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 <iwan.kawrakow@gmail.com>

1 files changed, 39 insertions, 0 deletions

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);