * src, size_t size) override {
+        file->write_raw(src, size);
+        size_written += size;
+    }
+
+    size_t get_size_written() override {
+        return size_written;
+    }
+};
+
+/** copy state data into either a buffer or file depending on the passed in context
+ *
+ * file context:
+ * llama_file file("/path", "wb");
+ * llama_data_file_context data_ctx(&file);
+ * llama_state_get_data(ctx, &data_ctx);
+ *
+ * buffer context:
+ * std::vector buf(max_size, 0);
+ * llama_data_buffer_context data_ctx(&buf.data());
+ * llama_state_get_data(ctx, &data_ctx);
+ *
+*/
+static void llama_state_get_data_internal(struct llama_context * ctx, llama_data_context * data_ctx) {
+    llama_synchronize(ctx);
+
+    // copy rng
+    {
+        std::ostringstream rng_ss;
+        rng_ss << ctx->sampling.rng;
+
+        const std::string & rng_str  = rng_ss.str();
+        const size_t        rng_size = rng_str.size();
+
+        GGML_ASSERT(rng_size <= LLAMA_MAX_RNG_STATE);
+
+        data_ctx->write(&rng_size,      sizeof(rng_size));
+        data_ctx->write(rng_str.data(), rng_size);
+    }
+
+    // copy outputs
+    {
+        // Can't use ctx->n_outputs because it's not for the
+        // entire last batch when n_ubatch is smaller than n_batch
+        size_t n_outputs = 0;
+
+        // copy output ids
+        {
+            std::vector output_pos;
+
+            const size_t    n_batch = ctx->cparams.n_batch;
+            const auto & output_ids = ctx->output_ids;
+
+            output_pos.resize(ctx->output_size);
+
+            // build a more compact representation of the output ids
+            for (size_t i = 0; i < n_batch; ++i) {
+                // map an output id to a position in the batch
+                int32_t pos = output_ids[i];
+                if (pos >= 0) {
+                    if ((size_t) pos >= n_outputs) {
+                        n_outputs = pos + 1;
+                    }
+                    GGML_ASSERT((size_t) pos < ctx->output_size);
+                    output_pos[pos] = i;
+                }
+            }
+
+            data_ctx->write(&n_outputs, sizeof(n_outputs));
+
+            if (n_outputs) {
+                data_ctx->write(output_pos.data(), n_outputs * sizeof(int32_t));
+            }
+        }
+
+        // copy logits
+        {
+            const size_t logits_size = std::min(ctx->logits_size, n_outputs * ctx->model.hparams.n_vocab);
+
+            data_ctx->write(&logits_size, sizeof(logits_size));
+
+            if (logits_size) {
+                data_ctx->write(ctx->logits, logits_size * sizeof(float));
+            }
+        }
+
+        // copy embeddings
+        {
+            const size_t embeddings_size = std::min(ctx->embd_size, n_outputs * ctx->model.hparams.n_embd);
+
+            data_ctx->write(&embeddings_size, sizeof(embeddings_size));
+
+            if (embeddings_size) {
+                data_ctx->write(ctx->embd, embeddings_size * sizeof(float));
+            }
+        }
+    }
+
+    // copy kv cache
+    {
+        const auto & kv_self = ctx->kv_self;
+        const auto & hparams = ctx->model.hparams;
+
+        const uint32_t n_layer      = hparams.n_layer;
+
+        // NOTE: kv_size and kv_buf_size are mostly used for sanity checks
+        const uint32_t kv_head     = llama_kv_cache_cell_max(kv_self);
+        const uint32_t kv_size     = kv_self.size;
+        const size_t   kv_buf_size = kv_self.total_size() / (kv_size ? kv_size : 1) * kv_head;
+        const uint32_t kv_used     = kv_self.used;
+        const uint32_t v_trans     = kv_self.v_trans ? 1 : 0;
+
+        data_ctx->write(&kv_buf_size, sizeof(kv_buf_size));
+        data_ctx->write(&kv_head,     sizeof(kv_head));
+        data_ctx->write(&kv_size,     sizeof(kv_size));
+        data_ctx->write(&kv_used,     sizeof(kv_used));
+        data_ctx->write(&v_trans,     sizeof(v_trans));
+
+        if (kv_buf_size) {
+            const size_t pre_kv_buf_size = data_ctx->get_size_written();
+
+            std::vector tmp_buf;
+            for (int il = 0; il < (int) n_layer; ++il) {
+                const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
+                const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+
+                const size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_head);
+
+                tmp_buf.resize(k_size);
+                ggml_backend_tensor_get(kv_self.k_l[il], tmp_buf.data(), 0, tmp_buf.size());
+                data_ctx->write(tmp_buf.data(), tmp_buf.size());
+
+                if (kv_self.recurrent || !kv_self.v_trans) {
+                    // v is contiguous for recurrent models
+                    // TODO: use other tensors for state models than k and v
+                    const size_t v_size = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*kv_head);
+
+                    tmp_buf.resize(v_size);
+                    ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), 0, tmp_buf.size());
+                    data_ctx->write(tmp_buf.data(), tmp_buf.size());
+                    continue;
+                }
+
+                // v is not contiguous, copy row by row
+                const size_t v_row_size   = ggml_row_size(kv_self.v_l[il]->type, kv_head);
+                const size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, kv_size);
+
+                tmp_buf.resize(v_row_size);
+                for (int ir = 0; ir < (int) n_embd_v_gqa; ++ir) {
+                    ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), ir*v_row_stride, tmp_buf.size());
+                    data_ctx->write(tmp_buf.data(), tmp_buf.size());
+                }
+            }
+            GGML_ASSERT(kv_buf_size == data_ctx->get_size_written() - pre_kv_buf_size);
+        }
+
+        for (uint32_t i = 0; i < kv_head; ++i) {
+            const auto & cell = kv_self.cells[i];
+
+            const llama_pos pos         = cell.pos;
+            const size_t    seq_id_size = cell.seq_id.size();
+
+            data_ctx->write(&pos,         sizeof(pos));
+            data_ctx->write(&seq_id_size, sizeof(seq_id_size));
+
+            for (auto seq_id : cell.seq_id) {
+                data_ctx->write(&seq_id, sizeof(seq_id));
+            }
+        }
+    }
+}
+
+size_t llama_state_get_data(struct llama_context * ctx, uint8_t * dst) {
+    llama_data_buffer_context data_ctx(dst);
+    llama_state_get_data_internal(ctx, &data_ctx);
+
+    return data_ctx.get_size_written();
+}
+
+// Sets the state reading from the specified source address
+size_t llama_state_set_data(struct llama_context * ctx, const uint8_t * src) {
+    llama_synchronize(ctx);
+
+    const uint8_t * inp = src;
+
+    // set rng
+    {
+        size_t rng_size;
+        memcpy(&rng_size, inp, sizeof(rng_size)); inp += sizeof(rng_size);
+
+        GGML_ASSERT(rng_size <= LLAMA_MAX_RNG_STATE);
+
+        std::string rng_str((const char *)inp, rng_size); inp += rng_size;
+
+        std::istringstream rng_ss(rng_str);
+        rng_ss >> ctx->sampling.rng;
+
+        GGML_ASSERT(!rng_ss.fail());
+    }
+
+    // set output ids
+    {
+        size_t n_outputs;
+        std::vector output_pos;
+
+        memcpy(&n_outputs, inp, sizeof(n_outputs)); inp += sizeof(n_outputs);
+
+        GGML_ASSERT(n_outputs <= llama_output_reserve(*ctx, n_outputs));
+
+        if (n_outputs) {
+            output_pos.resize(n_outputs);
+            memcpy(output_pos.data(), inp, n_outputs * sizeof(int32_t));
+            inp += n_outputs * sizeof(int32_t);
+
+            for (int32_t i = 0; i < (int32_t) output_pos.size(); ++i) {
+                int32_t id = output_pos[i];
+                GGML_ASSERT((uint32_t) id < ctx->cparams.n_batch);
+                ctx->output_ids[id] = i;
+            }
+
+            ctx->n_outputs = n_outputs;
+        }
+    }
+
+    // set logits
+    {
+        size_t logits_size;
+
+        memcpy(&logits_size, inp, sizeof(logits_size)); inp += sizeof(logits_size);
+
+        GGML_ASSERT(ctx->logits_size >= logits_size);
+
+        if (logits_size) {
+            memcpy(ctx->logits, inp, logits_size * sizeof(float));
+            inp += logits_size * sizeof(float);
+        }
+    }
+
+    // set embeddings
+    {
+        size_t embeddings_size;
+
+        memcpy(&embeddings_size, inp, sizeof(embeddings_size)); inp += sizeof(embeddings_size);
+
+        GGML_ASSERT(ctx->embd_size >= embeddings_size);
+
+        if (embeddings_size) {
+            memcpy(ctx->embd, inp, embeddings_size * sizeof(float));
+            inp += embeddings_size * sizeof(float);
+        }
+    }
+
+    // set kv cache
+    {
+        const auto & kv_self = ctx->kv_self;
+        const auto & hparams = ctx->model.hparams;
+
+        const uint32_t n_layer      = hparams.n_layer;
+
+        size_t   kv_buf_size;
+        uint32_t kv_head;
+        uint32_t kv_size;
+        uint32_t kv_used;
+        uint32_t v_trans;
+
+        memcpy(&kv_buf_size, inp, sizeof(kv_buf_size)); inp += sizeof(kv_buf_size);
+        memcpy(&kv_head,     inp, sizeof(kv_head));     inp += sizeof(kv_head);
+        memcpy(&kv_size,     inp, sizeof(kv_size));     inp += sizeof(kv_size);
+        memcpy(&kv_used,     inp, sizeof(kv_used));     inp += sizeof(kv_used);
+        memcpy(&v_trans,     inp, sizeof(v_trans));     inp += sizeof(v_trans);
+
+        GGML_ASSERT(kv_self.v_trans == (bool) v_trans); // incompatible V transposition
+
+        if (kv_self.size != kv_size) {
+            // the KV cache needs to be big enough to load all the KV cells from the saved state
+            GGML_ASSERT(kv_self.size >= kv_head);
+
+            LLAMA_LOG_INFO("%s: state contains %d KV cells, was saved with kv_size=%d, but is loaded with kv_size=%d (fine, but different)\n",
+                __func__, kv_head, kv_size, kv_self.size);
+        }
+
+        llama_kv_cache_clear(ctx);
+
+        if (kv_buf_size) {
+            const size_t pre_kv_buf_size = inp - src;
+
+            GGML_ASSERT(kv_self.total_size() >= kv_buf_size);
+
+            for (int il = 0; il < (int) n_layer; ++il) {
+                const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
+                const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+
+                const size_t k_size = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_head);
+
+                ggml_backend_tensor_set(kv_self.k_l[il], inp, 0, k_size);
+                inp += k_size;
+
+                if (kv_self.recurrent || !kv_self.v_trans) {
+                    // v is contiguous for recurrent models
+                    // TODO: use other tensors for state models than k and v
+                    const size_t v_size = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*kv_head);
+
+                    ggml_backend_tensor_set(kv_self.v_l[il], inp, 0, v_size);
+                    inp += v_size;
+                    continue;
+                }
+
+                // v is not contiguous, copy row by row
+                const size_t v_row_size   = ggml_row_size(kv_self.v_l[il]->type, kv_head);
+                const size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, kv_self.size);
+
+                for (int ir = 0; ir < (int) n_embd_v_gqa; ++ir) {
+                    ggml_backend_tensor_set(kv_self.v_l[il], inp, ir*v_row_stride, v_row_size);
+                    inp += v_row_size;
+                }
+            }
+            GGML_ASSERT(kv_buf_size == inp - src - pre_kv_buf_size);
+        }
+
+        ctx->kv_self.head = kv_head;
+        ctx->kv_self.used = kv_used;
+
+        for (uint32_t i = 0; i < kv_head; ++i) {
+            llama_pos pos;
+            size_t    seq_id_size;
+
+            memcpy(&pos,         inp, sizeof(pos));         inp += sizeof(pos);
+            memcpy(&seq_id_size, inp, sizeof(seq_id_size)); inp += sizeof(seq_id_size);
+
+            ctx->kv_self.cells[i].pos = pos;
+
+            llama_seq_id seq_id;
+
+            for (size_t j = 0; j < seq_id_size; ++j) {
+                memcpy(&seq_id, inp, sizeof(seq_id)); inp += sizeof(seq_id);
+                ctx->kv_self.cells[i].seq_id.insert(seq_id);
+            }
+        }
+    }
+
+    const size_t nread    = inp - src;
+    const size_t max_size = llama_state_get_size(ctx);
+
+    GGML_ASSERT(nread <= max_size);
+
+    return nread;
+}
+
+static bool llama_state_load_file_internal(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
+    llama_file file(path_session, "rb");
+
+    // sanity checks
+    {
+        const uint32_t magic   = file.read_u32();
+        const uint32_t version = file.read_u32();
+
+        if (magic != LLAMA_SESSION_MAGIC || version != LLAMA_SESSION_VERSION) {
+            LLAMA_LOG_ERROR("%s : unknown (magic, version) for session file: %08x, %08x\n", __func__, magic, version);
+            return false;
+        }
+
+        llama_hparams session_hparams;
+        file.read_raw(&session_hparams, sizeof(llama_hparams));
+
+        if (session_hparams != ctx->model.hparams) {
+            LLAMA_LOG_INFO("%s : model hparams didn't match from session file!\n", __func__);
+            return false;
+        }
+    }
+
+    // load the prompt
+    {
+        const uint32_t n_token_count = file.read_u32();
+
+        if (n_token_count > n_token_capacity) {
+            LLAMA_LOG_ERROR("%s : token count in session file exceeded capacity! %u > %zu\n", __func__, n_token_count, n_token_capacity);
+            return false;
+        }
+
+        file.read_raw(tokens_out, sizeof(llama_token) * n_token_count);
+        *n_token_count_out = n_token_count;
+    }
+
+    // restore the context state
+    {
+        const size_t n_state_size_cur = file.size - file.tell();
+        const size_t n_state_size_max = llama_state_get_size(ctx);
+
+        if (n_state_size_cur > n_state_size_max) {
+            LLAMA_LOG_ERROR("%s : the state size in session file is too big! max %zu, got %zu\n", __func__, n_state_size_max, n_state_size_cur);
+            return false;
+        }
+
+        std::vector state_data(n_state_size_max);
+        file.read_raw(state_data.data(), n_state_size_cur);
+
+        llama_state_set_data(ctx, state_data.data());
+    }
+
+    return true;
+}
+
+bool llama_state_load_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
+    try {
+        return llama_state_load_file_internal(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out);
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("error loading session file: %s\n", err.what());
+        return false;
+    }
+}
+
+static bool llama_state_save_file_internal(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) {
+    llama_file file(path_session, "wb");
+
+    file.write_u32(LLAMA_SESSION_MAGIC);
+    file.write_u32(LLAMA_SESSION_VERSION);
+
+    file.write_raw(&ctx->model.hparams, sizeof(llama_hparams));
+
+    // save the prompt
+    file.write_u32((uint32_t) n_token_count);
+    file.write_raw(tokens, sizeof(llama_token) * n_token_count);
+
+    // save the context state using stream saving
+    llama_data_file_context data_ctx(&file);
+    llama_state_get_data_internal(ctx, &data_ctx);
+
+    return true;
+}
+
+bool llama_state_save_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) {
+    try {
+        return llama_state_save_file_internal(ctx, path_session, tokens, n_token_count);
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("error saving session file: %s\n", err.what());
+        return false;
+    }
+}
+
+size_t llama_state_seq_get_size(struct llama_context* ctx, llama_seq_id seq_id) {
+    // save the size of size_t as a uint32_t for safety check
+    const size_t size_t_size_size = sizeof(uint32_t);
+
+    // other values
+    const size_t s_cell_count_size = sizeof(uint32_t);
+    const size_t s_layer_count_size = sizeof(uint32_t);
+    const size_t n_embd_v_gqa_size = sizeof(uint32_t);
+
+    size_t s_cell_count = 0;
+    size_t s_cell_data_size = 0;
+    const auto & kv_self = ctx->kv_self;
+    const auto & hparams = ctx->model.hparams;
+
+    const uint32_t n_layer = hparams.n_layer;
+
+    for (uint32_t i = 0; i < kv_self.size; ++i) {
+        const auto & cell = kv_self.cells[i];
+        if (cell.seq_id.count(seq_id) > 0) {
+            ++s_cell_count;
+            s_cell_data_size += sizeof(llama_pos);
+        }
+    }
+
+    for (int il = 0; il < (int)n_layer; ++il) {
+        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
+        const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+
+        // types of keys and values
+        s_cell_data_size += sizeof(int32_t) * 2;
+        // k_size_row and v_size_el values of layer
+        s_cell_data_size += sizeof(size_t) * 2;
+
+        // keys
+        const size_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa);
+        s_cell_data_size += k_size_row * s_cell_count;
+
+        // values (transposed)
+        const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type);
+        s_cell_data_size += v_size_el * s_cell_count * n_embd_v_gqa;
+    }
+
+    const size_t s_total = (
+        size_t_size_size +
+        s_cell_count_size +
+        s_layer_count_size +
+        n_embd_v_gqa_size +
+        s_cell_data_size
+        );
+
+    return s_total;
+}
+
+static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llama_data_context & data_ctx, llama_seq_id seq_id) {
+    llama_synchronize(ctx);
+
+    const auto & kv_self = ctx->kv_self;
+    GGML_ASSERT(!kv_self.recurrent); // not implemented
+
+    // Save the size of size_t as a uint32_t for safety check
+    const uint32_t size_t_size = sizeof(size_t);
+    data_ctx.write(&size_t_size, sizeof(size_t_size));
+
+    std::vector> cell_ranges; // ranges, from inclusive, to exclusive
+    uint32_t cell_count = 0;
+
+    // Count the number of cells with the specified seq_id
+    // Find all the ranges of cells with this seq id
+    {
+        uint32_t cell_range_begin = kv_self.size;
+        for (uint32_t i = 0; i < kv_self.size; ++i) {
+            const auto & cell = kv_self.cells[i];
+            if (cell.has_seq_id(seq_id)) {
+                ++cell_count;
+                if (cell_range_begin == kv_self.size) {
+                    cell_range_begin = i;
+                }
+            }
+            else {
+                if (cell_range_begin != kv_self.size) {
+                    cell_ranges.emplace_back(cell_range_begin, i);
+                    cell_range_begin = kv_self.size;
+                }
+            }
+        }
+        if (cell_range_begin != kv_self.size) {
+            cell_ranges.emplace_back(cell_range_begin, kv_self.size);
+        }
+
+        // DEBUG CHECK: Sum of cell counts in ranges should equal the total cell count
+        uint32_t cell_count_check = 0;
+        for (const auto & range : cell_ranges) {
+            cell_count_check += range.second - range.first;
+        }
+        GGML_ASSERT(cell_count == cell_count_check);
+    }
+
+    // Write the cell count
+    data_ctx.write(&cell_count, sizeof(cell_count));
+
+    const auto & hparams = ctx->model.hparams;
+    const uint32_t n_layer = hparams.n_layer;
+
+    // Write the layer count
+    data_ctx.write(&n_layer, sizeof(n_layer));
+
+    // Write n_embd_v_gqa (reference value)
+    {
+        const uint32_t n_embd_v_gqa_ref = hparams.n_embd_v_gqa() + hparams.n_embd_k_s();
+        data_ctx.write(&n_embd_v_gqa_ref, sizeof(n_embd_v_gqa_ref));
+    }
+
+    // Iterate the ranges and write all the pos (this is the token position in the prompt)
+    for (const auto & range : cell_ranges) {
+        for (uint32_t i = range.first; i < range.second; ++i) {
+            const auto & cell = kv_self.cells[i];
+            data_ctx.write(&cell.pos, sizeof(cell.pos));
+        }
+    }
+
+    // Iterate and write all the keys first, each row is a cell
+    // Get whole range at a time
+    std::vector tmp_buf;
+    for (int il = 0; il < (int)n_layer; ++il) {
+        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
+
+        // Write key type
+        const int32_t k_type_i = (int32_t)kv_self.k_l[il]->type;
+        data_ctx.write(&k_type_i, sizeof(k_type_i));
+
+        // Write row size of key
+        const size_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa);
+        data_ctx.write(&k_size_row, sizeof(k_size_row));
+
+        // Read each range of cells of k_size length each into tmp_buf and write out
+        for (const auto & range : cell_ranges) {
+            const size_t range_size = range.second - range.first;
+            tmp_buf.resize(range_size * k_size_row);
+            ggml_backend_tensor_get(kv_self.k_l[il], tmp_buf.data(), range.first * k_size_row, range_size * k_size_row);
+            data_ctx.write(tmp_buf.data(), tmp_buf.size());
+        }
+    }
+
+    // TODO: simplify, reduce copy-paste
+    if (!kv_self.v_trans) {
+        for (int il = 0; il < (int)n_layer; ++il) {
+            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+
+            // Write value type
+            const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type;
+            data_ctx.write(&v_type_i, sizeof(v_type_i));
+
+            // Write row size of value
+            const size_t v_size_row = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa);
+            data_ctx.write(&v_size_row, sizeof(v_size_row));
+
+            // Read each range of cells of v_size length each into tmp_buf and write out
+            for (const auto & range : cell_ranges) {
+                const size_t range_size = range.second - range.first;
+                tmp_buf.resize(range_size * v_size_row);
+                ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), range.first * v_size_row, range_size * v_size_row);
+                data_ctx.write(tmp_buf.data(), tmp_buf.size());
+            }
+        }
+    } else {
+        // For the values, they are transposed, so we also need the element size and get the element ranges from each row
+        const uint32_t kv_size = kv_self.size;
+        for (int il = 0; il < (int)n_layer; ++il) {
+            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+
+            // Write value type
+            const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type;
+            data_ctx.write(&v_type_i, sizeof(v_type_i));
+
+            // Write element size
+            const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type);
+            data_ctx.write(&v_size_el, sizeof(v_size_el));
+
+            // For each row, we get the element values of each cell
+            for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
+                // Read each range of cells of v_size_el length each into tmp_buf and write out
+                for (const auto & range : cell_ranges) {
+                    const size_t range_size = range.second - range.first;
+                    const size_t src_offset = (range.first + j * kv_size) * v_size_el;
+                    tmp_buf.resize(range_size * v_size_el);
+                    ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), src_offset, tmp_buf.size());
+                    data_ctx.write(tmp_buf.data(), tmp_buf.size());
+                }
+            }
+        }
+    }
+
+    return data_ctx.get_size_written();
+}
+
+size_t llama_state_seq_get_data(struct llama_context* ctx, uint8_t* dst, llama_seq_id seq_id) {
+    llama_data_buffer_context data_ctx(dst);
+    return llama_state_seq_get_data_internal(ctx, data_ctx, seq_id);
+}
+
+size_t llama_state_seq_set_data(struct llama_context * ctx, const uint8_t * src, llama_seq_id dest_seq_id) {
+    llama_synchronize(ctx);
+
+    auto & kv_self = ctx->kv_self;
+    GGML_ASSERT(!kv_self.recurrent); // not implemented
+
+    // Wipe the slot
+    llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1);
+
+    const uint8_t * inp = src;
+
+    // Read size of size_t
+    uint32_t size_t_size;
+    memcpy(&size_t_size, inp, sizeof(size_t_size));
+    inp += sizeof(size_t_size);
+    if (size_t_size != sizeof(size_t)) {
+        LLAMA_LOG_ERROR("%s: size_t size mismatch\n", __func__);
+        return 0;
+    }
+
+    // Read the cell count
+    uint32_t cell_count;
+    memcpy(&cell_count, inp, sizeof(cell_count));
+    inp += sizeof(cell_count);
+
+    // Read the layer count
+    uint32_t n_layer_ref;
+    memcpy(&n_layer_ref, inp, sizeof(n_layer_ref));
+    inp += sizeof(n_layer_ref);
+
+    // Read n_embd_v_gqa
+    uint32_t n_embd_v_gqa_ref;
+    memcpy(&n_embd_v_gqa_ref, inp, sizeof(n_embd_v_gqa_ref));
+    inp += sizeof(n_embd_v_gqa_ref);
+
+    // Sanity check model compatibility
+    const auto & hparams = ctx->model.hparams;
+    const uint32_t n_layer = hparams.n_layer;
+
+    if (n_layer != n_layer_ref) {
+        LLAMA_LOG_ERROR("%s: mismatched n_layer (%d != %d)\n", __func__, n_layer, n_layer_ref);
+        return 0;
+    }
+
+    if (hparams.n_embd_v_gqa() != n_embd_v_gqa_ref) {
+        LLAMA_LOG_ERROR("%s: mismatched n_embd_v_gqa (%d != %d)\n", __func__, hparams.n_embd_v_gqa(), n_embd_v_gqa_ref);
+        return 0;
+    }
+
+    // Allocate the new cells for the slot
+    if (cell_count) {
+        llama_batch batch = llama_batch_init(cell_count, 0, 1);
+        batch.n_tokens = cell_count;
+        for (uint32_t i = 0; i < cell_count; ++i) {
+            llama_pos pos;
+            memcpy(&pos, inp, sizeof(pos));
+            inp += sizeof(pos);
+
+            batch.pos[i] = pos;
+            batch.n_seq_id[i] = 1;
+            batch.seq_id[i][0] = dest_seq_id;
+        }
+        if (!llama_kv_cache_find_slot(kv_self, batch)) {
+            llama_batch_free(batch);
+            LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__);
+            return 0;
+        }
+
+        // DEBUG CHECK: kv_self.head should be our first cell, kv_self.head + cell_count - 1 should be our last cell (verify seq_id and pos values)
+        // Assume that this is one contiguous block of cells
+        GGML_ASSERT(kv_self.head + cell_count <= kv_self.size);
+        GGML_ASSERT(kv_self.cells[kv_self.head].pos == batch.pos[0]);
+        GGML_ASSERT(kv_self.cells[kv_self.head + cell_count - 1].pos == batch.pos[cell_count - 1]);
+        GGML_ASSERT(kv_self.cells[kv_self.head].has_seq_id(dest_seq_id));
+        GGML_ASSERT(kv_self.cells[kv_self.head + cell_count - 1].has_seq_id(dest_seq_id));
+
+        // Cleanup
+        llama_batch_free(batch);
+    }
+
+    const uint32_t kv_size = kv_self.size;
+    const uint32_t kv_head = kv_self.head;
+
+    // For each layer, read the keys for each cell, one row is one cell, read as one contiguous blo
+    for (int il = 0; il < (int)n_layer; ++il) {
+        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
+
+        // Read type of key
+        int32_t k_type_i_ref;
+        memcpy(&k_type_i_ref, inp, sizeof(k_type_i_ref));
+        inp += sizeof(k_type_i_ref);
+        const int32_t k_type_i = (int32_t)kv_self.k_l[il]->type;
+        if (k_type_i != k_type_i_ref) {
+            llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1);
+            LLAMA_LOG_ERROR("%s: mismatched key type (%d != %d, layer %d)\n", __func__, k_type_i, k_type_i_ref, il);
+            return 0;
+        }
+
+        // Read row size of key
+        size_t k_size_row_ref;
+        memcpy(&k_size_row_ref, inp, sizeof(k_size_row_ref));
+        inp += sizeof(k_size_row_ref);
+        const size_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa);
+        if (k_size_row != k_size_row_ref) {
+            llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1);
+            LLAMA_LOG_ERROR("%s: mismatched key row size (%zu != %zu, layer %d)\n", __func__, k_size_row, k_size_row_ref, il);
+            return 0;
+        }
+
+        if (cell_count) {
+            // Read and set the keys for the whole cell range
+            ggml_backend_tensor_set(kv_self.k_l[il], inp, kv_head * k_size_row, cell_count * k_size_row);
+            inp += cell_count * k_size_row;
+        }
+    }
+
+    // TODO: simplify, reduce copy-paste
+    if (!kv_self.v_trans) {
+        for (int il = 0; il < (int)n_layer; ++il) {
+            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+
+            // Read type of value
+            int32_t v_type_i_ref;
+            memcpy(&v_type_i_ref, inp, sizeof(v_type_i_ref));
+            inp += sizeof(v_type_i_ref);
+            const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type;
+            if (v_type_i != v_type_i_ref) {
+                llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1);
+                LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il);
+                return 0;
+            }
+
+            // Read row size of value
+            size_t v_size_row_ref;
+            memcpy(&v_size_row_ref, inp, sizeof(v_size_row_ref));
+            inp += sizeof(v_size_row_ref);
+            const size_t v_size_row = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa);
+            if (v_size_row != v_size_row_ref) {
+                llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1);
+                LLAMA_LOG_ERROR("%s: mismatched value row size (%zu != %zu, layer %d)\n", __func__, v_size_row, v_size_row_ref, il);
+                return 0;
+            }
+
+            if (cell_count) {
+                // Read and set the values for the whole cell range
+                ggml_backend_tensor_set(kv_self.v_l[il], inp, kv_head * v_size_row, cell_count * v_size_row);
+                inp += cell_count * v_size_row;
+            }
+        }
+    } else {
+        // For each layer, read the values for each cell (transposed)
+        for (int il = 0; il < (int)n_layer; ++il) {
+            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+
+            // Read type of value
+            int32_t v_type_i_ref;
+            memcpy(&v_type_i_ref, inp, sizeof(v_type_i_ref));
+            inp += sizeof(v_type_i_ref);
+            const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type;
+            if (v_type_i != v_type_i_ref) {
+                llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1);
+                LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il);
+                return 0;
+            }
+
+            // Read element size of value
+            size_t v_size_el_ref;
+            memcpy(&v_size_el_ref, inp, sizeof(v_size_el_ref));
+            inp += sizeof(v_size_el_ref);
+            const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type);
+            if (v_size_el != v_size_el_ref) {
+                llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1);
+                LLAMA_LOG_ERROR("%s: mismatched value element size (%zu != %zu, layer %d)\n", __func__, v_size_el, v_size_el_ref, il);
+                return 0;
+            }
+
+            if (cell_count) {
+                // For each row in the transposed matrix, read the values for the whole cell range
+                for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
+                    const size_t dst_offset = (kv_head + j * kv_size) * v_size_el;
+                    ggml_backend_tensor_set(kv_self.v_l[il], inp, dst_offset, cell_count * v_size_el);
+                    inp += cell_count * v_size_el;
+                }
+            }
+        }
+    }
+
+    const size_t nread = inp - src;
+
+    return nread;
+}
+
+static size_t llama_state_seq_save_file_internal(struct llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) {
+    llama_file file(filepath, "wb");
+
+    file.write_u32(LLAMA_STATE_SEQ_MAGIC);
+    file.write_u32(LLAMA_STATE_SEQ_VERSION);
+
+    // save the prompt
+    file.write_u32((uint32_t)n_token_count);
+    file.write_raw(tokens, sizeof(llama_token) * n_token_count);
+
+    // save the context state using stream saving
+    llama_data_file_context data_ctx(&file);
+    llama_state_seq_get_data_internal(ctx, data_ctx, seq_id);
+
+    const size_t res = file.tell();
+    GGML_ASSERT(res == sizeof(uint32_t) * 3 + sizeof(llama_token) * n_token_count + data_ctx.get_size_written());
+    return res;
+}
+
+static size_t llama_state_seq_load_file_internal(struct llama_context * ctx, const char * filepath, llama_seq_id dest_seq_id, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
+    llama_file file(filepath, "rb");
+
+    // version checks
+    {
+        const uint32_t magic   = file.read_u32();
+        const uint32_t version = file.read_u32();
+
+        if (magic != LLAMA_STATE_SEQ_MAGIC || version != LLAMA_STATE_SEQ_VERSION) {
+            LLAMA_LOG_ERROR("%s: unknown (magic, version) for sequence state file: %08x, %08x\n", __func__, magic, version);
+            return 0;
+        }
+    }
+
+    // load the prompt
+    {
+        const uint32_t n_token_count = file.read_u32();
+
+        if (n_token_count > n_token_capacity) {
+            LLAMA_LOG_ERROR("%s: token count in sequence state file exceeded capacity! %u > %zu\n", __func__, n_token_count, n_token_capacity);
+            return 0;
+        }
+
+        file.read_raw(tokens_out, sizeof(llama_token) * n_token_count);
+        *n_token_count_out = n_token_count;
+    }
+
+    // restore the context state
+    {
+        const size_t state_size = file.size - file.tell();
+        std::vector state_data(state_size);
+        file.read_raw(state_data.data(), state_size);
+        const size_t nread = llama_state_seq_set_data(ctx, state_data.data(), dest_seq_id);
+        if (!nread) {
+            LLAMA_LOG_ERROR("%s: failed to restore sequence state\n", __func__);
+            return 0;
+        }
+        GGML_ASSERT(nread <= state_size);
+        GGML_ASSERT(nread + sizeof(uint32_t) * 3 + sizeof(llama_token) * *n_token_count_out == file.tell());
+    }
+
+    return file.tell();
+}
+
+size_t llama_state_seq_save_file(struct llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) {
+    try {
+        return llama_state_seq_save_file_internal(ctx, filepath, seq_id, tokens, n_token_count);
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("error saving sequence state file: %s\n", err.what());
+        return 0;
+    }
+}
+
+size_t llama_state_seq_load_file(struct llama_context * ctx, const char * filepath, llama_seq_id dest_seq_id, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
+    try {
+        return llama_state_seq_load_file_internal(ctx, filepath, dest_seq_id, tokens_out, n_token_capacity, n_token_count_out);
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("error loading sequence state file: %s\n", err.what());
+        return 0;
+    }
+}
+
+void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch) {
+    ctx->cparams.n_threads       = n_threads;
+    ctx->cparams.n_threads_batch = n_threads_batch;
+}
+
+uint32_t llama_n_threads(struct llama_context * ctx) {
+    return ctx->cparams.n_threads;
+}
+
+uint32_t llama_n_threads_batch(struct llama_context * ctx) {
+    return ctx->cparams.n_threads_batch;
+}
+
+void llama_set_abort_callback(struct llama_context * ctx, bool (*abort_callback)(void * data), void * abort_callback_data) {
+    ctx->abort_callback      = abort_callback;
+    ctx->abort_callback_data = abort_callback_data;
+}
+
+void llama_set_embeddings(struct llama_context * ctx, bool embeddings) {
+    ctx->cparams.embeddings = embeddings;
+}
+
+void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn) {
+    ctx->cparams.causal_attn = causal_attn;
+}
+
+struct llama_batch llama_batch_get_one(
+             llama_token * tokens,
+                 int32_t   n_tokens,
+               llama_pos   pos_0,
+            llama_seq_id   seq_id) {
+    return {
+        /*n_tokens       =*/ n_tokens,
+        /*tokens         =*/ tokens,
+        /*embd           =*/ nullptr,
+        /*pos            =*/ nullptr,
+        /*n_seq_id       =*/ nullptr,
+        /*seq_id         =*/ nullptr,
+        /*logits         =*/ nullptr,
+        /*all_pos_0      =*/ pos_0,
+        /*all_pos_1      =*/ 1,
+        /*all_seq_id     =*/ seq_id,
+    };
+}
+
+struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_t n_seq_max) {
+    llama_batch batch = { 0, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, 0, 0, 0, };
+
+    if (embd) {
+        batch.embd = (float *) malloc(sizeof(float) * n_tokens_alloc * embd);
+    } else {
+        batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens_alloc);
+    }
+
+    batch.pos      = (llama_pos *)     malloc(sizeof(llama_pos)      * n_tokens_alloc);
+    batch.n_seq_id = (int32_t *)       malloc(sizeof(int32_t)        * n_tokens_alloc);
+    batch.seq_id   = (llama_seq_id **) malloc(sizeof(llama_seq_id *) * (n_tokens_alloc + 1));
+    for (int i = 0; i < n_tokens_alloc; ++i) {
+        batch.seq_id[i] = (llama_seq_id *) malloc(sizeof(llama_seq_id) * n_seq_max);
+    }
+    batch.seq_id[n_tokens_alloc] = nullptr;
+
+    batch.logits   = (int8_t *)        malloc(sizeof(int8_t)         * n_tokens_alloc);
+
+    return batch;
+}
+
+void llama_batch_free(struct llama_batch batch) {
+    if (batch.token)    free(batch.token);
+    if (batch.embd)     free(batch.embd);
+    if (batch.pos)      free(batch.pos);
+    if (batch.n_seq_id) free(batch.n_seq_id);
+    if (batch.seq_id) {
+        for (int i = 0; batch.seq_id[i] != nullptr; ++i) {
+            free(batch.seq_id[i]);
+        }
+        free(batch.seq_id);
+    }
+    if (batch.logits)   free(batch.logits);
+}
+
+int32_t llama_encode(
+        struct llama_context * ctx,
+          struct llama_batch   batch) {
+    const int ret = llama_encode_internal(*ctx, batch);
+    if (ret < 0) {
+        LLAMA_LOG_ERROR("%s: failed to encode, ret = %d\n", __func__, ret);
+    }
+
+    return ret;
+}
+
+int32_t llama_decode(
+        struct llama_context * ctx,
+          struct llama_batch   batch) {
+    const int ret = llama_decode_internal(*ctx, batch);
+    if (ret < 0) {
+        LLAMA_LOG_ERROR("%s: failed to decode, ret = %d\n", __func__, ret);
+    }
+
+    return ret;
+}
+
+void llama_synchronize(struct llama_context * ctx) {
+    ggml_backend_sched_synchronize(ctx->sched);
+
+    // FIXME: if multiple single tokens are evaluated without a synchronization,
+    // the stats will be added to the prompt evaluation stats
+    // this should only happen when using batch size 1 to evaluate a batch
+
+    // add the evaluation to the stats
+    if (ctx->n_queued_tokens == 1) {
+        ctx->t_eval_us += ggml_time_us() - ctx->t_compute_start_us;
+        ctx->n_eval++;
+    } else if (ctx->n_queued_tokens > 1) {
+        ctx->t_p_eval_us += ggml_time_us() - ctx->t_compute_start_us;
+        ctx->n_p_eval += ctx->n_queued_tokens;
+    }
+
+    // get a more accurate load time, upon first eval
+    if (ctx->n_queued_tokens > 0 && !ctx->has_evaluated_once) {
+        ctx->t_load_us = ggml_time_us() - ctx->t_start_us;
+        ctx->has_evaluated_once = true;
+    }
+
+    ctx->n_queued_tokens = 0;
+    ctx->t_compute_start_us = 0;
+}
+
+float * llama_get_logits(struct llama_context * ctx) {
+    llama_synchronize(ctx);
+
+    return ctx->logits;
+}
+
+float * llama_get_logits_ith(struct llama_context * ctx, int32_t i) {
+    int32_t j = -1;
+    llama_synchronize(ctx);
+
+    try {
+        if (ctx->logits == nullptr) {
+            throw std::runtime_error("no logits");
+        }
+
+        if (i < 0) {
+            j = ctx->n_outputs + i;
+            if (j < 0) {
+                throw std::runtime_error(format("negative index out of range [0, %d)", ctx->n_outputs));
+            }
+        } else if ((size_t) i >= ctx->output_ids.size()) {
+            throw std::runtime_error(format("out of range [0, %lu)", ctx->output_ids.size()));
+        } else {
+            j = ctx->output_ids[i];
+        }
+
+        if (j < 0) {
+            throw std::runtime_error(format("batch.logits[%d] != true", i));
+        }
+        if (j >= ctx->n_outputs) {
+            // This should not happen
+            throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, ctx->n_outputs));
+        }
+
+        return ctx->logits + j*ctx->model.hparams.n_vocab;
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what());
+#ifndef NDEBUG
+        GGML_ASSERT(false);
+#endif
+        return nullptr;
+    }
+}
+
+float * llama_get_embeddings(struct llama_context * ctx) {
+    llama_synchronize(ctx);
+
+    return ctx->embd;
+}
+
+float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i) {
+    int32_t j = -1;
+
+    llama_synchronize(ctx);
+
+    try {
+        if (ctx->embd == nullptr) {
+            throw std::runtime_error("no embeddings");
+        }
+
+        if (i < 0) {
+            j = ctx->n_outputs + i;
+            if (j < 0) {
+                throw std::runtime_error(format("negative index out of range [0, %d)", ctx->n_outputs));
+            }
+        } else if ((size_t) i >= ctx->output_ids.size()) {
+            throw std::runtime_error(format("out of range [0, %lu)", ctx->output_ids.size()));
+        } else {
+            j = ctx->output_ids[i];
+        }
+
+        if (j < 0) {
+            throw std::runtime_error(format("batch.logits[%d] != true", i));
+        }
+        if (j >= ctx->n_outputs) {
+            // This should not happen
+            throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, ctx->n_outputs));
+        }
+
+        return ctx->embd + j*ctx->model.hparams.n_embd;
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what());
+#ifndef NDEBUG
+        GGML_ASSERT(false);
+#endif
+        return nullptr;
+    }
+}
+
+float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id) {
+    llama_synchronize(ctx);
+
+    auto it = ctx->embd_seq.find(seq_id);
+    if (it == ctx->embd_seq.end()) {
+        return nullptr;
+    }
+
+    return it->second.data();
+}
+
+//
+// vocab
+//
+
+const char * llama_token_get_text(const struct llama_model * model, llama_token token) {
+    return llama_token_get_text_impl(model->vocab, token);
+}
+
+float llama_token_get_score(const struct llama_model * model, llama_token token) {
+    return llama_token_get_score_impl(model->vocab, token);
+}
+
+enum llama_token_attr llama_token_get_attr(const struct llama_model * model, llama_token token) {
+    return llama_token_get_attr_impl(model->vocab, token);
+}
+
+bool llama_token_is_eog(const struct llama_model * model, llama_token token) {
+    return llama_token_is_eog_impl(model->vocab, token);
+}
+
+bool llama_token_is_control(const struct llama_model * model, llama_token token) {
+    return llama_token_is_control_impl(model->vocab, token);
+}
+
+llama_token llama_token_bos(const struct llama_model * model) {
+    return llama_token_bos_impl(model->vocab);
+}
+
+llama_token llama_token_eos(const struct llama_model * model) {
+    return llama_token_eos_impl(model->vocab);
+}
+
+llama_token llama_token_cls(const struct llama_model * model) {
+    return llama_token_cls_impl(model->vocab);
+}
+
+llama_token llama_token_sep(const struct llama_model * model) {
+    return llama_token_sep_impl(model->vocab);
+}
+
+llama_token llama_token_nl (const struct llama_model * model) {
+    return llama_token_nl_impl(model->vocab);
+}
+
+llama_token llama_token_pad(const struct llama_model * model) {
+    return llama_token_pad_impl(model->vocab);
+}
+
+int32_t llama_add_bos_token(const struct llama_model * model) {
+    return llama_add_bos_token_impl(model->vocab);
+}
+
+int32_t llama_add_eos_token(const struct llama_model * model) {
+    return llama_add_eos_token_impl(model->vocab);
+}
+
+llama_token llama_token_prefix(const struct llama_model * model) {
+    return llama_token_prefix_impl(model->vocab);
+}
+
+llama_token llama_token_middle(const struct llama_model * model) {
+    return llama_token_middle_impl(model->vocab);
+}
+
+llama_token llama_token_suffix(const struct llama_model * model) {
+    return llama_token_suffix_impl(model->vocab);
+}
+
+llama_token llama_token_eot(const struct llama_model * model) {
+    return llama_token_eot_impl(model->vocab);
+}
+
+//
+// tokenization
+//
+
+int32_t llama_tokenize(
+    const struct llama_model * model,
+                  const char * text,
+                     int32_t   text_len,
+                 llama_token * tokens,
+                     int32_t   n_tokens_max,
+                        bool   add_special,
+                        bool   parse_special) {
+    return llama_tokenize_impl(model->vocab, text, text_len, tokens, n_tokens_max, add_special, parse_special);
+}
+
+int32_t llama_token_to_piece(
+    const struct llama_model * model,
+                 llama_token   token,
+                        char * buf,
+                     int32_t   length,
+                     int32_t   lstrip,
+                        bool   special) {
+    return llama_token_to_piece_impl(model->vocab, token, buf, length, lstrip, special);
+}
+
+int32_t llama_detokenize(
+    const struct llama_model * model,
+           const llama_token * tokens,
+                     int32_t   n_tokens,
+                        char * text,
+                     int32_t   text_len_max,
+                        bool   remove_special,
+                        bool   unparse_special) {
+    return llama_detokenize_impl(model->vocab, tokens, n_tokens, text, text_len_max, remove_special, unparse_special);
+}
+
+//
+// chat templates
+//
+
+// Simple version of "llama_apply_chat_template" that only works with strings
+// This function uses heuristic checks to determine commonly used template. It is not a jinja parser.
+static int32_t llama_chat_apply_template_internal(
+    const std::string & tmpl,
+    const std::vector & chat,
+    std::string & dest, bool add_ass) {
+    // Taken from the research: https://github.com/ggerganov/llama.cpp/issues/5527
+    std::stringstream ss;
+    auto tmpl_contains = [&tmpl](std::string haystack) -> bool {
+        return tmpl.find(haystack) != std::string::npos;
+    };
+    if (tmpl == "chatml" || tmpl_contains("<|im_start|>")) {
+        // chatml template
+        for (auto message : chat) {
+            ss << "<|im_start|>" << message->role << "\n" << message->content << "<|im_end|>\n";
+        }
+        if (add_ass) {
+            ss << "<|im_start|>assistant\n";
+        }
+    } else if (tmpl == "llama2" || tmpl == "mistral" || tmpl_contains("[INST]")) {
+        // llama2 template and its variants
+        // [variant] support system message
+        bool support_system_message = tmpl_contains("<>") || tmpl == "mistral";
+        // [variant] space before + after response
+        bool space_around_response = tmpl_contains("' ' + eos_token");
+        // [variant] add BOS inside history
+        bool add_bos_inside_history = tmpl_contains("bos_token + '[INST]");
+        // [variant] trim spaces from the input message
+        bool strip_message = tmpl_contains("content.strip()");
+        // construct the prompt
+        bool is_inside_turn = true; // skip BOS at the beginning
+        ss << "[INST] ";
+        for (auto message : chat) {
+            std::string content = strip_message ? trim(message->content) : message->content;
+            std::string role(message->role);
+            if (!is_inside_turn) {
+                is_inside_turn = true;
+                ss << (add_bos_inside_history ? "[INST] " : "[INST] ");
+            }
+            if (role == "system") {
+                if (support_system_message) {
+                    ss << "<>\n" << content << "\n<>\n\n";
+                } else {
+                    // if the model does not support system message, we still include it in the first message, but without <>
+                    ss << content << "\n";
+                }
+            } else if (role == "user") {
+                ss << content << " [/INST]";
+            } else {
+                ss << (space_around_response ? " " : "") << content << (space_around_response ? " " : "") << "";
+                is_inside_turn = false;
+            }
+        }
+        // llama2 templates seem to not care about "add_generation_prompt"
+    } else if (tmpl == "phi3" || (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>"))) {
+        // Phi 3
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>\n" << message->content << "<|end|>\n";
+        }
+        if (add_ass) {
+            ss << "<|assistant|>\n";
+        }
+    } else if (tmpl == "zephyr" || tmpl_contains("<|user|>")) {
+        // zephyr template
+        for (auto message : chat) {
+            ss << "<|" << message->role << "|>" << "\n" << message->content << "<|endoftext|>\n";
+        }
+        if (add_ass) {
+            ss << "<|assistant|>\n";
+        }
+    } else if (tmpl == "monarch" || tmpl_contains("bos_token + message['role']")) {
+        // mlabonne/AlphaMonarch-7B template (the  is included inside history)
+        for (auto message : chat) {
+            std::string bos = (message == chat.front()) ? "" : ""; // skip BOS for first message
+            ss << bos << message->role << "\n" << message->content << "\n";
+        }
+        if (add_ass) {
+            ss << "assistant\n";
+        }
+    } else if (tmpl == "gemma" || tmpl == "gemma2" || tmpl_contains("")) {
+        // google/gemma-7b-it
+        std::string system_prompt = "";
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                // there is no system message for gemma, but we will merge it with user prompt, so nothing is broken
+                system_prompt = trim(message->content);
+                continue;
+            }
+            // in gemma, "assistant" is "model"
+            role = role == "assistant" ? "model" : message->role;
+            ss << "" << role << "\n";
+            if (!system_prompt.empty() && role != "model") {
+                ss << system_prompt << "\n\n";
+                system_prompt = "";
+            }
+            ss << trim(message->content) << "\n";
+        }
+        if (add_ass) {
+            ss << "model\n";
+        }
+    } else if (tmpl == "orion" || tmpl_contains("'\\n\\nAssistant: ' + eos_token")) {
+        // OrionStarAI/Orion-14B-Chat
+        std::string system_prompt = "";
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                // there is no system message support, we will merge it with user prompt
+                system_prompt = message->content;
+                continue;
+            } else if (role == "user") {
+                ss << "Human: ";
+                if (!system_prompt.empty()) {
+                    ss << system_prompt << "\n\n";
+                    system_prompt = "";
+                }
+                ss << message->content << "\n\nAssistant: ";
+            } else {
+                ss << message->content << "";
+            }
+        }
+    } else if (tmpl == "openchat" || tmpl_contains("GPT4 Correct ")) {
+        // openchat/openchat-3.5-0106,
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << message->content << "<|end_of_turn|>";
+            } else {
+                role[0] = toupper(role[0]);
+                ss << "GPT4 Correct " << role << ": " << message->content << "<|end_of_turn|>";
+            }
+        }
+        if (add_ass) {
+            ss << "GPT4 Correct Assistant:";
+        }
+    } else if (tmpl == "vicuna" || tmpl == "vicuna-orca" || (tmpl_contains("USER: ") && tmpl_contains("ASSISTANT: "))) {
+        // eachadea/vicuna-13b-1.1 (and Orca variant)
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                // Orca-Vicuna variant uses a system prefix
+                if (tmpl == "vicuna-orca" || tmpl_contains("SYSTEM: ")) {
+                    ss << "SYSTEM: " << message->content << "\n";
+                } else {
+                    ss << message->content << "\n\n";
+                }
+            } else if (role == "user") {
+                ss << "USER: " << message->content << "\n";
+            } else if (role == "assistant") {
+                ss << "ASSISTANT: " << message->content << "\n";
+            }
+        }
+        if (add_ass) {
+            ss << "ASSISTANT:";
+        }
+    } else if (tmpl == "deepseek" || (tmpl_contains("### Instruction:") && tmpl_contains("<|EOT|>"))) {
+        // deepseek-ai/deepseek-coder-33b-instruct
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << message->content;
+            } else if (role == "user") {
+                ss << "### Instruction:\n" << message->content << "\n";
+            } else if (role == "assistant") {
+                ss << "### Response:\n" << message->content << "\n<|EOT|>\n";
+            }
+        }
+        if (add_ass) {
+            ss << "### Response:\n";
+        }
+    } else if (tmpl == "command-r" || (tmpl_contains("<|START_OF_TURN_TOKEN|>") && tmpl_contains("<|USER_TOKEN|>"))) {
+        // CohereForAI/c4ai-command-r-plus
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>";
+            } else if (role == "user") {
+                ss << "<|START_OF_TURN_TOKEN|><|USER_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>";
+            } else if (role == "assistant") {
+                ss << "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>" << trim(message->content) << "<|END_OF_TURN_TOKEN|>";
+            }
+        }
+        if (add_ass) {
+            ss << "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>";
+        }
+    } else if (tmpl == "llama3" || (tmpl_contains("<|start_header_id|>") && tmpl_contains("<|end_header_id|>"))) {
+        // Llama 3
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|start_header_id|>" << role << "<|end_header_id|>\n\n" << trim(message->content) << "<|eot_id|>";
+        }
+        if (add_ass) {
+            ss << "<|start_header_id|>assistant<|end_header_id|>\n\n";
+        }
+    } else if (tmpl == "chatglm3" || tmpl_contains("[gMASK]sop")) {
+        // chatglm3-6b
+        ss << "[gMASK]" << "sop";
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>" << "\n " << message->content;
+        }
+        if (add_ass) {
+            ss << "<|assistant|>";
+        }
+    } else if (tmpl == "chatglm4" || tmpl_contains("[gMASK]")) {
+        ss << "[gMASK]" << "";
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>" << "\n" << message->content;
+        }
+        if (add_ass) {
+            ss << "<|assistant|>";
+        }
+    } else if (tmpl == "minicpm" || tmpl_contains(LU8("<用户>"))) {
+        // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "user") {
+                ss << LU8("<用户>");
+                ss << trim(message->content);
+                ss << "";
+            } else {
+                ss << trim(message->content);
+            }
+        }
+    } else if (tmpl == "deepseek2" || tmpl_contains("'Assistant: ' + message['content'] + eos_token")) {
+        // DeepSeek-V2
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << message->content << "\n\n";
+            } else if (role == "user") {
+                ss << "User: " << message->content << "\n\n";
+            } else if (role == "assistant") {
+                ss << "Assistant: " << message->content << LU8("<｜end▁of▁sentence｜>");
+            }
+        }
+        if (add_ass) {
+            ss << "Assistant:";
+        }
+    } else {
+        // template not supported
+        return -1;
+    }
+    dest = ss.str();
+    return dest.size();
+}
+
+int32_t llama_chat_apply_template(
+                const struct llama_model * model,
+                              const char * tmpl,
+         const struct llama_chat_message * chat,
+                                  size_t   n_msg,
+                                    bool   add_ass,
+                                    char * buf,
+                                 int32_t   length) {
+    std::string curr_tmpl(tmpl == nullptr ? "" : tmpl);
+    if (tmpl == nullptr) {
+        GGML_ASSERT(model != nullptr);
+        // load template from model
+        std::vector model_template(2048, 0); // longest known template is about 1200 bytes
+        std::string template_key = "tokenizer.chat_template";
+        int32_t res = llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size());
+        if (res < 0) {
+            // worst case: there is no information about template, we will use chatml by default
+            curr_tmpl = "chatml"; // see llama_chat_apply_template_internal
+        } else {
+            curr_tmpl = std::string(model_template.data(), model_template.size());
+        }
+    }
+
+    // format the chat to string
+    std::vector chat_vec;
+    chat_vec.resize(n_msg);
+    for (size_t i = 0; i < n_msg; i++) {
+        chat_vec[i] = &chat[i];
+    }
+
+    std::string formatted_chat;
+    int32_t res = llama_chat_apply_template_internal(curr_tmpl, chat_vec, formatted_chat, add_ass);
+    if (res < 0) {
+        return res;
+    }
+    if (buf && length > 0) {
+        strncpy(buf, formatted_chat.c_str(), length);
+    }
+    return res;
+}
+
+//
+// grammar
+//
+
+struct llama_grammar * llama_grammar_init(
+        const llama_grammar_element ** rules,
+        size_t    n_rules,
+        size_t    start_rule_index) {
+    return llama_grammar_init_impl(rules, n_rules, start_rule_index);
+}
+
+void llama_grammar_free(struct llama_grammar * grammar) {
+    llama_grammar_free_impl(grammar);
+}
+
+struct llama_grammar * llama_grammar_copy(const struct llama_grammar * grammar) {
+    return llama_grammar_copy_impl(grammar);
+}
+
+void llama_grammar_sample(
+      const struct llama_grammar * grammar,
+      const struct llama_context * ctx,
+          llama_token_data_array * candidates) {
+    llama_grammar_sample_impl(grammar, &ctx->model.vocab, &ctx->sampling, candidates);
+}
+
+void llama_sample_grammar(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+      const struct llama_grammar * grammar) {
+    llama_grammar_sample(grammar, ctx, candidates);
+}
+
+void llama_grammar_accept_token(
+            struct llama_grammar * grammar,
+            struct llama_context * ctx,
+                     llama_token   token) {
+    llama_grammar_accept_token_impl(grammar, &ctx->model.vocab, &ctx->sampling, token);
+}
+
+//
+// sampling
+//
+
+void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed) {
+    llama_set_rng_seed_impl(&ctx->sampling, seed);
+}
+
+void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates) {
+    llama_sample_softmax_impl(ctx ? &ctx->sampling : nullptr, candidates);
+}
+
+void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int32_t k, size_t min_keep) {
+    llama_sample_top_k_impl(ctx ? &ctx->sampling : nullptr, candidates, k, min_keep);
+}
+
+void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep) {
+    llama_sample_top_p_impl(ctx ? &ctx->sampling : nullptr, candidates, p, min_keep);
+}
+
+void llama_sample_min_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep) {
+    llama_sample_min_p_impl(ctx ? &ctx->sampling : nullptr, candidates, p, min_keep);
+}
+
+void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep) {
+    llama_sample_tail_free_impl(ctx ? &ctx->sampling : nullptr, candidates, z, min_keep);
+}
+
+void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep) {
+    llama_sample_typical_impl(ctx ? &ctx->sampling : nullptr, candidates, p, min_keep);
+}
+
+void llama_sample_entropy(struct llama_context * ctx, llama_token_data_array * candidates_p, float min_temp, float max_temp, float exponent_val) {
+    llama_sample_entropy_impl(ctx ? &ctx->sampling : nullptr, candidates_p, min_temp, max_temp, exponent_val);
+}
+
+void llama_sample_temp(struct llama_context * ctx, llama_token_data_array * candidates_p, float temp) {
+    llama_sample_temp_impl(ctx ? &ctx->sampling : nullptr, candidates_p, temp);
+}
+
+void llama_sample_repetition_penalties(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates,
+               const llama_token * last_tokens,
+                          size_t   penalty_last_n,
+                           float   penalty_repeat,
+                           float   penalty_freq,
+                           float   penalty_present) {
+    llama_sample_repetition_penalties_impl(ctx ? &ctx->sampling : nullptr, candidates, last_tokens, penalty_last_n, penalty_repeat, penalty_freq, penalty_present);
+}
+
+void llama_sample_apply_guidance(
+          struct llama_context * ctx,
+                         float * logits,
+                         float * logits_guidance,
+                         float   scale) {
+    llama_sample_apply_guidance_impl(&ctx->sampling, logits, logits_guidance, scale);
+}
+
+llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int32_t m, float * mu) {
+    return llama_sample_token_mirostat_impl(&ctx->sampling, candidates, tau, eta, m, mu);
+}
+
+llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu) {
+    return llama_sample_token_mirostat_v2_impl(ctx ? &ctx->sampling : nullptr, candidates, tau, eta, mu);
+}
+
+llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates) {
+    return llama_sample_token_greedy_impl(ctx ? &ctx->sampling : nullptr, candidates);
+}
+
+llama_token llama_sample_token_with_rng(struct llama_context * ctx, llama_token_data_array * candidates, std::mt19937 & rng) {
+    return llama_sample_token_with_rng_impl(&ctx->sampling, candidates, rng);
+}
+
+llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates) {
+    return llama_sample_token_with_rng_impl(&ctx->sampling, candidates, ctx->sampling.rng);
+}
+
+int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count) {
+    static const char * const SPLIT_PATH_FORMAT = "%s-%05d-of-%05d.gguf";
+    if (snprintf(split_path, maxlen, SPLIT_PATH_FORMAT, path_prefix, split_no + 1, split_count)) {
+        return strlen(split_path);
+    }
+    return 0;
+}
+
+int llama_split_prefix(char * dest, size_t maxlen, const char * split_path, int split_no, int split_count) {
+    std::string str_split_path(split_path);
+    char postfix[32];
+    snprintf(postfix, 32, "-%05d-of-%05d.gguf", split_no + 1, split_count);
+    std::string str_postfix(postfix);
+
+    // check if dest ends with postfix
+    int size_prefix = str_split_path.size() - str_postfix.size();
+    if (size_prefix > 0 && str_split_path.find(str_postfix, size_prefix) != std::string::npos) {
+        snprintf(dest, std::min((size_t) size_prefix + 1, maxlen), "%s", split_path);
+        return size_prefix;
+    }
+
+    return 0;
+}
+
+struct llama_timings llama_get_timings(struct llama_context * ctx) {
+    struct llama_timings result = {
+        /*.t_start_ms  =*/ 1e-3 * ctx->t_start_us,
+        /*.t_end_ms    =*/ 1.00 * ggml_time_ms(),
+        /*.t_load_ms   =*/ 1e-3 * ctx->t_load_us,
+        /*.t_sample_ms =*/ 1e-3 * ctx->sampling.t_sample_us,
+        /*.t_p_eval_ms =*/ 1e-3 * ctx->t_p_eval_us,
+        /*.t_eval_ms   =*/ 1e-3 * ctx->t_eval_us,
+
+        /*.n_sample =*/ std::max(1, ctx->sampling.n_sample),
+        /*.n_p_eval =*/ std::max(0, ctx->n_p_eval),
+        /*.n_eval   =*/ std::max(1, ctx->n_eval),
+    };
+
+    return result;
+}
+
+void llama_print_timings(struct llama_context * ctx) {
+    const llama_timings timings = llama_get_timings(ctx);
+
+    LLAMA_LOG_INFO("\n");
+    LLAMA_LOG_INFO("%s:        load time = %10.2f ms\n", __func__, timings.t_load_ms);
+    LLAMA_LOG_INFO("%s:      sample time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+            __func__, timings.t_sample_ms, timings.n_sample, timings.t_sample_ms / timings.n_sample, 1e3 / timings.t_sample_ms * timings.n_sample);
+    LLAMA_LOG_INFO("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
+            __func__, timings.t_p_eval_ms, timings.n_p_eval, timings.t_p_eval_ms / timings.n_p_eval, 1e3 / timings.t_p_eval_ms * timings.n_p_eval);
+    LLAMA_LOG_INFO("%s:        eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+            __func__, timings.t_eval_ms, timings.n_eval, timings.t_eval_ms / timings.n_eval, 1e3 / timings.t_eval_ms * timings.n_eval);
+    LLAMA_LOG_INFO("%s:       total time = %10.2f ms / %5d tokens\n", __func__, (timings.t_end_ms - timings.t_start_ms), (timings.n_p_eval + timings.n_eval));
+}
+
+void llama_reset_timings(struct llama_context * ctx) {
+    ctx->t_start_us  = ggml_time_us();
+    ctx->t_eval_us   = ctx->n_eval   = 0;
+    ctx->t_p_eval_us = ctx->n_p_eval = 0;
+
+    ctx->sampling.reset_timings();
+}
+
+const char * llama_print_system_info(void) {
+    static std::string s;
+
+    s  = "";
+    s += "AVX = "         + std::to_string(ggml_cpu_has_avx())         + " | ";
+    s += "AVX_VNNI = "    + std::to_string(ggml_cpu_has_avx_vnni())    + " | ";
+    s += "AVX2 = "        + std::to_string(ggml_cpu_has_avx2())        + " | ";
+    s += "AVX512 = "      + std::to_string(ggml_cpu_has_avx512())      + " | ";
+    s += "AVX512_VBMI = " + std::to_string(ggml_cpu_has_avx512_vbmi()) + " | ";
+    s += "AVX512_VNNI = " + std::to_string(ggml_cpu_has_avx512_vnni()) + " | ";
+    s += "AVX512_BF16 = " + std::to_string(ggml_cpu_has_avx512_bf16()) + " | ";
+    s += "FMA = "         + std::to_string(ggml_cpu_has_fma())         + " | ";
+    s += "NEON = "        + std::to_string(ggml_cpu_has_neon())        + " | ";
+    s += "SVE = "         + std::to_string(ggml_cpu_has_sve())         + " | ";
+    s += "ARM_FMA = "     + std::to_string(ggml_cpu_has_arm_fma())     + " | ";
+    s += "F16C = "        + std::to_string(ggml_cpu_has_f16c())        + " | ";
+    s += "FP16_VA = "     + std::to_string(ggml_cpu_has_fp16_va())     + " | ";
+    s += "WASM_SIMD = "   + std::to_string(ggml_cpu_has_wasm_simd())   + " | ";
+    s += "BLAS = "        + std::to_string(ggml_cpu_has_blas())        + " | ";
+    s += "SSE3 = "        + std::to_string(ggml_cpu_has_sse3())        + " | ";
+    s += "SSSE3 = "       + std::to_string(ggml_cpu_has_ssse3())       + " | ";
+    s += "VSX = "         + std::to_string(ggml_cpu_has_vsx())         + " | ";
+    s += "MATMUL_INT8 = " + std::to_string(ggml_cpu_has_matmul_int8()) + " | ";
+    s += "LLAMAFILE = "   + std::to_string(ggml_cpu_has_llamafile())   + " | ";
+
+    return s.c_str();
+}
+
+void llama_dump_timing_info_yaml(FILE * stream, const llama_context * ctx) {
+    fprintf(stream, "\n");
+    fprintf(stream, "###########\n");
+    fprintf(stream, "# Timings #\n");
+    fprintf(stream, "###########\n");
+    fprintf(stream, "\n");
+
+    fprintf(stream, "mst_eval: %.2f  # ms / token during generation\n",
+            1.0e-3 * ctx->t_eval_us / ctx->n_eval);
+    fprintf(stream, "mst_p_eval: %.2f  # ms / token during prompt processing\n",
+            1.0e-3 * ctx->t_p_eval_us / ctx->n_p_eval);
+    fprintf(stream, "mst_sample: %.2f  # ms / token during sampling\n",
+            1.0e-3 * ctx->sampling.t_sample_us / ctx->sampling.n_sample);
+    fprintf(stream, "n_eval: %d  # number of tokens generated (excluding the first one)\n", ctx->n_eval);
+    fprintf(stream, "n_p_eval: %d  # number of tokens processed in batches at the beginning\n", ctx->n_p_eval);
+    fprintf(stream, "n_sample: %d  # number of sampled tokens\n", ctx->sampling.n_sample);
+    fprintf(stream, "t_eval_us: %" PRId64 "  # total microseconds spent generating tokens\n", ctx->t_eval_us);
+    fprintf(stream, "t_load_us: %" PRId64 "  # total microseconds spent loading the model\n", ctx->t_load_us);
+    fprintf(stream, "t_p_eval_us: %" PRId64 "  # total microseconds spent prompt processing\n", ctx->t_p_eval_us);
+    fprintf(stream, "t_sample_us: %" PRId64 "  # total microseconds spent sampling\n", ctx->sampling.t_sample_us);
+    fprintf(stream, "ts_eval: %.2f  # tokens / second during generation\n",
+            1.0e6 * ctx->n_eval / ctx->t_eval_us);
+    fprintf(stream, "ts_p_eval: %.2f  # tokens / second during prompt processing\n",
+            1.0e6 * ctx->n_p_eval / ctx->t_p_eval_us);
+    fprintf(stream, "ts_sample: %.2f  # tokens / second during sampling\n",
+            1.0e6 * ctx->sampling.n_sample / ctx->sampling.t_sample_us);
+}
+
+// For internal test use
+const std::vector> & llama_internal_get_tensor_map(
+    struct llama_context * ctx
+) {
+    return ctx->model.tensors_by_name;
+}
+
+void llama_log_set(ggml_log_callback log_callback, void * user_data) {
+    g_state.log_callback = log_callback ? log_callback : llama_log_callback_default;
+    g_state.log_callback_user_data = user_data;
+#ifdef GGML_USE_METAL
+    ggml_backend_metal_log_set_callback(g_state.log_callback, g_state.log_callback_user_data);
+#elif defined(GGML_USE_CUDA)
+    ggml_backend_cuda_log_set_callback(g_state.log_callback, g_state.log_callback_user_data);
+#elif defined(GGML_USE_CANN)
+    ggml_backend_cann_log_set_callback(g_state.log_callback, g_state.log_callback_user_data);
+#endif
+}
+
+static void llama_log_internal_v(ggml_log_level level, const char * format, va_list args) {
+    va_list args_copy;
+    va_copy(args_copy, args);
+    char buffer[128];
+    int len = vsnprintf(buffer, 128, format, args);
+    if (len < 128) {
+        g_state.log_callback(level, buffer, g_state.log_callback_user_data);
+    } else {
+        char* buffer2 = new char[len+1];
+        vsnprintf(buffer2, len+1, format, args_copy);
+        buffer2[len] = 0;
+        g_state.log_callback(level, buffer2, g_state.log_callback_user_data);
+        delete[] buffer2;
+    }
+    va_end(args_copy);
+}
+
+void llama_log_internal(ggml_log_level level, const char * format, ...) {
+    va_list args;
+    va_start(args, format);
+    llama_log_internal_v(level, format, args);
+    va_end(args);
+}
+
+void llama_log_callback_default(ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) user_data;
+    fputs(text, stderr);
+    fflush(stderr);
+}
diff --git a/src/unicode-data.cpp b/src/unicode-data.cpp
new file mode 100644
index 00000000..02bdf782
--- /dev/null
+++ b/src/unicode-data.cpp
@@ -0,0 +1,7032 @@
+// generated with scripts/gen-unicode-data.py
+
+#include "unicode-data.h"
+
+#include 
+#include 
+#include 
+#include 
+
+const std::vector> unicode_ranges_flags = {  // start, flags // last=next_start-1
+{0x000000, 0x0080},
+{0x000020, 0x0008},
+{0x000021, 0x0020},
+{0x000024, 0x0040},
+{0x000025, 0x0020},
+{0x00002B, 0x0040},
+{0x00002C, 0x0020},
+{0x000030, 0x0002},
+{0x00003A, 0x0020},
+{0x00003C, 0x0040},
+{0x00003F, 0x0020},
+{0x000041, 0x0004},
+{0x00005B, 0x0020},
+{0x00005E, 0x0040},
+{0x00005F, 0x0020},
+{0x000060, 0x0040},
+{0x000061, 0x0004},
+{0x00007B, 0x0020},
+{0x00007C, 0x0040},
+{0x00007D, 0x0020},
+{0x00007E, 0x0040},
+{0x00007F, 0x0080},
+{0x0000A0, 0x0008},
+{0x0000A1, 0x0020},
+{0x0000A2, 0x0040},
+{0x0000A7, 0x0020},
+{0x0000A8, 0x0040},
+{0x0000AA, 0x0004},
+{0x0000AB, 0x0020},
+{0x0000AC, 0x0040},
+{0x0000AD, 0x0080},
+{0x0000AE, 0x0040},
+{0x0000B2, 0x0002},
+{0x0000B4, 0x0040},
+{0x0000B5, 0x0004},
+{0x0000B6, 0x0020},
+{0x0000B8, 0x0040},
+{0x0000B9, 0x0002},
+{0x0000BA, 0x0004},
+{0x0000BB, 0x0020},
+{0x0000BC, 0x0002},
+{0x0000BF, 0x0020},
+{0x0000C0, 0x0004},
+{0x0000D7, 0x0040},
+{0x0000D8, 0x0004},
+{0x0000F7, 0x0040},
+{0x0000F8, 0x0004},
+{0x0002C2, 0x0040},
+{0x0002C6, 0x0004},
+{0x0002D2, 0x0040},
+{0x0002E0, 0x0004},
+{0x0002E5, 0x0040},
+{0x0002EC, 0x0004},
+{0x0002ED, 0x0040},
+{0x0002EE, 0x0004},
+{0x0002EF, 0x0040},
+{0x000300, 0x0010},
+{0x000370, 0x0004},
+{0x000375, 0x0040},
+{0x000376, 0x0004},
+{0x000378, 0x0001},
+{0x00037A, 0x0004},
+{0x00037E, 0x0020},
+{0x00037F, 0x0004},
+{0x000380, 0x0001},
+{0x000384, 0x0040},
+{0x000386, 0x0004},
+{0x000387, 0x0020},
+{0x000388, 0x0004},
+{0x00038B, 0x0001},
+{0x00038C, 0x0004},
+{0x00038D, 0x0001},
+{0x00038E, 0x0004},
+{0x0003A2, 0x0001},
+{0x0003A3, 0x0004},
+{0x0003F6, 0x0040},
+{0x0003F7, 0x0004},
+{0x000482, 0x0040},
+{0x000483, 0x0010},
+{0x00048A, 0x0004},
+{0x000530, 0x0001},
+{0x000531, 0x0004},
+{0x000557, 0x0001},
+{0x000559, 0x0004},
+{0x00055A, 0x0020},
+{0x000560, 0x0004},
+{0x000589, 0x0020},
+{0x00058B, 0x0001},
+{0x00058D, 0x0040},
+{0x000590, 0x0001},
+{0x000591, 0x0010},
+{0x0005BE, 0x0020},
+{0x0005BF, 0x0010},
+{0x0005C0, 0x0020},
+{0x0005C1, 0x0010},
+{0x0005C3, 0x0020},
+{0x0005C4, 0x0010},
+{0x0005C6, 0x0020},
+{0x0005C7, 0x0010},
+{0x0005C8, 0x0001},
+{0x0005D0, 0x0004},
+{0x0005EB, 0x0001},
+{0x0005EF, 0x0004},
+{0x0005F3, 0x0020},
+{0x0005F5, 0x0001},
+{0x000600, 0x0080},
+{0x000606, 0x0040},
+{0x000609, 0x0020},
+{0x00060B, 0x0040},
+{0x00060C, 0x0020},
+{0x00060E, 0x0040},
+{0x000610, 0x0010},
+{0x00061B, 0x0020},
+{0x00061C, 0x0080},
+{0x00061D, 0x0020},
+{0x000620, 0x0004},
+{0x00064B, 0x0010},
+{0x000660, 0x0002},
+{0x00066A, 0x0020},
+{0x00066E, 0x0004},
+{0x000670, 0x0010},
+{0x000671, 0x0004},
+{0x0006D4, 0x0020},
+{0x0006D5, 0x0004},
+{0x0006D6, 0x0010},
+{0x0006DD, 0x0080},
+{0x0006DE, 0x0040},
+{0x0006DF, 0x0010},
+{0x0006E5, 0x0004},
+{0x0006E7, 0x0010},
+{0x0006E9, 0x0040},
+{0x0006EA, 0x0010},
+{0x0006EE, 0x0004},
+{0x0006F0, 0x0002},
+{0x0006FA, 0x0004},
+{0x0006FD, 0x0040},
+{0x0006FF, 0x0004},
+{0x000700, 0x0020},
+{0x00070E, 0x0001},
+{0x00070F, 0x0080},
+{0x000710, 0x0004},
+{0x000711, 0x0010},
+{0x000712, 0x0004},
+{0x000730, 0x0010},
+{0x00074B, 0x0001},
+{0x00074D, 0x0004},
+{0x0007A6, 0x0010},
+{0x0007B1, 0x0004},
+{0x0007B2, 0x0001},
+{0x0007C0, 0x0002},
+{0x0007CA, 0x0004},
+{0x0007EB, 0x0010},
+{0x0007F4, 0x0004},
+{0x0007F6, 0x0040},
+{0x0007F7, 0x0020},
+{0x0007FA, 0x0004},
+{0x0007FB, 0x0001},
+{0x0007FD, 0x0010},
+{0x0007FE, 0x0040},
+{0x000800, 0x0004},
+{0x000816, 0x0010},
+{0x00081A, 0x0004},
+{0x00081B, 0x0010},
+{0x000824, 0x0004},
+{0x000825, 0x0010},
+{0x000828, 0x0004},
+{0x000829, 0x0010},
+{0x00082E, 0x0001},
+{0x000830, 0x0020},
+{0x00083F, 0x0001},
+{0x000840, 0x0004},
+{0x000859, 0x0010},
+{0x00085C, 0x0001},
+{0x00085E, 0x0020},
+{0x00085F, 0x0001},
+{0x000860, 0x0004},
+{0x00086B, 0x0001},
+{0x000870, 0x0004},
+{0x000888, 0x0040},
+{0x000889, 0x0004},
+{0x00088F, 0x0001},
+{0x000890, 0x0080},
+{0x000892, 0x0001},
+{0x000898, 0x0010},
+{0x0008A0, 0x0004},
+{0x0008CA, 0x0010},
+{0x0008E2, 0x0080},
+{0x0008E3, 0x0010},
+{0x000904, 0x0004},
+{0x00093A, 0x0010},
+{0x00093D, 0x0004},
+{0x00093E, 0x0010},
+{0x000950, 0x0004},
+{0x000951, 0x0010},
+{0x000958, 0x0004},
+{0x000962, 0x0010},
+{0x000964, 0x0020},
+{0x000966, 0x0002},
+{0x000970, 0x0020},
+{0x000971, 0x0004},
+{0x000981, 0x0010},
+{0x000984, 0x0001},
+{0x000985, 0x0004},
+{0x00098D, 0x0001},
+{0x00098F, 0x0004},
+{0x000991, 0x0001},
+{0x000993, 0x0004},
+{0x0009A9, 0x0001},
+{0x0009AA, 0x0004},
+{0x0009B1, 0x0001},
+{0x0009B2, 0x0004},
+{0x0009B3, 0x0001},
+{0x0009B6, 0x0004},
+{0x0009BA, 0x0001},
+{0x0009BC, 0x0010},
+{0x0009BD, 0x0004},
+{0x0009BE, 0x0010},
+{0x0009C5, 0x0001},
+{0x0009C7, 0x0010},
+{0x0009C9, 0x0001},
+{0x0009CB, 0x0010},
+{0x0009CE, 0x0004},
+{0x0009CF, 0x0001},
+{0x0009D7, 0x0010},
+{0x0009D8, 0x0001},
+{0x0009DC, 0x0004},
+{0x0009DE, 0x0001},
+{0x0009DF, 0x0004},
+{0x0009E2, 0x0010},
+{0x0009E4, 0x0001},
+{0x0009E6, 0x0002},
+{0x0009F0, 0x0004},
+{0x0009F2, 0x0040},
+{0x0009F4, 0x0002},
+{0x0009FA, 0x0040},
+{0x0009FC, 0x0004},
+{0x0009FD, 0x0020},
+{0x0009FE, 0x0010},
+{0x0009FF, 0x0001},
+{0x000A01, 0x0010},
+{0x000A04, 0x0001},
+{0x000A05, 0x0004},
+{0x000A0B, 0x0001},
+{0x000A0F, 0x0004},
+{0x000A11, 0x0001},
+{0x000A13, 0x0004},
+{0x000A29, 0x0001},
+{0x000A2A, 0x0004},
+{0x000A31, 0x0001},
+{0x000A32, 0x0004},
+{0x000A34, 0x0001},
+{0x000A35, 0x0004},
+{0x000A37, 0x0001},
+{0x000A38, 0x0004},
+{0x000A3A, 0x0001},
+{0x000A3C, 0x0010},
+{0x000A3D, 0x0001},
+{0x000A3E, 0x0010},
+{0x000A43, 0x0001},
+{0x000A47, 0x0010},
+{0x000A49, 0x0001},
+{0x000A4B, 0x0010},
+{0x000A4E, 0x0001},
+{0x000A51, 0x0010},
+{0x000A52, 0x0001},
+{0x000A59, 0x0004},
+{0x000A5D, 0x0001},
+{0x000A5E, 0x0004},
+{0x000A5F, 0x0001},
+{0x000A66, 0x0002},
+{0x000A70, 0x0010},
+{0x000A72, 0x0004},
+{0x000A75, 0x0010},
+{0x000A76, 0x0020},
+{0x000A77, 0x0001},
+{0x000A81, 0x0010},
+{0x000A84, 0x0001},
+{0x000A85, 0x0004},
+{0x000A8E, 0x0001},
+{0x000A8F, 0x0004},
+{0x000A92, 0x0001},
+{0x000A93, 0x0004},
+{0x000AA9, 0x0001},
+{0x000AAA, 0x0004},
+{0x000AB1, 0x0001},
+{0x000AB2, 0x0004},
+{0x000AB4, 0x0001},
+{0x000AB5, 0x0004},
+{0x000ABA, 0x0001},
+{0x000ABC, 0x0010},
+{0x000ABD, 0x0004},
+{0x000ABE, 0x0010},
+{0x000AC6, 0x0001},
+{0x000AC7, 0x0010},
+{0x000ACA, 0x0001},
+{0x000ACB, 0x0010},
+{0x000ACE, 0x0001},
+{0x000AD0, 0x0004},
+{0x000AD1, 0x0001},
+{0x000AE0, 0x0004},
+{0x000AE2, 0x0010},
+{0x000AE4, 0x0001},
+{0x000AE6, 0x0002},
+{0x000AF0, 0x0020},
+{0x000AF1, 0x0040},
+{0x000AF2, 0x0001},
+{0x000AF9, 0x0004},
+{0x000AFA, 0x0010},
+{0x000B00, 0x0001},
+{0x000B01, 0x0010},
+{0x000B04, 0x0001},
+{0x000B05, 0x0004},
+{0x000B0D, 0x0001},
+{0x000B0F, 0x0004},
+{0x000B11, 0x0001},
+{0x000B13, 0x0004},
+{0x000B29, 0x0001},
+{0x000B2A, 0x0004},
+{0x000B31, 0x0001},
+{0x000B32, 0x0004},
+{0x000B34, 0x0001},
+{0x000B35, 0x0004},
+{0x000B3A, 0x0001},
+{0x000B3C, 0x0010},
+{0x000B3D, 0x0004},
+{0x000B3E, 0x0010},
+{0x000B45, 0x0001},
+{0x000B47, 0x0010},
+{0x000B49, 0x0001},
+{0x000B4B, 0x0010},
+{0x000B4E, 0x0001},
+{0x000B55, 0x0010},
+{0x000B58, 0x0001},
+{0x000B5C, 0x0004},
+{0x000B5E, 0x0001},
+{0x000B5F, 0x0004},
+{0x000B62, 0x0010},
+{0x000B64, 0x0001},
+{0x000B66, 0x0002},
+{0x000B70, 0x0040},
+{0x000B71, 0x0004},
+{0x000B72, 0x0002},
+{0x000B78, 0x0001},
+{0x000B82, 0x0010},
+{0x000B83, 0x0004},
+{0x000B84, 0x0001},
+{0x000B85, 0x0004},
+{0x000B8B, 0x0001},
+{0x000B8E, 0x0004},
+{0x000B91, 0x0001},
+{0x000B92, 0x0004},
+{0x000B96, 0x0001},
+{0x000B99, 0x0004},
+{0x000B9B, 0x0001},
+{0x000B9C, 0x0004},
+{0x000B9D, 0x0001},
+{0x000B9E, 0x0004},
+{0x000BA0, 0x0001},
+{0x000BA3, 0x0004},
+{0x000BA5, 0x0001},
+{0x000BA8, 0x0004},
+{0x000BAB, 0x0001},
+{0x000BAE, 0x0004},
+{0x000BBA, 0x0001},
+{0x000BBE, 0x0010},
+{0x000BC3, 0x0001},
+{0x000BC6, 0x0010},
+{0x000BC9, 0x0001},
+{0x000BCA, 0x0010},
+{0x000BCE, 0x0001},
+{0x000BD0, 0x0004},
+{0x000BD1, 0x0001},
+{0x000BD7, 0x0010},
+{0x000BD8, 0x0001},
+{0x000BE6, 0x0002},
+{0x000BF3, 0x0040},
+{0x000BFB, 0x0001},
+{0x000C00, 0x0010},
+{0x000C05, 0x0004},
+{0x000C0D, 0x0001},
+{0x000C0E, 0x0004},
+{0x000C11, 0x0001},
+{0x000C12, 0x0004},
+{0x000C29, 0x0001},
+{0x000C2A, 0x0004},
+{0x000C3A, 0x0001},
+{0x000C3C, 0x0010},
+{0x000C3D, 0x0004},
+{0x000C3E, 0x0010},
+{0x000C45, 0x0001},
+{0x000C46, 0x0010},
+{0x000C49, 0x0001},
+{0x000C4A, 0x0010},
+{0x000C4E, 0x0001},
+{0x000C55, 0x0010},
+{0x000C57, 0x0001},
+{0x000C58, 0x0004},
+{0x000C5B, 0x0001},
+{0x000C5D, 0x0004},
+{0x000C5E, 0x0001},
+{0x000C60, 0x0004},
+{0x000C62, 0x0010},
+{0x000C64, 0x0001},
+{0x000C66, 0x0002},
+{0x000C70, 0x0001},
+{0x000C77, 0x0020},
+{0x000C78, 0x0002},
+{0x000C7F, 0x0040},
+{0x000C80, 0x0004},
+{0x000C81, 0x0010},
+{0x000C84, 0x0020},
+{0x000C85, 0x0004},
+{0x000C8D, 0x0001},
+{0x000C8E, 0x0004},
+{0x000C91, 0x0001},
+{0x000C92, 0x0004},
+{0x000CA9, 0x0001},
+{0x000CAA, 0x0004},
+{0x000CB4, 0x0001},
+{0x000CB5, 0x0004},
+{0x000CBA, 0x0001},
+{0x000CBC, 0x0010},
+{0x000CBD, 0x0004},
+{0x000CBE, 0x0010},
+{0x000CC5, 0x0001},
+{0x000CC6, 0x0010},
+{0x000CC9, 0x0001},
+{0x000CCA, 0x0010},
+{0x000CCE, 0x0001},
+{0x000CD5, 0x0010},
+{0x000CD7, 0x0001},
+{0x000CDD, 0x0004},
+{0x000CDF, 0x0001},
+{0x000CE0, 0x0004},
+{0x000CE2, 0x0010},
+{0x000CE4, 0x0001},
+{0x000CE6, 0x0002},
+{0x000CF0, 0x0001},
+{0x000CF1, 0x0004},
+{0x000CF3, 0x0010},
+{0x000CF4, 0x0001},
+{0x000D00, 0x0010},
+{0x000D04, 0x0004},
+{0x000D0D, 0x0001},
+{0x000D0E, 0x0004},
+{0x000D11, 0x0001},
+{0x000D12, 0x0004},
+{0x000D3B, 0x0010},
+{0x000D3D, 0x0004},
+{0x000D3E, 0x0010},
+{0x000D45, 0x0001},
+{0x000D46, 0x0010},
+{0x000D49, 0x0001},
+{0x000D4A, 0x0010},
+{0x000D4E, 0x0004},
+{0x000D4F, 0x0040},
+{0x000D50, 0x0001},
+{0x000D54, 0x0004},
+{0x000D57, 0x0010},
+{0x000D58, 0x0002},
+{0x000D5F, 0x0004},
+{0x000D62, 0x0010},
+{0x000D64, 0x0001},
+{0x000D66, 0x0002},
+{0x000D79, 0x0040},
+{0x000D7A, 0x0004},
+{0x000D80, 0x0001},
+{0x000D81, 0x0010},
+{0x000D84, 0x0001},
+{0x000D85, 0x0004},
+{0x000D97, 0x0001},
+{0x000D9A, 0x0004},
+{0x000DB2, 0x0001},
+{0x000DB3, 0x0004},
+{0x000DBC, 0x0001},
+{0x000DBD, 0x0004},
+{0x000DBE, 0x0001},
+{0x000DC0, 0x0004},
+{0x000DC7, 0x0001},
+{0x000DCA, 0x0010},
+{0x000DCB, 0x0001},
+{0x000DCF, 0x0010},
+{0x000DD5, 0x0001},
+{0x000DD6, 0x0010},
+{0x000DD7, 0x0001},
+{0x000DD8, 0x0010},
+{0x000DE0, 0x0001},
+{0x000DE6, 0x0002},
+{0x000DF0, 0x0001},
+{0x000DF2, 0x0010},
+{0x000DF4, 0x0020},
+{0x000DF5, 0x0001},
+{0x000E01, 0x0004},
+{0x000E31, 0x0010},
+{0x000E32, 0x0004},
+{0x000E34, 0x0010},
+{0x000E3B, 0x0001},
+{0x000E3F, 0x0040},
+{0x000E40, 0x0004},
+{0x000E47, 0x0010},
+{0x000E4F, 0x0020},
+{0x000E50, 0x0002},
+{0x000E5A, 0x0020},
+{0x000E5C, 0x0001},
+{0x000E81, 0x0004},
+{0x000E83, 0x0001},
+{0x000E84, 0x0004},
+{0x000E85, 0x0001},
+{0x000E86, 0x0004},
+{0x000E8B, 0x0001},
+{0x000E8C, 0x0004},
+{0x000EA4, 0x0001},
+{0x000EA5, 0x0004},
+{0x000EA6, 0x0001},
+{0x000EA7, 0x0004},
+{0x000EB1, 0x0010},
+{0x000EB2, 0x0004},
+{0x000EB4, 0x0010},
+{0x000EBD, 0x0004},
+{0x000EBE, 0x0001},
+{0x000EC0, 0x0004},
+{0x000EC5, 0x0001},
+{0x000EC6, 0x0004},
+{0x000EC7, 0x0001},
+{0x000EC8, 0x0010},
+{0x000ECF, 0x0001},
+{0x000ED0, 0x0002},
+{0x000EDA, 0x0001},
+{0x000EDC, 0x0004},
+{0x000EE0, 0x0001},
+{0x000F00, 0x0004},
+{0x000F01, 0x0040},
+{0x000F04, 0x0020},
+{0x000F13, 0x0040},
+{0x000F14, 0x0020},
+{0x000F15, 0x0040},
+{0x000F18, 0x0010},
+{0x000F1A, 0x0040},
+{0x000F20, 0x0002},
+{0x000F34, 0x0040},
+{0x000F35, 0x0010},
+{0x000F36, 0x0040},
+{0x000F37, 0x0010},
+{0x000F38, 0x0040},
+{0x000F39, 0x0010},
+{0x000F3A, 0x0020},
+{0x000F3E, 0x0010},
+{0x000F40, 0x0004},
+{0x000F48, 0x0001},
+{0x000F49, 0x0004},
+{0x000F6D, 0x0001},
+{0x000F71, 0x0010},
+{0x000F85, 0x0020},
+{0x000F86, 0x0010},
+{0x000F88, 0x0004},
+{0x000F8D, 0x0010},
+{0x000F98, 0x0001},
+{0x000F99, 0x0010},
+{0x000FBD, 0x0001},
+{0x000FBE, 0x0040},
+{0x000FC6, 0x0010},
+{0x000FC7, 0x0040},
+{0x000FCD, 0x0001},
+{0x000FCE, 0x0040},
+{0x000FD0, 0x0020},
+{0x000FD5, 0x0040},
+{0x000FD9, 0x0020},
+{0x000FDB, 0x0001},
+{0x001000, 0x0004},
+{0x00102B, 0x0010},
+{0x00103F, 0x0004},
+{0x001040, 0x0002},
+{0x00104A, 0x0020},
+{0x001050, 0x0004},
+{0x001056, 0x0010},
+{0x00105A, 0x0004},
+{0x00105E, 0x0010},
+{0x001061, 0x0004},
+{0x001062, 0x0010},
+{0x001065, 0x0004},
+{0x001067, 0x0010},
+{0x00106E, 0x0004},
+{0x001071, 0x0010},
+{0x001075, 0x0004},
+{0x001082, 0x0010},
+{0x00108E, 0x0004},
+{0x00108F, 0x0010},
+{0x001090, 0x0002},
+{0x00109A, 0x0010},
+{0x00109E, 0x0040},
+{0x0010A0, 0x0004},
+{0x0010C6, 0x0001},
+{0x0010C7, 0x0004},
+{0x0010C8, 0x0001},
+{0x0010CD, 0x0004},
+{0x0010CE, 0x0001},
+{0x0010D0, 0x0004},
+{0x0010FB, 0x0020},
+{0x0010FC, 0x0004},
+{0x001249, 0x0001},
+{0x00124A, 0x0004},
+{0x00124E, 0x0001},
+{0x001250, 0x0004},
+{0x001257, 0x0001},
+{0x001258, 0x0004},
+{0x001259, 0x0001},
+{0x00125A, 0x0004},
+{0x00125E, 0x0001},
+{0x001260, 0x0004},
+{0x001289, 0x0001},
+{0x00128A, 0x0004},
+{0x00128E, 0x0001},
+{0x001290, 0x0004},
+{0x0012B1, 0x0001},
+{0x0012B2, 0x0004},
+{0x0012B6, 0x0001},
+{0x0012B8, 0x0004},
+{0x0012BF, 0x0001},
+{0x0012C0, 0x0004},
+{0x0012C1, 0x0001},
+{0x0012C2, 0x0004},
+{0x0012C6, 0x0001},
+{0x0012C8, 0x0004},
+{0x0012D7, 0x0001},
+{0x0012D8, 0x0004},
+{0x001311, 0x0001},
+{0x001312, 0x0004},
+{0x001316, 0x0001},
+{0x001318, 0x0004},
+{0x00135B, 0x0001},
+{0x00135D, 0x0010},
+{0x001360, 0x0020},
+{0x001369, 0x0002},
+{0x00137D, 0x0001},
+{0x001380, 0x0004},
+{0x001390, 0x0040},
+{0x00139A, 0x0001},
+{0x0013A0, 0x0004},
+{0x0013F6, 0x0001},
+{0x0013F8, 0x0004},
+{0x0013FE, 0x0001},
+{0x001400, 0x0020},
+{0x001401, 0x0004},
+{0x00166D, 0x0040},
+{0x00166E, 0x0020},
+{0x00166F, 0x0004},
+{0x001680, 0x0008},
+{0x001681, 0x0004},
+{0x00169B, 0x0020},
+{0x00169D, 0x0001},
+{0x0016A0, 0x0004},
+{0x0016EB, 0x0020},
+{0x0016EE, 0x0002},
+{0x0016F1, 0x0004},
+{0x0016F9, 0x0001},
+{0x001700, 0x0004},
+{0x001712, 0x0010},
+{0x001716, 0x0001},
+{0x00171F, 0x0004},
+{0x001732, 0x0010},
+{0x001735, 0x0020},
+{0x001737, 0x0001},
+{0x001740, 0x0004},
+{0x001752, 0x0010},
+{0x001754, 0x0001},
+{0x001760, 0x0004},
+{0x00176D, 0x0001},
+{0x00176E, 0x0004},
+{0x001771, 0x0001},
+{0x001772, 0x0010},
+{0x001774, 0x0001},
+{0x001780, 0x0004},
+{0x0017B4, 0x0010},
+{0x0017D4, 0x0020},
+{0x0017D7, 0x0004},
+{0x0017D8, 0x0020},
+{0x0017DB, 0x0040},
+{0x0017DC, 0x0004},
+{0x0017DD, 0x0010},
+{0x0017DE, 0x0001},
+{0x0017E0, 0x0002},
+{0x0017EA, 0x0001},
+{0x0017F0, 0x0002},
+{0x0017FA, 0x0001},
+{0x001800, 0x0020},
+{0x00180B, 0x0010},
+{0x00180E, 0x0080},
+{0x00180F, 0x0010},
+{0x001810, 0x0002},
+{0x00181A, 0x0001},
+{0x001820, 0x0004},
+{0x001879, 0x0001},
+{0x001880, 0x0004},
+{0x001885, 0x0010},
+{0x001887, 0x0004},
+{0x0018A9, 0x0010},
+{0x0018AA, 0x0004},
+{0x0018AB, 0x0001},
+{0x0018B0, 0x0004},
+{0x0018F6, 0x0001},
+{0x001900, 0x0004},
+{0x00191F, 0x0001},
+{0x001920, 0x0010},
+{0x00192C, 0x0001},
+{0x001930, 0x0010},
+{0x00193C, 0x0001},
+{0x001940, 0x0040},
+{0x001941, 0x0001},
+{0x001944, 0x0020},
+{0x001946, 0x0002},
+{0x001950, 0x0004},
+{0x00196E, 0x0001},
+{0x001970, 0x0004},
+{0x001975, 0x0001},
+{0x001980, 0x0004},
+{0x0019AC, 0x0001},
+{0x0019B0, 0x0004},
+{0x0019CA, 0x0001},
+{0x0019D0, 0x0002},
+{0x0019DB, 0x0001},
+{0x0019DE, 0x0040},
+{0x001A00, 0x0004},
+{0x001A17, 0x0010},
+{0x001A1C, 0x0001},
+{0x001A1E, 0x0020},
+{0x001A20, 0x0004},
+{0x001A55, 0x0010},
+{0x001A5F, 0x0001},
+{0x001A60, 0x0010},
+{0x001A7D, 0x0001},
+{0x001A7F, 0x0010},
+{0x001A80, 0x0002},
+{0x001A8A, 0x0001},
+{0x001A90, 0x0002},
+{0x001A9A, 0x0001},
+{0x001AA0, 0x0020},
+{0x001AA7, 0x0004},
+{0x001AA8, 0x0020},
+{0x001AAE, 0x0001},
+{0x001AB0, 0x0010},
+{0x001ACF, 0x0001},
+{0x001B00, 0x0010},
+{0x001B05, 0x0004},
+{0x001B34, 0x0010},
+{0x001B45, 0x0004},
+{0x001B4D, 0x0001},
+{0x001B50, 0x0002},
+{0x001B5A, 0x0020},
+{0x001B61, 0x0040},
+{0x001B6B, 0x0010},
+{0x001B74, 0x0040},
+{0x001B7D, 0x0020},
+{0x001B7F, 0x0001},
+{0x001B80, 0x0010},
+{0x001B83, 0x0004},
+{0x001BA1, 0x0010},
+{0x001BAE, 0x0004},
+{0x001BB0, 0x0002},
+{0x001BBA, 0x0004},
+{0x001BE6, 0x0010},
+{0x001BF4, 0x0001},
+{0x001BFC, 0x0020},
+{0x001C00, 0x0004},
+{0x001C24, 0x0010},
+{0x001C38, 0x0001},
+{0x001C3B, 0x0020},
+{0x001C40, 0x0002},
+{0x001C4A, 0x0001},
+{0x001C4D, 0x0004},
+{0x001C50, 0x0002},
+{0x001C5A, 0x0004},
+{0x001C7E, 0x0020},
+{0x001C80, 0x0004},
+{0x001C89, 0x0001},
+{0x001C90, 0x0004},
+{0x001CBB, 0x0001},
+{0x001CBD, 0x0004},
+{0x001CC0, 0x0020},
+{0x001CC8, 0x0001},
+{0x001CD0, 0x0010},
+{0x001CD3, 0x0020},
+{0x001CD4, 0x0010},
+{0x001CE9, 0x0004},
+{0x001CED, 0x0010},
+{0x001CEE, 0x0004},
+{0x001CF4, 0x0010},
+{0x001CF5, 0x0004},
+{0x001CF7, 0x0010},
+{0x001CFA, 0x0004},
+{0x001CFB, 0x0001},
+{0x001D00, 0x0004},
+{0x001DC0, 0x0010},
+{0x001E00, 0x0004},
+{0x001F16, 0x0001},
+{0x001F18, 0x0004},
+{0x001F1E, 0x0001},
+{0x001F20, 0x0004},
+{0x001F46, 0x0001},
+{0x001F48, 0x0004},
+{0x001F4E, 0x0001},
+{0x001F50, 0x0004},
+{0x001F58, 0x0001},
+{0x001F59, 0x0004},
+{0x001F5A, 0x0001},
+{0x001F5B, 0x0004},
+{0x001F5C, 0x0001},
+{0x001F5D, 0x0004},
+{0x001F5E, 0x0001},
+{0x001F5F, 0x0004},
+{0x001F7E, 0x0001},
+{0x001F80, 0x0004},
+{0x001FB5, 0x0001},
+{0x001FB6, 0x0004},
+{0x001FBD, 0x0040},
+{0x001FBE, 0x0004},
+{0x001FBF, 0x0040},
+{0x001FC2, 0x0004},
+{0x001FC5, 0x0001},
+{0x001FC6, 0x0004},
+{0x001FCD, 0x0040},
+{0x001FD0, 0x0004},
+{0x001FD4, 0x0001},
+{0x001FD6, 0x0004},
+{0x001FDC, 0x0001},
+{0x001FDD, 0x0040},
+{0x001FE0, 0x0004},
+{0x001FED, 0x0040},
+{0x001FF0, 0x0001},
+{0x001FF2, 0x0004},
+{0x001FF5, 0x0001},
+{0x001FF6, 0x0004},
+{0x001FFD, 0x0040},
+{0x001FFF, 0x0001},
+{0x002000, 0x0008},
+{0x00200B, 0x0080},
+{0x002010, 0x0020},
+{0x002028, 0x0008},
+{0x00202A, 0x0080},
+{0x00202F, 0x0008},
+{0x002030, 0x0020},
+{0x002044, 0x0040},
+{0x002045, 0x0020},
+{0x002052, 0x0040},
+{0x002053, 0x0020},
+{0x00205F, 0x0008},
+{0x002060, 0x0080},
+{0x002065, 0x0001},
+{0x002066, 0x0080},
+{0x002070, 0x0002},
+{0x002071, 0x0004},
+{0x002072, 0x0001},
+{0x002074, 0x0002},
+{0x00207A, 0x0040},
+{0x00207D, 0x0020},
+{0x00207F, 0x0004},
+{0x002080, 0x0002},
+{0x00208A, 0x0040},
+{0x00208D, 0x0020},
+{0x00208F, 0x0001},
+{0x002090, 0x0004},
+{0x00209D, 0x0001},
+{0x0020A0, 0x0040},
+{0x0020C1, 0x0001},
+{0x0020D0, 0x0010},
+{0x0020F1, 0x0001},
+{0x002100, 0x0040},
+{0x002102, 0x0004},
+{0x002103, 0x0040},
+{0x002107, 0x0004},
+{0x002108, 0x0040},
+{0x00210A, 0x0004},
+{0x002114, 0x0040},
+{0x002115, 0x0004},
+{0x002116, 0x0040},
+{0x002119, 0x0004},
+{0x00211E, 0x0040},
+{0x002124, 0x0004},
+{0x002125, 0x0040},
+{0x002126, 0x0004},
+{0x002127, 0x0040},
+{0x002128, 0x0004},
+{0x002129, 0x0040},
+{0x00212A, 0x0004},
+{0x00212E, 0x0040},
+{0x00212F, 0x0004},
+{0x00213A, 0x0040},
+{0x00213C, 0x0004},
+{0x002140, 0x0040},
+{0x002145, 0x0004},
+{0x00214A, 0x0040},
+{0x00214E, 0x0004},
+{0x00214F, 0x0040},
+{0x002150, 0x0002},
+{0x002183, 0x0004},
+{0x002185, 0x0002},
+{0x00218A, 0x0040},
+{0x00218C, 0x0001},
+{0x002190, 0x0040},
+{0x002308, 0x0020},
+{0x00230C, 0x0040},
+{0x002329, 0x0020},
+{0x00232B, 0x0040},
+{0x002427, 0x0001},
+{0x002440, 0x0040},
+{0x00244B, 0x0001},
+{0x002460, 0x0002},
+{0x00249C, 0x0040},
+{0x0024EA, 0x0002},
+{0x002500, 0x0040},
+{0x002768, 0x0020},
+{0x002776, 0x0002},
+{0x002794, 0x0040},
+{0x0027C5, 0x0020},
+{0x0027C7, 0x0040},
+{0x0027E6, 0x0020},
+{0x0027F0, 0x0040},
+{0x002983, 0x0020},
+{0x002999, 0x0040},
+{0x0029D8, 0x0020},
+{0x0029DC, 0x0040},
+{0x0029FC, 0x0020},
+{0x0029FE, 0x0040},
+{0x002B74, 0x0001},
+{0x002B76, 0x0040},
+{0x002B96, 0x0001},
+{0x002B97, 0x0040},
+{0x002C00, 0x0004},
+{0x002CE5, 0x0040},
+{0x002CEB, 0x0004},
+{0x002CEF, 0x0010},
+{0x002CF2, 0x0004},
+{0x002CF4, 0x0001},
+{0x002CF9, 0x0020},
+{0x002CFD, 0x0002},
+{0x002CFE, 0x0020},
+{0x002D00, 0x0004},
+{0x002D26, 0x0001},
+{0x002D27, 0x0004},
+{0x002D28, 0x0001},
+{0x002D2D, 0x0004},
+{0x002D2E, 0x0001},
+{0x002D30, 0x0004},
+{0x002D68, 0x0001},
+{0x002D6F, 0x0004},
+{0x002D70, 0x0020},
+{0x002D71, 0x0001},
+{0x002D7F, 0x0010},
+{0x002D80, 0x0004},
+{0x002D97, 0x0001},
+{0x002DA0, 0x0004},
+{0x002DA7, 0x0001},
+{0x002DA8, 0x0004},
+{0x002DAF, 0x0001},
+{0x002DB0, 0x0004},
+{0x002DB7, 0x0001},
+{0x002DB8, 0x0004},
+{0x002DBF, 0x0001},
+{0x002DC0, 0x0004},
+{0x002DC7, 0x0001},
+{0x002DC8, 0x0004},
+{0x002DCF, 0x0001},
+{0x002DD0, 0x0004},
+{0x002DD7, 0x0001},
+{0x002DD8, 0x0004},
+{0x002DDF, 0x0001},
+{0x002DE0, 0x0010},
+{0x002E00, 0x0020},
+{0x002E2F, 0x0004},
+{0x002E30, 0x0020},
+{0x002E50, 0x0040},
+{0x002E52, 0x0020},
+{0x002E5E, 0x0001},
+{0x002E80, 0x0040},
+{0x002E9A, 0x0001},
+{0x002E9B, 0x0040},
+{0x002EF4, 0x0001},
+{0x002F00, 0x0040},
+{0x002FD6, 0x0001},
+{0x002FF0, 0x0040},
+{0x003000, 0x0008},
+{0x003001, 0x0020},
+{0x003004, 0x0040},
+{0x003005, 0x0004},
+{0x003007, 0x0002},
+{0x003008, 0x0020},
+{0x003012, 0x0040},
+{0x003014, 0x0020},
+{0x003020, 0x0040},
+{0x003021, 0x0002},
+{0x00302A, 0x0010},
+{0x003030, 0x0020},
+{0x003031, 0x0004},
+{0x003036, 0x0040},
+{0x003038, 0x0002},
+{0x00303B, 0x0004},
+{0x00303D, 0x0020},
+{0x00303E, 0x0040},
+{0x003040, 0x0001},
+{0x003041, 0x0004},
+{0x003097, 0x0001},
+{0x003099, 0x0010},
+{0x00309B, 0x0040},
+{0x00309D, 0x0004},
+{0x0030A0, 0x0020},
+{0x0030A1, 0x0004},
+{0x0030FB, 0x0020},
+{0x0030FC, 0x0004},
+{0x003100, 0x0001},
+{0x003105, 0x0004},
+{0x003130, 0x0001},
+{0x003131, 0x0004},
+{0x00318F, 0x0001},
+{0x003190, 0x0040},
+{0x003192, 0x0002},
+{0x003196, 0x0040},
+{0x0031A0, 0x0004},
+{0x0031C0, 0x0040},
+{0x0031E4, 0x0001},
+{0x0031EF, 0x0040},
+{0x0031F0, 0x0004},
+{0x003200, 0x0040},
+{0x00321F, 0x0001},
+{0x003220, 0x0002},
+{0x00322A, 0x0040},
+{0x003248, 0x0002},
+{0x003250, 0x0040},
+{0x003251, 0x0002},
+{0x003260, 0x0040},
+{0x003280, 0x0002},
+{0x00328A, 0x0040},
+{0x0032B1, 0x0002},
+{0x0032C0, 0x0040},
+{0x003400, 0x0004},
+{0x004DC0, 0x0040},
+{0x004E00, 0x0004},
+{0x00A48D, 0x0001},
+{0x00A490, 0x0040},
+{0x00A4C7, 0x0001},
+{0x00A4D0, 0x0004},
+{0x00A4FE, 0x0020},
+{0x00A500, 0x0004},
+{0x00A60D, 0x0020},
+{0x00A610, 0x0004},
+{0x00A620, 0x0002},
+{0x00A62A, 0x0004},
+{0x00A62C, 0x0001},
+{0x00A640, 0x0004},
+{0x00A66F, 0x0010},
+{0x00A673, 0x0020},
+{0x00A674, 0x0010},
+{0x00A67E, 0x0020},
+{0x00A67F, 0x0004},
+{0x00A69E, 0x0010},
+{0x00A6A0, 0x0004},
+{0x00A6E6, 0x0002},
+{0x00A6F0, 0x0010},
+{0x00A6F2, 0x0020},
+{0x00A6F8, 0x0001},
+{0x00A700, 0x0040},
+{0x00A717, 0x0004},
+{0x00A720, 0x0040},
+{0x00A722, 0x0004},
+{0x00A789, 0x0040},
+{0x00A78B, 0x0004},
+{0x00A7CB, 0x0001},
+{0x00A7D0, 0x0004},
+{0x00A7D2, 0x0001},
+{0x00A7D3, 0x0004},
+{0x00A7D4, 0x0001},
+{0x00A7D5, 0x0004},
+{0x00A7DA, 0x0001},
+{0x00A7F2, 0x0004},
+{0x00A802, 0x0010},
+{0x00A803, 0x0004},
+{0x00A806, 0x0010},
+{0x00A807, 0x0004},
+{0x00A80B, 0x0010},
+{0x00A80C, 0x0004},
+{0x00A823, 0x0010},
+{0x00A828, 0x0040},
+{0x00A82C, 0x0010},
+{0x00A82D, 0x0001},
+{0x00A830, 0x0002},
+{0x00A836, 0x0040},
+{0x00A83A, 0x0001},
+{0x00A840, 0x0004},
+{0x00A874, 0x0020},
+{0x00A878, 0x0001},
+{0x00A880, 0x0010},
+{0x00A882, 0x0004},
+{0x00A8B4, 0x0010},
+{0x00A8C6, 0x0001},
+{0x00A8CE, 0x0020},
+{0x00A8D0, 0x0002},
+{0x00A8DA, 0x0001},
+{0x00A8E0, 0x0010},
+{0x00A8F2, 0x0004},
+{0x00A8F8, 0x0020},
+{0x00A8FB, 0x0004},
+{0x00A8FC, 0x0020},
+{0x00A8FD, 0x0004},
+{0x00A8FF, 0x0010},
+{0x00A900, 0x0002},
+{0x00A90A, 0x0004},
+{0x00A926, 0x0010},
+{0x00A92E, 0x0020},
+{0x00A930, 0x0004},
+{0x00A947, 0x0010},
+{0x00A954, 0x0001},
+{0x00A95F, 0x0020},
+{0x00A960, 0x0004},
+{0x00A97D, 0x0001},
+{0x00A980, 0x0010},
+{0x00A984, 0x0004},
+{0x00A9B3, 0x0010},
+{0x00A9C1, 0x0020},
+{0x00A9CE, 0x0001},
+{0x00A9CF, 0x0004},
+{0x00A9D0, 0x0002},
+{0x00A9DA, 0x0001},
+{0x00A9DE, 0x0020},
+{0x00A9E0, 0x0004},
+{0x00A9E5, 0x0010},
+{0x00A9E6, 0x0004},
+{0x00A9F0, 0x0002},
+{0x00A9FA, 0x0004},
+{0x00A9FF, 0x0001},
+{0x00AA00, 0x0004},
+{0x00AA29, 0x0010},
+{0x00AA37, 0x0001},
+{0x00AA40, 0x0004},
+{0x00AA43, 0x0010},
+{0x00AA44, 0x0004},
+{0x00AA4C, 0x0010},
+{0x00AA4E, 0x0001},
+{0x00AA50, 0x0002},
+{0x00AA5A, 0x0001},
+{0x00AA5C, 0x0020},
+{0x00AA60, 0x0004},
+{0x00AA77, 0x0040},
+{0x00AA7A, 0x0004},
+{0x00AA7B, 0x0010},
+{0x00AA7E, 0x0004},
+{0x00AAB0, 0x0010},
+{0x00AAB1, 0x0004},
+{0x00AAB2, 0x0010},
+{0x00AAB5, 0x0004},
+{0x00AAB7, 0x0010},
+{0x00AAB9, 0x0004},
+{0x00AABE, 0x0010},
+{0x00AAC0, 0x0004},
+{0x00AAC1, 0x0010},
+{0x00AAC2, 0x0004},
+{0x00AAC3, 0x0001},
+{0x00AADB, 0x0004},
+{0x00AADE, 0x0020},
+{0x00AAE0, 0x0004},
+{0x00AAEB, 0x0010},
+{0x00AAF0, 0x0020},
+{0x00AAF2, 0x0004},
+{0x00AAF5, 0x0010},
+{0x00AAF7, 0x0001},
+{0x00AB01, 0x0004},
+{0x00AB07, 0x0001},
+{0x00AB09, 0x0004},
+{0x00AB0F, 0x0001},
+{0x00AB11, 0x0004},
+{0x00AB17, 0x0001},
+{0x00AB20, 0x0004},
+{0x00AB27, 0x0001},
+{0x00AB28, 0x0004},
+{0x00AB2F, 0x0001},
+{0x00AB30, 0x0004},
+{0x00AB5B, 0x0040},
+{0x00AB5C, 0x0004},
+{0x00AB6A, 0x0040},
+{0x00AB6C, 0x0001},
+{0x00AB70, 0x0004},
+{0x00ABE3, 0x0010},
+{0x00ABEB, 0x0020},
+{0x00ABEC, 0x0010},
+{0x00ABEE, 0x0001},
+{0x00ABF0, 0x0002},
+{0x00ABFA, 0x0001},
+{0x00AC00, 0x0004},
+{0x00D7A4, 0x0001},
+{0x00D7B0, 0x0004},
+{0x00D7C7, 0x0001},
+{0x00D7CB, 0x0004},
+{0x00D7FC, 0x0001},
+{0x00D800, 0x0080},
+{0x00F900, 0x0004},
+{0x00FA6E, 0x0001},
+{0x00FA70, 0x0004},
+{0x00FADA, 0x0001},
+{0x00FB00, 0x0004},
+{0x00FB07, 0x0001},
+{0x00FB13, 0x0004},
+{0x00FB18, 0x0001},
+{0x00FB1D, 0x0004},
+{0x00FB1E, 0x0010},
+{0x00FB1F, 0x0004},
+{0x00FB29, 0x0040},
+{0x00FB2A, 0x0004},
+{0x00FB37, 0x0001},
+{0x00FB38, 0x0004},
+{0x00FB3D, 0x0001},
+{0x00FB3E, 0x0004},
+{0x00FB3F, 0x0001},
+{0x00FB40, 0x0004},
+{0x00FB42, 0x0001},
+{0x00FB43, 0x0004},
+{0x00FB45, 0x0001},
+{0x00FB46, 0x0004},
+{0x00FBB2, 0x0040},
+{0x00FBC3, 0x0001},
+{0x00FBD3, 0x0004},
+{0x00FD3E, 0x0020},
+{0x00FD40, 0x0040},
+{0x00FD50, 0x0004},
+{0x00FD90, 0x0001},
+{0x00FD92, 0x0004},
+{0x00FDC8, 0x0001},
+{0x00FDCF, 0x0040},
+{0x00FDD0, 0x0001},
+{0x00FDF0, 0x0004},
+{0x00FDFC, 0x0040},
+{0x00FE00, 0x0010},
+{0x00FE10, 0x0020},
+{0x00FE1A, 0x0001},
+{0x00FE20, 0x0010},
+{0x00FE30, 0x0020},
+{0x00FE53, 0x0001},
+{0x00FE54, 0x0020},
+{0x00FE62, 0x0040},
+{0x00FE63, 0x0020},
+{0x00FE64, 0x0040},
+{0x00FE67, 0x0001},
+{0x00FE68, 0x0020},
+{0x00FE69, 0x0040},
+{0x00FE6A, 0x0020},
+{0x00FE6C, 0x0001},
+{0x00FE70, 0x0004},
+{0x00FE75, 0x0001},
+{0x00FE76, 0x0004},
+{0x00FEFD, 0x0001},
+{0x00FEFF, 0x0080},
+{0x00FF00, 0x0001},
+{0x00FF01, 0x0020},
+{0x00FF04, 0x0040},
+{0x00FF05, 0x0020},
+{0x00FF0B, 0x0040},
+{0x00FF0C, 0x0020},
+{0x00FF10, 0x0002},
+{0x00FF1A, 0x0020},
+{0x00FF1C, 0x0040},
+{0x00FF1F, 0x0020},
+{0x00FF21, 0x0004},
+{0x00FF3B, 0x0020},
+{0x00FF3E, 0x0040},
+{0x00FF3F, 0x0020},
+{0x00FF40, 0x0040},
+{0x00FF41, 0x0004},
+{0x00FF5B, 0x0020},
+{0x00FF5C, 0x0040},
+{0x00FF5D, 0x0020},
+{0x00FF5E, 0x0040},
+{0x00FF5F, 0x0020},
+{0x00FF66, 0x0004},
+{0x00FFBF, 0x0001},
+{0x00FFC2, 0x0004},
+{0x00FFC8, 0x0001},
+{0x00FFCA, 0x0004},
+{0x00FFD0, 0x0001},
+{0x00FFD2, 0x0004},
+{0x00FFD8, 0x0001},
+{0x00FFDA, 0x0004},
+{0x00FFDD, 0x0001},
+{0x00FFE0, 0x0040},
+{0x00FFE7, 0x0001},
+{0x00FFE8, 0x0040},
+{0x00FFEF, 0x0001},
+{0x00FFF9, 0x0080},
+{0x00FFFC, 0x0040},
+{0x00FFFE, 0x0001},
+{0x010000, 0x0004},
+{0x01000C, 0x0001},
+{0x01000D, 0x0004},
+{0x010027, 0x0001},
+{0x010028, 0x0004},
+{0x01003B, 0x0001},
+{0x01003C, 0x0004},
+{0x01003E, 0x0001},
+{0x01003F, 0x0004},
+{0x01004E, 0x0001},
+{0x010050, 0x0004},
+{0x01005E, 0x0001},
+{0x010080, 0x0004},
+{0x0100FB, 0x0001},
+{0x010100, 0x0020},
+{0x010103, 0x0001},
+{0x010107, 0x0002},
+{0x010134, 0x0001},
+{0x010137, 0x0040},
+{0x010140, 0x0002},
+{0x010179, 0x0040},
+{0x01018A, 0x0002},
+{0x01018C, 0x0040},
+{0x01018F, 0x0001},
+{0x010190, 0x0040},
+{0x01019D, 0x0001},
+{0x0101A0, 0x0040},
+{0x0101A1, 0x0001},
+{0x0101D0, 0x0040},
+{0x0101FD, 0x0010},
+{0x0101FE, 0x0001},
+{0x010280, 0x0004},
+{0x01029D, 0x0001},
+{0x0102A0, 0x0004},
+{0x0102D1, 0x0001},
+{0x0102E0, 0x0010},
+{0x0102E1, 0x0002},
+{0x0102FC, 0x0001},
+{0x010300, 0x0004},
+{0x010320, 0x0002},
+{0x010324, 0x0001},
+{0x01032D, 0x0004},
+{0x010341, 0x0002},
+{0x010342, 0x0004},
+{0x01034A, 0x0002},
+{0x01034B, 0x0001},
+{0x010350, 0x0004},
+{0x010376, 0x0010},
+{0x01037B, 0x0001},
+{0x010380, 0x0004},
+{0x01039E, 0x0001},
+{0x01039F, 0x0020},
+{0x0103A0, 0x0004},
+{0x0103C4, 0x0001},
+{0x0103C8, 0x0004},
+{0x0103D0, 0x0020},
+{0x0103D1, 0x0002},
+{0x0103D6, 0x0001},
+{0x010400, 0x0004},
+{0x01049E, 0x0001},
+{0x0104A0, 0x0002},
+{0x0104AA, 0x0001},
+{0x0104B0, 0x0004},
+{0x0104D4, 0x0001},
+{0x0104D8, 0x0004},
+{0x0104FC, 0x0001},
+{0x010500, 0x0004},
+{0x010528, 0x0001},
+{0x010530, 0x0004},
+{0x010564, 0x0001},
+{0x01056F, 0x0020},
+{0x010570, 0x0004},
+{0x01057B, 0x0001},
+{0x01057C, 0x0004},
+{0x01058B, 0x0001},
+{0x01058C, 0x0004},
+{0x010593, 0x0001},
+{0x010594, 0x0004},
+{0x010596, 0x0001},
+{0x010597, 0x0004},
+{0x0105A2, 0x0001},
+{0x0105A3, 0x0004},
+{0x0105B2, 0x0001},
+{0x0105B3, 0x0004},
+{0x0105BA, 0x0001},
+{0x0105BB, 0x0004},
+{0x0105BD, 0x0001},
+{0x010600, 0x0004},
+{0x010737, 0x0001},
+{0x010740, 0x0004},
+{0x010756, 0x0001},
+{0x010760, 0x0004},
+{0x010768, 0x0001},
+{0x010780, 0x0004},
+{0x010786, 0x0001},
+{0x010787, 0x0004},
+{0x0107B1, 0x0001},
+{0x0107B2, 0x0004},
+{0x0107BB, 0x0001},
+{0x010800, 0x0004},
+{0x010806, 0x0001},
+{0x010808, 0x0004},
+{0x010809, 0x0001},
+{0x01080A, 0x0004},
+{0x010836, 0x0001},
+{0x010837, 0x0004},
+{0x010839, 0x0001},
+{0x01083C, 0x0004},
+{0x01083D, 0x0001},
+{0x01083F, 0x0004},
+{0x010856, 0x0001},
+{0x010857, 0x0020},
+{0x010858, 0x0002},
+{0x010860, 0x0004},
+{0x010877, 0x0040},
+{0x010879, 0x0002},
+{0x010880, 0x0004},
+{0x01089F, 0x0001},
+{0x0108A7, 0x0002},
+{0x0108B0, 0x0001},
+{0x0108E0, 0x0004},
+{0x0108F3, 0x0001},
+{0x0108F4, 0x0004},
+{0x0108F6, 0x0001},
+{0x0108FB, 0x0002},
+{0x010900, 0x0004},
+{0x010916, 0x0002},
+{0x01091C, 0x0001},
+{0x01091F, 0x0020},
+{0x010920, 0x0004},
+{0x01093A, 0x0001},
+{0x01093F, 0x0020},
+{0x010940, 0x0001},
+{0x010980, 0x0004},
+{0x0109B8, 0x0001},
+{0x0109BC, 0x0002},
+{0x0109BE, 0x0004},
+{0x0109C0, 0x0002},
+{0x0109D0, 0x0001},
+{0x0109D2, 0x0002},
+{0x010A00, 0x0004},
+{0x010A01, 0x0010},
+{0x010A04, 0x0001},
+{0x010A05, 0x0010},
+{0x010A07, 0x0001},
+{0x010A0C, 0x0010},
+{0x010A10, 0x0004},
+{0x010A14, 0x0001},
+{0x010A15, 0x0004},
+{0x010A18, 0x0001},
+{0x010A19, 0x0004},
+{0x010A36, 0x0001},
+{0x010A38, 0x0010},
+{0x010A3B, 0x0001},
+{0x010A3F, 0x0010},
+{0x010A40, 0x0002},
+{0x010A49, 0x0001},
+{0x010A50, 0x0020},
+{0x010A59, 0x0001},
+{0x010A60, 0x0004},
+{0x010A7D, 0x0002},
+{0x010A7F, 0x0020},
+{0x010A80, 0x0004},
+{0x010A9D, 0x0002},
+{0x010AA0, 0x0001},
+{0x010AC0, 0x0004},
+{0x010AC8, 0x0040},
+{0x010AC9, 0x0004},
+{0x010AE5, 0x0010},
+{0x010AE7, 0x0001},
+{0x010AEB, 0x0002},
+{0x010AF0, 0x0020},
+{0x010AF7, 0x0001},
+{0x010B00, 0x0004},
+{0x010B36, 0x0001},
+{0x010B39, 0x0020},
+{0x010B40, 0x0004},
+{0x010B56, 0x0001},
+{0x010B58, 0x0002},
+{0x010B60, 0x0004},
+{0x010B73, 0x0001},
+{0x010B78, 0x0002},
+{0x010B80, 0x0004},
+{0x010B92, 0x0001},
+{0x010B99, 0x0020},
+{0x010B9D, 0x0001},
+{0x010BA9, 0x0002},
+{0x010BB0, 0x0001},
+{0x010C00, 0x0004},
+{0x010C49, 0x0001},
+{0x010C80, 0x0004},
+{0x010CB3, 0x0001},
+{0x010CC0, 0x0004},
+{0x010CF3, 0x0001},
+{0x010CFA, 0x0002},
+{0x010D00, 0x0004},
+{0x010D24, 0x0010},
+{0x010D28, 0x0001},
+{0x010D30, 0x0002},
+{0x010D3A, 0x0001},
+{0x010E60, 0x0002},
+{0x010E7F, 0x0001},
+{0x010E80, 0x0004},
+{0x010EAA, 0x0001},
+{0x010EAB, 0x0010},
+{0x010EAD, 0x0020},
+{0x010EAE, 0x0001},
+{0x010EB0, 0x0004},
+{0x010EB2, 0x0001},
+{0x010EFD, 0x0010},
+{0x010F00, 0x0004},
+{0x010F1D, 0x0002},
+{0x010F27, 0x0004},
+{0x010F28, 0x0001},
+{0x010F30, 0x0004},
+{0x010F46, 0x0010},
+{0x010F51, 0x0002},
+{0x010F55, 0x0020},
+{0x010F5A, 0x0001},
+{0x010F70, 0x0004},
+{0x010F82, 0x0010},
+{0x010F86, 0x0020},
+{0x010F8A, 0x0001},
+{0x010FB0, 0x0004},
+{0x010FC5, 0x0002},
+{0x010FCC, 0x0001},
+{0x010FE0, 0x0004},
+{0x010FF7, 0x0001},
+{0x011000, 0x0010},
+{0x011003, 0x0004},
+{0x011038, 0x0010},
+{0x011047, 0x0020},
+{0x01104E, 0x0001},
+{0x011052, 0x0002},
+{0x011070, 0x0010},
+{0x011071, 0x0004},
+{0x011073, 0x0010},
+{0x011075, 0x0004},
+{0x011076, 0x0001},
+{0x01107F, 0x0010},
+{0x011083, 0x0004},
+{0x0110B0, 0x0010},
+{0x0110BB, 0x0020},
+{0x0110BD, 0x0080},
+{0x0110BE, 0x0020},
+{0x0110C2, 0x0010},
+{0x0110C3, 0x0001},
+{0x0110CD, 0x0080},
+{0x0110CE, 0x0001},
+{0x0110D0, 0x0004},
+{0x0110E9, 0x0001},
+{0x0110F0, 0x0002},
+{0x0110FA, 0x0001},
+{0x011100, 0x0010},
+{0x011103, 0x0004},
+{0x011127, 0x0010},
+{0x011135, 0x0001},
+{0x011136, 0x0002},
+{0x011140, 0x0020},
+{0x011144, 0x0004},
+{0x011145, 0x0010},
+{0x011147, 0x0004},
+{0x011148, 0x0001},
+{0x011150, 0x0004},
+{0x011173, 0x0010},
+{0x011174, 0x0020},
+{0x011176, 0x0004},
+{0x011177, 0x0001},
+{0x011180, 0x0010},
+{0x011183, 0x0004},
+{0x0111B3, 0x0010},
+{0x0111C1, 0x0004},
+{0x0111C5, 0x0020},
+{0x0111C9, 0x0010},
+{0x0111CD, 0x0020},
+{0x0111CE, 0x0010},
+{0x0111D0, 0x0002},
+{0x0111DA, 0x0004},
+{0x0111DB, 0x0020},
+{0x0111DC, 0x0004},
+{0x0111DD, 0x0020},
+{0x0111E0, 0x0001},
+{0x0111E1, 0x0002},
+{0x0111F5, 0x0001},
+{0x011200, 0x0004},
+{0x011212, 0x0001},
+{0x011213, 0x0004},
+{0x01122C, 0x0010},
+{0x011238, 0x0020},
+{0x01123E, 0x0010},
+{0x01123F, 0x0004},
+{0x011241, 0x0010},
+{0x011242, 0x0001},
+{0x011280, 0x0004},
+{0x011287, 0x0001},
+{0x011288, 0x0004},
+{0x011289, 0x0001},
+{0x01128A, 0x0004},
+{0x01128E, 0x0001},
+{0x01128F, 0x0004},
+{0x01129E, 0x0001},
+{0x01129F, 0x0004},
+{0x0112A9, 0x0020},
+{0x0112AA, 0x0001},
+{0x0112B0, 0x0004},
+{0x0112DF, 0x0010},
+{0x0112EB, 0x0001},
+{0x0112F0, 0x0002},
+{0x0112FA, 0x0001},
+{0x011300, 0x0010},
+{0x011304, 0x0001},
+{0x011305, 0x0004},
+{0x01130D, 0x0001},
+{0x01130F, 0x0004},
+{0x011311, 0x0001},
+{0x011313, 0x0004},
+{0x011329, 0x0001},
+{0x01132A, 0x0004},
+{0x011331, 0x0001},
+{0x011332, 0x0004},
+{0x011334, 0x0001},
+{0x011335, 0x0004},
+{0x01133A, 0x0001},
+{0x01133B, 0x0010},
+{0x01133D, 0x0004},
+{0x01133E, 0x0010},
+{0x011345, 0x0001},
+{0x011347, 0x0010},
+{0x011349, 0x0001},
+{0x01134B, 0x0010},
+{0x01134E, 0x0001},
+{0x011350, 0x0004},
+{0x011351, 0x0001},
+{0x011357, 0x0010},
+{0x011358, 0x0001},
+{0x01135D, 0x0004},
+{0x011362, 0x0010},
+{0x011364, 0x0001},
+{0x011366, 0x0010},
+{0x01136D, 0x0001},
+{0x011370, 0x0010},
+{0x011375, 0x0001},
+{0x011400, 0x0004},
+{0x011435, 0x0010},
+{0x011447, 0x0004},
+{0x01144B, 0x0020},
+{0x011450, 0x0002},
+{0x01145A, 0x0020},
+{0x01145C, 0x0001},
+{0x01145D, 0x0020},
+{0x01145E, 0x0010},
+{0x01145F, 0x0004},
+{0x011462, 0x0001},
+{0x011480, 0x0004},
+{0x0114B0, 0x0010},
+{0x0114C4, 0x0004},
+{0x0114C6, 0x0020},
+{0x0114C7, 0x0004},
+{0x0114C8, 0x0001},
+{0x0114D0, 0x0002},
+{0x0114DA, 0x0001},
+{0x011580, 0x0004},
+{0x0115AF, 0x0010},
+{0x0115B6, 0x0001},
+{0x0115B8, 0x0010},
+{0x0115C1, 0x0020},
+{0x0115D8, 0x0004},
+{0x0115DC, 0x0010},
+{0x0115DE, 0x0001},
+{0x011600, 0x0004},
+{0x011630, 0x0010},
+{0x011641, 0x0020},
+{0x011644, 0x0004},
+{0x011645, 0x0001},
+{0x011650, 0x0002},
+{0x01165A, 0x0001},
+{0x011660, 0x0020},
+{0x01166D, 0x0001},
+{0x011680, 0x0004},
+{0x0116AB, 0x0010},
+{0x0116B8, 0x0004},
+{0x0116B9, 0x0020},
+{0x0116BA, 0x0001},
+{0x0116C0, 0x0002},
+{0x0116CA, 0x0001},
+{0x011700, 0x0004},
+{0x01171B, 0x0001},
+{0x01171D, 0x0010},
+{0x01172C, 0x0001},
+{0x011730, 0x0002},
+{0x01173C, 0x0020},
+{0x01173F, 0x0040},
+{0x011740, 0x0004},
+{0x011747, 0x0001},
+{0x011800, 0x0004},
+{0x01182C, 0x0010},
+{0x01183B, 0x0020},
+{0x01183C, 0x0001},
+{0x0118A0, 0x0004},
+{0x0118E0, 0x0002},
+{0x0118F3, 0x0001},
+{0x0118FF, 0x0004},
+{0x011907, 0x0001},
+{0x011909, 0x0004},
+{0x01190A, 0x0001},
+{0x01190C, 0x0004},
+{0x011914, 0x0001},
+{0x011915, 0x0004},
+{0x011917, 0x0001},
+{0x011918, 0x0004},
+{0x011930, 0x0010},
+{0x011936, 0x0001},
+{0x011937, 0x0010},
+{0x011939, 0x0001},
+{0x01193B, 0x0010},
+{0x01193F, 0x0004},
+{0x011940, 0x0010},
+{0x011941, 0x0004},
+{0x011942, 0x0010},
+{0x011944, 0x0020},
+{0x011947, 0x0001},
+{0x011950, 0x0002},
+{0x01195A, 0x0001},
+{0x0119A0, 0x0004},
+{0x0119A8, 0x0001},
+{0x0119AA, 0x0004},
+{0x0119D1, 0x0010},
+{0x0119D8, 0x0001},
+{0x0119DA, 0x0010},
+{0x0119E1, 0x0004},
+{0x0119E2, 0x0020},
+{0x0119E3, 0x0004},
+{0x0119E4, 0x0010},
+{0x0119E5, 0x0001},
+{0x011A00, 0x0004},
+{0x011A01, 0x0010},
+{0x011A0B, 0x0004},
+{0x011A33, 0x0010},
+{0x011A3A, 0x0004},
+{0x011A3B, 0x0010},
+{0x011A3F, 0x0020},
+{0x011A47, 0x0010},
+{0x011A48, 0x0001},
+{0x011A50, 0x0004},
+{0x011A51, 0x0010},
+{0x011A5C, 0x0004},
+{0x011A8A, 0x0010},
+{0x011A9A, 0x0020},
+{0x011A9D, 0x0004},
+{0x011A9E, 0x0020},
+{0x011AA3, 0x0001},
+{0x011AB0, 0x0004},
+{0x011AF9, 0x0001},
+{0x011B00, 0x0020},
+{0x011B0A, 0x0001},
+{0x011C00, 0x0004},
+{0x011C09, 0x0001},
+{0x011C0A, 0x0004},
+{0x011C2F, 0x0010},
+{0x011C37, 0x0001},
+{0x011C38, 0x0010},
+{0x011C40, 0x0004},
+{0x011C41, 0x0020},
+{0x011C46, 0x0001},
+{0x011C50, 0x0002},
+{0x011C6D, 0x0001},
+{0x011C70, 0x0020},
+{0x011C72, 0x0004},
+{0x011C90, 0x0001},
+{0x011C92, 0x0010},
+{0x011CA8, 0x0001},
+{0x011CA9, 0x0010},
+{0x011CB7, 0x0001},
+{0x011D00, 0x0004},
+{0x011D07, 0x0001},
+{0x011D08, 0x0004},
+{0x011D0A, 0x0001},
+{0x011D0B, 0x0004},
+{0x011D31, 0x0010},
+{0x011D37, 0x0001},
+{0x011D3A, 0x0010},
+{0x011D3B, 0x0001},
+{0x011D3C, 0x0010},
+{0x011D3E, 0x0001},
+{0x011D3F, 0x0010},
+{0x011D46, 0x0004},
+{0x011D47, 0x0010},
+{0x011D48, 0x0001},
+{0x011D50, 0x0002},
+{0x011D5A, 0x0001},
+{0x011D60, 0x0004},
+{0x011D66, 0x0001},
+{0x011D67, 0x0004},
+{0x011D69, 0x0001},
+{0x011D6A, 0x0004},
+{0x011D8A, 0x0010},
+{0x011D8F, 0x0001},
+{0x011D90, 0x0010},
+{0x011D92, 0x0001},
+{0x011D93, 0x0010},
+{0x011D98, 0x0004},
+{0x011D99, 0x0001},
+{0x011DA0, 0x0002},
+{0x011DAA, 0x0001},
+{0x011EE0, 0x0004},
+{0x011EF3, 0x0010},
+{0x011EF7, 0x0020},
+{0x011EF9, 0x0001},
+{0x011F00, 0x0010},
+{0x011F02, 0x0004},
+{0x011F03, 0x0010},
+{0x011F04, 0x0004},
+{0x011F11, 0x0001},
+{0x011F12, 0x0004},
+{0x011F34, 0x0010},
+{0x011F3B, 0x0001},
+{0x011F3E, 0x0010},
+{0x011F43, 0x0020},
+{0x011F50, 0x0002},
+{0x011F5A, 0x0001},
+{0x011FB0, 0x0004},
+{0x011FB1, 0x0001},
+{0x011FC0, 0x0002},
+{0x011FD5, 0x0040},
+{0x011FF2, 0x0001},
+{0x011FFF, 0x0020},
+{0x012000, 0x0004},
+{0x01239A, 0x0001},
+{0x012400, 0x0002},
+{0x01246F, 0x0001},
+{0x012470, 0x0020},
+{0x012475, 0x0001},
+{0x012480, 0x0004},
+{0x012544, 0x0001},
+{0x012F90, 0x0004},
+{0x012FF1, 0x0020},
+{0x012FF3, 0x0001},
+{0x013000, 0x0004},
+{0x013430, 0x0080},
+{0x013440, 0x0010},
+{0x013441, 0x0004},
+{0x013447, 0x0010},
+{0x013456, 0x0001},
+{0x014400, 0x0004},
+{0x014647, 0x0001},
+{0x016800, 0x0004},
+{0x016A39, 0x0001},
+{0x016A40, 0x0004},
+{0x016A5F, 0x0001},
+{0x016A60, 0x0002},
+{0x016A6A, 0x0001},
+{0x016A6E, 0x0020},
+{0x016A70, 0x0004},
+{0x016ABF, 0x0001},
+{0x016AC0, 0x0002},
+{0x016ACA, 0x0001},
+{0x016AD0, 0x0004},
+{0x016AEE, 0x0001},
+{0x016AF0, 0x0010},
+{0x016AF5, 0x0020},
+{0x016AF6, 0x0001},
+{0x016B00, 0x0004},
+{0x016B30, 0x0010},
+{0x016B37, 0x0020},
+{0x016B3C, 0x0040},
+{0x016B40, 0x0004},
+{0x016B44, 0x0020},
+{0x016B45, 0x0040},
+{0x016B46, 0x0001},
+{0x016B50, 0x0002},
+{0x016B5A, 0x0001},
+{0x016B5B, 0x0002},
+{0x016B62, 0x0001},
+{0x016B63, 0x0004},
+{0x016B78, 0x0001},
+{0x016B7D, 0x0004},
+{0x016B90, 0x0001},
+{0x016E40, 0x0004},
+{0x016E80, 0x0002},
+{0x016E97, 0x0020},
+{0x016E9B, 0x0001},
+{0x016F00, 0x0004},
+{0x016F4B, 0x0001},
+{0x016F4F, 0x0010},
+{0x016F50, 0x0004},
+{0x016F51, 0x0010},
+{0x016F88, 0x0001},
+{0x016F8F, 0x0010},
+{0x016F93, 0x0004},
+{0x016FA0, 0x0001},
+{0x016FE0, 0x0004},
+{0x016FE2, 0x0020},
+{0x016FE3, 0x0004},
+{0x016FE4, 0x0010},
+{0x016FE5, 0x0001},
+{0x016FF0, 0x0010},
+{0x016FF2, 0x0001},
+{0x017000, 0x0004},
+{0x0187F8, 0x0001},
+{0x018800, 0x0004},
+{0x018CD6, 0x0001},
+{0x018D00, 0x0004},
+{0x018D09, 0x0001},
+{0x01AFF0, 0x0004},
+{0x01AFF4, 0x0001},
+{0x01AFF5, 0x0004},
+{0x01AFFC, 0x0001},
+{0x01AFFD, 0x0004},
+{0x01AFFF, 0x0001},
+{0x01B000, 0x0004},
+{0x01B123, 0x0001},
+{0x01B132, 0x0004},
+{0x01B133, 0x0001},
+{0x01B150, 0x0004},
+{0x01B153, 0x0001},
+{0x01B155, 0x0004},
+{0x01B156, 0x0001},
+{0x01B164, 0x0004},
+{0x01B168, 0x0001},
+{0x01B170, 0x0004},
+{0x01B2FC, 0x0001},
+{0x01BC00, 0x0004},
+{0x01BC6B, 0x0001},
+{0x01BC70, 0x0004},
+{0x01BC7D, 0x0001},
+{0x01BC80, 0x0004},
+{0x01BC89, 0x0001},
+{0x01BC90, 0x0004},
+{0x01BC9A, 0x0001},
+{0x01BC9C, 0x0040},
+{0x01BC9D, 0x0010},
+{0x01BC9F, 0x0020},
+{0x01BCA0, 0x0080},
+{0x01BCA4, 0x0001},
+{0x01CF00, 0x0010},
+{0x01CF2E, 0x0001},
+{0x01CF30, 0x0010},
+{0x01CF47, 0x0001},
+{0x01CF50, 0x0040},
+{0x01CFC4, 0x0001},
+{0x01D000, 0x0040},
+{0x01D0F6, 0x0001},
+{0x01D100, 0x0040},
+{0x01D127, 0x0001},
+{0x01D129, 0x0040},
+{0x01D165, 0x0010},
+{0x01D16A, 0x0040},
+{0x01D16D, 0x0010},
+{0x01D173, 0x0080},
+{0x01D17B, 0x0010},
+{0x01D183, 0x0040},
+{0x01D185, 0x0010},
+{0x01D18C, 0x0040},
+{0x01D1AA, 0x0010},
+{0x01D1AE, 0x0040},
+{0x01D1EB, 0x0001},
+{0x01D200, 0x0040},
+{0x01D242, 0x0010},
+{0x01D245, 0x0040},
+{0x01D246, 0x0001},
+{0x01D2C0, 0x0002},
+{0x01D2D4, 0x0001},
+{0x01D2E0, 0x0002},
+{0x01D2F4, 0x0001},
+{0x01D300, 0x0040},
+{0x01D357, 0x0001},
+{0x01D360, 0x0002},
+{0x01D379, 0x0001},
+{0x01D400, 0x0004},
+{0x01D455, 0x0001},
+{0x01D456, 0x0004},
+{0x01D49D, 0x0001},
+{0x01D49E, 0x0004},
+{0x01D4A0, 0x0001},
+{0x01D4A2, 0x0004},
+{0x01D4A3, 0x0001},
+{0x01D4A5, 0x0004},
+{0x01D4A7, 0x0001},
+{0x01D4A9, 0x0004},
+{0x01D4AD, 0x0001},
+{0x01D4AE, 0x0004},
+{0x01D4BA, 0x0001},
+{0x01D4BB, 0x0004},
+{0x01D4BC, 0x0001},
+{0x01D4BD, 0x0004},
+{0x01D4C4, 0x0001},
+{0x01D4C5, 0x0004},
+{0x01D506, 0x0001},
+{0x01D507, 0x0004},
+{0x01D50B, 0x0001},
+{0x01D50D, 0x0004},
+{0x01D515, 0x0001},
+{0x01D516, 0x0004},
+{0x01D51D, 0x0001},
+{0x01D51E, 0x0004},
+{0x01D53A, 0x0001},
+{0x01D53B, 0x0004},
+{0x01D53F, 0x0001},
+{0x01D540, 0x0004},
+{0x01D545, 0x0001},
+{0x01D546, 0x0004},
+{0x01D547, 0x0001},
+{0x01D54A, 0x0004},
+{0x01D551, 0x0001},
+{0x01D552, 0x0004},
+{0x01D6A6, 0x0001},
+{0x01D6A8, 0x0004},
+{0x01D6C1, 0x0040},
+{0x01D6C2, 0x0004},
+{0x01D6DB, 0x0040},
+{0x01D6DC, 0x0004},
+{0x01D6FB, 0x0040},
+{0x01D6FC, 0x0004},
+{0x01D715, 0x0040},
+{0x01D716, 0x0004},
+{0x01D735, 0x0040},
+{0x01D736, 0x0004},
+{0x01D74F, 0x0040},
+{0x01D750, 0x0004},
+{0x01D76F, 0x0040},
+{0x01D770, 0x0004},
+{0x01D789, 0x0040},
+{0x01D78A, 0x0004},
+{0x01D7A9, 0x0040},
+{0x01D7AA, 0x0004},
+{0x01D7C3, 0x0040},
+{0x01D7C4, 0x0004},
+{0x01D7CC, 0x0001},
+{0x01D7CE, 0x0002},
+{0x01D800, 0x0040},
+{0x01DA00, 0x0010},
+{0x01DA37, 0x0040},
+{0x01DA3B, 0x0010},
+{0x01DA6D, 0x0040},
+{0x01DA75, 0x0010},
+{0x01DA76, 0x0040},
+{0x01DA84, 0x0010},
+{0x01DA85, 0x0040},
+{0x01DA87, 0x0020},
+{0x01DA8C, 0x0001},
+{0x01DA9B, 0x0010},
+{0x01DAA0, 0x0001},
+{0x01DAA1, 0x0010},
+{0x01DAB0, 0x0001},
+{0x01DF00, 0x0004},
+{0x01DF1F, 0x0001},
+{0x01DF25, 0x0004},
+{0x01DF2B, 0x0001},
+{0x01E000, 0x0010},
+{0x01E007, 0x0001},
+{0x01E008, 0x0010},
+{0x01E019, 0x0001},
+{0x01E01B, 0x0010},
+{0x01E022, 0x0001},
+{0x01E023, 0x0010},
+{0x01E025, 0x0001},
+{0x01E026, 0x0010},
+{0x01E02B, 0x0001},
+{0x01E030, 0x0004},
+{0x01E06E, 0x0001},
+{0x01E08F, 0x0010},
+{0x01E090, 0x0001},
+{0x01E100, 0x0004},
+{0x01E12D, 0x0001},
+{0x01E130, 0x0010},
+{0x01E137, 0x0004},
+{0x01E13E, 0x0001},
+{0x01E140, 0x0002},
+{0x01E14A, 0x0001},
+{0x01E14E, 0x0004},
+{0x01E14F, 0x0040},
+{0x01E150, 0x0001},
+{0x01E290, 0x0004},
+{0x01E2AE, 0x0010},
+{0x01E2AF, 0x0001},
+{0x01E2C0, 0x0004},
+{0x01E2EC, 0x0010},
+{0x01E2F0, 0x0002},
+{0x01E2FA, 0x0001},
+{0x01E2FF, 0x0040},
+{0x01E300, 0x0001},
+{0x01E4D0, 0x0004},
+{0x01E4EC, 0x0010},
+{0x01E4F0, 0x0002},
+{0x01E4FA, 0x0001},
+{0x01E7E0, 0x0004},
+{0x01E7E7, 0x0001},
+{0x01E7E8, 0x0004},
+{0x01E7EC, 0x0001},
+{0x01E7ED, 0x0004},
+{0x01E7EF, 0x0001},
+{0x01E7F0, 0x0004},
+{0x01E7FF, 0x0001},
+{0x01E800, 0x0004},
+{0x01E8C5, 0x0001},
+{0x01E8C7, 0x0002},
+{0x01E8D0, 0x0010},
+{0x01E8D7, 0x0001},
+{0x01E900, 0x0004},
+{0x01E944, 0x0010},
+{0x01E94B, 0x0004},
+{0x01E94C, 0x0001},
+{0x01E950, 0x0002},
+{0x01E95A, 0x0001},
+{0x01E95E, 0x0020},
+{0x01E960, 0x0001},
+{0x01EC71, 0x0002},
+{0x01ECAC, 0x0040},
+{0x01ECAD, 0x0002},
+{0x01ECB0, 0x0040},
+{0x01ECB1, 0x0002},
+{0x01ECB5, 0x0001},
+{0x01ED01, 0x0002},
+{0x01ED2E, 0x0040},
+{0x01ED2F, 0x0002},
+{0x01ED3E, 0x0001},
+{0x01EE00, 0x0004},
+{0x01EE04, 0x0001},
+{0x01EE05, 0x0004},
+{0x01EE20, 0x0001},
+{0x01EE21, 0x0004},
+{0x01EE23, 0x0001},
+{0x01EE24, 0x0004},
+{0x01EE25, 0x0001},
+{0x01EE27, 0x0004},
+{0x01EE28, 0x0001},
+{0x01EE29, 0x0004},
+{0x01EE33, 0x0001},
+{0x01EE34, 0x0004},
+{0x01EE38, 0x0001},
+{0x01EE39, 0x0004},
+{0x01EE3A, 0x0001},
+{0x01EE3B, 0x0004},
+{0x01EE3C, 0x0001},
+{0x01EE42, 0x0004},
+{0x01EE43, 0x0001},
+{0x01EE47, 0x0004},
+{0x01EE48, 0x0001},
+{0x01EE49, 0x0004},
+{0x01EE4A, 0x0001},
+{0x01EE4B, 0x0004},
+{0x01EE4C, 0x0001},
+{0x01EE4D, 0x0004},
+{0x01EE50, 0x0001},
+{0x01EE51, 0x0004},
+{0x01EE53, 0x0001},
+{0x01EE54, 0x0004},
+{0x01EE55, 0x0001},
+{0x01EE57, 0x0004},
+{0x01EE58, 0x0001},
+{0x01EE59, 0x0004},
+{0x01EE5A, 0x0001},
+{0x01EE5B, 0x0004},
+{0x01EE5C, 0x0001},
+{0x01EE5D, 0x0004},
+{0x01EE5E, 0x0001},
+{0x01EE5F, 0x0004},
+{0x01EE60, 0x0001},
+{0x01EE61, 0x0004},
+{0x01EE63, 0x0001},
+{0x01EE64, 0x0004},
+{0x01EE65, 0x0001},
+{0x01EE67, 0x0004},
+{0x01EE6B, 0x0001},
+{0x01EE6C, 0x0004},
+{0x01EE73, 0x0001},
+{0x01EE74, 0x0004},
+{0x01EE78, 0x0001},
+{0x01EE79, 0x0004},
+{0x01EE7D, 0x0001},
+{0x01EE7E, 0x0004},
+{0x01EE7F, 0x0001},
+{0x01EE80, 0x0004},
+{0x01EE8A, 0x0001},
+{0x01EE8B, 0x0004},
+{0x01EE9C, 0x0001},
+{0x01EEA1, 0x0004},
+{0x01EEA4, 0x0001},
+{0x01EEA5, 0x0004},
+{0x01EEAA, 0x0001},
+{0x01EEAB, 0x0004},
+{0x01EEBC, 0x0001},
+{0x01EEF0, 0x0040},
+{0x01EEF2, 0x0001},
+{0x01F000, 0x0040},
+{0x01F02C, 0x0001},
+{0x01F030, 0x0040},
+{0x01F094, 0x0001},
+{0x01F0A0, 0x0040},
+{0x01F0AF, 0x0001},
+{0x01F0B1, 0x0040},
+{0x01F0C0, 0x0001},
+{0x01F0C1, 0x0040},
+{0x01F0D0, 0x0001},
+{0x01F0D1, 0x0040},
+{0x01F0F6, 0x0001},
+{0x01F100, 0x0002},
+{0x01F10D, 0x0040},
+{0x01F1AE, 0x0001},
+{0x01F1E6, 0x0040},
+{0x01F203, 0x0001},
+{0x01F210, 0x0040},
+{0x01F23C, 0x0001},
+{0x01F240, 0x0040},
+{0x01F249, 0x0001},
+{0x01F250, 0x0040},
+{0x01F252, 0x0001},
+{0x01F260, 0x0040},
+{0x01F266, 0x0001},
+{0x01F300, 0x0040},
+{0x01F6D8, 0x0001},
+{0x01F6DC, 0x0040},
+{0x01F6ED, 0x0001},
+{0x01F6F0, 0x0040},
+{0x01F6FD, 0x0001},
+{0x01F700, 0x0040},
+{0x01F777, 0x0001},
+{0x01F77B, 0x0040},
+{0x01F7DA, 0x0001},
+{0x01F7E0, 0x0040},
+{0x01F7EC, 0x0001},
+{0x01F7F0, 0x0040},
+{0x01F7F1, 0x0001},
+{0x01F800, 0x0040},
+{0x01F80C, 0x0001},
+{0x01F810, 0x0040},
+{0x01F848, 0x0001},
+{0x01F850, 0x0040},
+{0x01F85A, 0x0001},
+{0x01F860, 0x0040},
+{0x01F888, 0x0001},
+{0x01F890, 0x0040},
+{0x01F8AE, 0x0001},
+{0x01F8B0, 0x0040},
+{0x01F8B2, 0x0001},
+{0x01F900, 0x0040},
+{0x01FA54, 0x0001},
+{0x01FA60, 0x0040},
+{0x01FA6E, 0x0001},
+{0x01FA70, 0x0040},
+{0x01FA7D, 0x0001},
+{0x01FA80, 0x0040},
+{0x01FA89, 0x0001},
+{0x01FA90, 0x0040},
+{0x01FABE, 0x0001},
+{0x01FABF, 0x0040},
+{0x01FAC6, 0x0001},
+{0x01FACE, 0x0040},
+{0x01FADC, 0x0001},
+{0x01FAE0, 0x0040},
+{0x01FAE9, 0x0001},
+{0x01FAF0, 0x0040},
+{0x01FAF9, 0x0001},
+{0x01FB00, 0x0040},
+{0x01FB93, 0x0001},
+{0x01FB94, 0x0040},
+{0x01FBCB, 0x0001},
+{0x01FBF0, 0x0002},
+{0x01FBFA, 0x0001},
+{0x020000, 0x0004},
+{0x02A6E0, 0x0001},
+{0x02A700, 0x0004},
+{0x02B73A, 0x0001},
+{0x02B740, 0x0004},
+{0x02B81E, 0x0001},
+{0x02B820, 0x0004},
+{0x02CEA2, 0x0001},
+{0x02CEB0, 0x0004},
+{0x02EBE1, 0x0001},
+{0x02EBF0, 0x0004},
+{0x02EE5E, 0x0001},
+{0x02F800, 0x0004},
+{0x02FA1E, 0x0001},
+{0x030000, 0x0004},
+{0x03134B, 0x0001},
+{0x031350, 0x0004},
+{0x0323B0, 0x0001},
+{0x0E0001, 0x0080},
+{0x0E0002, 0x0001},
+{0x0E0020, 0x0080},
+{0x0E0080, 0x0001},
+{0x0E0100, 0x0010},
+{0x0E01F0, 0x0001},
+{0x0F0000, 0x0080},
+{0x0FFFFE, 0x0001},
+{0x100000, 0x0080},
+{0x10FFFE, 0x0001},
+{0x110000, 0x0000},
+};
+
+const std::unordered_set unicode_set_whitespace = {
+0x000009,
+0x00000A,
+0x00000B,
+0x00000C,
+0x00000D,
+0x000020,
+0x000085,
+0x0000A0,
+0x001680,
+0x002000,
+0x002001,
+0x002002,
+0x002003,
+0x002004,
+0x002005,
+0x002006,
+0x002007,
+0x002008,
+0x002009,
+0x00200A,
+0x002028,
+0x002029,
+0x00202F,
+0x00205F,
+0x003000,
+};
+
+const std::unordered_map unicode_map_lowercase = {
+{0x000041, 0x000061},
+{0x000042, 0x000062},
+{0x000043, 0x000063},
+{0x000044, 0x000064},
+{0x000045, 0x000065},
+{0x000046, 0x000066},
+{0x000047, 0x000067},
+{0x000048, 0x000068},
+{0x000049, 0x000069},
+{0x00004A, 0x00006A},
+{0x00004B, 0x00006B},
+{0x00004C, 0x00006C},
+{0x00004D, 0x00006D},
+{0x00004E, 0x00006E},
+{0x00004F, 0x00006F},
+{0x000050, 0x000070},
+{0x000051, 0x000071},
+{0x000052, 0x000072},
+{0x000053, 0x000073},
+{0x000054, 0x000074},
+{0x000055, 0x000075},
+{0x000056, 0x000076},
+{0x000057, 0x000077},
+{0x000058, 0x000078},
+{0x000059, 0x000079},
+{0x00005A, 0x00007A},
+{0x0000C0, 0x0000E0},
+{0x0000C1, 0x0000E1},
+{0x0000C2, 0x0000E2},
+{0x0000C3, 0x0000E3},
+{0x0000C4, 0x0000E4},
+{0x0000C5, 0x0000E5},
+{0x0000C6, 0x0000E6},
+{0x0000C7, 0x0000E7},
+{0x0000C8, 0x0000E8},
+{0x0000C9, 0x0000E9},
+{0x0000CA, 0x0000EA},
+{0x0000CB, 0x0000EB},
+{0x0000CC, 0x0000EC},
+{0x0000CD, 0x0000ED},
+{0x0000CE, 0x0000EE},
+{0x0000CF, 0x0000EF},
+{0x0000D0, 0x0000F0},
+{0x0000D1, 0x0000F1},
+{0x0000D2, 0x0000F2},
+{0x0000D3, 0x0000F3},
+{0x0000D4, 0x0000F4},
+{0x0000D5, 0x0000F5},
+{0x0000D6, 0x0000F6},
+{0x0000D8, 0x0000F8},
+{0x0000D9, 0x0000F9},
+{0x0000DA, 0x0000FA},
+{0x0000DB, 0x0000FB},
+{0x0000DC, 0x0000FC},
+{0x0000DD, 0x0000FD},
+{0x0000DE, 0x0000FE},
+{0x000100, 0x000101},
+{0x000102, 0x000103},
+{0x000104, 0x000105},
+{0x000106, 0x000107},
+{0x000108, 0x000109},
+{0x00010A, 0x00010B},
+{0x00010C, 0x00010D},
+{0x00010E, 0x00010F},
+{0x000110, 0x000111},
+{0x000112, 0x000113},
+{0x000114, 0x000115},
+{0x000116, 0x000117},
+{0x000118, 0x000119},
+{0x00011A, 0x00011B},
+{0x00011C, 0x00011D},
+{0x00011E, 0x00011F},
+{0x000120, 0x000121},
+{0x000122, 0x000123},
+{0x000124, 0x000125},
+{0x000126, 0x000127},
+{0x000128, 0x000129},
+{0x00012A, 0x00012B},
+{0x00012C, 0x00012D},
+{0x00012E, 0x00012F},
+{0x000130, 0x000069},
+{0x000132, 0x000133},
+{0x000134, 0x000135},
+{0x000136, 0x000137},
+{0x000139, 0x00013A},
+{0x00013B, 0x00013C},
+{0x00013D, 0x00013E},
+{0x00013F, 0x000140},
+{0x000141, 0x000142},
+{0x000143, 0x000144},
+{0x000145, 0x000146},
+{0x000147, 0x000148},
+{0x00014A, 0x00014B},
+{0x00014C, 0x00014D},
+{0x00014E, 0x00014F},
+{0x000150, 0x000151},
+{0x000152, 0x000153},
+{0x000154, 0x000155},
+{0x000156, 0x000157},
+{0x000158, 0x000159},
+{0x00015A, 0x00015B},
+{0x00015C, 0x00015D},
+{0x00015E, 0x00015F},
+{0x000160, 0x000161},
+{0x000162, 0x000163},
+{0x000164, 0x000165},
+{0x000166, 0x000167},
+{0x000168, 0x000169},
+{0x00016A, 0x00016B},
+{0x00016C, 0x00016D},
+{0x00016E, 0x00016F},
+{0x000170, 0x000171},
+{0x000172, 0x000173},
+{0x000174, 0x000175},
+{0x000176, 0x000177},
+{0x000178, 0x0000FF},
+{0x000179, 0x00017A},
+{0x00017B, 0x00017C},
+{0x00017D, 0x00017E},
+{0x000181, 0x000253},
+{0x000182, 0x000183},
+{0x000184, 0x000185},
+{0x000186, 0x000254},
+{0x000187, 0x000188},
+{0x000189, 0x000256},
+{0x00018A, 0x000257},
+{0x00018B, 0x00018C},
+{0x00018E, 0x0001DD},
+{0x00018F, 0x000259},
+{0x000190, 0x00025B},
+{0x000191, 0x000192},
+{0x000193, 0x000260},
+{0x000194, 0x000263},
+{0x000196, 0x000269},
+{0x000197, 0x000268},
+{0x000198, 0x000199},
+{0x00019C, 0x00026F},
+{0x00019D, 0x000272},
+{0x00019F, 0x000275},
+{0x0001A0, 0x0001A1},
+{0x0001A2, 0x0001A3},
+{0x0001A4, 0x0001A5},
+{0x0001A6, 0x000280},
+{0x0001A7, 0x0001A8},
+{0x0001A9, 0x000283},
+{0x0001AC, 0x0001AD},
+{0x0001AE, 0x000288},
+{0x0001AF, 0x0001B0},
+{0x0001B1, 0x00028A},
+{0x0001B2, 0x00028B},
+{0x0001B3, 0x0001B4},
+{0x0001B5, 0x0001B6},
+{0x0001B7, 0x000292},
+{0x0001B8, 0x0001B9},
+{0x0001BC, 0x0001BD},
+{0x0001C4, 0x0001C6},
+{0x0001C5, 0x0001C6},
+{0x0001C7, 0x0001C9},
+{0x0001C8, 0x0001C9},
+{0x0001CA, 0x0001CC},
+{0x0001CB, 0x0001CC},
+{0x0001CD, 0x0001CE},
+{0x0001CF, 0x0001D0},
+{0x0001D1, 0x0001D2},
+{0x0001D3, 0x0001D4},
+{0x0001D5, 0x0001D6},
+{0x0001D7, 0x0001D8},
+{0x0001D9, 0x0001DA},
+{0x0001DB, 0x0001DC},
+{0x0001DE, 0x0001DF},
+{0x0001E0, 0x0001E1},
+{0x0001E2, 0x0001E3},
+{0x0001E4, 0x0001E5},
+{0x0001E6, 0x0001E7},
+{0x0001E8, 0x0001E9},
+{0x0001EA, 0x0001EB},
+{0x0001EC, 0x0001ED},
+{0x0001EE, 0x0001EF},
+{0x0001F1, 0x0001F3},
+{0x0001F2, 0x0001F3},
+{0x0001F4, 0x0001F5},
+{0x0001F6, 0x000195},
+{0x0001F7, 0x0001BF},
+{0x0001F8, 0x0001F9},
+{0x0001FA, 0x0001FB},
+{0x0001FC, 0x0001FD},
+{0x0001FE, 0x0001FF},
+{0x000200, 0x000201},
+{0x000202, 0x000203},
+{0x000204, 0x000205},
+{0x000206, 0x000207},
+{0x000208, 0x000209},
+{0x00020A, 0x00020B},
+{0x00020C, 0x00020D},
+{0x00020E, 0x00020F},
+{0x000210, 0x000211},
+{0x000212, 0x000213},
+{0x000214, 0x000215},
+{0x000216, 0x000217},
+{0x000218, 0x000219},
+{0x00021A, 0x00021B},
+{0x00021C, 0x00021D},
+{0x00021E, 0x00021F},
+{0x000220, 0x00019E},
+{0x000222, 0x000223},
+{0x000224, 0x000225},
+{0x000226, 0x000227},
+{0x000228, 0x000229},
+{0x00022A, 0x00022B},
+{0x00022C, 0x00022D},
+{0x00022E, 0x00022F},
+{0x000230, 0x000231},
+{0x000232, 0x000233},
+{0x00023A, 0x002C65},
+{0x00023B, 0x00023C},
+{0x00023D, 0x00019A},
+{0x00023E, 0x002C66},
+{0x000241, 0x000242},
+{0x000243, 0x000180},
+{0x000244, 0x000289},
+{0x000245, 0x00028C},
+{0x000246, 0x000247},
+{0x000248, 0x000249},
+{0x00024A, 0x00024B},
+{0x00024C, 0x00024D},
+{0x00024E, 0x00024F},
+{0x000370, 0x000371},
+{0x000372, 0x000373},
+{0x000376, 0x000377},
+{0x00037F, 0x0003F3},
+{0x000386, 0x0003AC},
+{0x000388, 0x0003AD},
+{0x000389, 0x0003AE},
+{0x00038A, 0x0003AF},
+{0x00038C, 0x0003CC},
+{0x00038E, 0x0003CD},
+{0x00038F, 0x0003CE},
+{0x000391, 0x0003B1},
+{0x000392, 0x0003B2},
+{0x000393, 0x0003B3},
+{0x000394, 0x0003B4},
+{0x000395, 0x0003B5},
+{0x000396, 0x0003B6},
+{0x000397, 0x0003B7},
+{0x000398, 0x0003B8},
+{0x000399, 0x0003B9},
+{0x00039A, 0x0003BA},
+{0x00039B, 0x0003BB},
+{0x00039C, 0x0003BC},
+{0x00039D, 0x0003BD},
+{0x00039E, 0x0003BE},
+{0x00039F, 0x0003BF},
+{0x0003A0, 0x0003C0},
+{0x0003A1, 0x0003C1},
+{0x0003A3, 0x0003C3},
+{0x0003A4, 0x0003C4},
+{0x0003A5, 0x0003C5},
+{0x0003A6, 0x0003C6},
+{0x0003A7, 0x0003C7},
+{0x0003A8, 0x0003C8},
+{0x0003A9, 0x0003C9},
+{0x0003AA, 0x0003CA},
+{0x0003AB, 0x0003CB},
+{0x0003CF, 0x0003D7},
+{0x0003D8, 0x0003D9},
+{0x0003DA, 0x0003DB},
+{0x0003DC, 0x0003DD},
+{0x0003DE, 0x0003DF},
+{0x0003E0, 0x0003E1},
+{0x0003E2, 0x0003E3},
+{0x0003E4, 0x0003E5},
+{0x0003E6, 0x0003E7},
+{0x0003E8, 0x0003E9},
+{0x0003EA, 0x0003EB},
+{0x0003EC, 0x0003ED},
+{0x0003EE, 0x0003EF},
+{0x0003F4, 0x0003B8},
+{0x0003F7, 0x0003F8},
+{0x0003F9, 0x0003F2},
+{0x0003FA, 0x0003FB},
+{0x0003FD, 0x00037B},
+{0x0003FE, 0x00037C},
+{0x0003FF, 0x00037D},
+{0x000400, 0x000450},
+{0x000401, 0x000451},
+{0x000402, 0x000452},
+{0x000403, 0x000453},
+{0x000404, 0x000454},
+{0x000405, 0x000455},
+{0x000406, 0x000456},
+{0x000407, 0x000457},
+{0x000408, 0x000458},
+{0x000409, 0x000459},
+{0x00040A, 0x00045A},
+{0x00040B, 0x00045B},
+{0x00040C, 0x00045C},
+{0x00040D, 0x00045D},
+{0x00040E, 0x00045E},
+{0x00040F, 0x00045F},
+{0x000410, 0x000430},
+{0x000411, 0x000431},
+{0x000412, 0x000432},
+{0x000413, 0x000433},
+{0x000414, 0x000434},
+{0x000415, 0x000435},
+{0x000416, 0x000436},
+{0x000417, 0x000437},
+{0x000418, 0x000438},
+{0x000419, 0x000439},
+{0x00041A, 0x00043A},
+{0x00041B, 0x00043B},
+{0x00041C, 0x00043C},
+{0x00041D, 0x00043D},
+{0x00041E, 0x00043E},
+{0x00041F, 0x00043F},
+{0x000420, 0x000440},
+{0x000421, 0x000441},
+{0x000422, 0x000442},
+{0x000423, 0x000443},
+{0x000424, 0x000444},
+{0x000425, 0x000445},
+{0x000426, 0x000446},
+{0x000427, 0x000447},
+{0x000428, 0x000448},
+{0x000429, 0x000449},
+{0x00042A, 0x00044A},
+{0x00042B, 0x00044B},
+{0x00042C, 0x00044C},
+{0x00042D, 0x00044D},
+{0x00042E, 0x00044E},
+{0x00042F, 0x00044F},
+{0x000460, 0x000461},
+{0x000462, 0x000463},
+{0x000464, 0x000465},
+{0x000466, 0x000467},
+{0x000468, 0x000469},
+{0x00046A, 0x00046B},
+{0x00046C, 0x00046D},
+{0x00046E, 0x00046F},
+{0x000470, 0x000471},
+{0x000472, 0x000473},
+{0x000474, 0x000475},
+{0x000476, 0x000477},
+{0x000478, 0x000479},
+{0x00047A, 0x00047B},
+{0x00047C, 0x00047D},
+{0x00047E, 0x00047F},
+{0x000480, 0x000481},
+{0x00048A, 0x00048B},
+{0x00048C, 0x00048D},
+{0x00048E, 0x00048F},
+{0x000490, 0x000491},
+{0x000492, 0x000493},
+{0x000494, 0x000495},
+{0x000496, 0x000497},
+{0x000498, 0x000499},
+{0x00049A, 0x00049B},
+{0x00049C, 0x00049D},
+{0x00049E, 0x00049F},
+{0x0004A0, 0x0004A1},
+{0x0004A2, 0x0004A3},
+{0x0004A4, 0x0004A5},
+{0x0004A6, 0x0004A7},
+{0x0004A8, 0x0004A9},
+{0x0004AA, 0x0004AB},
+{0x0004AC, 0x0004AD},
+{0x0004AE, 0x0004AF},
+{0x0004B0, 0x0004B1},
+{0x0004B2, 0x0004B3},
+{0x0004B4, 0x0004B5},
+{0x0004B6, 0x0004B7},
+{0x0004B8, 0x0004B9},
+{0x0004BA, 0x0004BB},
+{0x0004BC, 0x0004BD},
+{0x0004BE, 0x0004BF},
+{0x0004C0, 0x0004CF},
+{0x0004C1, 0x0004C2},
+{0x0004C3, 0x0004C4},
+{0x0004C5, 0x0004C6},
+{0x0004C7, 0x0004C8},
+{0x0004C9, 0x0004CA},
+{0x0004CB, 0x0004CC},
+{0x0004CD, 0x0004CE},
+{0x0004D0, 0x0004D1},
+{0x0004D2, 0x0004D3},
+{0x0004D4, 0x0004D5},
+{0x0004D6, 0x0004D7},
+{0x0004D8, 0x0004D9},
+{0x0004DA, 0x0004DB},
+{0x0004DC, 0x0004DD},
+{0x0004DE, 0x0004DF},
+{0x0004E0, 0x0004E1},
+{0x0004E2, 0x0004E3},
+{0x0004E4, 0x0004E5},
+{0x0004E6, 0x0004E7},
+{0x0004E8, 0x0004E9},
+{0x0004EA, 0x0004EB},
+{0x0004EC, 0x0004ED},
+{0x0004EE, 0x0004EF},
+{0x0004F0, 0x0004F1},
+{0x0004F2, 0x0004F3},
+{0x0004F4, 0x0004F5},
+{0x0004F6, 0x0004F7},
+{0x0004F8, 0x0004F9},
+{0x0004FA, 0x0004FB},
+{0x0004FC, 0x0004FD},
+{0x0004FE, 0x0004FF},
+{0x000500, 0x000501},
+{0x000502, 0x000503},
+{0x000504, 0x000505},
+{0x000506, 0x000507},
+{0x000508, 0x000509},
+{0x00050A, 0x00050B},
+{0x00050C, 0x00050D},
+{0x00050E, 0x00050F},
+{0x000510, 0x000511},
+{0x000512, 0x000513},
+{0x000514, 0x000515},
+{0x000516, 0x000517},
+{0x000518, 0x000519},
+{0x00051A, 0x00051B},
+{0x00051C, 0x00051D},
+{0x00051E, 0x00051F},
+{0x000520, 0x000521},
+{0x000522, 0x000523},
+{0x000524, 0x000525},
+{0x000526, 0x000527},
+{0x000528, 0x000529},
+{0x00052A, 0x00052B},
+{0x00052C, 0x00052D},
+{0x00052E, 0x00052F},
+{0x000531, 0x000561},
+{0x000532, 0x000562},
+{0x000533, 0x000563},
+{0x000534, 0x000564},
+{0x000535, 0x000565},
+{0x000536, 0x000566},
+{0x000537, 0x000567},
+{0x000538, 0x000568},
+{0x000539, 0x000569},
+{0x00053A, 0x00056A},
+{0x00053B, 0x00056B},
+{0x00053C, 0x00056C},
+{0x00053D, 0x00056D},
+{0x00053E, 0x00056E},
+{0x00053F, 0x00056F},
+{0x000540, 0x000570},
+{0x000541, 0x000571},
+{0x000542, 0x000572},
+{0x000543, 0x000573},
+{0x000544, 0x000574},
+{0x000545, 0x000575},
+{0x000546, 0x000576},
+{0x000547, 0x000577},
+{0x000548, 0x000578},
+{0x000549, 0x000579},
+{0x00054A, 0x00057A},
+{0x00054B, 0x00057B},
+{0x00054C, 0x00057C},
+{0x00054D, 0x00057D},
+{0x00054E, 0x00057E},
+{0x00054F, 0x00057F},
+{0x000550, 0x000580},
+{0x000551, 0x000581},
+{0x000552, 0x000582},
+{0x000553, 0x000583},
+{0x000554, 0x000584},
+{0x000555, 0x000585},
+{0x000556, 0x000586},
+{0x0010A0, 0x002D00},
+{0x0010A1, 0x002D01},
+{0x0010A2, 0x002D02},
+{0x0010A3, 0x002D03},
+{0x0010A4, 0x002D04},
+{0x0010A5, 0x002D05},
+{0x0010A6, 0x002D06},
+{0x0010A7, 0x002D07},
+{0x0010A8, 0x002D08},
+{0x0010A9, 0x002D09},
+{0x0010AA, 0x002D0A},
+{0x0010AB, 0x002D0B},
+{0x0010AC, 0x002D0C},
+{0x0010AD, 0x002D0D},
+{0x0010AE, 0x002D0E},
+{0x0010AF, 0x002D0F},
+{0x0010B0, 0x002D10},
+{0x0010B1, 0x002D11},
+{0x0010B2, 0x002D12},
+{0x0010B3, 0x002D13},
+{0x0010B4, 0x002D14},
+{0x0010B5, 0x002D15},
+{0x0010B6, 0x002D16},
+{0x0010B7, 0x002D17},
+{0x0010B8, 0x002D18},
+{0x0010B9, 0x002D19},
+{0x0010BA, 0x002D1A},
+{0x0010BB, 0x002D1B},
+{0x0010BC, 0x002D1C},
+{0x0010BD, 0x002D1D},
+{0x0010BE, 0x002D1E},
+{0x0010BF, 0x002D1F},
+{0x0010C0, 0x002D20},
+{0x0010C1, 0x002D21},
+{0x0010C2, 0x002D22},
+{0x0010C3, 0x002D23},
+{0x0010C4, 0x002D24},
+{0x0010C5, 0x002D25},
+{0x0010C7, 0x002D27},
+{0x0010CD, 0x002D2D},
+{0x0013A0, 0x00AB70},
+{0x0013A1, 0x00AB71},
+{0x0013A2, 0x00AB72},
+{0x0013A3, 0x00AB73},
+{0x0013A4, 0x00AB74},
+{0x0013A5, 0x00AB75},
+{0x0013A6, 0x00AB76},
+{0x0013A7, 0x00AB77},
+{0x0013A8, 0x00AB78},
+{0x0013A9, 0x00AB79},
+{0x0013AA, 0x00AB7A},
+{0x0013AB, 0x00AB7B},
+{0x0013AC, 0x00AB7C},
+{0x0013AD, 0x00AB7D},
+{0x0013AE, 0x00AB7E},
+{0x0013AF, 0x00AB7F},
+{0x0013B0, 0x00AB80},
+{0x0013B1, 0x00AB81},
+{0x0013B2, 0x00AB82},
+{0x0013B3, 0x00AB83},
+{0x0013B4, 0x00AB84},
+{0x0013B5, 0x00AB85},
+{0x0013B6, 0x00AB86},
+{0x0013B7, 0x00AB87},
+{0x0013B8, 0x00AB88},
+{0x0013B9, 0x00AB89},
+{0x0013BA, 0x00AB8A},
+{0x0013BB, 0x00AB8B},
+{0x0013BC, 0x00AB8C},
+{0x0013BD, 0x00AB8D},
+{0x0013BE, 0x00AB8E},
+{0x0013BF, 0x00AB8F},
+{0x0013C0, 0x00AB90},
+{0x0013C1, 0x00AB91},
+{0x0013C2, 0x00AB92},
+{0x0013C3, 0x00AB93},
+{0x0013C4, 0x00AB94},
+{0x0013C5, 0x00AB95},
+{0x0013C6, 0x00AB96},
+{0x0013C7, 0x00AB97},
+{0x0013C8, 0x00AB98},
+{0x0013C9, 0x00AB99},
+{0x0013CA, 0x00AB9A},
+{0x0013CB, 0x00AB9B},
+{0x0013CC, 0x00AB9C},
+{0x0013CD, 0x00AB9D},
+{0x0013CE, 0x00AB9E},
+{0x0013CF, 0x00AB9F},
+{0x0013D0, 0x00ABA0},
+{0x0013D1, 0x00ABA1},
+{0x0013D2, 0x00ABA2},
+{0x0013D3, 0x00ABA3},
+{0x0013D4, 0x00ABA4},
+{0x0013D5, 0x00ABA5},
+{0x0013D6, 0x00ABA6},
+{0x0013D7, 0x00ABA7},
+{0x0013D8, 0x00ABA8},
+{0x0013D9, 0x00ABA9},
+{0x0013DA, 0x00ABAA},
+{0x0013DB, 0x00ABAB},
+{0x0013DC, 0x00ABAC},
+{0x0013DD, 0x00ABAD},
+{0x0013DE, 0x00ABAE},
+{0x0013DF, 0x00ABAF},
+{0x0013E0, 0x00ABB0},
+{0x0013E1, 0x00ABB1},
+{0x0013E2, 0x00ABB2},
+{0x0013E3, 0x00ABB3},
+{0x0013E4, 0x00ABB4},
+{0x0013E5, 0x00ABB5},
+{0x0013E6, 0x00ABB6},
+{0x0013E7, 0x00ABB7},
+{0x0013E8, 0x00ABB8},
+{0x0013E9, 0x00ABB9},
+{0x0013EA, 0x00ABBA},
+{0x0013EB, 0x00ABBB},
+{0x0013EC, 0x00ABBC},
+{0x0013ED, 0x00ABBD},
+{0x0013EE, 0x00ABBE},
+{0x0013EF, 0x00ABBF},
+{0x0013F0, 0x0013F8},
+{0x0013F1, 0x0013F9},
+{0x0013F2, 0x0013FA},
+{0x0013F3, 0x0013FB},
+{0x0013F4, 0x0013FC},
+{0x0013F5, 0x0013FD},
+{0x001C90, 0x0010D0},
+{0x001C91, 0x0010D1},
+{0x001C92, 0x0010D2},
+{0x001C93, 0x0010D3},
+{0x001C94, 0x0010D4},
+{0x001C95, 0x0010D5},
+{0x001C96, 0x0010D6},
+{0x001C97, 0x0010D7},
+{0x001C98, 0x0010D8},
+{0x001C99, 0x0010D9},
+{0x001C9A, 0x0010DA},
+{0x001C9B, 0x0010DB},
+{0x001C9C, 0x0010DC},
+{0x001C9D, 0x0010DD},
+{0x001C9E, 0x0010DE},
+{0x001C9F, 0x0010DF},
+{0x001CA0, 0x0010E0},
+{0x001CA1, 0x0010E1},
+{0x001CA2, 0x0010E2},
+{0x001CA3, 0x0010E3},
+{0x001CA4, 0x0010E4},
+{0x001CA5, 0x0010E5},
+{0x001CA6, 0x0010E6},
+{0x001CA7, 0x0010E7},
+{0x001CA8, 0x0010E8},
+{0x001CA9, 0x0010E9},
+{0x001CAA, 0x0010EA},
+{0x001CAB, 0x0010EB},
+{0x001CAC, 0x0010EC},
+{0x001CAD, 0x0010ED},
+{0x001CAE, 0x0010EE},
+{0x001CAF, 0x0010EF},
+{0x001CB0, 0x0010F0},
+{0x001CB1, 0x0010F1},
+{0x001CB2, 0x0010F2},
+{0x001CB3, 0x0010F3},
+{0x001CB4, 0x0010F4},
+{0x001CB5, 0x0010F5},
+{0x001CB6, 0x0010F6},
+{0x001CB7, 0x0010F7},
+{0x001CB8, 0x0010F8},
+{0x001CB9, 0x0010F9},
+{0x001CBA, 0x0010FA},
+{0x001CBD, 0x0010FD},
+{0x001CBE, 0x0010FE},
+{0x001CBF, 0x0010FF},
+{0x001E00, 0x001E01},
+{0x001E02, 0x001E03},
+{0x001E04, 0x001E05},
+{0x001E06, 0x001E07},
+{0x001E08, 0x001E09},
+{0x001E0A, 0x001E0B},
+{0x001E0C, 0x001E0D},
+{0x001E0E, 0x001E0F},
+{0x001E10, 0x001E11},
+{0x001E12, 0x001E13},
+{0x001E14, 0x001E15},
+{0x001E16, 0x001E17},
+{0x001E18, 0x001E19},
+{0x001E1A, 0x001E1B},
+{0x001E1C, 0x001E1D},
+{0x001E1E, 0x001E1F},
+{0x001E20, 0x001E21},
+{0x001E22, 0x001E23},
+{0x001E24, 0x001E25},
+{0x001E26, 0x001E27},
+{0x001E28, 0x001E29},
+{0x001E2A, 0x001E2B},
+{0x001E2C, 0x001E2D},
+{0x001E2E, 0x001E2F},
+{0x001E30, 0x001E31},
+{0x001E32, 0x001E33},
+{0x001E34, 0x001E35},
+{0x001E36, 0x001E37},
+{0x001E38, 0x001E39},
+{0x001E3A, 0x001E3B},
+{0x001E3C, 0x001E3D},
+{0x001E3E, 0x001E3F},
+{0x001E40, 0x001E41},
+{0x001E42, 0x001E43},
+{0x001E44, 0x001E45},
+{0x001E46, 0x001E47},
+{0x001E48, 0x001E49},
+{0x001E4A, 0x001E4B},
+{0x001E4C, 0x001E4D},
+{0x001E4E, 0x001E4F},
+{0x001E50, 0x001E51},
+{0x001E52, 0x001E53},
+{0x001E54, 0x001E55},
+{0x001E56, 0x001E57},
+{0x001E58, 0x001E59},
+{0x001E5A, 0x001E5B},
+{0x001E5C, 0x001E5D},
+{0x001E5E, 0x001E5F},
+{0x001E60, 0x001E61},
+{0x001E62, 0x001E63},
+{0x001E64, 0x001E65},
+{0x001E66, 0x001E67},
+{0x001E68, 0x001E69},
+{0x001E6A, 0x001E6B},
+{0x001E6C, 0x001E6D},
+{0x001E6E, 0x001E6F},
+{0x001E70, 0x001E71},
+{0x001E72, 0x001E73},
+{0x001E74, 0x001E75},
+{0x001E76, 0x001E77},
+{0x001E78, 0x001E79},
+{0x001E7A, 0x001E7B},
+{0x001E7C, 0x001E7D},
+{0x001E7E, 0x001E7F},
+{0x001E80, 0x001E81},
+{0x001E82, 0x001E83},
+{0x001E84, 0x001E85},
+{0x001E86, 0x001E87},
+{0x001E88, 0x001E89},
+{0x001E8A, 0x001E8B},
+{0x001E8C, 0x001E8D},
+{0x001E8E, 0x001E8F},
+{0x001E90, 0x001E91},
+{0x001E92, 0x001E93},
+{0x001E94, 0x001E95},
+{0x001E9E, 0x0000DF},
+{0x001EA0, 0x001EA1},
+{0x001EA2, 0x001EA3},
+{0x001EA4, 0x001EA5},
+{0x001EA6, 0x001EA7},
+{0x001EA8, 0x001EA9},
+{0x001EAA, 0x001EAB},
+{0x001EAC, 0x001EAD},
+{0x001EAE, 0x001EAF},
+{0x001EB0, 0x001EB1},
+{0x001EB2, 0x001EB3},
+{0x001EB4, 0x001EB5},
+{0x001EB6, 0x001EB7},
+{0x001EB8, 0x001EB9},
+{0x001EBA, 0x001EBB},
+{0x001EBC, 0x001EBD},
+{0x001EBE, 0x001EBF},
+{0x001EC0, 0x001EC1},
+{0x001EC2, 0x001EC3},
+{0x001EC4, 0x001EC5},
+{0x001EC6, 0x001EC7},
+{0x001EC8, 0x001EC9},
+{0x001ECA, 0x001ECB},
+{0x001ECC, 0x001ECD},
+{0x001ECE, 0x001ECF},
+{0x001ED0, 0x001ED1},
+{0x001ED2, 0x001ED3},
+{0x001ED4, 0x001ED5},
+{0x001ED6, 0x001ED7},
+{0x001ED8, 0x001ED9},
+{0x001EDA, 0x001EDB},
+{0x001EDC, 0x001EDD},
+{0x001EDE, 0x001EDF},
+{0x001EE0, 0x001EE1},
+{0x001EE2, 0x001EE3},
+{0x001EE4, 0x001EE5},
+{0x001EE6, 0x001EE7},
+{0x001EE8, 0x001EE9},
+{0x001EEA, 0x001EEB},
+{0x001EEC, 0x001EED},
+{0x001EEE, 0x001EEF},
+{0x001EF0, 0x001EF1},
+{0x001EF2, 0x001EF3},
+{0x001EF4, 0x001EF5},
+{0x001EF6, 0x001EF7},
+{0x001EF8, 0x001EF9},
+{0x001EFA, 0x001EFB},
+{0x001EFC, 0x001EFD},
+{0x001EFE, 0x001EFF},
+{0x001F08, 0x001F00},
+{0x001F09, 0x001F01},
+{0x001F0A, 0x001F02},
+{0x001F0B, 0x001F03},
+{0x001F0C, 0x001F04},
+{0x001F0D, 0x001F05},
+{0x001F0E, 0x001F06},
+{0x001F0F, 0x001F07},
+{0x001F18, 0x001F10},
+{0x001F19, 0x001F11},
+{0x001F1A, 0x001F12},
+{0x001F1B, 0x001F13},
+{0x001F1C, 0x001F14},
+{0x001F1D, 0x001F15},
+{0x001F28, 0x001F20},
+{0x001F29, 0x001F21},
+{0x001F2A, 0x001F22},
+{0x001F2B, 0x001F23},
+{0x001F2C, 0x001F24},
+{0x001F2D, 0x001F25},
+{0x001F2E, 0x001F26},
+{0x001F2F, 0x001F27},
+{0x001F38, 0x001F30},
+{0x001F39, 0x001F31},
+{0x001F3A, 0x001F32},
+{0x001F3B, 0x001F33},
+{0x001F3C, 0x001F34},
+{0x001F3D, 0x001F35},
+{0x001F3E, 0x001F36},
+{0x001F3F, 0x001F37},
+{0x001F48, 0x001F40},
+{0x001F49, 0x001F41},
+{0x001F4A, 0x001F42},
+{0x001F4B, 0x001F43},
+{0x001F4C, 0x001F44},
+{0x001F4D, 0x001F45},
+{0x001F59, 0x001F51},
+{0x001F5B, 0x001F53},
+{0x001F5D, 0x001F55},
+{0x001F5F, 0x001F57},
+{0x001F68, 0x001F60},
+{0x001F69, 0x001F61},
+{0x001F6A, 0x001F62},
+{0x001F6B, 0x001F63},
+{0x001F6C, 0x001F64},
+{0x001F6D, 0x001F65},
+{0x001F6E, 0x001F66},
+{0x001F6F, 0x001F67},
+{0x001F88, 0x001F80},
+{0x001F89, 0x001F81},
+{0x001F8A, 0x001F82},
+{0x001F8B, 0x001F83},
+{0x001F8C, 0x001F84},
+{0x001F8D, 0x001F85},
+{0x001F8E, 0x001F86},
+{0x001F8F, 0x001F87},
+{0x001F98, 0x001F90},
+{0x001F99, 0x001F91},
+{0x001F9A, 0x001F92},
+{0x001F9B, 0x001F93},
+{0x001F9C, 0x001F94},
+{0x001F9D, 0x001F95},
+{0x001F9E, 0x001F96},
+{0x001F9F, 0x001F97},
+{0x001FA8, 0x001FA0},
+{0x001FA9, 0x001FA1},
+{0x001FAA, 0x001FA2},
+{0x001FAB, 0x001FA3},
+{0x001FAC, 0x001FA4},
+{0x001FAD, 0x001FA5},
+{0x001FAE, 0x001FA6},
+{0x001FAF, 0x001FA7},
+{0x001FB8, 0x001FB0},
+{0x001FB9, 0x001FB1},
+{0x001FBA, 0x001F70},
+{0x001FBB, 0x001F71},
+{0x001FBC, 0x001FB3},
+{0x001FC8, 0x001F72},
+{0x001FC9, 0x001F73},
+{0x001FCA, 0x001F74},
+{0x001FCB, 0x001F75},
+{0x001FCC, 0x001FC3},
+{0x001FD8, 0x001FD0},
+{0x001FD9, 0x001FD1},
+{0x001FDA, 0x001F76},
+{0x001FDB, 0x001F77},
+{0x001FE8, 0x001FE0},
+{0x001FE9, 0x001FE1},
+{0x001FEA, 0x001F7A},
+{0x001FEB, 0x001F7B},
+{0x001FEC, 0x001FE5},
+{0x001FF8, 0x001F78},
+{0x001FF9, 0x001F79},
+{0x001FFA, 0x001F7C},
+{0x001FFB, 0x001F7D},
+{0x001FFC, 0x001FF3},
+{0x002126, 0x0003C9},
+{0x00212A, 0x00006B},
+{0x00212B, 0x0000E5},
+{0x002132, 0x00214E},
+{0x002160, 0x002170},
+{0x002161, 0x002171},
+{0x002162, 0x002172},
+{0x002163, 0x002173},
+{0x002164, 0x002174},
+{0x002165, 0x002175},
+{0x002166, 0x002176},
+{0x002167, 0x002177},
+{0x002168, 0x002178},
+{0x002169, 0x002179},
+{0x00216A, 0x00217A},
+{0x00216B, 0x00217B},
+{0x00216C, 0x00217C},
+{0x00216D, 0x00217D},
+{0x00216E, 0x00217E},
+{0x00216F, 0x00217F},
+{0x002183, 0x002184},
+{0x0024B6, 0x0024D0},
+{0x0024B7, 0x0024D1},
+{0x0024B8, 0x0024D2},
+{0x0024B9, 0x0024D3},
+{0x0024BA, 0x0024D4},
+{0x0024BB, 0x0024D5},
+{0x0024BC, 0x0024D6},
+{0x0024BD, 0x0024D7},
+{0x0024BE, 0x0024D8},
+{0x0024BF, 0x0024D9},
+{0x0024C0, 0x0024DA},
+{0x0024C1, 0x0024DB},
+{0x0024C2, 0x0024DC},
+{0x0024C3, 0x0024DD},
+{0x0024C4, 0x0024DE},
+{0x0024C5, 0x0024DF},
+{0x0024C6, 0x0024E0},
+{0x0024C7, 0x0024E1},
+{0x0024C8, 0x0024E2},
+{0x0024C9, 0x0024E3},
+{0x0024CA, 0x0024E4},
+{0x0024CB, 0x0024E5},
+{0x0024CC, 0x0024E6},
+{0x0024CD, 0x0024E7},
+{0x0024CE, 0x0024E8},
+{0x0024CF, 0x0024E9},
+{0x002C00, 0x002C30},
+{0x002C01, 0x002C31},
+{0x002C02, 0x002C32},
+{0x002C03, 0x002C33},
+{0x002C04, 0x002C34},
+{0x002C05, 0x002C35},
+{0x002C06, 0x002C36},
+{0x002C07, 0x002C37},
+{0x002C08, 0x002C38},
+{0x002C09, 0x002C39},
+{0x002C0A, 0x002C3A},
+{0x002C0B, 0x002C3B},
+{0x002C0C, 0x002C3C},
+{0x002C0D, 0x002C3D},
+{0x002C0E, 0x002C3E},
+{0x002C0F, 0x002C3F},
+{0x002C10, 0x002C40},
+{0x002C11, 0x002C41},
+{0x002C12, 0x002C42},
+{0x002C13, 0x002C43},
+{0x002C14, 0x002C44},
+{0x002C15, 0x002C45},
+{0x002C16, 0x002C46},
+{0x002C17, 0x002C47},
+{0x002C18, 0x002C48},
+{0x002C19, 0x002C49},
+{0x002C1A, 0x002C4A},
+{0x002C1B, 0x002C4B},
+{0x002C1C, 0x002C4C},
+{0x002C1D, 0x002C4D},
+{0x002C1E, 0x002C4E},
+{0x002C1F, 0x002C4F},
+{0x002C20, 0x002C50},
+{0x002C21, 0x002C51},
+{0x002C22, 0x002C52},
+{0x002C23, 0x002C53},
+{0x002C24, 0x002C54},
+{0x002C25, 0x002C55},
+{0x002C26, 0x002C56},
+{0x002C27, 0x002C57},
+{0x002C28, 0x002C58},
+{0x002C29, 0x002C59},
+{0x002C2A, 0x002C5A},
+{0x002C2B, 0x002C5B},
+{0x002C2C, 0x002C5C},
+{0x002C2D, 0x002C5D},
+{0x002C2E, 0x002C5E},
+{0x002C2F, 0x002C5F},
+{0x002C60, 0x002C61},
+{0x002C62, 0x00026B},
+{0x002C63, 0x001D7D},
+{0x002C64, 0x00027D},
+{0x002C67, 0x002C68},
+{0x002C69, 0x002C6A},
+{0x002C6B, 0x002C6C},
+{0x002C6D, 0x000251},
+{0x002C6E, 0x000271},
+{0x002C6F, 0x000250},
+{0x002C70, 0x000252},
+{0x002C72, 0x002C73},
+{0x002C75, 0x002C76},
+{0x002C7E, 0x00023F},
+{0x002C7F, 0x000240},
+{0x002C80, 0x002C81},
+{0x002C82, 0x002C83},
+{0x002C84, 0x002C85},
+{0x002C86, 0x002C87},
+{0x002C88, 0x002C89},
+{0x002C8A, 0x002C8B},
+{0x002C8C, 0x002C8D},
+{0x002C8E, 0x002C8F},
+{0x002C90, 0x002C91},
+{0x002C92, 0x002C93},
+{0x002C94, 0x002C95},
+{0x002C96, 0x002C97},
+{0x002C98, 0x002C99},
+{0x002C9A, 0x002C9B},
+{0x002C9C, 0x002C9D},
+{0x002C9E, 0x002C9F},
+{0x002CA0, 0x002CA1},
+{0x002CA2, 0x002CA3},
+{0x002CA4, 0x002CA5},
+{0x002CA6, 0x002CA7},
+{0x002CA8, 0x002CA9},
+{0x002CAA, 0x002CAB},
+{0x002CAC, 0x002CAD},
+{0x002CAE, 0x002CAF},
+{0x002CB0, 0x002CB1},
+{0x002CB2, 0x002CB3},
+{0x002CB4, 0x002CB5},
+{0x002CB6, 0x002CB7},
+{0x002CB8, 0x002CB9},
+{0x002CBA, 0x002CBB},
+{0x002CBC, 0x002CBD},
+{0x002CBE, 0x002CBF},
+{0x002CC0, 0x002CC1},
+{0x002CC2, 0x002CC3},
+{0x002CC4, 0x002CC5},
+{0x002CC6, 0x002CC7},
+{0x002CC8, 0x002CC9},
+{0x002CCA, 0x002CCB},
+{0x002CCC, 0x002CCD},
+{0x002CCE, 0x002CCF},
+{0x002CD0, 0x002CD1},
+{0x002CD2, 0x002CD3},
+{0x002CD4, 0x002CD5},
+{0x002CD6, 0x002CD7},
+{0x002CD8, 0x002CD9},
+{0x002CDA, 0x002CDB},
+{0x002CDC, 0x002CDD},
+{0x002CDE, 0x002CDF},
+{0x002CE0, 0x002CE1},
+{0x002CE2, 0x002CE3},
+{0x002CEB, 0x002CEC},
+{0x002CED, 0x002CEE},
+{0x002CF2, 0x002CF3},
+{0x00A640, 0x00A641},
+{0x00A642, 0x00A643},
+{0x00A644, 0x00A645},
+{0x00A646, 0x00A647},
+{0x00A648, 0x00A649},
+{0x00A64A, 0x00A64B},
+{0x00A64C, 0x00A64D},
+{0x00A64E, 0x00A64F},
+{0x00A650, 0x00A651},
+{0x00A652, 0x00A653},
+{0x00A654, 0x00A655},
+{0x00A656, 0x00A657},
+{0x00A658, 0x00A659},
+{0x00A65A, 0x00A65B},
+{0x00A65C, 0x00A65D},
+{0x00A65E, 0x00A65F},
+{0x00A660, 0x00A661},
+{0x00A662, 0x00A663},
+{0x00A664, 0x00A665},
+{0x00A666, 0x00A667},
+{0x00A668, 0x00A669},
+{0x00A66A, 0x00A66B},
+{0x00A66C, 0x00A66D},
+{0x00A680, 0x00A681},
+{0x00A682, 0x00A683},
+{0x00A684, 0x00A685},
+{0x00A686, 0x00A687},
+{0x00A688, 0x00A689},
+{0x00A68A, 0x00A68B},
+{0x00A68C, 0x00A68D},
+{0x00A68E, 0x00A68F},
+{0x00A690, 0x00A691},
+{0x00A692, 0x00A693},
+{0x00A694, 0x00A695},
+{0x00A696, 0x00A697},
+{0x00A698, 0x00A699},
+{0x00A69A, 0x00A69B},
+{0x00A722, 0x00A723},
+{0x00A724, 0x00A725},
+{0x00A726, 0x00A727},
+{0x00A728, 0x00A729},
+{0x00A72A, 0x00A72B},
+{0x00A72C, 0x00A72D},
+{0x00A72E, 0x00A72F},
+{0x00A732, 0x00A733},
+{0x00A734, 0x00A735},
+{0x00A736, 0x00A737},
+{0x00A738, 0x00A739},
+{0x00A73A, 0x00A73B},
+{0x00A73C, 0x00A73D},
+{0x00A73E, 0x00A73F},
+{0x00A740, 0x00A741},
+{0x00A742, 0x00A743},
+{0x00A744, 0x00A745},
+{0x00A746, 0x00A747},
+{0x00A748, 0x00A749},
+{0x00A74A, 0x00A74B},
+{0x00A74C, 0x00A74D},
+{0x00A74E, 0x00A74F},
+{0x00A750, 0x00A751},
+{0x00A752, 0x00A753},
+{0x00A754, 0x00A755},
+{0x00A756, 0x00A757},
+{0x00A758, 0x00A759},
+{0x00A75A, 0x00A75B},
+{0x00A75C, 0x00A75D},
+{0x00A75E, 0x00A75F},
+{0x00A760, 0x00A761},
+{0x00A762, 0x00A763},
+{0x00A764, 0x00A765},
+{0x00A766, 0x00A767},
+{0x00A768, 0x00A769},
+{0x00A76A, 0x00A76B},
+{0x00A76C, 0x00A76D},
+{0x00A76E, 0x00A76F},
+{0x00A779, 0x00A77A},
+{0x00A77B, 0x00A77C},
+{0x00A77D, 0x001D79},
+{0x00A77E, 0x00A77F},
+{0x00A780, 0x00A781},
+{0x00A782, 0x00A783},
+{0x00A784, 0x00A785},
+{0x00A786, 0x00A787},
+{0x00A78B, 0x00A78C},
+{0x00A78D, 0x000265},
+{0x00A790, 0x00A791},
+{0x00A792, 0x00A793},
+{0x00A796, 0x00A797},
+{0x00A798, 0x00A799},
+{0x00A79A, 0x00A79B},
+{0x00A79C, 0x00A79D},
+{0x00A79E, 0x00A79F},
+{0x00A7A0, 0x00A7A1},
+{0x00A7A2, 0x00A7A3},
+{0x00A7A4, 0x00A7A5},
+{0x00A7A6, 0x00A7A7},
+{0x00A7A8, 0x00A7A9},
+{0x00A7AA, 0x000266},
+{0x00A7AB, 0x00025C},
+{0x00A7AC, 0x000261},
+{0x00A7AD, 0x00026C},
+{0x00A7AE, 0x00026A},
+{0x00A7B0, 0x00029E},
+{0x00A7B1, 0x000287},
+{0x00A7B2, 0x00029D},
+{0x00A7B3, 0x00AB53},
+{0x00A7B4, 0x00A7B5},
+{0x00A7B6, 0x00A7B7},
+{0x00A7B8, 0x00A7B9},
+{0x00A7BA, 0x00A7BB},
+{0x00A7BC, 0x00A7BD},
+{0x00A7BE, 0x00A7BF},
+{0x00A7C0, 0x00A7C1},
+{0x00A7C2, 0x00A7C3},
+{0x00A7C4, 0x00A794},
+{0x00A7C5, 0x000282},
+{0x00A7C6, 0x001D8E},
+{0x00A7C7, 0x00A7C8},
+{0x00A7C9, 0x00A7CA},
+{0x00A7D0, 0x00A7D1},
+{0x00A7D6, 0x00A7D7},
+{0x00A7D8, 0x00A7D9},
+{0x00A7F5, 0x00A7F6},
+{0x00FF21, 0x00FF41},
+{0x00FF22, 0x00FF42},
+{0x00FF23, 0x00FF43},
+{0x00FF24, 0x00FF44},
+{0x00FF25, 0x00FF45},
+{0x00FF26, 0x00FF46},
+{0x00FF27, 0x00FF47},
+{0x00FF28, 0x00FF48},
+{0x00FF29, 0x00FF49},
+{0x00FF2A, 0x00FF4A},
+{0x00FF2B, 0x00FF4B},
+{0x00FF2C, 0x00FF4C},
+{0x00FF2D, 0x00FF4D},
+{0x00FF2E, 0x00FF4E},
+{0x00FF2F, 0x00FF4F},
+{0x00FF30, 0x00FF50},
+{0x00FF31, 0x00FF51},
+{0x00FF32, 0x00FF52},
+{0x00FF33, 0x00FF53},
+{0x00FF34, 0x00FF54},
+{0x00FF35, 0x00FF55},
+{0x00FF36, 0x00FF56},
+{0x00FF37, 0x00FF57},
+{0x00FF38, 0x00FF58},
+{0x00FF39, 0x00FF59},
+{0x00FF3A, 0x00FF5A},
+{0x010400, 0x010428},
+{0x010401, 0x010429},
+{0x010402, 0x01042A},
+{0x010403, 0x01042B},
+{0x010404, 0x01042C},
+{0x010405, 0x01042D},
+{0x010406, 0x01042E},
+{0x010407, 0x01042F},
+{0x010408, 0x010430},
+{0x010409, 0x010431},
+{0x01040A, 0x010432},
+{0x01040B, 0x010433},
+{0x01040C, 0x010434},
+{0x01040D, 0x010435},
+{0x01040E, 0x010436},
+{0x01040F, 0x010437},
+{0x010410, 0x010438},
+{0x010411, 0x010439},
+{0x010412, 0x01043A},
+{0x010413, 0x01043B},
+{0x010414, 0x01043C},
+{0x010415, 0x01043D},
+{0x010416, 0x01043E},
+{0x010417, 0x01043F},
+{0x010418, 0x010440},
+{0x010419, 0x010441},
+{0x01041A, 0x010442},
+{0x01041B, 0x010443},
+{0x01041C, 0x010444},
+{0x01041D, 0x010445},
+{0x01041E, 0x010446},
+{0x01041F, 0x010447},
+{0x010420, 0x010448},
+{0x010421, 0x010449},
+{0x010422, 0x01044A},
+{0x010423, 0x01044B},
+{0x010424, 0x01044C},
+{0x010425, 0x01044D},
+{0x010426, 0x01044E},
+{0x010427, 0x01044F},
+{0x0104B0, 0x0104D8},
+{0x0104B1, 0x0104D9},
+{0x0104B2, 0x0104DA},
+{0x0104B3, 0x0104DB},
+{0x0104B4, 0x0104DC},
+{0x0104B5, 0x0104DD},
+{0x0104B6, 0x0104DE},
+{0x0104B7, 0x0104DF},
+{0x0104B8, 0x0104E0},
+{0x0104B9, 0x0104E1},
+{0x0104BA, 0x0104E2},
+{0x0104BB, 0x0104E3},
+{0x0104BC, 0x0104E4},
+{0x0104BD, 0x0104E5},
+{0x0104BE, 0x0104E6},
+{0x0104BF, 0x0104E7},
+{0x0104C0, 0x0104E8},
+{0x0104C1, 0x0104E9},
+{0x0104C2, 0x0104EA},
+{0x0104C3, 0x0104EB},
+{0x0104C4, 0x0104EC},
+{0x0104C5, 0x0104ED},
+{0x0104C6, 0x0104EE},
+{0x0104C7, 0x0104EF},
+{0x0104C8, 0x0104F0},
+{0x0104C9, 0x0104F1},
+{0x0104CA, 0x0104F2},
+{0x0104CB, 0x0104F3},
+{0x0104CC, 0x0104F4},
+{0x0104CD, 0x0104F5},
+{0x0104CE, 0x0104F6},
+{0x0104CF, 0x0104F7},
+{0x0104D0, 0x0104F8},
+{0x0104D1, 0x0104F9},
+{0x0104D2, 0x0104FA},
+{0x0104D3, 0x0104FB},
+{0x010570, 0x010597},
+{0x010571, 0x010598},
+{0x010572, 0x010599},
+{0x010573, 0x01059A},
+{0x010574, 0x01059B},
+{0x010575, 0x01059C},
+{0x010576, 0x01059D},
+{0x010577, 0x01059E},
+{0x010578, 0x01059F},
+{0x010579, 0x0105A0},
+{0x01057A, 0x0105A1},
+{0x01057C, 0x0105A3},
+{0x01057D, 0x0105A4},
+{0x01057E, 0x0105A5},
+{0x01057F, 0x0105A6},
+{0x010580, 0x0105A7},
+{0x010581, 0x0105A8},
+{0x010582, 0x0105A9},
+{0x010583, 0x0105AA},
+{0x010584, 0x0105AB},
+{0x010585, 0x0105AC},
+{0x010586, 0x0105AD},
+{0x010587, 0x0105AE},
+{0x010588, 0x0105AF},
+{0x010589, 0x0105B0},
+{0x01058A, 0x0105B1},
+{0x01058C, 0x0105B3},
+{0x01058D, 0x0105B4},
+{0x01058E, 0x0105B5},
+{0x01058F, 0x0105B6},
+{0x010590, 0x0105B7},
+{0x010591, 0x0105B8},
+{0x010592, 0x0105B9},
+{0x010594, 0x0105BB},
+{0x010595, 0x0105BC},
+{0x010C80, 0x010CC0},
+{0x010C81, 0x010CC1},
+{0x010C82, 0x010CC2},
+{0x010C83, 0x010CC3},
+{0x010C84, 0x010CC4},
+{0x010C85, 0x010CC5},
+{0x010C86, 0x010CC6},
+{0x010C87, 0x010CC7},
+{0x010C88, 0x010CC8},
+{0x010C89, 0x010CC9},
+{0x010C8A, 0x010CCA},
+{0x010C8B, 0x010CCB},
+{0x010C8C, 0x010CCC},
+{0x010C8D, 0x010CCD},
+{0x010C8E, 0x010CCE},
+{0x010C8F, 0x010CCF},
+{0x010C90, 0x010CD0},
+{0x010C91, 0x010CD1},
+{0x010C92, 0x010CD2},
+{0x010C93, 0x010CD3},
+{0x010C94, 0x010CD4},
+{0x010C95, 0x010CD5},
+{0x010C96, 0x010CD6},
+{0x010C97, 0x010CD7},
+{0x010C98, 0x010CD8},
+{0x010C99, 0x010CD9},
+{0x010C9A, 0x010CDA},
+{0x010C9B, 0x010CDB},
+{0x010C9C, 0x010CDC},
+{0x010C9D, 0x010CDD},
+{0x010C9E, 0x010CDE},
+{0x010C9F, 0x010CDF},
+{0x010CA0, 0x010CE0},
+{0x010CA1, 0x010CE1},
+{0x010CA2, 0x010CE2},
+{0x010CA3, 0x010CE3},
+{0x010CA4, 0x010CE4},
+{0x010CA5, 0x010CE5},
+{0x010CA6, 0x010CE6},
+{0x010CA7, 0x010CE7},
+{0x010CA8, 0x010CE8},
+{0x010CA9, 0x010CE9},
+{0x010CAA, 0x010CEA},
+{0x010CAB, 0x010CEB},
+{0x010CAC, 0x010CEC},
+{0x010CAD, 0x010CED},
+{0x010CAE, 0x010CEE},
+{0x010CAF, 0x010CEF},
+{0x010CB0, 0x010CF0},
+{0x010CB1, 0x010CF1},
+{0x010CB2, 0x010CF2},
+{0x0118A0, 0x0118C0},
+{0x0118A1, 0x0118C1},
+{0x0118A2, 0x0118C2},
+{0x0118A3, 0x0118C3},
+{0x0118A4, 0x0118C4},
+{0x0118A5, 0x0118C5},
+{0x0118A6, 0x0118C6},
+{0x0118A7, 0x0118C7},
+{0x0118A8, 0x0118C8},
+{0x0118A9, 0x0118C9},
+{0x0118AA, 0x0118CA},
+{0x0118AB, 0x0118CB},
+{0x0118AC, 0x0118CC},
+{0x0118AD, 0x0118CD},
+{0x0118AE, 0x0118CE},
+{0x0118AF, 0x0118CF},
+{0x0118B0, 0x0118D0},
+{0x0118B1, 0x0118D1},
+{0x0118B2, 0x0118D2},
+{0x0118B3, 0x0118D3},
+{0x0118B4, 0x0118D4},
+{0x0118B5, 0x0118D5},
+{0x0118B6, 0x0118D6},
+{0x0118B7, 0x0118D7},
+{0x0118B8, 0x0118D8},
+{0x0118B9, 0x0118D9},
+{0x0118BA, 0x0118DA},
+{0x0118BB, 0x0118DB},
+{0x0118BC, 0x0118DC},
+{0x0118BD, 0x0118DD},
+{0x0118BE, 0x0118DE},
+{0x0118BF, 0x0118DF},
+{0x016E40, 0x016E60},
+{0x016E41, 0x016E61},
+{0x016E42, 0x016E62},
+{0x016E43, 0x016E63},
+{0x016E44, 0x016E64},
+{0x016E45, 0x016E65},
+{0x016E46, 0x016E66},
+{0x016E47, 0x016E67},
+{0x016E48, 0x016E68},
+{0x016E49, 0x016E69},
+{0x016E4A, 0x016E6A},
+{0x016E4B, 0x016E6B},
+{0x016E4C, 0x016E6C},
+{0x016E4D, 0x016E6D},
+{0x016E4E, 0x016E6E},
+{0x016E4F, 0x016E6F},
+{0x016E50, 0x016E70},
+{0x016E51, 0x016E71},
+{0x016E52, 0x016E72},
+{0x016E53, 0x016E73},
+{0x016E54, 0x016E74},
+{0x016E55, 0x016E75},
+{0x016E56, 0x016E76},
+{0x016E57, 0x016E77},
+{0x016E58, 0x016E78},
+{0x016E59, 0x016E79},
+{0x016E5A, 0x016E7A},
+{0x016E5B, 0x016E7B},
+{0x016E5C, 0x016E7C},
+{0x016E5D, 0x016E7D},
+{0x016E5E, 0x016E7E},
+{0x016E5F, 0x016E7F},
+{0x01E900, 0x01E922},
+{0x01E901, 0x01E923},
+{0x01E902, 0x01E924},
+{0x01E903, 0x01E925},
+{0x01E904, 0x01E926},
+{0x01E905, 0x01E927},
+{0x01E906, 0x01E928},
+{0x01E907, 0x01E929},
+{0x01E908, 0x01E92A},
+{0x01E909, 0x01E92B},
+{0x01E90A, 0x01E92C},
+{0x01E90B, 0x01E92D},
+{0x01E90C, 0x01E92E},
+{0x01E90D, 0x01E92F},
+{0x01E90E, 0x01E930},
+{0x01E90F, 0x01E931},
+{0x01E910, 0x01E932},
+{0x01E911, 0x01E933},
+{0x01E912, 0x01E934},
+{0x01E913, 0x01E935},
+{0x01E914, 0x01E936},
+{0x01E915, 0x01E937},
+{0x01E916, 0x01E938},
+{0x01E917, 0x01E939},
+{0x01E918, 0x01E93A},
+{0x01E919, 0x01E93B},
+{0x01E91A, 0x01E93C},
+{0x01E91B, 0x01E93D},
+{0x01E91C, 0x01E93E},
+{0x01E91D, 0x01E93F},
+{0x01E91E, 0x01E940},
+{0x01E91F, 0x01E941},
+{0x01E920, 0x01E942},
+{0x01E921, 0x01E943},
+};
+
+const std::unordered_map unicode_map_uppercase = {
+{0x000061, 0x000041},
+{0x000062, 0x000042},
+{0x000063, 0x000043},
+{0x000064, 0x000044},
+{0x000065, 0x000045},
+{0x000066, 0x000046},
+{0x000067, 0x000047},
+{0x000068, 0x000048},
+{0x000069, 0x000049},
+{0x00006A, 0x00004A},
+{0x00006B, 0x00004B},
+{0x00006C, 0x00004C},
+{0x00006D, 0x00004D},
+{0x00006E, 0x00004E},
+{0x00006F, 0x00004F},
+{0x000070, 0x000050},
+{0x000071, 0x000051},
+{0x000072, 0x000052},
+{0x000073, 0x000053},
+{0x000074, 0x000054},
+{0x000075, 0x000055},
+{0x000076, 0x000056},
+{0x000077, 0x000057},
+{0x000078, 0x000058},
+{0x000079, 0x000059},
+{0x00007A, 0x00005A},
+{0x0000B5, 0x00039C},
+{0x0000E0, 0x0000C0},
+{0x0000E1, 0x0000C1},
+{0x0000E2, 0x0000C2},
+{0x0000E3, 0x0000C3},
+{0x0000E4, 0x0000C4},
+{0x0000E5, 0x0000C5},
+{0x0000E6, 0x0000C6},
+{0x0000E7, 0x0000C7},
+{0x0000E8, 0x0000C8},
+{0x0000E9, 0x0000C9},
+{0x0000EA, 0x0000CA},
+{0x0000EB, 0x0000CB},
+{0x0000EC, 0x0000CC},
+{0x0000ED, 0x0000CD},
+{0x0000EE, 0x0000CE},
+{0x0000EF, 0x0000CF},
+{0x0000F0, 0x0000D0},
+{0x0000F1, 0x0000D1},
+{0x0000F2, 0x0000D2},
+{0x0000F3, 0x0000D3},
+{0x0000F4, 0x0000D4},
+{0x0000F5, 0x0000D5},
+{0x0000F6, 0x0000D6},
+{0x0000F8, 0x0000D8},
+{0x0000F9, 0x0000D9},
+{0x0000FA, 0x0000DA},
+{0x0000FB, 0x0000DB},
+{0x0000FC, 0x0000DC},
+{0x0000FD, 0x0000DD},
+{0x0000FE, 0x0000DE},
+{0x0000FF, 0x000178},
+{0x000101, 0x000100},
+{0x000103, 0x000102},
+{0x000105, 0x000104},
+{0x000107, 0x000106},
+{0x000109, 0x000108},
+{0x00010B, 0x00010A},
+{0x00010D, 0x00010C},
+{0x00010F, 0x00010E},
+{0x000111, 0x000110},
+{0x000113, 0x000112},
+{0x000115, 0x000114},
+{0x000117, 0x000116},
+{0x000119, 0x000118},
+{0x00011B, 0x00011A},
+{0x00011D, 0x00011C},
+{0x00011F, 0x00011E},
+{0x000121, 0x000120},
+{0x000123, 0x000122},
+{0x000125, 0x000124},
+{0x000127, 0x000126},
+{0x000129, 0x000128},
+{0x00012B, 0x00012A},
+{0x00012D, 0x00012C},
+{0x00012F, 0x00012E},
+{0x000131, 0x000049},
+{0x000133, 0x000132},
+{0x000135, 0x000134},
+{0x000137, 0x000136},
+{0x00013A, 0x000139},
+{0x00013C, 0x00013B},
+{0x00013E, 0x00013D},
+{0x000140, 0x00013F},
+{0x000142, 0x000141},
+{0x000144, 0x000143},
+{0x000146, 0x000145},
+{0x000148, 0x000147},
+{0x00014B, 0x00014A},
+{0x00014D, 0x00014C},
+{0x00014F, 0x00014E},
+{0x000151, 0x000150},
+{0x000153, 0x000152},
+{0x000155, 0x000154},
+{0x000157, 0x000156},
+{0x000159, 0x000158},
+{0x00015B, 0x00015A},
+{0x00015D, 0x00015C},
+{0x00015F, 0x00015E},
+{0x000161, 0x000160},
+{0x000163, 0x000162},
+{0x000165, 0x000164},
+{0x000167, 0x000166},
+{0x000169, 0x000168},
+{0x00016B, 0x00016A},
+{0x00016D, 0x00016C},
+{0x00016F, 0x00016E},
+{0x000171, 0x000170},
+{0x000173, 0x000172},
+{0x000175, 0x000174},
+{0x000177, 0x000176},
+{0x00017A, 0x000179},
+{0x00017C, 0x00017B},
+{0x00017E, 0x00017D},
+{0x00017F, 0x000053},
+{0x000180, 0x000243},
+{0x000183, 0x000182},
+{0x000185, 0x000184},
+{0x000188, 0x000187},
+{0x00018C, 0x00018B},
+{0x000192, 0x000191},
+{0x000195, 0x0001F6},
+{0x000199, 0x000198},
+{0x00019A, 0x00023D},
+{0x00019E, 0x000220},
+{0x0001A1, 0x0001A0},
+{0x0001A3, 0x0001A2},
+{0x0001A5, 0x0001A4},
+{0x0001A8, 0x0001A7},
+{0x0001AD, 0x0001AC},
+{0x0001B0, 0x0001AF},
+{0x0001B4, 0x0001B3},
+{0x0001B6, 0x0001B5},
+{0x0001B9, 0x0001B8},
+{0x0001BD, 0x0001BC},
+{0x0001BF, 0x0001F7},
+{0x0001C5, 0x0001C4},
+{0x0001C6, 0x0001C4},
+{0x0001C8, 0x0001C7},
+{0x0001C9, 0x0001C7},
+{0x0001CB, 0x0001CA},
+{0x0001CC, 0x0001CA},
+{0x0001CE, 0x0001CD},
+{0x0001D0, 0x0001CF},
+{0x0001D2, 0x0001D1},
+{0x0001D4, 0x0001D3},
+{0x0001D6, 0x0001D5},
+{0x0001D8, 0x0001D7},
+{0x0001DA, 0x0001D9},
+{0x0001DC, 0x0001DB},
+{0x0001DD, 0x00018E},
+{0x0001DF, 0x0001DE},
+{0x0001E1, 0x0001E0},
+{0x0001E3, 0x0001E2},
+{0x0001E5, 0x0001E4},
+{0x0001E7, 0x0001E6},
+{0x0001E9, 0x0001E8},
+{0x0001EB, 0x0001EA},
+{0x0001ED, 0x0001EC},
+{0x0001EF, 0x0001EE},
+{0x0001F2, 0x0001F1},
+{0x0001F3, 0x0001F1},
+{0x0001F5, 0x0001F4},
+{0x0001F9, 0x0001F8},
+{0x0001FB, 0x0001FA},
+{0x0001FD, 0x0001FC},
+{0x0001FF, 0x0001FE},
+{0x000201, 0x000200},
+{0x000203, 0x000202},
+{0x000205, 0x000204},
+{0x000207, 0x000206},
+{0x000209, 0x000208},
+{0x00020B, 0x00020A},
+{0x00020D, 0x00020C},
+{0x00020F, 0x00020E},
+{0x000211, 0x000210},
+{0x000213, 0x000212},
+{0x000215, 0x000214},
+{0x000217, 0x000216},
+{0x000219, 0x000218},
+{0x00021B, 0x00021A},
+{0x00021D, 0x00021C},
+{0x00021F, 0x00021E},
+{0x000223, 0x000222},
+{0x000225, 0x000224},
+{0x000227, 0x000226},
+{0x000229, 0x000228},
+{0x00022B, 0x00022A},
+{0x00022D, 0x00022C},
+{0x00022F, 0x00022E},
+{0x000231, 0x000230},
+{0x000233, 0x000232},
+{0x00023C, 0x00023B},
+{0x00023F, 0x002C7E},
+{0x000240, 0x002C7F},
+{0x000242, 0x000241},
+{0x000247, 0x000246},
+{0x000249, 0x000248},
+{0x00024B, 0x00024A},
+{0x00024D, 0x00024C},
+{0x00024F, 0x00024E},
+{0x000250, 0x002C6F},
+{0x000251, 0x002C6D},
+{0x000252, 0x002C70},
+{0x000253, 0x000181},
+{0x000254, 0x000186},
+{0x000256, 0x000189},
+{0x000257, 0x00018A},
+{0x000259, 0x00018F},
+{0x00025B, 0x000190},
+{0x00025C, 0x00A7AB},
+{0x000260, 0x000193},
+{0x000261, 0x00A7AC},
+{0x000263, 0x000194},
+{0x000265, 0x00A78D},
+{0x000266, 0x00A7AA},
+{0x000268, 0x000197},
+{0x000269, 0x000196},
+{0x00026A, 0x00A7AE},
+{0x00026B, 0x002C62},
+{0x00026C, 0x00A7AD},
+{0x00026F, 0x00019C},
+{0x000271, 0x002C6E},
+{0x000272, 0x00019D},
+{0x000275, 0x00019F},
+{0x00027D, 0x002C64},
+{0x000280, 0x0001A6},
+{0x000282, 0x00A7C5},
+{0x000283, 0x0001A9},
+{0x000287, 0x00A7B1},
+{0x000288, 0x0001AE},
+{0x000289, 0x000244},
+{0x00028A, 0x0001B1},
+{0x00028B, 0x0001B2},
+{0x00028C, 0x000245},
+{0x000292, 0x0001B7},
+{0x00029D, 0x00A7B2},
+{0x00029E, 0x00A7B0},
+{0x000345, 0x000399},
+{0x000371, 0x000370},
+{0x000373, 0x000372},
+{0x000377, 0x000376},
+{0x00037B, 0x0003FD},
+{0x00037C, 0x0003FE},
+{0x00037D, 0x0003FF},
+{0x0003AC, 0x000386},
+{0x0003AD, 0x000388},
+{0x0003AE, 0x000389},
+{0x0003AF, 0x00038A},
+{0x0003B1, 0x000391},
+{0x0003B2, 0x000392},
+{0x0003B3, 0x000393},
+{0x0003B4, 0x000394},
+{0x0003B5, 0x000395},
+{0x0003B6, 0x000396},
+{0x0003B7, 0x000397},
+{0x0003B8, 0x000398},
+{0x0003B9, 0x000399},
+{0x0003BA, 0x00039A},
+{0x0003BB, 0x00039B},
+{0x0003BC, 0x00039C},
+{0x0003BD, 0x00039D},
+{0x0003BE, 0x00039E},
+{0x0003BF, 0x00039F},
+{0x0003C0, 0x0003A0},
+{0x0003C1, 0x0003A1},
+{0x0003C2, 0x0003A3},
+{0x0003C3, 0x0003A3},
+{0x0003C4, 0x0003A4},
+{0x0003C5, 0x0003A5},
+{0x0003C6, 0x0003A6},
+{0x0003C7, 0x0003A7},
+{0x0003C8, 0x0003A8},
+{0x0003C9, 0x0003A9},
+{0x0003CA, 0x0003AA},
+{0x0003CB, 0x0003AB},
+{0x0003CC, 0x00038C},
+{0x0003CD, 0x00038E},
+{0x0003CE, 0x00038F},
+{0x0003D0, 0x000392},
+{0x0003D1, 0x000398},
+{0x0003D5, 0x0003A6},
+{0x0003D6, 0x0003A0},
+{0x0003D7, 0x0003CF},
+{0x0003D9, 0x0003D8},
+{0x0003DB, 0x0003DA},
+{0x0003DD, 0x0003DC},
+{0x0003DF, 0x0003DE},
+{0x0003E1, 0x0003E0},
+{0x0003E3, 0x0003E2},
+{0x0003E5, 0x0003E4},
+{0x0003E7, 0x0003E6},
+{0x0003E9, 0x0003E8},
+{0x0003EB, 0x0003EA},
+{0x0003ED, 0x0003EC},
+{0x0003EF, 0x0003EE},
+{0x0003F0, 0x00039A},
+{0x0003F1, 0x0003A1},
+{0x0003F2, 0x0003F9},
+{0x0003F3, 0x00037F},
+{0x0003F5, 0x000395},
+{0x0003F8, 0x0003F7},
+{0x0003FB, 0x0003FA},
+{0x000430, 0x000410},
+{0x000431, 0x000411},
+{0x000432, 0x000412},
+{0x000433, 0x000413},
+{0x000434, 0x000414},
+{0x000435, 0x000415},
+{0x000436, 0x000416},
+{0x000437, 0x000417},
+{0x000438, 0x000418},
+{0x000439, 0x000419},
+{0x00043A, 0x00041A},
+{0x00043B, 0x00041B},
+{0x00043C, 0x00041C},
+{0x00043D, 0x00041D},
+{0x00043E, 0x00041E},
+{0x00043F, 0x00041F},
+{0x000440, 0x000420},
+{0x000441, 0x000421},
+{0x000442, 0x000422},
+{0x000443, 0x000423},
+{0x000444, 0x000424},
+{0x000445, 0x000425},
+{0x000446, 0x000426},
+{0x000447, 0x000427},
+{0x000448, 0x000428},
+{0x000449, 0x000429},
+{0x00044A, 0x00042A},
+{0x00044B, 0x00042B},
+{0x00044C, 0x00042C},
+{0x00044D, 0x00042D},
+{0x00044E, 0x00042E},
+{0x00044F, 0x00042F},
+{0x000450, 0x000400},
+{0x000451, 0x000401},
+{0x000452, 0x000402},
+{0x000453, 0x000403},
+{0x000454, 0x000404},
+{0x000455, 0x000405},
+{0x000456, 0x000406},
+{0x000457, 0x000407},
+{0x000458, 0x000408},
+{0x000459, 0x000409},
+{0x00045A, 0x00040A},
+{0x00045B, 0x00040B},
+{0x00045C, 0x00040C},
+{0x00045D, 0x00040D},
+{0x00045E, 0x00040E},
+{0x00045F, 0x00040F},
+{0x000461, 0x000460},
+{0x000463, 0x000462},
+{0x000465, 0x000464},
+{0x000467, 0x000466},
+{0x000469, 0x000468},
+{0x00046B, 0x00046A},
+{0x00046D, 0x00046C},
+{0x00046F, 0x00046E},
+{0x000471, 0x000470},
+{0x000473, 0x000472},
+{0x000475, 0x000474},
+{0x000477, 0x000476},
+{0x000479, 0x000478},
+{0x00047B, 0x00047A},
+{0x00047D, 0x00047C},
+{0x00047F, 0x00047E},
+{0x000481, 0x000480},
+{0x00048B, 0x00048A},
+{0x00048D, 0x00048C},
+{0x00048F, 0x00048E},
+{0x000491, 0x000490},
+{0x000493, 0x000492},
+{0x000495, 0x000494},
+{0x000497, 0x000496},
+{0x000499, 0x000498},
+{0x00049B, 0x00049A},
+{0x00049D, 0x00049C},
+{0x00049F, 0x00049E},
+{0x0004A1, 0x0004A0},
+{0x0004A3, 0x0004A2},
+{0x0004A5, 0x0004A4},
+{0x0004A7, 0x0004A6},
+{0x0004A9, 0x0004A8},
+{0x0004AB, 0x0004AA},
+{0x0004AD, 0x0004AC},
+{0x0004AF, 0x0004AE},
+{0x0004B1, 0x0004B0},
+{0x0004B3, 0x0004B2},
+{0x0004B5, 0x0004B4},
+{0x0004B7, 0x0004B6},
+{0x0004B9, 0x0004B8},
+{0x0004BB, 0x0004BA},
+{0x0004BD, 0x0004BC},
+{0x0004BF, 0x0004BE},
+{0x0004C2, 0x0004C1},
+{0x0004C4, 0x0004C3},
+{0x0004C6, 0x0004C5},
+{0x0004C8, 0x0004C7},
+{0x0004CA, 0x0004C9},
+{0x0004CC, 0x0004CB},
+{0x0004CE, 0x0004CD},
+{0x0004CF, 0x0004C0},
+{0x0004D1, 0x0004D0},
+{0x0004D3, 0x0004D2},
+{0x0004D5, 0x0004D4},
+{0x0004D7, 0x0004D6},
+{0x0004D9, 0x0004D8},
+{0x0004DB, 0x0004DA},
+{0x0004DD, 0x0004DC},
+{0x0004DF, 0x0004DE},
+{0x0004E1, 0x0004E0},
+{0x0004E3, 0x0004E2},
+{0x0004E5, 0x0004E4},
+{0x0004E7, 0x0004E6},
+{0x0004E9, 0x0004E8},
+{0x0004EB, 0x0004EA},
+{0x0004ED, 0x0004EC},
+{0x0004EF, 0x0004EE},
+{0x0004F1, 0x0004F0},
+{0x0004F3, 0x0004F2},
+{0x0004F5, 0x0004F4},
+{0x0004F7, 0x0004F6},
+{0x0004F9, 0x0004F8},
+{0x0004FB, 0x0004FA},
+{0x0004FD, 0x0004FC},
+{0x0004FF, 0x0004FE},
+{0x000501, 0x000500},
+{0x000503, 0x000502},
+{0x000505, 0x000504},
+{0x000507, 0x000506},
+{0x000509, 0x000508},
+{0x00050B, 0x00050A},
+{0x00050D, 0x00050C},
+{0x00050F, 0x00050E},
+{0x000511, 0x000510},
+{0x000513, 0x000512},
+{0x000515, 0x000514},
+{0x000517, 0x000516},
+{0x000519, 0x000518},
+{0x00051B, 0x00051A},
+{0x00051D, 0x00051C},
+{0x00051F, 0x00051E},
+{0x000521, 0x000520},
+{0x000523, 0x000522},
+{0x000525, 0x000524},
+{0x000527, 0x000526},
+{0x000529, 0x000528},
+{0x00052B, 0x00052A},
+{0x00052D, 0x00052C},
+{0x00052F, 0x00052E},
+{0x000561, 0x000531},
+{0x000562, 0x000532},
+{0x000563, 0x000533},
+{0x000564, 0x000534},
+{0x000565, 0x000535},
+{0x000566, 0x000536},
+{0x000567, 0x000537},
+{0x000568, 0x000538},
+{0x000569, 0x000539},
+{0x00056A, 0x00053A},
+{0x00056B, 0x00053B},
+{0x00056C, 0x00053C},
+{0x00056D, 0x00053D},
+{0x00056E, 0x00053E},
+{0x00056F, 0x00053F},
+{0x000570, 0x000540},
+{0x000571, 0x000541},
+{0x000572, 0x000542},
+{0x000573, 0x000543},
+{0x000574, 0x000544},
+{0x000575, 0x000545},
+{0x000576, 0x000546},
+{0x000577, 0x000547},
+{0x000578, 0x000548},
+{0x000579, 0x000549},
+{0x00057A, 0x00054A},
+{0x00057B, 0x00054B},
+{0x00057C, 0x00054C},
+{0x00057D, 0x00054D},
+{0x00057E, 0x00054E},
+{0x00057F, 0x00054F},
+{0x000580, 0x000550},
+{0x000581, 0x000551},
+{0x000582, 0x000552},
+{0x000583, 0x000553},
+{0x000584, 0x000554},
+{0x000585, 0x000555},
+{0x000586, 0x000556},
+{0x0010D0, 0x001C90},
+{0x0010D1, 0x001C91},
+{0x0010D2, 0x001C92},
+{0x0010D3, 0x001C93},
+{0x0010D4, 0x001C94},
+{0x0010D5, 0x001C95},
+{0x0010D6, 0x001C96},
+{0x0010D7, 0x001C97},
+{0x0010D8, 0x001C98},
+{0x0010D9, 0x001C99},
+{0x0010DA, 0x001C9A},
+{0x0010DB, 0x001C9B},
+{0x0010DC, 0x001C9C},
+{0x0010DD, 0x001C9D},
+{0x0010DE, 0x001C9E},
+{0x0010DF, 0x001C9F},
+{0x0010E0, 0x001CA0},
+{0x0010E1, 0x001CA1},
+{0x0010E2, 0x001CA2},
+{0x0010E3, 0x001CA3},
+{0x0010E4, 0x001CA4},
+{0x0010E5, 0x001CA5},
+{0x0010E6, 0x001CA6},
+{0x0010E7, 0x001CA7},
+{0x0010E8, 0x001CA8},
+{0x0010E9, 0x001CA9},
+{0x0010EA, 0x001CAA},
+{0x0010EB, 0x001CAB},
+{0x0010EC, 0x001CAC},
+{0x0010ED, 0x001CAD},
+{0x0010EE, 0x001CAE},
+{0x0010EF, 0x001CAF},
+{0x0010F0, 0x001CB0},
+{0x0010F1, 0x001CB1},
+{0x0010F2, 0x001CB2},
+{0x0010F3, 0x001CB3},
+{0x0010F4, 0x001CB4},
+{0x0010F5, 0x001CB5},
+{0x0010F6, 0x001CB6},
+{0x0010F7, 0x001CB7},
+{0x0010F8, 0x001CB8},
+{0x0010F9, 0x001CB9},
+{0x0010FA, 0x001CBA},
+{0x0010FD, 0x001CBD},
+{0x0010FE, 0x001CBE},
+{0x0010FF, 0x001CBF},
+{0x0013F8, 0x0013F0},
+{0x0013F9, 0x0013F1},
+{0x0013FA, 0x0013F2},
+{0x0013FB, 0x0013F3},
+{0x0013FC, 0x0013F4},
+{0x0013FD, 0x0013F5},
+{0x001C80, 0x000412},
+{0x001C81, 0x000414},
+{0x001C82, 0x00041E},
+{0x001C83, 0x000421},
+{0x001C84, 0x000422},
+{0x001C85, 0x000422},
+{0x001C86, 0x00042A},
+{0x001C87, 0x000462},
+{0x001C88, 0x00A64A},
+{0x001D79, 0x00A77D},
+{0x001D7D, 0x002C63},
+{0x001D8E, 0x00A7C6},
+{0x001E01, 0x001E00},
+{0x001E03, 0x001E02},
+{0x001E05, 0x001E04},
+{0x001E07, 0x001E06},
+{0x001E09, 0x001E08},
+{0x001E0B, 0x001E0A},
+{0x001E0D, 0x001E0C},
+{0x001E0F, 0x001E0E},
+{0x001E11, 0x001E10},
+{0x001E13, 0x001E12},
+{0x001E15, 0x001E14},
+{0x001E17, 0x001E16},
+{0x001E19, 0x001E18},
+{0x001E1B, 0x001E1A},
+{0x001E1D, 0x001E1C},
+{0x001E1F, 0x001E1E},
+{0x001E21, 0x001E20},
+{0x001E23, 0x001E22},
+{0x001E25, 0x001E24},
+{0x001E27, 0x001E26},
+{0x001E29, 0x001E28},
+{0x001E2B, 0x001E2A},
+{0x001E2D, 0x001E2C},
+{0x001E2F, 0x001E2E},
+{0x001E31, 0x001E30},
+{0x001E33, 0x001E32},
+{0x001E35, 0x001E34},
+{0x001E37, 0x001E36},
+{0x001E39, 0x001E38},
+{0x001E3B, 0x001E3A},
+{0x001E3D, 0x001E3C},
+{0x001E3F, 0x001E3E},
+{0x001E41, 0x001E40},
+{0x001E43, 0x001E42},
+{0x001E45, 0x001E44},
+{0x001E47, 0x001E46},
+{0x001E49, 0x001E48},
+{0x001E4B, 0x001E4A},
+{0x001E4D, 0x001E4C},
+{0x001E4F, 0x001E4E},
+{0x001E51, 0x001E50},
+{0x001E53, 0x001E52},
+{0x001E55, 0x001E54},
+{0x001E57, 0x001E56},
+{0x001E59, 0x001E58},
+{0x001E5B, 0x001E5A},
+{0x001E5D, 0x001E5C},
+{0x001E5F, 0x001E5E},
+{0x001E61, 0x001E60},
+{0x001E63, 0x001E62},
+{0x001E65, 0x001E64},
+{0x001E67, 0x001E66},
+{0x001E69, 0x001E68},
+{0x001E6B, 0x001E6A},
+{0x001E6D, 0x001E6C},
+{0x001E6F, 0x001E6E},
+{0x001E71, 0x001E70},
+{0x001E73, 0x001E72},
+{0x001E75, 0x001E74},
+{0x001E77, 0x001E76},
+{0x001E79, 0x001E78},
+{0x001E7B, 0x001E7A},
+{0x001E7D, 0x001E7C},
+{0x001E7F, 0x001E7E},
+{0x001E81, 0x001E80},
+{0x001E83, 0x001E82},
+{0x001E85, 0x001E84},
+{0x001E87, 0x001E86},
+{0x001E89, 0x001E88},
+{0x001E8B, 0x001E8A},
+{0x001E8D, 0x001E8C},
+{0x001E8F, 0x001E8E},
+{0x001E91, 0x001E90},
+{0x001E93, 0x001E92},
+{0x001E95, 0x001E94},
+{0x001E9B, 0x001E60},
+{0x001EA1, 0x001EA0},
+{0x001EA3, 0x001EA2},
+{0x001EA5, 0x001EA4},
+{0x001EA7, 0x001EA6},
+{0x001EA9, 0x001EA8},
+{0x001EAB, 0x001EAA},
+{0x001EAD, 0x001EAC},
+{0x001EAF, 0x001EAE},
+{0x001EB1, 0x001EB0},
+{0x001EB3, 0x001EB2},
+{0x001EB5, 0x001EB4},
+{0x001EB7, 0x001EB6},
+{0x001EB9, 0x001EB8},
+{0x001EBB, 0x001EBA},
+{0x001EBD, 0x001EBC},
+{0x001EBF, 0x001EBE},
+{0x001EC1, 0x001EC0},
+{0x001EC3, 0x001EC2},
+{0x001EC5, 0x001EC4},
+{0x001EC7, 0x001EC6},
+{0x001EC9, 0x001EC8},
+{0x001ECB, 0x001ECA},
+{0x001ECD, 0x001ECC},
+{0x001ECF, 0x001ECE},
+{0x001ED1, 0x001ED0},
+{0x001ED3, 0x001ED2},
+{0x001ED5, 0x001ED4},
+{0x001ED7, 0x001ED6},
+{0x001ED9, 0x001ED8},
+{0x001EDB, 0x001EDA},
+{0x001EDD, 0x001EDC},
+{0x001EDF, 0x001EDE},
+{0x001EE1, 0x001EE0},
+{0x001EE3, 0x001EE2},
+{0x001EE5, 0x001EE4},
+{0x001EE7, 0x001EE6},
+{0x001EE9, 0x001EE8},
+{0x001EEB, 0x001EEA},
+{0x001EED, 0x001EEC},
+{0x001EEF, 0x001EEE},
+{0x001EF1, 0x001EF0},
+{0x001EF3, 0x001EF2},
+{0x001EF5, 0x001EF4},
+{0x001EF7, 0x001EF6},
+{0x001EF9, 0x001EF8},
+{0x001EFB, 0x001EFA},
+{0x001EFD, 0x001EFC},
+{0x001EFF, 0x001EFE},
+{0x001F00, 0x001F08},
+{0x001F01, 0x001F09},
+{0x001F02, 0x001F0A},
+{0x001F03, 0x001F0B},
+{0x001F04, 0x001F0C},
+{0x001F05, 0x001F0D},
+{0x001F06, 0x001F0E},
+{0x001F07, 0x001F0F},
+{0x001F10, 0x001F18},
+{0x001F11, 0x001F19},
+{0x001F12, 0x001F1A},
+{0x001F13, 0x001F1B},
+{0x001F14, 0x001F1C},
+{0x001F15, 0x001F1D},
+{0x001F20, 0x001F28},
+{0x001F21, 0x001F29},
+{0x001F22, 0x001F2A},
+{0x001F23, 0x001F2B},
+{0x001F24, 0x001F2C},
+{0x001F25, 0x001F2D},
+{0x001F26, 0x001F2E},
+{0x001F27, 0x001F2F},
+{0x001F30, 0x001F38},
+{0x001F31, 0x001F39},
+{0x001F32, 0x001F3A},
+{0x001F33, 0x001F3B},
+{0x001F34, 0x001F3C},
+{0x001F35, 0x001F3D},
+{0x001F36, 0x001F3E},
+{0x001F37, 0x001F3F},
+{0x001F40, 0x001F48},
+{0x001F41, 0x001F49},
+{0x001F42, 0x001F4A},
+{0x001F43, 0x001F4B},
+{0x001F44, 0x001F4C},
+{0x001F45, 0x001F4D},
+{0x001F51, 0x001F59},
+{0x001F53, 0x001F5B},
+{0x001F55, 0x001F5D},
+{0x001F57, 0x001F5F},
+{0x001F60, 0x001F68},
+{0x001F61, 0x001F69},
+{0x001F62, 0x001F6A},
+{0x001F63, 0x001F6B},
+{0x001F64, 0x001F6C},
+{0x001F65, 0x001F6D},
+{0x001F66, 0x001F6E},
+{0x001F67, 0x001F6F},
+{0x001F70, 0x001FBA},
+{0x001F71, 0x001FBB},
+{0x001F72, 0x001FC8},
+{0x001F73, 0x001FC9},
+{0x001F74, 0x001FCA},
+{0x001F75, 0x001FCB},
+{0x001F76, 0x001FDA},
+{0x001F77, 0x001FDB},
+{0x001F78, 0x001FF8},
+{0x001F79, 0x001FF9},
+{0x001F7A, 0x001FEA},
+{0x001F7B, 0x001FEB},
+{0x001F7C, 0x001FFA},
+{0x001F7D, 0x001FFB},
+{0x001F80, 0x001F88},
+{0x001F81, 0x001F89},
+{0x001F82, 0x001F8A},
+{0x001F83, 0x001F8B},
+{0x001F84, 0x001F8C},
+{0x001F85, 0x001F8D},
+{0x001F86, 0x001F8E},
+{0x001F87, 0x001F8F},
+{0x001F90, 0x001F98},
+{0x001F91, 0x001F99},
+{0x001F92, 0x001F9A},
+{0x001F93, 0x001F9B},
+{0x001F94, 0x001F9C},
+{0x001F95, 0x001F9D},
+{0x001F96, 0x001F9E},
+{0x001F97, 0x001F9F},
+{0x001FA0, 0x001FA8},
+{0x001FA1, 0x001FA9},
+{0x001FA2, 0x001FAA},
+{0x001FA3, 0x001FAB},
+{0x001FA4, 0x001FAC},
+{0x001FA5, 0x001FAD},
+{0x001FA6, 0x001FAE},
+{0x001FA7, 0x001FAF},
+{0x001FB0, 0x001FB8},
+{0x001FB1, 0x001FB9},
+{0x001FB3, 0x001FBC},
+{0x001FBE, 0x000399},
+{0x001FC3, 0x001FCC},
+{0x001FD0, 0x001FD8},
+{0x001FD1, 0x001FD9},
+{0x001FE0, 0x001FE8},
+{0x001FE1, 0x001FE9},
+{0x001FE5, 0x001FEC},
+{0x001FF3, 0x001FFC},
+{0x00214E, 0x002132},
+{0x002170, 0x002160},
+{0x002171, 0x002161},
+{0x002172, 0x002162},
+{0x002173, 0x002163},
+{0x002174, 0x002164},
+{0x002175, 0x002165},
+{0x002176, 0x002166},
+{0x002177, 0x002167},
+{0x002178, 0x002168},
+{0x002179, 0x002169},
+{0x00217A, 0x00216A},
+{0x00217B, 0x00216B},
+{0x00217C, 0x00216C},
+{0x00217D, 0x00216D},
+{0x00217E, 0x00216E},
+{0x00217F, 0x00216F},
+{0x002184, 0x002183},
+{0x0024D0, 0x0024B6},
+{0x0024D1, 0x0024B7},
+{0x0024D2, 0x0024B8},
+{0x0024D3, 0x0024B9},
+{0x0024D4, 0x0024BA},
+{0x0024D5, 0x0024BB},
+{0x0024D6, 0x0024BC},
+{0x0024D7, 0x0024BD},
+{0x0024D8, 0x0024BE},
+{0x0024D9, 0x0024BF},
+{0x0024DA, 0x0024C0},
+{0x0024DB, 0x0024C1},
+{0x0024DC, 0x0024C2},
+{0x0024DD, 0x0024C3},
+{0x0024DE, 0x0024C4},
+{0x0024DF, 0x0024C5},
+{0x0024E0, 0x0024C6},
+{0x0024E1, 0x0024C7},
+{0x0024E2, 0x0024C8},
+{0x0024E3, 0x0024C9},
+{0x0024E4, 0x0024CA},
+{0x0024E5, 0x0024CB},
+{0x0024E6, 0x0024CC},
+{0x0024E7, 0x0024CD},
+{0x0024E8, 0x0024CE},
+{0x0024E9, 0x0024CF},
+{0x002C30, 0x002C00},
+{0x002C31, 0x002C01},
+{0x002C32, 0x002C02},
+{0x002C33, 0x002C03},
+{0x002C34, 0x002C04},
+{0x002C35, 0x002C05},
+{0x002C36, 0x002C06},
+{0x002C37, 0x002C07},
+{0x002C38, 0x002C08},
+{0x002C39, 0x002C09},
+{0x002C3A, 0x002C0A},
+{0x002C3B, 0x002C0B},
+{0x002C3C, 0x002C0C},
+{0x002C3D, 0x002C0D},
+{0x002C3E, 0x002C0E},
+{0x002C3F, 0x002C0F},
+{0x002C40, 0x002C10},
+{0x002C41, 0x002C11},
+{0x002C42, 0x002C12},
+{0x002C43, 0x002C13},
+{0x002C44, 0x002C14},
+{0x002C45, 0x002C15},
+{0x002C46, 0x002C16},
+{0x002C47, 0x002C17},
+{0x002C48, 0x002C18},
+{0x002C49, 0x002C19},
+{0x002C4A, 0x002C1A},
+{0x002C4B, 0x002C1B},
+{0x002C4C, 0x002C1C},
+{0x002C4D, 0x002C1D},
+{0x002C4E, 0x002C1E},
+{0x002C4F, 0x002C1F},
+{0x002C50, 0x002C20},
+{0x002C51, 0x002C21},
+{0x002C52, 0x002C22},
+{0x002C53, 0x002C23},
+{0x002C54, 0x002C24},
+{0x002C55, 0x002C25},
+{0x002C56, 0x002C26},
+{0x002C57, 0x002C27},
+{0x002C58, 0x002C28},
+{0x002C59, 0x002C29},
+{0x002C5A, 0x002C2A},
+{0x002C5B, 0x002C2B},
+{0x002C5C, 0x002C2C},
+{0x002C5D, 0x002C2D},
+{0x002C5E, 0x002C2E},
+{0x002C5F, 0x002C2F},
+{0x002C61, 0x002C60},
+{0x002C65, 0x00023A},
+{0x002C66, 0x00023E},
+{0x002C68, 0x002C67},
+{0x002C6A, 0x002C69},
+{0x002C6C, 0x002C6B},
+{0x002C73, 0x002C72},
+{0x002C76, 0x002C75},
+{0x002C81, 0x002C80},
+{0x002C83, 0x002C82},
+{0x002C85, 0x002C84},
+{0x002C87, 0x002C86},
+{0x002C89, 0x002C88},
+{0x002C8B, 0x002C8A},
+{0x002C8D, 0x002C8C},
+{0x002C8F, 0x002C8E},
+{0x002C91, 0x002C90},
+{0x002C93, 0x002C92},
+{0x002C95, 0x002C94},
+{0x002C97, 0x002C96},
+{0x002C99, 0x002C98},
+{0x002C9B, 0x002C9A},
+{0x002C9D, 0x002C9C},
+{0x002C9F, 0x002C9E},
+{0x002CA1, 0x002CA0},
+{0x002CA3, 0x002CA2},
+{0x002CA5, 0x002CA4},
+{0x002CA7, 0x002CA6},
+{0x002CA9, 0x002CA8},
+{0x002CAB, 0x002CAA},
+{0x002CAD, 0x002CAC},
+{0x002CAF, 0x002CAE},
+{0x002CB1, 0x002CB0},
+{0x002CB3, 0x002CB2},
+{0x002CB5, 0x002CB4},
+{0x002CB7, 0x002CB6},
+{0x002CB9, 0x002CB8},
+{0x002CBB, 0x002CBA},
+{0x002CBD, 0x002CBC},
+{0x002CBF, 0x002CBE},
+{0x002CC1, 0x002CC0},
+{0x002CC3, 0x002CC2},
+{0x002CC5, 0x002CC4},
+{0x002CC7, 0x002CC6},
+{0x002CC9, 0x002CC8},
+{0x002CCB, 0x002CCA},
+{0x002CCD, 0x002CCC},
+{0x002CCF, 0x002CCE},
+{0x002CD1, 0x002CD0},
+{0x002CD3, 0x002CD2},
+{0x002CD5, 0x002CD4},
+{0x002CD7, 0x002CD6},
+{0x002CD9, 0x002CD8},
+{0x002CDB, 0x002CDA},
+{0x002CDD, 0x002CDC},
+{0x002CDF, 0x002CDE},
+{0x002CE1, 0x002CE0},
+{0x002CE3, 0x002CE2},
+{0x002CEC, 0x002CEB},
+{0x002CEE, 0x002CED},
+{0x002CF3, 0x002CF2},
+{0x002D00, 0x0010A0},
+{0x002D01, 0x0010A1},
+{0x002D02, 0x0010A2},
+{0x002D03, 0x0010A3},
+{0x002D04, 0x0010A4},
+{0x002D05, 0x0010A5},
+{0x002D06, 0x0010A6},
+{0x002D07, 0x0010A7},
+{0x002D08, 0x0010A8},
+{0x002D09, 0x0010A9},
+{0x002D0A, 0x0010AA},
+{0x002D0B, 0x0010AB},
+{0x002D0C, 0x0010AC},
+{0x002D0D, 0x0010AD},
+{0x002D0E, 0x0010AE},
+{0x002D0F, 0x0010AF},
+{0x002D10, 0x0010B0},
+{0x002D11, 0x0010B1},
+{0x002D12, 0x0010B2},
+{0x002D13, 0x0010B3},
+{0x002D14, 0x0010B4},
+{0x002D15, 0x0010B5},
+{0x002D16, 0x0010B6},
+{0x002D17, 0x0010B7},
+{0x002D18, 0x0010B8},
+{0x002D19, 0x0010B9},
+{0x002D1A, 0x0010BA},
+{0x002D1B, 0x0010BB},
+{0x002D1C, 0x0010BC},
+{0x002D1D, 0x0010BD},
+{0x002D1E, 0x0010BE},
+{0x002D1F, 0x0010BF},
+{0x002D20, 0x0010C0},
+{0x002D21, 0x0010C1},
+{0x002D22, 0x0010C2},
+{0x002D23, 0x0010C3},
+{0x002D24, 0x0010C4},
+{0x002D25, 0x0010C5},
+{0x002D27, 0x0010C7},
+{0x002D2D, 0x0010CD},
+{0x00A641, 0x00A640},
+{0x00A643, 0x00A642},
+{0x00A645, 0x00A644},
+{0x00A647, 0x00A646},
+{0x00A649, 0x00A648},
+{0x00A64B, 0x00A64A},
+{0x00A64D, 0x00A64C},
+{0x00A64F, 0x00A64E},
+{0x00A651, 0x00A650},
+{0x00A653, 0x00A652},
+{0x00A655, 0x00A654},
+{0x00A657, 0x00A656},
+{0x00A659, 0x00A658},
+{0x00A65B, 0x00A65A},
+{0x00A65D, 0x00A65C},
+{0x00A65F, 0x00A65E},
+{0x00A661, 0x00A660},
+{0x00A663, 0x00A662},
+{0x00A665, 0x00A664},
+{0x00A667, 0x00A666},
+{0x00A669, 0x00A668},
+{0x00A66B, 0x00A66A},
+{0x00A66D, 0x00A66C},
+{0x00A681, 0x00A680},
+{0x00A683, 0x00A682},
+{0x00A685, 0x00A684},
+{0x00A687, 0x00A686},
+{0x00A689, 0x00A688},
+{0x00A68B, 0x00A68A},
+{0x00A68D, 0x00A68C},
+{0x00A68F, 0x00A68E},
+{0x00A691, 0x00A690},
+{0x00A693, 0x00A692},
+{0x00A695, 0x00A694},
+{0x00A697, 0x00A696},
+{0x00A699, 0x00A698},
+{0x00A69B, 0x00A69A},
+{0x00A723, 0x00A722},
+{0x00A725, 0x00A724},
+{0x00A727, 0x00A726},
+{0x00A729, 0x00A728},
+{0x00A72B, 0x00A72A},
+{0x00A72D, 0x00A72C},
+{0x00A72F, 0x00A72E},
+{0x00A733, 0x00A732},
+{0x00A735, 0x00A734},
+{0x00A737, 0x00A736},
+{0x00A739, 0x00A738},
+{0x00A73B, 0x00A73A},
+{0x00A73D, 0x00A73C},
+{0x00A73F, 0x00A73E},
+{0x00A741, 0x00A740},
+{0x00A743, 0x00A742},
+{0x00A745, 0x00A744},
+{0x00A747, 0x00A746},
+{0x00A749, 0x00A748},
+{0x00A74B, 0x00A74A},
+{0x00A74D, 0x00A74C},
+{0x00A74F, 0x00A74E},
+{0x00A751, 0x00A750},
+{0x00A753, 0x00A752},
+{0x00A755, 0x00A754},
+{0x00A757, 0x00A756},
+{0x00A759, 0x00A758},
+{0x00A75B, 0x00A75A},
+{0x00A75D, 0x00A75C},
+{0x00A75F, 0x00A75E},
+{0x00A761, 0x00A760},
+{0x00A763, 0x00A762},
+{0x00A765, 0x00A764},
+{0x00A767, 0x00A766},
+{0x00A769, 0x00A768},
+{0x00A76B, 0x00A76A},
+{0x00A76D, 0x00A76C},
+{0x00A76F, 0x00A76E},
+{0x00A77A, 0x00A779},
+{0x00A77C, 0x00A77B},
+{0x00A77F, 0x00A77E},
+{0x00A781, 0x00A780},
+{0x00A783, 0x00A782},
+{0x00A785, 0x00A784},
+{0x00A787, 0x00A786},
+{0x00A78C, 0x00A78B},
+{0x00A791, 0x00A790},
+{0x00A793, 0x00A792},
+{0x00A794, 0x00A7C4},
+{0x00A797, 0x00A796},
+{0x00A799, 0x00A798},
+{0x00A79B, 0x00A79A},
+{0x00A79D, 0x00A79C},
+{0x00A79F, 0x00A79E},
+{0x00A7A1, 0x00A7A0},
+{0x00A7A3, 0x00A7A2},
+{0x00A7A5, 0x00A7A4},
+{0x00A7A7, 0x00A7A6},
+{0x00A7A9, 0x00A7A8},
+{0x00A7B5, 0x00A7B4},
+{0x00A7B7, 0x00A7B6},
+{0x00A7B9, 0x00A7B8},
+{0x00A7BB, 0x00A7BA},
+{0x00A7BD, 0x00A7BC},
+{0x00A7BF, 0x00A7BE},
+{0x00A7C1, 0x00A7C0},
+{0x00A7C3, 0x00A7C2},
+{0x00A7C8, 0x00A7C7},
+{0x00A7CA, 0x00A7C9},
+{0x00A7D1, 0x00A7D0},
+{0x00A7D7, 0x00A7D6},
+{0x00A7D9, 0x00A7D8},
+{0x00A7F6, 0x00A7F5},
+{0x00AB53, 0x00A7B3},
+{0x00AB70, 0x0013A0},
+{0x00AB71, 0x0013A1},
+{0x00AB72, 0x0013A2},
+{0x00AB73, 0x0013A3},
+{0x00AB74, 0x0013A4},
+{0x00AB75, 0x0013A5},
+{0x00AB76, 0x0013A6},
+{0x00AB77, 0x0013A7},
+{0x00AB78, 0x0013A8},
+{0x00AB79, 0x0013A9},
+{0x00AB7A, 0x0013AA},
+{0x00AB7B, 0x0013AB},
+{0x00AB7C, 0x0013AC},
+{0x00AB7D, 0x0013AD},
+{0x00AB7E, 0x0013AE},
+{0x00AB7F, 0x0013AF},
+{0x00AB80, 0x0013B0},
+{0x00AB81, 0x0013B1},
+{0x00AB82, 0x0013B2},
+{0x00AB83, 0x0013B3},
+{0x00AB84, 0x0013B4},
+{0x00AB85, 0x0013B5},
+{0x00AB86, 0x0013B6},
+{0x00AB87, 0x0013B7},
+{0x00AB88, 0x0013B8},
+{0x00AB89, 0x0013B9},
+{0x00AB8A, 0x0013BA},
+{0x00AB8B, 0x0013BB},
+{0x00AB8C, 0x0013BC},
+{0x00AB8D, 0x0013BD},
+{0x00AB8E, 0x0013BE},
+{0x00AB8F, 0x0013BF},
+{0x00AB90, 0x0013C0},
+{0x00AB91, 0x0013C1},
+{0x00AB92, 0x0013C2},
+{0x00AB93, 0x0013C3},
+{0x00AB94, 0x0013C4},
+{0x00AB95, 0x0013C5},
+{0x00AB96, 0x0013C6},
+{0x00AB97, 0x0013C7},
+{0x00AB98, 0x0013C8},
+{0x00AB99, 0x0013C9},
+{0x00AB9A, 0x0013CA},
+{0x00AB9B, 0x0013CB},
+{0x00AB9C, 0x0013CC},
+{0x00AB9D, 0x0013CD},
+{0x00AB9E, 0x0013CE},
+{0x00AB9F, 0x0013CF},
+{0x00ABA0, 0x0013D0},
+{0x00ABA1, 0x0013D1},
+{0x00ABA2, 0x0013D2},
+{0x00ABA3, 0x0013D3},
+{0x00ABA4, 0x0013D4},
+{0x00ABA5, 0x0013D5},
+{0x00ABA6, 0x0013D6},
+{0x00ABA7, 0x0013D7},
+{0x00ABA8, 0x0013D8},
+{0x00ABA9, 0x0013D9},
+{0x00ABAA, 0x0013DA},
+{0x00ABAB, 0x0013DB},
+{0x00ABAC, 0x0013DC},
+{0x00ABAD, 0x0013DD},
+{0x00ABAE, 0x0013DE},
+{0x00ABAF, 0x0013DF},
+{0x00ABB0, 0x0013E0},
+{0x00ABB1, 0x0013E1},
+{0x00ABB2, 0x0013E2},
+{0x00ABB3, 0x0013E3},
+{0x00ABB4, 0x0013E4},
+{0x00ABB5, 0x0013E5},
+{0x00ABB6, 0x0013E6},
+{0x00ABB7, 0x0013E7},
+{0x00ABB8, 0x0013E8},
+{0x00ABB9, 0x0013E9},
+{0x00ABBA, 0x0013EA},
+{0x00ABBB, 0x0013EB},
+{0x00ABBC, 0x0013EC},
+{0x00ABBD, 0x0013ED},
+{0x00ABBE, 0x0013EE},
+{0x00ABBF, 0x0013EF},
+{0x00FF41, 0x00FF21},
+{0x00FF42, 0x00FF22},
+{0x00FF43, 0x00FF23},
+{0x00FF44, 0x00FF24},
+{0x00FF45, 0x00FF25},
+{0x00FF46, 0x00FF26},
+{0x00FF47, 0x00FF27},
+{0x00FF48, 0x00FF28},
+{0x00FF49, 0x00FF29},
+{0x00FF4A, 0x00FF2A},
+{0x00FF4B, 0x00FF2B},
+{0x00FF4C, 0x00FF2C},
+{0x00FF4D, 0x00FF2D},
+{0x00FF4E, 0x00FF2E},
+{0x00FF4F, 0x00FF2F},
+{0x00FF50, 0x00FF30},
+{0x00FF51, 0x00FF31},
+{0x00FF52, 0x00FF32},
+{0x00FF53, 0x00FF33},
+{0x00FF54, 0x00FF34},
+{0x00FF55, 0x00FF35},
+{0x00FF56, 0x00FF36},
+{0x00FF57, 0x00FF37},
+{0x00FF58, 0x00FF38},
+{0x00FF59, 0x00FF39},
+{0x00FF5A, 0x00FF3A},
+{0x010428, 0x010400},
+{0x010429, 0x010401},
+{0x01042A, 0x010402},
+{0x01042B, 0x010403},
+{0x01042C, 0x010404},
+{0x01042D, 0x010405},
+{0x01042E, 0x010406},
+{0x01042F, 0x010407},
+{0x010430, 0x010408},
+{0x010431, 0x010409},
+{0x010432, 0x01040A},
+{0x010433, 0x01040B},
+{0x010434, 0x01040C},
+{0x010435, 0x01040D},
+{0x010436, 0x01040E},
+{0x010437, 0x01040F},
+{0x010438, 0x010410},
+{0x010439, 0x010411},
+{0x01043A, 0x010412},
+{0x01043B, 0x010413},
+{0x01043C, 0x010414},
+{0x01043D, 0x010415},
+{0x01043E, 0x010416},
+{0x01043F, 0x010417},
+{0x010440, 0x010418},
+{0x010441, 0x010419},
+{0x010442, 0x01041A},
+{0x010443, 0x01041B},
+{0x010444, 0x01041C},
+{0x010445, 0x01041D},
+{0x010446, 0x01041E},
+{0x010447, 0x01041F},
+{0x010448, 0x010420},
+{0x010449, 0x010421},
+{0x01044A, 0x010422},
+{0x01044B, 0x010423},
+{0x01044C, 0x010424},
+{0x01044D, 0x010425},
+{0x01044E, 0x010426},
+{0x01044F, 0x010427},
+{0x0104D8, 0x0104B0},
+{0x0104D9, 0x0104B1},
+{0x0104DA, 0x0104B2},
+{0x0104DB, 0x0104B3},
+{0x0104DC, 0x0104B4},
+{0x0104DD, 0x0104B5},
+{0x0104DE, 0x0104B6},
+{0x0104DF, 0x0104B7},
+{0x0104E0, 0x0104B8},
+{0x0104E1, 0x0104B9},
+{0x0104E2, 0x0104BA},
+{0x0104E3, 0x0104BB},
+{0x0104E4, 0x0104BC},
+{0x0104E5, 0x0104BD},
+{0x0104E6, 0x0104BE},
+{0x0104E7, 0x0104BF},
+{0x0104E8, 0x0104C0},
+{0x0104E9, 0x0104C1},
+{0x0104EA, 0x0104C2},
+{0x0104EB, 0x0104C3},
+{0x0104EC, 0x0104C4},
+{0x0104ED, 0x0104C5},
+{0x0104EE, 0x0104C6},
+{0x0104EF, 0x0104C7},
+{0x0104F0, 0x0104C8},
+{0x0104F1, 0x0104C9},
+{0x0104F2, 0x0104CA},
+{0x0104F3, 0x0104CB},
+{0x0104F4, 0x0104CC},
+{0x0104F5, 0x0104CD},
+{0x0104F6, 0x0104CE},
+{0x0104F7, 0x0104CF},
+{0x0104F8, 0x0104D0},
+{0x0104F9, 0x0104D1},
+{0x0104FA, 0x0104D2},
+{0x0104FB, 0x0104D3},
+{0x010597, 0x010570},
+{0x010598, 0x010571},
+{0x010599, 0x010572},
+{0x01059A, 0x010573},
+{0x01059B, 0x010574},
+{0x01059C, 0x010575},
+{0x01059D, 0x010576},
+{0x01059E, 0x010577},
+{0x01059F, 0x010578},
+{0x0105A0, 0x010579},
+{0x0105A1, 0x01057A},
+{0x0105A3, 0x01057C},
+{0x0105A4, 0x01057D},
+{0x0105A5, 0x01057E},
+{0x0105A6, 0x01057F},
+{0x0105A7, 0x010580},
+{0x0105A8, 0x010581},
+{0x0105A9, 0x010582},
+{0x0105AA, 0x010583},
+{0x0105AB, 0x010584},
+{0x0105AC, 0x010585},
+{0x0105AD, 0x010586},
+{0x0105AE, 0x010587},
+{0x0105AF, 0x010588},
+{0x0105B0, 0x010589},
+{0x0105B1, 0x01058A},
+{0x0105B3, 0x01058C},
+{0x0105B4, 0x01058D},
+{0x0105B5, 0x01058E},
+{0x0105B6, 0x01058F},
+{0x0105B7, 0x010590},
+{0x0105B8, 0x010591},
+{0x0105B9, 0x010592},
+{0x0105BB, 0x010594},
+{0x0105BC, 0x010595},
+{0x010CC0, 0x010C80},
+{0x010CC1, 0x010C81},
+{0x010CC2, 0x010C82},
+{0x010CC3, 0x010C83},
+{0x010CC4, 0x010C84},
+{0x010CC5, 0x010C85},
+{0x010CC6, 0x010C86},
+{0x010CC7, 0x010C87},
+{0x010CC8, 0x010C88},
+{0x010CC9, 0x010C89},
+{0x010CCA, 0x010C8A},
+{0x010CCB, 0x010C8B},
+{0x010CCC, 0x010C8C},
+{0x010CCD, 0x010C8D},
+{0x010CCE, 0x010C8E},
+{0x010CCF, 0x010C8F},
+{0x010CD0, 0x010C90},
+{0x010CD1, 0x010C91},
+{0x010CD2, 0x010C92},
+{0x010CD3, 0x010C93},
+{0x010CD4, 0x010C94},
+{0x010CD5, 0x010C95},
+{0x010CD6, 0x010C96},
+{0x010CD7, 0x010C97},
+{0x010CD8, 0x010C98},
+{0x010CD9, 0x010C99},
+{0x010CDA, 0x010C9A},
+{0x010CDB, 0x010C9B},
+{0x010CDC, 0x010C9C},
+{0x010CDD, 0x010C9D},
+{0x010CDE, 0x010C9E},
+{0x010CDF, 0x010C9F},
+{0x010CE0, 0x010CA0},
+{0x010CE1, 0x010CA1},
+{0x010CE2, 0x010CA2},
+{0x010CE3, 0x010CA3},
+{0x010CE4, 0x010CA4},
+{0x010CE5, 0x010CA5},
+{0x010CE6, 0x010CA6},
+{0x010CE7, 0x010CA7},
+{0x010CE8, 0x010CA8},
+{0x010CE9, 0x010CA9},
+{0x010CEA, 0x010CAA},
+{0x010CEB, 0x010CAB},
+{0x010CEC, 0x010CAC},
+{0x010CED, 0x010CAD},
+{0x010CEE, 0x010CAE},
+{0x010CEF, 0x010CAF},
+{0x010CF0, 0x010CB0},
+{0x010CF1, 0x010CB1},
+{0x010CF2, 0x010CB2},
+{0x0118C0, 0x0118A0},
+{0x0118C1, 0x0118A1},
+{0x0118C2, 0x0118A2},
+{0x0118C3, 0x0118A3},
+{0x0118C4, 0x0118A4},
+{0x0118C5, 0x0118A5},
+{0x0118C6, 0x0118A6},
+{0x0118C7, 0x0118A7},
+{0x0118C8, 0x0118A8},
+{0x0118C9, 0x0118A9},
+{0x0118CA, 0x0118AA},
+{0x0118CB, 0x0118AB},
+{0x0118CC, 0x0118AC},
+{0x0118CD, 0x0118AD},
+{0x0118CE, 0x0118AE},
+{0x0118CF, 0x0118AF},
+{0x0118D0, 0x0118B0},
+{0x0118D1, 0x0118B1},
+{0x0118D2, 0x0118B2},
+{0x0118D3, 0x0118B3},
+{0x0118D4, 0x0118B4},
+{0x0118D5, 0x0118B5},
+{0x0118D6, 0x0118B6},
+{0x0118D7, 0x0118B7},
+{0x0118D8, 0x0118B8},
+{0x0118D9, 0x0118B9},
+{0x0118DA, 0x0118BA},
+{0x0118DB, 0x0118BB},
+{0x0118DC, 0x0118BC},
+{0x0118DD, 0x0118BD},
+{0x0118DE, 0x0118BE},
+{0x0118DF, 0x0118BF},
+{0x016E60, 0x016E40},
+{0x016E61, 0x016E41},
+{0x016E62, 0x016E42},
+{0x016E63, 0x016E43},
+{0x016E64, 0x016E44},
+{0x016E65, 0x016E45},
+{0x016E66, 0x016E46},
+{0x016E67, 0x016E47},
+{0x016E68, 0x016E48},
+{0x016E69, 0x016E49},
+{0x016E6A, 0x016E4A},
+{0x016E6B, 0x016E4B},
+{0x016E6C, 0x016E4C},
+{0x016E6D, 0x016E4D},
+{0x016E6E, 0x016E4E},
+{0x016E6F, 0x016E4F},
+{0x016E70, 0x016E50},
+{0x016E71, 0x016E51},
+{0x016E72, 0x016E52},
+{0x016E73, 0x016E53},
+{0x016E74, 0x016E54},
+{0x016E75, 0x016E55},
+{0x016E76, 0x016E56},
+{0x016E77, 0x016E57},
+{0x016E78, 0x016E58},
+{0x016E79, 0x016E59},
+{0x016E7A, 0x016E5A},
+{0x016E7B, 0x016E5B},
+{0x016E7C, 0x016E5C},
+{0x016E7D, 0x016E5D},
+{0x016E7E, 0x016E5E},
+{0x016E7F, 0x016E5F},
+{0x01E922, 0x01E900},
+{0x01E923, 0x01E901},
+{0x01E924, 0x01E902},
+{0x01E925, 0x01E903},
+{0x01E926, 0x01E904},
+{0x01E927, 0x01E905},
+{0x01E928, 0x01E906},
+{0x01E929, 0x01E907},
+{0x01E92A, 0x01E908},
+{0x01E92B, 0x01E909},
+{0x01E92C, 0x01E90A},
+{0x01E92D, 0x01E90B},
+{0x01E92E, 0x01E90C},
+{0x01E92F, 0x01E90D},
+{0x01E930, 0x01E90E},
+{0x01E931, 0x01E90F},
+{0x01E932, 0x01E910},
+{0x01E933, 0x01E911},
+{0x01E934, 0x01E912},
+{0x01E935, 0x01E913},
+{0x01E936, 0x01E914},
+{0x01E937, 0x01E915},
+{0x01E938, 0x01E916},
+{0x01E939, 0x01E917},
+{0x01E93A, 0x01E918},
+{0x01E93B, 0x01E919},
+{0x01E93C, 0x01E91A},
+{0x01E93D, 0x01E91B},
+{0x01E93E, 0x01E91C},
+{0x01E93F, 0x01E91D},
+{0x01E940, 0x01E91E},
+{0x01E941, 0x01E91F},
+{0x01E942, 0x01E920},
+{0x01E943, 0x01E921},
+};
+
+const std::vector unicode_ranges_nfd = {  // start, last, nfd
+{0x000000, 0x000000, 0x000000},
+{0x0000C0, 0x0000C5, 0x000041},
+{0x0000C7, 0x0000C7, 0x000043},
+{0x0000C8, 0x0000CB, 0x000045},
+{0x0000CC, 0x0000CF, 0x000049},
+{0x0000D1, 0x0000D1, 0x00004E},
+{0x0000D2, 0x0000D6, 0x00004F},
+{0x0000D9, 0x0000DC, 0x000055},
+{0x0000DD, 0x0000DD, 0x000059},
+{0x0000E0, 0x0000E5, 0x000061},
+{0x0000E7, 0x0000E7, 0x000063},
+{0x0000E8, 0x0000EB, 0x000065},
+{0x0000EC, 0x0000EF, 0x000069},
+{0x0000F1, 0x0000F1, 0x00006E},
+{0x0000F2, 0x0000F6, 0x00006F},
+{0x0000F9, 0x0000FC, 0x000075},
+{0x0000FD, 0x0000FD, 0x000079},
+{0x0000FF, 0x0000FF, 0x000079},
+{0x000100, 0x000100, 0x000041},
+{0x000101, 0x000101, 0x000061},
+{0x000102, 0x000102, 0x000041},
+{0x000103, 0x000103, 0x000061},
+{0x000104, 0x000104, 0x000041},
+{0x000105, 0x000105, 0x000061},
+{0x000106, 0x000106, 0x000043},
+{0x000107, 0x000107, 0x000063},
+{0x000108, 0x000108, 0x000043},
+{0x000109, 0x000109, 0x000063},
+{0x00010A, 0x00010A, 0x000043},
+{0x00010B, 0x00010B, 0x000063},
+{0x00010C, 0x00010C, 0x000043},
+{0x00010D, 0x00010D, 0x000063},
+{0x00010E, 0x00010E, 0x000044},
+{0x00010F, 0x00010F, 0x000064},
+{0x000112, 0x000112, 0x000045},
+{0x000113, 0x000113, 0x000065},
+{0x000114, 0x000114, 0x000045},
+{0x000115, 0x000115, 0x000065},
+{0x000116, 0x000116, 0x000045},
+{0x000117, 0x000117, 0x000065},
+{0x000118, 0x000118, 0x000045},
+{0x000119, 0x000119, 0x000065},
+{0x00011A, 0x00011A, 0x000045},
+{0x00011B, 0x00011B, 0x000065},
+{0x00011C, 0x00011C, 0x000047},
+{0x00011D, 0x00011D, 0x000067},
+{0x00011E, 0x00011E, 0x000047},
+{0x00011F, 0x00011F, 0x000067},
+{0x000120, 0x000120, 0x000047},
+{0x000121, 0x000121, 0x000067},
+{0x000122, 0x000122, 0x000047},
+{0x000123, 0x000123, 0x000067},
+{0x000124, 0x000124, 0x000048},
+{0x000125, 0x000125, 0x000068},
+{0x000128, 0x000128, 0x000049},
+{0x000129, 0x000129, 0x000069},
+{0x00012A, 0x00012A, 0x000049},
+{0x00012B, 0x00012B, 0x000069},
+{0x00012C, 0x00012C, 0x000049},
+{0x00012D, 0x00012D, 0x000069},
+{0x00012E, 0x00012E, 0x000049},
+{0x00012F, 0x00012F, 0x000069},
+{0x000130, 0x000130, 0x000049},
+{0x000134, 0x000134, 0x00004A},
+{0x000135, 0x000135, 0x00006A},
+{0x000136, 0x000136, 0x00004B},
+{0x000137, 0x000137, 0x00006B},
+{0x000139, 0x000139, 0x00004C},
+{0x00013A, 0x00013A, 0x00006C},
+{0x00013B, 0x00013B, 0x00004C},
+{0x00013C, 0x00013C, 0x00006C},
+{0x00013D, 0x00013D, 0x00004C},
+{0x00013E, 0x00013E, 0x00006C},
+{0x000143, 0x000143, 0x00004E},
+{0x000144, 0x000144, 0x00006E},
+{0x000145, 0x000145, 0x00004E},
+{0x000146, 0x000146, 0x00006E},
+{0x000147, 0x000147, 0x00004E},
+{0x000148, 0x000148, 0x00006E},
+{0x00014C, 0x00014C, 0x00004F},
+{0x00014D, 0x00014D, 0x00006F},
+{0x00014E, 0x00014E, 0x00004F},
+{0x00014F, 0x00014F, 0x00006F},
+{0x000150, 0x000150, 0x00004F},
+{0x000151, 0x000151, 0x00006F},
+{0x000154, 0x000154, 0x000052},
+{0x000155, 0x000155, 0x000072},
+{0x000156, 0x000156, 0x000052},
+{0x000157, 0x000157, 0x000072},
+{0x000158, 0x000158, 0x000052},
+{0x000159, 0x000159, 0x000072},
+{0x00015A, 0x00015A, 0x000053},
+{0x00015B, 0x00015B, 0x000073},
+{0x00015C, 0x00015C, 0x000053},
+{0x00015D, 0x00015D, 0x000073},
+{0x00015E, 0x00015E, 0x000053},
+{0x00015F, 0x00015F, 0x000073},
+{0x000160, 0x000160, 0x000053},
+{0x000161, 0x000161, 0x000073},
+{0x000162, 0x000162, 0x000054},
+{0x000163, 0x000163, 0x000074},
+{0x000164, 0x000164, 0x000054},
+{0x000165, 0x000165, 0x000074},
+{0x000168, 0x000168, 0x000055},
+{0x000169, 0x000169, 0x000075},
+{0x00016A, 0x00016A, 0x000055},
+{0x00016B, 0x00016B, 0x000075},
+{0x00016C, 0x00016C, 0x000055},
+{0x00016D, 0x00016D, 0x000075},
+{0x00016E, 0x00016E, 0x000055},
+{0x00016F, 0x00016F, 0x000075},
+{0x000170, 0x000170, 0x000055},
+{0x000171, 0x000171, 0x000075},
+{0x000172, 0x000172, 0x000055},
+{0x000173, 0x000173, 0x000075},
+{0x000174, 0x000174, 0x000057},
+{0x000175, 0x000175, 0x000077},
+{0x000176, 0x000176, 0x000059},
+{0x000177, 0x000177, 0x000079},
+{0x000178, 0x000178, 0x000059},
+{0x000179, 0x000179, 0x00005A},
+{0x00017A, 0x00017A, 0x00007A},
+{0x00017B, 0x00017B, 0x00005A},
+{0x00017C, 0x00017C, 0x00007A},
+{0x00017D, 0x00017D, 0x00005A},
+{0x00017E, 0x00017E, 0x00007A},
+{0x0001A0, 0x0001A0, 0x00004F},
+{0x0001A1, 0x0001A1, 0x00006F},
+{0x0001AF, 0x0001AF, 0x000055},
+{0x0001B0, 0x0001B0, 0x000075},
+{0x0001CD, 0x0001CD, 0x000041},
+{0x0001CE, 0x0001CE, 0x000061},
+{0x0001CF, 0x0001CF, 0x000049},
+{0x0001D0, 0x0001D0, 0x000069},
+{0x0001D1, 0x0001D1, 0x00004F},
+{0x0001D2, 0x0001D2, 0x00006F},
+{0x0001D3, 0x0001D3, 0x000055},
+{0x0001D4, 0x0001D4, 0x000075},
+{0x0001D5, 0x0001D5, 0x000055},
+{0x0001D6, 0x0001D6, 0x000075},
+{0x0001D7, 0x0001D7, 0x000055},
+{0x0001D8, 0x0001D8, 0x000075},
+{0x0001D9, 0x0001D9, 0x000055},
+{0x0001DA, 0x0001DA, 0x000075},
+{0x0001DB, 0x0001DB, 0x000055},
+{0x0001DC, 0x0001DC, 0x000075},
+{0x0001DE, 0x0001DE, 0x000041},
+{0x0001DF, 0x0001DF, 0x000061},
+{0x0001E0, 0x0001E0, 0x000041},
+{0x0001E1, 0x0001E1, 0x000061},
+{0x0001E2, 0x0001E2, 0x0000C6},
+{0x0001E3, 0x0001E3, 0x0000E6},
+{0x0001E6, 0x0001E6, 0x000047},
+{0x0001E7, 0x0001E7, 0x000067},
+{0x0001E8, 0x0001E8, 0x00004B},
+{0x0001E9, 0x0001E9, 0x00006B},
+{0x0001EA, 0x0001EA, 0x00004F},
+{0x0001EB, 0x0001EB, 0x00006F},
+{0x0001EC, 0x0001EC, 0x00004F},
+{0x0001ED, 0x0001ED, 0x00006F},
+{0x0001EE, 0x0001EE, 0x0001B7},
+{0x0001EF, 0x0001EF, 0x000292},
+{0x0001F0, 0x0001F0, 0x00006A},
+{0x0001F4, 0x0001F4, 0x000047},
+{0x0001F5, 0x0001F5, 0x000067},
+{0x0001F8, 0x0001F8, 0x00004E},
+{0x0001F9, 0x0001F9, 0x00006E},
+{0x0001FA, 0x0001FA, 0x000041},
+{0x0001FB, 0x0001FB, 0x000061},
+{0x0001FC, 0x0001FC, 0x0000C6},
+{0x0001FD, 0x0001FD, 0x0000E6},
+{0x0001FE, 0x0001FE, 0x0000D8},
+{0x0001FF, 0x0001FF, 0x0000F8},
+{0x000200, 0x000200, 0x000041},
+{0x000201, 0x000201, 0x000061},
+{0x000202, 0x000202, 0x000041},
+{0x000203, 0x000203, 0x000061},
+{0x000204, 0x000204, 0x000045},
+{0x000205, 0x000205, 0x000065},
+{0x000206, 0x000206, 0x000045},
+{0x000207, 0x000207, 0x000065},
+{0x000208, 0x000208, 0x000049},
+{0x000209, 0x000209, 0x000069},
+{0x00020A, 0x00020A, 0x000049},
+{0x00020B, 0x00020B, 0x000069},
+{0x00020C, 0x00020C, 0x00004F},
+{0x00020D, 0x00020D, 0x00006F},
+{0x00020E, 0x00020E, 0x00004F},
+{0x00020F, 0x00020F, 0x00006F},
+{0x000210, 0x000210, 0x000052},
+{0x000211, 0x000211, 0x000072},
+{0x000212, 0x000212, 0x000052},
+{0x000213, 0x000213, 0x000072},
+{0x000214, 0x000214, 0x000055},
+{0x000215, 0x000215, 0x000075},
+{0x000216, 0x000216, 0x000055},
+{0x000217, 0x000217, 0x000075},
+{0x000218, 0x000218, 0x000053},
+{0x000219, 0x000219, 0x000073},
+{0x00021A, 0x00021A, 0x000054},
+{0x00021B, 0x00021B, 0x000074},
+{0x00021E, 0x00021E, 0x000048},
+{0x00021F, 0x00021F, 0x000068},
+{0x000226, 0x000226, 0x000041},
+{0x000227, 0x000227, 0x000061},
+{0x000228, 0x000228, 0x000045},
+{0x000229, 0x000229, 0x000065},
+{0x00022A, 0x00022A, 0x00004F},
+{0x00022B, 0x00022B, 0x00006F},
+{0x00022C, 0x00022C, 0x00004F},
+{0x00022D, 0x00022D, 0x00006F},
+{0x00022E, 0x00022E, 0x00004F},
+{0x00022F, 0x00022F, 0x00006F},
+{0x000230, 0x000230, 0x00004F},
+{0x000231, 0x000231, 0x00006F},
+{0x000232, 0x000232, 0x000059},
+{0x000233, 0x000233, 0x000079},
+{0x000340, 0x000340, 0x000300},
+{0x000341, 0x000341, 0x000301},
+{0x000343, 0x000343, 0x000313},
+{0x000344, 0x000344, 0x000308},
+{0x000374, 0x000374, 0x0002B9},
+{0x00037E, 0x00037E, 0x00003B},
+{0x000385, 0x000385, 0x0000A8},
+{0x000386, 0x000386, 0x000391},
+{0x000387, 0x000387, 0x0000B7},
+{0x000388, 0x000388, 0x000395},
+{0x000389, 0x000389, 0x000397},
+{0x00038A, 0x00038A, 0x000399},
+{0x00038C, 0x00038C, 0x00039F},
+{0x00038E, 0x00038E, 0x0003A5},
+{0x00038F, 0x00038F, 0x0003A9},
+{0x000390, 0x000390, 0x0003B9},
+{0x0003AA, 0x0003AA, 0x000399},
+{0x0003AB, 0x0003AB, 0x0003A5},
+{0x0003AC, 0x0003AC, 0x0003B1},
+{0x0003AD, 0x0003AD, 0x0003B5},
+{0x0003AE, 0x0003AE, 0x0003B7},
+{0x0003AF, 0x0003AF, 0x0003B9},
+{0x0003B0, 0x0003B0, 0x0003C5},
+{0x0003CA, 0x0003CA, 0x0003B9},
+{0x0003CB, 0x0003CB, 0x0003C5},
+{0x0003CC, 0x0003CC, 0x0003BF},
+{0x0003CD, 0x0003CD, 0x0003C5},
+{0x0003CE, 0x0003CE, 0x0003C9},
+{0x0003D3, 0x0003D4, 0x0003D2},
+{0x000400, 0x000401, 0x000415},
+{0x000403, 0x000403, 0x000413},
+{0x000407, 0x000407, 0x000406},
+{0x00040C, 0x00040C, 0x00041A},
+{0x00040D, 0x00040D, 0x000418},
+{0x00040E, 0x00040E, 0x000423},
+{0x000419, 0x000419, 0x000418},
+{0x000439, 0x000439, 0x000438},
+{0x000450, 0x000451, 0x000435},
+{0x000453, 0x000453, 0x000433},
+{0x000457, 0x000457, 0x000456},
+{0x00045C, 0x00045C, 0x00043A},
+{0x00045D, 0x00045D, 0x000438},
+{0x00045E, 0x00045E, 0x000443},
+{0x000476, 0x000476, 0x000474},
+{0x000477, 0x000477, 0x000475},
+{0x0004C1, 0x0004C1, 0x000416},
+{0x0004C2, 0x0004C2, 0x000436},
+{0x0004D0, 0x0004D0, 0x000410},
+{0x0004D1, 0x0004D1, 0x000430},
+{0x0004D2, 0x0004D2, 0x000410},
+{0x0004D3, 0x0004D3, 0x000430},
+{0x0004D6, 0x0004D6, 0x000415},
+{0x0004D7, 0x0004D7, 0x000435},
+{0x0004DA, 0x0004DA, 0x0004D8},
+{0x0004DB, 0x0004DB, 0x0004D9},
+{0x0004DC, 0x0004DC, 0x000416},
+{0x0004DD, 0x0004DD, 0x000436},
+{0x0004DE, 0x0004DE, 0x000417},
+{0x0004DF, 0x0004DF, 0x000437},
+{0x0004E2, 0x0004E2, 0x000418},
+{0x0004E3, 0x0004E3, 0x000438},
+{0x0004E4, 0x0004E4, 0x000418},
+{0x0004E5, 0x0004E5, 0x000438},
+{0x0004E6, 0x0004E6, 0x00041E},
+{0x0004E7, 0x0004E7, 0x00043E},
+{0x0004EA, 0x0004EA, 0x0004E8},
+{0x0004EB, 0x0004EB, 0x0004E9},
+{0x0004EC, 0x0004EC, 0x00042D},
+{0x0004ED, 0x0004ED, 0x00044D},
+{0x0004EE, 0x0004EE, 0x000423},
+{0x0004EF, 0x0004EF, 0x000443},
+{0x0004F0, 0x0004F0, 0x000423},
+{0x0004F1, 0x0004F1, 0x000443},
+{0x0004F2, 0x0004F2, 0x000423},
+{0x0004F3, 0x0004F3, 0x000443},
+{0x0004F4, 0x0004F4, 0x000427},
+{0x0004F5, 0x0004F5, 0x000447},
+{0x0004F8, 0x0004F8, 0x00042B},
+{0x0004F9, 0x0004F9, 0x00044B},
+{0x000622, 0x000623, 0x000627},
+{0x000624, 0x000624, 0x000648},
+{0x000625, 0x000625, 0x000627},
+{0x000626, 0x000626, 0x00064A},
+{0x0006C0, 0x0006C0, 0x0006D5},
+{0x0006C2, 0x0006C2, 0x0006C1},
+{0x0006D3, 0x0006D3, 0x0006D2},
+{0x000929, 0x000929, 0x000928},
+{0x000931, 0x000931, 0x000930},
+{0x000934, 0x000934, 0x000933},
+{0x000958, 0x000958, 0x000915},
+{0x000959, 0x000959, 0x000916},
+{0x00095A, 0x00095A, 0x000917},
+{0x00095B, 0x00095B, 0x00091C},
+{0x00095C, 0x00095C, 0x000921},
+{0x00095D, 0x00095D, 0x000922},
+{0x00095E, 0x00095E, 0x00092B},
+{0x00095F, 0x00095F, 0x00092F},
+{0x0009CB, 0x0009CC, 0x0009C7},
+{0x0009DC, 0x0009DC, 0x0009A1},
+{0x0009DD, 0x0009DD, 0x0009A2},
+{0x0009DF, 0x0009DF, 0x0009AF},
+{0x000A33, 0x000A33, 0x000A32},
+{0x000A36, 0x000A36, 0x000A38},
+{0x000A59, 0x000A59, 0x000A16},
+{0x000A5A, 0x000A5A, 0x000A17},
+{0x000A5B, 0x000A5B, 0x000A1C},
+{0x000A5E, 0x000A5E, 0x000A2B},
+{0x000B48, 0x000B48, 0x000B47},
+{0x000B4B, 0x000B4C, 0x000B47},
+{0x000B5C, 0x000B5C, 0x000B21},
+{0x000B5D, 0x000B5D, 0x000B22},
+{0x000B94, 0x000B94, 0x000B92},
+{0x000BCA, 0x000BCA, 0x000BC6},
+{0x000BCB, 0x000BCB, 0x000BC7},
+{0x000BCC, 0x000BCC, 0x000BC6},
+{0x000C48, 0x000C48, 0x000C46},
+{0x000CC0, 0x000CC0, 0x000CBF},
+{0x000CC7, 0x000CC8, 0x000CC6},
+{0x000CCA, 0x000CCB, 0x000CC6},
+{0x000D4A, 0x000D4A, 0x000D46},
+{0x000D4B, 0x000D4B, 0x000D47},
+{0x000D4C, 0x000D4C, 0x000D46},
+{0x000DDA, 0x000DDA, 0x000DD9},
+{0x000DDC, 0x000DDE, 0x000DD9},
+{0x000F43, 0x000F43, 0x000F42},
+{0x000F4D, 0x000F4D, 0x000F4C},
+{0x000F52, 0x000F52, 0x000F51},
+{0x000F57, 0x000F57, 0x000F56},
+{0x000F5C, 0x000F5C, 0x000F5B},
+{0x000F69, 0x000F69, 0x000F40},
+{0x000F73, 0x000F73, 0x000F71},
+{0x000F75, 0x000F75, 0x000F71},
+{0x000F76, 0x000F76, 0x000FB2},
+{0x000F78, 0x000F78, 0x000FB3},
+{0x000F81, 0x000F81, 0x000F71},
+{0x000F93, 0x000F93, 0x000F92},
+{0x000F9D, 0x000F9D, 0x000F9C},
+{0x000FA2, 0x000FA2, 0x000FA1},
+{0x000FA7, 0x000FA7, 0x000FA6},
+{0x000FAC, 0x000FAC, 0x000FAB},
+{0x000FB9, 0x000FB9, 0x000F90},
+{0x001026, 0x001026, 0x001025},
+{0x001B06, 0x001B06, 0x001B05},
+{0x001B08, 0x001B08, 0x001B07},
+{0x001B0A, 0x001B0A, 0x001B09},
+{0x001B0C, 0x001B0C, 0x001B0B},
+{0x001B0E, 0x001B0E, 0x001B0D},
+{0x001B12, 0x001B12, 0x001B11},
+{0x001B3B, 0x001B3B, 0x001B3A},
+{0x001B3D, 0x001B3D, 0x001B3C},
+{0x001B40, 0x001B40, 0x001B3E},
+{0x001B41, 0x001B41, 0x001B3F},
+{0x001B43, 0x001B43, 0x001B42},
+{0x001E00, 0x001E00, 0x000041},
+{0x001E01, 0x001E01, 0x000061},
+{0x001E02, 0x001E02, 0x000042},
+{0x001E03, 0x001E03, 0x000062},
+{0x001E04, 0x001E04, 0x000042},
+{0x001E05, 0x001E05, 0x000062},
+{0x001E06, 0x001E06, 0x000042},
+{0x001E07, 0x001E07, 0x000062},
+{0x001E08, 0x001E08, 0x000043},
+{0x001E09, 0x001E09, 0x000063},
+{0x001E0A, 0x001E0A, 0x000044},
+{0x001E0B, 0x001E0B, 0x000064},
+{0x001E0C, 0x001E0C, 0x000044},
+{0x001E0D, 0x001E0D, 0x000064},
+{0x001E0E, 0x001E0E, 0x000044},
+{0x001E0F, 0x001E0F, 0x000064},
+{0x001E10, 0x001E10, 0x000044},
+{0x001E11, 0x001E11, 0x000064},
+{0x001E12, 0x001E12, 0x000044},
+{0x001E13, 0x001E13, 0x000064},
+{0x001E14, 0x001E14, 0x000045},
+{0x001E15, 0x001E15, 0x000065},
+{0x001E16, 0x001E16, 0x000045},
+{0x001E17, 0x001E17, 0x000065},
+{0x001E18, 0x001E18, 0x000045},
+{0x001E19, 0x001E19, 0x000065},
+{0x001E1A, 0x001E1A, 0x000045},
+{0x001E1B, 0x001E1B, 0x000065},
+{0x001E1C, 0x001E1C, 0x000045},
+{0x001E1D, 0x001E1D, 0x000065},
+{0x001E1E, 0x001E1E, 0x000046},
+{0x001E1F, 0x001E1F, 0x000066},
+{0x001E20, 0x001E20, 0x000047},
+{0x001E21, 0x001E21, 0x000067},
+{0x001E22, 0x001E22, 0x000048},
+{0x001E23, 0x001E23, 0x000068},
+{0x001E24, 0x001E24, 0x000048},
+{0x001E25, 0x001E25, 0x000068},
+{0x001E26, 0x001E26, 0x000048},
+{0x001E27, 0x001E27, 0x000068},
+{0x001E28, 0x001E28, 0x000048},
+{0x001E29, 0x001E29, 0x000068},
+{0x001E2A, 0x001E2A, 0x000048},
+{0x001E2B, 0x001E2B, 0x000068},
+{0x001E2C, 0x001E2C, 0x000049},
+{0x001E2D, 0x001E2D, 0x000069},
+{0x001E2E, 0x001E2E, 0x000049},
+{0x001E2F, 0x001E2F, 0x000069},
+{0x001E30, 0x001E30, 0x00004B},
+{0x001E31, 0x001E31, 0x00006B},
+{0x001E32, 0x001E32, 0x00004B},
+{0x001E33, 0x001E33, 0x00006B},
+{0x001E34, 0x001E34, 0x00004B},
+{0x001E35, 0x001E35, 0x00006B},
+{0x001E36, 0x001E36, 0x00004C},
+{0x001E37, 0x001E37, 0x00006C},
+{0x001E38, 0x001E38, 0x00004C},
+{0x001E39, 0x001E39, 0x00006C},
+{0x001E3A, 0x001E3A, 0x00004C},
+{0x001E3B, 0x001E3B, 0x00006C},
+{0x001E3C, 0x001E3C, 0x00004C},
+{0x001E3D, 0x001E3D, 0x00006C},
+{0x001E3E, 0x001E3E, 0x00004D},
+{0x001E3F, 0x001E3F, 0x00006D},
+{0x001E40, 0x001E40, 0x00004D},
+{0x001E41, 0x001E41, 0x00006D},
+{0x001E42, 0x001E42, 0x00004D},
+{0x001E43, 0x001E43, 0x00006D},
+{0x001E44, 0x001E44, 0x00004E},
+{0x001E45, 0x001E45, 0x00006E},
+{0x001E46, 0x001E46, 0x00004E},
+{0x001E47, 0x001E47, 0x00006E},
+{0x001E48, 0x001E48, 0x00004E},
+{0x001E49, 0x001E49, 0x00006E},
+{0x001E4A, 0x001E4A, 0x00004E},
+{0x001E4B, 0x001E4B, 0x00006E},
+{0x001E4C, 0x001E4C, 0x00004F},
+{0x001E4D, 0x001E4D, 0x00006F},
+{0x001E4E, 0x001E4E, 0x00004F},
+{0x001E4F, 0x001E4F, 0x00006F},
+{0x001E50, 0x001E50, 0x00004F},
+{0x001E51, 0x001E51, 0x00006F},
+{0x001E52, 0x001E52, 0x00004F},
+{0x001E53, 0x001E53, 0x00006F},
+{0x001E54, 0x001E54, 0x000050},
+{0x001E55, 0x001E55, 0x000070},
+{0x001E56, 0x001E56, 0x000050},
+{0x001E57, 0x001E57, 0x000070},
+{0x001E58, 0x001E58, 0x000052},
+{0x001E59, 0x001E59, 0x000072},
+{0x001E5A, 0x001E5A, 0x000052},
+{0x001E5B, 0x001E5B, 0x000072},
+{0x001E5C, 0x001E5C, 0x000052},
+{0x001E5D, 0x001E5D, 0x000072},
+{0x001E5E, 0x001E5E, 0x000052},
+{0x001E5F, 0x001E5F, 0x000072},
+{0x001E60, 0x001E60, 0x000053},
+{0x001E61, 0x001E61, 0x000073},
+{0x001E62, 0x001E62, 0x000053},
+{0x001E63, 0x001E63, 0x000073},
+{0x001E64, 0x001E64, 0x000053},
+{0x001E65, 0x001E65, 0x000073},
+{0x001E66, 0x001E66, 0x000053},
+{0x001E67, 0x001E67, 0x000073},
+{0x001E68, 0x001E68, 0x000053},
+{0x001E69, 0x001E69, 0x000073},
+{0x001E6A, 0x001E6A, 0x000054},
+{0x001E6B, 0x001E6B, 0x000074},
+{0x001E6C, 0x001E6C, 0x000054},
+{0x001E6D, 0x001E6D, 0x000074},
+{0x001E6E, 0x001E6E, 0x000054},
+{0x001E6F, 0x001E6F, 0x000074},
+{0x001E70, 0x001E70, 0x000054},
+{0x001E71, 0x001E71, 0x000074},
+{0x001E72, 0x001E72, 0x000055},
+{0x001E73, 0x001E73, 0x000075},
+{0x001E74, 0x001E74, 0x000055},
+{0x001E75, 0x001E75, 0x000075},
+{0x001E76, 0x001E76, 0x000055},
+{0x001E77, 0x001E77, 0x000075},
+{0x001E78, 0x001E78, 0x000055},
+{0x001E79, 0x001E79, 0x000075},
+{0x001E7A, 0x001E7A, 0x000055},
+{0x001E7B, 0x001E7B, 0x000075},
+{0x001E7C, 0x001E7C, 0x000056},
+{0x001E7D, 0x001E7D, 0x000076},
+{0x001E7E, 0x001E7E, 0x000056},
+{0x001E7F, 0x001E7F, 0x000076},
+{0x001E80, 0x001E80, 0x000057},
+{0x001E81, 0x001E81, 0x000077},
+{0x001E82, 0x001E82, 0x000057},
+{0x001E83, 0x001E83, 0x000077},
+{0x001E84, 0x001E84, 0x000057},
+{0x001E85, 0x001E85, 0x000077},
+{0x001E86, 0x001E86, 0x000057},
+{0x001E87, 0x001E87, 0x000077},
+{0x001E88, 0x001E88, 0x000057},
+{0x001E89, 0x001E89, 0x000077},
+{0x001E8A, 0x001E8A, 0x000058},
+{0x001E8B, 0x001E8B, 0x000078},
+{0x001E8C, 0x001E8C, 0x000058},
+{0x001E8D, 0x001E8D, 0x000078},
+{0x001E8E, 0x001E8E, 0x000059},
+{0x001E8F, 0x001E8F, 0x000079},
+{0x001E90, 0x001E90, 0x00005A},
+{0x001E91, 0x001E91, 0x00007A},
+{0x001E92, 0x001E92, 0x00005A},
+{0x001E93, 0x001E93, 0x00007A},
+{0x001E94, 0x001E94, 0x00005A},
+{0x001E95, 0x001E95, 0x00007A},
+{0x001E96, 0x001E96, 0x000068},
+{0x001E97, 0x001E97, 0x000074},
+{0x001E98, 0x001E98, 0x000077},
+{0x001E99, 0x001E99, 0x000079},
+{0x001E9B, 0x001E9B, 0x00017F},
+{0x001EA0, 0x001EA0, 0x000041},
+{0x001EA1, 0x001EA1, 0x000061},
+{0x001EA2, 0x001EA2, 0x000041},
+{0x001EA3, 0x001EA3, 0x000061},
+{0x001EA4, 0x001EA4, 0x000041},
+{0x001EA5, 0x001EA5, 0x000061},
+{0x001EA6, 0x001EA6, 0x000041},
+{0x001EA7, 0x001EA7, 0x000061},
+{0x001EA8, 0x001EA8, 0x000041},
+{0x001EA9, 0x001EA9, 0x000061},
+{0x001EAA, 0x001EAA, 0x000041},
+{0x001EAB, 0x001EAB, 0x000061},
+{0x001EAC, 0x001EAC, 0x000041},
+{0x001EAD, 0x001EAD, 0x000061},
+{0x001EAE, 0x001EAE, 0x000041},
+{0x001EAF, 0x001EAF, 0x000061},
+{0x001EB0, 0x001EB0, 0x000041},
+{0x001EB1, 0x001EB1, 0x000061},
+{0x001EB2, 0x001EB2, 0x000041},
+{0x001EB3, 0x001EB3, 0x000061},
+{0x001EB4, 0x001EB4, 0x000041},
+{0x001EB5, 0x001EB5, 0x000061},
+{0x001EB6, 0x001EB6, 0x000041},
+{0x001EB7, 0x001EB7, 0x000061},
+{0x001EB8, 0x001EB8, 0x000045},
+{0x001EB9, 0x001EB9, 0x000065},
+{0x001EBA, 0x001EBA, 0x000045},
+{0x001EBB, 0x001EBB, 0x000065},
+{0x001EBC, 0x001EBC, 0x000045},
+{0x001EBD, 0x001EBD, 0x000065},
+{0x001EBE, 0x001EBE, 0x000045},
+{0x001EBF, 0x001EBF, 0x000065},
+{0x001EC0, 0x001EC0, 0x000045},
+{0x001EC1, 0x001EC1, 0x000065},
+{0x001EC2, 0x001EC2, 0x000045},
+{0x001EC3, 0x001EC3, 0x000065},
+{0x001EC4, 0x001EC4, 0x000045},
+{0x001EC5, 0x001EC5, 0x000065},
+{0x001EC6, 0x001EC6, 0x000045},
+{0x001EC7, 0x001EC7, 0x000065},
+{0x001EC8, 0x001EC8, 0x000049},
+{0x001EC9, 0x001EC9, 0x000069},
+{0x001ECA, 0x001ECA, 0x000049},
+{0x001ECB, 0x001ECB, 0x000069},
+{0x001ECC, 0x001ECC, 0x00004F},
+{0x001ECD, 0x001ECD, 0x00006F},
+{0x001ECE, 0x001ECE, 0x00004F},
+{0x001ECF, 0x001ECF, 0x00006F},
+{0x001ED0, 0x001ED0, 0x00004F},
+{0x001ED1, 0x001ED1, 0x00006F},
+{0x001ED2, 0x001ED2, 0x00004F},
+{0x001ED3, 0x001ED3, 0x00006F},
+{0x001ED4, 0x001ED4, 0x00004F},
+{0x001ED5, 0x001ED5, 0x00006F},
+{0x001ED6, 0x001ED6, 0x00004F},
+{0x001ED7, 0x001ED7, 0x00006F},
+{0x001ED8, 0x001ED8, 0x00004F},
+{0x001ED9, 0x001ED9, 0x00006F},
+{0x001EDA, 0x001EDA, 0x00004F},
+{0x001EDB, 0x001EDB, 0x00006F},
+{0x001EDC, 0x001EDC, 0x00004F},
+{0x001EDD, 0x001EDD, 0x00006F},
+{0x001EDE, 0x001EDE, 0x00004F},
+{0x001EDF, 0x001EDF, 0x00006F},
+{0x001EE0, 0x001EE0, 0x00004F},
+{0x001EE1, 0x001EE1, 0x00006F},
+{0x001EE2, 0x001EE2, 0x00004F},
+{0x001EE3, 0x001EE3, 0x00006F},
+{0x001EE4, 0x001EE4, 0x000055},
+{0x001EE5, 0x001EE5, 0x000075},
+{0x001EE6, 0x001EE6, 0x000055},
+{0x001EE7, 0x001EE7, 0x000075},
+{0x001EE8, 0x001EE8, 0x000055},
+{0x001EE9, 0x001EE9, 0x000075},
+{0x001EEA, 0x001EEA, 0x000055},
+{0x001EEB, 0x001EEB, 0x000075},
+{0x001EEC, 0x001EEC, 0x000055},
+{0x001EED, 0x001EED, 0x000075},
+{0x001EEE, 0x001EEE, 0x000055},
+{0x001EEF, 0x001EEF, 0x000075},
+{0x001EF0, 0x001EF0, 0x000055},
+{0x001EF1, 0x001EF1, 0x000075},
+{0x001EF2, 0x001EF2, 0x000059},
+{0x001EF3, 0x001EF3, 0x000079},
+{0x001EF4, 0x001EF4, 0x000059},
+{0x001EF5, 0x001EF5, 0x000079},
+{0x001EF6, 0x001EF6, 0x000059},
+{0x001EF7, 0x001EF7, 0x000079},
+{0x001EF8, 0x001EF8, 0x000059},
+{0x001EF9, 0x001EF9, 0x000079},
+{0x001F00, 0x001F07, 0x0003B1},
+{0x001F08, 0x001F0F, 0x000391},
+{0x001F10, 0x001F15, 0x0003B5},
+{0x001F18, 0x001F1D, 0x000395},
+{0x001F20, 0x001F27, 0x0003B7},
+{0x001F28, 0x001F2F, 0x000397},
+{0x001F30, 0x001F37, 0x0003B9},
+{0x001F38, 0x001F3F, 0x000399},
+{0x001F40, 0x001F45, 0x0003BF},
+{0x001F48, 0x001F4D, 0x00039F},
+{0x001F50, 0x001F57, 0x0003C5},
+{0x001F59, 0x001F59, 0x0003A5},
+{0x001F5B, 0x001F5B, 0x0003A5},
+{0x001F5D, 0x001F5D, 0x0003A5},
+{0x001F5F, 0x001F5F, 0x0003A5},
+{0x001F60, 0x001F67, 0x0003C9},
+{0x001F68, 0x001F6F, 0x0003A9},
+{0x001F70, 0x001F71, 0x0003B1},
+{0x001F72, 0x001F73, 0x0003B5},
+{0x001F74, 0x001F75, 0x0003B7},
+{0x001F76, 0x001F77, 0x0003B9},
+{0x001F78, 0x001F79, 0x0003BF},
+{0x001F7A, 0x001F7B, 0x0003C5},
+{0x001F7C, 0x001F7D, 0x0003C9},
+{0x001F80, 0x001F87, 0x0003B1},
+{0x001F88, 0x001F8F, 0x000391},
+{0x001F90, 0x001F97, 0x0003B7},
+{0x001F98, 0x001F9F, 0x000397},
+{0x001FA0, 0x001FA7, 0x0003C9},
+{0x001FA8, 0x001FAF, 0x0003A9},
+{0x001FB0, 0x001FB4, 0x0003B1},
+{0x001FB6, 0x001FB7, 0x0003B1},
+{0x001FB8, 0x001FBC, 0x000391},
+{0x001FBE, 0x001FBE, 0x0003B9},
+{0x001FC1, 0x001FC1, 0x0000A8},
+{0x001FC2, 0x001FC4, 0x0003B7},
+{0x001FC6, 0x001FC7, 0x0003B7},
+{0x001FC8, 0x001FC9, 0x000395},
+{0x001FCA, 0x001FCC, 0x000397},
+{0x001FCD, 0x001FCF, 0x001FBF},
+{0x001FD0, 0x001FD3, 0x0003B9},
+{0x001FD6, 0x001FD7, 0x0003B9},
+{0x001FD8, 0x001FDB, 0x000399},
+{0x001FDD, 0x001FDF, 0x001FFE},
+{0x001FE0, 0x001FE3, 0x0003C5},
+{0x001FE4, 0x001FE5, 0x0003C1},
+{0x001FE6, 0x001FE7, 0x0003C5},
+{0x001FE8, 0x001FEB, 0x0003A5},
+{0x001FEC, 0x001FEC, 0x0003A1},
+{0x001FED, 0x001FEE, 0x0000A8},
+{0x001FEF, 0x001FEF, 0x000060},
+{0x001FF2, 0x001FF4, 0x0003C9},
+{0x001FF6, 0x001FF7, 0x0003C9},
+{0x001FF8, 0x001FF9, 0x00039F},
+{0x001FFA, 0x001FFC, 0x0003A9},
+{0x001FFD, 0x001FFD, 0x0000B4},
+{0x002000, 0x002000, 0x002002},
+{0x002001, 0x002001, 0x002003},
+{0x002126, 0x002126, 0x0003A9},
+{0x00212A, 0x00212A, 0x00004B},
+{0x00212B, 0x00212B, 0x000041},
+{0x00219A, 0x00219A, 0x002190},
+{0x00219B, 0x00219B, 0x002192},
+{0x0021AE, 0x0021AE, 0x002194},
+{0x0021CD, 0x0021CD, 0x0021D0},
+{0x0021CE, 0x0021CE, 0x0021D4},
+{0x0021CF, 0x0021CF, 0x0021D2},
+{0x002204, 0x002204, 0x002203},
+{0x002209, 0x002209, 0x002208},
+{0x00220C, 0x00220C, 0x00220B},
+{0x002224, 0x002224, 0x002223},
+{0x002226, 0x002226, 0x002225},
+{0x002241, 0x002241, 0x00223C},
+{0x002244, 0x002244, 0x002243},
+{0x002247, 0x002247, 0x002245},
+{0x002249, 0x002249, 0x002248},
+{0x002260, 0x002260, 0x00003D},
+{0x002262, 0x002262, 0x002261},
+{0x00226D, 0x00226D, 0x00224D},
+{0x00226E, 0x00226E, 0x00003C},
+{0x00226F, 0x00226F, 0x00003E},
+{0x002270, 0x002270, 0x002264},
+{0x002271, 0x002271, 0x002265},
+{0x002274, 0x002274, 0x002272},
+{0x002275, 0x002275, 0x002273},
+{0x002278, 0x002278, 0x002276},
+{0x002279, 0x002279, 0x002277},
+{0x002280, 0x002280, 0x00227A},
+{0x002281, 0x002281, 0x00227B},
+{0x002284, 0x002284, 0x002282},
+{0x002285, 0x002285, 0x002283},
+{0x002288, 0x002288, 0x002286},
+{0x002289, 0x002289, 0x002287},
+{0x0022AC, 0x0022AC, 0x0022A2},
+{0x0022AD, 0x0022AD, 0x0022A8},
+{0x0022AE, 0x0022AE, 0x0022A9},
+{0x0022AF, 0x0022AF, 0x0022AB},
+{0x0022E0, 0x0022E0, 0x00227C},
+{0x0022E1, 0x0022E1, 0x00227D},
+{0x0022E2, 0x0022E2, 0x002291},
+{0x0022E3, 0x0022E3, 0x002292},
+{0x0022EA, 0x0022EA, 0x0022B2},
+{0x0022EB, 0x0022EB, 0x0022B3},
+{0x0022EC, 0x0022EC, 0x0022B4},
+{0x0022ED, 0x0022ED, 0x0022B5},
+{0x002329, 0x002329, 0x003008},
+{0x00232A, 0x00232A, 0x003009},
+{0x002ADC, 0x002ADC, 0x002ADD},
+{0x00304C, 0x00304C, 0x00304B},
+{0x00304E, 0x00304E, 0x00304D},
+{0x003050, 0x003050, 0x00304F},
+{0x003052, 0x003052, 0x003051},
+{0x003054, 0x003054, 0x003053},
+{0x003056, 0x003056, 0x003055},
+{0x003058, 0x003058, 0x003057},
+{0x00305A, 0x00305A, 0x003059},
+{0x00305C, 0x00305C, 0x00305B},
+{0x00305E, 0x00305E, 0x00305D},
+{0x003060, 0x003060, 0x00305F},
+{0x003062, 0x003062, 0x003061},
+{0x003065, 0x003065, 0x003064},
+{0x003067, 0x003067, 0x003066},
+{0x003069, 0x003069, 0x003068},
+{0x003070, 0x003071, 0x00306F},
+{0x003073, 0x003074, 0x003072},
+{0x003076, 0x003077, 0x003075},
+{0x003079, 0x00307A, 0x003078},
+{0x00307C, 0x00307D, 0x00307B},
+{0x003094, 0x003094, 0x003046},
+{0x00309E, 0x00309E, 0x00309D},
+{0x0030AC, 0x0030AC, 0x0030AB},
+{0x0030AE, 0x0030AE, 0x0030AD},
+{0x0030B0, 0x0030B0, 0x0030AF},
+{0x0030B2, 0x0030B2, 0x0030B1},
+{0x0030B4, 0x0030B4, 0x0030B3},
+{0x0030B6, 0x0030B6, 0x0030B5},
+{0x0030B8, 0x0030B8, 0x0030B7},
+{0x0030BA, 0x0030BA, 0x0030B9},
+{0x0030BC, 0x0030BC, 0x0030BB},
+{0x0030BE, 0x0030BE, 0x0030BD},
+{0x0030C0, 0x0030C0, 0x0030BF},
+{0x0030C2, 0x0030C2, 0x0030C1},
+{0x0030C5, 0x0030C5, 0x0030C4},
+{0x0030C7, 0x0030C7, 0x0030C6},
+{0x0030C9, 0x0030C9, 0x0030C8},
+{0x0030D0, 0x0030D1, 0x0030CF},
+{0x0030D3, 0x0030D4, 0x0030D2},
+{0x0030D6, 0x0030D7, 0x0030D5},
+{0x0030D9, 0x0030DA, 0x0030D8},
+{0x0030DC, 0x0030DD, 0x0030DB},
+{0x0030F4, 0x0030F4, 0x0030A6},
+{0x0030F7, 0x0030F7, 0x0030EF},
+{0x0030F8, 0x0030F8, 0x0030F0},
+{0x0030F9, 0x0030F9, 0x0030F1},
+{0x0030FA, 0x0030FA, 0x0030F2},
+{0x0030FE, 0x0030FE, 0x0030FD},
+{0x00AC00, 0x00AE4B, 0x001100},
+{0x00AE4C, 0x00B097, 0x001101},
+{0x00B098, 0x00B2E3, 0x001102},
+{0x00B2E4, 0x00B52F, 0x001103},
+{0x00B530, 0x00B77B, 0x001104},
+{0x00B77C, 0x00B9C7, 0x001105},
+{0x00B9C8, 0x00BC13, 0x001106},
+{0x00BC14, 0x00BE5F, 0x001107},
+{0x00BE60, 0x00C0AB, 0x001108},
+{0x00C0AC, 0x00C2F7, 0x001109},
+{0x00C2F8, 0x00C543, 0x00110A},
+{0x00C544, 0x00C78F, 0x00110B},
+{0x00C790, 0x00C9DB, 0x00110C},
+{0x00C9DC, 0x00CC27, 0x00110D},
+{0x00CC28, 0x00CE73, 0x00110E},
+{0x00CE74, 0x00D0BF, 0x00110F},
+{0x00D0C0, 0x00D30B, 0x001110},
+{0x00D30C, 0x00D557, 0x001111},
+{0x00D558, 0x00D7A3, 0x001112},
+{0x00F900, 0x00F900, 0x008C48},
+{0x00F901, 0x00F901, 0x0066F4},
+{0x00F902, 0x00F902, 0x008ECA},
+{0x00F903, 0x00F903, 0x008CC8},
+{0x00F904, 0x00F904, 0x006ED1},
+{0x00F905, 0x00F905, 0x004E32},
+{0x00F906, 0x00F906, 0x0053E5},
+{0x00F907, 0x00F908, 0x009F9C},
+{0x00F909, 0x00F909, 0x005951},
+{0x00F90A, 0x00F90A, 0x0091D1},
+{0x00F90B, 0x00F90B, 0x005587},
+{0x00F90C, 0x00F90C, 0x005948},
+{0x00F90D, 0x00F90D, 0x0061F6},
+{0x00F90E, 0x00F90E, 0x007669},
+{0x00F90F, 0x00F90F, 0x007F85},
+{0x00F910, 0x00F910, 0x00863F},
+{0x00F911, 0x00F911, 0x0087BA},
+{0x00F912, 0x00F912, 0x0088F8},
+{0x00F913, 0x00F913, 0x00908F},
+{0x00F914, 0x00F914, 0x006A02},
+{0x00F915, 0x00F915, 0x006D1B},
+{0x00F916, 0x00F916, 0x0070D9},
+{0x00F917, 0x00F917, 0x0073DE},
+{0x00F918, 0x00F918, 0x00843D},
+{0x00F919, 0x00F919, 0x00916A},
+{0x00F91A, 0x00F91A, 0x0099F1},
+{0x00F91B, 0x00F91B, 0x004E82},
+{0x00F91C, 0x00F91C, 0x005375},
+{0x00F91D, 0x00F91D, 0x006B04},
+{0x00F91E, 0x00F91E, 0x00721B},
+{0x00F91F, 0x00F91F, 0x00862D},
+{0x00F920, 0x00F920, 0x009E1E},
+{0x00F921, 0x00F921, 0x005D50},
+{0x00F922, 0x00F922, 0x006FEB},
+{0x00F923, 0x00F923, 0x0085CD},
+{0x00F924, 0x00F924, 0x008964},
+{0x00F925, 0x00F925, 0x0062C9},
+{0x00F926, 0x00F926, 0x0081D8},
+{0x00F927, 0x00F927, 0x00881F},
+{0x00F928, 0x00F928, 0x005ECA},
+{0x00F929, 0x00F929, 0x006717},
+{0x00F92A, 0x00F92A, 0x006D6A},
+{0x00F92B, 0x00F92B, 0x0072FC},
+{0x00F92C, 0x00F92C, 0x0090CE},
+{0x00F92D, 0x00F92D, 0x004F86},
+{0x00F92E, 0x00F92E, 0x0051B7},
+{0x00F92F, 0x00F92F, 0x0052DE},
+{0x00F930, 0x00F930, 0x0064C4},
+{0x00F931, 0x00F931, 0x006AD3},
+{0x00F932, 0x00F932, 0x007210},
+{0x00F933, 0x00F933, 0x0076E7},
+{0x00F934, 0x00F934, 0x008001},
+{0x00F935, 0x00F935, 0x008606},
+{0x00F936, 0x00F936, 0x00865C},
+{0x00F937, 0x00F937, 0x008DEF},
+{0x00F938, 0x00F938, 0x009732},
+{0x00F939, 0x00F939, 0x009B6F},
+{0x00F93A, 0x00F93A, 0x009DFA},
+{0x00F93B, 0x00F93B, 0x00788C},
+{0x00F93C, 0x00F93C, 0x00797F},
+{0x00F93D, 0x00F93D, 0x007DA0},
+{0x00F93E, 0x00F93E, 0x0083C9},
+{0x00F93F, 0x00F93F, 0x009304},
+{0x00F940, 0x00F940, 0x009E7F},
+{0x00F941, 0x00F941, 0x008AD6},
+{0x00F942, 0x00F942, 0x0058DF},
+{0x00F943, 0x00F943, 0x005F04},
+{0x00F944, 0x00F944, 0x007C60},
+{0x00F945, 0x00F945, 0x00807E},
+{0x00F946, 0x00F946, 0x007262},
+{0x00F947, 0x00F947, 0x0078CA},
+{0x00F948, 0x00F948, 0x008CC2},
+{0x00F949, 0x00F949, 0x0096F7},
+{0x00F94A, 0x00F94A, 0x0058D8},
+{0x00F94B, 0x00F94B, 0x005C62},
+{0x00F94C, 0x00F94C, 0x006A13},
+{0x00F94D, 0x00F94D, 0x006DDA},
+{0x00F94E, 0x00F94E, 0x006F0F},
+{0x00F94F, 0x00F94F, 0x007D2F},
+{0x00F950, 0x00F950, 0x007E37},
+{0x00F951, 0x00F951, 0x00964B},
+{0x00F952, 0x00F952, 0x0052D2},
+{0x00F953, 0x00F953, 0x00808B},
+{0x00F954, 0x00F954, 0x0051DC},
+{0x00F955, 0x00F955, 0x0051CC},
+{0x00F956, 0x00F956, 0x007A1C},
+{0x00F957, 0x00F957, 0x007DBE},
+{0x00F958, 0x00F958, 0x0083F1},
+{0x00F959, 0x00F959, 0x009675},
+{0x00F95A, 0x00F95A, 0x008B80},
+{0x00F95B, 0x00F95B, 0x0062CF},
+{0x00F95C, 0x00F95C, 0x006A02},
+{0x00F95D, 0x00F95D, 0x008AFE},
+{0x00F95E, 0x00F95E, 0x004E39},
+{0x00F95F, 0x00F95F, 0x005BE7},
+{0x00F960, 0x00F960, 0x006012},
+{0x00F961, 0x00F961, 0x007387},
+{0x00F962, 0x00F962, 0x007570},
+{0x00F963, 0x00F963, 0x005317},
+{0x00F964, 0x00F964, 0x0078FB},
+{0x00F965, 0x00F965, 0x004FBF},
+{0x00F966, 0x00F966, 0x005FA9},
+{0x00F967, 0x00F967, 0x004E0D},
+{0x00F968, 0x00F968, 0x006CCC},
+{0x00F969, 0x00F969, 0x006578},
+{0x00F96A, 0x00F96A, 0x007D22},
+{0x00F96B, 0x00F96B, 0x0053C3},
+{0x00F96C, 0x00F96C, 0x00585E},
+{0x00F96D, 0x00F96D, 0x007701},
+{0x00F96E, 0x00F96E, 0x008449},
+{0x00F96F, 0x00F96F, 0x008AAA},
+{0x00F970, 0x00F970, 0x006BBA},
+{0x00F971, 0x00F971, 0x008FB0},
+{0x00F972, 0x00F972, 0x006C88},
+{0x00F973, 0x00F973, 0x0062FE},
+{0x00F974, 0x00F974, 0x0082E5},
+{0x00F975, 0x00F975, 0x0063A0},
+{0x00F976, 0x00F976, 0x007565},
+{0x00F977, 0x00F977, 0x004EAE},
+{0x00F978, 0x00F978, 0x005169},
+{0x00F979, 0x00F979, 0x0051C9},
+{0x00F97A, 0x00F97A, 0x006881},
+{0x00F97B, 0x00F97B, 0x007CE7},
+{0x00F97C, 0x00F97C, 0x00826F},
+{0x00F97D, 0x00F97D, 0x008AD2},
+{0x00F97E, 0x00F97E, 0x0091CF},
+{0x00F97F, 0x00F97F, 0x0052F5},
+{0x00F980, 0x00F980, 0x005442},
+{0x00F981, 0x00F981, 0x005973},
+{0x00F982, 0x00F982, 0x005EEC},
+{0x00F983, 0x00F983, 0x0065C5},
+{0x00F984, 0x00F984, 0x006FFE},
+{0x00F985, 0x00F985, 0x00792A},
+{0x00F986, 0x00F986, 0x0095AD},
+{0x00F987, 0x00F987, 0x009A6A},
+{0x00F988, 0x00F988, 0x009E97},
+{0x00F989, 0x00F989, 0x009ECE},
+{0x00F98A, 0x00F98A, 0x00529B},
+{0x00F98B, 0x00F98B, 0x0066C6},
+{0x00F98C, 0x00F98C, 0x006B77},
+{0x00F98D, 0x00F98D, 0x008F62},
+{0x00F98E, 0x00F98E, 0x005E74},
+{0x00F98F, 0x00F98F, 0x006190},
+{0x00F990, 0x00F990, 0x006200},
+{0x00F991, 0x00F991, 0x00649A},
+{0x00F992, 0x00F992, 0x006F23},
+{0x00F993, 0x00F993, 0x007149},
+{0x00F994, 0x00F994, 0x007489},
+{0x00F995, 0x00F995, 0x0079CA},
+{0x00F996, 0x00F996, 0x007DF4},
+{0x00F997, 0x00F997, 0x00806F},
+{0x00F998, 0x00F998, 0x008F26},
+{0x00F999, 0x00F999, 0x0084EE},
+{0x00F99A, 0x00F99A, 0x009023},
+{0x00F99B, 0x00F99B, 0x00934A},
+{0x00F99C, 0x00F99C, 0x005217},
+{0x00F99D, 0x00F99D, 0x0052A3},
+{0x00F99E, 0x00F99E, 0x0054BD},
+{0x00F99F, 0x00F99F, 0x0070C8},
+{0x00F9A0, 0x00F9A0, 0x0088C2},
+{0x00F9A1, 0x00F9A1, 0x008AAA},
+{0x00F9A2, 0x00F9A2, 0x005EC9},
+{0x00F9A3, 0x00F9A3, 0x005FF5},
+{0x00F9A4, 0x00F9A4, 0x00637B},
+{0x00F9A5, 0x00F9A5, 0x006BAE},
+{0x00F9A6, 0x00F9A6, 0x007C3E},
+{0x00F9A7, 0x00F9A7, 0x007375},
+{0x00F9A8, 0x00F9A8, 0x004EE4},
+{0x00F9A9, 0x00F9A9, 0x0056F9},
+{0x00F9AA, 0x00F9AA, 0x005BE7},
+{0x00F9AB, 0x00F9AB, 0x005DBA},
+{0x00F9AC, 0x00F9AC, 0x00601C},
+{0x00F9AD, 0x00F9AD, 0x0073B2},
+{0x00F9AE, 0x00F9AE, 0x007469},
+{0x00F9AF, 0x00F9AF, 0x007F9A},
+{0x00F9B0, 0x00F9B0, 0x008046},
+{0x00F9B1, 0x00F9B1, 0x009234},
+{0x00F9B2, 0x00F9B2, 0x0096F6},
+{0x00F9B3, 0x00F9B3, 0x009748},
+{0x00F9B4, 0x00F9B4, 0x009818},
+{0x00F9B5, 0x00F9B5, 0x004F8B},
+{0x00F9B6, 0x00F9B6, 0x0079AE},
+{0x00F9B7, 0x00F9B7, 0x0091B4},
+{0x00F9B8, 0x00F9B8, 0x0096B8},
+{0x00F9B9, 0x00F9B9, 0x0060E1},
+{0x00F9BA, 0x00F9BA, 0x004E86},
+{0x00F9BB, 0x00F9BB, 0x0050DA},
+{0x00F9BC, 0x00F9BC, 0x005BEE},
+{0x00F9BD, 0x00F9BD, 0x005C3F},
+{0x00F9BE, 0x00F9BE, 0x006599},
+{0x00F9BF, 0x00F9BF, 0x006A02},
+{0x00F9C0, 0x00F9C0, 0x0071CE},
+{0x00F9C1, 0x00F9C1, 0x007642},
+{0x00F9C2, 0x00F9C2, 0x0084FC},
+{0x00F9C3, 0x00F9C3, 0x00907C},
+{0x00F9C4, 0x00F9C4, 0x009F8D},
+{0x00F9C5, 0x00F9C5, 0x006688},
+{0x00F9C6, 0x00F9C6, 0x00962E},
+{0x00F9C7, 0x00F9C7, 0x005289},
+{0x00F9C8, 0x00F9C8, 0x00677B},
+{0x00F9C9, 0x00F9C9, 0x0067F3},
+{0x00F9CA, 0x00F9CA, 0x006D41},
+{0x00F9CB, 0x00F9CB, 0x006E9C},
+{0x00F9CC, 0x00F9CC, 0x007409},
+{0x00F9CD, 0x00F9CD, 0x007559},
+{0x00F9CE, 0x00F9CE, 0x00786B},
+{0x00F9CF, 0x00F9CF, 0x007D10},
+{0x00F9D0, 0x00F9D0, 0x00985E},
+{0x00F9D1, 0x00F9D1, 0x00516D},
+{0x00F9D2, 0x00F9D2, 0x00622E},
+{0x00F9D3, 0x00F9D3, 0x009678},
+{0x00F9D4, 0x00F9D4, 0x00502B},
+{0x00F9D5, 0x00F9D5, 0x005D19},
+{0x00F9D6, 0x00F9D6, 0x006DEA},
+{0x00F9D7, 0x00F9D7, 0x008F2A},
+{0x00F9D8, 0x00F9D8, 0x005F8B},
+{0x00F9D9, 0x00F9D9, 0x006144},
+{0x00F9DA, 0x00F9DA, 0x006817},
+{0x00F9DB, 0x00F9DB, 0x007387},
+{0x00F9DC, 0x00F9DC, 0x009686},
+{0x00F9DD, 0x00F9DD, 0x005229},
+{0x00F9DE, 0x00F9DE, 0x00540F},
+{0x00F9DF, 0x00F9DF, 0x005C65},
+{0x00F9E0, 0x00F9E0, 0x006613},
+{0x00F9E1, 0x00F9E1, 0x00674E},
+{0x00F9E2, 0x00F9E2, 0x0068A8},
+{0x00F9E3, 0x00F9E3, 0x006CE5},
+{0x00F9E4, 0x00F9E4, 0x007406},
+{0x00F9E5, 0x00F9E5, 0x0075E2},
+{0x00F9E6, 0x00F9E6, 0x007F79},
+{0x00F9E7, 0x00F9E7, 0x0088CF},
+{0x00F9E8, 0x00F9E8, 0x0088E1},
+{0x00F9E9, 0x00F9E9, 0x0091CC},
+{0x00F9EA, 0x00F9EA, 0x0096E2},
+{0x00F9EB, 0x00F9EB, 0x00533F},
+{0x00F9EC, 0x00F9EC, 0x006EBA},
+{0x00F9ED, 0x00F9ED, 0x00541D},
+{0x00F9EE, 0x00F9EE, 0x0071D0},
+{0x00F9EF, 0x00F9EF, 0x007498},
+{0x00F9F0, 0x00F9F0, 0x0085FA},
+{0x00F9F1, 0x00F9F1, 0x0096A3},
+{0x00F9F2, 0x00F9F2, 0x009C57},
+{0x00F9F3, 0x00F9F3, 0x009E9F},
+{0x00F9F4, 0x00F9F4, 0x006797},
+{0x00F9F5, 0x00F9F5, 0x006DCB},
+{0x00F9F6, 0x00F9F6, 0x0081E8},
+{0x00F9F7, 0x00F9F7, 0x007ACB},
+{0x00F9F8, 0x00F9F8, 0x007B20},
+{0x00F9F9, 0x00F9F9, 0x007C92},
+{0x00F9FA, 0x00F9FA, 0x0072C0},
+{0x00F9FB, 0x00F9FB, 0x007099},
+{0x00F9FC, 0x00F9FC, 0x008B58},
+{0x00F9FD, 0x00F9FD, 0x004EC0},
+{0x00F9FE, 0x00F9FE, 0x008336},
+{0x00F9FF, 0x00F9FF, 0x00523A},
+{0x00FA00, 0x00FA00, 0x005207},
+{0x00FA01, 0x00FA01, 0x005EA6},
+{0x00FA02, 0x00FA02, 0x0062D3},
+{0x00FA03, 0x00FA03, 0x007CD6},
+{0x00FA04, 0x00FA04, 0x005B85},
+{0x00FA05, 0x00FA05, 0x006D1E},
+{0x00FA06, 0x00FA06, 0x0066B4},
+{0x00FA07, 0x00FA07, 0x008F3B},
+{0x00FA08, 0x00FA08, 0x00884C},
+{0x00FA09, 0x00FA09, 0x00964D},
+{0x00FA0A, 0x00FA0A, 0x00898B},
+{0x00FA0B, 0x00FA0B, 0x005ED3},
+{0x00FA0C, 0x00FA0C, 0x005140},
+{0x00FA0D, 0x00FA0D, 0x0055C0},
+{0x00FA10, 0x00FA10, 0x00585A},
+{0x00FA12, 0x00FA12, 0x006674},
+{0x00FA15, 0x00FA15, 0x0051DE},
+{0x00FA16, 0x00FA16, 0x00732A},
+{0x00FA17, 0x00FA17, 0x0076CA},
+{0x00FA18, 0x00FA18, 0x00793C},
+{0x00FA19, 0x00FA19, 0x00795E},
+{0x00FA1A, 0x00FA1A, 0x007965},
+{0x00FA1B, 0x00FA1B, 0x00798F},
+{0x00FA1C, 0x00FA1C, 0x009756},
+{0x00FA1D, 0x00FA1D, 0x007CBE},
+{0x00FA1E, 0x00FA1E, 0x007FBD},
+{0x00FA20, 0x00FA20, 0x008612},
+{0x00FA22, 0x00FA22, 0x008AF8},
+{0x00FA25, 0x00FA25, 0x009038},
+{0x00FA26, 0x00FA26, 0x0090FD},
+{0x00FA2A, 0x00FA2A, 0x0098EF},
+{0x00FA2B, 0x00FA2B, 0x0098FC},
+{0x00FA2C, 0x00FA2C, 0x009928},
+{0x00FA2D, 0x00FA2D, 0x009DB4},
+{0x00FA2E, 0x00FA2E, 0x0090DE},
+{0x00FA2F, 0x00FA2F, 0x0096B7},
+{0x00FA30, 0x00FA30, 0x004FAE},
+{0x00FA31, 0x00FA31, 0x0050E7},
+{0x00FA32, 0x00FA32, 0x00514D},
+{0x00FA33, 0x00FA33, 0x0052C9},
+{0x00FA34, 0x00FA34, 0x0052E4},
+{0x00FA35, 0x00FA35, 0x005351},
+{0x00FA36, 0x00FA36, 0x00559D},
+{0x00FA37, 0x00FA37, 0x005606},
+{0x00FA38, 0x00FA38, 0x005668},
+{0x00FA39, 0x00FA39, 0x005840},
+{0x00FA3A, 0x00FA3A, 0x0058A8},
+{0x00FA3B, 0x00FA3B, 0x005C64},
+{0x00FA3C, 0x00FA3C, 0x005C6E},
+{0x00FA3D, 0x00FA3D, 0x006094},
+{0x00FA3E, 0x00FA3E, 0x006168},
+{0x00FA3F, 0x00FA3F, 0x00618E},
+{0x00FA40, 0x00FA40, 0x0061F2},
+{0x00FA41, 0x00FA41, 0x00654F},
+{0x00FA42, 0x00FA42, 0x0065E2},
+{0x00FA43, 0x00FA43, 0x006691},
+{0x00FA44, 0x00FA44, 0x006885},
+{0x00FA45, 0x00FA45, 0x006D77},
+{0x00FA46, 0x00FA46, 0x006E1A},
+{0x00FA47, 0x00FA47, 0x006F22},
+{0x00FA48, 0x00FA48, 0x00716E},
+{0x00FA49, 0x00FA49, 0x00722B},
+{0x00FA4A, 0x00FA4A, 0x007422},
+{0x00FA4B, 0x00FA4B, 0x007891},
+{0x00FA4C, 0x00FA4C, 0x00793E},
+{0x00FA4D, 0x00FA4D, 0x007949},
+{0x00FA4E, 0x00FA4E, 0x007948},
+{0x00FA4F, 0x00FA4F, 0x007950},
+{0x00FA50, 0x00FA50, 0x007956},
+{0x00FA51, 0x00FA51, 0x00795D},
+{0x00FA52, 0x00FA52, 0x00798D},
+{0x00FA53, 0x00FA53, 0x00798E},
+{0x00FA54, 0x00FA54, 0x007A40},
+{0x00FA55, 0x00FA55, 0x007A81},
+{0x00FA56, 0x00FA56, 0x007BC0},
+{0x00FA57, 0x00FA57, 0x007DF4},
+{0x00FA58, 0x00FA58, 0x007E09},
+{0x00FA59, 0x00FA59, 0x007E41},
+{0x00FA5A, 0x00FA5A, 0x007F72},
+{0x00FA5B, 0x00FA5B, 0x008005},
+{0x00FA5C, 0x00FA5C, 0x0081ED},
+{0x00FA5D, 0x00FA5E, 0x008279},
+{0x00FA5F, 0x00FA5F, 0x008457},
+{0x00FA60, 0x00FA60, 0x008910},
+{0x00FA61, 0x00FA61, 0x008996},
+{0x00FA62, 0x00FA62, 0x008B01},
+{0x00FA63, 0x00FA63, 0x008B39},
+{0x00FA64, 0x00FA64, 0x008CD3},
+{0x00FA65, 0x00FA65, 0x008D08},
+{0x00FA66, 0x00FA66, 0x008FB6},
+{0x00FA67, 0x00FA67, 0x009038},
+{0x00FA68, 0x00FA68, 0x0096E3},
+{0x00FA69, 0x00FA69, 0x0097FF},
+{0x00FA6A, 0x00FA6A, 0x00983B},
+{0x00FA6B, 0x00FA6B, 0x006075},
+{0x00FA6C, 0x00FA6C, 0x0242EE},
+{0x00FA6D, 0x00FA6D, 0x008218},
+{0x00FA70, 0x00FA70, 0x004E26},
+{0x00FA71, 0x00FA71, 0x0051B5},
+{0x00FA72, 0x00FA72, 0x005168},
+{0x00FA73, 0x00FA73, 0x004F80},
+{0x00FA74, 0x00FA74, 0x005145},
+{0x00FA75, 0x00FA75, 0x005180},
+{0x00FA76, 0x00FA76, 0x0052C7},
+{0x00FA77, 0x00FA77, 0x0052FA},
+{0x00FA78, 0x00FA78, 0x00559D},
+{0x00FA79, 0x00FA79, 0x005555},
+{0x00FA7A, 0x00FA7A, 0x005599},
+{0x00FA7B, 0x00FA7B, 0x0055E2},
+{0x00FA7C, 0x00FA7C, 0x00585A},
+{0x00FA7D, 0x00FA7D, 0x0058B3},
+{0x00FA7E, 0x00FA7E, 0x005944},
+{0x00FA7F, 0x00FA7F, 0x005954},
+{0x00FA80, 0x00FA80, 0x005A62},
+{0x00FA81, 0x00FA81, 0x005B28},
+{0x00FA82, 0x00FA82, 0x005ED2},
+{0x00FA83, 0x00FA83, 0x005ED9},
+{0x00FA84, 0x00FA84, 0x005F69},
+{0x00FA85, 0x00FA85, 0x005FAD},
+{0x00FA86, 0x00FA86, 0x0060D8},
+{0x00FA87, 0x00FA87, 0x00614E},
+{0x00FA88, 0x00FA88, 0x006108},
+{0x00FA89, 0x00FA89, 0x00618E},
+{0x00FA8A, 0x00FA8A, 0x006160},
+{0x00FA8B, 0x00FA8B, 0x0061F2},
+{0x00FA8C, 0x00FA8C, 0x006234},
+{0x00FA8D, 0x00FA8D, 0x0063C4},
+{0x00FA8E, 0x00FA8E, 0x00641C},
+{0x00FA8F, 0x00FA8F, 0x006452},
+{0x00FA90, 0x00FA90, 0x006556},
+{0x00FA91, 0x00FA91, 0x006674},
+{0x00FA92, 0x00FA92, 0x006717},
+{0x00FA93, 0x00FA93, 0x00671B},
+{0x00FA94, 0x00FA94, 0x006756},
+{0x00FA95, 0x00FA95, 0x006B79},
+{0x00FA96, 0x00FA96, 0x006BBA},
+{0x00FA97, 0x00FA97, 0x006D41},
+{0x00FA98, 0x00FA98, 0x006EDB},
+{0x00FA99, 0x00FA99, 0x006ECB},
+{0x00FA9A, 0x00FA9A, 0x006F22},
+{0x00FA9B, 0x00FA9B, 0x00701E},
+{0x00FA9C, 0x00FA9C, 0x00716E},
+{0x00FA9D, 0x00FA9D, 0x0077A7},
+{0x00FA9E, 0x00FA9E, 0x007235},
+{0x00FA9F, 0x00FA9F, 0x0072AF},
+{0x00FAA0, 0x00FAA0, 0x00732A},
+{0x00FAA1, 0x00FAA1, 0x007471},
+{0x00FAA2, 0x00FAA2, 0x007506},
+{0x00FAA3, 0x00FAA3, 0x00753B},
+{0x00FAA4, 0x00FAA4, 0x00761D},
+{0x00FAA5, 0x00FAA5, 0x00761F},
+{0x00FAA6, 0x00FAA6, 0x0076CA},
+{0x00FAA7, 0x00FAA7, 0x0076DB},
+{0x00FAA8, 0x00FAA8, 0x0076F4},
+{0x00FAA9, 0x00FAA9, 0x00774A},
+{0x00FAAA, 0x00FAAA, 0x007740},
+{0x00FAAB, 0x00FAAB, 0x0078CC},
+{0x00FAAC, 0x00FAAC, 0x007AB1},
+{0x00FAAD, 0x00FAAD, 0x007BC0},
+{0x00FAAE, 0x00FAAE, 0x007C7B},
+{0x00FAAF, 0x00FAAF, 0x007D5B},
+{0x00FAB0, 0x00FAB0, 0x007DF4},
+{0x00FAB1, 0x00FAB1, 0x007F3E},
+{0x00FAB2, 0x00FAB2, 0x008005},
+{0x00FAB3, 0x00FAB3, 0x008352},
+{0x00FAB4, 0x00FAB4, 0x0083EF},
+{0x00FAB5, 0x00FAB5, 0x008779},
+{0x00FAB6, 0x00FAB6, 0x008941},
+{0x00FAB7, 0x00FAB7, 0x008986},
+{0x00FAB8, 0x00FAB8, 0x008996},
+{0x00FAB9, 0x00FAB9, 0x008ABF},
+{0x00FABA, 0x00FABA, 0x008AF8},
+{0x00FABB, 0x00FABB, 0x008ACB},
+{0x00FABC, 0x00FABC, 0x008B01},
+{0x00FABD, 0x00FABD, 0x008AFE},
+{0x00FABE, 0x00FABE, 0x008AED},
+{0x00FABF, 0x00FABF, 0x008B39},
+{0x00FAC0, 0x00FAC0, 0x008B8A},
+{0x00FAC1, 0x00FAC1, 0x008D08},
+{0x00FAC2, 0x00FAC2, 0x008F38},
+{0x00FAC3, 0x00FAC3, 0x009072},
+{0x00FAC4, 0x00FAC4, 0x009199},
+{0x00FAC5, 0x00FAC5, 0x009276},
+{0x00FAC6, 0x00FAC6, 0x00967C},
+{0x00FAC7, 0x00FAC7, 0x0096E3},
+{0x00FAC8, 0x00FAC8, 0x009756},
+{0x00FAC9, 0x00FAC9, 0x0097DB},
+{0x00FACA, 0x00FACA, 0x0097FF},
+{0x00FACB, 0x00FACB, 0x00980B},
+{0x00FACC, 0x00FACC, 0x00983B},
+{0x00FACD, 0x00FACD, 0x009B12},
+{0x00FACE, 0x00FACE, 0x009F9C},
+{0x00FACF, 0x00FACF, 0x02284A},
+{0x00FAD0, 0x00FAD0, 0x022844},
+{0x00FAD1, 0x00FAD1, 0x0233D5},
+{0x00FAD2, 0x00FAD2, 0x003B9D},
+{0x00FAD3, 0x00FAD3, 0x004018},
+{0x00FAD4, 0x00FAD4, 0x004039},
+{0x00FAD5, 0x00FAD5, 0x025249},
+{0x00FAD6, 0x00FAD6, 0x025CD0},
+{0x00FAD7, 0x00FAD7, 0x027ED3},
+{0x00FAD8, 0x00FAD8, 0x009F43},
+{0x00FAD9, 0x00FAD9, 0x009F8E},
+{0x00FB1D, 0x00FB1D, 0x0005D9},
+{0x00FB1F, 0x00FB1F, 0x0005F2},
+{0x00FB2A, 0x00FB2D, 0x0005E9},
+{0x00FB2E, 0x00FB30, 0x0005D0},
+{0x00FB31, 0x00FB31, 0x0005D1},
+{0x00FB32, 0x00FB32, 0x0005D2},
+{0x00FB33, 0x00FB33, 0x0005D3},
+{0x00FB34, 0x00FB34, 0x0005D4},
+{0x00FB35, 0x00FB35, 0x0005D5},
+{0x00FB36, 0x00FB36, 0x0005D6},
+{0x00FB38, 0x00FB38, 0x0005D8},
+{0x00FB39, 0x00FB39, 0x0005D9},
+{0x00FB3A, 0x00FB3A, 0x0005DA},
+{0x00FB3B, 0x00FB3B, 0x0005DB},
+{0x00FB3C, 0x00FB3C, 0x0005DC},
+{0x00FB3E, 0x00FB3E, 0x0005DE},
+{0x00FB40, 0x00FB40, 0x0005E0},
+{0x00FB41, 0x00FB41, 0x0005E1},
+{0x00FB43, 0x00FB43, 0x0005E3},
+{0x00FB44, 0x00FB44, 0x0005E4},
+{0x00FB46, 0x00FB46, 0x0005E6},
+{0x00FB47, 0x00FB47, 0x0005E7},
+{0x00FB48, 0x00FB48, 0x0005E8},
+{0x00FB49, 0x00FB49, 0x0005E9},
+{0x00FB4A, 0x00FB4A, 0x0005EA},
+{0x00FB4B, 0x00FB4B, 0x0005D5},
+{0x00FB4C, 0x00FB4C, 0x0005D1},
+{0x00FB4D, 0x00FB4D, 0x0005DB},
+{0x00FB4E, 0x00FB4E, 0x0005E4},
+{0x01109A, 0x01109A, 0x011099},
+{0x01109C, 0x01109C, 0x01109B},
+{0x0110AB, 0x0110AB, 0x0110A5},
+{0x01112E, 0x01112E, 0x011131},
+{0x01112F, 0x01112F, 0x011132},
+{0x01134B, 0x01134C, 0x011347},
+{0x0114BB, 0x0114BC, 0x0114B9},
+{0x0114BE, 0x0114BE, 0x0114B9},
+{0x0115BA, 0x0115BA, 0x0115B8},
+{0x0115BB, 0x0115BB, 0x0115B9},
+{0x011938, 0x011938, 0x011935},
+{0x01D15E, 0x01D15E, 0x01D157},
+{0x01D15F, 0x01D164, 0x01D158},
+{0x01D1BB, 0x01D1BB, 0x01D1B9},
+{0x01D1BC, 0x01D1BC, 0x01D1BA},
+{0x01D1BD, 0x01D1BD, 0x01D1B9},
+{0x01D1BE, 0x01D1BE, 0x01D1BA},
+{0x01D1BF, 0x01D1BF, 0x01D1B9},
+{0x01D1C0, 0x01D1C0, 0x01D1BA},
+{0x02F800, 0x02F800, 0x004E3D},
+{0x02F801, 0x02F801, 0x004E38},
+{0x02F802, 0x02F802, 0x004E41},
+{0x02F803, 0x02F803, 0x020122},
+{0x02F804, 0x02F804, 0x004F60},
+{0x02F805, 0x02F805, 0x004FAE},
+{0x02F806, 0x02F806, 0x004FBB},
+{0x02F807, 0x02F807, 0x005002},
+{0x02F808, 0x02F808, 0x00507A},
+{0x02F809, 0x02F809, 0x005099},
+{0x02F80A, 0x02F80A, 0x0050E7},
+{0x02F80B, 0x02F80B, 0x0050CF},
+{0x02F80C, 0x02F80C, 0x00349E},
+{0x02F80D, 0x02F80D, 0x02063A},
+{0x02F80E, 0x02F80E, 0x00514D},
+{0x02F80F, 0x02F80F, 0x005154},
+{0x02F810, 0x02F810, 0x005164},
+{0x02F811, 0x02F811, 0x005177},
+{0x02F812, 0x02F812, 0x02051C},
+{0x02F813, 0x02F813, 0x0034B9},
+{0x02F814, 0x02F814, 0x005167},
+{0x02F815, 0x02F815, 0x00518D},
+{0x02F816, 0x02F816, 0x02054B},
+{0x02F817, 0x02F817, 0x005197},
+{0x02F818, 0x02F818, 0x0051A4},
+{0x02F819, 0x02F819, 0x004ECC},
+{0x02F81A, 0x02F81A, 0x0051AC},
+{0x02F81B, 0x02F81B, 0x0051B5},
+{0x02F81C, 0x02F81C, 0x0291DF},
+{0x02F81D, 0x02F81D, 0x0051F5},
+{0x02F81E, 0x02F81E, 0x005203},
+{0x02F81F, 0x02F81F, 0x0034DF},
+{0x02F820, 0x02F820, 0x00523B},
+{0x02F821, 0x02F821, 0x005246},
+{0x02F822, 0x02F822, 0x005272},
+{0x02F823, 0x02F823, 0x005277},
+{0x02F824, 0x02F824, 0x003515},
+{0x02F825, 0x02F825, 0x0052C7},
+{0x02F826, 0x02F826, 0x0052C9},
+{0x02F827, 0x02F827, 0x0052E4},
+{0x02F828, 0x02F828, 0x0052FA},
+{0x02F829, 0x02F829, 0x005305},
+{0x02F82A, 0x02F82A, 0x005306},
+{0x02F82B, 0x02F82B, 0x005317},
+{0x02F82C, 0x02F82C, 0x005349},
+{0x02F82D, 0x02F82D, 0x005351},
+{0x02F82E, 0x02F82E, 0x00535A},
+{0x02F82F, 0x02F82F, 0x005373},
+{0x02F830, 0x02F830, 0x00537D},
+{0x02F831, 0x02F833, 0x00537F},
+{0x02F834, 0x02F834, 0x020A2C},
+{0x02F835, 0x02F835, 0x007070},
+{0x02F836, 0x02F836, 0x0053CA},
+{0x02F837, 0x02F837, 0x0053DF},
+{0x02F838, 0x02F838, 0x020B63},
+{0x02F839, 0x02F839, 0x0053EB},
+{0x02F83A, 0x02F83A, 0x0053F1},
+{0x02F83B, 0x02F83B, 0x005406},
+{0x02F83C, 0x02F83C, 0x00549E},
+{0x02F83D, 0x02F83D, 0x005438},
+{0x02F83E, 0x02F83E, 0x005448},
+{0x02F83F, 0x02F83F, 0x005468},
+{0x02F840, 0x02F840, 0x0054A2},
+{0x02F841, 0x02F841, 0x0054F6},
+{0x02F842, 0x02F842, 0x005510},
+{0x02F843, 0x02F843, 0x005553},
+{0x02F844, 0x02F844, 0x005563},
+{0x02F845, 0x02F846, 0x005584},
+{0x02F847, 0x02F847, 0x005599},
+{0x02F848, 0x02F848, 0x0055AB},
+{0x02F849, 0x02F849, 0x0055B3},
+{0x02F84A, 0x02F84A, 0x0055C2},
+{0x02F84B, 0x02F84B, 0x005716},
+{0x02F84C, 0x02F84C, 0x005606},
+{0x02F84D, 0x02F84D, 0x005717},
+{0x02F84E, 0x02F84E, 0x005651},
+{0x02F84F, 0x02F84F, 0x005674},
+{0x02F850, 0x02F850, 0x005207},
+{0x02F851, 0x02F851, 0x0058EE},
+{0x02F852, 0x02F852, 0x0057CE},
+{0x02F853, 0x02F853, 0x0057F4},
+{0x02F854, 0x02F854, 0x00580D},
+{0x02F855, 0x02F855, 0x00578B},
+{0x02F856, 0x02F856, 0x005832},
+{0x02F857, 0x02F857, 0x005831},
+{0x02F858, 0x02F858, 0x0058AC},
+{0x02F859, 0x02F859, 0x0214E4},
+{0x02F85A, 0x02F85A, 0x0058F2},
+{0x02F85B, 0x02F85B, 0x0058F7},
+{0x02F85C, 0x02F85C, 0x005906},
+{0x02F85D, 0x02F85D, 0x00591A},
+{0x02F85E, 0x02F85E, 0x005922},
+{0x02F85F, 0x02F85F, 0x005962},
+{0x02F860, 0x02F860, 0x0216A8},
+{0x02F861, 0x02F861, 0x0216EA},
+{0x02F862, 0x02F862, 0x0059EC},
+{0x02F863, 0x02F863, 0x005A1B},
+{0x02F864, 0x02F864, 0x005A27},
+{0x02F865, 0x02F865, 0x0059D8},
+{0x02F866, 0x02F866, 0x005A66},
+{0x02F867, 0x02F867, 0x0036EE},
+{0x02F868, 0x02F868, 0x0036FC},
+{0x02F869, 0x02F869, 0x005B08},
+{0x02F86A, 0x02F86B, 0x005B3E},
+{0x02F86C, 0x02F86C, 0x0219C8},
+{0x02F86D, 0x02F86D, 0x005BC3},
+{0x02F86E, 0x02F86E, 0x005BD8},
+{0x02F86F, 0x02F86F, 0x005BE7},
+{0x02F870, 0x02F870, 0x005BF3},
+{0x02F871, 0x02F871, 0x021B18},
+{0x02F872, 0x02F872, 0x005BFF},
+{0x02F873, 0x02F873, 0x005C06},
+{0x02F874, 0x02F874, 0x005F53},
+{0x02F875, 0x02F875, 0x005C22},
+{0x02F876, 0x02F876, 0x003781},
+{0x02F877, 0x02F877, 0x005C60},
+{0x02F878, 0x02F878, 0x005C6E},
+{0x02F879, 0x02F879, 0x005CC0},
+{0x02F87A, 0x02F87A, 0x005C8D},
+{0x02F87B, 0x02F87B, 0x021DE4},
+{0x02F87C, 0x02F87C, 0x005D43},
+{0x02F87D, 0x02F87D, 0x021DE6},
+{0x02F87E, 0x02F87E, 0x005D6E},
+{0x02F87F, 0x02F87F, 0x005D6B},
+{0x02F880, 0x02F880, 0x005D7C},
+{0x02F881, 0x02F881, 0x005DE1},
+{0x02F882, 0x02F882, 0x005DE2},
+{0x02F883, 0x02F883, 0x00382F},
+{0x02F884, 0x02F884, 0x005DFD},
+{0x02F885, 0x02F885, 0x005E28},
+{0x02F886, 0x02F886, 0x005E3D},
+{0x02F887, 0x02F887, 0x005E69},
+{0x02F888, 0x02F888, 0x003862},
+{0x02F889, 0x02F889, 0x022183},
+{0x02F88A, 0x02F88A, 0x00387C},
+{0x02F88B, 0x02F88B, 0x005EB0},
+{0x02F88C, 0x02F88C, 0x005EB3},
+{0x02F88D, 0x02F88D, 0x005EB6},
+{0x02F88E, 0x02F88E, 0x005ECA},
+{0x02F88F, 0x02F88F, 0x02A392},
+{0x02F890, 0x02F890, 0x005EFE},
+{0x02F891, 0x02F892, 0x022331},
+{0x02F893, 0x02F893, 0x008201},
+{0x02F894, 0x02F895, 0x005F22},
+{0x02F896, 0x02F896, 0x0038C7},
+{0x02F897, 0x02F897, 0x0232B8},
+{0x02F898, 0x02F898, 0x0261DA},
+{0x02F899, 0x02F899, 0x005F62},
+{0x02F89A, 0x02F89A, 0x005F6B},
+{0x02F89B, 0x02F89B, 0x0038E3},
+{0x02F89C, 0x02F89C, 0x005F9A},
+{0x02F89D, 0x02F89D, 0x005FCD},
+{0x02F89E, 0x02F89E, 0x005FD7},
+{0x02F89F, 0x02F89F, 0x005FF9},
+{0x02F8A0, 0x02F8A0, 0x006081},
+{0x02F8A1, 0x02F8A1, 0x00393A},
+{0x02F8A2, 0x02F8A2, 0x00391C},
+{0x02F8A3, 0x02F8A3, 0x006094},
+{0x02F8A4, 0x02F8A4, 0x0226D4},
+{0x02F8A5, 0x02F8A5, 0x0060C7},
+{0x02F8A6, 0x02F8A6, 0x006148},
+{0x02F8A7, 0x02F8A7, 0x00614C},
+{0x02F8A8, 0x02F8A8, 0x00614E},
+{0x02F8A9, 0x02F8A9, 0x00614C},
+{0x02F8AA, 0x02F8AA, 0x00617A},
+{0x02F8AB, 0x02F8AB, 0x00618E},
+{0x02F8AC, 0x02F8AC, 0x0061B2},
+{0x02F8AD, 0x02F8AD, 0x0061A4},
+{0x02F8AE, 0x02F8AE, 0x0061AF},
+{0x02F8AF, 0x02F8AF, 0x0061DE},
+{0x02F8B0, 0x02F8B0, 0x0061F2},
+{0x02F8B1, 0x02F8B1, 0x0061F6},
+{0x02F8B2, 0x02F8B2, 0x006210},
+{0x02F8B3, 0x02F8B3, 0x00621B},
+{0x02F8B4, 0x02F8B4, 0x00625D},
+{0x02F8B5, 0x02F8B5, 0x0062B1},
+{0x02F8B6, 0x02F8B6, 0x0062D4},
+{0x02F8B7, 0x02F8B7, 0x006350},
+{0x02F8B8, 0x02F8B8, 0x022B0C},
+{0x02F8B9, 0x02F8B9, 0x00633D},
+{0x02F8BA, 0x02F8BA, 0x0062FC},
+{0x02F8BB, 0x02F8BB, 0x006368},
+{0x02F8BC, 0x02F8BC, 0x006383},
+{0x02F8BD, 0x02F8BD, 0x0063E4},
+{0x02F8BE, 0x02F8BE, 0x022BF1},
+{0x02F8BF, 0x02F8BF, 0x006422},
+{0x02F8C0, 0x02F8C0, 0x0063C5},
+{0x02F8C1, 0x02F8C1, 0x0063A9},
+{0x02F8C2, 0x02F8C2, 0x003A2E},
+{0x02F8C3, 0x02F8C3, 0x006469},
+{0x02F8C4, 0x02F8C4, 0x00647E},
+{0x02F8C5, 0x02F8C5, 0x00649D},
+{0x02F8C6, 0x02F8C6, 0x006477},
+{0x02F8C7, 0x02F8C7, 0x003A6C},
+{0x02F8C8, 0x02F8C8, 0x00654F},
+{0x02F8C9, 0x02F8C9, 0x00656C},
+{0x02F8CA, 0x02F8CA, 0x02300A},
+{0x02F8CB, 0x02F8CB, 0x0065E3},
+{0x02F8CC, 0x02F8CC, 0x0066F8},
+{0x02F8CD, 0x02F8CD, 0x006649},
+{0x02F8CE, 0x02F8CE, 0x003B19},
+{0x02F8CF, 0x02F8CF, 0x006691},
+{0x02F8D0, 0x02F8D0, 0x003B08},
+{0x02F8D1, 0x02F8D1, 0x003AE4},
+{0x02F8D2, 0x02F8D2, 0x005192},
+{0x02F8D3, 0x02F8D3, 0x005195},
+{0x02F8D4, 0x02F8D4, 0x006700},
+{0x02F8D5, 0x02F8D5, 0x00669C},
+{0x02F8D6, 0x02F8D6, 0x0080AD},
+{0x02F8D7, 0x02F8D7, 0x0043D9},
+{0x02F8D8, 0x02F8D8, 0x006717},
+{0x02F8D9, 0x02F8D9, 0x00671B},
+{0x02F8DA, 0x02F8DA, 0x006721},
+{0x02F8DB, 0x02F8DB, 0x00675E},
+{0x02F8DC, 0x02F8DC, 0x006753},
+{0x02F8DD, 0x02F8DD, 0x0233C3},
+{0x02F8DE, 0x02F8DE, 0x003B49},
+{0x02F8DF, 0x02F8DF, 0x0067FA},
+{0x02F8E0, 0x02F8E0, 0x006785},
+{0x02F8E1, 0x02F8E1, 0x006852},
+{0x02F8E2, 0x02F8E2, 0x006885},
+{0x02F8E3, 0x02F8E3, 0x02346D},
+{0x02F8E4, 0x02F8E4, 0x00688E},
+{0x02F8E5, 0x02F8E5, 0x00681F},
+{0x02F8E6, 0x02F8E6, 0x006914},
+{0x02F8E7, 0x02F8E7, 0x003B9D},
+{0x02F8E8, 0x02F8E8, 0x006942},
+{0x02F8E9, 0x02F8E9, 0x0069A3},
+{0x02F8EA, 0x02F8EA, 0x0069EA},
+{0x02F8EB, 0x02F8EB, 0x006AA8},
+{0x02F8EC, 0x02F8EC, 0x0236A3},
+{0x02F8ED, 0x02F8ED, 0x006ADB},
+{0x02F8EE, 0x02F8EE, 0x003C18},
+{0x02F8EF, 0x02F8EF, 0x006B21},
+{0x02F8F0, 0x02F8F0, 0x0238A7},
+{0x02F8F1, 0x02F8F1, 0x006B54},
+{0x02F8F2, 0x02F8F2, 0x003C4E},
+{0x02F8F3, 0x02F8F3, 0x006B72},
+{0x02F8F4, 0x02F8F4, 0x006B9F},
+{0x02F8F5, 0x02F8F5, 0x006BBA},
+{0x02F8F6, 0x02F8F6, 0x006BBB},
+{0x02F8F7, 0x02F8F7, 0x023A8D},
+{0x02F8F8, 0x02F8F8, 0x021D0B},
+{0x02F8F9, 0x02F8F9, 0x023AFA},
+{0x02F8FA, 0x02F8FA, 0x006C4E},
+{0x02F8FB, 0x02F8FB, 0x023CBC},
+{0x02F8FC, 0x02F8FC, 0x006CBF},
+{0x02F8FD, 0x02F8FD, 0x006CCD},
+{0x02F8FE, 0x02F8FE, 0x006C67},
+{0x02F8FF, 0x02F8FF, 0x006D16},
+{0x02F900, 0x02F900, 0x006D3E},
+{0x02F901, 0x02F901, 0x006D77},
+{0x02F902, 0x02F902, 0x006D41},
+{0x02F903, 0x02F903, 0x006D69},
+{0x02F904, 0x02F904, 0x006D78},
+{0x02F905, 0x02F905, 0x006D85},
+{0x02F906, 0x02F906, 0x023D1E},
+{0x02F907, 0x02F907, 0x006D34},
+{0x02F908, 0x02F908, 0x006E2F},
+{0x02F909, 0x02F909, 0x006E6E},
+{0x02F90A, 0x02F90A, 0x003D33},
+{0x02F90B, 0x02F90B, 0x006ECB},
+{0x02F90C, 0x02F90C, 0x006EC7},
+{0x02F90D, 0x02F90D, 0x023ED1},
+{0x02F90E, 0x02F90E, 0x006DF9},
+{0x02F90F, 0x02F90F, 0x006F6E},
+{0x02F910, 0x02F910, 0x023F5E},
+{0x02F911, 0x02F911, 0x023F8E},
+{0x02F912, 0x02F912, 0x006FC6},
+{0x02F913, 0x02F913, 0x007039},
+{0x02F914, 0x02F914, 0x00701E},
+{0x02F915, 0x02F915, 0x00701B},
+{0x02F916, 0x02F916, 0x003D96},
+{0x02F917, 0x02F917, 0x00704A},
+{0x02F918, 0x02F918, 0x00707D},
+{0x02F919, 0x02F919, 0x007077},
+{0x02F91A, 0x02F91A, 0x0070AD},
+{0x02F91B, 0x02F91B, 0x020525},
+{0x02F91C, 0x02F91C, 0x007145},
+{0x02F91D, 0x02F91D, 0x024263},
+{0x02F91E, 0x02F91E, 0x00719C},
+{0x02F91F, 0x02F91F, 0x0243AB},
+{0x02F920, 0x02F920, 0x007228},
+{0x02F921, 0x02F921, 0x007235},
+{0x02F922, 0x02F922, 0x007250},
+{0x02F923, 0x02F923, 0x024608},
+{0x02F924, 0x02F924, 0x007280},
+{0x02F925, 0x02F925, 0x007295},
+{0x02F926, 0x02F926, 0x024735},
+{0x02F927, 0x02F927, 0x024814},
+{0x02F928, 0x02F928, 0x00737A},
+{0x02F929, 0x02F929, 0x00738B},
+{0x02F92A, 0x02F92A, 0x003EAC},
+{0x02F92B, 0x02F92B, 0x0073A5},
+{0x02F92C, 0x02F92D, 0x003EB8},
+{0x02F92E, 0x02F92E, 0x007447},
+{0x02F92F, 0x02F92F, 0x00745C},
+{0x02F930, 0x02F930, 0x007471},
+{0x02F931, 0x02F931, 0x007485},
+{0x02F932, 0x02F932, 0x0074CA},
+{0x02F933, 0x02F933, 0x003F1B},
+{0x02F934, 0x02F934, 0x007524},
+{0x02F935, 0x02F935, 0x024C36},
+{0x02F936, 0x02F936, 0x00753E},
+{0x02F937, 0x02F937, 0x024C92},
+{0x02F938, 0x02F938, 0x007570},
+{0x02F939, 0x02F939, 0x02219F},
+{0x02F93A, 0x02F93A, 0x007610},
+{0x02F93B, 0x02F93B, 0x024FA1},
+{0x02F93C, 0x02F93C, 0x024FB8},
+{0x02F93D, 0x02F93D, 0x025044},
+{0x02F93E, 0x02F93E, 0x003FFC},
+{0x02F93F, 0x02F93F, 0x004008},
+{0x02F940, 0x02F940, 0x0076F4},
+{0x02F941, 0x02F941, 0x0250F3},
+{0x02F942, 0x02F942, 0x0250F2},
+{0x02F943, 0x02F943, 0x025119},
+{0x02F944, 0x02F944, 0x025133},
+{0x02F945, 0x02F945, 0x00771E},
+{0x02F946, 0x02F947, 0x00771F},
+{0x02F948, 0x02F948, 0x00774A},
+{0x02F949, 0x02F949, 0x004039},
+{0x02F94A, 0x02F94A, 0x00778B},
+{0x02F94B, 0x02F94B, 0x004046},
+{0x02F94C, 0x02F94C, 0x004096},
+{0x02F94D, 0x02F94D, 0x02541D},
+{0x02F94E, 0x02F94E, 0x00784E},
+{0x02F94F, 0x02F94F, 0x00788C},
+{0x02F950, 0x02F950, 0x0078CC},
+{0x02F951, 0x02F951, 0x0040E3},
+{0x02F952, 0x02F952, 0x025626},
+{0x02F953, 0x02F953, 0x007956},
+{0x02F954, 0x02F954, 0x02569A},
+{0x02F955, 0x02F955, 0x0256C5},
+{0x02F956, 0x02F956, 0x00798F},
+{0x02F957, 0x02F957, 0x0079EB},
+{0x02F958, 0x02F958, 0x00412F},
+{0x02F959, 0x02F959, 0x007A40},
+{0x02F95A, 0x02F95A, 0x007A4A},
+{0x02F95B, 0x02F95B, 0x007A4F},
+{0x02F95C, 0x02F95C, 0x02597C},
+{0x02F95D, 0x02F95E, 0x025AA7},
+{0x02F95F, 0x02F95F, 0x007AEE},
+{0x02F960, 0x02F960, 0x004202},
+{0x02F961, 0x02F961, 0x025BAB},
+{0x02F962, 0x02F962, 0x007BC6},
+{0x02F963, 0x02F963, 0x007BC9},
+{0x02F964, 0x02F964, 0x004227},
+{0x02F965, 0x02F965, 0x025C80},
+{0x02F966, 0x02F966, 0x007CD2},
+{0x02F967, 0x02F967, 0x0042A0},
+{0x02F968, 0x02F968, 0x007CE8},
+{0x02F969, 0x02F969, 0x007CE3},
+{0x02F96A, 0x02F96A, 0x007D00},
+{0x02F96B, 0x02F96B, 0x025F86},
+{0x02F96C, 0x02F96C, 0x007D63},
+{0x02F96D, 0x02F96D, 0x004301},
+{0x02F96E, 0x02F96E, 0x007DC7},
+{0x02F96F, 0x02F96F, 0x007E02},
+{0x02F970, 0x02F970, 0x007E45},
+{0x02F971, 0x02F971, 0x004334},
+{0x02F972, 0x02F972, 0x026228},
+{0x02F973, 0x02F973, 0x026247},
+{0x02F974, 0x02F974, 0x004359},
+{0x02F975, 0x02F975, 0x0262D9},
+{0x02F976, 0x02F976, 0x007F7A},
+{0x02F977, 0x02F977, 0x02633E},
+{0x02F978, 0x02F978, 0x007F95},
+{0x02F979, 0x02F979, 0x007FFA},
+{0x02F97A, 0x02F97A, 0x008005},
+{0x02F97B, 0x02F97B, 0x0264DA},
+{0x02F97C, 0x02F97C, 0x026523},
+{0x02F97D, 0x02F97D, 0x008060},
+{0x02F97E, 0x02F97E, 0x0265A8},
+{0x02F97F, 0x02F97F, 0x008070},
+{0x02F980, 0x02F980, 0x02335F},
+{0x02F981, 0x02F981, 0x0043D5},
+{0x02F982, 0x02F982, 0x0080B2},
+{0x02F983, 0x02F983, 0x008103},
+{0x02F984, 0x02F984, 0x00440B},
+{0x02F985, 0x02F985, 0x00813E},
+{0x02F986, 0x02F986, 0x005AB5},
+{0x02F987, 0x02F987, 0x0267A7},
+{0x02F988, 0x02F988, 0x0267B5},
+{0x02F989, 0x02F989, 0x023393},
+{0x02F98A, 0x02F98A, 0x02339C},
+{0x02F98B, 0x02F98B, 0x008201},
+{0x02F98C, 0x02F98C, 0x008204},
+{0x02F98D, 0x02F98D, 0x008F9E},
+{0x02F98E, 0x02F98E, 0x00446B},
+{0x02F98F, 0x02F98F, 0x008291},
+{0x02F990, 0x02F990, 0x00828B},
+{0x02F991, 0x02F991, 0x00829D},
+{0x02F992, 0x02F992, 0x0052B3},
+{0x02F993, 0x02F993, 0x0082B1},
+{0x02F994, 0x02F994, 0x0082B3},
+{0x02F995, 0x02F995, 0x0082BD},
+{0x02F996, 0x02F996, 0x0082E6},
+{0x02F997, 0x02F997, 0x026B3C},
+{0x02F998, 0x02F998, 0x0082E5},
+{0x02F999, 0x02F999, 0x00831D},
+{0x02F99A, 0x02F99A, 0x008363},
+{0x02F99B, 0x02F99B, 0x0083AD},
+{0x02F99C, 0x02F99C, 0x008323},
+{0x02F99D, 0x02F99D, 0x0083BD},
+{0x02F99E, 0x02F99E, 0x0083E7},
+{0x02F99F, 0x02F99F, 0x008457},
+{0x02F9A0, 0x02F9A0, 0x008353},
+{0x02F9A1, 0x02F9A1, 0x0083CA},
+{0x02F9A2, 0x02F9A2, 0x0083CC},
+{0x02F9A3, 0x02F9A3, 0x0083DC},
+{0x02F9A4, 0x02F9A4, 0x026C36},
+{0x02F9A5, 0x02F9A5, 0x026D6B},
+{0x02F9A6, 0x02F9A6, 0x026CD5},
+{0x02F9A7, 0x02F9A7, 0x00452B},
+{0x02F9A8, 0x02F9A8, 0x0084F1},
+{0x02F9A9, 0x02F9A9, 0x0084F3},
+{0x02F9AA, 0x02F9AA, 0x008516},
+{0x02F9AB, 0x02F9AB, 0x0273CA},
+{0x02F9AC, 0x02F9AC, 0x008564},
+{0x02F9AD, 0x02F9AD, 0x026F2C},
+{0x02F9AE, 0x02F9AE, 0x00455D},
+{0x02F9AF, 0x02F9AF, 0x004561},
+{0x02F9B0, 0x02F9B0, 0x026FB1},
+{0x02F9B1, 0x02F9B1, 0x0270D2},
+{0x02F9B2, 0x02F9B2, 0x00456B},
+{0x02F9B3, 0x02F9B3, 0x008650},
+{0x02F9B4, 0x02F9B4, 0x00865C},
+{0x02F9B5, 0x02F9B5, 0x008667},
+{0x02F9B6, 0x02F9B6, 0x008669},
+{0x02F9B7, 0x02F9B7, 0x0086A9},
+{0x02F9B8, 0x02F9B8, 0x008688},
+{0x02F9B9, 0x02F9B9, 0x00870E},
+{0x02F9BA, 0x02F9BA, 0x0086E2},
+{0x02F9BB, 0x02F9BB, 0x008779},
+{0x02F9BC, 0x02F9BC, 0x008728},
+{0x02F9BD, 0x02F9BD, 0x00876B},
+{0x02F9BE, 0x02F9BE, 0x008786},
+{0x02F9BF, 0x02F9BF, 0x0045D7},
+{0x02F9C0, 0x02F9C0, 0x0087E1},
+{0x02F9C1, 0x02F9C1, 0x008801},
+{0x02F9C2, 0x02F9C2, 0x0045F9},
+{0x02F9C3, 0x02F9C3, 0x008860},
+{0x02F9C4, 0x02F9C4, 0x008863},
+{0x02F9C5, 0x02F9C5, 0x027667},
+{0x02F9C6, 0x02F9C6, 0x0088D7},
+{0x02F9C7, 0x02F9C7, 0x0088DE},
+{0x02F9C8, 0x02F9C8, 0x004635},
+{0x02F9C9, 0x02F9C9, 0x0088FA},
+{0x02F9CA, 0x02F9CA, 0x0034BB},
+{0x02F9CB, 0x02F9CB, 0x0278AE},
+{0x02F9CC, 0x02F9CC, 0x027966},
+{0x02F9CD, 0x02F9CD, 0x0046BE},
+{0x02F9CE, 0x02F9CE, 0x0046C7},
+{0x02F9CF, 0x02F9CF, 0x008AA0},
+{0x02F9D0, 0x02F9D0, 0x008AED},
+{0x02F9D1, 0x02F9D1, 0x008B8A},
+{0x02F9D2, 0x02F9D2, 0x008C55},
+{0x02F9D3, 0x02F9D3, 0x027CA8},
+{0x02F9D4, 0x02F9D4, 0x008CAB},
+{0x02F9D5, 0x02F9D5, 0x008CC1},
+{0x02F9D6, 0x02F9D6, 0x008D1B},
+{0x02F9D7, 0x02F9D7, 0x008D77},
+{0x02F9D8, 0x02F9D8, 0x027F2F},
+{0x02F9D9, 0x02F9D9, 0x020804},
+{0x02F9DA, 0x02F9DA, 0x008DCB},
+{0x02F9DB, 0x02F9DB, 0x008DBC},
+{0x02F9DC, 0x02F9DC, 0x008DF0},
+{0x02F9DD, 0x02F9DD, 0x0208DE},
+{0x02F9DE, 0x02F9DE, 0x008ED4},
+{0x02F9DF, 0x02F9DF, 0x008F38},
+{0x02F9E0, 0x02F9E0, 0x0285D2},
+{0x02F9E1, 0x02F9E1, 0x0285ED},
+{0x02F9E2, 0x02F9E2, 0x009094},
+{0x02F9E3, 0x02F9E3, 0x0090F1},
+{0x02F9E4, 0x02F9E4, 0x009111},
+{0x02F9E5, 0x02F9E5, 0x02872E},
+{0x02F9E6, 0x02F9E6, 0x00911B},
+{0x02F9E7, 0x02F9E7, 0x009238},
+{0x02F9E8, 0x02F9E8, 0x0092D7},
+{0x02F9E9, 0x02F9E9, 0x0092D8},
+{0x02F9EA, 0x02F9EA, 0x00927C},
+{0x02F9EB, 0x02F9EB, 0x0093F9},
+{0x02F9EC, 0x02F9EC, 0x009415},
+{0x02F9ED, 0x02F9ED, 0x028BFA},
+{0x02F9EE, 0x02F9EE, 0x00958B},
+{0x02F9EF, 0x02F9EF, 0x004995},
+{0x02F9F0, 0x02F9F0, 0x0095B7},
+{0x02F9F1, 0x02F9F1, 0x028D77},
+{0x02F9F2, 0x02F9F2, 0x0049E6},
+{0x02F9F3, 0x02F9F3, 0x0096C3},
+{0x02F9F4, 0x02F9F4, 0x005DB2},
+{0x02F9F5, 0x02F9F5, 0x009723},
+{0x02F9F6, 0x02F9F6, 0x029145},
+{0x02F9F7, 0x02F9F7, 0x02921A},
+{0x02F9F8, 0x02F9F8, 0x004A6E},
+{0x02F9F9, 0x02F9F9, 0x004A76},
+{0x02F9FA, 0x02F9FA, 0x0097E0},
+{0x02F9FB, 0x02F9FB, 0x02940A},
+{0x02F9FC, 0x02F9FC, 0x004AB2},
+{0x02F9FD, 0x02F9FD, 0x029496},
+{0x02F9FE, 0x02F9FF, 0x00980B},
+{0x02FA00, 0x02FA00, 0x009829},
+{0x02FA01, 0x02FA01, 0x0295B6},
+{0x02FA02, 0x02FA02, 0x0098E2},
+{0x02FA03, 0x02FA03, 0x004B33},
+{0x02FA04, 0x02FA04, 0x009929},
+{0x02FA05, 0x02FA05, 0x0099A7},
+{0x02FA06, 0x02FA06, 0x0099C2},
+{0x02FA07, 0x02FA07, 0x0099FE},
+{0x02FA08, 0x02FA08, 0x004BCE},
+{0x02FA09, 0x02FA09, 0x029B30},
+{0x02FA0A, 0x02FA0A, 0x009B12},
+{0x02FA0B, 0x02FA0B, 0x009C40},
+{0x02FA0C, 0x02FA0C, 0x009CFD},
+{0x02FA0D, 0x02FA0D, 0x004CCE},
+{0x02FA0E, 0x02FA0E, 0x004CED},
+{0x02FA0F, 0x02FA0F, 0x009D67},
+{0x02FA10, 0x02FA10, 0x02A0CE},
+{0x02FA11, 0x02FA11, 0x004CF8},
+{0x02FA12, 0x02FA12, 0x02A105},
+{0x02FA13, 0x02FA13, 0x02A20E},
+{0x02FA14, 0x02FA14, 0x02A291},
+{0x02FA15, 0x02FA15, 0x009EBB},
+{0x02FA16, 0x02FA16, 0x004D56},
+{0x02FA17, 0x02FA17, 0x009EF9},
+{0x02FA18, 0x02FA18, 0x009EFE},
+{0x02FA19, 0x02FA19, 0x009F05},
+{0x02FA1A, 0x02FA1A, 0x009F0F},
+{0x02FA1B, 0x02FA1B, 0x009F16},
+{0x02FA1C, 0x02FA1C, 0x009F3B},
+{0x02FA1D, 0x02FA1D, 0x02A600},
+};
diff --git a/src/unicode-data.h b/src/unicode-data.h
new file mode 100644
index 00000000..e27fe177
--- /dev/null
+++ b/src/unicode-data.h
@@ -0,0 +1,20 @@
+#pragma once
+
+#include 
+#include 
+#include 
+#include 
+
+struct range_nfd {
+    uint32_t first;
+    uint32_t last;
+    uint32_t nfd;
+};
+
+static const uint32_t MAX_CODEPOINTS = 0x110000;
+
+extern const std::vector> unicode_ranges_flags;
+extern const std::unordered_set unicode_set_whitespace;
+extern const std::unordered_map unicode_map_lowercase;
+extern const std::unordered_map unicode_map_uppercase;
+extern const std::vector unicode_ranges_nfd;
diff --git a/src/unicode.cpp b/src/unicode.cpp
new file mode 100644
index 00000000..46650bff
--- /dev/null
+++ b/src/unicode.cpp
@@ -0,0 +1,818 @@
+#if defined(_MSC_VER)
+#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
+#endif
+
+#include "unicode.h"
+#include "unicode-data.h"
+
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+#include 
+
+size_t unicode_len_utf8(char src) {
+    const size_t lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
+    uint8_t highbits = static_cast(src) >> 4;
+    return lookup[highbits];
+}
+
+static std::string unicode_cpts_to_utf8(const std::vector & cps) {
+    std::string result;
+    for (size_t i = 0; i < cps.size(); ++i) {
+        result.append(unicode_cpt_to_utf8(cps[i]));
+    }
+    return result;
+}
+
+uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset) {
+    assert(offset < utf8.size());
+    if (!(utf8[offset + 0] & 0x80)) {
+        auto result = utf8[offset + 0];
+        offset += 1;
+        return result;
+    }
+    if (!(utf8[offset + 0] & 0x40)) {
+        throw std::invalid_argument("invalid character");
+    }
+    if (!(utf8[offset + 0] & 0x20)) {
+        if (offset + 1 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80)) {
+            throw std::invalid_argument("invalid character");
+        }
+        auto result = ((utf8[offset + 0] & 0x1f) << 6) | (utf8[offset + 1] & 0x3f);
+        offset += 2;
+        return result;
+    }
+    if (!(utf8[offset + 0] & 0x10)) {
+        if (offset + 2 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80)) {
+            throw std::invalid_argument("invalid character");
+        }
+        auto result = ((utf8[offset + 0] & 0x0f) << 12) | ((utf8[offset + 1] & 0x3f) << 6) | (utf8[offset + 2] & 0x3f);
+        offset += 3;
+        return result;
+    }
+    if (!(utf8[offset + 0] & 0x08)) {
+        if (offset + 3 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80) || !((utf8[offset + 3] & 0xc0) == 0x80)) {
+            throw std::invalid_argument("invalid character");
+        }
+        auto result = ((utf8[offset + 0] & 0x07) << 18) | ((utf8[offset + 1] & 0x3f) << 12) | ((utf8[offset + 2] & 0x3f) << 6) | (utf8[offset + 3] & 0x3f);
+        offset += 4;
+        return result;
+    }
+    throw std::invalid_argument("failed to convert utf8 to codepoint");
+}
+
+//static std::vector unicode_cpt_to_utf16(uint32_t cp) {
+//    std::vector result;
+//    if (/* 0x0000 <= cp && */ cp <= 0xffff) {
+//        result.emplace_back(cp);
+//        return result;
+//    }
+//    if (0x10000 <= cp && cp <= 0x10ffff) {
+//        result.emplace_back(0xd800 | ((cp - 0x10000) >> 10));
+//        result.emplace_back(0xdc00 | ((cp - 0x10000) & 0x03ff));
+//        return result;
+//    }
+//    throw std::invalid_argument("failed to convert codepoint to utf16");
+//}
+
+//static std::vector unicode_cpts_to_utf16(const std::vector & cps) {
+//    std::vector result;
+//    for (size_t i = 0; i < cps.size(); ++i) {
+//        auto temp = unicode_cpt_to_utf16(cps[i]);
+//        result.insert(result.end(), temp.begin(), temp.end());
+//    }
+//    return result;
+//}
+
+//static uint32_t unicode_cpt_from_utf16(const std::vector & utf16, size_t & offset) {
+//    assert(offset < utf16.size());
+//    if (((utf16[0] >> 10) << 10) != 0xd800) {
+//        auto result = utf16[offset + 0];
+//        offset += 1;
+//        return result;
+//    }
+//
+//    if (offset + 1 >= utf16.size() || !((utf16[1] & 0xdc00) == 0xdc00)) {
+//        throw std::invalid_argument("invalid character");
+//    }
+//
+//    auto result = 0x10000 + (((utf16[0] & 0x03ff) << 10) | (utf16[1] & 0x03ff));
+//    offset += 2;
+//    return result;
+//}
+
+//static std::vector unicode_cpts_from_utf16(const std::vector & utf16) {
+//    std::vector result;
+//    size_t offset = 0;
+//    while (offset < utf16.size()) {
+//        result.push_back(unicode_cpt_from_utf16(utf16, offset));
+//    }
+//    return result;
+//}
+
+static std::vector unicode_cpt_flags_array() {
+    std::vector cpt_flags(MAX_CODEPOINTS, codepoint_flags::UNDEFINED);
+
+    assert (unicode_ranges_flags.front().first == 0);
+    assert (unicode_ranges_flags.back().first == MAX_CODEPOINTS);
+    for (size_t i = 1; i < unicode_ranges_flags.size(); ++i) {
+        const auto range_ini = unicode_ranges_flags[i-1];  // codepoint_ini, flags
+        const auto range_end = unicode_ranges_flags[i];    // codepoint_end, flags
+        for (uint32_t cpt = range_ini.first; cpt < range_end.first; ++cpt) {
+            cpt_flags[cpt] = range_ini.second;
+        }
+    }
+
+    for (auto cpt : unicode_set_whitespace) {
+        cpt_flags[cpt].is_whitespace = true;
+    }
+
+    for (auto p : unicode_map_lowercase) {
+        cpt_flags[p.second].is_lowercase = true;
+    }
+
+    for (auto p : unicode_map_uppercase) {
+        cpt_flags[p.second].is_uppercase = true;
+    }
+
+    for (auto &range : unicode_ranges_nfd) {  // start, last, nfd
+        cpt_flags[range.nfd].is_nfd = true;
+    }
+
+    return cpt_flags;
+}
+
+static std::unordered_map unicode_byte_to_utf8_map() {
+    std::unordered_map map;
+    for (int ch = 0x21; ch <= 0x7E; ++ch) {  // u'!' to u'~'
+        assert(0 <= ch && ch < 256);
+        map[ch] = unicode_cpt_to_utf8(ch);
+    }
+    for (int ch = 0xA1; ch <= 0xAC; ++ch) {  // u'¡' to u'¬'
+        assert(0 <= ch && ch < 256);
+        map[ch] = unicode_cpt_to_utf8(ch);
+    }
+    for (int ch = 0xAE; ch <= 0xFF; ++ch) {  // u'®' to u'ÿ'
+        assert(0 <= ch && ch < 256);
+        map[ch] = unicode_cpt_to_utf8(ch);
+    }
+    auto n = 0;
+    for (int ch = 0; ch < 256; ++ch) {
+        if (map.find(ch) == map.end()) {
+            map[ch] = unicode_cpt_to_utf8(256 + n);
+            ++n;
+        }
+    }
+    return map;
+}
+
+static std::unordered_map unicode_utf8_to_byte_map() {
+    std::unordered_map map;
+    for (int ch = 0x21; ch <= 0x7E; ++ch) {  // u'!' to u'~'
+        assert(0 <= ch && ch < 256);
+        map[unicode_cpt_to_utf8(ch)] = ch;
+    }
+    for (int ch = 0xA1; ch <= 0xAC; ++ch) {  // u'¡' to u'¬'
+        assert(0 <= ch && ch < 256);
+        map[unicode_cpt_to_utf8(ch)] = ch;
+    }
+    for (int ch = 0xAE; ch <= 0xFF; ++ch) {  // u'®' to u'ÿ'
+        assert(0 <= ch && ch < 256);
+        map[unicode_cpt_to_utf8(ch)] = ch;
+    }
+    auto n = 0;
+    for (int ch = 0; ch < 256; ++ch) {
+        if (map.find(unicode_cpt_to_utf8(ch)) == map.end()) {
+            map[unicode_cpt_to_utf8(256 + n)] = ch;
+            ++n;
+        }
+    }
+    return map;
+}
+
+static inline std::wstring unicode_wstring_from_utf8(const std::string & s) {
+    std::wstring_convert> conv;
+    return conv.from_bytes(s);
+}
+
+static std::vector unicode_byte_encoding_process(const std::vector & bpe_words) {
+    std::vector bpe_encoded_words;
+    for (const auto & word : bpe_words) {
+        std::string text_utf;
+        auto utf_word =  unicode_cpts_from_utf8(word);
+        for (size_t i = 0; i < utf_word.size(); ++i) {
+            text_utf += unicode_cpt_to_utf8(utf_word[i]);
+        }
+
+        std::string encoded_token;
+        for (char & c : text_utf) {
+            encoded_token += unicode_byte_to_utf8(c);
+        }
+        bpe_encoded_words.emplace_back(encoded_token);
+    }
+    return bpe_encoded_words;
+}
+
+// GPT2 system regex:  's|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+
+static std::vector unicode_regex_split_custom_gpt2(const std::string & text, const std::vector & offsets) {
+    std::vector bpe_offsets; // store the offset of each word
+    bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
+
+    const auto cpts = unicode_cpts_from_utf8(text);
+
+    size_t start = 0;
+    for (auto offset : offsets) {
+        const size_t offset_ini = start;
+        const size_t offset_end = start + offset;
+        assert(offset_end <= cpts.size());
+        start = offset_end;
+
+        static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;
+        auto _get_cpt = [&] (const size_t pos) -> uint32_t {
+            return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
+        };
+
+        auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : codepoint_flags{};
+        };
+
+        size_t _prev_end = offset_ini;
+        auto _add_token = [&] (const size_t end) -> size_t {
+            assert(_prev_end <= end && end <= offset_end);
+            size_t len = end - _prev_end;
+            if (len > 0) {
+                bpe_offsets.push_back(len);
+            }
+            _prev_end = end;
+            //if (len > 0) {
+            //    std::string s = "";
+            //    for(size_t p = end-len; p < end; p++)
+            //        s += unicode_cpt_to_utf8(cpts[p]);
+            //    printf(">>> '%s'\n", s.c_str());
+            //}
+            return len;
+        };
+
+        for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
+            const uint32_t cpt = _get_cpt(pos);
+            const auto flags = _get_flags(pos);
+
+            // regex: 's|'t|'re|'ve|'m|'ll|'d
+            if (cpt == '\'' && pos+1 < offset_end) {
+                uint32_t cpt_next = _get_cpt(pos+1);
+                if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
+                    pos += _add_token(pos+2);
+                    continue;
+                }
+                if (pos+2 < offset_end) {
+                    uint32_t cpt_next_next = _get_cpt(pos+2);
+                    if ((cpt_next == 'r' && cpt_next_next == 'e') ||
+                        (cpt_next == 'v' && cpt_next_next == 'e') ||
+                        (cpt_next == 'l' && cpt_next_next == 'l')) {
+                        pos += _add_token(pos+3);
+                        continue;
+                    }
+                }
+            }
+
+            auto flags2 = (cpt == ' ' ? _get_flags(pos+1) : flags);
+            // regex: ?\p{L}+
+            if (flags2.is_letter) {
+                pos += (cpt == ' ');
+                while (flags2.is_letter) {
+                    flags2 = _get_flags(++pos);
+                }
+                _add_token(pos);
+                continue;
+            }
+            // regex: ?\p{N}+
+            if (flags2.is_number) {
+                pos += (cpt == ' ');
+                while (flags2.is_number) {
+                    flags2 = _get_flags(++pos);
+                }
+                _add_token(pos);
+                continue;
+            }
+            // regex: ?[^\s\p{L}\p{N}]+
+            if (!(flags2.is_whitespace | flags2.is_letter | flags2.is_number) && flags2.as_uint()) {
+                pos += (cpt == ' ');
+                while (!(flags2.is_whitespace | flags2.is_letter | flags2.is_number) && flags2.as_uint()) {
+                    flags2 = _get_flags(++pos);
+                }
+                _add_token(pos);
+                continue;
+            }
+
+            size_t num_whitespaces = 0;
+            while (_get_flags(pos+num_whitespaces).is_whitespace) {
+                num_whitespaces++;
+            }
+
+            // regex: \s+(?!\S)
+            if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != OUT_OF_RANGE) {
+                pos += num_whitespaces - 1;
+                _add_token(pos);
+                continue;
+            }
+
+            // regex: \s+
+            if (num_whitespaces > 0) {
+                pos += num_whitespaces;
+                _add_token(pos);
+                continue;
+            }
+
+            // no matches
+            _add_token(++pos);
+        }
+    }
+
+    return bpe_offsets;
+}
+
+// LLAMA3 system regex: "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"
+static std::vector unicode_regex_split_custom_llama3(const std::string & text, const std::vector & offsets) {
+    std::vector bpe_offsets; // store the offset of each word
+    bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
+
+    const auto cpts = unicode_cpts_from_utf8(text);
+
+    size_t start = 0;
+    for (auto offset : offsets) {
+        const size_t offset_ini = start;
+        const size_t offset_end = start + offset;
+        assert(offset_end <= cpts.size());
+        start = offset_end;
+
+        static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;
+        auto _get_cpt = [&] (const size_t pos) -> uint32_t {
+            return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
+        };
+
+        auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : codepoint_flags{};
+        };
+
+        size_t _prev_end = offset_ini;
+        auto _add_token = [&] (const size_t end) -> size_t {
+            assert(_prev_end <= end && end <= offset_end);
+            size_t len = end - _prev_end;
+            if (len > 0) {
+                bpe_offsets.push_back(len);
+            }
+            _prev_end = end;
+            //if (len > 0) {
+            //    std::string s = "";
+            //    for(size_t p = end-len; p < end; p++)
+            //        s += unicode_cpt_to_utf8(cpts[p]);
+            //    printf(">>> '%s'\n", s.c_str());
+            //}
+            return len;
+        };
+
+        for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
+            const uint32_t cpt = _get_cpt(pos);
+            const auto flags = _get_flags(pos);
+
+            // regex: (?i:'s|'t|'re|'ve|'m|'ll|'d) // case insensitive
+            if (cpt == '\'' && pos+1 < offset_end) {
+                uint32_t cpt_next = unicode_tolower(_get_cpt(pos+1));
+                if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
+                    pos += _add_token(pos+2);
+                    continue;
+                }
+                if (pos+2 < offset_end) {
+                    uint32_t cpt_next_next = unicode_tolower(_get_cpt(pos+2));
+                    if ((cpt_next == 'r' && cpt_next_next == 'e') ||
+                        (cpt_next == 'v' && cpt_next_next == 'e') ||
+                        (cpt_next == 'l' && cpt_next_next == 'l')) {
+                        pos += _add_token(pos+3);
+                        continue;
+                    }
+                }
+            }
+
+            // regex: [^\r\n\p{L}\p{N}]?\p{L}+
+            if (!(cpt == '\r' || cpt == '\n' || flags.is_number)) {
+                if (flags.is_letter || _get_flags(pos+1).is_letter) {  // one or more letters
+                    pos++;
+                    while (_get_flags(pos).is_letter) {
+                        pos++;
+                    }
+                    _add_token(pos);
+                    continue;
+                }
+            }
+
+            // regex: \p{N}{1,3}
+            if (flags.is_number) {
+                size_t ini = pos;
+                while (_get_flags(pos).is_number) {
+                    if (++pos - ini >= 3 ) {
+                        _add_token(pos);
+                        ini = pos;
+                    }
+                }
+                _add_token(pos);
+                continue;
+            }
+
+            // regex: ?[^\s\p{L}\p{N}]+[\r\n]*
+            auto flags2 = (cpt == ' ' ? _get_flags(pos+1) : flags);
+            if (!(flags2.is_whitespace | flags2.is_letter | flags2.is_number) && flags.as_uint()) {
+                pos += (cpt == ' ');
+                while (!(flags2.is_whitespace | flags2.is_letter | flags2.is_number) && flags2.as_uint()) {
+                    flags2 = _get_flags(++pos);
+                }
+                uint32_t cpt2 = _get_cpt(pos);
+                while (cpt2 == '\r' || cpt2 == '\n') {
+                    cpt2 = _get_cpt(++pos);
+                }
+                _add_token(pos);
+                continue;
+            }
+
+            size_t num_whitespaces = 0;
+            size_t last_end_r_or_n = 0;
+            while (_get_flags(pos+num_whitespaces).is_whitespace) {
+                uint32_t cpt2 = _get_cpt(pos+num_whitespaces);
+                if (cpt2 == '\r' || cpt2 == '\n') {
+                    last_end_r_or_n = pos + num_whitespaces + 1;
+                }
+                num_whitespaces++;
+            }
+
+            // regex: \s*[\r\n]+
+            if (last_end_r_or_n > 0) {
+                pos = last_end_r_or_n;
+                _add_token(pos);
+                continue;
+            }
+
+            // regex: \s+(?!\S)
+            if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != OUT_OF_RANGE) {
+                pos += num_whitespaces - 1;
+                _add_token(pos);
+                continue;
+            }
+
+            // regex: \s+
+            if (num_whitespaces > 0) {
+                pos += num_whitespaces;
+                _add_token(pos);
+                continue;
+            }
+
+            // no matches
+            _add_token(++pos);
+        }
+    }
+
+    return bpe_offsets;
+}
+
+// use std::wregex to split the text
+static std::vector unicode_regex_split_stl(const std::wstring & wtext, const std::wstring & regex_expr, const std::vector & offsets) {
+    std::wregex expr(regex_expr);
+    std::vector bpe_offsets; // store the offset of each word
+    bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
+    size_t start = 0;
+    for (auto offset : offsets) {
+        std::wcregex_iterator it(wtext.data() + start, wtext.data() + start + offset, expr);
+        std::wcregex_iterator end;
+
+        int64_t start_idx = 0;
+        while (it != end) {
+            std::wcmatch match = *it;
+            if (match.position() > start_idx) {
+                bpe_offsets.emplace_back(match.position() - start_idx);
+            }
+            bpe_offsets.emplace_back(match.length());
+            start_idx = match.position() + match.length();
+            ++it;
+        }
+
+        if (start_idx < (int64_t) offset) {
+            bpe_offsets.emplace_back(offset - start_idx);
+        }
+        start += offset;
+    }
+
+    return bpe_offsets;
+}
+
+// use std::regex to split the text
+static std::vector unicode_regex_split_stl(const std::string & text, const std::string & regex_expr, const std::vector & offsets) {
+    std::regex expr(regex_expr);
+    std::vector bpe_offsets; // store the offset of each word
+    bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
+    size_t start = 0;
+    for (auto offset : offsets) {
+        std::cregex_iterator it(text.data() + start, text.data() + start + offset, expr);
+        std::cregex_iterator end;
+
+        int64_t start_idx = 0;
+        while (it != end) {
+            std::cmatch match = *it;
+            if (match.position() > start_idx) {
+                bpe_offsets.emplace_back(match.position() - start_idx);
+            }
+            bpe_offsets.emplace_back(match.length());
+            start_idx = match.position() + match.length();
+            ++it;
+        }
+
+        if (start_idx < (int64_t) offset) {
+            bpe_offsets.emplace_back(offset - start_idx);
+        }
+        start += offset;
+    }
+
+    return bpe_offsets;
+}
+
+static std::vector unicode_regex_split_custom(const std::string & text, const std::string & regex_expr, const std::vector & offsets) {
+    std::vector bpe_offsets;
+
+    if (regex_expr == "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)") {
+        bpe_offsets = unicode_regex_split_custom_gpt2(text, offsets);
+    } else if (
+            regex_expr == "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+" ||
+            regex_expr == "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+") {
+
+        bpe_offsets = unicode_regex_split_custom_llama3(text, offsets);
+    }
+
+    return bpe_offsets;
+}
+
+//
+// interface
+//
+
+std::string unicode_cpt_to_utf8(uint32_t cp) {
+    std::string result;
+
+    if (/* 0x00 <= cp && */ cp <= 0x7f) {
+        result.push_back(cp);
+        return result;
+    }
+    if (0x80 <= cp && cp <= 0x7ff) {
+        result.push_back(0xc0 | ((cp >> 6) & 0x1f));
+        result.push_back(0x80 | (cp & 0x3f));
+        return result;
+    }
+    if (0x800 <= cp && cp <= 0xffff) {
+        result.push_back(0xe0 | ((cp >> 12) & 0x0f));
+        result.push_back(0x80 | ((cp >> 6) & 0x3f));
+        result.push_back(0x80 | (cp & 0x3f));
+        return result;
+    }
+    if (0x10000 <= cp && cp <= 0x10ffff) {
+        result.push_back(0xf0 | ((cp >> 18) & 0x07));
+        result.push_back(0x80 | ((cp >> 12) & 0x3f));
+        result.push_back(0x80 | ((cp >> 6) & 0x3f));
+        result.push_back(0x80 | (cp & 0x3f));
+        return result;
+    }
+
+    throw std::invalid_argument("invalid codepoint");
+}
+
+std::vector unicode_cpts_normalize_nfd(const std::vector & cpts) {
+    auto comp = [] (const uint32_t cpt, const range_nfd & range) {
+        return cpt < range.first;
+    };
+    std::vector result(cpts.size());
+    for (size_t i = 0; i < cpts.size(); ++i) {
+        const uint32_t cpt = cpts[i];
+        auto it = std::upper_bound(unicode_ranges_nfd.cbegin(), unicode_ranges_nfd.cend(), cpt, comp) - 1;
+        result[i] = (it->first <= cpt && cpt <= it->last) ? it->nfd : cpt;
+    }
+    return result;
+}
+
+std::vector unicode_cpts_from_utf8(const std::string & utf8) {
+    std::vector result;
+    result.reserve(utf8.size());
+    size_t offset = 0;
+    while (offset < utf8.size()) {
+        result.push_back(unicode_cpt_from_utf8(utf8, offset));
+    }
+    return result;
+}
+
+codepoint_flags unicode_cpt_flags(const uint32_t cp) {
+    static const codepoint_flags undef(codepoint_flags::UNDEFINED);
+    static const auto cpt_flags = unicode_cpt_flags_array();
+    return cp < cpt_flags.size() ? cpt_flags[cp] : undef;
+}
+
+codepoint_flags unicode_cpt_flags(const std::string & utf8) {
+    static const codepoint_flags undef(codepoint_flags::UNDEFINED);
+    if (utf8.empty()) {
+        return undef;  // undefined
+    }
+    size_t offset = 0;
+    return unicode_cpt_flags(unicode_cpt_from_utf8(utf8, offset));
+}
+
+std::string unicode_byte_to_utf8(uint8_t byte) {
+    static std::unordered_map map = unicode_byte_to_utf8_map();
+    return map.at(byte);
+}
+
+uint8_t unicode_utf8_to_byte(const std::string & utf8) {
+    static std::unordered_map map = unicode_utf8_to_byte_map();
+    return map.at(utf8);
+}
+
+uint32_t unicode_tolower(uint32_t cp) {
+    auto it = unicode_map_lowercase.find(cp);
+    return it == unicode_map_lowercase.end() ? cp : it->second;
+}
+
+std::vector unicode_regex_split(const std::string & text, const std::vector & regex_exprs) {
+    // unicode categories
+    static const std::map k_ucat_enum = {
+        { "\\p{N}", codepoint_flags::NUMBER },
+        { "\\p{L}", codepoint_flags::LETTER },
+        { "\\p{P}", codepoint_flags::PUNCTUATION },
+    };
+
+    static const std::map k_ucat_cpt = {
+        { codepoint_flags::NUMBER,        0xD1 },
+        { codepoint_flags::LETTER,        0xD2 },
+        { codepoint_flags::PUNCTUATION,   0xD3 },
+    };
+
+    static const std::map k_ucat_map = {
+        { codepoint_flags::NUMBER,        "\x30-\x39" }, // 0-9
+        { codepoint_flags::LETTER,        "\x41-\x5A\x61-\x7A" }, // A-Za-z
+        { codepoint_flags::PUNCTUATION,   "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
+    };
+
+    // compute collapsed codepoints only if needed by at least one regex
+    bool need_collapse = false;
+    for (auto & regex_expr : regex_exprs) {
+        // search for unicode categories
+        for (const auto & ucat : k_ucat_enum) {
+            if (std::string::npos != regex_expr.find(ucat.first)) {
+                need_collapse = true;
+                break;
+            }
+        }
+    }
+
+    const auto cpts = unicode_cpts_from_utf8(text);
+
+    // generate a "collapsed" representation of the text, where all codepoints are replaced by a single byte
+    // ref: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2081479935
+    std::string text_collapsed;
+    if (need_collapse) {
+        // collapse all unicode categories
+        text_collapsed.resize(cpts.size());
+
+        for (size_t i = 0; i < cpts.size(); ++i) {
+            // keep single-byte codepoints as is
+            if (cpts[i] < 128) {
+                text_collapsed[i] = cpts[i];
+                continue;
+            }
+
+            const auto flags = unicode_cpt_flags(cpts[i]);
+
+            if (flags.is_whitespace) {
+                //NOTE: C++ std::regex \s does not mach 0x85, Rust and Python regex does.
+                //text_collapsed[i] = (char) 0x85;  //  as whitespace fallback
+                text_collapsed[i] = (char) 0x0B;    //  as whitespace fallback
+            } else if (k_ucat_cpt.find(flags.category_flag()) != k_ucat_cpt.end()) {
+                text_collapsed[i] = k_ucat_cpt.at(flags.category_flag());
+            } else {
+                text_collapsed[i] = (char) 0xD0; // fallback
+            }
+        }
+    }
+
+    std::vector bpe_offsets = { cpts.size() };
+
+    for (auto & regex_expr : regex_exprs) {
+        // first, see if we have an efficient custom regex implementation
+        auto tmp = unicode_regex_split_custom(text, regex_expr, bpe_offsets);
+
+        if (!tmp.empty()) {
+            bpe_offsets = std::move(tmp);
+            continue;
+        }
+
+        // fallback to general-purpose std::regex / std::wregex
+        try {
+            // if a unicode category is used in the regex, we use the collapsed text and replace the unicode category
+            // with the corresponding collapsed representation
+            bool use_collapsed = false;
+            for (auto & ucat : k_ucat_enum) {
+                if (std::string::npos != regex_expr.find(ucat.first)) {
+                    use_collapsed = true;
+                    break;
+                }
+            }
+
+            if (use_collapsed) {
+                // sanity-check that the original regex does not contain any non-ASCII characters
+                const auto cpts_regex = unicode_cpts_from_utf8(regex_expr);
+                for (size_t i = 0; i < cpts_regex.size(); ++i) {
+                    if (cpts_regex[i] >= 128) {
+                        throw std::runtime_error("Regex includes both unicode categories and non-ASCII characters - not supported");
+                    }
+                }
+
+                // generate a collapsed representation of the regex
+                std::string regex_expr_collapsed;
+
+                // track if we are inside [], because nested [] are not allowed
+                bool inside = false;
+                for (size_t i = 0; i < regex_expr.size(); ++i) {
+                    if (regex_expr[i] == '[' && (i == 0 || regex_expr[i - 1] != '\\')) {
+                        regex_expr_collapsed += '[';
+                        inside = true;
+                        continue;
+                    }
+
+                    if (inside && regex_expr[i] == ']' && regex_expr[i - 1] != '\\') {
+                        regex_expr_collapsed += ']';
+                        inside = false;
+                        continue;
+                    }
+
+                    if (regex_expr[i + 0] == '\\' && i + 4 < regex_expr.size() &&
+                        regex_expr[i + 1] == 'p' &&
+                        regex_expr[i + 2] == '{' &&
+                        regex_expr[i + 4] == '}') {
+                        const std::string pat = regex_expr.substr(i, 5);
+                        if (k_ucat_enum.find(pat) != k_ucat_enum.end()) {
+                            if (!inside) {
+                                regex_expr_collapsed += '[';
+                            }
+                            regex_expr_collapsed += k_ucat_cpt.at(k_ucat_enum.at(pat));
+                            regex_expr_collapsed += k_ucat_map.at(k_ucat_enum.at(pat));
+                            if (!inside) {
+                                regex_expr_collapsed += ']';
+                            }
+                            i += 4;
+                            continue;
+                        }
+                    }
+
+                    regex_expr_collapsed += regex_expr[i];
+                }
+
+                //printf("text_collapsed: %s\n", text_collapsed.c_str());
+                //printf("regex_expr_collapsed: %s\n", regex_expr_collapsed.c_str());
+                bpe_offsets = unicode_regex_split_stl(text_collapsed, regex_expr_collapsed, bpe_offsets);
+            } else {
+                // no unicode category used, we can use std::wregex directly
+                const std::wstring wregex_expr = unicode_wstring_from_utf8(regex_expr);
+
+                // std::wregex \s does not mach non-ASCII whitespaces, using 0x0B as fallback
+                std::wstring wtext(cpts.begin(), cpts.end());
+                for (size_t i = 0; i < wtext.size(); ++i) {
+                    if (wtext[i] > 0x7F && unicode_cpt_flags(wtext[i]).is_whitespace) {
+                        wtext[i] = 0x0B;
+                    }
+                }
+
+                //printf("text: %s\n", text.c_str());
+                //printf("regex_expr: %s\n", regex_expr.c_str());
+                bpe_offsets = unicode_regex_split_stl(wtext, wregex_expr, bpe_offsets);
+            }
+        } catch (std::regex_error & e) {
+            fprintf(stderr, "Failed to process regex: '%s'\n", regex_expr.c_str());
+            fprintf(stderr, "Regex error: %s\n", e.what());
+            throw std::runtime_error("Failed to process regex");
+        }
+    }
+
+    std::vector bpe_words;
+    bpe_words.reserve(bpe_offsets.size()); // reserve memory for the approximate size
+
+    size_t start = 0;
+    for (size_t & offset : bpe_offsets) {
+        bpe_words.emplace_back();
+        for (size_t i = start; i < start + offset; ++i) {
+            bpe_words.back() += unicode_cpt_to_utf8(cpts[i]);
+        }
+        start += offset;
+    }
+
+    return unicode_byte_encoding_process(bpe_words);
+}
diff --git a/src/unicode.h b/src/unicode.h
new file mode 100644
index 00000000..008532a2
--- /dev/null
+++ b/src/unicode.h
@@ -0,0 +1,67 @@
+#pragma once
+
+#include 
+#include 
+#include 
+
+// TODO: prefix all symbols with "llama_"
+
+struct codepoint_flags {
+    enum {
+        UNDEFINED       = 0x0001,
+        NUMBER          = 0x0002,  // regex: \p{N}
+        LETTER          = 0x0004,  // regex: \p{L}
+        SEPARATOR       = 0x0008,  // regex: \p{Z}
+        ACCENT_MARK     = 0x0010,  // regex: \p{M}
+        PUNCTUATION     = 0x0020,  // regex: \p{P}
+        SYMBOL          = 0x0040,  // regex: \p{S}
+        CONTROL         = 0x0080,  // regex: \p{C}
+        MASK_CATEGORIES = 0x00FF,
+    };
+
+    // codepoint type
+    uint16_t is_undefined   : 1;
+    uint16_t is_number      : 1;  // regex: \p{N}
+    uint16_t is_letter      : 1;  // regex: \p{L}
+    uint16_t is_separator   : 1;  // regex: \p{Z}
+    uint16_t is_accent_mark : 1;  // regex: \p{M}
+    uint16_t is_punctuation : 1;  // regex: \p{P}
+    uint16_t is_symbol      : 1;  // regex: \p{S}
+    uint16_t is_control     : 1;  // regex: \p{C}
+    // helper flags
+    uint16_t is_whitespace  : 1;  // regex: \s
+    uint16_t is_lowercase   : 1;
+    uint16_t is_uppercase   : 1;
+    uint16_t is_nfd         : 1;
+
+    // decode from uint16
+    inline codepoint_flags(const uint16_t flags=0) {
+        *reinterpret_cast(this) = flags;
+    }
+
+    inline uint16_t as_uint() const {
+        return *reinterpret_cast(this);
+    }
+
+    inline uint16_t category_flag() const {
+        return this->as_uint() & MASK_CATEGORIES;
+    }
+};
+
+size_t unicode_len_utf8(char src);
+
+std::string unicode_cpt_to_utf8(uint32_t cp);
+uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset);
+std::vector unicode_cpts_from_utf8(const std::string & utf8);
+
+std::vector unicode_cpts_normalize_nfd(const std::vector & cpts);
+
+codepoint_flags unicode_cpt_flags(const uint32_t cp);
+codepoint_flags unicode_cpt_flags(const std::string & utf8);
+
+std::string unicode_byte_to_utf8(uint8_t byte);
+uint8_t unicode_utf8_to_byte(const std::string & utf8);
+
+uint32_t unicode_tolower(uint32_t cp);
+
+std::vector unicode_regex_split(const std::string & text, const std::vector & regex_exprs);
-- 
cgit v1.2.3