tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-27 03:55:20 +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

62
examples/server/server.cpp
View File

@ -131,9 +131,9 @@ struct slot_params {
lora.push_back({{"id", i}, {"scale", this->lora[i].scale}});
}
std::vector<std::string> grammar_trigger_words;
for (const auto & trigger : sampling.grammar_trigger_words) {
grammar_trigger_words.push_back(trigger.word);
auto grammar_triggers = json::array();
for (const auto & trigger : sampling.grammar_triggers) {
grammar_triggers.push_back(trigger.to_json<json>());
}
return json {
@ -170,8 +170,8 @@ struct slot_params {
{"n_probs", sampling.n_probs},
{"min_keep", sampling.min_keep},
{"grammar", sampling.grammar},
{"grammar_trigger_words", grammar_trigger_words},
{"grammar_trigger_tokens", sampling.grammar_trigger_tokens},
{"grammar_lazy", sampling.grammar_lazy},
{"grammar_triggers", grammar_triggers},
{"preserved_tokens", sampling.preserved_tokens},
{"chat_format", common_chat_format_name(oaicompat_chat_format)},
{"samplers", samplers},
@ -356,24 +356,6 @@ struct server_task {
}
{
const auto grammar_triggers = data.find("grammar_triggers");
if (grammar_triggers != data.end()) {
for (const auto & t : *grammar_triggers) {
common_grammar_trigger trigger;
trigger.word = t.at("word");
trigger.at_start = t.at("at_start");
auto ids = common_tokenize(vocab, trigger.word, /* add_special= */ false, /* parse_special= */ true);
if (ids.size() == 1) {
SRV_DBG("Grammar trigger token: %d (`%s`)\n", ids[0], trigger.word.c_str());
params.sampling.grammar_trigger_tokens.push_back(ids[0]);
params.sampling.preserved_tokens.insert(ids[0]);
continue;
}
SRV_DBG("Grammar trigger word: `%s`\n", trigger.word.c_str());
params.sampling.grammar_trigger_words.push_back(trigger);
}
}
const auto preserved_tokens = data.find("preserved_tokens");
if (preserved_tokens != data.end()) {
for (const auto & t : *preserved_tokens) {
@ -383,12 +365,38 @@ struct server_task {
params.sampling.preserved_tokens.insert(ids[0]);
} else {
// This may happen when using a tool call style meant for a model with special tokens to preserve on a model without said tokens.
SRV_WRN("Not preserved because more than 1 token (wrong chat template override?): %s\n", t.get<std::string>().c_str());
SRV_DBG("Not preserved because more than 1 token: %s\n", t.get<std::string>().c_str());
}
}
}
if (params.sampling.grammar_lazy) {
GGML_ASSERT(params.sampling.grammar_trigger_tokens.size() > 0 || params.sampling.grammar_trigger_words.size() > 0);
const auto grammar_triggers = data.find("grammar_triggers");
if (grammar_triggers != data.end()) {
for (const auto & t : *grammar_triggers) {
auto ct = common_grammar_trigger::from_json(t);
if (ct.type == COMMON_GRAMMAR_TRIGGER_TYPE_WORD) {
const auto & word = ct.value;
auto ids = common_tokenize(vocab, word, /* add_special= */ false, /* parse_special= */ true);
if (ids.size() == 1) {
auto token = ids[0];
if (std::find(params.sampling.preserved_tokens.begin(), params.sampling.preserved_tokens.end(), (llama_token) token) == params.sampling.preserved_tokens.end()) {
throw std::runtime_error("Grammar trigger word should be marked as preserved token: " + word);
}
SRV_DBG("Grammar trigger token: %d (`%s`)\n", token, word.c_str());
common_grammar_trigger trigger;
trigger.type = COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN;
trigger.value = (llama_token) token;
params.sampling.grammar_triggers.push_back(trigger);
} else {
SRV_DBG("Grammar trigger word: `%s`\n", word.c_str());
params.sampling.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, word});
}
} else {
params.sampling.grammar_triggers.push_back(ct);
}
}
}
if (params.sampling.grammar_lazy && params.sampling.grammar_triggers.empty()) {
throw std::runtime_error("Error: no triggers set for lazy grammar!");
}
}
@ -2045,7 +2053,7 @@ struct server_context {
if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) {
// Might be better to reject the request with a 400 ?
SLT_WRN(slot, "n_predict = %d exceeds server configuration, setting to %d", slot.params.n_predict, slot.n_predict);
SLT_WRN(slot, "n_predict = %d exceeds server configuration, setting to %d\n", slot.params.n_predict, slot.n_predict);
slot.params.n_predict = slot.n_predict;
}