ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

server: add --reasoning-budget 0 to disable thinking (incl. qwen3 w/ enable_thinking:false) (#13771)

Browse source 
---------

Co-authored-by: ochafik <ochafik@google.com>
Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com>
This commit is contained in:
Olivier Chafik 2025-05-26 00:30:51 +01:00 • committed by GitHub
parent 2f099b510f
commit e121edc432
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
12 changed files with 277 additions and 107 deletions
Download patch file Download diff file Expand all files Collapse all files

34
tools/server/tests/unit/test_template.py
View file

@ -25,6 +25,40 @@ def create_server():
server.n_slots = 1
@pytest.mark.parametrize("tools", [None, [], [TEST_TOOL]])
@pytest.mark.parametrize("template_name,reasoning_budget,expected_end", [
("deepseek-ai-DeepSeek-R1-Distill-Qwen-32B", None, "<think>\n"),
("deepseek-ai-DeepSeek-R1-Distill-Qwen-32B", -1, "<think>\n"),
("deepseek-ai-DeepSeek-R1-Distill-Qwen-32B", 0, "<think>\n</think>"),
("Qwen-Qwen3-0.6B", -1, "<|im_start|>assistant\n"),
("Qwen-Qwen3-0.6B", 0, "<|im_start|>assistant\n<think>\n\n</think>\n\n"),
("Qwen-QwQ-32B", -1, "<|im_start|>assistant\n<think>\n"),
("Qwen-QwQ-32B", 0, "<|im_start|>assistant\n<think>\n</think>"),
("CohereForAI-c4ai-command-r7b-12-2024-tool_use", -1, "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>"),
("CohereForAI-c4ai-command-r7b-12-2024-tool_use", 0, "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|><|START_THINKING|><|END_THINKING|>"),
])
def test_reasoning_budget(template_name: str, reasoning_budget: int | None, expected_end: str, tools: list[dict]):
global server
server.jinja = True
server.reasoning_budget = reasoning_budget
server.chat_template_file = f'../../../models/templates/{template_name}.jinja'
server.start(timeout_seconds=TIMEOUT_SERVER_START)
res = server.make_request("POST", "/apply-template", data={
"messages": [
{"role": "user", "content": "What is today?"},
],
"tools": tools,
})
assert res.status_code == 200
prompt = res.body["prompt"]
assert prompt.endswith(expected_end), f"Expected prompt to end with '{expected_end}', got '{prompt}'"
@pytest.mark.parametrize("tools", [None, [], [TEST_TOOL]])
@pytest.mark.parametrize("template_name,format", [
("meta-llama-Llama-3.3-70B-Instruct", "%d %b %Y"),

5
tools/server/tests/utils.py
View file

@ -84,7 +84,8 @@ class ServerProcess:
draft_max: int | None = None
no_webui: bool | None = None
jinja: bool | None = None
reasoning_format: Literal['deepseek', 'none'] | None = None
reasoning_format: Literal['deepseek', 'none', 'nothink'] | None = None
reasoning_budget: int | None = None
chat_template: str | None = None
chat_template_file: str | None = None
server_path: str | None = None
@ -191,6 +192,8 @@ class ServerProcess:
server_args.append("--jinja")
if self.reasoning_format is not None:
server_args.extend(("--reasoning-format", self.reasoning_format))
if self.reasoning_budget is not None:
server_args.extend(("--reasoning-budget", self.reasoning_budget))
if self.chat_template:
server_args.extend(["--chat-template", self.chat_template])
if self.chat_template_file: