tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-29 04:35:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

tool-call: fix Qwen 2.5 Coder support, add micro benchmarks, support trigger patterns for lazy grammars (#12034)

Browse Source 
* sampler: turn lazy grammar trigger words to regexes

* add scripts/tool_bench.sh & .py

* constrain llama json output regardless of function name if matches at beginning

* update relaxed newline space rule in grammar tests

* support add_generation_prompt query parameter (useful for /apply_template)

* Update src/llama-grammar.cpp

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

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Olivier Chafik
2025-03-05 13:05:13 +00:00
committed by GitHub
parent fa31c438e0
commit 669912d9a5
26 changed files with 1314 additions and 408 deletions
Download Patch File Download Diff File Expand all files Collapse all files

44
src/llama-grammar.cpp
View File

@ -969,7 +969,7 @@ struct llama_grammar * llama_grammar_init_impl(
/* .awaiting_trigger = */ false,
/* .trigger_buffer = */ "",
/* .trigger_tokens = */ {},
/* .trigger_words = */ {},
/* .trigger_patterns = */ {},
};
}
@ -978,19 +978,15 @@ struct llama_grammar * llama_grammar_init_impl(
const char * grammar_str,
const char * grammar_root,
bool lazy,
const char ** trigger_words,
size_t num_trigger_words,
const char ** trigger_patterns,
size_t num_trigger_patterns,
const llama_token * trigger_tokens,
size_t num_trigger_tokens) {
llama_grammar_parser parser;
// if there is a grammar, parse it
if (!parser.parse(grammar_str)) {
return nullptr;
}
// will be empty (default) if there are parse errors
if (parser.rules.empty()) {
// rules will be empty (default) if there are parse errors
if (!parser.parse(grammar_str) || parser.rules.empty()) {
fprintf(stderr, "%s: failed to parse grammar\n", __func__);
return nullptr;
}
@ -1054,14 +1050,16 @@ struct llama_grammar * llama_grammar_init_impl(
} while (true);
std::vector<llama_token> vec_trigger_tokens;
std::vector<std::string> vec_trigger_words;
std::vector<llama_grammar_trigger_pattern> vec_trigger_patterns;
for (size_t i = 0; i < num_trigger_tokens; i++) {
GGML_ASSERT(trigger_tokens != nullptr);
vec_trigger_tokens.push_back(trigger_tokens[i]);
}
for (size_t i = 0; i < num_trigger_words; i++) {
GGML_ASSERT(trigger_words != nullptr);
vec_trigger_words.push_back(trigger_words[i]);
for (size_t i = 0; i < num_trigger_patterns; i++) {
GGML_ASSERT(trigger_patterns != nullptr);
auto & trigger = vec_trigger_patterns.emplace_back();
trigger.pattern = trigger_patterns[i];
trigger.regex = std::regex(trigger.pattern);
}
// Important: vec_rules has to be moved here, not copied, because stacks contains
@ -1076,7 +1074,7 @@ struct llama_grammar * llama_grammar_init_impl(
/* .awaiting_trigger = */ lazy,
/* .trigger_buffer = */ "",
std::move(vec_trigger_tokens),
std::move(vec_trigger_words),
std::move(vec_trigger_patterns),
};
}
@ -1089,7 +1087,7 @@ void llama_grammar_free_impl(struct llama_grammar * grammar) {
}
struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & grammar) {
llama_grammar * result = new llama_grammar {
auto * result = new llama_grammar {
grammar.vocab,
grammar.rules,
grammar.stacks,
@ -1098,7 +1096,7 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
grammar.awaiting_trigger,
grammar.trigger_buffer,
grammar.trigger_tokens,
grammar.trigger_words,
grammar.trigger_patterns,
};
// redirect elements in stacks to point to new rules
@ -1173,16 +1171,18 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
LLAMA_LOG_DEBUG("Grammar triggered on token %u (`%s`)", token, piece.c_str());
return;
} else {
// TODO: consider a smarter incremental substring search algorithm (store last position to search from).
grammar.trigger_buffer += piece;
for (const auto & word : grammar.trigger_words) {
auto pos = grammar.trigger_buffer.find(word);
if (pos != std::string::npos) {
std::smatch match;
for (const auto & trigger_pattern : grammar.trigger_patterns) {
if (std::regex_match(grammar.trigger_buffer, match, trigger_pattern.regex)) {
grammar.awaiting_trigger = false;
auto constrained_str = grammar.trigger_buffer.substr(pos);
// get from the first match to the end of the string
auto constrained_str = grammar.trigger_buffer.substr(match.position(1));
// std::string constrained_str(match[1].first, grammar.trigger_buffer.end());
grammar.trigger_buffer.clear();
llama_grammar_accept_str(grammar, constrained_str);
LLAMA_LOG_DEBUG("Grammar triggered on word `%s`", word.c_str());
LLAMA_LOG_DEBUG("Grammar triggered on regex: '%s'\n", constrained_str.c_str());
return;
}
}