Tool call support (generic + native for Llama, Functionary, Hermes, Mistral, Firefunction, DeepSeek) w/ lazy grammars (#9639)
--------- Co-authored-by: Xuan Son Nguyen <thichthat@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
This commit is contained in:
Olivier Chafik
GitHub
parent
27d135c970
commit
8b576b6c55
48 changed files with 3861 additions and 156 deletions
|
|
@ -1117,6 +1117,82 @@ curl http://localhost:8080/v1/chat/completions \
|
|||
}'
|
||||
```
|
||||
|
||||
... and even tool usage (needs `--jinja` flag):
|
||||
|
||||
```shell
|
||||
llama-server --jinja -hfr lmstudio-community/Meta-Llama-3.1-8B-Instruct-GGUF -hff Meta-Llama-3.1-8B-Instruct-Q5_K_M.gguf -fa
|
||||
|
||||
# https://huggingface.co/meetkai/functionary-medium-v3.2
|
||||
llama-server --jinja -hfr bartowski/functionary-medium-v3.2-GGUF -hff functionary-medium-v3.2-IQ4_XS.gguf -fa
|
||||
|
||||
# https://huggingface.co/meetkai/functionary-medium-v3.1
|
||||
llama-server --jinja -hfr meetkai/functionary-medium-v3.1-GGUF -hff functionary-medium-llama-3.1.Q4_0.gguf -fa
|
||||
|
||||
curl http://localhost:8080/v1/chat/completions -d '{
|
||||
"model": "gpt-3.5-turbo",
|
||||
"tools": [
|
||||
{
|
||||
"type":"function",
|
||||
"function":{
|
||||
"name":"get_current_weather",
|
||||
"description":"Get the current weather in a given location",
|
||||
"parameters":{
|
||||
"type":"object",
|
||||
"properties":{
|
||||
"location":{
|
||||
"type":"string",
|
||||
"description":"The city and state, e.g. San Francisco, CA"
|
||||
}
|
||||
},
|
||||
"required":["location"]
|
||||
}
|
||||
}
|
||||
}
|
||||
],
|
||||
"messages": [
|
||||
{
|
||||
"role": "user",
|
||||
"content": "What is the weather like in Istanbul?."
|
||||
}
|
||||
]
|
||||
}'
|
||||
```
|
||||
|
||||
<details>
|
||||
<summary>Show output</summary>
|
||||
|
||||
```json
|
||||
{
|
||||
"choices": [
|
||||
{
|
||||
"finish_reason": "tool",
|
||||
"index": 0,
|
||||
"message": {
|
||||
"content": null,
|
||||
"tool_calls": [
|
||||
{
|
||||
"name": "python",
|
||||
"arguments": "{\"code\":\" \\nprint(\\\"Hello, World!\\\")\"}"
|
||||
}
|
||||
],
|
||||
"role": "assistant"
|
||||
}
|
||||
}
|
||||
],
|
||||
"created": 1727287211,
|
||||
"model": "gpt-3.5-turbo",
|
||||
"object": "chat.completion",
|
||||
"usage": {
|
||||
"completion_tokens": 16,
|
||||
"prompt_tokens": 44,
|
||||
"total_tokens": 60
|
||||
},
|
||||
"id": "chatcmpl-Htbgh9feMmGM0LEH2hmQvwsCxq3c6Ni8"
|
||||
}
|
||||
```
|
||||
|
||||
</details>
|
||||
|
||||
### POST `/v1/embeddings`: OpenAI-compatible embeddings API
|
||||
|
||||
This endpoint requires that the model uses a pooling different than type `none`. The embeddings are normalized using the Eucledian norm.
|
||||
|
|
|
|||
|
|
@ -113,10 +113,11 @@ struct slot_params {
|
|||
struct common_params_speculative speculative;
|
||||
|
||||
// OAI-compat fields
|
||||
bool verbose = false;
|
||||
oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE;
|
||||
std::string oaicompat_model;
|
||||
std::string oaicompat_cmpl_id;
|
||||
bool verbose = false;
|
||||
oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE;
|
||||
std::string oaicompat_model;
|
||||
std::string oaicompat_cmpl_id;
|
||||
common_chat_format oaicompat_chat_format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
|
||||
|
||||
json to_json() const {
|
||||
std::vector<std::string> samplers;
|
||||
|
|
@ -164,6 +165,8 @@ struct slot_params {
|
|||
{"n_probs", sampling.n_probs},
|
||||
{"min_keep", sampling.min_keep},
|
||||
{"grammar", sampling.grammar},
|
||||
// {"grammar_trigger_words", sampling.grammar_trigger_words},
|
||||
{"grammar_trigger_tokens", sampling.grammar_trigger_tokens},
|
||||
{"samplers", samplers},
|
||||
{"speculative.n_max", speculative.n_max},
|
||||
{"speculative.n_min", speculative.n_min},
|
||||
|
|
@ -325,12 +328,50 @@ struct server_task {
|
|||
if (data.contains("json_schema") && !data.contains("grammar")) {
|
||||
try {
|
||||
auto schema = json_value(data, "json_schema", json::object());
|
||||
params.sampling.grammar = json_schema_to_grammar(schema);
|
||||
LOG_DBG("JSON schema: %s\n", schema.dump(2).c_str());
|
||||
params.sampling.grammar = json_schema_to_grammar(schema);
|
||||
LOG_DBG("Converted grammar: %s\n", params.sampling.grammar.c_str());
|
||||
} catch (const std::exception & e) {
|
||||
throw std::runtime_error(std::string("\"json_schema\": ") + e.what());
|
||||
}
|
||||
} else {
|
||||
params.sampling.grammar = json_value(data, "grammar", defaults.sampling.grammar);
|
||||
params.sampling.grammar = json_value(data, "grammar", defaults.sampling.grammar);
|
||||
LOG_DBG("Grammar: %s\n", params.sampling.grammar.c_str());
|
||||
params.sampling.grammar_lazy = json_value(data, "grammar_lazy", defaults.sampling.grammar_lazy);
|
||||
LOG_DBG("Grammar lazy: %s\n", params.sampling.grammar_lazy ? "true" : "false");
|
||||
}
|
||||
|
||||
{
|
||||
auto it = data.find("chat_format");
|
||||
if (it != data.end()) {
|
||||
params.oaicompat_chat_format = static_cast<common_chat_format>(it->get<int>());
|
||||
LOG_DBG("Chat format: %s\n", common_chat_format_name(params.oaicompat_chat_format).c_str());
|
||||
} else {
|
||||
params.oaicompat_chat_format = defaults.oaicompat_chat_format;
|
||||
}
|
||||
}
|
||||
|
||||
{
|
||||
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) {
|
||||
LOG_DBG("Grammar trigger token: %d (`%s`)\n", ids[0], trigger.word.c_str());
|
||||
params.sampling.grammar_trigger_tokens.push_back(ids[0]);
|
||||
continue;
|
||||
}
|
||||
LOG_DBG("Grammar trigger word: `%s`\n", trigger.word.c_str());
|
||||
params.sampling.grammar_trigger_words.push_back(trigger);
|
||||
}
|
||||
}
|
||||
if (params.sampling.grammar_lazy) {
|
||||
GGML_ASSERT(params.sampling.grammar_trigger_tokens.size() > 0 || params.sampling.grammar_trigger_words.size() > 0);
|
||||
}
|
||||
}
|
||||
|
||||
{
|
||||
|
|
@ -382,22 +423,12 @@ struct server_task {
|
|||
}
|
||||
|
||||
{
|
||||
const auto & samplers = data.find("samplers");
|
||||
const auto samplers = data.find("samplers");
|
||||
if (samplers != data.end()) {
|
||||
if (samplers->is_array()) {
|
||||
std::vector<std::string> sampler_names;
|
||||
for (const auto & name : *samplers) {
|
||||
if (name.is_string()) {
|
||||
sampler_names.emplace_back(name);
|
||||
}
|
||||
}
|
||||
params.sampling.samplers = common_sampler_types_from_names(sampler_names, false);
|
||||
params.sampling.samplers = common_sampler_types_from_names(*samplers, false);
|
||||
} else if (samplers->is_string()){
|
||||
std::string sampler_string;
|
||||
for (const auto & name : *samplers) {
|
||||
sampler_string += name;
|
||||
}
|
||||
params.sampling.samplers = common_sampler_types_from_chars(sampler_string);
|
||||
params.sampling.samplers = common_sampler_types_from_chars(samplers->get<std::string>());
|
||||
}
|
||||
} else {
|
||||
params.sampling.samplers = defaults.sampling.samplers;
|
||||
|
|
@ -544,7 +575,7 @@ struct completion_token_output {
|
|||
struct server_task_result_cmpl_final : server_task_result {
|
||||
int index = 0;
|
||||
|
||||
std::string content;
|
||||
std::string content;
|
||||
llama_tokens tokens;
|
||||
|
||||
bool stream;
|
||||
|
|
@ -566,10 +597,11 @@ struct server_task_result_cmpl_final : server_task_result {
|
|||
slot_params generation_params;
|
||||
|
||||
// OAI-compat fields
|
||||
bool verbose = false;
|
||||
oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE;
|
||||
std::string oaicompat_model;
|
||||
std::string oaicompat_cmpl_id;
|
||||
bool verbose = false;
|
||||
oaicompat_type oaicompat = OAICOMPAT_TYPE_NONE;
|
||||
std::string oaicompat_model;
|
||||
std::string oaicompat_cmpl_id;
|
||||
common_chat_format oaicompat_chat_format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
|
||||
|
||||
virtual int get_index() override {
|
||||
return index;
|
||||
|
|
@ -663,18 +695,38 @@ struct server_task_result_cmpl_final : server_task_result {
|
|||
|
||||
json to_json_oaicompat_chat() {
|
||||
std::string finish_reason = "length";
|
||||
common_chat_msg message;
|
||||
if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
|
||||
finish_reason = "stop";
|
||||
message = common_chat_parse(content, oaicompat_chat_format);
|
||||
finish_reason = message.tool_calls.empty() ? "stop" : "tool_calls";
|
||||
} else {
|
||||
message.content = content;
|
||||
}
|
||||
|
||||
json choice = json{
|
||||
json tool_calls;
|
||||
if (!message.tool_calls.empty()) {
|
||||
tool_calls = json::array();
|
||||
for (const auto & tc : message.tool_calls) {
|
||||
tool_calls.push_back({
|
||||
{"type", "function"},
|
||||
{"function", {
|
||||
{"name", tc.name},
|
||||
{"arguments", tc.arguments},
|
||||
}},
|
||||
{"id", tc.id.empty() ? json() : json(tc.id)},
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
json choice {
|
||||
{"finish_reason", finish_reason},
|
||||
{"index", 0},
|
||||
{"message", json {
|
||||
{"content", content},
|
||||
{"role", "assistant"}
|
||||
}
|
||||
}};
|
||||
{"content", message.content},
|
||||
{"tool_calls", tool_calls},
|
||||
{"role", "assistant"},
|
||||
}},
|
||||
};
|
||||
|
||||
if (!stream && probs_output.size() > 0) {
|
||||
choice["logprobs"] = json{
|
||||
|
|
@ -716,7 +768,7 @@ struct server_task_result_cmpl_final : server_task_result {
|
|||
finish_reason = "stop";
|
||||
}
|
||||
|
||||
json choice = json{
|
||||
json choice = json {
|
||||
{"finish_reason", finish_reason},
|
||||
{"index", 0},
|
||||
{"delta", json::object()}
|
||||
|
|
@ -1191,6 +1243,8 @@ struct server_slot {
|
|||
|
||||
llama_token sampled;
|
||||
|
||||
common_chat_format chat_format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
|
||||
|
||||
// stats
|
||||
size_t n_sent_text = 0; // number of sent text character
|
||||
|
||||
|
|
@ -1815,17 +1869,16 @@ struct server_context {
|
|||
|
||||
if (use_jinja) {
|
||||
auto templates = common_chat_templates_from_model(model, "");
|
||||
common_chat_inputs inputs;
|
||||
inputs.messages = json::array({{
|
||||
{"role", "user"},
|
||||
{"content", "test"},
|
||||
}});
|
||||
GGML_ASSERT(templates.template_default);
|
||||
try {
|
||||
templates.template_default->apply({{
|
||||
{"role", "user"},
|
||||
{"content", "test"},
|
||||
}}, json(), true);
|
||||
common_chat_params_init(*templates.template_default, inputs);
|
||||
if (templates.template_tool_use) {
|
||||
templates.template_tool_use->apply({{
|
||||
{"role", "user"},
|
||||
{"content", "test"},
|
||||
}}, json(), true);
|
||||
common_chat_params_init(*templates.template_tool_use, inputs);
|
||||
}
|
||||
return true;
|
||||
} catch (const std::exception & e) {
|
||||
|
|
@ -2275,11 +2328,11 @@ struct server_context {
|
|||
res->id_slot = slot.id;
|
||||
|
||||
res->index = slot.index;
|
||||
res->content = slot.generated_text;
|
||||
res->tokens = slot.generated_tokens;
|
||||
res->content = std::move(slot.generated_text);
|
||||
res->tokens = std::move(slot.generated_tokens);
|
||||
res->timings = slot.get_timings();
|
||||
res->prompt = common_detokenize(ctx, slot.prompt_tokens, true);
|
||||
res->response_fields = slot.params.response_fields;
|
||||
res->response_fields = std::move(slot.params.response_fields);
|
||||
|
||||
res->truncated = slot.truncated;
|
||||
res->n_decoded = slot.n_decoded;
|
||||
|
|
@ -2290,12 +2343,12 @@ struct server_context {
|
|||
res->stop = slot.stop;
|
||||
res->post_sampling_probs = slot.params.post_sampling_probs;
|
||||
|
||||
res->verbose = slot.params.verbose;
|
||||
res->stream = slot.params.stream;
|
||||
res->oaicompat = slot.params.oaicompat;
|
||||
res->oaicompat_model = slot.params.oaicompat_model;
|
||||
res->oaicompat_cmpl_id = slot.params.oaicompat_cmpl_id;
|
||||
|
||||
res->verbose = slot.params.verbose;
|
||||
res->stream = slot.params.stream;
|
||||
res->oaicompat = slot.params.oaicompat;
|
||||
res->oaicompat_model = slot.params.oaicompat_model;
|
||||
res->oaicompat_cmpl_id = slot.params.oaicompat_cmpl_id;
|
||||
res->oaicompat_chat_format = slot.params.oaicompat_chat_format;
|
||||
// populate res.probs_output
|
||||
if (slot.params.sampling.n_probs > 0) {
|
||||
if (!slot.params.stream && slot.stop == STOP_TYPE_WORD) {
|
||||
|
|
@ -2773,6 +2826,11 @@ struct server_context {
|
|||
// track if given slot can be batched with slots already in the batch
|
||||
server_slot * slot_batched = nullptr;
|
||||
|
||||
auto accept_special_token = [&](server_slot & slot, llama_token token) {
|
||||
const auto & trigger_tokens = slot.params.sampling.grammar_trigger_tokens;
|
||||
return params_base.special || std::find(trigger_tokens.begin(), trigger_tokens.end(), token) != trigger_tokens.end();
|
||||
};
|
||||
|
||||
// frist, add sampled tokens from any ongoing sequences
|
||||
for (auto & slot : slots) {
|
||||
if (slot.state != SLOT_STATE_GENERATING) {
|
||||
|
|
@ -3136,7 +3194,7 @@ struct server_context {
|
|||
|
||||
completion_token_output result;
|
||||
result.tok = id;
|
||||
result.text_to_send = common_token_to_piece(ctx, result.tok, params_base.special);
|
||||
result.text_to_send = common_token_to_piece(ctx, result.tok, accept_special_token(slot, result.tok));
|
||||
result.prob = 1.0f; // TODO: set it here instead of doing inside populate_token_probs
|
||||
|
||||
if (slot.params.sampling.n_probs > 0) {
|
||||
|
|
@ -3225,7 +3283,7 @@ struct server_context {
|
|||
completion_token_output result;
|
||||
|
||||
result.tok = ids[i];
|
||||
result.text_to_send = common_token_to_piece(ctx, result.tok, params_base.special);
|
||||
result.text_to_send = common_token_to_piece(ctx, result.tok, accept_special_token(slot, result.tok));
|
||||
result.prob = 1.0f; // set later
|
||||
|
||||
// TODO: set result.probs
|
||||
|
|
@ -3722,6 +3780,8 @@ int main(int argc, char ** argv) {
|
|||
{ "total_slots", ctx_server.params_base.n_parallel },
|
||||
{ "model_path", ctx_server.params_base.model },
|
||||
{ "chat_template", ctx_server.chat_templates.template_default->source() },
|
||||
{ "bos_token", ctx_server.chat_templates.template_default->bos_token() },
|
||||
{ "eos_token", ctx_server.chat_templates.template_default->eos_token() },
|
||||
{ "build_info", build_info },
|
||||
};
|
||||
if (ctx_server.params_base.use_jinja && ctx_server.chat_templates.template_tool_use) {
|
||||
|
|
@ -3763,7 +3823,9 @@ int main(int argc, char ** argv) {
|
|||
std::vector<server_task> tasks;
|
||||
|
||||
try {
|
||||
std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, data.at("prompt"), true, true);
|
||||
const auto & prompt = data.at("prompt");
|
||||
LOG_DBG("Prompt: %s\n", prompt.is_string() ? prompt.get<std::string>().c_str() : prompt.dump(2).c_str());
|
||||
std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true);
|
||||
tasks.reserve(tokenized_prompts.size());
|
||||
for (size_t i = 0; i < tokenized_prompts.size(); i++) {
|
||||
server_task task = server_task(type);
|
||||
|
|
@ -3779,8 +3841,8 @@ int main(int argc, char ** argv) {
|
|||
task.id_selected_slot = json_value(data, "id_slot", -1);
|
||||
|
||||
// OAI-compat
|
||||
task.params.oaicompat = oaicompat;
|
||||
task.params.oaicompat_cmpl_id = completion_id;
|
||||
task.params.oaicompat = oaicompat;
|
||||
task.params.oaicompat_cmpl_id = completion_id;
|
||||
// oaicompat_model is already populated by params_from_json_cmpl
|
||||
|
||||
tasks.push_back(task);
|
||||
|
|
@ -3949,14 +4011,14 @@ int main(int argc, char ** argv) {
|
|||
};
|
||||
|
||||
const auto handle_chat_completions = [&ctx_server, ¶ms, &res_error, &handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
|
||||
LOG_DBG("request: %s\n", req.body.c_str());
|
||||
if (ctx_server.params_base.embedding) {
|
||||
res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
|
||||
return;
|
||||
}
|
||||
|
||||
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;
|
||||
json data = oaicompat_completion_params_parse(body, chat_template, params.use_jinja);
|
||||
json data = oaicompat_completion_params_parse(body, params.use_jinja, ctx_server.chat_templates);
|
||||
|
||||
return handle_completions_impl(
|
||||
SERVER_TASK_TYPE_COMPLETION,
|
||||
|
|
@ -3966,6 +4028,13 @@ int main(int argc, char ** argv) {
|
|||
OAICOMPAT_TYPE_CHAT);
|
||||
};
|
||||
|
||||
// same with handle_chat_completions, but without inference part
|
||||
const auto handle_apply_template = [&ctx_server, ¶ms, &res_ok](const httplib::Request & req, httplib::Response & res) {
|
||||
auto body = json::parse(req.body);
|
||||
json data = oaicompat_completion_params_parse(body, params.use_jinja, ctx_server.chat_templates);
|
||||
res_ok(res, {{ "prompt", std::move(data.at("prompt")) }});
|
||||
};
|
||||
|
||||
const auto handle_models = [¶ms, &ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) {
|
||||
json models = {
|
||||
{"object", "list"},
|
||||
|
|
@ -4124,14 +4193,6 @@ int main(int argc, char ** argv) {
|
|||
res_ok(res, root);
|
||||
};
|
||||
|
||||
const auto handle_apply_template = [&ctx_server, ¶ms, &res_ok](const httplib::Request & req, httplib::Response & res) {
|
||||
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;
|
||||
json data = oaicompat_completion_params_parse(body, chat_template, params.use_jinja);
|
||||
|
||||
res_ok(res, {{ "prompt", data.at("prompt") }});
|
||||
};
|
||||
|
||||
const auto handle_embeddings = [&handle_embeddings_impl](const httplib::Request & req, httplib::Response & res) {
|
||||
handle_embeddings_impl(req, res, OAICOMPAT_TYPE_NONE);
|
||||
};
|
||||
|
|
|
|||
|
|
@ -31,8 +31,9 @@ It's possible to override some scenario steps values with environment variables:
|
|||
| `LLAMA_SERVER_BIN_PATH` | to change the server binary path, default: `../../../build/bin/llama-server` |
|
||||
| `DEBUG` | to enable steps and server verbose mode `--verbose` |
|
||||
| `N_GPU_LAYERS` | number of model layers to offload to VRAM `-ngl --n-gpu-layers` |
|
||||
| `LLAMA_CACHE` | by default server tests re-download models to the `tmp` subfolder. Set this to your cache (e.g. `$HOME/Library/Caches/llama.cpp` on Mac or `$HOME/.cache/llama.cpp` on Unix) to avoid this |
|
||||
|
||||
To run slow tests:
|
||||
To run slow tests (will download many models, make sure to set `LLAMA_CACHE` if needed):
|
||||
|
||||
```shell
|
||||
SLOW_TESTS=1 ./tests.sh
|
||||
|
|
@ -44,10 +45,16 @@ To run with stdout/stderr display in real time (verbose output, but useful for d
|
|||
DEBUG=1 ./tests.sh -s -v -x
|
||||
```
|
||||
|
||||
To run single test unit:
|
||||
To run all the tests in a file:
|
||||
|
||||
```shell
|
||||
./tests.sh unit/test_{name of test case here}.py -v -x
|
||||
./tests.sh unit/test_chat_completion.py.py -v -x
|
||||
```
|
||||
|
||||
To run a single test:
|
||||
|
||||
```shell
|
||||
./tests.sh unit/test_chat_completion.py::test_invalid_chat_completion_req
|
||||
```
|
||||
|
||||
Hint: You can compile and run test in single command, useful for local developement:
|
||||
|
|
|
|||
4
examples/server/tests/pytest.ini
Normal file
4
examples/server/tests/pytest.ini
Normal file
|
|
@ -0,0 +1,4 @@
|
|||
[pytest]
|
||||
markers =
|
||||
slow: marks tests as slow (deselect with '-m "not slow"')
|
||||
serial
|
||||
|
|
@ -6,9 +6,18 @@ cd $SCRIPT_DIR
|
|||
|
||||
set -eu
|
||||
|
||||
if [[ "${SLOW_TESTS:-0}" == 1 ]]; then
|
||||
# Slow tests for tool calls need quite a few models ahead of time to avoid timing out.
|
||||
python $SCRIPT_DIR/../../../scripts/fetch_server_test_models.py
|
||||
fi
|
||||
|
||||
if [ $# -lt 1 ]
|
||||
then
|
||||
pytest -v -x
|
||||
if [[ "${SLOW_TESTS:-0}" == 1 ]]; then
|
||||
pytest -v -x
|
||||
else
|
||||
pytest -v -x -m "not slow"
|
||||
fi
|
||||
else
|
||||
pytest "$@"
|
||||
fi
|
||||
|
|
|
|||
|
|
@ -2,7 +2,7 @@ import pytest
|
|||
from openai import OpenAI
|
||||
from utils import *
|
||||
|
||||
server = ServerPreset.tinyllama2()
|
||||
server: ServerProcess
|
||||
|
||||
@pytest.fixture(autouse=True)
|
||||
def create_server():
|
||||
|
|
@ -13,11 +13,12 @@ def create_server():
|
|||
@pytest.mark.parametrize(
|
||||
"model,system_prompt,user_prompt,max_tokens,re_content,n_prompt,n_predicted,finish_reason,jinja,chat_template",
|
||||
[
|
||||
(None, "Book", "What is the best book", 8, "(Suddenly)+", 77, 8, "length", False, None),
|
||||
(None, "Book", "What is the best book", 8, "(Suddenly)+", 77, 8, "length", True, None),
|
||||
(None, "Book", "What is the best book", 8, "^ blue", 23, 8, "length", True, "This is not a chat template, it is"),
|
||||
(None, "Book", "What is the best book", 8, "(Suddenly)+|\\{ \" Sarax.", 77, 8, "length", False, None),
|
||||
("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length", False, None),
|
||||
("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length", True, None),
|
||||
# TODO: fix testing of non-tool jinja mode
|
||||
# (None, "Book", "What is the best book", 8, "(Suddenly)+|\\{ \" Sarax.", 77, 8, "length", True, None),
|
||||
# (None, "Book", "What is the best book", 8, "I want to play with", 23, 8, "length", True, "This is not a chat template, it is"),
|
||||
# ("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length", True, None),
|
||||
]
|
||||
)
|
||||
def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, finish_reason, jinja, chat_template):
|
||||
|
|
|
|||
352
examples/server/tests/unit/test_tool_call.py
Normal file
352
examples/server/tests/unit/test_tool_call.py
Normal file
|
|
@ -0,0 +1,352 @@
|
|||
import pytest
|
||||
from utils import *
|
||||
|
||||
server: ServerProcess
|
||||
|
||||
TIMEOUT_SERVER_START = 15*60
|
||||
TIMEOUT_HTTP_REQUEST = 60
|
||||
|
||||
@pytest.fixture(autouse=True)
|
||||
def create_server():
|
||||
global server
|
||||
server = ServerPreset.tinyllama2()
|
||||
server.model_alias = "tinyllama-2-tool-call"
|
||||
server.server_port = 8081
|
||||
|
||||
|
||||
TEST_TOOL = {
|
||||
"type":"function",
|
||||
"function": {
|
||||
"name": "test",
|
||||
"description": "",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"success": {"type": "boolean", "const": True},
|
||||
},
|
||||
"required": ["success"]
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
PYTHON_TOOL = {
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "python",
|
||||
"description": "Runs code in an ipython interpreter and returns the result of the execution after 60 seconds.",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"code": {
|
||||
"type": "string",
|
||||
"description": "The code to run in the ipython interpreter."
|
||||
}
|
||||
},
|
||||
"required": ["code"]
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
WEATHER_TOOL = {
|
||||
"type":"function",
|
||||
"function":{
|
||||
"name":"get_current_weather",
|
||||
"description":"Get the current weather in a given location",
|
||||
"parameters":{
|
||||
"type":"object",
|
||||
"properties":{
|
||||
"location":{
|
||||
"type":"string",
|
||||
"description":"The city and country/state, e.g. 'San Francisco, CA', or 'Paris, France'"
|
||||
}
|
||||
},
|
||||
"required":["location"]
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
def do_test_completion_with_required_tool_tiny(template_name: str, tool: dict, argument_key: str | None):
|
||||
n_predict = 512
|
||||
global server
|
||||
# server = ServerPreset.stories15m_moe()
|
||||
server.jinja = True
|
||||
server.n_predict = n_predict
|
||||
server.chat_template_file = f'../../../models/templates/{template_name}.jinja'
|
||||
server.start(timeout_seconds=TIMEOUT_SERVER_START)
|
||||
res = server.make_request("POST", "/chat/completions", data={
|
||||
"max_tokens": n_predict,
|
||||
"messages": [
|
||||
{"role": "system", "content": "You are a coding assistant."},
|
||||
{"role": "user", "content": "Write an example"},
|
||||
],
|
||||
"tool_choice": "required",
|
||||
"tools": [tool],
|
||||
"parallel_tool_calls": False,
|
||||
"temperature": 0.0,
|
||||
"top_k": 1,
|
||||
"top_p": 1.0,
|
||||
})
|
||||
assert res.status_code == 200, f"Expected status code 200, got {res.status_code}"
|
||||
choice = res.body["choices"][0]
|
||||
tool_calls = choice["message"].get("tool_calls")
|
||||
assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}'
|
||||
tool_call = tool_calls[0]
|
||||
expected_function_name = "python" if tool["type"] == "code_interpreter" else tool["function"]["name"]
|
||||
assert expected_function_name == tool_call["function"]["name"]
|
||||
actual_arguments = tool_call["function"]["arguments"]
|
||||
assert isinstance(actual_arguments, str)
|
||||
if argument_key is not None:
|
||||
actual_arguments = json.loads(actual_arguments)
|
||||
assert argument_key in actual_arguments, f"tool arguments: {json.dumps(actual_arguments)}, expected: {argument_key}"
|
||||
|
||||
|
||||
@pytest.mark.parametrize("template_name,tool,argument_key", [
|
||||
("google-gemma-2-2b-it", TEST_TOOL, "success"),
|
||||
("meta-llama-Llama-3.3-70B-Instruct", TEST_TOOL, "success"),
|
||||
("meta-llama-Llama-3.3-70B-Instruct", PYTHON_TOOL, "code"),
|
||||
])
|
||||
def test_completion_with_required_tool_tiny_fast(template_name: str, tool: dict, argument_key: str | None):
|
||||
do_test_completion_with_required_tool_tiny(template_name, tool, argument_key)
|
||||
|
||||
|
||||
@pytest.mark.slow
|
||||
@pytest.mark.parametrize("template_name,tool,argument_key", [
|
||||
("meta-llama-Llama-3.1-8B-Instruct", TEST_TOOL, "success"),
|
||||
("meta-llama-Llama-3.1-8B-Instruct", PYTHON_TOOL, "code"),
|
||||
("meetkai-functionary-medium-v3.1", TEST_TOOL, "success"),
|
||||
("meetkai-functionary-medium-v3.1", PYTHON_TOOL, "code"),
|
||||
("meetkai-functionary-medium-v3.2", TEST_TOOL, "success"),
|
||||
("meetkai-functionary-medium-v3.2", PYTHON_TOOL, "code"),
|
||||
("NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use", TEST_TOOL, "success"),
|
||||
("NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use", PYTHON_TOOL, "code"),
|
||||
("meta-llama-Llama-3.2-3B-Instruct", TEST_TOOL, "success"),
|
||||
("meta-llama-Llama-3.2-3B-Instruct", PYTHON_TOOL, "code"),
|
||||
("mistralai-Mistral-Nemo-Instruct-2407", TEST_TOOL, "success"),
|
||||
("mistralai-Mistral-Nemo-Instruct-2407", PYTHON_TOOL, "code"),
|
||||
("NousResearch-Hermes-3-Llama-3.1-8B-tool_use", TEST_TOOL, "success"),
|
||||
("NousResearch-Hermes-3-Llama-3.1-8B-tool_use", PYTHON_TOOL, "code"),
|
||||
("deepseek-ai-DeepSeek-R1-Distill-Llama-8B", TEST_TOOL, "success"),
|
||||
("deepseek-ai-DeepSeek-R1-Distill-Llama-8B", PYTHON_TOOL, "code"),
|
||||
("fireworks-ai-llama-3-firefunction-v2", TEST_TOOL, "success"),
|
||||
("fireworks-ai-llama-3-firefunction-v2", PYTHON_TOOL, "code"),
|
||||
])
|
||||
def test_completion_with_required_tool_tiny_slow(template_name: str, tool: dict, argument_key: str | None):
|
||||
do_test_completion_with_required_tool_tiny(template_name, tool, argument_key)
|
||||
|
||||
|
||||
@pytest.mark.slow
|
||||
@pytest.mark.parametrize("tool,argument_key,hf_repo,template_override", [
|
||||
(TEST_TOOL, "success", "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
|
||||
(PYTHON_TOOL, "code", "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
|
||||
(TEST_TOOL, "success", "bartowski/gemma-2-2b-it-GGUF:Q4_K_M", None),
|
||||
(PYTHON_TOOL, "code", "bartowski/gemma-2-2b-it-GGUF:Q4_K_M", None),
|
||||
(TEST_TOOL, "success", "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
|
||||
(PYTHON_TOOL, "code", "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
|
||||
(TEST_TOOL, "success", "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", None),
|
||||
(PYTHON_TOOL, "code", "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", None),
|
||||
(TEST_TOOL, "success", "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
|
||||
(PYTHON_TOOL, "code", "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
|
||||
(TEST_TOOL, "success", "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-3-Llama-3.1-8B", "tool_use")),
|
||||
(PYTHON_TOOL, "code", "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-3-Llama-3.1-8B", "tool_use")),
|
||||
(TEST_TOOL, "success", "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
|
||||
(PYTHON_TOOL, "code", "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
|
||||
(TEST_TOOL, "success", "bartowski/functionary-small-v3.2-GGUF:Q8_0", ("meetkai/functionary-medium-v3.2", None)),
|
||||
(PYTHON_TOOL, "code", "bartowski/functionary-small-v3.2-GGUF:Q8_0", ("meetkai/functionary-medium-v3.2", None)),
|
||||
(TEST_TOOL, "success", "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)),
|
||||
(PYTHON_TOOL, "code", "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)),
|
||||
(TEST_TOOL, "success", "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)),
|
||||
(PYTHON_TOOL, "code", "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)),
|
||||
# TODO: fix these
|
||||
# (TEST_TOOL, "success", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
|
||||
# (PYTHON_TOOL, "code", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
|
||||
])
|
||||
def test_completion_with_required_tool_real_model(tool: dict, argument_key: str | None, hf_repo: str, template_override: Tuple[str, str | None] | None):
|
||||
n_predict = 512
|
||||
server.n_slots = 1
|
||||
server.jinja = True
|
||||
server.n_ctx = 8192
|
||||
server.n_predict = n_predict
|
||||
server.model_hf_repo = hf_repo
|
||||
server.model_hf_file = None
|
||||
if template_override:
|
||||
(template_hf_repo, template_variant) = template_override
|
||||
server.chat_template_file = f"../../../models/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_chat_template.py {template_hf_repo} {template_variant} > {server.chat_template_file}` to download the template."
|
||||
server.start(timeout_seconds=TIMEOUT_SERVER_START)
|
||||
res = server.make_request("POST", "/chat/completions", data={
|
||||
"max_tokens": n_predict,
|
||||
"messages": [
|
||||
{"role": "system", "content": "You are a coding assistant."},
|
||||
{"role": "user", "content": "Write an example"},
|
||||
],
|
||||
"tool_choice": "required",
|
||||
"tools": [tool],
|
||||
"parallel_tool_calls": False,
|
||||
"temperature": 0.0,
|
||||
"top_k": 1,
|
||||
"top_p": 1.0,
|
||||
}, timeout=TIMEOUT_HTTP_REQUEST)
|
||||
assert res.status_code == 200, f"Expected status code 200, got {res.status_code}"
|
||||
choice = res.body["choices"][0]
|
||||
tool_calls = choice["message"].get("tool_calls")
|
||||
assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}'
|
||||
tool_call = tool_calls[0]
|
||||
expected_function_name = "python" if tool["type"] == "code_interpreter" else tool["function"]["name"]
|
||||
assert expected_function_name == tool_call["function"]["name"]
|
||||
actual_arguments = tool_call["function"]["arguments"]
|
||||
assert isinstance(actual_arguments, str)
|
||||
if argument_key is not None:
|
||||
actual_arguments = json.loads(actual_arguments)
|
||||
assert argument_key in actual_arguments, f"tool arguments: {json.dumps(actual_arguments)}, expected: {argument_key}"
|
||||
|
||||
|
||||
def do_test_completion_without_tool_call(template_name: str, n_predict: int, tools: list[dict], tool_choice: str | None):
|
||||
global server
|
||||
server.jinja = True
|
||||
server.n_predict = n_predict
|
||||
server.chat_template_file = f'../../../models/templates/{template_name}.jinja'
|
||||
server.start(timeout_seconds=TIMEOUT_SERVER_START)
|
||||
res = server.make_request("POST", "/chat/completions", data={
|
||||
"max_tokens": n_predict,
|
||||
"messages": [
|
||||
{"role": "system", "content": "You are a coding assistant."},
|
||||
{"role": "user", "content": "say hello world with python"},
|
||||
],
|
||||
"tools": tools if tools else None,
|
||||
"tool_choice": tool_choice,
|
||||
"temperature": 0.0,
|
||||
"top_k": 1,
|
||||
"top_p": 1.0,
|
||||
}, timeout=TIMEOUT_HTTP_REQUEST)
|
||||
assert res.status_code == 200, f"Expected status code 200, got {res.status_code}"
|
||||
choice = res.body["choices"][0]
|
||||
assert choice["message"].get("tool_calls") is None, f'Expected no tool call in {choice["message"]}'
|
||||
|
||||
|
||||
@pytest.mark.parametrize("template_name,n_predict,tools,tool_choice", [
|
||||
("meta-llama-Llama-3.3-70B-Instruct", 128, [], None),
|
||||
("meta-llama-Llama-3.3-70B-Instruct", 128, [TEST_TOOL], None),
|
||||
("meta-llama-Llama-3.3-70B-Instruct", 128, [PYTHON_TOOL], 'none'),
|
||||
])
|
||||
def test_completion_without_tool_call_fast(template_name: str, n_predict: int, tools: list[dict], tool_choice: str | None):
|
||||
do_test_completion_without_tool_call(template_name, n_predict, tools, tool_choice)
|
||||
|
||||
|
||||
@pytest.mark.slow
|
||||
@pytest.mark.parametrize("template_name,n_predict,tools,tool_choice", [
|
||||
("meetkai-functionary-medium-v3.2", 256, [], None),
|
||||
("meetkai-functionary-medium-v3.2", 256, [TEST_TOOL], None),
|
||||
("meetkai-functionary-medium-v3.2", 256, [PYTHON_TOOL], 'none'),
|
||||
("meetkai-functionary-medium-v3.1", 256, [], None),
|
||||
("meetkai-functionary-medium-v3.1", 256, [TEST_TOOL], None),
|
||||
("meetkai-functionary-medium-v3.1", 256, [PYTHON_TOOL], 'none'),
|
||||
("meta-llama-Llama-3.2-3B-Instruct", 256, [], None),
|
||||
("meta-llama-Llama-3.2-3B-Instruct", 256, [TEST_TOOL], None),
|
||||
("meta-llama-Llama-3.2-3B-Instruct", 256, [PYTHON_TOOL], 'none'),
|
||||
])
|
||||
def test_completion_without_tool_call_slow(template_name: str, n_predict: int, tools: list[dict], tool_choice: str | None):
|
||||
do_test_completion_without_tool_call(template_name, n_predict, tools, tool_choice)
|
||||
|
||||
|
||||
@pytest.mark.slow
|
||||
@pytest.mark.parametrize("hf_repo,template_override", [
|
||||
("bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
|
||||
("bartowski/gemma-2-2b-it-GGUF:Q4_K_M", None),
|
||||
("bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
|
||||
("bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", None),
|
||||
("bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
|
||||
("bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-3-Llama-3.1-8B", "tool_use")),
|
||||
("bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
|
||||
("bartowski/functionary-small-v3.2-GGUF:Q8_0", ("meetkai/functionary-medium-v3.2", None)),
|
||||
("bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)),
|
||||
# ("bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)),
|
||||
# ("bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
|
||||
])
|
||||
def test_weather_tool_call(hf_repo: str, template_override: Tuple[str, str | None] | None):
|
||||
global server
|
||||
server.n_slots = 1
|
||||
server.jinja = True
|
||||
server.n_ctx = 8192
|
||||
server.n_predict = 512
|
||||
server.model_hf_repo = hf_repo
|
||||
server.model_hf_file = None
|
||||
if template_override:
|
||||
(template_hf_repo, template_variant) = template_override
|
||||
server.chat_template_file = f"../../../models/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_chat_template.py {template_hf_repo} {template_variant} > {server.chat_template_file}` to download the template."
|
||||
server.start(timeout_seconds=TIMEOUT_SERVER_START)
|
||||
res = server.make_request("POST", "/chat/completions", data={
|
||||
"max_tokens": 256,
|
||||
"messages": [
|
||||
{"role": "user", "content": "What is the weather in Istanbul?"},
|
||||
],
|
||||
"tools": [WEATHER_TOOL],
|
||||
}, timeout=TIMEOUT_HTTP_REQUEST)
|
||||
assert res.status_code == 200, f"Expected status code 200, got {res.status_code}"
|
||||
choice = res.body["choices"][0]
|
||||
tool_calls = choice["message"].get("tool_calls")
|
||||
assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}'
|
||||
tool_call = tool_calls[0]
|
||||
assert tool_call["function"]["name"] == WEATHER_TOOL["function"]["name"]
|
||||
actual_arguments = json.loads(tool_call["function"]["arguments"])
|
||||
assert 'location' in actual_arguments, f"location not found in {json.dumps(actual_arguments)}"
|
||||
location = actual_arguments["location"]
|
||||
assert isinstance(location, str), f"Expected location to be a string, got {type(location)}: {json.dumps(location)}"
|
||||
assert re.match('^Istanbul(, (TR|Turkey|Türkiye))?$', location), f'Expected Istanbul for location, got {location}'
|
||||
|
||||
|
||||
@pytest.mark.slow
|
||||
@pytest.mark.parametrize("expected_arguments_override,hf_repo,template_override", [
|
||||
(None, "bartowski/gemma-2-2b-it-GGUF:Q4_K_M", None),
|
||||
(None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
|
||||
(None, "bartowski/functionary-small-v3.2-GGUF:Q8_0", ("meetkai-functionary-medium-v3.2", None)),
|
||||
('{"code":"print("}', "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
|
||||
(None, "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama-Llama-3.2-3B-Instruct", None)),
|
||||
('{"code":"print("}', "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", ("meta-llama-Llama-3.2-3B-Instruct", None)),
|
||||
(None, "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", None),
|
||||
(None, "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
|
||||
(None, "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", ("NousResearch-Hermes-3-Llama-3.1-8B", "tool_use")),
|
||||
(None, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
|
||||
# (None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
|
||||
])
|
||||
def test_hello_world_tool_call(expected_arguments_override: str | None, hf_repo: str, template_override: Tuple[str, str | None] | None):
|
||||
global server
|
||||
server.n_slots = 1
|
||||
server.jinja = True
|
||||
server.n_ctx = 8192
|
||||
server.n_predict = 128
|
||||
server.model_hf_repo = hf_repo
|
||||
server.model_hf_file = None
|
||||
if template_override:
|
||||
(template_hf_repo, template_variant) = template_override
|
||||
server.chat_template_file = f"../../../models/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_chat_template.py {template_hf_repo} {template_variant} > {server.chat_template_file}` to download the template."
|
||||
server.start(timeout_seconds=TIMEOUT_SERVER_START)
|
||||
res = server.make_request("POST", "/chat/completions", data={
|
||||
"max_tokens": 256,
|
||||
"messages": [
|
||||
{"role": "system", "content": "You are a coding assistant."},
|
||||
{"role": "user", "content": "say hello world with python"},
|
||||
],
|
||||
"tools": [PYTHON_TOOL],
|
||||
# Note: without these greedy params, Functionary v3.2 writes `def hello_world():\n print("Hello, World!")\nhello_world()` which is correct but a pain to test.
|
||||
"temperature": 0.0,
|
||||
"top_k": 1,
|
||||
"top_p": 1.0,
|
||||
}, timeout=TIMEOUT_HTTP_REQUEST)
|
||||
assert res.status_code == 200, f"Expected status code 200, got {res.status_code}"
|
||||
choice = res.body["choices"][0]
|
||||
tool_calls = choice["message"].get("tool_calls")
|
||||
assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}'
|
||||
tool_call = tool_calls[0]
|
||||
assert tool_call["function"]["name"] == PYTHON_TOOL["function"]["name"]
|
||||
actual_arguments = tool_call["function"]["arguments"]
|
||||
if expected_arguments_override is not None:
|
||||
assert actual_arguments == expected_arguments_override
|
||||
else:
|
||||
actual_arguments = json.loads(actual_arguments)
|
||||
assert 'code' in actual_arguments, f"code not found in {json.dumps(actual_arguments)}"
|
||||
code = actual_arguments["code"]
|
||||
assert isinstance(code, str), f"Expected code to be a string, got {type(code)}: {json.dumps(code)}"
|
||||
assert re.match(r'''print\(("[Hh]ello,? [Ww]orld!?"|'[Hh]ello,? [Ww]orld!?')\)''', code), f'Expected hello world, got {code}'
|
||||
|
|
@ -26,7 +26,7 @@ from re import RegexFlag
|
|||
import wget
|
||||
|
||||
|
||||
DEFAULT_HTTP_TIMEOUT = 10 if "LLAMA_SANITIZE" not in os.environ else 30
|
||||
DEFAULT_HTTP_TIMEOUT = 12 if "LLAMA_SANITIZE" not in os.environ else 30
|
||||
|
||||
|
||||
class ServerResponse:
|
||||
|
|
@ -41,7 +41,7 @@ class ServerProcess:
|
|||
server_port: int = 8080
|
||||
server_host: str = "127.0.0.1"
|
||||
model_hf_repo: str = "ggml-org/models"
|
||||
model_hf_file: str = "tinyllamas/stories260K.gguf"
|
||||
model_hf_file: str | None = "tinyllamas/stories260K.gguf"
|
||||
model_alias: str = "tinyllama-2"
|
||||
temperature: float = 0.8
|
||||
seed: int = 42
|
||||
|
|
@ -191,7 +191,7 @@ class ServerProcess:
|
|||
creationflags=flags,
|
||||
stdout=sys.stdout,
|
||||
stderr=sys.stdout,
|
||||
env={**os.environ, "LLAMA_CACHE": "tmp"},
|
||||
env={**os.environ, "LLAMA_CACHE": "tmp"} if "LLAMA_CACHE" not in os.environ else None,
|
||||
)
|
||||
server_instances.add(self)
|
||||
|
||||
|
|
|
|||
|
|
@ -17,6 +17,7 @@
|
|||
#define JSON_ASSERT GGML_ASSERT
|
||||
#include "json.hpp"
|
||||
#include "minja.hpp"
|
||||
#include "chat.hpp"
|
||||
#include "chat-template.hpp"
|
||||
|
||||
#include <random>
|
||||
|
|
@ -376,7 +377,7 @@ inline std::string format_chat(const common_chat_template & tmpl, const std::vec
|
|||
throw std::runtime_error("Missing 'content' (ref: https://github.com/ggerganov/llama.cpp/issues/8367)");
|
||||
}
|
||||
|
||||
chat.push_back({role, content});
|
||||
chat.push_back({role, content, /* tool_calls= */ {}});
|
||||
}
|
||||
|
||||
const auto formatted_chat = common_chat_apply_template(tmpl, chat, true, /* use_jinja= */ false);
|
||||
|
|
@ -483,14 +484,13 @@ static bool ends_with(const std::string & str, const std::string & suffix) {
|
|||
|
||||
static size_t find_partial_stop_string(const std::string &stop, const std::string &text) {
|
||||
if (!text.empty() && !stop.empty()) {
|
||||
const char text_last_char = text.back();
|
||||
for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) {
|
||||
if (stop[char_index] == text_last_char) {
|
||||
const std::string current_partial = stop.substr(0, char_index + 1);
|
||||
if (ends_with(text, current_partial)) {
|
||||
return text.size() - char_index - 1;
|
||||
}
|
||||
auto it = std::find(stop.rbegin(), stop.rend(), text.back());
|
||||
while (it != stop.rend()) {
|
||||
size_t length = std::distance(it, stop.rend());
|
||||
if (text.length() >= length && 0 == text.compare(text.length() - length, length, stop)) {
|
||||
return text.length() - length;
|
||||
}
|
||||
it = std::find(std::next(it), stop.rend(), text.back());
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -580,21 +580,30 @@ static json oaicompat_completion_params_parse(const json & body) {
|
|||
|
||||
static json oaicompat_completion_params_parse(
|
||||
const json & body, /* openai api json semantics */
|
||||
const common_chat_template & tmpl,
|
||||
bool use_jinja)
|
||||
bool use_jinja,
|
||||
const common_chat_templates & chat_templates)
|
||||
{
|
||||
json llama_params;
|
||||
const auto & tmpl = body.contains("tools") && chat_templates.template_tool_use
|
||||
? *chat_templates.template_tool_use
|
||||
: *chat_templates.template_default;
|
||||
|
||||
auto tools = json_value(body, "tools", json());
|
||||
auto has_tools = tools.is_array() && !tools.empty();
|
||||
auto stream = json_value(body, "stream", false);
|
||||
|
||||
if (has_tools) {
|
||||
if (use_jinja) {
|
||||
LOG_WRN("tools param is not fully supported yet\n");
|
||||
} else {
|
||||
if (tools.is_array() && !tools.empty()) {
|
||||
if (stream) {
|
||||
throw std::runtime_error("Cannot use tools with stream");
|
||||
}
|
||||
if (!use_jinja) {
|
||||
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()) {
|
||||
|
|
@ -619,7 +628,38 @@ static json oaicompat_completion_params_parse(
|
|||
|
||||
// Apply chat template to the list of messages
|
||||
if (use_jinja) {
|
||||
llama_params["prompt"] = tmpl.apply(body.at("messages"), tools, /* add_generation_prompt= */ true);
|
||||
auto tool_choice = json_value(body, "tool_choice", std::string("auto"));
|
||||
if (tool_choice != "none" && tool_choice != "auto" && tool_choice != "required") {
|
||||
throw std::runtime_error("Invalid tool_choice: " + tool_choice);
|
||||
}
|
||||
if (tool_choice != "none" && llama_params.contains("grammar")) {
|
||||
throw std::runtime_error("Cannot use custom grammar constraints with tools.");
|
||||
}
|
||||
common_chat_inputs inputs;
|
||||
inputs.messages = body.at("messages");
|
||||
inputs.tools = tools;
|
||||
inputs.tool_choice = tool_choice;
|
||||
inputs.parallel_tool_calls = json_value(body, "parallel_tool_calls", false);
|
||||
inputs.stream = stream;
|
||||
// TODO: support mixing schema w/ tools beyond generic format.
|
||||
inputs.json_schema = json_value(llama_params, "json_schema", json::object());
|
||||
auto chat_params = common_chat_params_init(tmpl, inputs);
|
||||
|
||||
llama_params["chat_format"] = static_cast<int>(chat_params.format);
|
||||
llama_params["prompt"] = chat_params.prompt;
|
||||
llama_params["grammar"] = chat_params.grammar;
|
||||
llama_params["grammar_lazy"] = chat_params.grammar_lazy;
|
||||
auto grammar_triggers = json::array();
|
||||
for (const auto & trigger : chat_params.grammar_triggers) {
|
||||
grammar_triggers.push_back({
|
||||
{"word", trigger.word},
|
||||
{"at_start", trigger.at_start},
|
||||
});
|
||||
}
|
||||
llama_params["grammar_triggers"] = grammar_triggers;
|
||||
for (const auto & stop : chat_params.additional_stops) {
|
||||
llama_params["stop"].push_back(stop);
|
||||
}
|
||||
} else {
|
||||
llama_params["prompt"] = format_chat(tmpl, body.at("messages"));
|
||||
}
|
||||
|
|
@ -638,14 +678,6 @@ 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
|
||||
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
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue