@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / fb455ed547

Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 8, 10:00 UTC

[V0 Deprecation] Remove code related to per-request logits processors (#34400)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

Cyrus Leung committed Feb 12, 2026 at 20:44 UTC fb455ed547a63e97e15deccfc493f8eef7a2da5c

12 files changed +15 -122

tests/entrypoints/openai/test_chat_error.py

-1

@@ -45,7 +45,6 @@ class MockModelConfig:

45 multimodal_config = MultiModalConfig()

46 hf_config = MockHFConfig()

47 hf_text_config = MockHFConfig()

48 - logits_processor_pattern = None

48 logits_processors: list[str] | None = None

49 diff_sampling_param: dict | None = None

50 allowed_local_media_path: str = ""

tests/entrypoints/openai/test_completion_error.py

-1

@@ -44,7 +44,6 @@ class MockModelConfig:

44 tokenizer_revision = None

45 multimodal_config = MultiModalConfig()

46 hf_config = MockHFConfig()

47 - logits_processor_pattern = None

47 logits_processors: list[str] | None = None

48 diff_sampling_param: dict | None = None

49 allowed_local_media_path: str = ""

tests/entrypoints/openai/test_lora_resolvers.py

-1

@@ -45,7 +45,6 @@ class MockModelConfig:

45 multimodal_config: MultiModalConfig = field(default_factory=MultiModalConfig)

46 hf_config: MockHFConfig = field(default_factory=MockHFConfig)

47 logits_processors: list[str] | None = None

48 - logits_processor_pattern: str | None = None

48 diff_sampling_param: dict | None = None

49 allowed_local_media_path: str = ""

50 allowed_media_domains: list[str] | None = None

tests/entrypoints/openai/test_serving_chat.py

-1

@@ -521,7 +521,6 @@ class MockModelConfig:

521 hf_config = MockHFConfig()

522 hf_text_config = MockHFConfig()

523 logits_processors: list[str] | None = None

524 - logits_processor_pattern = None

524 diff_sampling_param: dict | None = None

525 allowed_local_media_path: str = ""

526 allowed_media_domains: list[str] | None = None

tests/v1/sample/test_sampling_params_e2e.py

-14

@@ -144,20 +144,6 @@ def test_bad_words(llm):

144 assert not contains_bad_word(new_text, new_tokens, bad_words_2)

145

146

147 -def test_logits_processor(llm):

148 - """Check that we reject logits processor."""

149 -

150 - # This sample logits processor gives infinite score to the i-th token,

151 - # where i is the length of the input sequence.

152 - # We therefore expect the output token sequence to be [0, 1, 2, ...]

153 - def pick_ith(token_ids, logits):

154 - logits[len(token_ids)] = float("inf")

155 - return logits

156 -

157 - with pytest.raises(ValueError):

158 - _ = llm.generate(PROMPT, SamplingParams(logits_processors=[pick_ith]))

159 -

160 -

147 def test_allowed_token_ids(llm):

148 """Check that we can use allowed_token_ids."""

149

vllm/config/model.py

-5

@@ -252,10 +252,6 @@ class ModelConfig:

252 hf_overrides: HfOverrides = field(default_factory=dict)

253 """If a dictionary, contains arguments to be forwarded to the Hugging Face

254 config. If a callable, it is called to update the HuggingFace config."""

255 - logits_processor_pattern: str | None = None

256 - """Optional regex pattern specifying valid logits processor qualified names

257 - that can be passed with the `logits_processors` extra completion argument.

258 - Defaults to `None`, which allows no processors."""

255 generation_config: str = "auto"

256 """The folder path to the generation config. Defaults to `"auto"`, the

257 generation config will be loaded from model path. If set to `"vllm"`, no

@@ -342,7 +338,6 @@ class ModelConfig:

338 "config_format",

339 "hf_token",

340 "hf_overrides",

345 - "logits_processor_pattern",

341 "override_attention_dtype",

342 "logits_processors",

343 "io_processor_plugin",

vllm/engine/arg_utils.py

+2 -11

@@ -508,8 +508,6 @@ class EngineArgs:

508 reasoning_parser: str = StructuredOutputsConfig.reasoning_parser

509 reasoning_parser_plugin: str | None = None

510

511 - logits_processor_pattern: str | None = ModelConfig.logits_processor_pattern

512 -

511 speculative_config: dict[str, Any] | None = None

512

513 show_hidden_metrics_for_version: str | None = (

@@ -710,9 +708,6 @@ class EngineArgs:

708 )

709 model_group.add_argument("--hf-overrides", **model_kwargs["hf_overrides"])

710 model_group.add_argument("--pooler-config", **model_kwargs["pooler_config"])

713 - model_group.add_argument(

714 - "--logits-processor-pattern", **model_kwargs["logits_processor_pattern"]

715 - )

711 model_group.add_argument(

712 "--generation-config", **model_kwargs["generation_config"]

713 )

@@ -1320,7 +1315,6 @@ class EngineArgs:

1315 mm_encoder_tp_mode=self.mm_encoder_tp_mode,

1316 mm_encoder_attn_backend=self.mm_encoder_attn_backend,

1317 pooler_config=self.pooler_config,

1323 - logits_processor_pattern=self.logits_processor_pattern,

1318 generation_config=self.generation_config,

1319 override_generation_config=self.override_generation_config,

1320 enable_sleep_mode=self.enable_sleep_mode,

@@ -1429,7 +1423,7 @@ class EngineArgs:

1423 self.model_weights = model_config.model_weights

1424 self.tokenizer = model_config.tokenizer

1425

1432 - self._check_feature_supported(model_config)

1426 + self._check_feature_supported()

1427 self._set_default_chunked_prefill_and_prefix_caching_args(model_config)

1428 self._set_default_max_num_seqs_and_batched_tokens_args(

1429 usage_context, model_config

@@ -1831,11 +1825,8 @@ class EngineArgs:

1825

1826 return config

1827

1834 - def _check_feature_supported(self, model_config: ModelConfig):

1828 + def _check_feature_supported(self):

1829 """Raise an error if the feature is not supported."""

1836 - if self.logits_processor_pattern != EngineArgs.logits_processor_pattern:

1837 - _raise_unsupported_error(feature_name="--logits-processor-pattern")

1838 -

1830 # No Concurrent Partial Prefills so far.

1831 if (

1832 self.max_num_partial_prefills != SchedulerConfig.max_num_partial_prefills

vllm/entrypoints/openai/chat_completion/protocol.py

+3 -19

@@ -26,13 +26,11 @@ from vllm.entrypoints.openai.engine.protocol import (

26 FunctionCall,

27 FunctionDefinition,

28 LegacyStructuralTagResponseFormat,

29 - LogitsProcessors,

29 OpenAIBaseModel,

30 StreamOptions,

31 StructuralTagResponseFormat,

32 ToolCall,

33 UsageInfo,

35 - get_logits_processors,

34 )

35 from vllm.exceptions import VLLMValidationError

36 from vllm.logger import init_logger

@@ -293,19 +291,7 @@ class ChatCompletionRequest(OpenAIBaseModel):

291 "through out the inference process and return in response."

292 ),

293 )

296 - logits_processors: LogitsProcessors | None = Field(

297 - default=None,

298 - description=(

299 - "A list of either qualified names of logits processors, or "

300 - "constructor objects, to apply when sampling. A constructor is "

301 - "a JSON object with a required 'qualname' field specifying the "

302 - "qualified name of the processor class/factory, and optional "

303 - "'args' and 'kwargs' fields containing positional and keyword "

304 - "arguments. For example: {'qualname': "

305 - "'my_module.MyLogitsProcessor', 'args': [1, 2], 'kwargs': "

306 - "{'param': 'value'}}."

307 - ),

308 - )

294 +

295 return_tokens_as_token_ids: bool | None = Field(

296 default=None,

297 description=(

@@ -324,6 +310,7 @@ class ChatCompletionRequest(OpenAIBaseModel):

310 "need to map generated text back to input tokens."

311 ),

312 )

313 +

314 cache_salt: str | None = Field(

315 default=None,

316 description=(

@@ -335,6 +322,7 @@ class ChatCompletionRequest(OpenAIBaseModel):

322 "to 256 bit)."

323 ),

324 )

325 +

326 kv_transfer_params: dict[str, Any] | None = Field(

327 default=None,

328 description="KVTransfer parameters used for disaggregated serving.",

@@ -417,7 +405,6 @@ class ChatCompletionRequest(OpenAIBaseModel):

405 def to_sampling_params(

406 self,

407 max_tokens: int,

420 - logits_processor_pattern: str | None,

408 default_sampling_params: dict,

409 ) -> SamplingParams:

410 # Default parameters

@@ -502,9 +489,6 @@ class ChatCompletionRequest(OpenAIBaseModel):

489 min_tokens=self.min_tokens,

490 skip_special_tokens=self.skip_special_tokens,

491 spaces_between_special_tokens=self.spaces_between_special_tokens,

505 - logits_processors=get_logits_processors(

506 - self.logits_processors, logits_processor_pattern

507 - ),

492 include_stop_str_in_output=self.include_stop_str_in_output,

493 truncate_prompt_tokens=self.truncate_prompt_tokens,

494 output_kind=RequestOutputKind.DELTA

vllm/entrypoints/openai/chat_completion/serving.py

-9

@@ -86,7 +86,6 @@ from vllm.tool_parsers import ToolParser

86 from vllm.tool_parsers.mistral_tool_parser import MistralToolCall

87 from vllm.tool_parsers.utils import partial_json_loads

88 from vllm.utils.collection_utils import as_list

89 -from vllm.v1.sample.logits_processor import validate_logits_processors_parameters

90 logger = init_logger(__name__)

@@ -130,9 +129,6 @@ class OpenAIServingChat(OpenAIServing):

129 self.enable_log_outputs = enable_log_outputs

130 self.enable_log_deltas = enable_log_deltas

131

133 - # set up logits processors

134 - self.logits_processors = self.model_config.logits_processors

135 -

132 # set up reasoning parser

133 self.reasoning_parser_cls = ParserManager.get_reasoning_parser(

134 reasoning_parser_name=reasoning_parser

@@ -403,13 +399,8 @@ class OpenAIServingChat(OpenAIServing):

399 else:

400 sampling_params = request.to_sampling_params(

401 max_tokens,

406 - self.model_config.logits_processor_pattern,

402 self.default_sampling_params,

403 )

409 - validate_logits_processors_parameters(

410 - self.logits_processors,

411 - sampling_params,

412 - )

404

405 self._log_inputs(

406 sub_request_id,

vllm/entrypoints/openai/completion/protocol.py

-19

@@ -15,12 +15,10 @@ from vllm.config import ModelConfig

15 from vllm.entrypoints.openai.engine.protocol import (

16 AnyResponseFormat,

17 LegacyStructuralTagResponseFormat,

18 - LogitsProcessors,

18 OpenAIBaseModel,

19 StreamOptions,

20 StructuralTagResponseFormat,

21 UsageInfo,

23 - get_logits_processors,

22 )

23 from vllm.exceptions import VLLMValidationError

24 from vllm.logger import init_logger

@@ -117,19 +115,6 @@ class CompletionRequest(OpenAIBaseModel):

115 "through out the inference process and return in response."

116 ),

117 )

120 - logits_processors: LogitsProcessors | None = Field(

121 - default=None,

122 - description=(

123 - "A list of either qualified names of logits processors, or "

124 - "constructor objects, to apply when sampling. A constructor is "

125 - "a JSON object with a required 'qualname' field specifying the "

126 - "qualified name of the processor class/factory, and optional "

127 - "'args' and 'kwargs' fields containing positional and keyword "

128 - "arguments. For example: {'qualname': "

129 - "'my_module.MyLogitsProcessor', 'args': [1, 2], 'kwargs': "

130 - "{'param': 'value'}}."

131 - ),

132 - )

118

119 return_tokens_as_token_ids: bool | None = Field(

120 default=None,

@@ -221,7 +206,6 @@ class CompletionRequest(OpenAIBaseModel):

206 def to_sampling_params(

207 self,

208 max_tokens: int,

224 - logits_processor_pattern: str | None,

209 default_sampling_params: dict | None = None,

210 ) -> SamplingParams:

211 if default_sampling_params is None:

@@ -312,9 +296,6 @@ class CompletionRequest(OpenAIBaseModel):

296 skip_special_tokens=self.skip_special_tokens,

297 spaces_between_special_tokens=self.spaces_between_special_tokens,

298 include_stop_str_in_output=self.include_stop_str_in_output,

315 - logits_processors=get_logits_processors(

316 - self.logits_processors, logits_processor_pattern

317 - ),

299 truncate_prompt_tokens=self.truncate_prompt_tokens,

300 output_kind=RequestOutputKind.DELTA

301 if self.stream

vllm/entrypoints/openai/completion/serving.py

-9

@@ -42,7 +42,6 @@ from vllm.sampling_params import BeamSearchParams, SamplingParams

42 from vllm.tokenizers import TokenizerLike

43 from vllm.utils.async_utils import merge_async_iterators

44 from vllm.utils.collection_utils import as_list

45 -from vllm.v1.sample.logits_processor import validate_logits_processors_parameters

46 logger = init_logger(__name__)

@@ -67,9 +66,6 @@ class OpenAIServingCompletion(OpenAIServing):

66 log_error_stack=log_error_stack,

67 )

70 - # set up logits processors

71 - self.logits_processors = self.model_config.logits_processors

72 -

69 self.enable_prompt_tokens_details = enable_prompt_tokens_details

70 self.enable_force_include_usage = enable_force_include_usage

@@ -178,13 +174,8 @@ class OpenAIServingCompletion(OpenAIServing):

174 else:

175 sampling_params = request.to_sampling_params(

176 max_tokens,

181 - self.model_config.logits_processor_pattern,

177 self.default_sampling_params,

178 )

184 - validate_logits_processors_parameters(

185 - self.logits_processors,

186 - sampling_params,

187 - )

179

180 request_id_item = f"{request_id}-{i}"

181

vllm/sampling_params.py

+10 -32

@@ -15,7 +15,6 @@ from pydantic.dataclasses import dataclass

15 from vllm.config import ModelConfig, SpeculativeConfig, StructuredOutputsConfig

16 from vllm.exceptions import VLLMValidationError

17 from vllm.logger import init_logger

18 -from vllm.logits_process import LogitsProcessor

18 from vllm.tokenizers import TokenizerLike

19 from vllm.v1.serial_utils import PydanticMsgspecMixin

@@ -207,11 +206,6 @@ class SamplingParams(

206 """Whether to skip special tokens in the output."""

207 spaces_between_special_tokens: bool = True

208 """Whether to add spaces between special tokens in the output."""

210 - # `list[LogitsProcessor] | None` type. We use Any here because

211 - # `list[LogitsProcessor] | None` type is not supported by msgspec.

212 - logits_processors: Any | None = None

213 - """Functions that modify logits based on previously generated tokens, and

214 - optionally prompt tokens as a first argument."""

209 include_stop_str_in_output: bool = False

210 """Whether to include the stop strings in output text."""

211 truncate_prompt_tokens: Annotated[int, msgspec.Meta(ge=-1)] | None = None

@@ -277,7 +271,6 @@ class SamplingParams(

271 detokenize: bool = True,

272 skip_special_tokens: bool = True,

273 spaces_between_special_tokens: bool = True,

280 - logits_processors: list[LogitsProcessor] | None = None,

274 truncate_prompt_tokens: Annotated[int, msgspec.Meta(ge=-1)] | None = None,

275 output_kind: RequestOutputKind = RequestOutputKind.CUMULATIVE,

276 structured_outputs: StructuredOutputsParams | None = None,

@@ -318,7 +311,6 @@ class SamplingParams(

311 detokenize=detokenize,

312 skip_special_tokens=skip_special_tokens,

313 spaces_between_special_tokens=spaces_between_special_tokens,

321 - logits_processors=logits_processors,

314 truncate_prompt_tokens=truncate_prompt_tokens,

315 output_kind=output_kind,

316 structured_outputs=structured_outputs,

@@ -455,11 +447,6 @@ class SamplingParams(

447 parameter="prompt_logprobs",

448 value=self.prompt_logprobs,

449 )

458 - if self.logits_processors:

459 - # TODO: Remove `logits_processors` attribute

460 - raise ValueError(

461 - "vLLM V1 does not support per request user-provided logits processors."

462 - )

450 if self.truncate_prompt_tokens is not None and (

451 self.truncate_prompt_tokens == 0 or self.truncate_prompt_tokens < -1

452 ):

@@ -573,28 +560,11 @@ class SamplingParams(

560 return self._bad_words_token_ids

561

562 def clone(self) -> "SamplingParams":

576 - """Deep copy, but maybe not the LogitsProcessor objects.

577 -

578 - LogitsProcessor objects may contain an arbitrary, nontrivial amount of

579 - data that is expensive to copy. However, if not copied, the processor

580 - needs to support parallel decoding for multiple sequences

581 - See https://github.com/vllm-project/vllm/issues/3087

582 -

583 - If skip_clone is True, uses shallow copy instead of deep copy.

584 - """

585 -

563 + """If skip_clone is True, uses shallow copy instead of deep copy."""

564 if self.skip_clone:

565 return copy.copy(self)

566

589 - logit_processor_refs = (

590 - None

591 - if self.logits_processors is None

592 - else {

593 - id(lp): lp.clone() if hasattr(lp, "clone") else lp

594 - for lp in self.logits_processors

595 - }

596 - )

597 - return copy.deepcopy(self, memo=logit_processor_refs)

567 + return copy.deepcopy(self)

568

569 def verify(

570 self,

@@ -605,6 +575,7 @@ class SamplingParams(

575 ) -> None:

576 self._validate_logprobs(model_config)

577 self._validate_logit_bias(model_config)

578 + self._validate_logits_processors(model_config)

579 self._validate_allowed_token_ids(tokenizer)

580 self._validate_spec_decode(speculative_config)

581 self._validate_structured_outputs(structured_outputs_config, tokenizer)

@@ -658,6 +629,13 @@ class SamplingParams(

629 value=invalid_token_ids,

630 )

631

632 + def _validate_logits_processors(self, model_config: ModelConfig) -> None:

633 + from vllm.v1.sample.logits_processor import (

634 + validate_logits_processors_parameters,

635 + )

636 +

637 + validate_logits_processors_parameters(model_config.logits_processors, self)

638 +

639 def _validate_allowed_token_ids(self, tokenizer: TokenizerLike | None) -> None:

640 allowed_token_ids = self.allowed_token_ids

641 if allowed_token_ids is None: