@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 3b312fb792

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 10, 20:15 UTC

[Minor] Various small code cleanups/simplifications (#31508)

Signed-off-by: njhill <nickhill123@gmail.com>

Nick Hill committed Dec 29, 2025 at 22:42 UTC 3b312fb792d200ba4b1c3b4bf5ad06a46cd2fa1e

14 files changed +29 -55

vllm/config/model.py

+4 -4

@@ -1579,14 +1579,14 @@ class ModelConfig:

1579

1580 @property

1581 def is_hybrid(self) -> bool:

1582 + if not self._model_info.is_hybrid:

1583 + return False

1584 # Handle granite-4.0-micro case which uses hybrid config but does not

1585 # actually contain any non-attention layers.

1586 layer_types = getattr(self.hf_config, "layer_types", None)

1585 - if layer_types is not None and all(

1587 + return layer_types is None or not all(

1588 layer == "attention" for layer in layer_types

1587 - ):

1588 - return False

1589 - return self._model_info.is_hybrid

1589 + )

1590

1591 @property

1592 def has_noops(self) -> bool:

vllm/entrypoints/openai/serving_responses.py

-1

@@ -2005,7 +2005,6 @@ class OpenAIServingResponses(OpenAIServing):

2005 return event

2006

2007 async with AsyncExitStack() as exit_stack:

2008 - processer = None

2008 if self.use_harmony:

2009 # TODO: in streaming, we noticed this bug:

2010 # https://github.com/vllm-project/vllm/issues/25697

vllm/entrypoints/renderer.py

+2 -5

@@ -44,11 +44,8 @@ class RenderConfig:

44 def verify_truncate_prompt_tokens(self, model_config: ModelConfig) -> int | None:

45 """Validate and normalize `truncate_prompt_tokens` parameter."""

46 truncate_prompt_tokens = self.truncate_prompt_tokens

47 - if truncate_prompt_tokens is None:

48 - return None

49 -

50 - if truncate_prompt_tokens == 0:

51 - return 0

47 + if truncate_prompt_tokens is None or truncate_prompt_tokens == 0:

48 + return truncate_prompt_tokens

50 if truncate_prompt_tokens < 0:

51 truncate_prompt_tokens = model_config.max_model_len

vllm/inputs/preprocess.py

+1 -5

@@ -686,11 +686,7 @@ class InputPreprocessor:

686 mm_uuids: MultiModalUUIDDict | None = None,

687 ) -> ProcessorInputs:

688 """Preprocess the input prompt."""

689 - res = self._preprocess(

690 - prompt,

691 - tokenization_kwargs,

692 - mm_uuids=mm_uuids,

693 - )

689 + res = self._preprocess(prompt, tokenization_kwargs, mm_uuids=mm_uuids)

690

691 if self.mm_processor_cache and self.mm_cache_stats is not None:

692 delta = self.mm_processor_cache.make_stats(delta=True)

vllm/multimodal/inputs.py

+2 -11

@@ -171,10 +171,7 @@ class PlaceholderRange:

171

172 @cached_property

173 def embeds_cumsum(self) -> torch.Tensor | None:

174 - if self.is_embed is None:

175 - return None

176 -

177 - return self.is_embed.cumsum(dim=0)

174 + return None if self.is_embed is None else self.is_embed.cumsum(dim=0)

175

176 @cached_property

177 def get_num_embeds(self) -> int:

@@ -308,13 +305,7 @@ def batched_tensors_equal(a: BatchedTensorInputs, b: BatchedTensorInputs) -> boo

305 Equality check between

306 [`BatchedTensorInputs`][vllm.multimodal.inputs.BatchedTensorInputs] objects.

307 """

311 - for k in a:

312 - if k not in b:

313 - return False

314 - if not nested_tensors_equal(a[k], b[k]):

315 - return False

316 -

317 - return True

308 + return all(k in b and nested_tensors_equal(a[k], b[k]) for k in a)

309

310

311 @dataclass

vllm/v1/engine/core.py

+2 -4

@@ -487,10 +487,8 @@ class EngineCore:

487 request_ids = []

488 while not self.aborts_queue.empty():

489 ids = self.aborts_queue.get_nowait()

490 - if isinstance(ids, str):

491 - # Should be a list here, but also handle string just in case.

492 - ids = (ids,)

493 - request_ids.extend(ids)

490 + # Should be a list here, but also handle string just in case.

491 + request_ids.extend((ids,) if isinstance(ids, str) else ids)

492 # More efficient to abort all as a single batch.

493 self.abort_requests(request_ids)

494

vllm/v1/engine/input_processor.py

+1 -1

@@ -618,7 +618,7 @@ class InputProcessor:

618

619 tokenizer = self.tokenizer

620 if tokenizer is not None:

621 - max_input_id = max(prompt_ids or [], default=0)

621 + max_input_id = max(prompt_ids or (), default=0)

622

623 # NOTE: tokenizer.max_token_id is the tokenizer’s vocab size while

624 # self.model_config.get_vocab_size() is the model’s vocab size.

vllm/v1/engine/output_processor.py

+2 -7

@@ -339,10 +339,7 @@ class RequestState:

339 stop_reason=stop_reason if finished else None,

340 )

341

342 - def _new_pooling_output(

343 - self,

344 - pooling_output: torch.Tensor,

345 - ) -> PoolingOutput:

342 + def _new_pooling_output(self, pooling_output: torch.Tensor) -> PoolingOutput:

343 return PoolingOutput(data=pooling_output)

344

345

@@ -695,9 +692,7 @@ class OutputProcessor:

692 assert req_state.stats is not None

693 iteration_stats.update_from_finished_request(

694 finish_reason=finish_reason,

698 - num_prompt_tokens=length_from_prompt_token_ids_or_embeds(

699 - req_state.prompt_token_ids, req_state.prompt_embeds

700 - ),

695 + num_prompt_tokens=req_state.prompt_len,

696 max_tokens_param=req_state.max_tokens_param,

697 req_stats=req_state.stats,

698 num_cached_tokens=req_state.num_cached_tokens,

vllm/v1/executor/multiproc_executor.py

+1 -1

@@ -695,7 +695,7 @@ class WorkerProc:

695 worker = None

696 # tuple[Connection, Connection]

697 reader, ready_writer = kwargs.pop("ready_pipe")

698 - death_pipe = kwargs.pop("death_pipe", None)

698 + death_pipe: Connection | None = kwargs.pop("death_pipe", None)

699 shutdown_event = threading.Event()

700 # Start death monitoring thread if death_pipe is provided

701 if death_pipe is not None:

vllm/v1/request.py

+1 -2

@@ -211,8 +211,7 @@ class Request:

211

212 def get_num_encoder_embeds(self, input_id: int) -> int:

213 assert input_id < len(self.mm_features)

214 - num_embeds = self.mm_features[input_id].mm_position.get_num_embeds

215 - return num_embeds

214 + return self.mm_features[input_id].mm_position.get_num_embeds

215

216 def record_event(

217 self,

vllm/v1/structured_output/__init__.py

+8 -6

@@ -1,6 +1,8 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 +import itertools

4 import multiprocessing

5 +from collections.abc import Iterable

6 from concurrent.futures import Future, ThreadPoolExecutor

7 from typing import TYPE_CHECKING

@@ -172,7 +174,7 @@ class StructuredOutputManager:

174

175 def _fill_bitmasks(

176 self,

175 - batch: list[tuple[StructuredOutputGrammar, int, bool]],

177 + batch: Iterable[tuple[StructuredOutputGrammar, int, bool]],

178 ) -> None:

179 assert self._grammar_bitmask is not None

180 for grammar, index, apply_bitmask in batch:

@@ -265,16 +267,16 @@ class StructuredOutputManager:

267 apply_bitmask = self.should_fill_bitmask(request)

268

269 state_advancements = 0

268 - req_tokens = scheduled_spec_decode_tokens.get(req_id, [])

269 - for i, token in enumerate(req_tokens + [None]):

270 + req_tokens = scheduled_spec_decode_tokens.get(req_id, ())

271 + for token in itertools.chain(req_tokens, (None,)):

272 self._fill_bitmasks(

271 - [

273 + (

274 (

275 structured_output_request.grammar,

276 cumulative_index,

277 apply_bitmask,

276 - )

277 - ]

278 + ),

279 + )

280 )

281

282 if (

vllm/v1/structured_output/request.py

+3 -6

@@ -28,12 +28,9 @@ class StructuredOutputRequest:

28 if sampling_params is None:

29 return None

30 params = sampling_params.structured_outputs

31 - if params:

32 - if params.all_constraints_none():

33 - return None

34 - else:

35 - return StructuredOutputRequest(params=params)

36 - return None

31 + if not params or params.all_constraints_none():

32 + return None

33 + return StructuredOutputRequest(params=params)

35 def _check_grammar_completion(self) -> bool:

36 # NOTE: We have to lazy import to gate circular imports

vllm/v1/worker/gpu_input_batch.py

+1 -1

@@ -829,7 +829,7 @@ class InputBatch:

829 presence_penalties=self.presence_penalties[:num_reqs],

830 repetition_penalties=self.repetition_penalties[:num_reqs],

831 output_token_ids=output_token_ids,

832 - spec_token_ids=cast(list[list[int]], self.spec_token_ids),

832 + spec_token_ids=self.spec_token_ids,

833 no_penalties=self.no_penalties,

834 allowed_token_ids_mask=allowed_token_ids_mask,

835 bad_words_token_ids=self.bad_words_token_ids,

vllm/v1/worker/gpu_model_runner.py

+1 -1

@@ -1026,7 +1026,7 @@ class GPUModelRunner(

1026 each sequence, and a shifting is done during the next iteration

1027 based on the number of accepted tokens.

1028 """

1029 - if not self.model_config.is_hybrid or not self.speculative_config:

1029 + if not self.speculative_config or not self.model_config.is_hybrid:

1030 return

1031

1032 # Find the number of accepted tokens for each sequence.