38a267cdd5 · tduncan/Nvidia-Optimization-Stack-VLLM

tests/compile/correctness_e2e/test_sequence_parallel.py

+6 -2

@@ -348,10 +348,14 @@ def test_tp_sp_generation(

348 if use_inductor_graph_partition and not is_torch_equal_or_newer("2.9.0.dev"):

349 pytest.skip("inductor graph partition is only available in PyTorch 2.9+")

350

351 - # Skip FP8 SP-only test on sm89 (compute capability 8.9)

351 + # Skip FP8 SP-only test on sm89 (compute capability 8.9).

352 + # An unknown capability (None) is left to fail in the test body rather than

353 + # being silently skipped here.

354 + capability = current_platform.get_device_capability()

355 if (

356 "fp8" in model_id.lower()

354 - and current_platform.get_device_capability() < (9, 0)

357 + and capability is not None

358 + and capability < (9, 0)

359 and (not fuse_gemm_comms)

360 ):

361 pytest.skip("FP8 reduction support begins with sm90 capable devices.")

tests/compile/fullgraph/test_full_cudagraph.py

+4 -2

@@ -9,7 +9,7 @@ import pytest

9 from tests.utils import wait_for_gpu_memory_to_clear

10 from tests.v1.attention.utils import full_cg_backend_configs as backend_configs

11 from vllm import LLM, SamplingParams

12 -from vllm.config import CompilationConfig

12 +from vllm.config import CompilationConfig, CUDAGraphMode

13 from vllm.platforms import current_platform

14 from vllm.utils.torch_utils import is_torch_equal_or_newer

15 from vllm.v1.attention.backends.registry import AttentionBackendEnum

@@ -97,7 +97,9 @@ def llm_pair(request):

97 trust_remote_code=True,

98 max_model_len=1024,

99 max_num_seqs=128,

100 - compilation_config=CompilationConfig(cudagraph_mode="PIECEWISE"),

100 + compilation_config=CompilationConfig(

101 + cudagraph_mode=CUDAGraphMode.PIECEWISE

102 + ),

103 generation_config="vllm",

104 seed=42,

105 )

tests/compile/fullgraph/test_full_graph.py

+5 -3

@@ -75,7 +75,7 @@ def test_full_graph(

75 monkeypatch: pytest.MonkeyPatch,

76 model: str,

77 model_kwargs: dict[str, Any],

78 - compilation_mode: int,

78 + compilation_mode: CompilationMode,

79 ):

80 if (

81 "w8a8" in model

@@ -197,7 +197,7 @@ def test_custom_compile_config(

197 ],

198 )

199 def test_fp8_kv_scale_compile(

200 - compilation_mode: int,

200 + compilation_mode: CompilationMode,

201 model: str,

202 backend: AttentionBackendEnum | None,

203 ):

@@ -213,7 +213,9 @@ def test_fp8_kv_scale_compile(

213 run_model(compilation_mode, model, **model_kwargs)

214

215

216 -def run_model(compile_config: int | CompilationConfig, model: str, **model_kwargs):

216 +def run_model(

217 + compile_config: CompilationMode | CompilationConfig, model: str, **model_kwargs

218 +):

219 compilation_config = (

220 compile_config

221 if isinstance(compile_config, CompilationConfig)

tests/compile/fusions_e2e/conftest.py

+4 -2

@@ -12,7 +12,9 @@ from vllm.config import CompilationConfig, CompilationMode, CUDAGraphMode

12 from .common import FUSION_LOG_PATTERNS, AttentionBackendCase, Matches

13

14

15 -def run_model(compile_config: int | CompilationConfig, model: str, **model_kwargs):

15 +def run_model(

16 + compile_config: CompilationMode | CompilationConfig, model: str, **model_kwargs

17 +):

18 """Run a model with the given compilation config for E2E fusion tests."""

19 compilation_config = (

20 compile_config

@@ -54,7 +56,7 @@ def run_model(compile_config: int | CompilationConfig, model: str, **model_kwarg

56 )

57

58 # Fetch match table from each worker via RPC and sum across workers.

57 - worker_tables = llm.llm_engine.engine_core.collective_rpc(

59 + worker_tables: list[dict[str, int]] = llm.llm_engine.engine_core.collective_rpc(

60 "get_compilation_match_table"

61 )

62 combined: defaultdict[str, int] = defaultdict(int)

tests/config/test_bailing_mtp_config.py

+3 -1

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

1 # SPDX-License-Identifier: Apache-2.0

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

3

4 +from typing import get_args

5 +

6 from transformers import PretrainedConfig

7

8 from vllm.config.speculative import MTPModelTypes, SpeculativeConfig

@@ -34,7 +36,7 @@ def test_bailing_hybrid_mtp_hf_config_override():

36 assert overridden.model_type == "bailing_hybrid_mtp"

37 assert overridden.architectures == ["BailingMoeV25MTPModel"]

38 assert overridden.n_predict == 1

37 - assert "bailing_hybrid_mtp" in MTPModelTypes.__args__

39 + assert "bailing_hybrid_mtp" in get_args(MTPModelTypes)

40

41

42 def test_bailing_hybrid_mtp_model_arch_config():

tests/config/test_multimodal_config.py

+2 -2

@@ -15,13 +15,13 @@ from vllm.v1.attention.backends.registry import AttentionBackendEnum

15

16

17 def test_mm_encoder_attn_backend_str_conversion():

18 - config = MultiModalConfig(mm_encoder_attn_backend="FLASH_ATTN")

18 + config = MultiModalConfig(mm_encoder_attn_backend="FLASH_ATTN") # type: ignore[arg-type]

19 assert config.mm_encoder_attn_backend == AttentionBackendEnum.FLASH_ATTN

20

21

22 def test_mm_encoder_attn_backend_invalid():

23 with pytest.raises(ValueError):

24 - MultiModalConfig(mm_encoder_attn_backend="not_a_backend")

24 + MultiModalConfig(mm_encoder_attn_backend="not_a_backend") # type: ignore[arg-type]

25

26

27 def test_mm_encoder_attn_backend_hash_updates():

tests/config/test_speculative_draft_max_position_embeddings.py

+2 -1

@@ -14,6 +14,7 @@ methods only.

14 """

15

16 import logging

17 +from typing import Literal

18

19 import pytest

20 from transformers import PretrainedConfig

@@ -77,7 +78,7 @@ def test_override_ignores_missing_attribute(vllm_caplog: pytest.LogCaptureFixtur

78 @pytest.mark.cpu_test

79 @pytest.mark.parametrize("method", ["eagle", "eagle3"])

80 def test_eagle_draft_inherits_target_max_model_len(

80 - method: str, vllm_caplog: pytest.LogCaptureFixture

81 + method: Literal["eagle", "eagle3"], vllm_caplog: pytest.LogCaptureFixture

82 ):

83 target_model_config = ModelConfig(LLAMA3_TARGET)

84 assert target_model_config.max_model_len > 2048

tests/entrypoints/generate/generative_scoring/test_generative_scoring.py

+1 -1

@@ -285,7 +285,7 @@ class TestPromptBuilding:

285 )

286

287 for i, exp in enumerate(expected):

288 - assert engine_inputs[i]["prompt_token_ids"] == exp

288 + assert engine_inputs[i]["prompt_token_ids"] == exp # type: ignore[typeddict-item]

289

290

291 class TestGeneration:

tests/entrypoints/tool_parsers/test_hermes_tool_parser.py

+1 -1

@@ -45,7 +45,7 @@ class ServerConfig(TypedDict, total=False):

45 model: str

46 arguments: list[str]

47 model_arg: str

48 - tool_parser: ToolParser

48 + tool_parser: type[ToolParser]

49

50

51 CONFIGS: dict[str, ServerConfig] = {

tests/entrypoints/weight_transfer/test_weight_transfer_llm.py

+1

@@ -329,4 +329,5 @@ def test_weight_transfer_config_backend():

329 )

330

331 config = llm.llm_engine.vllm_config.weight_transfer_config

332 + assert config is not None

333 assert config.backend == "nccl"

tests/kernels/core/test_fused_qk_norm_rope.py

+2

@@ -35,10 +35,12 @@ def _apply_qk_norm_rope(

35

36 q_by_head = q.view(*q.shape[:-1], q.shape[-1] // head_dim, head_dim)

37 q_by_head = q_norm.forward_native(q_by_head)

38 + assert isinstance(q_by_head, torch.Tensor)

39 q = q_by_head.view(q.shape)

40

41 k_by_head = k.view(*k.shape[:-1], k.shape[-1] // head_dim, head_dim)

42 k_by_head = k_norm.forward_native(k_by_head)

43 + assert isinstance(k_by_head, torch.Tensor)

44 k = k_by_head.view(k.shape)

45

46 q, k = rope.forward_native(positions, q, k)

tests/kernels/core/test_vit_fp8_attn.py

+3 -4

@@ -34,10 +34,9 @@ NUM_HEADS = [16]

34 @pytest.fixture

35 def _fp8_attention():

36 """Create FP8-enabled MMEncoderAttention via config."""

37 - from types import SimpleNamespace

38 - from unittest.mock import patch

37 + from unittest.mock import MagicMock, patch

38

40 - from vllm.config import VllmConfig, set_current_vllm_config

39 + from vllm.config import ModelConfig, VllmConfig, set_current_vllm_config

40 from vllm.config.multimodal import MultiModalConfig

41

42 if not is_flashinfer_cudnn_fp8_prefill_attn_supported():

@@ -45,7 +44,7 @@ def _fp8_attention():

44

45 mm_config = MultiModalConfig(mm_encoder_attn_dtype="fp8")

46 vllm_config = VllmConfig()

48 - vllm_config.model_config = SimpleNamespace(multimodal_config=mm_config)

47 + vllm_config.model_config = MagicMock(spec=ModelConfig, multimodal_config=mm_config)

48

49 # MMEncoderAttention reads torch.get_default_dtype() during init

50 # to determine the output dtype. In real model loading this is bf16.

tests/kernels/core/test_vit_fp8_scaling.py

+5 -6

@@ -4,8 +4,7 @@

4

5 import contextlib

6 import json

7 -from types import SimpleNamespace

8 -from unittest.mock import patch

7 +from unittest.mock import MagicMock, patch

8

9 import pytest

10 import torch

@@ -48,7 +47,7 @@ def _build_attention(mm_config, attn_backend=None):

47 invokes ``process_weights_after_loading`` to simulate the model loader's

48 auto-scan. Yields ``None`` if FlashInfer cuDNN is not available.

49 """

51 - from vllm.config import VllmConfig, set_current_vllm_config

50 + from vllm.config import ModelConfig, VllmConfig, set_current_vllm_config

51 from vllm.model_executor.layers.attention.mm_encoder_attention import (

52 MMEncoderAttention,

53 )

@@ -65,7 +64,7 @@ def _build_attention(mm_config, attn_backend=None):

64 return

65

66 vllm_config = VllmConfig()

68 - vllm_config.model_config = SimpleNamespace(multimodal_config=mm_config)

67 + vllm_config.model_config = MagicMock(spec=ModelConfig, multimodal_config=mm_config)

68

69 with (

70 set_current_vllm_config(vllm_config),

@@ -212,7 +211,7 @@ def test_aiter_static_scales_loaded(tmp_path) -> None:

211

212 def test_static_scales_missing_layer(tmp_path) -> None:

213 """Verify error when requested layer is not in the scale file."""

215 - from vllm.config import VllmConfig, set_current_vllm_config

214 + from vllm.config import ModelConfig, VllmConfig, set_current_vllm_config

215 from vllm.config.multimodal import MultiModalConfig

216 from vllm.v1.attention.backends.registry import AttentionBackendEnum

217

@@ -228,7 +227,7 @@ def test_static_scales_missing_layer(tmp_path) -> None:

227 mm_encoder_fp8_scale_path=str(scale_file),

228 )

229 vllm_config = VllmConfig()

231 - vllm_config.model_config = SimpleNamespace(multimodal_config=mm_config)

230 + vllm_config.model_config = MagicMock(spec=ModelConfig, multimodal_config=mm_config)

231

232 from vllm.model_executor.layers.attention.mm_encoder_attention import (

233 MMEncoderAttention,

tests/kernels/mamba/cpu/test_cpu_gdn_ops.py

+24 -9

@@ -12,6 +12,7 @@ import vllm._custom_ops as ops

12 from vllm.model_executor.layers.mamba.ops.cpu import gdn_attention

13 from vllm.platforms import current_platform

14 from vllm.utils.torch_utils import set_random_seed

15 +from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadata

16

17 if not current_platform.is_cpu():

18 pytest.skip("skipping CPU-only tests", allow_module_level=True)

@@ -452,12 +453,17 @@ def test_spec_aware_mixed_routing_preserves_token_order(

453

454 spec_indices = torch.tensor([0, 2])

455 nonspec_indices = torch.tensor([1, 3])

455 - metadata = types.SimpleNamespace(

456 + metadata = GDNAttentionMetadata(

457 + num_prefills=1,

458 + num_prefill_tokens=0,

459 + num_decodes=0,

460 + num_decode_tokens=0,

461 + num_spec_decodes=0,

462 + num_spec_decode_tokens=0,

463 + num_actual_tokens=0,

464 spec_sequence_masks=torch.ones(1, dtype=torch.bool),

465 spec_token_indx=spec_indices,

466 non_spec_token_indx=nonspec_indices,

459 - num_prefills=1,

460 - num_decodes=0,

467 )

468 routed = []

469

@@ -501,13 +507,16 @@ def test_spec_aware_nonspec_materializes_state_indices(

507 state_indices = block_table[:, 0]

508 assert not state_indices.is_contiguous()

509

504 - metadata = types.SimpleNamespace(

505 - non_spec_state_indices_tensor=state_indices,

506 - non_spec_query_start_loc=torch.tensor([0, 2, 4], dtype=torch.int32),

507 - num_decodes=0,

508 - num_decode_tokens=0,

510 + metadata = GDNAttentionMetadata(

511 num_prefills=2,

512 num_prefill_tokens=4,

513 + num_decodes=0,

514 + num_decode_tokens=0,

515 + num_spec_decodes=0,

516 + num_spec_decode_tokens=0,

517 + num_actual_tokens=4,

518 + non_spec_state_indices_tensor=state_indices,

519 + non_spec_query_start_loc=torch.tensor([0, 2, 4], dtype=torch.int32),

520 has_initial_state=torch.tensor([False, False]),

521 )

522

@@ -575,8 +584,14 @@ def test_spec_forward_prepares_native_conv_metadata(

584 state_indices = block_table[:, 0]

585 accepted_counts = torch.tensor([1, 4], dtype=torch.int32)

586 assert not state_indices.is_contiguous()

578 - metadata = types.SimpleNamespace(

587 + metadata = GDNAttentionMetadata(

588 + num_prefills=0,

589 + num_prefill_tokens=0,

590 + num_decodes=0,

591 + num_decode_tokens=0,

592 num_spec_decodes=2,

593 + num_spec_decode_tokens=8,

594 + num_actual_tokens=8,

595 spec_state_indices_tensor=block_table,

596 spec_query_start_loc=torch.tensor([0, 4, 8], dtype=torch.int32),

597 num_accepted_tokens=accepted_counts,

tests/kernels/mamba/test_gdn_forward_core_split.py

+5 -1

@@ -174,7 +174,9 @@ def test_forward_core_split_matches_unified(

174 batch = BatchSpec(seq_lens=seq_lens, query_lens=query_lens)

175

176 builder = GDNAttentionMetadataBuilder(

177 - kv_cache_spec=MambaSpec(

177 + # GDNAttentionMetadataBuilder declares kv_cache_spec as AttentionSpec but

178 + # asserts isinstance(kv_cache_spec, MambaSpec). MambaSpec is correct here.

179 + kv_cache_spec=MambaSpec( # type: ignore[arg-type]

180 block_size=BLOCK_SIZE, shapes=((16, 64),), dtypes=(torch.float16,)

181 ),

182 layer_names=[PREFIX],

@@ -198,6 +200,7 @@ def test_forward_core_split_matches_unified(

200 # Full-batch chunk metadata for the unified reference path, built the same

201 # way the builder would for a non-split batch (backend-matched).

202 cu_full = meta_split.non_spec_query_start_loc

203 + assert cu_full is not None

204 if builder.gdn_prefill_backend == "cutedsl":

205 from vllm.model_executor.layers.mamba.ops.gdn_chunk_cutedsl import (

206 prepare_metadata_cutedsl,

@@ -227,6 +230,7 @@ def test_forward_core_split_matches_unified(

230 )

231

232 # Size the state pools from the indices the builder actually produced.

233 + assert meta_split.non_spec_state_indices_tensor is not None

234 pool_size = int(meta_split.non_spec_state_indices_tensor.max().item()) + 1

235 conv_state_shape, temporal_state_shape = (

236 MambaStateShapeCalculator.gated_delta_net_state_shape(

tests/kernels/mamba/test_precopy_mamba_align.py

+14 -7

@@ -21,7 +21,9 @@ The kernel must also no-op when ``src_col < 0`` (fresh request) or

21

22 from __future__ import annotations

23

24 +from collections.abc import Callable

25 from types import SimpleNamespace

26 +from typing import Any

27

28 import numpy as np

29 import torch

@@ -31,6 +33,8 @@ from vllm.platforms import current_platform

33 from vllm.v1.worker import mamba_utils as worker_mamba_utils

34 from vllm.v1.worker.mamba_utils import precopy_mamba_align_fused_kernel

35

36 +_parametrize: Callable[..., Callable[[Any], Any]]

37 +

38 try:

39 import pytest

40

@@ -40,17 +44,18 @@ try:

44 )

45 _parametrize = pytest.mark.parametrize

46 except ModuleNotFoundError: # allow running directly as ``python <thisfile>``

43 - pytest = None

47

48 def _cuda_required(fn):

49 return fn

50

48 - def _parametrize(_name, _values):

51 + def _no_parametrize(_name, _values):

52 def _deco(fn):

53 return fn

54

55 return _deco

56

57 + _parametrize = _no_parametrize

58 +

59

60 NUM_LAYERS = 3

61 CONV_WIDTH = 4 # conv_kernel - 1 + num_spec

@@ -308,8 +313,10 @@ def _make_preprocess_case(token_bias):

313 def test_preprocess_fused_align_matches_scalar_bookkeeping(monkeypatch, token_bias):

314 block_size = 4

315 mamba_spec = SimpleNamespace(block_size=block_size, num_speculative_blocks=1)

311 - cache_config = SimpleNamespace(enable_prefix_caching=True)

312 - kv_cache_config = SimpleNamespace()

316 + # Fakes stand in for real config/context objects. The fused path tests

317 + # below only touch the attributes preprocess_mamba actually reads.

318 + cache_config: Any = SimpleNamespace(enable_prefix_caching=True)

319 + kv_cache_config: Any = SimpleNamespace()

320 scalar_copy_calls = []

321

322 def fake_collect(

@@ -340,7 +347,7 @@ def test_preprocess_fused_align_matches_scalar_bookkeeping(monkeypatch, token_bi

347 scalar_case = _make_preprocess_case(token_bias)

348 fused_case = _make_preprocess_case(token_bias)

349

343 - scalar_copy_bufs = SimpleNamespace(

350 + scalar_copy_bufs: Any = SimpleNamespace(

351 mamba_group_ids=[0],

352 mamba_spec=mamba_spec,

353 offset=0,

@@ -357,8 +364,8 @@ def test_preprocess_fused_align_matches_scalar_bookkeeping(monkeypatch, token_bi

364 copy_bufs=scalar_copy_bufs,

365 )

366

360 - ctx = _FakePrecopyContext(len(fused_case[1].req_ids))

361 - fused_copy_bufs = SimpleNamespace(

367 + ctx: Any = _FakePrecopyContext(len(fused_case[1].req_ids))

368 + fused_copy_bufs: Any = SimpleNamespace(

369 mamba_group_ids=[0],

370 mamba_spec=mamba_spec,

371 offset=0,

tests/models/language/generation/test_mistral.py

+3 -3

@@ -261,7 +261,7 @@ def test_mistral_function_calling(vllm_runner, model: str, dtype: str) -> None:

261

262 model_output = outputs[0].outputs[0].text.strip()

263 assert model_output.startswith(tool_parser.bot_token), model_output

264 - parsed_message = tool_parser.extract_tool_calls(model_output, None)

264 + parsed_message = tool_parser.extract_tool_calls(model_output, None) # type: ignore[arg-type]

265

266 assert parsed_message.tools_called

267

@@ -304,7 +304,7 @@ def test_mistral_function_call_nested_json():

304

305 model_output = f"{parser.bot_token}get_current_weather{json.dumps(args_dict)}"

306

307 - parsed = parser.extract_tool_calls(model_output, None)

307 + parsed = parser.extract_tool_calls(model_output, None) # type: ignore[arg-type]

308

309 # Assertions: the tool call is detected and the full nested JSON is parsed

310 # without truncation.

@@ -337,7 +337,7 @@ def test_mistral_function_call_nested_json():

337 ]

338 )

339

340 - parsed = parser.extract_tool_calls(model_output, None)

340 + parsed = parser.extract_tool_calls(model_output, None) # type: ignore[arg-type]

341

342 # Assertions: the tool call is detected and the full nested JSON is parsed

343 # without truncation.

tests/models/quantization/test_bitsandbytes.py

+1 -1

@@ -323,7 +323,7 @@ def test_bitsandbytes_passes_revision_by_name():

323 return_value=["/folder/model.safetensors"],

324 ),

325 ):

326 - bnb.BitsAndBytesModelLoader._prepare_weights(fake_self, "org/model", "myrev")

326 + bnb.BitsAndBytesModelLoader._prepare_weights(fake_self, "org/model", "myrev") # type: ignore[arg-type]

327

328 mock_idx.assert_called_once()

329 assert mock_idx.call_args.kwargs.get("revision") == "myrev"

tests/plugins/bge_m3_sparse_plugin/bge_m3_sparse_processor/sparse_embeddings_processor.py

+3 -3

@@ -29,7 +29,7 @@ class BgeM3SparseEmbeddingsProcessor(

29 self.online_requests: dict[str, SparseEmbeddingCompletionRequestMixin] = {}

30 self.renderer: BaseRenderer = renderer

31 self.default_pooling_params = {}

32 - pooler_config: PoolerConfig = vllm_config.model_config.pooler_config

32 + pooler_config: PoolerConfig | None = vllm_config.model_config.pooler_config

33 if pooler_config is not None:

34 for param in ["use_activation", "dimensions"]:

35 if getattr(pooler_config, param, None) is None:

@@ -91,11 +91,11 @@ class BgeM3SparseEmbeddingsProcessor(

91 ) -> list[SparseEmbeddingTokenWeight]:

92 token_ids = sparse_embedding.keys()

93 token_weights = sparse_embedding.values()

94 - tokens = [None] * len(token_ids)

94 + tokens: Sequence[str | None] = [None] * len(token_ids)

95

96 if return_tokens and self.renderer is not None:

97 tokens = convert_ids_list_to_tokens(

98 - self.renderer.get_tokenizer(), token_ids

98 + self.renderer.get_tokenizer(), list(token_ids)

99 )

100 sparse_embedding_output: list[SparseEmbeddingTokenWeight] = []

101 for token_id, weight, token in zip(token_ids, token_weights, tokens):

tests/plugins/prithvi_io_processor_plugin/prithvi_io_processor/prithvi_processor.py

+6 -6

@@ -170,7 +170,7 @@ def load_image(

170 list of meta info for each image in *file_paths*

171 """

172

173 - imgs = []

173 + imgs: list[np.ndarray] = []

174 metas = []

175 temporal_coords = []

176 location_coords = []

@@ -209,11 +209,11 @@ def load_image(

209 except Exception:

210 logger.exception("Could not extract timestamp for %s", file)

211

212 - imgs = np.stack(imgs, axis=0) # num_frames, H, W, C

213 - imgs = np.moveaxis(imgs, -1, 0).astype("float32") # C, num_frames, H, W

214 - imgs = np.expand_dims(imgs, axis=0) # add batch di

212 + stacked = np.stack(imgs, axis=0) # num_frames, H, W, C

213 + stacked = np.moveaxis(stacked, -1, 0).astype("float32") # C, num_frames, H, W

214 + stacked = np.expand_dims(stacked, axis=0) # add batch di

215

216 - return imgs, temporal_coords, location_coords, metas

216 + return stacked, temporal_coords, location_coords, metas

217

218

219 class PrithviMultimodalDataProcessor(IOProcessor[ImagePrompt, ImageRequestOutput]):

@@ -327,7 +327,7 @@ class PrithviMultimodalDataProcessor(IOProcessor[ImagePrompt, ImageRequestOutput

327 }

328 )

329

330 - return prompts

330 + return prompts # type: ignore[return-value]

331

332 def post_process(

333 self,

tests/plugins/vllm_add_dummy_platform/vllm_add_dummy_platform/dummy_platform.py

+8 -11

@@ -6,6 +6,8 @@ from vllm.platforms.interface import Platform, PlatformEnum

6

7 if TYPE_CHECKING:

8 from vllm.config import VllmConfig

9 + from vllm.v1.attention.backends.registry import AttentionBackendEnum

10 + from vllm.v1.attention.selector import AttentionSelectorConfig

11 else:

12 VllmConfig = None

13

@@ -20,16 +22,11 @@ class DummyPlatform(Platform):

22 def check_and_update_config(cls, vllm_config: VllmConfig) -> None:

23 vllm_config.compilation_config.custom_ops = ["all"]

24

25 + @classmethod

26 def get_attn_backend_cls(

24 - self,

25 - backend_name,

26 - head_size,

27 - dtype,

28 - kv_cache_dtype,

29 - block_size,

30 - use_mla,

31 - has_sink,

32 - use_sparse,

33 - use_mm_prefix,

34 - ):

27 + cls,

28 + selected_backend: "AttentionBackendEnum",

29 + attn_selector_config: "AttentionSelectorConfig",

30 + num_heads: int | None = None,

31 + ) -> str:

32 return "vllm_add_dummy_platform.dummy_attention_backend.DummyAttentionBackend" # noqa E501

tests/plugins/vllm_add_dummy_stat_logger/dummy_stat_logger/dummy_stat_logger.py

+3 -1

@@ -17,7 +17,9 @@ class DummyStatLogger(StatLoggerBase):

17 self.logged = False

18 self.engine_initialized = False

19

20 - def record(self, scheduler_stats, iteration_stats, mm_cache_stats, engine_idx):

20 + def record(

21 + self, scheduler_stats, iteration_stats, mm_cache_stats=None, engine_idx=0

22 + ):

23 self.recorded.append(

24 (scheduler_stats, iteration_stats, mm_cache_stats, engine_idx)

25 )

tests/plugins_tests/gguf/test_gguf_plugin_multimodal.py

+1 -1

@@ -92,7 +92,7 @@ def run_multimodal_gguf_test(

92 num_logprobs: int,

93 ):

94 # Load images at runtime (inside subprocess) to avoid pickle issues

95 - images = [ImageAsset(name).pil_image for name in model.image_names]

95 + images = [ImageAsset(name).pil_image for name in model.image_names] # type: ignore[arg-type]

96 size_factors = [0.25, 0.5, 1.0]

97 inputs_per_image = [

98 (

tests/plugins_tests/lora_resolvers/test_hf_hub_resolver.py

+1

@@ -31,6 +31,7 @@ async def test_hf_resolver_with_direct_path():

31 assert hf_resolver is not None

32

33 lora_request = await hf_resolver.resolve_lora(LORA_REPO_MODEL_NAME, LORA_REPO)

34 + assert lora_request is not None

35 assert lora_request.lora_name == LORA_REPO

36 assert REPO_DOWNLOAD_DIR in lora_request.lora_path

37 assert "adapter_config.json" in os.listdir(lora_request.lora_path)

tests/spec_decode/test_custom_proposer.py

+1

@@ -30,6 +30,7 @@ class DummyDraftProposer:

30 Args:

31 vllm_config: vLLM configuration containing model and speculative settings.

32 """

33 + assert vllm_config.speculative_config is not None

34 self.num_speculative_tokens = (

35 vllm_config.speculative_config.num_speculative_tokens

36 )

tests/transformers_utils/test_processor.py

+2 -2

@@ -44,7 +44,7 @@ class _ProcWithUnpack:

44

45 def test_get_processor_kwargs_from_processor_unpack_path_returns_full_union():

46 proc = _ProcWithUnpack()

47 - keys = get_processor_kwargs_keys(get_processor_kwargs_type(proc))

47 + keys = get_processor_kwargs_keys(get_processor_kwargs_type(proc)) # type: ignore[arg-type]

48 _assert_has_all_expected(keys)

49

50

@@ -63,5 +63,5 @@ def test_get_processor_kwargs_from_processor_module_scan_returns_full_union():

63 assert hasattr(mod, "_FakeProcessorKwargs")

64

65 proc = _ProcWithoutUnpack()

66 - keys = get_processor_kwargs_keys(get_processor_kwargs_type(proc))

66 + keys = get_processor_kwargs_keys(get_processor_kwargs_type(proc)) # type: ignore[arg-type]

67 _assert_has_all_expected(keys)

tests/v1/distributed/test_async_llm_dp.py

+5 -3

@@ -12,13 +12,14 @@ import pytest

12

13 from vllm import SamplingParams

14 from vllm.config import VllmConfig

15 +from vllm.config.parallel import DataParallelBackend

16 from vllm.engine.arg_utils import AsyncEngineArgs

17 from vllm.inputs import PromptType

18 from vllm.outputs import RequestOutput

19 from vllm.platforms import current_platform

20 from vllm.sampling_params import RequestOutputKind

21 from vllm.v1.engine.async_llm import AsyncLLM

21 -from vllm.v1.engine.core_client import DPAsyncMPClient

22 +from vllm.v1.engine.core_client import DPLBAsyncMPClient

23 from vllm.v1.metrics.loggers import StatLoggerBase

24 from vllm.v1.metrics.stats import IterationStats, MultiModalCacheStats, SchedulerStats

25

@@ -82,7 +83,7 @@ async def generate(

83 async def test_load(

84 model: str,

85 output_kind: RequestOutputKind,

85 - data_parallel_backend: str,

86 + data_parallel_backend: DataParallelBackend,

87 async_scheduling: bool,

88 ):

89 if async_scheduling and data_parallel_backend == "ray":

@@ -162,7 +163,8 @@ async def test_load(

163 assert not engine.output_processor.has_unfinished_requests()

164

165 # testing internals here which may break

165 - core_client: DPAsyncMPClient = engine.engine_core

166 + core_client = engine.engine_core

167 + assert isinstance(core_client, DPLBAsyncMPClient)

168 # the engines only synchronize stopping every N steps so

169 # allow a small amount of time here.

170 for _ in range(10):

tests/v1/distributed/test_eagle_dp.py

+1 -1

@@ -61,7 +61,7 @@ async def test_run_eagle_dp(monkeypatch: pytest.MonkeyPatch, attn_backend: str):

61 data_parallel_backend="mp", # ray takes more time

62 trust_remote_code=True,

63 max_model_len=16384,

64 - attention_config={"backend": attn_backend},

64 + attention_config={"backend": attn_backend}, # type: ignore[arg-type]

65 )

66

67 eagle_engine_args = replace(

tests/v1/shutdown/test_processor_error.py

+8 -3

@@ -9,7 +9,7 @@ import pytest

9 from tests.v1.shutdown.utils import SHUTDOWN_TEST_TIMEOUT_SEC

10 from vllm import SamplingParams

11 from vllm.engine.arg_utils import AsyncEngineArgs

12 -from vllm.inputs import TokensPrompt

12 +from vllm.inputs import ExplicitEncoderDecoderPrompt

13 from vllm.sampling_params import RequestOutputKind

14 from vllm.v1.engine.async_llm import AsyncLLM

15 from vllm.v1.engine.exceptions import EngineGenerateError

@@ -29,9 +29,14 @@ async def test_async_llm_processor_error(model: str) -> None:

29 async_llm = AsyncLLM.from_engine_args(engine_args)

30

31 async def generate(request_id: str):

32 - # [] is not allowed and will raise a ValueError in Processor.

32 + # An encoder/decoder prompt is rejected by the Processor for a

33 + # decoder only model.

34 generator = async_llm.generate(

34 - TokensPrompt([]), request_id=request_id, sampling_params=SamplingParams()

35 + ExplicitEncoderDecoderPrompt(

36 + encoder_prompt="Hello my name is", decoder_prompt=None

37 + ),

38 + request_id=request_id,

39 + sampling_params=SamplingParams(),

40 )

41 try:

42 async for _ in generator:

tools/pre_commit/mypy.py

+41 -24

@@ -21,34 +21,55 @@ import sys

21

22 import regex as re

23

24 +# Paths verified clean under follow_imports="silent". Matched before

25 +# SEPARATE_GROUPS, so these files join the default group and are checked at the

26 +# stricter setting even while a parent directory remains in SEPARATE_GROUPS.

27 +#

28 +# Fixing a directory means moving it from SEPARATE_GROUPS to here. Without that

29 +# move the fixes are not enforced because "tests" claims every file below it.

30 +SILENT_GROUPS = [

31 + "tests/compile/correctness_e2e",

32 + "tests/compile/fullgraph",

33 + "tests/compile/fusions_e2e",

34 + "tests/config",

35 + "tests/entrypoints/generate",

36 + "tests/entrypoints/tool_parsers",

37 + "tests/entrypoints/weight_transfer",

38 + "tests/kernels/core",

39 + "tests/kernels/mamba",

40 + "tests/models/language",

41 + "tests/models/quantization",

42 + "tests/plugins/bge_m3_sparse_plugin",

43 + "tests/plugins/prithvi_io_processor_plugin",

44 + "tests/plugins/vllm_add_dummy_platform",

45 + "tests/plugins/vllm_add_dummy_stat_logger",

46 + "tests/plugins_tests/gguf",

47 + "tests/plugins_tests/lora_resolvers",

48 + "tests/spec_decode",

49 + "tests/transformers_utils",

50 + "tests/v1/distributed",

51 + "tests/v1/shutdown",

52 +]

53 +

54 # After fixing errors resulting from changing follow_imports

25 -# from "skip" to "silent", remove its directory from SEPARATE_GROUPS.

55 +# from "skip" to "silent", move its directory to SILENT_GROUPS.

56 SEPARATE_GROUPS = [

57 "tests",

58 "tests/benchmarks",

29 - "tests/compile/correctness_e2e",

30 - "tests/config",

59 "tests/compile",

32 - "tests/compile/fullgraph",

33 - "tests/compile/fusions_e2e",

60 "tests/compile/passes",

61 "tests/distributed",

62 "tests/entrypoints/anthropic",

37 - "tests/entrypoints/generate",

63 "tests/entrypoints/llm",

64 "tests/entrypoints/multimodal",

65 "tests/entrypoints/openai",

66 "tests/entrypoints/pooling",

67 "tests/entrypoints/serve",

68 "tests/entrypoints/speech_to_text",

44 - "tests/entrypoints/tool_parsers",

69 "tests/entrypoints/unit_tests",

46 - "tests/entrypoints/weight_transfer",

70 "tests/kernels",

71 "tests/kernels/attention",

49 - "tests/kernels/core",

72 "tests/kernels/helion",

51 - "tests/kernels/mamba",

73 "tests/kernels/moe",

74 "tests/kernels/quantization",

75 "tests/lora",

@@ -57,34 +78,23 @@ SEPARATE_GROUPS = [

78 "tests/model_executor/model_loader",

79 "tests/models",

80 "tests/models/test_initialization.py",

60 - "tests/models/language",

81 "tests/models/multimodal",

62 - "tests/models/quantization",

82 "tests/multimodal",

83 "tests/parser",

65 - "tests/plugins_tests/gguf",

66 - "tests/plugins_tests/lora_resolvers",

67 - "tests/plugins/bge_m3_sparse_plugin",

68 - "tests/plugins/prithvi_io_processor_plugin",

69 - "tests/plugins/vllm_add_dummy_platform",

70 - "tests/plugins/vllm_add_dummy_stat_logger",

84 "tests/plugins_tests",

85 "tests/quantization",

86 "tests/reasoning",

87 "tests/renderers",

88 "tests/samplers",

76 - "tests/spec_decode",

89 "tests/tokenizers_",

90 "tests/tool_parsers",

91 "tests/tool_use",

80 - "tests/transformers_utils",

92 "tests/utils_",

93 "tests/v1",

94 "tests/v1/attention",

95 "tests/v1/core",

96 "tests/v1/cudagraph",

97 "tests/v1/determinism",

87 - "tests/v1/distributed",

98 "tests/v1/e2e",

99 "tests/v1/ec_connector",

100 "tests/v1/engine",

@@ -94,7 +104,6 @@ SEPARATE_GROUPS = [

104 "tests/v1/logits_processors",

105 "tests/v1/metrics",

106 "tests/v1/sample",

97 - "tests/v1/shutdown",

107 "tests/v1/simple_kv_offload",

108 "tests/v1/spec_decode",

109 "tests/v1/streaming_input",

@@ -142,14 +151,22 @@ def group_files(changed_files: list[str]) -> dict[str, list[str]]:

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

152 """

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

145 - file_groups = {"": []}

154 + silent_pattern = re.compile(f"^({'|'.join(SILENT_GROUPS)}).*")

155 + file_groups: dict[str, list[str]] = {"": []}

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

157 + # Longest path first so a sub-directory is not shadowed by its parent

158 + separate_groups = sorted(SEPARATE_GROUPS, key=len, reverse=True)

159 for changed_file in changed_files:

160 # Skip files which should be ignored completely

161 if exclude_pattern.match(changed_file):

162 continue

163 + # Already-fixed paths go in the default group, which runs at the

164 + # stricter follow_imports setting from pyproject.toml

165 + if silent_pattern.match(changed_file):

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

167 + continue

168 # Group files by mypy call

152 - for directory in SEPARATE_GROUPS:

169 + for directory in separate_groups:

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

171 file_groups[directory].append(changed_file)

172 break

[MyPy][1/N] Fix mypy errors in some tests/ directories and enforce follow-imports=silent (#49570)