@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / fb946a7f89

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 9, 10:00 UTC

Make `mypy` opt-out instead of opt-in (#33205)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>

Harry Mellor committed Jan 29, 2026 at 09:12 UTC fb946a7f893da21808a016c86973bdeb2b3c6c95

11 files changed +35 -57

tools/pre_commit/mypy.py

+17 -40

@@ -23,39 +23,8 @@ import sys

24 import regex as re

26 -FILES = [

27 - "vllm/*.py",

28 - "vllm/assets",

29 - "vllm/compilation",

30 - "vllm/distributed",

31 - "vllm/engine",

32 - "vllm/entrypoints",

33 - "vllm/executor",

34 - "vllm/inputs",

35 - "vllm/logging_utils",

36 - "vllm/multimodal",

37 - "vllm/platforms",

38 - "vllm/plugins",

39 - "vllm/renderers",

40 - "vllm/tokenizers",

41 - "vllm/transformers_utils",

42 - "vllm/triton_utils",

43 - "vllm/usage",

44 - "vllm/utils",

45 - "vllm/worker",

46 - "vllm/v1/attention",

47 - "vllm/v1/core",

48 - "vllm/v1/engine",

49 - "vllm/v1/executor",

50 - "vllm/v1/metrics",

51 - "vllm/v1/pool",

52 - "vllm/v1/sample",

53 - "vllm/v1/structured_output",

54 - "vllm/v1/worker",

55 -]

56 -

26 # After fixing errors resulting from changing follow_imports

58 -# from "skip" to "silent", move the following directories to FILES

27 +# from "skip" to "silent", remove its directory from SEPARATE_GROUPS.

28 SEPARATE_GROUPS = [

29 "tests",

30 # v0 related

@@ -74,6 +43,16 @@ EXCLUDE = [

43 "vllm/model_executor/layers/fla/ops",

44 # Ignore triton kernels in ops.

45 "vllm/v1/attention/ops",

46 + # TODO: Remove these entries after fixing mypy errors.

47 + "vllm/benchmarks",

48 + "vllm/config",

49 + "vllm/device_allocator",

50 + "vllm/profiler",

51 + "vllm/reasoning",

52 + "vllm/tool_parser",

53 + "vllm/v1/cudagraph_dispatcher.py",

54 + "vllm/outputs.py",

55 + "vllm/logger.py",

56 ]

@@ -88,7 +67,6 @@ def group_files(changed_files: list[str]) -> dict[str, list[str]]:

67 A dictionary mapping file group names to lists of changed files.

68 """

69 exclude_pattern = re.compile(f"^{'|'.join(EXCLUDE)}.*")

91 - files_pattern = re.compile(f"^({'|'.join(FILES)}).*")

70 file_groups = {"": []}

71 file_groups.update({k: [] for k in SEPARATE_GROUPS})

72 for changed_file in changed_files:

@@ -96,14 +74,13 @@ def group_files(changed_files: list[str]) -> dict[str, list[str]]:

74 if exclude_pattern.match(changed_file):

75 continue

76 # Group files by mypy call

99 - if files_pattern.match(changed_file):

100 - file_groups[""].append(changed_file)

101 - continue

77 + for directory in SEPARATE_GROUPS:

78 + if re.match(f"^{directory}.*", changed_file):

79 + file_groups[directory].append(changed_file)

80 + break

81 else:

103 - for directory in SEPARATE_GROUPS:

104 - if re.match(f"^{directory}.*", changed_file):

105 - file_groups[directory].append(changed_file)

106 - break

82 + if changed_file.startswith("vllm/"):

83 + file_groups[""].append(changed_file)

84 return file_groups

vllm/_aiter_ops.py

+1 -1

@@ -349,7 +349,7 @@ def _rocm_aiter_mla_decode_fwd_impl(

349 ) -> None:

350 from aiter.mla import mla_decode_fwd

351

352 - kwargs = {

352 + kwargs: dict[str, float | torch.Tensor | None] = {

353 "sm_scale": sm_scale,

354 "logit_cap": logit_cap,

355 }

vllm/config/compilation.py

+2 -1

@@ -570,7 +570,7 @@ class CompilationConfig:

570 pass_config: PassConfig = field(default_factory=PassConfig)

571 """Custom inductor passes, see PassConfig for more details"""

572

573 - max_cudagraph_capture_size: int | None = field(default=None)

573 + max_cudagraph_capture_size: int = field(default=None)

574 """The maximum cudagraph capture size.

575

576 If cudagraph_capture_sizes is specified, this will be set to the largest

@@ -743,6 +743,7 @@ class CompilationConfig:

743 "level",

744 "mode",

745 "cudagraph_mode",

746 + "max_cudagraph_capture_size",

747 "use_inductor_graph_partition",

748 mode="wrap",

749 )

vllm/config/utils.py

+2 -2

@@ -9,7 +9,7 @@ import inspect

9 import json

10 import pathlib

11 import textwrap

12 -from collections.abc import Callable, Iterable, Mapping, Sequence, Set

12 +from collections.abc import Callable, Mapping, Sequence, Set

13 from dataclasses import MISSING, Field, dataclass, field, fields, is_dataclass, replace

14 from itertools import pairwise

15 from typing import TYPE_CHECKING, Any, Protocol, TypeVar

@@ -75,7 +75,7 @@ def get_field(cls: ConfigType, name: str) -> Field:

76 def getattr_iter(

77 object: object,

78 - names: Iterable[str],

78 + names: Sequence[str],

79 default: Any | None = None,

80 default_factory: Callable[[], Any] | None = None,

81 warn: bool = False,

vllm/env_override.py

+1 -1

@@ -382,7 +382,7 @@ def _patch_get_raw_stream_if_needed():

382 if hasattr(torch._C, "_cuda_getCurrentRawStream"):

383 from torch._C import _cuda_getCurrentRawStream as _get_raw_stream

384

385 - builtins.get_raw_stream = _get_raw_stream

385 + builtins.get_raw_stream = _get_raw_stream # type: ignore[attr-defined]

386

387

388 _patch_get_raw_stream_if_needed()

vllm/envs.py

@@ -1680,6 +1680,7 @@ def disable_envs_cache() -> None:

1680 global __getattr__

1681 # If __getattr__ is wrapped by functions.cache, unwrap the caching layer.

1682 if _is_envs_cache_enabled():

1683 + assert hasattr(__getattr__, "__wrapped__")

1684 __getattr__ = __getattr__.__wrapped__

1685

1686

vllm/forward_context.py

+1 -1

@@ -270,7 +270,7 @@ def create_forward_context(

270 cudagraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE,

271 batch_descriptor: BatchDescriptor | None = None,

272 ubatch_slices: UBatchSlices | None = None,

273 - slot_mapping: dict[str, torch.Tensor] | None = None,

273 + slot_mapping: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None = None,

274 additional_kwargs: dict[str, Any] | None = None,

275 skip_compiled: bool = False,

276 ):

vllm/logger.py

+1 -1

@@ -157,7 +157,7 @@ _METHODS_TO_PATCH = {

157

158

159 def _configure_vllm_root_logger() -> None:

160 - logging_config = dict[str, Any]()

160 + logging_config = dict[str, dict[str, Any] | Any]()

161

162 if not envs.VLLM_CONFIGURE_LOGGING and envs.VLLM_LOGGING_CONFIG_PATH:

163 raise RuntimeError(

vllm/logprobs.py

+3 -3

@@ -28,7 +28,7 @@ LogprobsOnePosition = dict[int, Logprob]

30 @dataclass

31 -class FlatLogprobs(MutableSequence[LogprobsOnePosition]):

31 +class FlatLogprobs(MutableSequence[LogprobsOnePosition | None]):

32 """

33 Flat logprobs of a request into multiple primitive type lists.

@@ -140,7 +140,7 @@ class FlatLogprobs(MutableSequence[LogprobsOnePosition]):

140 def __delitem__(self, item) -> None:

141 raise TypeError("Cannot delete logprobs from FlatLogprobs")

142

143 - def insert(self, item) -> None:

143 + def insert(self, index: int, value: dict[int, Logprob] | None) -> None:

144 raise TypeError("Cannot insert logprobs to FlatLogprobs")

145

146 def __iter__(self) -> Iterator[LogprobsOnePosition]:

@@ -161,7 +161,7 @@ SampleLogprobs = FlatLogprobs | list[LogprobsOnePosition]

161

162 def create_prompt_logprobs(flat_logprobs: bool) -> PromptLogprobs:

163 """Creates a container to store prompt logprobs for a request"""

164 - logprobs = FlatLogprobs() if flat_logprobs else []

164 + logprobs: PromptLogprobs = FlatLogprobs() if flat_logprobs else []

165 # NOTE: logprob of first prompt token is None.

166 logprobs.append(None)

167 return logprobs

vllm/profiler/utils.py

+5 -6

@@ -4,6 +4,7 @@

4 import dataclasses

5 from collections.abc import Callable

7 +from _typeshed import DataclassInstance

8 from torch._C._profiler import _EventType, _ProfilerEvent, _TensorMetadata

10 #

@@ -11,7 +12,7 @@ from torch._C._profiler import _EventType, _ProfilerEvent, _TensorMetadata

12 #

14 -def trim_string_front(string, width):

15 +def trim_string_front(string: str, width: int) -> str:

16 if len(string) > width:

17 offset = len(string) - width + 3

18 string = string[offset:]

@@ -20,7 +21,7 @@ def trim_string_front(string, width):

21 return string

23 -def trim_string_back(string, width):

24 +def trim_string_back(string: str, width: int) -> str:

25 if len(string) > width:

26 offset = len(string) - width + 3

27 string = string[:-offset]

@@ -30,15 +31,13 @@ def trim_string_back(string, width):

33 class TablePrinter:

33 - def __init__(

34 - self, row_cls: type[dataclasses.dataclass], column_widths: dict[str, int]

35 - ):

34 + def __init__(self, row_cls: type[DataclassInstance], column_widths: dict[str, int]):

35 self.row_cls = row_cls

36 self.fieldnames = [x.name for x in dataclasses.fields(row_cls)]

37 self.column_widths = column_widths

38 assert set(self.column_widths.keys()) == set(self.fieldnames)

41 - def print_table(self, rows: list[dataclasses.dataclass]):

40 + def print_table(self, rows: list[DataclassInstance]):

41 self._print_header()

42 self._print_line()

43 for row in rows:

vllm/v1/kv_cache_interface.py

+1 -1

@@ -98,7 +98,7 @@ class FullAttentionSpec(AttentionSpec):

98 In this case, we use FullAttentionSpec and record the sliding window size.

99 """

100

101 - head_size_v: int | None = None

101 + head_size_v: int = None # type: ignore[assignment]

102

103 sliding_window: int | None = None

104 """