vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

rm dead code + nits

Browse source
This commit is contained in:
Olivier Chafik 2025-01-29 15:57:58 +00:00
parent 18d5a1b2ca
commit 923c805d04
7 changed files with 25 additions and 35 deletions
Download patch file Download diff file Expand all files Collapse all files

6
examples/server/server.cpp
View file

@ -768,7 +768,6 @@ struct server_task_result_cmpl_partial : server_task_result {
oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE;
std::string oaicompat_model;
std::string oaicompat_cmpl_id;
std::shared_ptr<common_chat_parser> chat_parser;
virtual int get_index() override {
return index;
@ -1191,7 +1190,6 @@ struct server_slot {
std::string stopping_word;
std::shared_ptr<common_chat_parser> chat_parser;
// sampling
json json_schema;
@ -1200,6 +1198,8 @@ struct server_slot {
llama_token sampled;
common_chat_parser chat_parser;
// stats
size_t n_sent_text = 0; // number of sent text character
@ -3998,8 +3998,6 @@ int main(int argc, char ** argv) {
auto body = json::parse(req.body);
const auto & chat_template = body.contains("tools") && ctx_server.chat_templates.template_tool_use ? *ctx_server.chat_templates.template_tool_use : *ctx_server.chat_templates.template_default;
LOG_INF("Request: %s\n", body.dump(2).c_str());
json data = oaicompat_completion_params_parse(body, chat_template, params.use_jinja);
return handle_completions_impl(

6
examples/server/tests/unit/test_tool_call.py
View file

@ -154,7 +154,7 @@ def test_completion_with_required_tool_real_model(tool: dict, argument_key: str
if template_override:
(template_hf_repo, template_variant) = template_override
server.chat_template_file = f"../../../tests/chat/templates/{template_hf_repo.replace('/', '') + ('-' + template_variant if template_variant else '')}.jinja"
assert os.path.exists(server.chat_template_file), f"Template file {server.chat_template_file} does not exist. Run `python scripts/get_hf_chat_template.py {template_hf_repo} {template_variant} > {server.chat_template_file}` to download the template."
assert os.path.exists(server.chat_template_file), f"Template file {server.chat_template_file} does not exist. Run `python scripts/get_chat_template.py {template_hf_repo} {template_variant} > {server.chat_template_file}` to download the template."
server.start()
res = server.make_request("POST", "/chat/completions", data={
"max_tokens": n_predict,
@ -243,7 +243,7 @@ def test_weather_tool_call(hf_repo: str, hf_file: str, template_override: Tuple[
if template_override:
(template_hf_repo, template_variant) = template_override
server.chat_template_file = f"../../../tests/chat/templates/{template_hf_repo.replace('/', '') + ('-' + template_variant if template_variant else '')}.jinja"
assert os.path.exists(server.chat_template_file), f"Template file {server.chat_template_file} does not exist. Run `python scripts/get_hf_chat_template.py {template_hf_repo} {template_variant} > {server.chat_template_file}` to download the template."
assert os.path.exists(server.chat_template_file), f"Template file {server.chat_template_file} does not exist. Run `python scripts/get_chat_template.py {template_hf_repo} {template_variant} > {server.chat_template_file}` to download the template."
server.start(timeout_seconds=15*60)
res = server.make_request("POST", "/chat/completions", data={
"max_tokens": 256,
@ -292,7 +292,7 @@ def test_hello_world_tool_call(expected_arguments: str | None, hf_repo: str, hf_
if template_override:
(template_hf_repo, template_variant) = template_override
server.chat_template_file = f"../../../tests/chat/templates/{template_hf_repo.replace('/', '') + ('-' + template_variant if template_variant else '')}.jinja"
assert os.path.exists(server.chat_template_file), f"Template file {server.chat_template_file} does not exist. Run `python scripts/get_hf_chat_template.py {template_hf_repo} {template_variant} > {server.chat_template_file}` to download the template."
assert os.path.exists(server.chat_template_file), f"Template file {server.chat_template_file} does not exist. Run `python scripts/get_chat_template.py {template_hf_repo} {template_variant} > {server.chat_template_file}` to download the template."
server.start(timeout_seconds=15*60)
res = server.make_request("POST", "/chat/completions", data={
"max_tokens": 256,

16
examples/server/utils.hpp
View file

@ -596,6 +596,11 @@ static json oaicompat_completion_params_parse(
throw std::runtime_error("tools param requires --jinja flag");
}
}
if (!use_jinja) {
if (body.contains("tool_choice") && !body.at("tool_choice").is_null()) {
throw std::runtime_error("Unsupported param: tool_choice");
}
}
// Handle "stop" field
if (body.contains("stop") && body.at("stop").is_string()) {
@ -605,7 +610,6 @@ static json oaicompat_completion_params_parse(
}
// Handle "response_format" field
auto tool_choice = json_value(body, "tool_choice", std::string("auto"));
if (body.contains("response_format")) {
json response_format = json_value(body, "response_format", json::object());
std::string response_type = json_value(response_format, "type", std::string());
@ -649,16 +653,6 @@ static json oaicompat_completion_params_parse(
throw std::runtime_error("top_logprobs requires logprobs to be set to true");
}
// Params supported by OAI but unsupported by llama.cpp
if (!use_jinja) {
static const std::vector<std::string> unsupported_params { "tool_choice" };
for (const auto & param : unsupported_params) {
if (body.contains(param)) {
throw std::runtime_error("Unsupported param: " + param);
}
}
}
// Copy remaining properties to llama_params
// This allows user to use llama.cpp-specific params like "mirostat", ... via OAI endpoint.
// See "launch_slot_with_task()" for a complete list of params supported by llama.cpp

13
scripts/get_hf_chat_template.py → scripts/get_chat_template.py
View file

@ -4,12 +4,12 @@
If a model has multiple chat templates, you can specify the variant name.
Syntax:
./scripts/get_hf_chat_template.py model_id [variant]
./scripts/get_chat_template.py model_id [variant]
Examples:
./scripts/get_hf_chat_template.py NousResearch/Meta-Llama-3-8B-Instruct
./scripts/get_hf_chat_template.py NousResearch/Hermes-3-Llama-3.1-8B tool_use
./scripts/get_hf_chat_template.py meta-llama/Llama-3.2-3B-Instruct
./scripts/get_chat_template.py NousResearch/Meta-Llama-3-8B-Instruct
./scripts/get_chat_template.py NousResearch/Hermes-3-Llama-3.1-8B tool_use
./scripts/get_chat_template.py meta-llama/Llama-3.2-3B-Instruct
'''
import json
@ -17,7 +17,7 @@ import re
import sys
def get_hf_chat_template(model_id, variant=None):
def get_chat_template(model_id, variant=None):
try:
# Use huggingface_hub library if available.
# Allows access to gated models if the user has access and ran `huggingface-cli login`.
@ -69,9 +69,10 @@ def main(args):
model_id = args[0]
variant = None if len(args) < 2 else args[1]
template = get_hf_chat_template(model_id, variant)
template = get_chat_template(model_id, variant)
sys.stdout.write(template)
if __name__ == '__main__':
main(sys.argv[1:])

2
src/llama-grammar.cpp
View file

@ -560,7 +560,7 @@ bool llama_grammar_parser::parse(const char * src) {
}
}
} catch (const std::exception & err) {
fprintf(stderr, "\n%s: error parsing grammar: %s\n\n%s\n", __func__, err.what(), src);
fprintf(stderr, "%s: error parsing grammar: %s\n\n%s\n", __func__, err.what(), src);
rules.clear();
return false;
}

4
src/llama-grammar.h
View file

@ -118,8 +118,8 @@ struct llama_grammar {
// lazy grammars wait for trigger words or tokens before constraining the sampling.
// we still ahve trigger_tokens for non-lazy grammars to force printing of special trigger tokens.
// (useful e.g. for tool_choice=required)
bool lazy; // Useful when resetting
bool awaiting_trigger; // Initialized to lazy
bool lazy;
bool awaiting_trigger; // Initialized to true for lazy grammars only
std::string trigger_buffer; // Output buffered by lazy grammar. Will be cleared once trigger is found.
std::vector<llama_token> trigger_tokens; // Tokens that trigger a lazy grammar, or tokens to force printing of (even if special).
std::vector<std::string> trigger_words;

13
tests/test-chat-handler.cpp
View file

@ -169,9 +169,6 @@ struct delta_data {
};
static delta_data init_delta(const common_chat_template & tmpl, const std::vector<std::string> & end_tokens, const json & user_message, const json & delta_message, const json & tools) {
fprintf(stderr, "Template source: %s\n", tmpl.source().c_str());
fprintf(stderr, "Delta message: %s\n", delta_message.dump(2).c_str());
common_chat_params params;
params.parallel_tool_calls = true;
params.messages = json::array();
@ -209,12 +206,14 @@ static delta_data init_delta(const common_chat_template & tmpl, const std::vecto
return {delta, full_data.grammar, full_data.parser};
}
/*
Applies the template to 1 user message w/ add_generation_prompt=true, then w/ the test message w/ add_generation_prompt=false,
gets the diff, removes any end tokens and parses the result w/ the grammar, checking that
the parsed message is the same as the test_message
*/
static void test_template(const common_chat_template & tmpl, const std::vector<std::string> & end_tokens, const json & test_message, const json & tools = {}, const std::string & expected_delta = "", bool skip_grammar_test = false, bool skip_parser_test = false) {
// auto tool_call_style = common_tool_call_style_detect(tmpl);
common_chat_msg expected_msg = msg_from_json(test_message);
// Format the message: apply the template to 1 user message w/ add_generation_prompt=true, then w/ the extra message w/ add_generation_prompt=false,
// get the diff and try and parse it w/ the grammar.
auto user_message = json {
{"role", "user"},
{"content", "Hello, world!"}
@ -228,7 +227,6 @@ static void test_template(const common_chat_template & tmpl, const std::vector<s
params.tools = tools;
auto data = init_delta(tmpl, end_tokens, user_message, test_message, tools);
std::cout << "Full delta:\n```\n" << data.delta << "\n```" << std::endl;
if (!expected_delta.empty()) {
assert_equals(expected_delta, data.delta);
}
@ -495,7 +493,6 @@ static void test_template_output_parsers() {
}
int main() {
// test_parsing();
test_template_output_parsers();
std::cout << "\n[tool-call] All tests passed!" << std::endl;