llama : custom attention mask + parallel decoding + no context swaps (#3228)

* tests : verify that RoPE is "additive"

* llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask)

* ggml : ggml_rope now takes a vector with positions instead of n_past

* metal : add rope_f16 kernel + optimize cpy kernels

* llama : unified KV cache + batch inference API

* llama : add new llama_decode() API that works with llama_batch

* llama : add cell_max heuristic for more efficient kv_cache

* llama : extend llama_kv_cache API

* llama : more robust cell_max heuristic + wip shift

* metal : disable concurrency optimization

* llama : add llama_kv_cache_shift_seq + no more context swaps

* llama : apply K-cache roping for Falcon and Baichuan

* speculative : fix KV cache management

* parallel : example for serving multiple users in parallel

* parallel : disable hot-plug to avoid cache fragmentation

* fixes : speculative KV cache + llama worst-case graph

* llama : extend batch API to select which logits to output

* llama : fix worst case graph build

* ggml-cuda : update rope implementation for parallel decoding (#3254)

* ggml-cuda : update rope implementation for parallel decoding

* better solution for p0 computation

* fix rope

* simpler rope implementation

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* make : add parallel to build + fix static functions in llama.cpp

* simple : fix token counting

* parallel : various improvements

* llama : fix cell_max logic + rename functions

* parallel : try smaller batches when the KV cache is fragmented

* parallel : fix sequence termination criteria

* llama : silence errors KV cache errors

* parallel : remove new line from prompt

* parallel : process system prompt once + configurable paramters + llama API

* parallel : remove question with short answers

* parallel : count cache misses

* parallel : print misses on each request

* parallel : minor

* llama : fix n_kv to never become 0

* parallel : rename hot-plug to continuous-batching

* llama : improve llama_batch API + simplify parallel example

* simple : add parallel decoding support

* simple : improve comments + free batch

* ggml-cuda : add rope f16, restore performance with parallel decoding (#3272)

* ggml-cuda : add rope f16, restore performance

* offload KQ_mask with all models

* fix rope shift

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* llama : disable MPI for now

ggml-ci

* train : make KQ_pos memory buffer permanent via dummy scale op

* ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275)

ggml-ci

* parallel : fix bug (extra BOS) + smaller token_prev array

* parallel : fix cases where the input prompts can overflow the batch

* parallel : add disabled experimental batch chunking in powers of two

* llama : llama.h formatting + comments

* simple : add README.md

* llama : fix kv cache heuristic when context is less than 32

* parallel : fix crash when `-n -1`

* llama : simplify returns if/else branches

* metal : use mm kernels for batch size > 2

* examples : utilize new llama_get_logits_ith()

* examples : add example for batched decoding

* examples : do not eval prompt 2 times (close #3348)

* server : clear the KV cache beyond n_past before llama_decode

* server : avoid context swaps by shifting the KV cache

---------

Co-authored-by: slaren <slarengh@gmail.com>

1 files changed, 97 insertions, 39 deletions

diff --git a/examples/simple/simple.cpp b/examples/simple/simple.cpp
index 440d22ec..1616a4a7 100644
--- a/examples/simple/simple.cpp
+++ b/examples/simple/simple.cpp
@@ -26,12 +26,18 @@ int main(int argc, char ** argv) {
         params.prompt = "Hello my name is";
     }
 
+    // total length of the sequence including the prompt
+    const int n_len = 32;
+
     // init LLM
 
     llama_backend_init(params.numa);
 
     llama_context_params ctx_params = llama_context_default_params();
 
+    ctx_params.seed  = 1234;
+    ctx_params.n_ctx = 2048;
+
     llama_model * model = llama_load_model_from_file(params.model.c_str(), ctx_params);
 
     if (model == NULL) {
@@ -41,20 +47,31 @@ int main(int argc, char ** argv) {
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
     // tokenize the prompt
 
     std::vector<llama_token> tokens_list;
     tokens_list = ::llama_tokenize(ctx, params.prompt, true);
 
-    const int max_context_size     = llama_n_ctx(ctx);
-    const int max_tokens_list_size = max_context_size - 4;
+    const int n_ctx    = llama_n_ctx(ctx);
+    const int n_kv_req = tokens_list.size() + (n_len - tokens_list.size());
+
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, n_kv_req);
 
-    if ((int) tokens_list.size() > max_tokens_list_size) {
-        fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) tokens_list.size(), max_tokens_list_size);
+    // make sure the KV cache is big enough to hold all the prompt and generated tokens
+    if (n_kv_req > n_ctx) {
+        LOG_TEE("%s: error: n_kv_req > n_ctx, the required KV cache size is not big enough\n", __func__);
+        LOG_TEE("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
         return 1;
     }
 
-    fprintf(stderr, "\n\n");
+    // print the prompt token-by-token
+
+    fprintf(stderr, "\n");
 
     for (auto id : tokens_list) {
         fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
@@ -62,63 +79,104 @@ int main(int argc, char ** argv) {
 
     fflush(stderr);
 
-    // main loop
+    // create a llama_batch with size 512
+    // we use this object to submit token data for decoding
 
-    // The LLM keeps a contextual cache memory of previous token evaluation.
-    // Usually, once this cache is full, it is required to recompute a compressed context based on previous
-    // tokens (see "infinite text generation via context swapping" in the main example), but in this minimalist
-    // example, we will just stop the loop once this cache is full or once an end of stream is detected.
+    llama_batch batch = llama_batch_init(512, 0);
 
-    const int n_gen = std::min(32, max_context_size);
+    // evaluate the initial prompt
+    batch.n_tokens = tokens_list.size();
 
-    while (llama_get_kv_cache_token_count(ctx) < n_gen) {
-        // evaluate the transformer
+    for (int32_t i = 0; i < batch.n_tokens; i++) {
+        batch.token[i]  = tokens_list[i];
+        batch.pos[i]    = i;
+        batch.seq_id[i] = 0;
+        batch.logits[i] = false;
+    }
 
-        if (llama_eval(ctx, tokens_list.data(), int(tokens_list.size()), llama_get_kv_cache_token_count(ctx), params.n_threads)) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return 1;
-        }
+    // llama_decode will output logits only for the last token of the prompt
+    batch.logits[batch.n_tokens - 1] = true;
+
+    if (llama_decode(ctx, batch, params.n_threads) != 0) {
+        LOG_TEE("%s: llama_decode() failed\n", __func__);
+        return 1;
+    }
+
+    // main loop
 
-        tokens_list.clear();
+    int n_cur    = batch.n_tokens;
+    int n_decode = 0;
 
+    const auto t_main_start = ggml_time_us();
+
+    while (n_cur <= n_len) {
         // sample the next token
+        {
+            auto   n_vocab = llama_n_vocab(ctx);
+            auto * logits  = llama_get_logits_ith(ctx, batch.n_tokens - 1);
 
-        llama_token new_token_id = 0;
+            std::vector<llama_token_data> candidates;
+            candidates.reserve(n_vocab);
 
-        auto logits  = llama_get_logits(ctx);
-        auto n_vocab = llama_n_vocab(ctx);
+            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
+            }
 
-        std::vector<llama_token_data> candidates;
-        candidates.reserve(n_vocab);
+            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
 
-        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-            candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
-        }
+            // sample the most likely token
+            const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+
+            // is it an end of stream?
+            if (new_token_id == llama_token_eos(ctx) || n_cur == n_len) {
+                LOG_TEE("\n");
+
+                break;
+            }
 
-        llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+            LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
+            fflush(stdout);
 
-        new_token_id = llama_sample_token_greedy(ctx , &candidates_p);
+            // prepare the next batch
+            batch.n_tokens = 0;
 
-        // is it an end of stream ?
-        if (new_token_id == llama_token_eos(ctx)) {
-            fprintf(stderr, " [end of text]\n");
-            break;
+            // push this new token for next evaluation
+            batch.token [batch.n_tokens] = new_token_id;
+            batch.pos   [batch.n_tokens] = n_cur;
+            batch.seq_id[batch.n_tokens] = 0;
+            batch.logits[batch.n_tokens] = true;
+
+            batch.n_tokens += 1;
+
+            n_decode += 1;
         }
 
-        // print the new token :
-        printf("%s", llama_token_to_piece(ctx, new_token_id).c_str());
-        fflush(stdout);
+        n_cur += 1;
 
-        // push this new token for next evaluation
-        tokens_list.push_back(new_token_id);
+        // evaluate the current batch with the transformer model
+        if (llama_decode(ctx, batch, params.n_threads)) {
+            fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
+            return 1;
+        }
     }
 
+    LOG_TEE("\n");
+
+    const auto t_main_end = ggml_time_us();
+
+    LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
+            __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
+
+    llama_print_timings(ctx);
+
+    fprintf(stderr, "\n");
+
+    llama_batch_free(batch);
+
     llama_free(ctx);
     llama_free_model(model);
 
     llama_backend_free();
 
-    fprintf(stderr, "\n\n");
-
     return 0;
 }