5133e1d285 · tduncan/Nvidia-Optimization-Stack-VLLM

tests/entrypoints/openai/test_render_parity.py

+14

@@ -29,6 +29,7 @@ from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionReque

29 from vllm.entrypoints.openai.responses.protocol import ResponsesRequest

30 from vllm.entrypoints.openai.responses.serving import OpenAIServingResponses

31 from vllm.entrypoints.serve.engine.protocol import ErrorResponse

32 +from vllm.exceptions import VLLMValidationError

33 from vllm.inputs import tokens_input

34 from vllm.renderers.online_renderer import OnlineRenderer

35 from vllm.renderers.params import ChatParams

@@ -427,6 +428,19 @@ class TestReasoningRenderParity:

428 },

429 )

430

431 + async def test_reasoning_effort_none_rejected_by_harmony(self, online_renderer):

432 + """Harmony rejects reasoning_effort="none" as invalid user input."""

433 + request = ChatCompletionRequest(

434 + model=_MODEL,

435 + messages=_USER,

436 + reasoning_effort="none",

437 + )

438 + with pytest.raises(

439 + VLLMValidationError, match="Harmony does not support"

440 + ) as exc_info:

441 + online_renderer._make_request_with_harmony(request)

442 + assert exc_info.value.parameter == "reasoning_effort"

443 +

444 async def test_explicit_enable_thinking_not_overridden(

445 self, online_renderer, serving_responses

446 ):

tests/entrypoints/unit_tests/test_chat_utils.py

+10

@@ -21,6 +21,7 @@ from vllm.entrypoints.chat_utils import (

21 _postprocess_messages,

22 parse_chat_messages,

23 parse_chat_messages_async,

24 + validate_chat_template,

25 )

26 from vllm.exceptions import VLLMValidationError

27 from vllm.inputs import MultiModalDataDict, MultiModalUUIDDict

@@ -3138,3 +3139,12 @@ def test_tool_call_arguments_multiple_independent(caplog):

3139

3140 assert len(caplog.records) == 1

3141 assert "bad" in caplog.records[0].message

3142 +

3143 +

3144 +def test_validate_chat_template_rejects_invalid_type():

3145 + """A non-str/Path chat_template is invalid user input."""

3146 + with pytest.raises(

3147 + VLLMValidationError, match="not a valid chat template type"

3148 + ) as exc_info:

3149 + validate_chat_template(123) # type: ignore[arg-type]

3150 + assert exc_info.value.parameter == "chat_template"

tests/tokenizers_/test_mistral.py

+14 -1

@@ -11,6 +11,7 @@ from mistral_common.guidance.grammar_factory import GrammarFactory

11 from mistral_common.tokens.tokenizers.base import SpecialTokenPolicy, SpecialTokens

12

13 from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest

14 +from vllm.exceptions import VLLMValidationError

15 from vllm.tokenizers.mistral import (

16 MistralTokenizer,

17 _validate_apply_chat_template_args,

@@ -2279,8 +2280,20 @@ def test_validate_request_params_rejects_unsupported_effort() -> None:

2280 messages=[],

2281 reasoning_effort="medium",

2282 )

2282 - with pytest.raises(ValueError, match="reasoning_effort"):

2283 + with pytest.raises(VLLMValidationError, match="reasoning_effort") as exc_info:

2284 validate_request_params(request)

2285 + assert exc_info.value.parameter == "reasoning_effort"

2286 +

2287 +

2288 +def test_validate_request_params_rejects_chat_template() -> None:

2289 + request = ChatCompletionRequest(

2290 + model="test-model",

2291 + messages=[],

2292 + chat_template="{{ messages }}",

2293 + )

2294 + with pytest.raises(VLLMValidationError, match="chat_template") as exc_info:

2295 + validate_request_params(request)

2296 + assert exc_info.value.parameter == "chat_template"

2297

2298

2299 # ---------------------------------------------------------------------------

vllm/entrypoints/chat_utils.py

+4 -1

@@ -1454,7 +1454,10 @@ def validate_chat_template(chat_template: Path | str | None):

1454 )

1455

1456 else:

1457 - raise TypeError(f"{type(chat_template)} is not a valid chat template type")

1457 + raise VLLMValidationError(

1458 + f"{type(chat_template)} is not a valid chat template type",

1459 + parameter="chat_template",

1460 + )

1461

1462

1463 def _load_chat_template(

vllm/renderers/online_renderer.py

+4 -1

@@ -512,7 +512,10 @@ class OnlineRenderer:

512 assert not self.supports_browsing

513 assert not self.supports_code_interpreter

514 if (reasoning_effort := request.reasoning_effort) == "none":

515 - raise ValueError(f"Harmony does not support {reasoning_effort=}")

515 + raise VLLMValidationError(

516 + f"Harmony does not support {reasoning_effort=}",

517 + parameter="reasoning_effort",

518 + )

519 tools = request.tools if should_include_tools else None

520 messages.extend(

521 build_harmony_preamble(

vllm/renderers/params.py

+8 -2

@@ -445,9 +445,15 @@ class TokenizeParams:

445 return tokens

446

447 if tokenizer is None:

448 - raise ValueError("Cannot pad tokens when `skip_tokenizer_init=True`")

448 + raise VLLMValidationError(

449 + "Cannot pad tokens when `skip_tokenizer_init=True`",

450 + parameter="pad_prompt_tokens",

451 + )

452 if not isinstance(tokens, list):

450 - raise ValueError("Cannot pad tokens for embedding inputs")

453 + raise VLLMValidationError(

454 + "Cannot pad tokens for embedding inputs",

455 + parameter="pad_prompt_tokens",

456 + )

457

458 return tokens + [tokenizer.pad_token_id] * (pad_length - len(tokens))

459

vllm/tokenizers/mistral.py

+8 -3

@@ -34,6 +34,7 @@ from transformers.tokenization_mistral_common import MistralCommonBackend

34

35 from vllm.entrypoints.chat_utils import ChatCompletionMessageParam

36 from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest

37 +from vllm.exceptions import VLLMValidationError

38 from vllm.logger import init_logger

39 from vllm.tokenizers.protocol import TokenizerLike

40

@@ -144,15 +145,19 @@ def _validate_apply_chat_template_args(

145

146 def validate_request_params(request: "ChatCompletionRequest"):

147 if request.chat_template is not None or request.chat_template_kwargs is not None:

147 - raise ValueError("chat_template is not supported for Mistral tokenizers.")

148 + raise VLLMValidationError(

149 + "chat_template is not supported for Mistral tokenizers.",

150 + parameter="chat_template",

151 + )

152

153 if request.reasoning_effort and request.reasoning_effort not in list(

154 ReasoningEffort

155 ):

152 - raise ValueError(

156 + raise VLLMValidationError(

157 f"reasoning_effort={request.reasoning_effort} is not supported by "

158 "Mistral models. Supported values are: "

155 - f"{[e.value for e in ReasoningEffort]}."

159 + f"{[e.value for e in ReasoningEffort]}.",

160 + parameter="reasoning_effort",

161 )

162

163

[Frontend] Migrate input-validation errors to VLLMValidationError (harmony/mistral/chat_utils/params) (#55735)