[Reasoning] [Bugfix] Reject invalid thinking_token_budget values (#43402)

Signed-off-by: linzm1007 <linzm1007@126.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
linzm1007
2026-05-26 03:37:30 -07:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 5d09f471f4
commit 739af5c7e1
5 changed files with 152 additions and 5 deletions
@@ -0,0 +1,76 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from pydantic import ValidationError
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
from vllm.entrypoints.openai.completion.protocol import CompletionRequest
@pytest.mark.parametrize("raw_value", [-2, 0.6, 10.5])
def test_chat_completion_request_rejects_invalid_thinking_token_budget(raw_value):
with pytest.raises(ValidationError, match="thinking_token_budget"):
ChatCompletionRequest.model_validate(
{
"model": "qwen",
"messages": [{"role": "user", "content": "hello"}],
"thinking_token_budget": raw_value,
}
)
def test_chat_completion_request_accepts_valid_thinking_token_budget():
request = ChatCompletionRequest.model_validate(
{
"model": "qwen",
"messages": [{"role": "user", "content": "hello"}],
"thinking_token_budget": 10,
}
)
assert request.thinking_token_budget == 10
def test_chat_completion_request_accepts_minus_one_as_unlimited():
request = ChatCompletionRequest.model_validate(
{
"model": "qwen",
"messages": [{"role": "user", "content": "hello"}],
"thinking_token_budget": -1,
}
)
assert request.thinking_token_budget is None
@pytest.mark.parametrize("raw_value", [0.6, 3.14, -2])
def test_completion_request_rejects_invalid_thinking_token_budget(raw_value):
with pytest.raises(ValidationError, match="thinking_token_budget"):
CompletionRequest.model_validate(
{
"model": "qwen",
"prompt": "hello",
"thinking_token_budget": raw_value,
}
)
def test_completion_request_accepts_valid_thinking_token_budget():
request = CompletionRequest.model_validate(
{
"model": "qwen",
"prompt": "hello",
"thinking_token_budget": 5,
}
)
assert request.thinking_token_budget == 5
def test_completion_request_accepts_minus_one_as_unlimited():
request = CompletionRequest.model_validate(
{
"model": "qwen",
"prompt": "hello",
"thinking_token_budget": -1,
}
)
assert request.thinking_token_budget is None
+35 -1
View File
@@ -20,7 +20,7 @@ from tests.v1.sample.utils import (
)
from vllm.config import VllmConfig
from vllm.platforms import current_platform
from vllm.sampling_params import SamplingParams
from vllm.sampling_params import SamplingParams, validate_thinking_token_budget
from vllm.utils.platform_utils import is_pin_memory_available
from vllm.v1.sample.logits_processor import (
BatchUpdate,
@@ -1194,3 +1194,37 @@ def test_thinking_budget_enforced_without_penalties():
"Budget exceeded: in_end should be True so that apply_to_logits "
"forces the end token"
)
@pytest.mark.parametrize(
("raw_value", "expected"),
[
(None, None),
(-1, None),
(10, 10),
(0, 0),
],
)
def test_validate_thinking_token_budget(raw_value, expected):
assert validate_thinking_token_budget(raw_value) == expected
def test_sampling_params_minus_one_normalizes_to_none():
params = SamplingParams(thinking_token_budget=-1)
assert params.thinking_token_budget is None
@pytest.mark.parametrize("invalid_budget", [-2, 0.6, 10.5, True])
def test_validate_thinking_token_budget_rejects_invalid(invalid_budget):
from vllm.exceptions import VLLMValidationError
with pytest.raises(VLLMValidationError, match="thinking_token_budget"):
validate_thinking_token_budget(invalid_budget)
@pytest.mark.parametrize("invalid_budget", [-2, 0.6, 10.5])
def test_thinking_budget_invalid_budget_rejected(invalid_budget):
from vllm.exceptions import VLLMValidationError
with pytest.raises(VLLMValidationError, match="thinking_token_budget"):
SamplingParams(thinking_token_budget=invalid_budget)