@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 56e96b37e4

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 12, 19:45 UTC

[V0 Deprecation] Remove `best_of` (#29090)

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

Cyrus Leung committed Nov 21, 2025 at 11:40 UTC 56e96b37e4951946c06379b4891d8170e743dcc2

6 files changed +4 -65

docs/usage/v1_guide.md

+2 -2

@@ -169,8 +169,8 @@ As part of the major architectural rework in vLLM V1, several legacy features ha

169 - **best_of**: This feature has been deprecated due to limited usage. See details at [RFC #13361](https://github.com/vllm-project/vllm/issues/13361).

170 - **Per-Request Logits Processors**: In V0, users could pass custom

171 processing functions to adjust logits on a per-request basis. In vLLM V1, this

172 - feature has been deprecated. Instead, the design is moving toward supporting **global logits

173 - processors**, a feature the team is actively working on for future releases. See details at [RFC #13360](https://github.com/vllm-project/vllm/pull/13360).

172 + feature has been deprecated. Instead, we now support **global logits processors**

173 + which are set at startup time, see [RFC #17799](https://github.com/vllm-project/vllm/issues/17799).

174

175 ##### KV Cache features

176

tests/v1/sample/test_sampling_params_e2e.py

-8

@@ -22,14 +22,6 @@ def test_n_gt_1(llm):

22 assert len(outputs[0].outputs) == 3

25 -def test_best_of(llm):

26 - """Raise a ValueError since best_of is deprecated."""

27 -

28 - params = SamplingParams(n=2, best_of=3)

29 - with pytest.raises(ValueError):

30 - _ = llm.generate(PROMPT, params)

31 -

32 -

25 def test_penalties(llm):

26 """Check that we do not get errors if applied."""

vllm/entrypoints/openai/protocol.py

-4

@@ -565,7 +565,6 @@ class ChatCompletionRequest(OpenAIBaseModel):

565 user: str | None = None

566

567 # --8<-- [start:chat-completion-sampling-params]

568 - best_of: int | None = None

568 use_beam_search: bool = False

569 top_k: int | None = None

570 min_p: float | None = None

@@ -889,7 +888,6 @@ class ChatCompletionRequest(OpenAIBaseModel):

888 extra_args["kv_transfer_params"] = self.kv_transfer_params

889 return SamplingParams.from_optional(

890 n=self.n,

892 - best_of=self.best_of,

891 presence_penalty=self.presence_penalty,

892 frequency_penalty=self.frequency_penalty,

893 repetition_penalty=repetition_penalty,

@@ -1088,7 +1086,6 @@ class CompletionRequest(OpenAIBaseModel):

1086 # https://platform.openai.com/docs/api-reference/completions/create

1087 model: str | None = None

1088 prompt: list[int] | list[list[int]] | str | list[str] | None = None

1091 - best_of: int | None = None

1089 echo: bool | None = False

1090 frequency_penalty: float | None = 0.0

1091 logit_bias: dict[str, float] | None = None

@@ -1375,7 +1372,6 @@ class CompletionRequest(OpenAIBaseModel):

1372 extra_args["kv_transfer_params"] = self.kv_transfer_params

1373 return SamplingParams.from_optional(

1374 n=self.n,

1378 - best_of=self.best_of,

1375 presence_penalty=self.presence_penalty,

1376 frequency_penalty=self.frequency_penalty,

1377 repetition_penalty=repetition_penalty,

vllm/entrypoints/openai/serving_completion.py

+2 -8

@@ -250,14 +250,8 @@ class OpenAIServingCompletion(OpenAIServing):

250 model_name = self.models.model_name(lora_request)

251 num_prompts = len(engine_prompts)

252

253 - # Similar to the OpenAI API, when n != best_of, we do not stream the

254 - # results. Noting that best_of is only supported in V0. In addition,

255 - # we do not stream the results when use beam search.

256 - stream = (

257 - request.stream

258 - and (request.best_of is None or request.n == request.best_of)

259 - and not request.use_beam_search

260 - )

253 + # We do not stream the results when using beam search.

254 + stream = request.stream and not request.use_beam_search

255

256 # Streaming response

257 if stream:

vllm/sampling_params.py

-40

@@ -144,12 +144,6 @@ class SamplingParams(

144 are generated and streamed cumulatively per request. To see all `n`

145 outputs upon completion, use `output_kind=RequestOutputKind.FINAL_ONLY`

146 in `SamplingParams`."""

147 - best_of: int | None = None

148 - """Number of output sequences that are generated from the prompt. From

149 - these `best_of` sequences, the top `n` sequences are returned. `best_of`

150 - must be greater than or equal to `n`. By default, `best_of` is set to `n`.

151 - Warning, this is only supported in V0."""

152 - _real_n: int | None = None

147 presence_penalty: float = 0.0

148 """Penalizes new tokens based on whether they appear in the generated text

149 so far. Values > 0 encourage the model to use new tokens, while values < 0

@@ -265,7 +259,6 @@ class SamplingParams(

259 @staticmethod

260 def from_optional(

261 n: int | None = 1,

268 - best_of: int | None = None,

262 presence_penalty: float | None = 0.0,

263 frequency_penalty: float | None = 0.0,

264 repetition_penalty: float | None = 1.0,

@@ -315,7 +308,6 @@ class SamplingParams(

308

309 return SamplingParams(

310 n=1 if n is None else n,

318 - best_of=best_of,

311 presence_penalty=0.0 if presence_penalty is None else presence_penalty,

312 frequency_penalty=0.0 if frequency_penalty is None else frequency_penalty,

313 repetition_penalty=1.0

@@ -348,22 +340,6 @@ class SamplingParams(

340 )

341

342 def __post_init__(self) -> None:

351 - # how we deal with `best_of`:

352 - # if `best_of` is not set, we default to `n`;

353 - # if `best_of` is set, we set `n` to `best_of`,

354 - # and set `_real_n` to the original `n`.

355 - # when we return the result, we will check

356 - # if we need to return `n` or `_real_n` results

357 - if self.best_of:

358 - if self.best_of < self.n:

359 - raise ValueError(

360 - f"best_of must be greater than or equal to n, "

361 - f"got n={self.n} and best_of={self.best_of}."

362 - )

363 - if not self._real_n:

364 - self._real_n = self.n

365 - self.n = self.best_of

366 -

343 if 0 < self.temperature < _MAX_TEMP:

344 logger.warning(

345 "temperature %s is less than %s, which may cause numerical "

@@ -433,18 +409,6 @@ class SamplingParams(

409 raise ValueError(f"n must be an int, but is of type {type(self.n)}")

410 if self.n < 1:

411 raise ValueError(f"n must be at least 1, got {self.n}.")

436 - if self.best_of is not None:

437 - if not isinstance(self.best_of, int):

438 - raise ValueError(

439 - f"best_of must be an integer, got {type(self.best_of)}"

440 - )

441 - if self.best_of < 1:

442 - raise ValueError(f"best_of must be at least 1, got {self.best_of}")

443 - if self.best_of < self.n:

444 - raise ValueError(

445 - f"best_of must be greater than or equal to n, "

446 - f"got n={self.n} and best_of={self.best_of}."

447 - )

412 if not -2.0 <= self.presence_penalty <= 2.0:

413 raise ValueError(

414 f"presence_penalty must be in [-2, 2], got {self.presence_penalty}."

@@ -519,10 +483,6 @@ class SamplingParams(

483 "stop strings are only supported when detokenize is True. "

484 "Set detokenize=True to use stop."

485 )

522 - if self.best_of != self._real_n and self.output_kind == (

523 - RequestOutputKind.DELTA

524 - ):

525 - raise ValueError("best_of must equal n to use output_kind=DELTA")

486

487 def _verify_greedy_sampling(self) -> None:

488 if self.n > 1:

vllm/v1/engine/processor.py

-3

@@ -142,9 +142,6 @@ class Processor:

142 self,

143 params: SamplingParams,

144 ) -> None:

145 - # Best of not yet supported.

146 - if params.best_of is not None and params.best_of > 1:

147 - raise ValueError("vLLM V1 does not yet support best_of.")

145 # Logits processors not supported.

146 if params.logits_processors:

147 raise ValueError(