Files
llama.cpp/tools/server/tests/unit/test_template.py
T
Pascal a946eb541a server: answer reasoning_effort levels the template rejects with 400 (#32332)
The reasoning effort probe already renders every candidate level, but
it dropped the ones the template raises on, so a request with such a
level (e.g. max on Qwen3.8) reached the template and came back as a
500 on /v1/chat/completions, /v1/responses and /v1/messages.

Keep the raising levels in reasoning_effort_rejected and reject them
in oaicompat_chat_params_parse with an invalid_argument, which the
server answers with a 400 listing the supported levels. Aliases and
levels a template ignores are not rejected, so high on Qwen3.8
still renders as xhigh.

Also trim the output_config.effort test to a single level.
2026-11-16 06:55:56 +01:00

127 lines
4.8 KiB
Python

#!/usr/bin/env python
import pytest
# ensure grandparent path is in sys.path
from pathlib import Path
import sys
from unit.test_tool_call import TEST_TOOL
path = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(path))
import datetime
from utils import *
from typing import Literal
server: ServerProcess
@pytest.fixture(autouse=True)
def create_server():
global server
server = ServerPreset.tinyllama2()
server.model_alias = "tinyllama-2"
server.n_slots = 1
def test_reasoning_effort_rejected_by_template():
global server
server.jinja = True
# the template raises on max, and high is one of its levels
server.chat_template_file = '../../../models/templates/IFM-K2-Horizon.jinja'
server.start()
messages = [{"role": "user", "content": "Hello"}]
for path, data in [
("/chat/completions/input_tokens", {"messages": messages, "reasoning_effort": "max"}),
("/v1/responses/input_tokens", {"input": "Hello", "reasoning": {"effort": "max"}}),
("/v1/messages/count_tokens", {"model": "test", "messages": messages, "output_config": {"effort": "max"}}),
]:
res = server.make_request("POST", path, data=data)
assert res.status_code == 400, f"{path}: {res.body}"
res = server.make_request("POST", "/chat/completions/input_tokens", data={"messages": messages, "reasoning_effort": "high"})
assert res.status_code == 200, res.body
@pytest.mark.parametrize("tools", [None, [], [TEST_TOOL]])
@pytest.mark.parametrize("template_name,reasoning,expected_end", [
("deepseek-ai-DeepSeek-R1-Distill-Qwen-32B", "on", "<think>\n"),
("deepseek-ai-DeepSeek-R1-Distill-Qwen-32B","auto", "<think>\n"),
("deepseek-ai-DeepSeek-R1-Distill-Qwen-32B", "off", "<think>\n</think>"),
("Qwen-Qwen3-0.6B","auto", "<|im_start|>assistant\n"),
("Qwen-Qwen3-0.6B", "off", "<|im_start|>assistant\n<think>\n\n</think>\n\n"),
("Qwen-QwQ-32B","auto", "<|im_start|>assistant\n<think>\n"),
("Qwen-QwQ-32B", "off", "<|im_start|>assistant\n<think>\n</think>"),
("CohereForAI-c4ai-command-r7b-12-2024-tool_use","auto", "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>"),
("CohereForAI-c4ai-command-r7b-12-2024-tool_use", "off", "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|><|START_THINKING|><|END_THINKING|>"),
])
def test_reasoning(template_name: str, reasoning: Literal['on', 'off', 'auto'] | None, expected_end: str, tools: list[dict]):
global server
server.jinja = True
server.reasoning = reasoning
server.chat_template_file = f'../../../models/templates/{template_name}.jinja'
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"),
("fireworks-ai-llama-3-firefunction-v2", "%b %d %Y"),
])
def test_date_inside_prompt(template_name: str, format: str, tools: list[dict]):
global server
server.jinja = True
server.chat_template_file = f'../../../models/templates/{template_name}.jinja'
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"]
today_str = datetime.date.today().strftime(format)
assert today_str in prompt, f"Expected today's date ({today_str}) in content ({prompt})"
@pytest.mark.parametrize("add_generation_prompt", [False, True])
@pytest.mark.parametrize("template_name,expected_generation_prompt", [
("meta-llama-Llama-3.3-70B-Instruct", "<|start_header_id|>assistant<|end_header_id|>"),
])
def test_add_generation_prompt(template_name: str, expected_generation_prompt: str, add_generation_prompt: bool):
global server
server.jinja = True
server.chat_template_file = f'../../../models/templates/{template_name}.jinja'
server.start()
res = server.make_request("POST", "/apply-template", data={
"messages": [
{"role": "user", "content": "What is today?"},
],
"add_generation_prompt": add_generation_prompt,
})
assert res.status_code == 200
prompt = res.body["prompt"]
if add_generation_prompt:
assert expected_generation_prompt in prompt, f"Expected generation prompt ({expected_generation_prompt}) in content ({prompt})"
else:
assert expected_generation_prompt not in prompt, f"Did not expect generation prompt ({expected_generation_prompt}) in content ({prompt})"