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

Use --reasoning-format, remove forced thinking for now

Browse source
This commit is contained in:
ochafik 2025-02-08 17:58:33 +00:00
parent cc2c712cf9
commit c0f972bb45
10 changed files with 184 additions and 246 deletions
Download patch file Download diff file Expand all files Collapse all files

8
examples/server/README.md
View file

@ -127,7 +127,7 @@ The project is under active development, and we are [looking for feedback and co
| `--grammar-file FNAME` | file to read grammar from |
| `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
| `--jinja` | Enable experimental Jinja templating engine (required for tool use) |
| `--think` | Enable experimental thinking mode (extracts DeepSeek R1 & Command R7B's native thinking tags and forces any other model to think before responding, resulting thoughts are in the `reasoning_content` output field) (requires `--jinja`) |
| `--reasoning-format FORMAT` | Controls extraction of model thinking traces and the format / field in which they are returned (default: `deepseek`; allowed values: `deepseek`, `none`; requires `--jinja`). `none` will leave thinking traces inline in `message.content` in a model-specific format, while `deepseek` will return them separately under `message.reasoning_content` |
**Example-specific params**
@ -1224,10 +1224,10 @@ curl http://localhost:8080/v1/chat/completions \
# Native support for DeepSeek R1 works best w/ our own template (official template buggy)
llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q6_K_L --think \
llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q6_K_L \
--chat-template-file models/templates/llama-cpp-deepseek-r1.jinja
llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M --think \
llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M \
--chat-template-file models/templates/llama-cpp-deepseek-r1.jinja
# Native support requires the right template for these GGUFs:
@ -1241,7 +1241,7 @@ curl http://localhost:8080/v1/chat/completions \
llama-server --jinja -fa -hf bartowski/firefunction-v2-GGUF -hff firefunction-v2-IQ1_M.gguf \
--chat-template-file <( python scripts/get_chat_template.py fireworks-ai/llama-3-firefunction-v2 tool_use )
llama-server --jinja -fa -hf bartowski/c4ai-command-r7b-12-2024-GGUF:Q6_K_L --think \
llama-server --jinja -fa -hf bartowski/c4ai-command-r7b-12-2024-GGUF:Q6_K_L \
--chat-template-file <( python scripts/get_chat_template.py CohereForAI/c4ai-command-r7b-12-2024 tool_use )
# Generic format support

4
examples/server/server.cpp
View file

@ -4055,7 +4055,7 @@ int main(int argc, char ** argv) {
}
auto body = json::parse(req.body);
json data = oaicompat_completion_params_parse(body, params.use_jinja, params.think, ctx_server.chat_templates);
json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates);
return handle_completions_impl(
SERVER_TASK_TYPE_COMPLETION,
@ -4068,7 +4068,7 @@ int main(int argc, char ** argv) {
// same with handle_chat_completions, but without inference part
const auto handle_apply_template = [&ctx_server, &params, &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, params.think, ctx_server.chat_templates);
json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates);
res_ok(res, {{ "prompt", std::move(data.at("prompt")) }});
};

107
examples/server/tests/unit/test_tool_call.py
View file

@ -274,44 +274,44 @@ def test_completion_without_tool_call_slow(template_name: str, n_predict: int, t
@pytest.mark.slow
@pytest.mark.parametrize("think,hf_repo,template_override", [
(True, "bartowski/c4ai-command-r7b-12-2024-GGUF:Q4_K_M", ("CohereForAI/c4ai-command-r7b-12-2024", "tool_use")),
@pytest.mark.parametrize("reasoning_format,hf_repo,template_override", [
('deepseek', "bartowski/c4ai-command-r7b-12-2024-GGUF:Q4_K_M", ("CohereForAI/c4ai-command-r7b-12-2024", "tool_use")),
(False, "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
(False, "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
(None, "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"),
(False, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(False, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", "chatml"),
(False, "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", None),
(False, "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", None),
(None, "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", "chatml"),
(False, "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
(False, "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
(None, "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", "chatml"),
(False, "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-3-Llama-3.1-8B", "tool_use")),
(False, "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-3-Llama-3.1-8B", "tool_use")),
(None, "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", "chatml"),
(False, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
(False, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
(None, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", "chatml"),
(False, "bartowski/functionary-small-v3.2-GGUF:Q8_0", ("meetkai/functionary-medium-v3.2", None)),
(False, "bartowski/functionary-small-v3.2-GGUF:Q8_0", "chatml"),
(None, "bartowski/functionary-small-v3.2-GGUF:Q8_0", ("meetkai/functionary-medium-v3.2", None)),
(None, "bartowski/functionary-small-v3.2-GGUF:Q8_0", "chatml"),
(False, "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)),
(False, "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)),
(None, "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", "chatml"),
(True, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
('deepseek', "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
# Note: gemma-2-2b-it knows itself as "model", not "assistant", so we don't test the ill-suited chatml on it.
(False, "bartowski/gemma-2-2b-it-GGUF:Q4_K_M", None),
(None, "bartowski/gemma-2-2b-it-GGUF:Q4_K_M", None),
# ("bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)),
])
def test_weather(think: bool, hf_repo: str, template_override: Tuple[str, str | None] | None):
def test_weather(reasoning_format: Literal['deepseek', 'none'] | None, hf_repo: str, template_override: Tuple[str, str | None] | None):
global server
n_predict = 512
server.think = think
server.reasoning_format = reasoning_format
server.n_slots = 1
server.jinja = True
server.n_ctx = 8192
@ -440,19 +440,19 @@ def test_calc_result(result_override: str | None, n_predict: int, hf_repo: str,
@pytest.mark.slow
@pytest.mark.parametrize("n_predict,think,expect_content,expect_reasoning_content,hf_repo,template_override", [
(1024, True, "^The sum of 102 and 7 is 109.*", "^The user's request is straightforward.*", "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(128, False, "^The sum of 102 and 7 is 109.*", None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
@pytest.mark.parametrize("n_predict,reasoning_format,expect_content,expect_reasoning_content,hf_repo,template_override", [
# (1024, 'deepseek', "^The sum of 102 and 7 is 109.*", "^The user's request is straightforward.*", "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
# (128, None, "^The sum of 102 and 7 is 109.*", None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(1024, True, "To find the sum of.*", "I need to calculate the sum of 102 and 7.*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(1024, False, "<think>\nI need[\\s\\S\\r\\n]*?</think>\nTo find.*", None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(1024, 'deepseek', "To find the sum of.*", "I need to calculate the sum of 102 and 7.*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(1024, 'none', "<think>\nI need[\\s\\S\\r\\n]*?</think>\nTo find.*", None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(1024, True, "To find the sum of.*", "First, I need to add the tens place.*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
(1024, 'deepseek', "To find the sum of.*", "First, I need to add the tens place.*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
])
def test_thoughts(n_predict: int, think: bool, expect_content: str | None, expect_reasoning_content: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
def test_thoughts(n_predict: int, reasoning_format: Literal['deepseek', 'none'] | None, expect_content: str | None, expect_reasoning_content: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
global server
server.n_slots = 1
server.think = think
server.reasoning_format = reasoning_format
server.jinja = True
server.n_ctx = 8192 * 2
server.n_predict = n_predict
@ -489,45 +489,44 @@ def test_thoughts(n_predict: int, think: bool, expect_content: str | None, expec
@pytest.mark.slow
@pytest.mark.parametrize("think,expected_arguments_override,hf_repo,template_override", [
(True, None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(True, None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", "chatml"),
@pytest.mark.parametrize("expected_arguments_override,hf_repo,template_override", [
(None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
(None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", "chatml"),
(False, None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(False, None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None),
(None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", "chatml"),
(False, None, "bartowski/functionary-small-v3.2-GGUF:Q8_0", ("meetkai-functionary-medium-v3.2", None)),
(False, None, "bartowski/functionary-small-v3.2-GGUF:Q8_0", "chatml"),
(None, "bartowski/functionary-small-v3.2-GGUF:Q8_0", ("meetkai-functionary-medium-v3.2", None)),
(None, "bartowski/functionary-small-v3.2-GGUF:Q8_0", "chatml"),
(False, None, "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
(False, '{"code":"print("}', "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
('{"code":"print("}', "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"),
(False, '{"code":"print("}', "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama-Llama-3.2-3B-Instruct", None)),
(False, None, "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", "chatml"),
('{"code":"print("}', "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama-Llama-3.2-3B-Instruct", None)),
(None, "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", "chatml"),
(False, '{"code":"print("}', "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", ("meta-llama-Llama-3.2-3B-Instruct", None)),
(False, '{"code":"print("}', "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", "chatml"),
('{"code":"print("}', "bartowski/Llama-3.2-3B-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", "chatml"),
(False, None, "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", None),
(False, None, "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", None),
(None, "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", "chatml"),
(False, None, "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
(False, None, "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
(None, "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", "chatml"),
(False, None, "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", ("NousResearch-Hermes-3-Llama-3.1-8B", "tool_use")),
(False, None, "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", ("NousResearch-Hermes-3-Llama-3.1-8B", "tool_use")),
(None, "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", "chatml"),
(False, None, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
(False, None, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", "chatml"),
(None, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
(None, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", "chatml"),
# Note: gemma-2-2b-it knows itself as "model", not "assistant", so we don't test the ill-suited chatml on it.
(False, None, "bartowski/gemma-2-2b-it-GGUF:Q4_K_M", None),
(None, "bartowski/gemma-2-2b-it-GGUF:Q4_K_M", None),
])
def test_hello_world(think: bool, expected_arguments_override: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
def test_hello_world(reasoning_format: Literal['deepseek', 'none'] | None, expected_arguments_override: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
global server
server.n_slots = 1
server.jinja = True
server.think = think
server.n_ctx = 8192
server.n_predict = 512 # High because of DeepSeek R1
server.model_hf_repo = hf_repo

6
examples/server/tests/utils.py
View file

@ -78,7 +78,7 @@ class ServerProcess:
draft_max: int | None = None
no_webui: bool | None = None
jinja: bool | None = None
think: bool | None = None
reasoning_format: Literal['deepseek', 'none'] | None = None
chat_template: str | None = None
chat_template_file: str | None = None
@ -173,8 +173,8 @@ class ServerProcess:
server_args.append("--no-webui")
if self.jinja:
server_args.append("--jinja")
if self.think:
server_args.append("--think")
if self.reasoning_format:
server_args.append("--reasoning-format")
if self.chat_template:
server_args.extend(["--chat-template", self.chat_template])
if self.chat_template_file:

4
examples/server/utils.hpp
View file

@ -578,7 +578,7 @@ static json oaicompat_completion_params_parse(const json & body) {
static json oaicompat_completion_params_parse(
const json & body, /* openai api json semantics */
bool use_jinja,
bool think,
common_reasoning_format reasoning_format,
const common_chat_templates & chat_templates)
{
json llama_params;
@ -634,7 +634,7 @@ static json oaicompat_completion_params_parse(
throw std::runtime_error("Cannot use custom grammar constraints with tools.");
}
common_chat_inputs inputs;
inputs.think = think;
inputs.extract_reasoning = reasoning_format != COMMON_REASONING_FORMAT_NONE;
inputs.messages = body.at("messages");
inputs.tools = tools;
inputs.tool_choice = tool_choice;