fc2c801ad9 · tduncan/Nvidia-Optimization-Stack-VLLM

pyproject.toml

+5

@@ -117,6 +117,11 @@ ignore_missing_imports = true

117 check_untyped_defs = true

118 follow_imports = "silent"

119

120 +[[tool.mypy.overrides]]

121 +module = "tests.*"

122 +disable_error_code = ["arg-type", "assignment"]

123 +enable_error_code = ["method-assign"]

124 +

125 [tool.pytest.ini_options]

126 markers = [

127 "slow_test",

tests/benchmarks/test_bfcl_dataset.py

+2 -2

@@ -271,7 +271,7 @@ def test_bfcl_prompt_len_includes_tools(tmp_path: Path) -> None:

271 args.num_prompts = 1

272

273 with _patch_hf_api(fake_download):

274 - samples = get_samples(args, fake) # type: ignore[arg-type]

274 + samples = get_samples(args, fake)

275

276 assert len(samples) == 1

277 assert captured["tools"] is not None, (

@@ -311,7 +311,7 @@ def test_bfcl_prompt_len_falls_back_when_tokenizer_rejects_tools(

311 args.num_prompts = 1

312

313 with _patch_hf_api(fake_download):

314 - samples = get_samples(args, _LegacyTokenizer()) # type: ignore[arg-type]

314 + samples = get_samples(args, _LegacyTokenizer())

315

316 assert len(samples) == 1

317 assert samples[0].prompt_len > 0

tests/benchmarks/test_custom_dataset_chat_template_kwargs.py

+2 -2

@@ -58,7 +58,7 @@ def test_chat_template_kwargs_forwarded(tmp_path: Path) -> None:

58 _write_one(jsonl)

59

60 tok = _RecordingTokenizer()

61 - get_samples(_args(str(jsonl), {"thinking": True}), tok) # type: ignore[arg-type]

61 + get_samples(_args(str(jsonl), {"thinking": True}), tok)

62

63 assert tok.captured_kwargs == {"thinking": True}

64

@@ -70,6 +70,6 @@ def test_chat_template_kwargs_default_is_noop(tmp_path: Path) -> None:

70 _write_one(jsonl)

71

72 tok = _RecordingTokenizer()

73 - get_samples(_args(str(jsonl), None), tok) # type: ignore[arg-type]

73 + get_samples(_args(str(jsonl), None), tok)

74

75 assert tok.captured_kwargs == {}

tests/compile/passes/distributed/test_async_tp.py

+1 -1

@@ -285,7 +285,7 @@ def test_async_tp_pass_requires_full_graph_compilation():

285 ]

286

287 async_tp_pass = object.__new__(AsyncTPPass)

288 - async_tp_pass.compilation_config = vllm_config.compilation_config # type: ignore[assignment]

288 + async_tp_pass.compilation_config = vllm_config.compilation_config

289

290 with pytest.raises(

291 AssertionError, match="AsyncTPPass requires full-graph compilation"

tests/compile/passes/distributed/test_fusion_all_reduce.py

+2 -2

@@ -111,7 +111,7 @@ def test_fused_ar_workspace_hidden_dim(

111 target_hidden: int, draft_hidden: int | None, expected: int

112 ):

113 config = _make_workspace_hidden_config(target_hidden, draft_hidden)

114 - assert _fused_ar_workspace_hidden_dim(config) == expected # type: ignore[arg-type]

114 + assert _fused_ar_workspace_hidden_dim(config) == expected

115

116

117 def test_fused_ar_workspace_hidden_dim_spec_without_draft():

@@ -120,7 +120,7 @@ def test_fused_ar_workspace_hidden_dim_spec_without_draft():

120 model_config=SimpleNamespace(get_hidden_size=lambda: 2048),

121 speculative_config=SimpleNamespace(draft_model_config=None),

122 )

123 - assert _fused_ar_workspace_hidden_dim(config) == 2048 # type: ignore[arg-type]

123 + assert _fused_ar_workspace_hidden_dim(config) == 2048

124

125

126 class TestAllReduceRMSNormModel(torch.nn.Module):

tests/compile/passes/distributed/test_sequence_parallelism.py

+1 -1

@@ -222,7 +222,7 @@ def test_sequence_parallelism_pass_requires_full_graph_compilation():

222 ]

223

224 sequence_parallelism_pass = object.__new__(SequenceParallelismPass)

225 - sequence_parallelism_pass.compilation_config = vllm_config.compilation_config # type: ignore[assignment]

225 + sequence_parallelism_pass.compilation_config = vllm_config.compilation_config

226 sequence_parallelism_pass.min_token_num = 1

227

228 with pytest.raises(

tests/compile/passes/test_pass_manager.py

+1 -1

@@ -28,7 +28,7 @@ def test_bad_callable():

28 pass_manager.configure(config)

29

30 with pytest.raises(AssertionError):

31 - pass_manager.add(simple_callable) # type: ignore[arg-type]

31 + pass_manager.add(simple_callable)

32

33

34 # Pass that inherits from InductorPass

tests/compile/passes/test_silu_mul_quant_fusion.py

+1 -1

@@ -331,7 +331,7 @@ def test_fusion_silu_and_mul_quant(

331 backend = TestBackend(*passes)

332 model = model_class(

333 hidden_size=hidden_size,

334 - force_kernel=force_kernel, # type: ignore[arg-type]

334 + force_kernel=force_kernel,

335 x=x,

336 dtype=dtype,

337 )

tests/config/test_config_generation.py

+2 -2

@@ -133,7 +133,7 @@ def test_get_diff_sampling_param_includes_penalties():

133 },

134 )

135

136 - diff = ModelConfig.get_diff_sampling_param(fake) # type: ignore[arg-type]

136 + diff = ModelConfig.get_diff_sampling_param(fake)

137

138 assert diff["presence_penalty"] == 1.5

139 assert diff["frequency_penalty"] == 0.5

@@ -161,7 +161,7 @@ def test_generation_config_penalties_survive_diff_sampling_filter():

161 },

162 )

163

164 - diff = ModelConfig.get_diff_sampling_param(fake) # type: ignore[arg-type]

164 + diff = ModelConfig.get_diff_sampling_param(fake)

165

166 assert diff["presence_penalty"] == 1.1

167 assert diff["frequency_penalty"] == 0.7

tests/config/test_multimodal_config.py

+5 -5

@@ -40,18 +40,18 @@ def test_supports_multimodal_inputs(model_id, limit_mm_per_prompt, expected):

40

41

42 def test_mm_encoder_attn_backend_str_conversion():

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

43 + config = MultiModalConfig(mm_encoder_attn_backend="FLASH_ATTN")

44 assert config.mm_encoder_attn_backend == AttentionBackendEnum.FLASH_ATTN

45

46

47 def test_mm_encoder_attn_backend_invalid():

48 with pytest.raises(ValueError):

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

49 + MultiModalConfig(mm_encoder_attn_backend="not_a_backend")

50

51

52 def test_mm_hasher_algorithm_invalid():

53 with pytest.raises(ValueError, match="mm_hasher_algorithm"):

54 - MultiModalConfig(mm_hasher_algorithm="md5") # type: ignore[arg-type]

54 + MultiModalConfig(mm_hasher_algorithm="md5")

55

56

57 def test_mm_encoder_attn_backend_hash_updates():

@@ -560,7 +560,7 @@ def _resolve_mm_processor_device(

560 """Run the `auto` resolution and report where the processor ended up."""

561 mm_config = MultiModalConfig(

562 mm_processor_kwargs={} if device is None else {"device": device},

563 - mm_tensor_ipc=mm_tensor_ipc, # type: ignore[arg-type]

563 + mm_tensor_ipc=mm_tensor_ipc,

564 )

565 model_config = MagicMock(spec=ModelConfig)

566 model_config.multimodal_config = mm_config

@@ -662,7 +662,7 @@ def _resolve_mm_video_decode_device(

662 the resulting video media IO kwargs."""

663 mm_config = MultiModalConfig(

664 mm_processor_kwargs={} if device is None else {"device": device},

665 - mm_tensor_ipc=mm_tensor_ipc, # type: ignore[arg-type]

665 + mm_tensor_ipc=mm_tensor_ipc,

666 media_io_kwargs={} if video_kwargs is None else {"video": dict(video_kwargs)},

667 )

668 model_config = MagicMock(spec=ModelConfig)

tests/conftest.py

+1 -1

@@ -1163,7 +1163,7 @@ class VllmRunner:

1163 output_logprobs = sample.logprobs

1164 if include_prompt_token_ids:

1165 outputs.append(

1166 - ( # type: ignore[arg-type]

1166 + (

1167 output_ids,

1168 output_str,

1169 output_logprobs,

tests/entrypoints/anthropic/test_anthropic_messages_conversion.py

+1 -1

@@ -1447,7 +1447,7 @@ class TestDetectMergeInlineSystem:

1447 model_config=None,

1448 )

1449 assert (

1450 - AnthropicServingMessages._should_merge_inline_system(online_renderer) # type: ignore[arg-type]

1450 + AnthropicServingMessages._should_merge_inline_system(online_renderer)

1451 is expected

1452 )

1453 assert mock_logger.warning_once.called is expected

tests/entrypoints/cohere/test_cohere_chat_message.py

+2 -2

@@ -29,7 +29,7 @@ from vllm.entrypoints.cohere.cohere_chat_message import (

29 class TestCitationSource:

30 def test_invalid_type_rejected(self):

31 with pytest.raises(ValidationError):

32 - CitationSource(type="other") # type: ignore[arg-type]

32 + CitationSource(type="other")

33

34 def test_none_fields_excluded_from_dump(self):

35 s = CitationSource(type="document", id="d1")

@@ -65,7 +65,7 @@ class TestCitationSource:

65 class TestCitation:

66 def test_invalid_type_rejected(self):

67 with pytest.raises(ValidationError):

68 - Citation(type="OTHER") # type: ignore[arg-type]

68 + Citation(type="OTHER")

69

70 def test_dump_excludes_none_fields(self):

71 c = Citation(start=0, end=5, text="hello")

tests/entrypoints/cohere/test_protocol.py

+2 -2

@@ -136,7 +136,7 @@ class TestCohereChatV2Response:

136 with pytest.raises(ValidationError):

137 CohereChatV2Response(

138 id="r1",

139 - finish_reason="NOT_A_REASON", # type: ignore[arg-type]

139 + finish_reason="NOT_A_REASON",

140 message=msg,

141 )

142

@@ -246,5 +246,5 @@ class TestStreamingEventTypeField:

246 MessageStartEvent(

247 id="a",

248 delta={"message": {"role": "assistant"}},

249 - type="other", # type: ignore[arg-type]

249 + type="other",

250 )

tests/entrypoints/launchers/api_server/test_api_server_process_manager.py

+3 -3

@@ -308,7 +308,7 @@ def test_external_process_monitoring(api_server_args):

308 try:

309 wait_for_completion_or_failure(

310 api_server_manager=manager,

311 - coordinator=mock_coordinator, # type: ignore[arg-type]

311 + coordinator=mock_coordinator,

312 )

313 except Exception as e:

314 result["exception"] = e

@@ -377,7 +377,7 @@ def test_gather_actual_addresses_end_to_end():

377 manager = APIServerProcessManager(

378 listen_address=f"tcp://{host}:0",

379 sock=sock,

380 - args="test_args", # type: ignore[arg-type]

380 + args="test_args",

381 num_servers=num_servers,

382 input_addresses=placeholder_inputs,

383 output_addresses=placeholder_outputs,

@@ -424,7 +424,7 @@ def test_gather_actual_addresses_child_crash_before_report():

424 manager = APIServerProcessManager(

425 listen_address=f"tcp://{host}:0",

426 sock=sock,

427 - args="test_args", # type: ignore[arg-type]

427 + args="test_args",

428 num_servers=num_servers,

429 input_addresses=placeholder_inputs,

430 output_addresses=placeholder_outputs,

tests/entrypoints/multimodal/llm/test_mm_processor_kwargs.py

+1 -1

@@ -14,7 +14,7 @@ from vllm.outputs import RequestOutput

14

15 def _make_mock_llm() -> LLM:

16 llm = object.__new__(LLM)

17 - llm.model_config = SimpleNamespace( # type: ignore[assignment]

17 + llm.model_config = SimpleNamespace(

18 runner_type="generate", enable_prompt_embeds=False

19 )

20 return llm

tests/entrypoints/openai/test_render_parity.py

+1 -1

@@ -225,7 +225,7 @@ def online_renderer(model_config: MockModelConfig, request) -> OnlineRenderer:

225 renderer.tokenizer = MagicMock()

226

227 return OnlineRenderer(

228 - model_config=model_config, # type: ignore[arg-type]

228 + model_config=model_config,

229 renderer=renderer,

230 request_logger=None,

231 chat_template=None,

tests/entrypoints/pooling/embed/test_io_processor.py

+1 -1

@@ -411,7 +411,7 @@ class TestChunkedEmbeddingProcessing:

411 @classmethod

412 def _make_handler(cls):

413 handler = object.__new__(EmbedIOProcessor)

414 - handler.model_config = cls._FakeModelConfig() # type: ignore[assignment]

414 + handler.model_config = cls._FakeModelConfig()

415 handler.enable_chunked_processing = True

416 return handler

417

tests/entrypoints/pooling/scoring/test_io_processor_unit.py

+2 -2

@@ -47,7 +47,7 @@ def llm_reranker_processor() -> CrossEncoderIOProcessor:

47 )

48

49 processor = CrossEncoderIOProcessor.__new__(CrossEncoderIOProcessor)

50 - processor.model_config = SimpleNamespace(enable_prompt_embeds=False) # type: ignore[assignment]

50 + processor.model_config = SimpleNamespace(enable_prompt_embeds=False)

51 processor.tokenizer = TokenizersBackend(

52 tokenizer_object=backend,

53 unk_token="[UNK]",

@@ -96,7 +96,7 @@ def test_token_type_ids_stay_aligned_with_a_truncated_padded_prompt():

96 )

97

98 processed_prompt = tok_params.apply_post_tokenization(

99 - tokenizer, # type: ignore[arg-type]

99 + tokenizer,

100 prompt,

101 )

102 assert "prompt_token_ids" in processed_prompt

tests/entrypoints/pooling/scoring/test_jina_ranking_io_processor_unit.py

+2 -4

@@ -77,10 +77,8 @@ def offline_processor_and_context():

77 backend.pre_tokenizer = pre_tokenizers.WhitespaceSplit()

78 proc = JinaRankingIOProcessor.__new__(JinaRankingIOProcessor)

79 proc.tokenizer = TokenizersBackend(tokenizer_object=backend)

80 - proc.model_config = SimpleNamespace( # type: ignore[assignment]

81 - max_model_len=1024, is_encoder_decoder=False

82 - )

83 - proc.renderer = SimpleNamespace( # type: ignore[assignment]

80 + proc.model_config = SimpleNamespace(max_model_len=1024, is_encoder_decoder=False)

81 + proc.renderer = SimpleNamespace(

82 default_cmpl_tok_params=TokenizeParams(max_total_tokens=1024)

83 )

84 ctx = OfflineScoringInputsContext(

tests/entrypoints/pooling/test_io_processor.py

+2 -2

@@ -53,8 +53,8 @@ def test_rejects_mismatched_lora_requests(processor: PoolingIOProcessor):

53

54

55 def test_rejects_conflicting_pooling_task(processor: PoolingIOProcessor):

56 - processor.model_config = SimpleNamespace(is_encoder_decoder=False) # type: ignore[assignment]

57 - processor.renderer = SimpleNamespace( # type: ignore[assignment] # Tokenization defaults only.

56 + processor.model_config = SimpleNamespace(is_encoder_decoder=False)

57 + processor.renderer = SimpleNamespace(

58 default_cmpl_tok_params=TokenizeParams(max_total_tokens=None)

59 )

60 ctx = OfflineEncodeInputsContext(

tests/entrypoints/serve/middleware/test_authentication_middleware.py

+1 -1

@@ -87,7 +87,7 @@ def task_routes(request, monkeypatch) -> tuple[str, list[tuple[str, list[str]]]]

87 args,

88 app,

89 supported_tasks=(task,),

90 - model_config=MockModelConfig(), # type: ignore[arg-type]

90 + model_config=MockModelConfig(),

91 )

92

93 routes = get_all_http_routes(app)

tests/entrypoints/serve/tokenize/test_tokenization.py

+2 -2

@@ -107,7 +107,7 @@ async def test_tokenize_chat(

107 conversation=conversation,

108 tokenize=False,

109 )

110 - tokens = tokenizer.encode(prompt, add_special_tokens=add_special) # type: ignore[arg-type]

110 + tokens = tokenizer.encode(prompt, add_special_tokens=add_special)

111

112 response = requests.post(

113 server.url_for("tokenize"),

@@ -176,7 +176,7 @@ async def test_tokenize_chat_with_tools(

176 tools=tools,

177 tokenize=False,

178 )

179 - tokens = tokenizer.encode(prompt, add_special_tokens=add_special) # type: ignore[arg-type]

179 + tokens = tokenizer.encode(prompt, add_special_tokens=add_special)

180

181 response = requests.post(

182 server.url_for("tokenize"),

tests/entrypoints/speech_to_text/transcription/test_qwen3_asr_sanitize_prompt.py

+1 -1

@@ -47,7 +47,7 @@ def test_sanitize_handles_falsy_inputs() -> None:

47 assert _sanitize_transcription_user_text("") == ""

48 # The dataclass default for ``response_prefix`` is the empty string;

49 # the sanitizer must accept that without exception or extra work.

50 - assert _sanitize_transcription_user_text(None) == "" # type: ignore[arg-type]

50 + assert _sanitize_transcription_user_text(None) == ""

51

52

53 def test_sanitize_is_idempotent() -> None:

tests/entrypoints/unit_tests/test_chat_utils.py

+2 -2

@@ -216,7 +216,7 @@ def video_url():

216 @pytest.fixture(scope="module")

217 def audio_url():

218 audio = AudioAsset("mary_had_lamb")

219 - return encode_audio_url(*audio.audio_and_sample_rate) # type: ignore[arg-type]

219 + return encode_audio_url(*audio.audio_and_sample_rate)

220

221

222 def _assert_mm_data_is_image_input(

@@ -3194,5 +3194,5 @@ def test_validate_chat_template_rejects_invalid_type():

3194 with pytest.raises(

3195 VLLMValidationError, match="not a valid chat template type"

3196 ) as exc_info:

3197 - validate_chat_template(123) # type: ignore[arg-type]

3197 + validate_chat_template(123)

3198 assert exc_info.value.parameter == "chat_template"

tests/jit_monitor/test_hooks.py

+1 -1

@@ -149,7 +149,7 @@ def test_activate_logs_info():

149

150 def test_activate_rejects_unknown_mode():

151 with pytest.raises(ValueError, match="Unsupported JIT monitor mode"):

152 - jit_monitor.activate(mode="panic") # type: ignore[arg-type]

152 + jit_monitor.activate(mode="panic")

153

154

155 def test_activate_without_triton():

tests/kernels/attention/test_attention_selector.py

+4 -4

@@ -23,12 +23,12 @@ from vllm.platforms.interface import DeviceCapability

23 if current_platform.is_cuda():

24 from vllm.platforms.cuda import CudaPlatform

25 else:

26 - CudaPlatform = None # type: ignore[assignment] # Unavailable platform import.

26 + CudaPlatform = None

27

28 if current_platform.is_rocm():

29 from vllm.platforms.rocm import RocmPlatform

30 else:

31 - RocmPlatform = None # type: ignore[misc, assignment] # Unavailable platform import.

31 + RocmPlatform = None # type: ignore[misc] # Unavailable platform import.

32

33 from vllm.v1.attention.backend import AttentionType

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

@@ -356,14 +356,14 @@ def test_invalid_backend():

356 def test_auto_backend_string(auto_value: str):

357 """Test that 'auto' string value triggers automatic backend selection."""

358 # Using "auto" should result in backend=None (automatic selection)

359 - attention_config = AttentionConfig(backend=auto_value) # type: ignore[arg-type]

359 + attention_config = AttentionConfig(backend=auto_value)

360 assert attention_config.backend is None

361

362

363 def test_auto_backend_selection_behavior():

364 """Test that 'auto' backend behaves same as None (automatic selection)."""

365 # Create config with explicit "auto"

366 - auto_config = AttentionConfig(backend="auto") # type: ignore[arg-type]

366 + auto_config = AttentionConfig(backend="auto")

367

368 # Create config with None (default)

369 none_config = AttentionConfig(backend=None)

tests/kernels/attention/test_cpu_attn.py

+1 -1

@@ -836,7 +836,7 @@ def varlen_encoder_zentorch_sdpa(

836 key,

837 value,

838 output,

839 - SimpleNamespace(query_start_loc=query_start_loc, causal=False), # type: ignore[arg-type]

839 + SimpleNamespace(query_start_loc=query_start_loc, causal=False),

840 scale,

841 sliding_window if sliding_window is not None else -1,

842 alibi_slopes,

tests/kernels/attention/test_minimax_m3.py

+1 -1

@@ -1782,7 +1782,7 @@ def test_aiter_indexer_requires_the_aiter_attend(monkeypatch):

1782 num_index_heads=1,

1783 index_head_dim=HEAD_DIM,

1784 # The implementation also accepts this legacy fp8 alias.

1785 - indexer_kv_dtype="fp8_e4m3", # type: ignore[arg-type]

1785 + indexer_kv_dtype="fp8_e4m3",

1786 max_model_len=8192,

1787 )

1788

tests/kernels/attention/test_minimax_m3_msa_cutlass_sparse_decode.py

+5 -5

@@ -235,14 +235,14 @@ def test_msa_metadata_builder_prepares_cutlass_for_regular_decode(

235 builder.num_q_heads = 64

236 builder.num_kv_heads = 4

237 builder.topk_blocks = TOPK

238 - builder.kv_cache_spec = SimpleNamespace(num_kv_heads=4) # type: ignore[assignment]

238 + builder.kv_cache_spec = SimpleNamespace(num_kv_heads=4)

239 builder.kv_cache_dtype = "fp8_e4m3"

240 builder.decode_backend = "cutlass"

241 - builder.msa_cutlass_plan_cache = object() # type: ignore[assignment] # Mocked plan-cache sentinel.

241 + builder.msa_cutlass_plan_cache = object()

242

243 metadata = builder.build(

244 0,

245 - SimpleNamespace( # type: ignore[arg-type]

245 + SimpleNamespace(

246 seq_lens_cpu_upper_bound=torch.full((batch,), 257, dtype=torch.int32)

247 ),

248 )

@@ -377,7 +377,7 @@ def test_query_fp8_stays_valid_when_cutlass_plan_appears_on_replay(

377 with torch.cuda.stream(stream), override_forward_context(forward_context):

378 capture = BreakableCUDAGraphCapture()

379 with capture:

380 - query_fp8 = MiniMaxM3SparseAttention._allocate_query_fp8(attention, qkv) # type: ignore[arg-type]

380 + query_fp8 = MiniMaxM3SparseAttention._allocate_query_fp8(attention, qkv)

381 if query_fp8 is not None:

382 query_fp8.zero_()

383 run_attention(query_fp8)

@@ -387,7 +387,7 @@ def test_query_fp8_stays_valid_when_cutlass_plan_appears_on_replay(

387 assert capture.num_eager_breaks == 1

388 assert observed_ptrs == []

389

390 - decode.msa_cutlass = object() # type: ignore[assignment]

390 + decode.msa_cutlass = object()

391 for _ in range(3):

392 capture.replay()

393 stream.synchronize()

tests/kernels/attention/test_rocm_aiter_mla_causal_verify_mask.py

+1 -1

@@ -204,7 +204,7 @@ def _run_verify_block():

204 ),

205 ):

206 assert isinstance(impl, AiterMLAImpl)

207 - impl.forward_mqa((q_nope, q_pe), kv_cache, metadata, layer=None) # type: ignore[arg-type] # Spy decode path does not read layer.

207 + impl.forward_mqa((q_nope, q_pe), kv_cache, metadata, layer=None)

208

209 return metadata, captured

210

tests/kernels/attention/test_rocm_aiter_mla_fp8_prefill.py

+2 -2

@@ -121,7 +121,7 @@ def _build_prefill_metadata(

121

122 builder = object.__new__(AiterMLAMetadataBuilder)

123 builder.num_heads = num_heads

124 - builder.mla_dims = SimpleNamespace(v_head_dim=V_HEAD_DIM) # type: ignore[assignment]

124 + builder.mla_dims = SimpleNamespace(v_head_dim=V_HEAD_DIM)

125 builder._init_fp8_prefill_ps_buffers(

126 max_num_reqs=len(seq_lens),

127 max_prefill_qlen=max_q,

@@ -136,7 +136,7 @@ def _build_prefill_metadata(

136 prefill = SimpleNamespace(query_start_loc=qo_indptr, max_query_len=max_q)

137 metadata = SimpleNamespace(prefill=prefill, num_decodes=0)

138 common = SimpleNamespace(query_start_loc_cpu=qo_indptr_cpu)

139 - builder._build_fp8_prefill_ps_metadata(metadata, common) # type: ignore[arg-type]

139 + builder._build_fp8_prefill_ps_metadata(metadata, common)

140 return metadata, total_q, builder

141

142

tests/kernels/attention/test_rocm_aiter_mla_sink.py

+6 -6

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

170 SimpleNamespace(_q_scale=q_scale, _k_scale=kv_scale),

171 q,

172 kv,

173 - metadata, # type: ignore[arg-type]

173 + metadata,

174 )

175 kv_flat = kv_ref[:, 0]

176 references = []

@@ -244,7 +244,7 @@ def test_sparse_mla_sink_rejects_unsupported_aiter_dtypes(

244 SimpleNamespace(_q_scale=None, _k_scale=None),

245 torch.empty(1, 16, Q_HEAD_DIM, dtype=q_dtype),

246 torch.empty(1, 1, Q_HEAD_DIM, dtype=kv_dtype),

247 - metadata, # type: ignore[arg-type]

247 + metadata,

248 )

249

250

@@ -384,7 +384,7 @@ def test_sparse_mla_sink_matches_dense_attention_with_empty_rows_and_paged_cache

384 paged_kv_indices=torch.tensor(indices, dtype=torch.int32, device="cuda"),

385 paged_kv_indptr=lengths.cumsum(0).to(device="cuda", dtype=torch.int32),

386 )

387 - impl = ROCMAiterMLASparseImpl.__new__(ROCMAiterMLASparseImpl) # type: ignore[arg-type]

387 + impl = ROCMAiterMLASparseImpl.__new__(ROCMAiterMLASparseImpl)

388 impl.num_heads = num_heads

389 impl.head_size = head_dim

390 impl.kv_lora_rank = value_dim

@@ -441,9 +441,9 @@ def test_sparse_mla_backend_resolves_only_contiguous_layer_layouts(monkeypatch,

441 ]

442 if layout == "BLNHC":

443 with pytest.raises(ValueError, match="does not satisfy"):

444 - resolve_kv_cache_layout(config, supported) # type: ignore[arg-type]

444 + resolve_kv_cache_layout(config, supported)

445 else:

446 - assert resolve_kv_cache_layout(config, supported).name == layout # type: ignore[arg-type]

446 + assert resolve_kv_cache_layout(config, supported).name == layout

447

448

449 @pytest.mark.parametrize("dtype", [torch.bfloat16, torch.float16])

@@ -494,7 +494,7 @@ def test_sparse_mla_sink_forward_mqa_preserves_split_query(dtype):

494 actual, _ = impl.forward_mqa(

495 (q[..., :V_HEAD_DIM], q[..., V_HEAD_DIM:]),

496 kv,

497 - metadata, # type: ignore[arg-type]

497 + metadata,

498 SimpleNamespace(_q_scale=None, _k_scale=None),

499 )

500 references = [

tests/kernels/attention/test_rocm_aiter_mla_sparse_metadata_sync.py

+2 -2

@@ -39,10 +39,10 @@ def _make_builder():

39 topk_tokens = 4

40

41 builder.device = torch.device("cpu")

42 - builder.kv_cache_spec = SimpleNamespace(block_size=1) # type: ignore[assignment]

42 + builder.kv_cache_spec = SimpleNamespace(block_size=1)

43 builder.model_dtype = torch.bfloat16

44 builder.kv_cache_dtype = "fp8"

45 - builder.mla_dims = SimpleNamespace(kv_lora_rank=512, qk_rope_head_dim=64) # type: ignore[assignment]

45 + builder.mla_dims = SimpleNamespace(kv_lora_rank=512, qk_rope_head_dim=64)

46 builder.topk_tokens = topk_tokens

47 builder.req_id_per_token_buffer = torch.zeros(

48 max_num_batched_tokens, dtype=torch.int32, device="cpu"

tests/kernels/attention/test_rocm_paged_mxfp4_indexer.py

+2 -4

@@ -227,9 +227,7 @@ def _decode_metadata(case, rows, ratio, query_lens):

227 num_decode_tokens=len(rows),

228 num_prefills=0,

229 num_prefill_tokens=0,

230 - decode=types.SimpleNamespace( # type: ignore[arg-type]

231 - block_table=block_table, seq_lens=lens[:, None]

232 - ),

230 + decode=types.SimpleNamespace(block_table=block_table, seq_lens=lens[:, None]),

231 decode_row_lens=lens,

232 decode_block_ends=(lens + CAND_BLOCK - 1) // CAND_BLOCK,

233 decode_native=native,

@@ -280,7 +278,7 @@ def _prefill_metadata(

278 num_decode_tokens=0,

279 num_prefills=len(case.seq_lens),

280 num_prefill_tokens=len(rows),

283 - prefill=types.SimpleNamespace(chunks=chunks), # type: ignore[arg-type]

281 + prefill=types.SimpleNamespace(chunks=chunks),

282 prefill_plans=plans,

283 gather_launches=plan_gather_launches(chunks, plans, gather_rows),

284 )

tests/kernels/attention/test_rocm_triton_attn_dsv4.py

+2 -2

@@ -955,7 +955,7 @@ def test_rocm_capture_metadata_sets_adaptive_marker(monkeypatch) -> None:

955 )

956 builder = object.__new__(rocm_mod.DeepseekV4ROCMAiterMLASparseMetadataBuilder)

957

958 - actual = builder.build_for_cudagraph_capture(SimpleNamespace()) # type: ignore[arg-type] # Patched base method ignores metadata.

958 + actual = builder.build_for_cudagraph_capture(SimpleNamespace())

959

960 assert actual is metadata

961 assert actual.for_cudagraph_capture is _on_gfx950()

@@ -1563,7 +1563,7 @@ def test_dsv4_adaptive_mla_swa_metadata_graph_replay(monkeypatch) -> None:

1563 "sliding_window": window_size,

1564 },

1565 )

1566 - vllm_config.speculative_config = SimpleNamespace( # type: ignore[assignment]

1566 + vllm_config.speculative_config = SimpleNamespace(

1567 num_speculative_tokens=upper_query_len - 1,

1568 parallel_drafting=False,

1569 enable_adaptive_verification=True,

tests/kernels/mamba/test_gdn_forward_core_split.py

+1 -1

@@ -176,7 +176,7 @@ def test_forward_core_split_matches_unified(

176 builder = GDNAttentionMetadataBuilder(

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]

179 + kv_cache_spec=MambaSpec(

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

181 ),

182 layer_names=[PREFIX],

tests/kernels/moe/test_cutlass_moe.py

+1 -1

@@ -538,7 +538,7 @@ def run_8_bit(

538 assert num_local_experts is not None

539 return run_with_expert_maps(

540 num_experts,

541 - num_local_experts, # type: ignore[arg-type]

541 + num_local_experts,

542 quant_config,

543 **kwargs,

544 )

tests/kernels/moe/test_ocp_mx_moe.py

+1 -1

@@ -1724,7 +1724,7 @@ def test_rocm_mxfp4_moe_oracle(

1724 w13_conv, w2_conv, w13_scale_conv, w2_scale_conv, w13_bias_conv, w2_bias_conv = (

1725 convert_gpt_oss_weight_to_mxfp4_moe_kernel_format(

1726 mxfp4_backend=backend,

1727 - layer=layer, # type: ignore[arg-type]

1727 + layer=layer,

1728 w13_weight=w13_quant,

1729 w2_weight=w2_quant,

1730 w13_weight_scale=w13_scale,

tests/kernels/quantization/test_block_fp8.py

+1 -1

@@ -259,7 +259,7 @@ def test_w8a8_block_fp8_torch_scaled_mm_matmul():

259 stub = BlockWiseTorchFP8ScaledMMLinearKernel.__new__(

260 BlockWiseTorchFP8ScaledMMLinearKernel

261 )

262 - stub.config = types.SimpleNamespace(out_dtype=out_dtype) # type: ignore[assignment]

262 + stub.config = types.SimpleNamespace(out_dtype=out_dtype)

263 out = stub.apply_block_scaled_mm(

264 A_fp8_cuda.cuda(), B_fp8.cuda(), As_cuda.cuda(), Bs.cuda()

265 )

tests/kernels/quantization/test_cutlass_w4a8.py

+1 -1

@@ -208,7 +208,7 @@ def mm_test_helper(

208 a=tensors.a,

209 b_q=tensors.w_q,

210 b_group_scales=tensors.w_g_s,

211 - b_group_size=group_size, # type: ignore[arg-type]

211 + b_group_size=group_size,

212 b_channel_scales=tensors.w_ch_s,

213 a_token_scales=tensors.w_tok_s,

214 )

tests/kernels/quantization/test_marlin_tile_padding.py

+2 -4

@@ -525,15 +525,13 @@ def test_check_marlin_supports_layer_allow_tile_padding():

525 )

526

527 # Tile-misaligned but group-aligned: rejected strictly, allowed w/ padding

528 - layer: LinearBase = _FakeLinear( # type: ignore[assignment]

529 - 4640, 512, input_size=2048

530 - )

528 + layer: LinearBase = _FakeLinear(4640, 512, input_size=2048)

529 assert not check_marlin_supports_layer(layer, 128)

530 assert check_marlin_supports_layer(layer, 128, allow_tile_padding=True)

531 assert check_marlin_supports_layer(layer, -1, allow_tile_padding=True)

532

533 # A group straddling the TP shard cannot be fixed by padding

536 - layer = _FakeLinear(4608, 4672, input_size=18688) # type: ignore[assignment]

534 + layer = _FakeLinear(4608, 4672, input_size=18688)

535 assert not check_marlin_supports_layer(layer, 128, allow_tile_padding=True)

536

537

tests/kernels/quantization/test_nvfp4_emulation.py

+1 -1

@@ -206,7 +206,7 @@ def test_nvfp4_emulation_support_check_rejects_bias_and_lora(

206 in_dtype=torch.bfloat16,

207 device="cuda",

208 routing_method=RoutingMethodType.TopK,

209 - **config_kwargs, # type: ignore[arg-type]

209 + **config_kwargs,

210 )

211

212 supported, reason = Nvfp4QuantizationEmulationTritonExperts.is_supported_config(

tests/model_executor/layers/test_fused_shared_expert.py

+3 -3

@@ -413,7 +413,7 @@ def test_deepseek_v4_shared_expert_fse_uses_mtp_quantization_config_prefix(

413 )

414

415 compatible, reason = is_shared_expert_quant_fse_compatible(

416 - DeepseekV4Config(), # type: ignore[arg-type]

416 + DeepseekV4Config(),

417 "model.layers.2.ffn.experts",

418 "model.layers.2.ffn.shared_experts",

419 )

@@ -824,7 +824,7 @@ def test_models_fse_init(

824 model_type, quantization_config

825 )

826 vllm_config = VllmConfig()

827 - vllm_config.model_config = SimpleNamespace( # type: ignore[assignment]

827 + vllm_config.model_config = SimpleNamespace(

828 hf_config=config,

829 hf_text_config=config,

830 dtype=torch.bfloat16,

@@ -893,7 +893,7 @@ def test_models_fse_init(

893 DeepSeekV4MTP,

894 )

895

896 - vllm_config.speculative_config = SimpleNamespace( # type: ignore[assignment]

896 + vllm_config.speculative_config = SimpleNamespace(

897 draft_model_config=SimpleNamespace(hf_config=config),

898 method="mtp",

899 parallel_drafting=False,

tests/model_executor/layers/test_mla_short_prefill_indexer.py

+4 -4

@@ -46,7 +46,7 @@ def test_sparse_attention_refreshes_batch_state_inside_eager_segment(

46

47 with pytest.raises(BatchStateRefreshed):

48 DeepseekV32Attention._sparse_indexer_and_attn(

49 - layer, # type: ignore[arg-type]

49 + layer,

50 torch.empty(1, dtype=torch.long),

51 torch.empty(1, 1),

52 torch.empty(1, 1, 1),

@@ -80,7 +80,7 @@ def make_indexer_metadata(

80 num_decode_tokens=num_decode_tokens,

81 num_prefills=num_prefills,

82 num_prefill_tokens=num_prefill_tokens,

83 - prefill=SimpleNamespace(chunks=[]) if num_prefills else None, # type: ignore[arg-type]

83 + prefill=SimpleNamespace(chunks=[]) if num_prefills else None,

84 )

85

86

@@ -129,7 +129,7 @@ def test_short_prefill_updates_k_cache_before_scoring_decision(

129 slot_mapping=slot_mapping,

130 )

131 if indexer_metadata.num_decodes:

132 - indexer_metadata.decode = object() # type: ignore[assignment]

132 + indexer_metadata.decode = object()

133 mla_metadata = make_mla_metadata(

134 use_dense_mha=batch_kind != "force_mqa",

135 num_decode_tokens=mla_num_decode_tokens,

@@ -307,7 +307,7 @@ def test_deepseek_v32_dispatches_selected_mha(

307 output = torch.empty(2, 2)

308

309 DeepseekV32Attention._sparse_indexer_and_attn(

310 - layer, # type: ignore[arg-type]

310 + layer,

311 torch.arange(2),

312 torch.empty(2, 2),

313 q_nope,

tests/model_executor/layers/test_pooler_activations.py

+2 -2

@@ -224,11 +224,11 @@ class TestResolveClassifierActFn:

224 model_config = SimpleNamespace(

225 hf_config=SimpleNamespace(num_labels=3, problem_type="")

226 )

227 - result = resolve_classifier_act_fn(model_config, act_fn=None) # type: ignore[arg-type]

227 + result = resolve_classifier_act_fn(model_config, act_fn=None)

228 assert isinstance(result, PoolerClassify)

229 assert result.num_labels == 3

230

231 def test_passes_through_provided_act_fn(self):

232 custom = PoolerIdentity()

233 - result = resolve_classifier_act_fn(None, act_fn=custom) # type: ignore[arg-type]

233 + result = resolve_classifier_act_fn(None, act_fn=custom)

234 assert result is custom

tests/model_executor/layers/test_pooler_methods.py

+1 -1

@@ -320,7 +320,7 @@ def test_dispatch_seq_cls_honors_token_pooling_type(tok_pooling_type):

320 ),

321 )

322 classifier = torch.nn.Linear(4, 3)

323 - with set_current_vllm_config(config): # type: ignore[arg-type]

323 + with set_current_vllm_config(config):

324 pooler = DispatchPooler.for_seq_cls(pooler_config, classifier=classifier)

325

326 token_ids = [[2, 99, 3, 99], [4, 99], [5, 6]]

tests/model_executor/model_loader/tensorizer_loader/test_tensorizer.py

+1 -1

@@ -36,7 +36,7 @@ try:

36 import tensorizer

37 from tensorizer import EncryptionParams

38 except ImportError:

39 - tensorizer = PlaceholderModule("tensorizer") # type: ignore[assignment]

39 + tensorizer = PlaceholderModule("tensorizer")

40 EncryptionParams = tensorizer.placeholder_attr("EncryptionParams")

41

42

tests/model_executor/model_loader/test_weight_cache.py

+1 -1

@@ -368,4 +368,4 @@ def test_ipc_loader_copy_mode_reports_no_external_weight_memory():

368 load_format="ipc_cache", model_loader_extra_config={"mode": "copy"}

369 )

370 loader = IpcModelLoader(copy_mode)

371 - assert loader.get_external_weight_memory(None) == 0 # type: ignore[arg-type]

371 + assert loader.get_external_weight_memory(None) == 0

tests/models/language/generation/test_mistral.py

+3 -3

@@ -259,7 +259,7 @@ def test_mistral_function_calling(vllm_runner, model: str, dtype: str) -> None:

259

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

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

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

262 + parsed_message = tool_parser.extract_tool_calls(model_output, None)

263

264 assert parsed_message.tools_called

265

@@ -302,7 +302,7 @@ def test_mistral_function_call_nested_json():

302

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

304

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

305 + parsed = parser.extract_tool_calls(model_output, None)

306

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

308 # without truncation.

@@ -335,7 +335,7 @@ def test_mistral_function_call_nested_json():

335 ]

336 )

337

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

338 + parsed = parser.extract_tool_calls(model_output, None)

339

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

341 # without truncation.

tests/models/multimodal/generation/vlm_utils/model_utils.py

+1 -1

@@ -1193,7 +1193,7 @@ def voxtral_patch_hf_runner(hf_model: "HfRunner") -> "HfRunner":

1193 output.sequences = output.sequences[:, prompt_len:]

1194 return output

1195

1196 - hf_model.get_inputs = patched_get_inputs # type: ignore[method-assign, assignment]

1196 + hf_model.get_inputs = patched_get_inputs # type: ignore[method-assign]

1197 hf_model.model.generate = patched_generate # type: ignore[method-assign]

1198 return hf_model

1199

tests/models/multimodal/pooling/test_clip.py

+1 -1

@@ -132,7 +132,7 @@ def test_models(

132 _run_test(

133 hf_runner,

134 vllm_runner,

135 - input_cases, # type: ignore[arg-type]

135 + input_cases,

136 model,

137 dtype=dtype,

138 )

tests/models/multimodal/pooling/test_llama_nemotron_vl.py

+1 -1

@@ -190,7 +190,7 @@ def test_models(

190 _run_test(

191 hf_runner,

192 vllm_runner,

193 - input_cases, # type: ignore[arg-type]

193 + input_cases,

194 model,

195 dtype=dtype,

196 )

tests/models/multimodal/pooling/test_siglip.py

+1 -1

@@ -149,7 +149,7 @@ def test_models(

149 _run_test(

150 hf_runner,

151 vllm_runner,

152 - input_cases, # type: ignore[arg-type]

152 + input_cases,

153 model,

154 dtype=dtype,

155 )

tests/models/multimodal/processing/test_qwen2_vl.py

+2 -2

@@ -106,7 +106,7 @@ def test_merge_and_resolve_mm_processor_kwargs_preserves_request_alias_precedenc

106 mm_device_do_normalize=False,

107 )

108 model_config = SimpleNamespace(get_multimodal_config=lambda: mm_config)

109 - ctx = InputProcessingContext(model_config, tokenizer=None) # type: ignore[arg-type]

109 + ctx = InputProcessingContext(model_config, tokenizer=None)

110 info = Qwen2VLProcessingInfo(ctx)

111 monkeypatch.setattr(

112 info,

@@ -236,7 +236,7 @@ def test_get_vision_info_reads_merged_modality_scope(

236 do_resize=False,

237 image_processor=image_processor,

238 mm_kwargs={},

239 - modality=modality, # type: ignore[arg-type]

239 + modality=modality,

240 )

241

242 assert seen == [merged[scope]]

tests/models/transformers/test_backend.py

+2 -2

@@ -260,7 +260,7 @@ def test_quantization(

260 model,

261 model_impl="auto",

262 enforce_eager=True,

263 - **quantization_kwargs, # type: ignore[arg-type]

263 + **quantization_kwargs,

264 ) as vllm_model:

265 vllm_outputs = vllm_model.generate_greedy_logprobs(

266 example_prompts, max_tokens=max_tokens, num_logprobs=num_logprobs

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

270 model,

271 model_impl="transformers",

272 enforce_eager=True,

273 - **quantization_kwargs, # type: ignore[arg-type]

273 + **quantization_kwargs,

274 ) as vllm_model:

275 model_config = vllm_model.llm.llm_engine.model_config

276 assert model_config.using_transformers_backend()

tests/multimodal/media/test_video.py

+1 -1

@@ -385,7 +385,7 @@ def test_pynvvideocodec_unrelated_error_propagates(

385 PyNvVideoCodecVideoBackendMixin.decode_frames_pynvvideocodec(

386 None,

387 b"video",

388 - None, # type: ignore[arg-type] # Fails before target metadata is used.

388 + None,

389 )

390

391 assert exc_info.value is original_error

tests/multimodal/test_audio.py

+1 -1

@@ -179,7 +179,7 @@ def test_audio_resampler_invalid_method(dummy_audio):

179 # Validated eagerly so a bad method fails at construction, not on the

180 # first audio request.

181 with pytest.raises(ValueError, match="Invalid resampling method"):

182 - AudioResampler(target_sr=22050, method="invalid") # type: ignore[arg-type]

182 + AudioResampler(target_sr=22050, method="invalid")

183

184

185 def test_audio_resampler_no_target_sr(dummy_audio):

tests/multimodal/test_cache.py

+12 -12

@@ -276,14 +276,14 @@ def test_oversized_item_is_served_uncached():

276 item = MultiModalKwargsItem.dummy(nbytes=4096)

277 small = MultiModalKwargsItem.dummy(nbytes=64)

278

279 - p0_only = MultiModalProcessorOnlyCache(model_config) # type: ignore[arg-type]

279 + p0_only = MultiModalProcessorOnlyCache(model_config)

280 assert p0_only.get_and_update_item((item, []), "big")[0] is item

281 assert not p0_only.is_cached_item("big")

282 assert p0_only.get_and_update_item((small, []), "small")[0] is small

283 assert p0_only.is_cached_item("small")

284

285 - p0 = LruKeyReplicatedSenderCache(model_config) # type: ignore[arg-type]

286 - p1 = LruKeyReplicatedReceiverCache(model_config) # type: ignore[arg-type]

285 + p0 = LruKeyReplicatedSenderCache(model_config)

286 + p1 = LruKeyReplicatedReceiverCache(model_config)

287 assert item is not None

288 assert p0.get_and_update_item((item, []), "big")[0] is item

289 assert not p0.is_cached_item("big")

@@ -304,8 +304,8 @@ def test_mm_cache_miss_raises_and_recovers():

304 model or network.

305 """

306 model_config = _StubModelConfig(mm_processor_cache_gb=1)

307 - p0 = LruKeyReplicatedSenderCache(model_config) # type: ignore[arg-type]

308 - p1 = LruKeyReplicatedReceiverCache(model_config) # type: ignore[arg-type]

307 + p0 = LruKeyReplicatedSenderCache(model_config)

308 + p1 = LruKeyReplicatedReceiverCache(model_config)

309

310 mm_hash = "image_A"

311 item = MultiModalKwargsItem.dummy(nbytes=64)

@@ -348,7 +348,7 @@ def test_receiver_cache_replaces_stale_item_when_payload_resent():

348 and kills EngineCore. Prefer the fresh payload and keep it for later hits.

349 """

350 model_config = _StubModelConfig(mm_processor_cache_gb=1)

351 - p1 = LruKeyReplicatedReceiverCache(model_config) # type: ignore[arg-type]

351 + p1 = LruKeyReplicatedReceiverCache(model_config)

352 small = MultiModalKwargsItem.dummy(nbytes=64)

353 large = MultiModalKwargsItem.dummy(nbytes=256)

354 mm_hash = "shared-id"

@@ -361,7 +361,7 @@ def test_receiver_cache_replaces_stale_item_when_payload_resent():

361 def test_receiver_cache_features_keep_resent_payload():

362 """EngineCore updates features through get_and_update_features."""

363 model_config = _StubModelConfig(mm_processor_cache_gb=1)

364 - p1 = LruKeyReplicatedReceiverCache(model_config) # type: ignore[arg-type]

364 + p1 = LruKeyReplicatedReceiverCache(model_config)

365 small = MultiModalKwargsItem.dummy(nbytes=64)

366 large = MultiModalKwargsItem.dummy(nbytes=256)

367 mm_hash = "shared-id"

@@ -394,7 +394,7 @@ def test_mm_cache_miss_batches_all_drifted_hashes():

394 non-drifted items in the same request are still ingested.

395 """

396 model_config = _StubModelConfig(mm_processor_cache_gb=1)

397 - p1 = LruKeyReplicatedReceiverCache(model_config) # type: ignore[arg-type]

397 + p1 = LruKeyReplicatedReceiverCache(model_config)

398 item = MultiModalKwargsItem.dummy(nbytes=64)

399

400 def _feature(

@@ -442,8 +442,8 @@ def test_shm_receiver_handles_prefix_covered_items(monkeypatch):

442 parallel_config=SimpleNamespace(world_size=1),

443 model_config=_StubModelConfig(mm_processor_cache_gb=4 * MiB_bytes / GiB_bytes),

444 )

445 - p0 = ShmObjectStoreSenderCache(vllm_config) # type: ignore[arg-type]

446 - p1 = ShmObjectStoreReceiverCache(vllm_config, mp.Lock()) # type: ignore[arg-type]

445 + p0 = ShmObjectStoreSenderCache(vllm_config)

446 + p1 = ShmObjectStoreReceiverCache(vllm_config, mp.Lock())

447

448 def _feature(

449 mm_hash: str, data: MultiModalKwargsItem | None

@@ -1082,9 +1082,9 @@ async def test_release_kv_cache_resends_mm_payload(use_async, release_error):

1082

1083 async def call_release():

1084 if use_async:

1085 - await AsyncLLM.release_kv_cache_memory(engine) # type: ignore[arg-type]

1085 + await AsyncLLM.release_kv_cache_memory(engine)

1086 else:

1087 - LLMEngine.release_kv_cache_memory(engine) # type: ignore[arg-type]

1087 + LLMEngine.release_kv_cache_memory(engine)

1088

1089 if release_error:

1090 with pytest.raises(RuntimeError, match=release_error):

tests/multimodal/test_processing.py

+4 -4

@@ -1281,7 +1281,7 @@ def test_get_supported_mm_processor_kwargs_uses_supported_modalities(

1281 processor: SimpleNamespace,

1282 expected: dict[str, set[str]],

1283 ) -> None:

1284 - info = BaseProcessingInfo(SimpleNamespace()) # type: ignore[arg-type]

1284 + info = BaseProcessingInfo(SimpleNamespace())

1285 info.__dict__["supported_mm_limits"] = supported_mm_limits

1286 monkeypatch.setattr(info, "get_hf_processor", lambda **_: processor)

1287

@@ -1291,7 +1291,7 @@ def test_get_supported_mm_processor_kwargs_uses_supported_modalities(

1291 def test_supported_mm_processor_kwargs_is_cached(

1292 monkeypatch: pytest.MonkeyPatch,

1293 ) -> None:

1294 - info = BaseProcessingInfo(SimpleNamespace()) # type: ignore[arg-type]

1294 + info = BaseProcessingInfo(SimpleNamespace())

1295 expected = {"images_kwargs": {"size"}}

1296 calls = 0

1297

@@ -1442,7 +1442,7 @@ def test_get_merged_mm_kwargs_treats_empty_scopes_as_absent_before_routing(

1442 mm_device_do_normalize=False,

1443 )

1444 model_config = SimpleNamespace(get_multimodal_config=lambda: mm_config)

1445 - ctx = InputProcessingContext(model_config, tokenizer=None) # type: ignore[arg-type]

1445 + ctx = InputProcessingContext(model_config, tokenizer=None)

1446

1447 merged = ctx.get_merged_mm_kwargs(

1448 inference_kwargs,

@@ -1482,7 +1482,7 @@ def test_get_merged_mm_kwargs_merges_before_routing():

1482 mm_device_do_normalize=False,

1483 )

1484 model_config = SimpleNamespace(get_multimodal_config=lambda: mm_config)

1485 - ctx = InputProcessingContext(model_config, tokenizer=None) # type: ignore[arg-type]

1485 + ctx = InputProcessingContext(model_config, tokenizer=None)

1486

1487 merged = ctx.get_merged_mm_kwargs(

1488 {

tests/multimodal/test_registry.py

+1 -1

@@ -24,4 +24,4 @@ def test_create_processor_error_uses_served_model_name():

24 ValueError,

25 match="friendly-model-name is not a multimodal model",

26 ):

27 - MULTIMODAL_REGISTRY.create_processor(model_config) # type: ignore[arg-type]

27 + MULTIMODAL_REGISTRY.create_processor(model_config)

tests/multimodal/test_video.py

+1 -1

@@ -493,7 +493,7 @@ def test_pynvvideocodec_failed_rebuild_invalidates_decoder_slot():

493

494 old_decoder = FakeDecoder()

495 slot = PyNvVideoCodecDecoderSlot(FakeStream())

496 - slot.decoder = old_decoder # type: ignore[assignment] # Fake retained GPU decoder.

496 + slot.decoder = old_decoder

497 slot.source_path = "valid.mp4"

498

499 class FakeNvc:

tests/parser/cohere/test_structural_tags.py

+1 -1

@@ -347,7 +347,7 @@ class TestAdjustRequestFoldFromStructuredOutputs:

347 id="chat_completion_request",

348 ),

349 pytest.param(

350 - lambda: StructuredOutputsParams(json=[1, 2, 3]), # type: ignore[arg-type]

350 + lambda: StructuredOutputsParams(json=[1, 2, 3]),

351 id="structured_outputs_params",

352 ),

353 ],

tests/plugins_tests/bitsandbytes/test_bitsandbytes.py

+1 -1

@@ -271,7 +271,7 @@ def test_bitsandbytes_passes_revision_by_name():

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

272 ),

273 ):

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

274 + bnb.BitsAndBytesModelLoader._prepare_weights(fake_self, "org/model", "myrev")

275

276 mock_idx.assert_called_once()

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

tests/plugins_tests/bitsandbytes/test_transformers.py

+2 -2

@@ -43,7 +43,7 @@ def test_transformers_backend_quantization(

43 model,

44 model_impl="auto",

45 enforce_eager=True,

46 - **quantization_kwargs, # type: ignore[arg-type]

46 + **quantization_kwargs,

47 ) as vllm_model:

48 vllm_outputs = vllm_model.generate_greedy_logprobs(

49 example_prompts, max_tokens=max_tokens, num_logprobs=num_logprobs

@@ -53,7 +53,7 @@ def test_transformers_backend_quantization(

53 model,

54 model_impl="transformers",

55 enforce_eager=True,

56 - **quantization_kwargs, # type: ignore[arg-type]

56 + **quantization_kwargs,

57 ) as vllm_model:

58 model_config = vllm_model.llm.llm_engine.model_config

59 assert model_config.using_transformers_backend()

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] # type: ignore[arg-type]

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

96 size_factors = [0.25, 0.5, 1.0]

97 inputs_per_image = [

98 (

tests/plugins_tests/test_endpoint_plugins.py

+2 -2

@@ -172,7 +172,7 @@ async def test_init_state_is_noop_without_phase_a(monkeypatch: pytest.MonkeyPatc

172 monkeypatch.setenv("VLLM_PLUGINS", "dummy_admin_endpoint_plugin")

173

174 state = State()

175 - await init_endpoint_plugins_state(_FakeEngineClient(), state, _build_args()) # type: ignore[arg-type]

175 + await init_endpoint_plugins_state(_FakeEngineClient(), state, _build_args())

176

177 assert not hasattr(state, "dummy_engine_client")

178

@@ -223,7 +223,7 @@ async def test_endpoint_plugin_end_to_end(monkeypatch: pytest.MonkeyPatch):

223 )

224

225 fake_engine_client = _FakeEngineClient(rpc_result=["cfg-a", "cfg-b"])

226 - await init_endpoint_plugins_state(fake_engine_client, app.state, args) # type: ignore[arg-type]

226 + await init_endpoint_plugins_state(fake_engine_client, app.state, args)

227

228 assert app.state.dummy_engine_client is fake_engine_client

229

tests/plugins_tests/test_oot_registration_offline.py

+1 -1

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

92 )

93

94 first_token = llm.get_tokenizer().decode(0)

95 - outputs = llm.generate(prompts, sampling_params) # type: ignore[arg-type]

95 + outputs = llm.generate(prompts, sampling_params)

96

97 for output in outputs:

98 generated_text = output.outputs[0].text

tests/plugins_tests/test_stats_logger_plugins.py

+1 -1

@@ -24,7 +24,7 @@ def test_stat_logger_plugin_is_discovered(monkeypatch: pytest.MonkeyPatch):

24 vllm_config = VllmConfig()

25 instance = factories[0](vllm_config)

26 # The `is` check above narrows DummyStatLogger to a Callable

27 - assert isinstance(instance, DummyStatLogger) # type: ignore[arg-type]

27 + assert isinstance(instance, DummyStatLogger)

28

29

30 def test_no_plugins_loaded_if_env_empty(monkeypatch: pytest.MonkeyPatch):

tests/quantization/test_auto_gptq.py

+2 -2

@@ -123,8 +123,8 @@ def test_auto_gptq_moe_creates_zero_initialized_expert_biases():

123 method = object.__new__(AutoGPTQMoEMethod)

124 method.quant_config = AutoGPTQConfig(4, 128, False, True, False, {}, {})

125 method.input_dtype = None

126 - method.experts_cls = None # type: ignore[assignment] # Weight creation does not select experts.

127 - method.moe = SimpleNamespace(w13_num_shards=2) # type: ignore[assignment]

126 + method.experts_cls = None

127 + method.moe = SimpleNamespace(w13_num_shards=2)

128 layer = torch.nn.Module()

129

130 method.create_weights(

tests/quantization/test_auto_round.py

+6 -6

@@ -423,7 +423,7 @@ def test_inc_get_quant_method_unquantized_routed_experts_with_model_prefix(

423

424 config = make_config(extra_config={"model.moe_layer": {"bits": 16}})

425 layer = object.__new__(RoutedExperts)

426 - layer.moe_config = None # type: ignore[assignment] # No MoE config used by this stub.

426 + layer.moe_config = None

427

428 method = config.get_quant_method(layer, "moe_layer")

429

@@ -730,7 +730,7 @@ def test_wna16_cuda_low_bit_moe_routes_to_humming(monkeypatch, bits) -> None:

730 )

731

732 layer = object.__new__(RoutedExperts)

733 - layer.moe_config = DummyMoeConfig() # type: ignore[assignment]

733 + layer.moe_config = DummyMoeConfig()

734 layer_config = make_layer_config(bits=bits)

735 method = INCWna16Scheme().get_moe_method(

736 make_config(), layer, "model.layers.0.mlp", layer_config

@@ -861,7 +861,7 @@ def test_qwen3_30b_a3b_w4a16_autoround_routes_to_gptq_moe(

861

862 config = make_qwen3_autoround_config("qwen3_30b_a3b_w4a16")

863 layer = object.__new__(RoutedExperts)

864 - layer.moe_config = DummyMoeConfig() # type: ignore[assignment]

864 + layer.moe_config = DummyMoeConfig()

865

866 method = config.get_quant_method(layer, "model.layers.0.mlp")

867

@@ -890,7 +890,7 @@ def test_qwen3_30b_a3b_mxfp4_autoround_routes_to_mxfp4_moe(

890

891 config = make_qwen3_autoround_config("qwen3_30b_a3b_mxfp4")

892 layer = object.__new__(RoutedExperts)

893 - layer.moe_config = DummyMoeConfig() # type: ignore[assignment]

893 + layer.moe_config = DummyMoeConfig()

894

895 ignored_method = config.get_quant_method(

896 object.__new__(LinearBase), "model.layers.0.self_attn.q_proj"

@@ -1186,7 +1186,7 @@ def test_wna16_xpu_moe_routes_to_gptq_moe(monkeypatch) -> None:

1186 )

1187

1188 layer = object.__new__(RoutedExperts)

1189 - layer.moe_config = DummyMoeConfig() # type: ignore[assignment]

1189 + layer.moe_config = DummyMoeConfig()

1190 method = INCWna16Scheme().get_moe_method(

1191 make_config(),

1192 layer,

@@ -1669,7 +1669,7 @@ def test_inc_get_quant_method_unquantized_moe_returns_unquantized(

1669 when extra_config has bits >= 16."""

1670 config = make_config(extra_config={"layer": {"bits": 16}})

1671 layer = object.__new__(RoutedExperts)

1672 - layer.moe_config = None # type: ignore[assignment] # No MoE config used by this stub.

1672 + layer.moe_config = None

1673

1674 class DummyUnquantizedFusedMoEMethod:

1675 def __init__(self, moe_config) -> None:

tests/quantization/test_fp8.py

+2 -2

@@ -220,7 +220,7 @@ def test_deepseek_v41_vl_mapper_routes_linear_scales(

220 quant_config=SimpleNamespace(weight_block_size=weight_block_size)

221 )

222 resolved = model_module._linear_scale_param_name(

223 - vllm_config, # type: ignore[arg-type]

223 + vllm_config,

224 expert_dtype,

225 )

226 assert resolved == scale_name

@@ -798,7 +798,7 @@ def test_kv_cache_scale_sync_to_host_copies():

798 set_default_quant_scales(layer, register_buffer=True)

799 layer.kv_cache_dtype = "fp8"

800

801 - method = BaseKVCacheMethod(quant_config=None) # type: ignore[arg-type] # Scale synchronization does not use quant_config.

801 + method = BaseKVCacheMethod(quant_config=None)

802 method.create_weights(layer)

803 # 0.3 stays != 1.0 even after the fp8_fnuz x2 rescale.

804 checkpoint_scale = torch.tensor(0.3, dtype=torch.float32)

tests/quantization/test_moe_wna16.py

+4 -4

@@ -218,9 +218,9 @@ def test_gptq_nfirst_weights_pass_through_for_triton():

218

219 def test_moe_wna16_setup_forwards_selected_backend(monkeypatch):

220 method = object.__new__(MoeWNA16Method)

221 - method.experts_cls = object # type: ignore[assignment] # Opaque mocked-expert sentinel.

221 + method.experts_cls = object

222 method.wna16_backend = WNA16MoEBackend.HUMMING

223 - method.moe = object() # type: ignore[assignment] # Opaque mocked-config sentinel.

223 + method.moe = object()

224 quant_config = object()

225 monkeypatch.setattr(

226 method, "get_fused_moe_quant_config", lambda layer: quant_config

@@ -235,7 +235,7 @@ def test_moe_wna16_setup_forwards_selected_backend(monkeypatch):

235

236 monkeypatch.setattr(moe_wna16, "make_wna16_moe_kernel", fake_make_wna16_moe_kernel)

237

238 - method._setup_kernel(layer) # type: ignore[arg-type] # Routing-table stub.

238 + method._setup_kernel(layer)

239

240 assert method.moe_kernel is kernel

241 assert captured["backend"] == WNA16MoEBackend.HUMMING

@@ -280,7 +280,7 @@ def test_moe_wna16_uses_humming_quant_config(monkeypatch):

280 ),

281 )

282

283 - assert method.get_fused_moe_quant_config(layer) is quant_config # type: ignore[arg-type] # Identity sentinel.

283 + assert method.get_fused_moe_quant_config(layer) is quant_config

284

285

286 @pytest.mark.skipif(

tests/quantization/test_online.py

+4 -4

@@ -210,9 +210,9 @@ def test_online_nvfp4_reuses_kernel_when_weights_are_reprocessed(

210 monkeypatch,

211 ) -> None:

212 method = object.__new__(Nvfp4OnlineMoEMethod)

213 - method.moe = SimpleNamespace(is_act_and_mul=True) # type: ignore[assignment]

214 - method.nvfp4_backend = object() # type: ignore[assignment] # Opaque mocked-backend sentinel.

215 - method.experts_cls = object # type: ignore[assignment] # Opaque mocked-expert sentinel.

213 + method.moe = SimpleNamespace(is_act_and_mul=True)

214 + method.nvfp4_backend = object()

215 + method.experts_cls = object

216 method.moe_quant_config = None

217 method.moe_kernel = None

218

@@ -1432,7 +1432,7 @@ def test_online_int8_moe_w2_scale_matches_unsharded(monkeypatch) -> None:

1432 layer.w2_weight = torch.nn.Parameter(w2_in, requires_grad=False)

1433 layer.num_experts = layer.local_num_experts = w13.shape[0]

1434 method = SimpleNamespace(moe=SimpleNamespace(tp_size=moe_tp_size))

1435 - Int8OnlineMoEMethod._quantize_weights(method, layer) # type: ignore[arg-type]

1435 + Int8OnlineMoEMethod._quantize_weights(method, layer)

1436 return layer.w2_weight, layer.w2_scale

1437

1438 full_weight, full_scale = quantize(w2, 1)

tests/quantization/test_quark.py

+2 -2

@@ -2314,7 +2314,7 @@ class TestQuarkInt4Format:

2314 None,

2315 quant_config.quant_config["global_quant_config"]["weight"],

2316 quant_config.pack_method,

2317 - moe_config, # type: ignore[arg-type]

2317 + moe_config,

2318 )

2319

2320 layer = _FakeLayer(moe_config)

@@ -2326,7 +2326,7 @@ class TestQuarkInt4Format:

2326 requires_grad=False,

2327 )

2328 loader = method.get_weight_loader(

2329 - layer, # type: ignore[arg-type]

2329 + layer,

2330 weight_loader=None,

2331 )

2332 loader(

tests/quantization/test_trtllm_nvfp4_hidden_dim_padding.py

+1 -1

@@ -34,7 +34,7 @@ def test_shared_nvfp4_input_scales_have_writable_storage(monkeypatch):

34

35 outputs = prepare_nvfp4_moe_layer_for_fi_or_cutlass(

36 backend=NvFp4MoeBackend.FLASHINFER_CUTLASS,

37 - layer=layer, # type: ignore[arg-type]

37 + layer=layer,

38 w13=w13,

39 w13_scale=w13_scale,

40 w13_scale_2=weight_scale,

tests/quantization/test_turboquant.py

+1 -1

@@ -300,7 +300,7 @@ class TestTurboQuantKVCacheSpec:

300

301 # The layer builds an unpacked spec; the worker's spec-collection

302 # loop applies TQ slot packing via the backend's customize_spec hook.

303 - spec = Attention.get_kv_cache_spec(layer, vllm_config) # type: ignore[arg-type]

303 + spec = Attention.get_kv_cache_spec(layer, vllm_config)

304 assert isinstance(spec, FullAttentionSpec)

305 assert spec.kv_quant_mode.is_turboquant

306 assert spec.state_content_bytes is None

tests/reasoning/test_base_thinking_reasoning_parser.py

+1 -1

@@ -68,7 +68,7 @@ class TestBaseThinkingReasoningParserInit:

68 def test_initialization_with_missing_tokenizer(self):

69 """Test that initialization fails without tokenizer."""

70 with pytest.raises(ValueError, match="model tokenizer must be passed"):

71 - TestThinkingReasoningParser(None) # type: ignore[arg-type]

71 + TestThinkingReasoningParser(None)

72

73 def test_initialization_with_missing_tokens(self, test_tokenizer):

74 """Test that initialization fails when tokens are not in vocabulary."""

tests/renderers/test_cohere.py

+50 -50

@@ -81,7 +81,7 @@ class TestRoleToMelody:

81 # reject non-strings — any exception type is acceptable as long

82 # as we don't silently produce a malformed prompt.

83 with pytest.raises((AttributeError, TypeError, ValueError)):

84 - _role_to_melody(None) # type: ignore[arg-type]

84 + _role_to_melody(None)

85

86

87 # ======================================================================

@@ -135,7 +135,7 @@ class TestNormalizeToolCall:

135

136 def test_invalid_type_rejected(self):

137 with pytest.raises(TypeError, match="Unexpected tool_call value"):

138 - _normalize_tool_call(42) # type: ignore[arg-type]

138 + _normalize_tool_call(42)

139

140

141 # ======================================================================

@@ -249,7 +249,7 @@ class TestDocumentToMelody:

249

250 def test_invalid_type_rejected(self):

251 with pytest.raises(TypeError, match="Unsupported document type"):

252 - _document_to_melody(42) # type: ignore[arg-type]

252 + _document_to_melody(42)

253

254

255 # ======================================================================

@@ -302,7 +302,7 @@ class TestToolToMelody:

302

303 def test_invalid_type_rejected(self):

304 with pytest.raises(TypeError, match="Unsupported tool type"):

305 - _tool_to_melody(42) # type: ignore[arg-type]

305 + _tool_to_melody(42)

306

307

308 # ======================================================================

@@ -418,7 +418,7 @@ class TestConversationToMelody:

418 {"role": "user", "content": "hi"},

419 {"role": "assistant", "content": "hello"},

420 ]

421 - out = _conversation_to_melody_messages(conv) # type: ignore[arg-type]

421 + out = _conversation_to_melody_messages(conv)

422 assert out == [

423 {

424 "role": "user",

@@ -443,7 +443,7 @@ class TestConversationToMelody:

443 "reasoning": "thoughts",

444 }

445 ]

446 - out = _conversation_to_melody_messages(conv) # type: ignore[arg-type]

446 + out = _conversation_to_melody_messages(conv)

447 assert out[0]["content"] == [

448 {"type": MelodyContentType.THINKING, "thinking": "thoughts"},

449 {"type": MelodyContentType.TEXT, "text": "answer"},

@@ -457,7 +457,7 @@ class TestConversationToMelody:

457 "reasoning_content": "thoughts",

458 }

459 ]

460 - out = _conversation_to_melody_messages(conv) # type: ignore[arg-type]

460 + out = _conversation_to_melody_messages(conv)

461 assert out[0]["content"][0] == {

462 "type": MelodyContentType.THINKING,

463 "thinking": "thoughts",

@@ -474,7 +474,7 @@ class TestConversationToMelody:

474 "reasoning": "should be ignored",

475 }

476 ]

477 - out = _conversation_to_melody_messages(conv) # type: ignore[arg-type]

477 + out = _conversation_to_melody_messages(conv)

478 assert out[0]["content"] == [{"type": MelodyContentType.TEXT, "text": "hi"}]

479

480 def test_tool_calls_normalized(self):

@@ -491,7 +491,7 @@ class TestConversationToMelody:

491 ],

492 }

493 ]

494 - out = _conversation_to_melody_messages(conv) # type: ignore[arg-type]

494 + out = _conversation_to_melody_messages(conv)

495 assert out[0]["tool_calls"] == [

496 {"id": "c1", "name": "f", "parameters": '{"a":1}'}

497 ]

@@ -504,7 +504,7 @@ class TestConversationToMelody:

504 "tool_call_id": "c1",

505 }

506 ]

507 - out = _conversation_to_melody_messages(conv) # type: ignore[arg-type]

507 + out = _conversation_to_melody_messages(conv)

508 assert out[0]["tool_call_id"] == "c1"

509

510 def test_messages_citations_attached_by_index(self):

@@ -531,20 +531,20 @@ class TestConversationToMelody:

531 }

532 ]

533 }

534 - out = _conversation_to_melody_messages(conv, citations) # type: ignore[arg-type]

534 + out = _conversation_to_melody_messages(conv, citations)

535 assert "citations" not in out[0]

536 assert out[1]["citations"] == citations[1]

537

538 def test_messages_citations_none_is_a_no_op(self):

539 conv = [{"role": "assistant", "content": "a"}]

540 - out = _conversation_to_melody_messages(conv, None) # type: ignore[arg-type]

540 + out = _conversation_to_melody_messages(conv, None)

541 assert "citations" not in out[0]

542

543 def test_messages_citations_missing_index_is_a_no_op(self):

544 # A ``messages_citations`` dict whose key doesn't hit any

545 # message must not attach anything (and must not raise).

546 conv = [{"role": "assistant", "content": "a"}]

547 - out = _conversation_to_melody_messages(conv, {5: [{"anything": 1}]}) # type: ignore[arg-type]

547 + out = _conversation_to_melody_messages(conv, {5: [{"anything": 1}]})

548 assert "citations" not in out[0]

549

550 def test_tool_message_v2_content_replaces_default_content(self):

@@ -575,7 +575,7 @@ class TestConversationToMelody:

575 ]

576 }

577 out = _conversation_to_melody_messages(

578 - conv, # type: ignore[arg-type]

578 + conv,

579 tool_message_v2_content=v2_content,

580 )

581 # User message (no forwarded entry) still goes through the

@@ -597,7 +597,7 @@ class TestConversationToMelody:

597 ]

598 }

599 out = _conversation_to_melody_messages(

600 - conv, # type: ignore[arg-type]

600 + conv,

601 tool_message_v2_content=v2_content,

602 )

603 assert out[0]["content"] == [

@@ -607,7 +607,7 @@ class TestConversationToMelody:

607 def test_tool_message_v2_content_none_falls_back_to_default(self):

608 conv = [{"role": "tool", "tool_call_id": "c", "content": "plain text"}]

609 out = _conversation_to_melody_messages(

610 - conv, # type: ignore[arg-type]

610 + conv,

611 tool_message_v2_content=None,

612 )

613 assert out[0]["content"] == [{"type": "text", "text": "plain text"}]

@@ -626,7 +626,7 @@ class TestBuildRenderConfig:

626 # Bare kwargs -> cmd4 (the current Command A+ prompt format).

627 # Mirrors ``_DEFAULT_FORMAT`` in ``vllm/renderers/cohere.py`` and

628 # the ``--cohere-format`` CLI default.

629 - fmt, cfg = _build_render_config(self._conv(), {}) # type: ignore[arg-type]

629 + fmt, cfg = _build_render_config(self._conv(), {})

630 assert fmt == "cmd4"

631 assert cfg["use_jinja"] is True

632 assert isinstance(cfg["messages"], list)

@@ -634,12 +634,12 @@ class TestBuildRenderConfig:

634 assert "additional_template_fields" not in cfg

635

636 def test_explicit_cmd3(self):

637 - fmt, cfg = _build_render_config(self._conv(), {"cohere_format": "cmd3"}) # type: ignore[arg-type]

637 + fmt, cfg = _build_render_config(self._conv(), {"cohere_format": "cmd3"})

638 assert fmt == "cmd3"

639

640 def test_invalid_format_raises(self):

641 with pytest.raises(ValueError, match="Invalid cohere_format"):

642 - _build_render_config(self._conv(), {"cohere_format": "cmd5"}) # type: ignore[arg-type]

642 + _build_render_config(self._conv(), {"cohere_format": "cmd5"})

643

644 def test_documents_converted(self):

645 _, cfg = _build_render_config(

@@ -650,7 +650,7 @@ class TestBuildRenderConfig:

650 {"id": "d1", "data": {"text": "wrapped"}},

651 ]

652 },

653 - ) # type: ignore[arg-type]

653 + )

654 assert cfg["documents"] == [

655 {"text": "doc text"},

656 {"id": "d1", "text": "wrapped"},

@@ -665,7 +665,7 @@ class TestBuildRenderConfig:

665 {"type": "function", "function": {"name": "preferred"}}

666 ],

667 },

668 - ) # type: ignore[arg-type]

668 + )

669 names = [t["name"] for t in cfg["available_tools"]]

670 assert names == ["preferred"]

671

@@ -673,30 +673,30 @@ class TestBuildRenderConfig:

673 _, cfg = _build_render_config(

674 self._conv(),

675 {"tools": [{"type": "function", "function": {"name": "from_tools"}}]},

676 - ) # type: ignore[arg-type]

676 + )

677 assert [t["name"] for t in cfg["available_tools"]] == ["from_tools"]

678

679 @pytest.mark.parametrize("value", ["enabled", "disabled"])

680 def test_reasoning_type_direct(self, value):

681 - _, cfg = _build_render_config(self._conv(), {"reasoning_type": value}) # type: ignore[arg-type]

681 + _, cfg = _build_render_config(self._conv(), {"reasoning_type": value})

682 assert cfg["reasoning_type"] == value

683

684 def test_thinking_dict_shorthand_resolves_reasoning_type(self):

685 - _, cfg = _build_render_config(self._conv(), {"thinking": {"type": "enabled"}}) # type: ignore[arg-type]

685 + _, cfg = _build_render_config(self._conv(), {"thinking": {"type": "enabled"}})

686 assert cfg["reasoning_type"] == "enabled"

687

688 def test_thinking_shorthand_ignores_unknown_type(self):

689 - _, cfg = _build_render_config(self._conv(), {"thinking": {"type": "auto"}}) # type: ignore[arg-type]

689 + _, cfg = _build_render_config(self._conv(), {"thinking": {"type": "auto"}})

690 assert "reasoning_type" not in cfg

691

692 def test_dev_instruction_forwarded(self):

693 - _, cfg = _build_render_config(self._conv(), {"dev_instruction": "be brief"}) # type: ignore[arg-type]

693 + _, cfg = _build_render_config(self._conv(), {"dev_instruction": "be brief"})

694 assert cfg["dev_instruction"] == "be brief"

695

696 def test_response_format_json_object_sets_json_mode(self):

697 _, cfg = _build_render_config(

698 self._conv(), {"response_format": {"type": "json_object"}}

699 - ) # type: ignore[arg-type]

699 + )

700 assert cfg["json_mode"] is True

701 assert "json_schema" not in cfg

702

@@ -705,7 +705,7 @@ class TestBuildRenderConfig:

705 _, cfg = _build_render_config(

706 self._conv(),

707 {"response_format": {"type": "json_schema", "schema": schema}},

708 - ) # type: ignore[arg-type]

708 + )

709 # JSON-encoded for melody (string-only schema field).

710 assert cfg["json_schema"] == json.dumps(schema)

711

@@ -721,35 +721,35 @@ class TestBuildRenderConfig:

721 "schema": {"schema": inner},

722 }

723 },

724 - ) # type: ignore[arg-type]

724 + )

725 assert cfg["json_schema"] == json.dumps(inner)

726

727 def test_json_schema_kwarg_direct(self):

728 # Caller can also pass ``json_schema`` directly, both as dict and

729 # as a pre-stringified value.

730 - _, cfg = _build_render_config(self._conv(), {"json_schema": {"a": 1}}) # type: ignore[arg-type]

730 + _, cfg = _build_render_config(self._conv(), {"json_schema": {"a": 1}})

731 assert cfg["json_schema"] == '{"a": 1}'

732 _, cfg = _build_render_config(

733 self._conv(), {"json_schema": "raw-string-schema"}

734 - ) # type: ignore[arg-type]

734 + )

735 assert cfg["json_schema"] == "raw-string-schema"

736

737 def test_json_mode_kwarg_overrides(self):

738 - _, cfg = _build_render_config(self._conv(), {"json_mode": True}) # type: ignore[arg-type]

738 + _, cfg = _build_render_config(self._conv(), {"json_mode": True})

739 assert cfg["json_mode"] is True

740

741 def test_cmd3_safety_mode_lowercased(self):

742 _, cfg = _build_render_config(

743 self._conv(),

744 {"cohere_format": "cmd3", "safety_mode": "CONTEXTUAL"},

745 - ) # type: ignore[arg-type]

745 + )

746 assert cfg["safety_mode"] == "contextual"

747

748 def test_cmd3_citation_quality_direct(self):

749 _, cfg = _build_render_config(

750 self._conv(),

751 {"cohere_format": "cmd3", "citation_quality": "ACCURATE"},

752 - ) # type: ignore[arg-type]

752 + )

753 assert cfg["citation_quality"] == "accurate"

754

755 def test_cmd3_citation_quality_derived_from_citation_options(self):

@@ -758,20 +758,20 @@ class TestBuildRenderConfig:

758 _, cfg = _build_render_config(

759 self._conv(),

760 {"cohere_format": "cmd3", "citation_options": {"mode": "accurate"}},

761 - ) # type: ignore[arg-type]

761 + )

762 assert cfg["citation_quality"] == "on"

763

764 _, cfg = _build_render_config(

765 self._conv(),

766 {"cohere_format": "cmd3", "citation_options": {"mode": "off"}},

767 - ) # type: ignore[arg-type]

767 + )

768 assert cfg["citation_quality"] == "off"

769

770 def test_cmd3_skip_preamble_forwarded(self):

771 _, cfg = _build_render_config(

772 self._conv(),

773 {"cohere_format": "cmd3", "skip_preamble": True},

774 - ) # type: ignore[arg-type]

774 + )

775 assert cfg["skip_preamble"] is True

776

777 def test_cmd3_no_grounding_field(self):

@@ -779,7 +779,7 @@ class TestBuildRenderConfig:

779 _, cfg = _build_render_config(

780 self._conv(),

781 {"cohere_format": "cmd3", "grounding": "fast"},

782 - ) # type: ignore[arg-type]

782 + )

783 assert "grounding" not in cfg

784

785 @pytest.mark.parametrize(

@@ -800,7 +800,7 @@ class TestBuildRenderConfig:

800 _, cfg = _build_render_config(

801 self._conv(),

802 {"cohere_format": "cmd4", "grounding": raw},

803 - ) # type: ignore[arg-type]

803 + )

804 assert cfg["grounding"] == expected

805

806 @pytest.mark.parametrize(

@@ -818,7 +818,7 @@ class TestBuildRenderConfig:

818 "cohere_format": "cmd4",

819 "citation_options": {"mode": mode},

820 },

821 - ) # type: ignore[arg-type]

821 + )

822 assert cfg["grounding"] == expected

823

824 def test_cmd4_grounding_rejects_unknown_value(self):

@@ -826,7 +826,7 @@ class TestBuildRenderConfig:

826 _build_render_config(

827 self._conv(),

828 {"cohere_format": "cmd4", "grounding": "foobar"},

829 - ) # type: ignore[arg-type]

829 + )

830

831 def test_cmd4_platform_instruction(self):

832 _, cfg = _build_render_config(

@@ -835,7 +835,7 @@ class TestBuildRenderConfig:

835 "cohere_format": "cmd4",

836 "platform_instruction": "do this",

837 },

838 - ) # type: ignore[arg-type]

838 + )

839 assert cfg["platform_instruction"] == "do this"

840

841 def test_cmd4_no_safety_mode_field(self):

@@ -847,7 +847,7 @@ class TestBuildRenderConfig:

847 "safety_mode": "contextual",

848 "citation_quality": "on",

849 },

850 - ) # type: ignore[arg-type]

850 + )

851 assert "safety_mode" not in cfg

852 assert "citation_quality" not in cfg

853

@@ -862,7 +862,7 @@ class TestBuildRenderConfig:

862 "my_var": "x",

863 "documents": ["doc"], # consumed, must NOT leak through

864 },

865 - ) # type: ignore[arg-type]

865 + )

866 extras = cfg["additional_template_fields"]

867 assert extras == {"reasoning_effort": "low", "my_var": "x"}

868 # Sanity: the consumed key still produced its dedicated config slot.

@@ -875,7 +875,7 @@ class TestBuildRenderConfig:

875 _, cfg = _build_render_config(

876 self._conv(),

877 {"template_id": "tpl1"},

878 - ) # type: ignore[arg-type]

878 + )

879 assert cfg["template_id"] == "tpl1"

880 # use_jinja is always True, regardless of caller input.

881 assert cfg["use_jinja"] is True

@@ -896,7 +896,7 @@ class TestBuildRenderConfig:

896 _build_render_config(

897 self._conv(),

898 {cohere_only_key: "raw {{ jinja }}"},

899 - ) # type: ignore[arg-type]

899 + )

900

901 @pytest.mark.parametrize("cohere_only_key", ["template_jinja", "template"])

902 def test_cohere_only_template_kwargs_none_is_tolerated(self, cohere_only_key):

@@ -906,7 +906,7 @@ class TestBuildRenderConfig:

906 _, cfg = _build_render_config(

907 self._conv(),

908 {cohere_only_key: None},

909 - ) # type: ignore[arg-type]

909 + )

910 assert cohere_only_key not in cfg

911 assert "additional_template_fields" not in cfg

912

@@ -918,7 +918,7 @@ class TestBuildRenderConfig:

918 self._conv(),

919 {},

920 "raw {{ jinja }}",

921 - ) # type: ignore[arg-type]

921 + )

922 assert cfg["template_jinja"] == "raw {{ jinja }}"

923 assert cfg["use_jinja"] is True

924

@@ -927,7 +927,7 @@ class TestBuildRenderConfig:

927 self._conv(),

928 {},

929 None,

930 - ) # type: ignore[arg-type]

930 + )

931 assert "template_jinja" not in cfg

932

933 def test_tool_message_v2_content_key_picked_up_from_kwargs(self):

@@ -954,7 +954,7 @@ class TestBuildRenderConfig:

954 ]

955 }

956 }

957 - _, cfg = _build_render_config(conv, kwargs) # type: ignore[arg-type]

957 + _, cfg = _build_render_config(conv, kwargs)

958 tool_msg = cfg["messages"][1]

959 assert tool_msg["content"] == [{"type": "document", "document": {"r": "k"}}]

960 # And the key must be *consumed* -- not surfaced as a stray

tests/renderers/test_completions.py

+1 -1

@@ -139,7 +139,7 @@ class TestValidatePrompt:

139

140 with pytest.raises(TypeError, match="should be a list of integers"):

141 renderer.render_prompts(

142 - _preprocess_prompt(renderer.model_config, [[1, 2], ["foo", "bar"]]) # type: ignore[arg-type]

142 + _preprocess_prompt(renderer.model_config, [[1, 2], ["foo", "bar"]])

143 )

144

145

tests/samplers/test_beam_search.py

+3 -3

@@ -113,7 +113,7 @@ def test_beam_search_abort_returns_partial_outputs_and_continues_other_prompts(

113 {"type": "token", "prompt_token_ids": [token]} for token in [prompt_token, 2]

114 ]

115 params = BeamSearchParams(beam_width=2, max_tokens=3)

116 - outputs = llm.beam_search(prompts, params) # type: ignore[arg-type]

116 + outputs = llm.beam_search(prompts, params)

117

118 aborted, normal = outputs

119 expected_tokens = (

@@ -326,7 +326,7 @@ def test_beam_search_structured_output(

326

327 llm = LLM(

328 model=model,

329 - dtype=dtype, # type: ignore[arg-type]

329 + dtype=dtype,

330 max_model_len=512,

331 structured_outputs_config=dict(

332 backend="xgrammar",

@@ -345,7 +345,7 @@ def test_beam_search_structured_output(

345 "Generate a JSON object for a person with name and age:",

346 ]

347

348 - outputs = llm.beam_search(prompts, params) # type: ignore[arg-type]

348 + outputs = llm.beam_search(prompts, params)

349

350 assert len(outputs) == len(prompts)

351 for output in outputs:

tests/samplers/test_beam_search_online.py

+2 -2

@@ -57,8 +57,8 @@ class _EngineClient:

57

58

59 class _Serving(BeamSearchOnlineMixin):

60 - renderer = _Renderer() # type: ignore[assignment]

61 - engine_client = _EngineClient() # type: ignore[assignment]

60 + renderer = _Renderer()

61 + engine_client = _EngineClient()

62

63

64 @pytest.mark.asyncio

tests/standalone_tests/lazy_imports.py

+1 -1

@@ -17,6 +17,6 @@ module_names = ["torch._inductor.async_compile", "cv2"]

17 # if we import any modules during `import vllm`, there would be a

18 # hard error and nice stacktrace on the first import.

19 for module_name in module_names:

20 - sys.modules[module_name] = None # type: ignore[assignment]

20 + sys.modules[module_name] = None

21

22 import vllm # noqa

tests/test_audio_media_size_precheck.py

+1 -1

@@ -144,6 +144,6 @@ async def test_audio_connector_threads_byte_limit_to_http_reader(

144 assert kwargs["max_bytes"] == MiB_bytes

145 raise VLLMValidationError("Maximum file size exceeded")

146

147 - connector = MediaConnector(connection=_Connection()) # type: ignore[arg-type]

147 + connector = MediaConnector(connection=_Connection())

148 with pytest.raises(VLLMValidationError, match="Maximum file size exceeded"):

149 await connector.fetch_audio_async("https://example.test/audio")

tests/tokenizers_/test_mistral.py

+5 -5

@@ -975,7 +975,7 @@ class TestMistralTokenizer:

975 mistral_tokenizer: MistralTokenizer,

976 ):

977 assert mistral_tokenizer.batch_decode(

978 - [[]], # type: ignore[arg-type]

978 + [[]],

979 ) == [""]

980

981 def test_convert_tokens_to_string(self, mistral_tokenizer: MistralTokenizer):

@@ -2246,7 +2246,7 @@ class TestMistralTokenizer:

2246 ]

2247

2248 output = mistral_tokenizer.apply_chat_template(

2249 - messages, # type: ignore[arg-type]

2249 + messages,

2250 add_generation_prompt=True,

2251 )

2252 decoded = mistral_tokenizer.tokenizer.decode(output, SpecialTokenPolicy.KEEP)

@@ -2352,7 +2352,7 @@ def test_v15_apply_chat_template_passes_reasoning_effort_high(

2352 )

2353

2354 v15_mistral_tokenizer.apply_chat_template(

2355 - messages=_PASSTHROUGH_MESSAGES, # type: ignore[arg-type]

2355 + messages=_PASSTHROUGH_MESSAGES,

2356 reasoning_effort="high",

2357 )

2358

@@ -2375,7 +2375,7 @@ def test_v15_apply_chat_template_passes_reasoning_effort_none_by_default(

2375 fake_apply_chat_template,

2376 )

2377

2378 - v15_mistral_tokenizer.apply_chat_template(messages=_PASSTHROUGH_MESSAGES) # type: ignore[arg-type]

2378 + v15_mistral_tokenizer.apply_chat_template(messages=_PASSTHROUGH_MESSAGES)

2379

2380 assert "reasoning_effort" in captured_kwargs[-1]

2381 assert captured_kwargs[-1]["reasoning_effort"] is None

@@ -2398,7 +2398,7 @@ def test_pre_v15_apply_chat_template_omits_reasoning_effort(

2398 )

2399

2400 v13_mistral_tokenizer.apply_chat_template(

2401 - messages=_PASSTHROUGH_MESSAGES, # type: ignore[arg-type]

2401 + messages=_PASSTHROUGH_MESSAGES,

2402 reasoning_effort="high",

2403 )

2404

tests/tokenizers_/test_registry.py

+1 -1

@@ -205,7 +205,7 @@ def test_cached_tokenizer_from_config_registers_local_config(tmp_path: Path):

205 side_effect=lambda tokenizer: tokenizer,

206 ),

207 ):

208 - tokenizer = cached_tokenizer_from_config(model_config) # type: ignore[arg-type]

208 + tokenizer = cached_tokenizer_from_config(model_config)

209

210 assert tokenizer.is_fast is True

211 finally:

tests/tool_parsers/test_ernie45_moe_tool_parser.py

+2 -2

@@ -54,7 +54,7 @@ def test_extract_tool_calls_no_tools(ernie45_tool_parser):

54 model_output = "This is a test"

55 extracted_tool_calls = ernie45_tool_parser.extract_tool_calls(

56 model_output, request=None

57 - ) # type: ignore[arg-type]

57 + )

58 assert not extracted_tool_calls.tools_called

59 assert extracted_tool_calls.tool_calls == []

60 assert extracted_tool_calls.content == model_output

@@ -163,7 +163,7 @@ def test_extract_tool_calls(

163 ):

164 extracted_tool_calls = ernie45_tool_parser.extract_tool_calls(

165 model_output, request=None

166 - ) # type: ignore[arg-type]

166 + )

167 assert extracted_tool_calls.tools_called

168

169 assert_tool_calls(extracted_tool_calls.tool_calls, expected_tool_calls)

tests/tool_parsers/test_jamba_tool_parser.py

+3 -3

@@ -78,7 +78,7 @@ def stream_delta_message_generator(

78 previous_token_ids,

79 current_token_ids,

80 delta_token_ids,

81 - request=None, # type: ignore[arg-type]

81 + request=None,

82 )

83 if delta_message:

84 yield delta_message

@@ -95,7 +95,7 @@ def test_extract_tool_calls_no_tools(jamba_tool_parser):

95 model_output = "This is a test"

96 extracted_tool_calls = jamba_tool_parser.extract_tool_calls(

97 model_output, request=None

98 - ) # type: ignore[arg-type]

98 + )

99 assert not extracted_tool_calls.tools_called

100 assert extracted_tool_calls.tool_calls == []

101 assert extracted_tool_calls.content == model_output

@@ -166,7 +166,7 @@ def test_extract_tool_calls(

166 ):

167 extracted_tool_calls = jamba_tool_parser.extract_tool_calls(

168 model_output, request=None

169 - ) # type: ignore[arg-type]

169 + )

170 assert extracted_tool_calls.tools_called

171

172 assert_tool_calls(extracted_tool_calls.tool_calls, expected_tool_calls)

tests/tool_parsers/test_qwen3coder_tool_parser.py

+1 -1

@@ -225,7 +225,7 @@ def test_extract_tool_calls_no_tools(qwen3_tool_parser):

225 model_output = "This is a test response without any tool calls"

226 extracted_tool_calls = qwen3_tool_parser.extract_tool_calls(

227 model_output, request=None

228 - ) # type: ignore[arg-type]

228 + )

229 assert not extracted_tool_calls.tools_called

230 assert extracted_tool_calls.tool_calls == []

231 assert extracted_tool_calls.content == model_output

tests/tool_parsers/test_step3p5_tool_parser.py

+1 -1

@@ -170,7 +170,7 @@ def test_extract_tool_calls_no_tools(step3p5_tool_parser):

170 model_output = "This is a test response without any tool calls"

171 extracted_tool_calls = step3p5_tool_parser.extract_tool_calls(

172 model_output, request=None

173 - ) # type: ignore[arg-type]

173 + )

174 assert not extracted_tool_calls.tools_called

175 assert extracted_tool_calls.tool_calls == []

176 assert extracted_tool_calls.content == model_output

tests/tool_parsers/test_xlam_tool_parser.py

+4 -4

@@ -100,7 +100,7 @@ def test_extract_tool_calls_no_tools(xlam_tool_parser):

100 model_output = "This is a test"

101 extracted_tool_calls = xlam_tool_parser.extract_tool_calls(

102 model_output, request=None

103 - ) # type: ignore[arg-type]

103 + )

104 assert not extracted_tool_calls.tools_called

105 assert extracted_tool_calls.tool_calls == []

106 assert extracted_tool_calls.content == model_output

@@ -225,7 +225,7 @@ def test_extract_tool_calls(

225 ):

226 extracted_tool_calls = xlam_tool_parser.extract_tool_calls(

227 model_output, request=None

228 - ) # type: ignore[arg-type]

228 + )

229 assert extracted_tool_calls.tools_called

230

231 assert_tool_calls(extracted_tool_calls.tool_calls, expected_tool_calls)

@@ -263,7 +263,7 @@ def test_extract_tool_calls_list_structure(

263 """Test extraction of tool calls when the model outputs a list-structured tool call.""" # noqa: E501

264 extracted_tool_calls = xlam_tool_parser.extract_tool_calls(

265 model_output, request=None

266 - ) # type: ignore[arg-type]

266 + )

267 assert extracted_tool_calls.tools_called

268

269 assert_tool_calls(extracted_tool_calls.tool_calls, expected_tool_calls)

@@ -550,7 +550,7 @@ def test_extract_tool_calls_non_ascii(xlam_tool_parser, xlam_tokenizer, streamin

550 if delta.tool_calls and delta.tool_calls[0].function.arguments

551 )

552 else:

553 - extracted = xlam_tool_parser.extract_tool_calls(model_output, request=None) # type: ignore[arg-type]

553 + extracted = xlam_tool_parser.extract_tool_calls(model_output, request=None)

554 args = "".join(tc.function.arguments for tc in extracted.tool_calls)

555

556 assert "北京" in args

tests/tool_use/test_gemma4_responses_adjust_request.py

+1 -1

@@ -148,7 +148,7 @@ def test_tool_parser_adjust_request_builds_valid_response_text_config() -> None:

148 description string that the previous two-step construction injected.

149 """

150 parser = ToolParser.__new__(ToolParser)

151 - parser.model_tokenizer = None # type: ignore[assignment]

151 + parser.model_tokenizer = None

152

153 request = _build_responses_request(tool_choice="required")

154 ToolParser.adjust_request(parser, request)

tests/tool_use/test_kimi_k3_tool_parser.py

+1 -1

@@ -370,7 +370,7 @@ def test_streaming_split_markers_do_not_leak():

370 assert len(tool_deltas) == 1

371 assert tool_deltas[0].function is not None

372 assert tool_deltas[0].function.name == "calc"

373 - assert json.loads(tool_deltas[0].function.arguments) == {"x": 1} # type: ignore[arg-type]

373 + assert json.loads(tool_deltas[0].function.arguments) == {"x": 1}

374

375

376 def test_tool_call_ids_are_unique_across_messages():

tests/transformers_utils/processors/test_voxtral.py

+1 -1

@@ -98,7 +98,7 @@ def test_fetch_audio_rejects_unsupported_type(

98 feature_extractor: MistralCommonFeatureExtractor,

99 ):

100 with pytest.raises(TypeError, match="only a numpy array"):

101 - feature_extractor.fetch_audio(42) # type: ignore[arg-type]

101 + feature_extractor.fetch_audio(42)

102

103

104 def test_fetch_audio_str_delegates_to_load_audio(

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)) # type: ignore[arg-type]

47 + keys = get_processor_kwargs_keys(get_processor_kwargs_type(proc))

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)) # type: ignore[arg-type]

66 + keys = get_processor_kwargs_keys(get_processor_kwargs_type(proc))

67 _assert_has_all_expected(keys)

tests/utils.py

+1 -1

@@ -830,7 +830,7 @@ class RemoteOpenAIServerCustom(RemoteOpenAIServer):

830 self.proc: Process = cast(Any, ctx).Process(

831 target=_run_in_new_process_group,

832 args=(self.child_process_fxn, env_dict, model, vllm_serve_args),

833 - ) # type: ignore[assignment]

833 + )

834 with _temporarily_sanitized_pythonpath_env():

835 self.proc.start()

836

tests/utils_/test_argparse_utils.py

+9 -9

@@ -476,32 +476,32 @@ def test_compilation_config_mode_validator():

476 """Test that CompilationConfig.mode field validator converts strings to integers."""

477 from vllm.config.compilation import CompilationConfig, CompilationMode

478

479 - config = CompilationConfig(mode=0) # type: ignore[arg-type]

479 + config = CompilationConfig(mode=0)

480 assert config.mode == CompilationMode.NONE

481

482 - config = CompilationConfig(mode=3) # type: ignore[arg-type]

482 + config = CompilationConfig(mode=3)

483 assert config.mode == CompilationMode.VLLM_COMPILE

484

485 - config = CompilationConfig(mode="NONE") # type: ignore[arg-type]

485 + config = CompilationConfig(mode="NONE")

486 assert config.mode == CompilationMode.NONE

487

488 - config = CompilationConfig(mode="STOCK_TORCH_COMPILE") # type: ignore[arg-type]

488 + config = CompilationConfig(mode="STOCK_TORCH_COMPILE")

489 assert config.mode == CompilationMode.STOCK_TORCH_COMPILE

490

491 - config = CompilationConfig(mode="DYNAMO_TRACE_ONCE") # type: ignore[arg-type]

491 + config = CompilationConfig(mode="DYNAMO_TRACE_ONCE")

492 assert config.mode == CompilationMode.DYNAMO_TRACE_ONCE

493

494 - config = CompilationConfig(mode="VLLM_COMPILE") # type: ignore[arg-type]

494 + config = CompilationConfig(mode="VLLM_COMPILE")

495 assert config.mode == CompilationMode.VLLM_COMPILE

496

497 - config = CompilationConfig(mode="none") # type: ignore[arg-type]

497 + config = CompilationConfig(mode="none")

498 assert config.mode == CompilationMode.NONE

499

500 - config = CompilationConfig(mode="vllm_compile") # type: ignore[arg-type]

500 + config = CompilationConfig(mode="vllm_compile")

501 assert config.mode == CompilationMode.VLLM_COMPILE

502

503 with pytest.raises(ValidationError, match="Invalid compilation mode"):

504 - CompilationConfig(mode="INVALID_MODE") # type: ignore[arg-type]

504 + CompilationConfig(mode="INVALID_MODE")

505

506

507 def test_flat_product():

tests/v1/attention/test_mla_prefill_selector.py

+2 -2

@@ -511,14 +511,14 @@ class TestMLAPrefillBackendParsing:

511

512 def test_valid_string_parses_to_enum(self):

513 config = AttentionConfig(

514 - mla_prefill_backend="FLASH_ATTN", # type: ignore[arg-type]

514 + mla_prefill_backend="FLASH_ATTN",

515 )

516 assert config.mla_prefill_backend == MLAPrefillBackendEnum.FLASH_ATTN

517

518 def test_invalid_string_raises_error(self):

519 with pytest.raises(ValueError, match="Unknown MLA prefill backend"):

520 AttentionConfig(

521 - mla_prefill_backend="NONEXISTENT", # type: ignore[arg-type]

521 + mla_prefill_backend="NONEXISTENT",

522 )

523

524

tests/v1/attention/test_sparse_mla_backends.py

+3 -3

@@ -3575,7 +3575,7 @@ def test_flashmla_fp8_paths_accept_decode_subset(monkeypatch, use_mixed_batch: b

3575

3576 if use_mixed_batch:

3577 fp8_metadata = FlashMLASparseMetadata.FP8KernelMetadata(

3578 - scheduler_metadata=object(), # type: ignore[arg-type]

3578 + scheduler_metadata=object(),

3579 dummy_block_table=torch.empty(1, 1, dtype=torch.int32, device=DEVICE_TYPE),

3580 cache_lens=torch.empty(1, dtype=torch.int32, device=DEVICE_TYPE),

3581 )

@@ -3588,7 +3588,7 @@ def test_flashmla_fp8_paths_accept_decode_subset(monkeypatch, use_mixed_batch: b

3588 num_prefill_tokens=num_batch_tokens - num_decode_tokens,

3589 decode=FP8Meta.Decode(

3590 seq_lens=torch.empty(1, dtype=torch.int32, device=DEVICE_TYPE),

3591 - kernel_metadata=object(), # type: ignore[arg-type]

3591 + kernel_metadata=object(),

3592 decode_query_len=num_decode_tokens,

3593 ),

3594 prefill=FP8Meta.Prefill(

@@ -3774,7 +3774,7 @@ def test_fp8_mixed_batch_dcp_neutralizes_empty_rows(monkeypatch):

3774

3775 metadata = SimpleNamespace(

3776 fp8_extra_metadata=FlashMLASparseMetadata.FP8KernelMetadata(

3777 - scheduler_metadata=object(), # type: ignore[arg-type]

3777 + scheduler_metadata=object(),

3778 dummy_block_table=torch.empty(1, 1, dtype=torch.int32, device=DEVICE_TYPE),

3779 cache_lens=torch.empty(1, dtype=torch.int32, device=DEVICE_TYPE),

3780 ),

tests/v1/cudagraph/test_cudagraph_manager.py

+2 -2

@@ -134,7 +134,7 @@ def test_piecewise_capture_uses_pcp_dummy_slot_mappings():

134 device=torch.device("cpu"),

135 max_num_reqs=num_reqs,

136 max_num_tokens=num_tokens,

137 - block_tables=pcp_block_tables, # type: ignore[arg-type]

137 + block_tables=pcp_block_tables,

138 )

139

140 block_tables = MagicMock()

@@ -224,7 +224,7 @@ def test_speculator_capture_preserves_decode_query_bounds(

224 )

225 manager = SpeculatorCudaGraphManager.__new__(SpeculatorCudaGraphManager)

226 manager.max_num_reqs = num_reqs

227 - manager.vllm_config = SimpleNamespace( # type: ignore[assignment]

227 + manager.vllm_config = SimpleNamespace(

228 parallel_config=SimpleNamespace(data_parallel_size=1)

229 )

230 buffers = InputBuffers(num_reqs, num_tokens, torch.device("cpu"))

tests/v1/cudagraph/test_encoder_cudagraph.py

+4 -4

@@ -652,7 +652,7 @@ def test_eonly_capture_preserves_outputs_across_replay_and_fallback(profile_only

652 encoder.device = device

653 encoder.cudagraph_manager = manager

654 runner = object.__new__(MMEncoderModelRunner)

655 - runner.model_state = SimpleNamespace(encoder_runner=encoder) # type: ignore[assignment]

655 + runner.model_state = SimpleNamespace(encoder_runner=encoder)

656 # No decoder manager is installed: capture must be encoder-only.

657 with patch(

658 "vllm.v1.worker.mm_encoder_model_runner.lock_workspace", wraps=lock_workspace

@@ -673,7 +673,7 @@ def test_eonly_capture_preserves_outputs_across_replay_and_fallback(profile_only

673 with patch(

674 "vllm.v1.worker.gpu.ec_connector.get_ec_transfer", return_value=connector

675 ):

676 - ec = ActiveECConnector(SimpleNamespace(), cache) # type: ignore[arg-type]

676 + ec = ActiveECConnector(SimpleNamespace(), cache)

677 scheduled = cast(

678 SchedulerOutput,

679 SimpleNamespace(

@@ -716,7 +716,7 @@ def test_eonly_without_encoder_graph_skips_capture():

716 encoder = object.__new__(EncoderRunner)

717 encoder.cudagraph_manager = None

718 runner = object.__new__(MMEncoderModelRunner)

719 - runner.model_state = SimpleNamespace(encoder_runner=encoder) # type: ignore[assignment]

719 + runner.model_state = SimpleNamespace(encoder_runner=encoder)

720 assert runner.capture_model() == 0

721

722

@@ -1023,7 +1023,7 @@ class TestInitInvariantValidation:

1023 vllm_config = _MockVllmConfig(token_budgets, max_mm_items)

1024 model = _MockModel(min_budget, max_budget)

1025 return EncoderCudaGraphManager(

1026 - vllm_config=vllm_config, # type: ignore[arg-type]

1026 + vllm_config=vllm_config,

1027 device=torch.device("cpu"),

1028 dtype=torch.float32,

1029 model=model,

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}, # type: ignore[arg-type]

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

65 )

66

67 eagle_engine_args = replace(

tests/v1/engine/test_startup_watch_processes.py

+2 -2

@@ -304,9 +304,9 @@ def test_wait_for_engine_startup_reports_watched_process_exit():

304 wait_for_engine_startup(

305 handshake_socket,

306 [CoreEngine()],

307 - parallel_config, # type: ignore[arg-type]

307 + parallel_config,

308 coordinated_dp=False,

309 - cache_config=None, # type: ignore[arg-type]

309 + cache_config=None,

310 launch=launch,

311 )

312 finally:

tests/v1/executor/test_executor.py

+2 -2

@@ -126,7 +126,7 @@ def test_custom_executor_type_checking():

126 model=MODEL,

127 gpu_memory_utilization=0.2,

128 max_model_len=8192,

129 - distributed_executor_backend=Mock, # type: ignore[arg-type]

129 + distributed_executor_backend=Mock,

130 )

131 LLMEngine.from_engine_args(engine_args)

132 with pytest.raises(ValueError):

@@ -134,7 +134,7 @@ def test_custom_executor_type_checking():

134 model=MODEL,

135 gpu_memory_utilization=0.2,

136 max_model_len=8192,

137 - distributed_executor_backend=Mock, # type: ignore[arg-type]

137 + distributed_executor_backend=Mock,

138 )

139 AsyncLLM.from_engine_args(engine_args)

140

tests/v1/kv_connector/unit/offloading_connector/test_metrics.py

+9 -9

@@ -480,7 +480,7 @@ def test_reset():

480

481 def test_prom_metrics_observes_manager_counter():

482 prom_metrics = OffloadPromMetrics(

483 - vllm_config=_FakeVllmConfig(), # type: ignore[arg-type]

483 + vllm_config=_FakeVllmConfig(),

484 metric_types={

485 Gauge: _FakeMetric,

486 Counter: _FakeMetric,

@@ -506,7 +506,7 @@ def test_prom_metrics_observes_manager_counter():

506

507 def test_prom_metrics_observes_flat_transfer_metrics_and_legacy_metrics():

508 prom_metrics = OffloadPromMetrics(

509 - vllm_config=_FakeVllmConfig(), # type: ignore[arg-type]

509 + vllm_config=_FakeVllmConfig(),

510 metric_types={

511 Gauge: _FakeMetric,

512 Counter: _FakeMetric,

@@ -568,7 +568,7 @@ def test_prom_metrics_observes_manager_gauge_and_histogram():

568 return_value=_spec_cls_with_metric_definitions(metric_definitions),

569 ):

570 prom_metrics = OffloadPromMetrics(

571 - vllm_config=_FakeVllmConfig(store_threshold=0), # type: ignore[arg-type]

571 + vllm_config=_FakeVllmConfig(store_threshold=0),

572 metric_types={

573 Gauge: _FakeMetric,

574 Counter: _FakeMetric,

@@ -612,7 +612,7 @@ def test_prom_metrics_lazily_observes_labeled_metric():

612 return_value=_spec_cls_with_metric_definitions(metric_definitions),

613 ):

614 prom_metrics = OffloadPromMetrics(

615 - vllm_config=_FakeVllmConfig(store_threshold=0), # type: ignore[arg-type]

615 + vllm_config=_FakeVllmConfig(store_threshold=0),

616 metric_types={

617 Gauge: _FakeMetric,

618 Counter: _FakeMetric,

@@ -651,7 +651,7 @@ def test_prom_metrics_rejects_wrong_label_count():

651 return_value=_spec_cls_with_metric_definitions(metric_definitions),

652 ):

653 prom_metrics = OffloadPromMetrics(

654 - vllm_config=_FakeVllmConfig(store_threshold=0), # type: ignore[arg-type]

654 + vllm_config=_FakeVllmConfig(store_threshold=0),

655 metric_types={

656 Gauge: _FakeMetric,

657 Counter: _FakeMetric,

@@ -672,7 +672,7 @@ def test_prom_metrics_rejects_wrong_label_count():

672

673 def test_prom_metrics_uses_configured_manager_metrics():

674 prom_metrics = OffloadPromMetrics(

675 - vllm_config=_FakeVllmConfig(store_threshold=0), # type: ignore[arg-type]

675 + vllm_config=_FakeVllmConfig(store_threshold=0),

676 metric_types={

677 Gauge: _FakeMetric,

678 Counter: _FakeMetric,

@@ -693,7 +693,7 @@ def test_prom_metrics_registers_tiering_metrics_from_spec():

693 "spec_name": "TieringOffloadingSpec",

694 "secondary_tiers": [],

695 },

696 - ), # type: ignore[arg-type]

696 + ),

697 metric_types={

698 Gauge: _FakeMetric,

699 Counter: _FakeMetric,

@@ -741,7 +741,7 @@ def test_aggregate_into_empty_stats():

741 def test_prom_metrics_multi_engine_routing():

742 """Metrics are routed to the correct engine index."""

743 prom_metrics = OffloadPromMetrics(

744 - vllm_config=_FakeVllmConfig(), # type: ignore[arg-type]

744 + vllm_config=_FakeVllmConfig(),

745 metric_types={

746 Gauge: _FakeMetric,

747 Counter: _FakeMetric,

@@ -767,7 +767,7 @@ def test_prom_metrics_multi_engine_routing():

767 def test_prom_metrics_rejects_undeclared_metric():

768 """observe() asserts if a metric was never declared in metadata."""

769 prom_metrics = OffloadPromMetrics(

770 - vllm_config=_FakeVllmConfig(store_threshold=0), # type: ignore[arg-type]

770 + vllm_config=_FakeVllmConfig(store_threshold=0),

771 metric_types={

772 Gauge: _FakeMetric,

773 Counter: _FakeMetric,

tests/v1/kv_connector/unit/test_mooncake_store_worker.py

+1 -1

@@ -1543,7 +1543,7 @@ def test_store_sending_thread_reports_job_when_the_preamble_raises():

1543 # every dequeue leaves through the same exit.

1544 thread = _make_store_sending_thread(MagicMock())

1545 req = _make_store_req("req-a", [b"a0", b"a1"])

1546 - req.token_len_chunk = None # type: ignore[assignment]

1546 + req.token_len_chunk = None

1547

1548 with contextlib.suppress(TypeError):

1549 _run_store_req(thread, req)

tests/v1/kv_connector/unit/test_nixl_connector.py

+1 -1

@@ -114,7 +114,7 @@ def get_default_xfer_telemetry(

114 class AttributeDict(dict):

115 __slots__ = ()

116 __getattr__ = dict.__getitem__

117 - __setattr__ = dict.__setitem__ # type: ignore[assignment]

117 + __setattr__ = dict.__setitem__

118

119 # We can't instantiate nixlXferTelemetry because it's read only and

120 # ray env does not have NIXL, so we must fake it

tests/v1/kv_connector/unit/test_nixl_push_connector.py

+1 -1

@@ -1256,7 +1256,7 @@ class TestPushWriterNegative:

1256 w = _StubWriterWorker.fresh()

1257 for bogus_rid in (123, None, 4.5, b"bytes-not-str"):

1258 payload = _registration_data("placeholder")

1259 - payload["request_id"] = bogus_rid # type: ignore[assignment]

1259 + payload["request_id"] = bogus_rid

1260 notif = PUSH_REG_NOTIF_PREFIX + msgspec.msgpack.encode(payload)

1261 w._handle_push_reg_notif(notif)

1262 assert w._pending_d_registrations == {}

tests/v1/kv_offload/tiering/p2p/test_manager.py

+41 -41

@@ -349,10 +349,10 @@ class TestServeExternalRequests:

349 mgr._failed_serve_ctxs = [ctx]

350 sess_a = _RecordingSession()

351 sess_b = _RecordingSession()

352 - mgr._sessions = {"a": sess_a, "b": sess_b} # type: ignore[assignment]

352 + mgr._sessions = {"a": sess_a, "b": sess_b}

353

354 parent = _RecordingParent()

355 - mgr.serve_external_requests(parent) # type: ignore[arg-type]

355 + mgr.serve_external_requests(parent)

356

357 # Failed serve released and queue cleared.

358 assert parent.finished == ["p2p:peer:req-1:lu1"]

@@ -364,10 +364,10 @@ class TestServeExternalRequests:

364 def test_no_failed_serves_still_serves_sessions(self):

365 mgr = _make_manager()

366 sess = _RecordingSession()

367 - mgr._sessions = {"a": sess} # type: ignore[assignment]

367 + mgr._sessions = {"a": sess}

368

369 parent = _RecordingParent()

370 - mgr.serve_external_requests(parent) # type: ignore[arg-type]

370 + mgr.serve_external_requests(parent)

371

372 assert parent.finished == []

373 assert sess.served_with == [parent]

@@ -427,7 +427,7 @@ class TestSubmitStore:

427 forwards directly to that session rather than re-buffering."""

428 mgr = _make_manager()

429 bound = _FakeSession(peer_id="10.0.0.1:8000", connected=True)

430 - mgr._kv_to_session["req-1"] = bound # type: ignore[assignment]

430 + mgr._kv_to_session["req-1"] = bound

431 # Note: _sessions is intentionally untouched — this test isolates

432 # the kv_request_id → session fast path.

433 job = _job_metadata(

@@ -495,7 +495,7 @@ class TestSubmitLoad:

495 mgr = _make_manager()

496 peer_id = "10.0.0.1:8000"

497 existing = _FakeSession(peer_id=peer_id, connected=True)

498 - mgr._sessions[peer_id] = existing # type: ignore[assignment]

498 + mgr._sessions[peer_id] = existing

499 job = _job_metadata(

500 job_id=42,

501 keys=[b"k1", b"k2"],

@@ -581,7 +581,7 @@ class TestOnRequestFinished:

581 (FetchMsg received) routes via _kv_to_session and pops the entry."""

582 mgr = _make_manager()

583 bound = _FakeSession(peer_id="some-peer:1", connected=True)

584 - mgr._kv_to_session["req-1"] = bound # type: ignore[assignment]

584 + mgr._kv_to_session["req-1"] = bound

585 ctx = _req_context(kv_params=_remote_decoder_kv_params(kv_request_id="req-1"))

586 mgr.on_request_finished(ctx)

587 assert bound.finishes == ["req-1"]

@@ -718,8 +718,8 @@ class TestGetFinished:

718 def poll(self):

719 return []

720

721 - mgr._control = FakeControl() # type: ignore[assignment]

722 - mgr._data = None # type: ignore[assignment]

721 + mgr._control = FakeControl()

722 + mgr._data = None

723 return mgr

724

725 def test_drains_finished_jobs(self):

@@ -740,7 +740,7 @@ class TestGetFinished:

740 pass

741

742 mgr = self._make()

743 - mgr._data = FakeData() # type: ignore[assignment]

743 + mgr._data = FakeData()

744 dead = _FakeSession(

745 peer_id="dead:1234",

746 alive=False,

@@ -749,7 +749,7 @@ class TestGetFinished:

749 close_req_ids=["req-load"],

750 close_stores=[10, 11],

751 )

752 - mgr._sessions["dead:1234"] = dead # type: ignore[assignment]

752 + mgr._sessions["dead:1234"] = dead

753

754 results = list(mgr.get_finished_jobs())

755 # 2 baseline + 1 failed load + 2 failed stores

@@ -769,14 +769,14 @@ class TestGetFinished:

769 pass

770

771 mgr = self._make()

772 - mgr._data = FakeData() # type: ignore[assignment]

772 + mgr._data = FakeData()

773 dead = _FakeSession(

774 peer_id="dead:1234",

775 alive=False,

776 connected=True,

777 close_req_ids=["req-probe-1", "req-probe-2"],

778 )

779 - mgr._sessions["dead:1234"] = dead # type: ignore[assignment]

779 + mgr._sessions["dead:1234"] = dead

780

781 list(mgr.get_finished_jobs())

782 assert "dead:1234" not in mgr._sessions

@@ -854,7 +854,7 @@ class TestHasPendingWork:

854 """The result is the same regardless of session state — there is

855 no 'idle' branch."""

856 mgr = _make_manager()

857 - mgr._sessions["peer:1"] = _FakeSession(peer_id="peer:1") # type: ignore[assignment]

857 + mgr._sessions["peer:1"] = _FakeSession(peer_id="peer:1")

858 assert mgr.has_pending_work() is True

859

860

@@ -919,12 +919,12 @@ class TestShutdownDrain:

919 mgr = _make_manager()

920 data = _ShutdownFakeData(still_queue=still_queue)

921 control = _ShutdownFakeControl()

922 - mgr._data = data # type: ignore[assignment]

923 - mgr._control = control # type: ignore[assignment]

922 + mgr._data = data

923 + mgr._control = control

924 if inflight_ids:

925 session = _FakeSession(peer_id="peer:1", connected=True)

926 session._server._inflight = {tid: object() for tid in inflight_ids}

927 - mgr._sessions["peer:1"] = session # type: ignore[assignment]

927 + mgr._sessions["peer:1"] = session

928 return mgr, data, control

929

930 def test_shutdown_drains_inflight_via_wait_cancel(self):

@@ -1148,10 +1148,10 @@ def _build_paired_managers() -> tuple[P2PSecondaryTierManager, P2PSecondaryTierM

1148 ctrl_b = _LoopbackControl(mgr_b._local_id)

1149 ctrl_a.pair(ctrl_b)

1150

1151 - mgr_a._control = ctrl_a # type: ignore[assignment]

1152 - mgr_b._control = ctrl_b # type: ignore[assignment]

1153 - mgr_a._data = _FakeData(mgr_a._local_id) # type: ignore[assignment]

1154 - mgr_b._data = _FakeData(mgr_b._local_id) # type: ignore[assignment]

1151 + mgr_a._control = ctrl_a

1152 + mgr_b._control = ctrl_b

1153 + mgr_a._data = _FakeData(mgr_a._local_id)

1154 + mgr_b._data = _FakeData(mgr_b._local_id)

1155

1156 return mgr_a, mgr_b

1157

@@ -1334,7 +1334,7 @@ class TestAcceptNewPeers:

1334 mgr = _make_manager()

1335 peer_id = "10.0.0.1:8000"

1336 existing = _FakeSession(peer_id=peer_id, connected=True)

1337 - mgr._sessions[peer_id] = existing # type: ignore[assignment]

1337 + mgr._sessions[peer_id] = existing

1338

1339 new_conn = _RecordingConn(peer_id)

1340 mgr._accept_new_peers([new_conn])

@@ -1365,7 +1365,7 @@ class TestAcceptNewPeers:

1365 pass

1366

1367 mgr = _make_manager()

1368 - mgr._data = FakeData() # type: ignore[assignment]

1368 + mgr._data = FakeData()

1369 peer_id = "10.0.0.1:8000"

1370

1371 # Real ControlConnection-shaped fake: send/close/peer_id only.

@@ -1382,7 +1382,7 @@ class TestAcceptNewPeers:

1382 def close(self) -> None:

1383 self.alive = False

1384

1385 - mgr._accept_new_peers([_Conn(peer_id)]) # type: ignore[arg-type]

1385 + mgr._accept_new_peers([_Conn(peer_id)])

1386

1387 assert peer_id in mgr._sessions

1388 assert mgr._sessions[peer_id].connected is True

@@ -1411,7 +1411,7 @@ class TestPollOnce:

1411 loads=[LoadResult(job_id=11, kv_request_id="req-11", success=True)],

1412 stores=[StoreResult(job_id=22, success=True)],

1413 )

1414 - mgr._sessions[peer_alive] = alive # type: ignore[assignment]

1414 + mgr._sessions[peer_alive] = alive

1415

1416 # Dead session whose pending close() jobs surface as failures.

1417 peer_dead = "10.0.0.3:9999"

@@ -1423,7 +1423,7 @@ class TestPollOnce:

1423 close_req_ids=["req-33"],

1424 close_stores=[44],

1425 )

1426 - mgr._sessions[peer_dead] = dead # type: ignore[assignment]

1426 + mgr._sessions[peer_dead] = dead

1427

1428 class _Ctrl:

1429 def poll(self_inner):

@@ -1433,8 +1433,8 @@ class TestPollOnce:

1433 def remove_remote_peer(self_inner, pid):

1434 pass

1435

1436 - mgr._control = _Ctrl() # type: ignore[assignment]

1437 - mgr._data = _Data() # type: ignore[assignment]

1436 + mgr._control = _Ctrl()

1437 + mgr._data = _Data()

1438

1439 mgr._poll_once()

1440

@@ -1464,7 +1464,7 @@ class TestPollOnce:

1464 connected=True,

1465 new_fetch_ids=["req-1"],

1466 )

1467 - mgr._sessions[peer] = sess # type: ignore[assignment]

1467 + mgr._sessions[peer] = sess

1468 mgr._unbound_stores["req-1"] = [

1469 _UnboundStoreBatch(job_id=5, keys=[b"k1"], block_ids=[0]),

1470 _UnboundStoreBatch(job_id=6, keys=[b"k2"], block_ids=[1]),

@@ -1474,7 +1474,7 @@ class TestPollOnce:

1474 def poll(self_inner):

1475 return []

1476

1477 - mgr._control = _Ctrl() # type: ignore[assignment]

1477 + mgr._control = _Ctrl()

1478 mgr._poll_once()

1479

1480 assert mgr._kv_to_session["req-1"] is sess

@@ -1496,13 +1496,13 @@ class TestPollOnce:

1496 connected=True,

1497 new_fetch_ids=["req-fast"],

1498 )

1499 - mgr._sessions[peer] = sess # type: ignore[assignment]

1499 + mgr._sessions[peer] = sess

1500

1501 class _Ctrl:

1502 def poll(self_inner):

1503 return []

1504

1505 - mgr._control = _Ctrl() # type: ignore[assignment]

1505 + mgr._control = _Ctrl()

1506 mgr._poll_once()

1507

1508 assert mgr._kv_to_session["req-fast"] is sess

@@ -1519,13 +1519,13 @@ class TestPollOnce:

1519 connected=True,

1520 loads=[LoadResult(job_id=5, kv_request_id="req-5", success=False)],

1521 )

1522 - mgr._sessions[peer] = sess # type: ignore[assignment]

1522 + mgr._sessions[peer] = sess

1523

1524 class _Ctrl:

1525 def poll(self_inner):

1526 return []

1527

1528 - mgr._control = _Ctrl() # type: ignore[assignment]

1528 + mgr._control = _Ctrl()

1529

1530 mgr._poll_once()

1531

@@ -1549,26 +1549,26 @@ class TestDrainJobs:

1549 def test_returns_immediately_when_quiescent(self):

1550 """No sessions and no inflight: drain_jobs returns without sleeping."""

1551 mgr = _make_manager()

1552 - mgr._control = _DrainCtrl() # type: ignore[assignment]

1552 + mgr._control = _DrainCtrl()

1553

1554 sleeps: list[float] = []

1555 # If drain_jobs sleeps when nothing is pending, that's a regression.

1556 import vllm.v1.kv_offload.tiering.p2p.manager as m

1557

1558 original_sleep = m.time.sleep

1559 - m.time.sleep = lambda s: sleeps.append(s) # type: ignore[assignment]

1559 + m.time.sleep = lambda s: sleeps.append(s)

1560 try:

1561 mgr.drain_jobs()

1562 finally:

1563 - m.time.sleep = original_sleep # type: ignore[assignment]

1563 + m.time.sleep = original_sleep

1564

1565 assert sleeps == []

1566

1567 def test_returns_when_session_has_no_inflight_or_inbound(self):

1568 """A session with empty _inbound and _inflight does not block drain."""

1569 mgr = _make_manager()

1570 - mgr._control = _DrainCtrl() # type: ignore[assignment]

1571 - mgr._sessions["peer:1"] = _FakeSession(peer_id="peer:1") # type: ignore[assignment]

1570 + mgr._control = _DrainCtrl()

1571 + mgr._sessions["peer:1"] = _FakeSession(peer_id="peer:1")

1572 # Should return on the first iteration.

1573 mgr.drain_jobs()

1574

@@ -1576,10 +1576,10 @@ class TestDrainJobs:

1576 """A session that stays inflight past 5s triggers the warning, and

1577 once it clears the loop returns."""

1578 mgr = _make_manager()

1579 - mgr._control = _DrainCtrl() # type: ignore[assignment]

1579 + mgr._control = _DrainCtrl()

1580 sess = _FakeSession(peer_id="peer:1")

1581 sess._server._inflight = {1: object()} # non-empty

1582 - mgr._sessions["peer:1"] = sess # type: ignore[assignment]

1582 + mgr._sessions["peer:1"] = sess

1583

1584 # Synthetic monotonic clock: 100.0 for the start stamp, then 106.0

1585 # for the first elapsed-check (past the 5s warning threshold), then

tests/v1/kv_offload/tiering/p2p/test_sessions.py

+15 -15

@@ -288,17 +288,17 @@ def _make_session(

288 session = P2PSession(

289 peer_id=peer_id,

290 local_id=local_id,

291 - transport=transport, # type: ignore[arg-type]

291 + transport=transport,

292 local_block_len=transport.block_len,

293 local_hash_seed=local_hash_seed,

294 - conn=conn, # type: ignore[arg-type]

294 + conn=conn,

295 )

296 return session, conn, transport

297

298

299 def _serve(session: P2PSession, parent: FakeParent) -> None:

300 """Resolve enqueued inbound lookups, as the manager does each step."""

301 - session.serve_external_requests(parent) # type: ignore[arg-type]

301 + session.serve_external_requests(parent)

302

303

304 def _activate(

@@ -1951,7 +1951,7 @@ class TestFinishRequestServerSide:

1951 )

1952 session.poll()

1953 # Force write_blocks to fail on the next call.

1954 - transport.write_blocks = lambda *a, **kw: None # type: ignore[assignment]

1954 + transport.write_blocks = lambda *a, **kw: None # type: ignore[method-assign]

1955

1956 session.add_stored_blocks("req-1", [b"k1"], [0], job_id=42)

1957

@@ -2057,7 +2057,7 @@ class TestFinishRequestServerSide:

2057 assert outbound.finishing is False

2058

2059 # Round 2: write_blocks fails for k2 while transfer_1 is still inflight.

2060 - transport.write_blocks = lambda *a, **kw: None # type: ignore[assignment]

2060 + transport.write_blocks = lambda *a, **kw: None # type: ignore[method-assign]

2061 session.add_stored_blocks("req-1", [b"k2"], [1], job_id=200)

2062 # No new transfer was registered.

2063 assert list(session._server._inflight.keys()) == [tid_1]

@@ -2163,7 +2163,7 @@ class TestPendingSession:

2163 session = P2PSession(

2164 peer_id="peer:8000",

2165 local_id="local:9000",

2166 - transport=transport, # type: ignore[arg-type]

2166 + transport=transport,

2167 local_block_len=4096,

2168 local_hash_seed=_DEFAULT_HASH_SEED,

2169 conn=None,

@@ -2183,13 +2183,13 @@ class TestPendingSession:

2183 session = P2PSession(

2184 peer_id="peer:8000",

2185 local_id="local:9000",

2186 - transport=transport, # type: ignore[arg-type]

2186 + transport=transport,

2187 local_block_len=4096,

2188 local_hash_seed=_DEFAULT_HASH_SEED,

2189 conn=None,

2190 )

2191 conn = FakeConnection(peer_id="peer:8000")

2192 - session.attach_connection(conn) # type: ignore[arg-type]

2192 + session.attach_connection(conn)

2193 assert conn._sent

2194 assert conn._sent[0][TYPE_KEY] == ConnectMsg.TYPE

2195

@@ -2197,7 +2197,7 @@ class TestPendingSession:

2197 """attach_connection on an already-connected session raises."""

2198 session, conn, _ = _make_session()

2199 with pytest.raises(ValueError, match="already connected"):

2200 - session.attach_connection(FakeConnection()) # type: ignore[arg-type]

2200 + session.attach_connection(FakeConnection())

2201

2202 def test_pending_close_returns_pending_stores(self):

2203 """Closing a pending session reports buffered stores as failed."""

@@ -2205,7 +2205,7 @@ class TestPendingSession:

2205 session = P2PSession(

2206 peer_id="peer:8000",

2207 local_id="local:9000",

2208 - transport=transport, # type: ignore[arg-type]

2208 + transport=transport,

2209 local_block_len=4096,

2210 local_hash_seed=_DEFAULT_HASH_SEED,

2211 conn=None,

@@ -2313,7 +2313,7 @@ class TestAdversarial:

2313 def test_non_dict_message(self):

2314 session, conn, _ = _make_session()

2315 _activate(session, conn)

2316 - conn._inbox.append(42) # type: ignore[arg-type]

2316 + conn._inbox.append(42)

2317 result_ = session.poll()

2318 loads = result_.loads

2319 stores = result_.stores

@@ -2402,7 +2402,7 @@ class TestDispatchErrorHandling:

2402 def _boom(*args, **kwargs):

2403 raise RuntimeError("simulated internal bug")

2404

2405 - session._server.on_fetch = _boom # type: ignore[assignment]

2405 + session._server.on_fetch = _boom

2406 conn.enqueue(

2407 {

2408 TYPE_KEY: FetchMsg.TYPE,

@@ -2425,7 +2425,7 @@ class TestDispatchErrorHandling:

2425 def _boom(*args, **kwargs):

2426 raise RuntimeError("simulated internal bug")

2427

2428 - session._server.on_fetch = _boom # type: ignore[assignment]

2428 + session._server.on_fetch = _boom

2429 for _ in range(_MAX_CONSECUTIVE_DISPATCH_ERRORS):

2430 conn.enqueue(

2431 {

@@ -2454,7 +2454,7 @@ class TestDispatchErrorHandling:

2454 # Alternate (boom, success) (_MAX-1) times: counter rises to 1

2455 # then resets to 0 each cycle, never reaching the threshold.

2456 for _ in range(_MAX_CONSECUTIVE_DISPATCH_ERRORS - 1):

2457 - session._server.on_fetch = _boom # type: ignore[assignment]

2457 + session._server.on_fetch = _boom

2458 conn.enqueue(

2459 {

2460 TYPE_KEY: FetchMsg.TYPE,

@@ -2465,7 +2465,7 @@ class TestDispatchErrorHandling:

2465 }

2466 )

2467 session.poll()

2468 - session._server.on_fetch = original_on_fetch # type: ignore[assignment]

2468 + session._server.on_fetch = original_on_fetch

2469 # A benign no-op message (unknown type) dispatches cleanly

2470 # and resets the consecutive-error counter.

2471 conn.enqueue({TYPE_KEY: "unknown_for_test"})

tests/v1/sample/test_dry.py

+3 -3

@@ -340,7 +340,7 @@ def _make_v2_state(

340 device=device,

341 all_token_ids=SimpleNamespace(gpu=all_tokens),

342 )

343 - state = DryState(_make_vllm_config(), req_states) # type: ignore[arg-type]

343 + state = DryState(_make_vllm_config(), req_states)

344 if tokenizer is not None:

345 # Re-resolve the default set now that a tokenizer is installed.

346 state._tokenizer = tokenizer

@@ -446,7 +446,7 @@ def test_spec_decode_is_refused_at_construction():

446 )

447 config = _make_vllm_config(speculative_config=SimpleNamespace())

448 with pytest.raises(ValueError, match="speculative decoding"):

449 - DryState(config, req_states) # type: ignore[arg-type]

449 + DryState(config, req_states)

450

451

452 def test_v2_draft_expanded_logits_raise():

@@ -918,7 +918,7 @@ def test_loads_through_the_loader_by_fqcn():

918 )

919 procs = loader.build_custom_logits_processors(

920 _make_vllm_config(),

921 - req_states, # type: ignore[arg-type]

921 + req_states,

922 False,

923 [DRY_FQCN],

924 )

tests/v1/sample/test_logprobs.py

+3 -3

@@ -486,12 +486,12 @@ def test_prompt_logprob_token_ids_require_v2_model_runner():

486 params = SamplingParams(prompt_logprob_token_ids=[1, 2])

487 with patch.object(SamplingParams, "verify"):

488 with pytest.raises(VLLMValidationError, match="V2 model runner"):

489 - InputProcessor._validate_params(processor, params, ("generate",)) # type: ignore[arg-type]

489 + InputProcessor._validate_params(processor, params, ("generate",))

490 processor.vllm_config.use_v2_model_runner = True

491 - InputProcessor._validate_params(processor, params, ("generate",)) # type: ignore[arg-type]

491 + InputProcessor._validate_params(processor, params, ("generate",))

492 processor.vllm_config.cache_config.kv_sharing_fast_prefill = True

493 with pytest.raises(VLLMValidationError, match="fast-prefill"):

494 - InputProcessor._validate_params(processor, params, ("generate",)) # type: ignore[arg-type]

494 + InputProcessor._validate_params(processor, params, ("generate",))

495

496

497 def test_none_logprobs(vllm_model, example_prompts):

tests/v1/sample/test_trace_replay_params.py

+1 -1

@@ -105,7 +105,7 @@ def test_trace_decode_token_ids_rejects_speculative_decoding():

105 with pytest.raises(ValueError, match="not supported with speculative decoding"):

106 params._validate_trace_replay(

107 _make_model_config(vocab_size=100),

108 - speculative_config=object(), # type: ignore[arg-type] # Only non-None presence is tested.

108 + speculative_config=object(),

109 )

110

111

tests/v1/simple_kv_offload/test_worker.py

+3 -3

@@ -71,7 +71,7 @@ SLEEP_CYCLES = 50_000_000

71 def _make_worker(

72 kv_cache_config: KVCacheConfig | None, cpu_capacity_bytes: int

73 ) -> SimpleCPUOffloadWorker:

74 - return SimpleCPUOffloadWorker(None, kv_cache_config, cpu_capacity_bytes) # type: ignore[arg-type]

74 + return SimpleCPUOffloadWorker(None, kv_cache_config, cpu_capacity_bytes)

75

76

77 def _make_backend() -> tuple[DmaCopyBackend, torch.Tensor, torch.Tensor]:

@@ -121,7 +121,7 @@ def test_no_forward_step_completes_cpu_store(monkeypatch, use_v2):

121 monkeypatch.setattr(f"{module}.get_forward_context", lambda: None)

122 result = active.no_forward(output)

123 else:

124 - result = KVConnectorModelRunnerMixin.kv_connector_no_forward(output, None) # type: ignore[arg-type]

124 + result = KVConnectorModelRunnerMixin.kv_connector_no_forward(output, None)

125 completion = (

126 result.kv_connector_output.kv_connector_worker_meta

127 if result.kv_connector_output is not None

@@ -283,7 +283,7 @@ def test_transfer_hooks_pass_wait_event_for_store_only():

283 cpu_capacity_bytes=0,

284 )

285 recording = _RecordingBackend()

286 - worker._backend = recording # type: ignore[assignment]

286 + worker._backend = recording

287 worker._connector_metadata = SimpleCPUOffloadMetadata(

288 load_event=0,

289 load_gpu_blocks=[0],

tests/v1/spec_decode/test_adaptive_verification.py

+4 -6

@@ -27,7 +27,7 @@ def make_manager(

27 manager.num_speculative_steps = num_steps

28 manager._stale_confidences = [SimpleNamespace(np=confidences)] # type: ignore[list-item] # CPU-only confidence buffer.

29 manager._stale_idx = 0

30 - manager.req_states = SimpleNamespace( # type: ignore[assignment] # CPU-only request fixture.

30 + manager.req_states = SimpleNamespace(

31 req_id_to_index={"low": 0, "high": 1},

32 num_computed_tokens_np=np.ones(num_reqs, dtype=np.int32),

33 prefill_len=SimpleNamespace(np=np.ones(num_reqs, dtype=np.int32)),

@@ -153,11 +153,11 @@ def test_manager_checks_target_varlen_cudagraph_bound(

153 manager = maybe_create_adaptive_verification_manager(

154 enable_adaptive_verification=True,

155 attn_groups=groups,

156 - req_states=SimpleNamespace(num_speculative_steps=7), # type: ignore[arg-type]

156 + req_states=SimpleNamespace(num_speculative_steps=7),

157 query_start_loc=object(),

158 num_bonus_tokens=1,

159 max_total_logits=1,

160 - vllm_config=None, # type: ignore[arg-type]

160 + vllm_config=None,

161 target_layer_names={"target"},

162 additional_attn_cg_support=additional,

163 )

@@ -186,9 +186,7 @@ def test_budget_stops_where_marginal_drafts_stop_paying_for_themselves():

186

187 def test_profiled_batches_seed_cost_curves_via_consumer(monkeypatch):

188 manager = AdaptiveVerificationManager.__new__(AdaptiveVerificationManager)

189 - manager.req_states = SimpleNamespace( # type: ignore[assignment] # CPU-only request fixture.

190 - max_num_batched_tokens=4096, max_num_reqs=64

191 - )

189 + manager.req_states = SimpleNamespace(max_num_batched_tokens=4096, max_num_reqs=64)

190 manager.num_speculative_steps = 7

191 manager.num_bonus_tokens = 1

192 curves: dict[str, list[tuple[int, float]]] = {}

tests/v1/spec_decode/test_dflash2.py

+8 -8

@@ -98,7 +98,7 @@ def test_draft_quant_exclusions_include_global_layer_indices():

98 ]

99 )

100

101 - _add_global_draft_layer_exclusions(quant_config, 88, 5) # type: ignore[arg-type]

101 + _add_global_draft_layer_exclusions(quant_config, 88, 5)

102

103 assert "layers.88.mlp_conv*" in quant_config.exclude_modules

104 assert "*layers.92.self_attn.q_proj" in quant_config.exclude_modules

@@ -171,7 +171,7 @@ def test_selector_leaves_greedy_drafting_without_proposal_logits(monkeypatch):

171 allocating one here would claim a proposal the walk never sampled from.

172 """

173 _stub_base(monkeypatch, None)

174 - speculator = DFlash2Speculator(None, torch.device("cpu")) # type: ignore[arg-type] # Base initialization is stubbed.

174 + speculator = DFlash2Speculator(None, torch.device("cpu"))

175

176 assert speculator.draft_logits is None

177

@@ -183,7 +183,7 @@ def test_selector_asks_for_fp32_proposal_logits():

183 candidate row often enough that the walk and the rejection sampler checking it

184 would no longer read the same distribution.

185 """

186 - dtype, fill = DFlash2Speculator.draft_logits_spec(None, None) # type: ignore[arg-type] # Constant spec ignores self/config.

186 + dtype, fill = DFlash2Speculator.draft_logits_spec(None, None)

187

188 assert dtype is torch.float32

189 assert fill == float("-inf")

@@ -309,9 +309,9 @@ def test_candidate_model_decoder_layer_cls(monkeypatch, variant):

309 vllm_config.compilation_config = mock_current_vllm_config.compilation_config

310

311 # 3. Instantiate the model under meta device to avoid parameter allocation issues

312 - with set_current_vllm_config(mock_current_vllm_config), torch.device("meta"): # type: ignore[arg-type]

312 + with set_current_vllm_config(mock_current_vllm_config), torch.device("meta"):

313 model_cls = DFlash2Qwen3Model if variant == "dflash2" else LiLiCorr

314 - model = model_cls(vllm_config=vllm_config) # type: ignore[arg-type]

314 + model = model_cls(vllm_config=vllm_config)

315

316 # 4. Assert that the layers are DFlash2Qwen3DecoderLayer (the subclass)

317 assert len(model.layers) == 2

@@ -350,7 +350,7 @@ def test_conv_projections_use_draft_quant_config(monkeypatch):

350 quant_method="W4A16_NVFP4", is_checkpoint_nvfp4_serialized=True

351 )

352 layer = DFlash2Qwen3DecoderLayer(

353 - SimpleNamespace( # type: ignore[arg-type]

353 + SimpleNamespace(

354 speculative_config=SimpleNamespace(num_speculative_tokens=7),

355 model_config=SimpleNamespace(dtype=torch.bfloat16),

356 ),

@@ -434,7 +434,7 @@ def test_context_kv_uses_quantized_projection_fallback(monkeypatch):

434 for projection in projections

435 ]

436 qwen3_dflash.DFlashQwen3Model._build_context_kv_buffers(

437 - model, # type: ignore[arg-type]

437 + model,

438 layers_attn,

439 has_bias=False,

440 )

@@ -442,7 +442,7 @@ def test_context_kv_uses_quantized_projection_fallback(monkeypatch):

442 assert all(not hasattr(projection, "weight") for projection in projections)

443

444 actual_k, actual_v = qwen3_dflash.DFlashQwen3Model._project_context_kv(

445 - model, # type: ignore[arg-type]

445 + model,

446 context_states,

447 num_ctx=2,

448 num_layers=2,

tests/v1/spec_decode/test_dflash_lookahead.py

+4 -4

@@ -141,9 +141,9 @@ def test_dflash_drafter_window_reserves_bonus_token():

141 speculative_config=_dflash_speculative_config(NUM_SPECULATIVE_TOKENS),

142 )

143 # window = 4, so 96 fits (96 + 4 == 100) but 97 does not (97 + 4 == 101)

144 - assert input_fits_in_drafter(dflash_runner, SimpleNamespace(max_seq_len=96)) # type: ignore[arg-type]

145 - assert not input_fits_in_drafter(dflash_runner, SimpleNamespace(max_seq_len=97)) # type: ignore[arg-type]

146 - assert not input_fits_in_drafter(dflash_runner, None) # type: ignore[arg-type] # No metadata.

144 + assert input_fits_in_drafter(dflash_runner, SimpleNamespace(max_seq_len=96))

145 + assert not input_fits_in_drafter(dflash_runner, SimpleNamespace(max_seq_len=97))

146 + assert not input_fits_in_drafter(dflash_runner, None)

147

148 # Other drafters don't reserve the bonus token, so 97 fits (97 + 3 == 100).

149 plain_runner = SimpleNamespace(

@@ -151,4 +151,4 @@ def test_dflash_drafter_window_reserves_bonus_token():

151 effective_drafter_max_model_len=100,

152 speculative_config=SimpleNamespace(use_dflash=lambda: False),

153 )

154 - assert input_fits_in_drafter(plain_runner, SimpleNamespace(max_seq_len=97)) # type: ignore[arg-type]

154 + assert input_fits_in_drafter(plain_runner, SimpleNamespace(max_seq_len=97))

tests/v1/spec_decode/test_dsv4_mtp_local_argmax.py

+1 -3

@@ -158,9 +158,7 @@ def test_wrapper_forwards_spec_step_idx():

158

159 hidden_states = torch.empty(0)

160 wrapper = object.__new__(DeepSeekV4MTP)

161 - wrapper.model = SimpleNamespace( # type: ignore[assignment]

162 - get_top_tokens=lambda *args: args

163 - )

161 + wrapper.model = SimpleNamespace(get_top_tokens=lambda *args: args)

162

163 assert DeepSeekV4MTP.get_top_tokens(wrapper, hidden_states, 2) == (hidden_states, 2)

164 assert DeepSeekV4MTP.get_top_tokens(wrapper, hidden_states) == (hidden_states, 0)

tests/v1/spec_decode/test_extract_hidden_states.py

+4 -4

@@ -51,7 +51,7 @@ def _create_proposer(

51 target_parallel_config=ParallelConfig(),

52 method="extract_hidden_states",

53 num_speculative_tokens=num_speculative_tokens,

54 - draft_model_config={ # type: ignore[arg-type]

54 + draft_model_config={

55 "hf_config": {

56 "eagle_aux_hidden_state_layer_ids": layer_ids,

57 }

@@ -103,7 +103,7 @@ def test_proposer_initialization_missing_layer_ids():

103 target_parallel_config=ParallelConfig(),

104 method="extract_hidden_states",

105 num_speculative_tokens=1,

106 - draft_model_config={ # type: ignore[arg-type]

106 + draft_model_config={

107 "hf_config": {} # Missing eagle_aux_hidden_state_layer_ids

108 },

109 )

@@ -363,7 +363,7 @@ def test_extract_hidden_states_text_only_config_regression():

363 target_parallel_config=ParallelConfig(),

364 method="extract_hidden_states",

365 num_speculative_tokens=1,

366 - draft_model_config={ # type: ignore[arg-type]

366 + draft_model_config={

367 "hf_config": {

368 "eagle_aux_hidden_state_layer_ids": [1, 2, 3, 4],

369 }

@@ -450,7 +450,7 @@ def test_extract_hidden_states_speculative_config_vlm():

450 target_parallel_config=ParallelConfig(),

451 method="extract_hidden_states",

452 num_speculative_tokens=1,

453 - draft_model_config={ # type: ignore[arg-type]

453 + draft_model_config={

454 "hf_config": {

455 "eagle_aux_hidden_state_layer_ids": [1, 2],

456 }

tests/v1/spec_decode/test_lilicorr.py

+11 -11

@@ -190,17 +190,17 @@ def test_runtime_length_keeps_trained_checkpoint_geometry(monkeypatch):

190 for slots in (1, 3):

191 spec.num_speculative_tokens = slots

192 head = LiLiCorr(

193 - vllm_config=SimpleNamespace(speculative_config=spec) # type: ignore[arg-type]

193 + vllm_config=SimpleNamespace(speculative_config=spec)

194 ).lilicorr

195 assert head.slot_embedding.shape == (1, 1, 3, 1, 8)

196 assert head.relative_slot_bias.shape == (2, 7)

197 spec.draft_model_config.hf_config.dflash_config["lilicorr_enabled"] = False

198 with pytest.raises(ValueError, match="lilicorr_enabled"):

199 - LiLiCorr(vllm_config=SimpleNamespace(speculative_config=spec)) # type: ignore[arg-type]

199 + LiLiCorr(vllm_config=SimpleNamespace(speculative_config=spec))

200 spec.draft_model_config.hf_config.dflash_config["lilicorr_enabled"] = True

201 spec.num_speculative_tokens = 4

202 with pytest.raises(ValueError, match="trained block_size"):

203 - LiLiCorr(vllm_config=SimpleNamespace(speculative_config=spec)) # type: ignore[arg-type]

203 + LiLiCorr(vllm_config=SimpleNamespace(speculative_config=spec))

204

205

206 @pytest.mark.parametrize("tp_size", [1, 2])

@@ -269,8 +269,8 @@ def test_candidates_use_global_partition_and_exclude_padding(monkeypatch, tp_siz

269 )

270 monkeypatch.setattr(logits_processor, "_topk", lambda x, k: x.topk(k, dim=-1))

271 ids, values = logits_processor.LogitsProcessor.get_top_k_tokens(

272 - fake, # type: ignore[arg-type]

273 - lm_head, # type: ignore[arg-type]

272 + fake,

273 + lm_head,

274 torch.empty(2, 1),

275 2,

276 return_log_probs=True,

@@ -291,13 +291,13 @@ def test_context_anchor_is_last_committed_normalized_feature():

291 anchor_valid=torch.ones(4, dtype=torch.bool),

292 )

293 batch = SimpleNamespace(num_reqs=2, query_start_loc=torch.tensor([0, 2, 6]))

294 - LiLiCorrSpeculator.prepare_context_anchor(spec, batch, torch.tensor([0, 2])) # type: ignore[arg-type]

294 + LiLiCorrSpeculator.prepare_context_anchor(spec, batch, torch.tensor([0, 2]))

295 torch.testing.assert_close(spec.anchor_hidden[:2], norm(hidden[[1, 3]]))

296 assert spec.anchor_valid.tolist() == [True, True, False, False]

297 assert not spec.anchor_hidden[2:].any()

298 # A smaller reordered batch must not retain an old request's anchor.

299 batch = SimpleNamespace(num_reqs=1, query_start_loc=torch.tensor([0, 3]))

300 - LiLiCorrSpeculator.prepare_context_anchor(spec, batch, torch.tensor([1])) # type: ignore[arg-type]

300 + LiLiCorrSpeculator.prepare_context_anchor(spec, batch, torch.tensor([1]))

301 torch.testing.assert_close(spec.anchor_hidden[0], norm(hidden[1]))

302 assert not spec.anchor_hidden[1:].any()

303

@@ -321,7 +321,7 @@ def test_candidate_generation_routes_scores_and_adaptive_inputs(monkeypatch, lil

321

322 monkeypatch.setattr(DFlashSpeculator, "__init__", init_base)

323 cls = LiLiCorrSpeculator if lilicorr else DFlash2Speculator

324 - spec = cls(None, torch.device("cpu")) # type: ignore[arg-type]

324 + spec = cls(None, torch.device("cpu"))

325 hidden = torch.arange(36).float().view(9, 4)

326 spec._run_model = lambda *args: hidden

327 spec.sample_indices = torch.tensor([1, 2, 4, 5, 7, 8])

@@ -334,7 +334,7 @@ def test_candidate_generation_routes_scores_and_adaptive_inputs(monkeypatch, lil

334 spec.draft_logits = None

335 spec.use_fp64_gumbel = False

336 spec.enable_adaptive_verification = True

337 - spec.input_buffers = SimpleNamespace(input_ids=torch.arange(9)) # type: ignore[assignment]

337 + spec.input_buffers = SimpleNamespace(input_ids=torch.arange(9))

338 candidates = torch.arange(24).view(6, 4)

339 log_probs = candidates.float().log_softmax(-1)

340 scores = torch.randn(3, 2, 4, 4)

@@ -636,7 +636,7 @@ def owned_head_model(monkeypatch):

636 compilation_config=CompilationConfig(mode=CompilationMode.NONE),

637 )

638 ):

639 - model = LiLiCorrForCausalLM(vllm_config=vllm_config) # type: ignore[arg-type]

639 + model = LiLiCorrForCausalLM(vllm_config=vllm_config)

640 return model, vllm_config

641

642 return build

@@ -746,7 +746,7 @@ def test_quantized_head_calls_methods_without_reading_packed_weights(monkeypatch

746 block_size=4,

747 rms_norm_eps=1e-6,

748 config=_config(hidden_size=8),

749 - quant_config=quant_config, # type: ignore[arg-type]

749 + quant_config=quant_config,

750 prefix="model.lilicorr",

751 )

752 with torch.no_grad():

tests/v1/spec_decode/test_llm_base_proposer.py

+1 -1

@@ -53,7 +53,7 @@ def _make_proposer(

53 monkeypatch.setattr(llm_base_proposer, "AttentionGroup", _FakeAttentionGroup)

54

55 proposer = EagleProposer.__new__(EagleProposer)

56 - proposer.vllm_config = None # type: ignore[assignment] # Patched backend discovery does not read config.

56 + proposer.vllm_config = None

57 proposer.device = None

58 proposer._draft_attn_layer_names = set(layer_names)

59 proposer.kv_cache_gid = -1

tests/v1/spec_decode/test_llm_base_proposer_sampling.py

+1 -1

@@ -85,7 +85,7 @@ def test_compute_probs_and_sample_next_token_uses_fp64_exponential_race():

85 def test_mtp_model_returns_tuple(architecture: str, expected: bool):

86 proposer = object.__new__(SpecDecodeBaseProposer)

87 proposer.method = "mtp"

88 - proposer.draft_model_config = SimpleNamespace( # type: ignore[assignment] # Architecture-only config.

88 + proposer.draft_model_config = SimpleNamespace(

89 hf_config=SimpleNamespace(architectures=[architecture])

90 )

91

tests/v1/spec_decode/test_rejection_sampler_utils.py

+1 -1

@@ -700,7 +700,7 @@ def test_verify_rejects_unproposed_drafts():

700 first_pos = inputs["pos"].view(num_trials, K + 1)[:, 0]

701 prefill_len = first_pos + 1

702 rejection_sampler = object.__new__(RejectionSampler)

703 - rejection_sampler.sampler = SimpleNamespace( # type: ignore[assignment]

703 + rejection_sampler.sampler = SimpleNamespace(

704 apply_sampling_params=lambda logits, *args: logits,

705 sampling_states=SimpleNamespace(

706 temperature=SimpleNamespace(gpu=inputs["temperature"]),

tests/v1/structured_output/test_backend_xgrammar_lark.py

+2 -2

@@ -48,9 +48,9 @@ def _validate(

48 ) -> StructuredOutputsParams:

49 params = SamplingParams(structured_outputs=StructuredOutputsParams(grammar=grammar))

50 params._validate_structured_outputs(

51 - _StubModelConfig(), # type: ignore[arg-type]

51 + _StubModelConfig(),

52 StructuredOutputsConfig(backend=backend),

53 - tokenizer=tokenizer or object(), # type: ignore[arg-type]

53 + tokenizer=tokenizer or object(),

54 )

55 assert params.structured_outputs is not None

56 return params.structured_outputs

tests/v1/structured_output/test_backend_xgrammar_plamo3.py

+3 -3

@@ -51,7 +51,7 @@ def make_vllm_config() -> VllmConfig:

51 config = VllmConfig(

52 structured_outputs_config=StructuredOutputsConfig(backend="xgrammar"),

53 )

54 - config.model_config = SimpleNamespace( # type: ignore[assignment]

54 + config.model_config = SimpleNamespace(

55 hf_config=SimpleNamespace(model_type="plamo3")

56 )

57 return config

@@ -73,7 +73,7 @@ def test_plamo3_python_tokenizer_compiles_and_accepts_structural_tag():

73 tokenizer = Plamo3Tokenizer(["<eos>", *call_tokens])

74 backend = XgrammarBackend(

75 make_vllm_config(),

76 - tokenizer=tokenizer, # type: ignore[arg-type]

76 + tokenizer=tokenizer,

77 vocab_size=len(tokenizer.get_vocab()),

78 )

79 tools = [

@@ -107,7 +107,7 @@ def test_plamo3_grammar_matches_decoded_bytes(value):

107 tokenizer = Plamo3Tokenizer(list(dict.fromkeys(["<eos>", value, *pieces])))

108 backend = XgrammarBackend(

109 make_vllm_config(),

110 - tokenizer=tokenizer, # type: ignore[arg-type]

110 + tokenizer=tokenizer,

111 vocab_size=len(tokenizer.get_vocab()),

112 )

113 for tokens in ([value], pieces):

tests/v1/structured_output/test_structured_output_manager.py

+1 -1

@@ -108,7 +108,7 @@ def _build_harness(

108 )

109 manager = StructuredOutputManager(vllm_config)

110 if use_reasoner:

111 - manager.reasoner_cls = MockReasoner # type: ignore[assignment]

111 + manager.reasoner_cls = MockReasoner

112

113 sampling_params = SamplingParams(

114 structured_outputs=structured_outputs

tests/v1/structured_output/test_validation.py

+10 -10

@@ -35,7 +35,7 @@ def test_structured_outputs_rejected_for_diffusion_models():

35 )

36 with pytest.raises(VLLMValidationError, match="not yet supported for diffusion"):

37 params._validate_structured_outputs(

38 - _StubModelConfig(is_diffusion=True), # type: ignore[arg-type]

38 + _StubModelConfig(is_diffusion=True),

39 StructuredOutputsConfig(),

40 tokenizer=None,

41 )

@@ -45,7 +45,7 @@ def test_plain_request_allowed_for_diffusion_models():

45 """Requests without structured outputs are unaffected by the guard."""

46 params = SamplingParams()

47 params._validate_structured_outputs(

48 - _StubModelConfig(is_diffusion=True), # type: ignore[arg-type]

48 + _StubModelConfig(is_diffusion=True),

49 StructuredOutputsConfig(),

50 tokenizer=None,

51 )

@@ -71,9 +71,9 @@ def test_degenerate_structured_outputs_rejected(structured_outputs, match):

71 params = SamplingParams(structured_outputs=structured_outputs)

72 with pytest.raises(VLLMValidationError, match=match):

73 params._validate_structured_outputs(

74 - _StubModelConfig(is_diffusion=False), # type: ignore[arg-type]

74 + _StubModelConfig(is_diffusion=False),

75 StructuredOutputsConfig(),

76 - tokenizer=object(), # type: ignore[arg-type]

76 + tokenizer=object(),

77 )

78

79

@@ -97,9 +97,9 @@ def test_regex_with_nul_byte_rejected(regex):

97 # (which would otherwise catch the error and fall back to another backend).

98 with pytest.raises(VLLMValidationError, match="NUL"):

99 params._validate_structured_outputs(

100 - _StubModelConfig(is_diffusion=False), # type: ignore[arg-type]

100 + _StubModelConfig(is_diffusion=False),

101 StructuredOutputsConfig(),

102 - tokenizer=object(), # type: ignore[arg-type]

102 + tokenizer=object(),

103 )

104

105 # The xgrammar backend also rejects it directly (defense in depth), before

@@ -133,9 +133,9 @@ def test_unsupported_grammar_is_a_client_error(backend, structured_outputs):

133 params = SamplingParams(structured_outputs=structured_outputs)

134 with pytest.raises(VLLMClientError):

135 params._validate_structured_outputs(

136 - _StubModelConfig(is_diffusion=False), # type: ignore[arg-type]

136 + _StubModelConfig(is_diffusion=False),

137 StructuredOutputsConfig(backend=backend),

138 - tokenizer=object(), # type: ignore[arg-type]

138 + tokenizer=object(),

139 )

140

141

@@ -180,9 +180,9 @@ def test_auto_backend_falls_back_on_unsupported_schema(schema, expected_backend)

180 """`auto` falls back on rejection, so it must catch what the validators raise."""

181 params = SamplingParams(structured_outputs=StructuredOutputsParams(json=schema))

182 params._validate_structured_outputs(

183 - _StubModelConfig(is_diffusion=False), # type: ignore[arg-type]

183 + _StubModelConfig(is_diffusion=False),

184 StructuredOutputsConfig(backend="auto"),

185 - tokenizer=object(), # type: ignore[arg-type]

185 + tokenizer=object(),

186 )

187 assert params.structured_outputs is not None

188 assert params.structured_outputs._backend == expected_backend

tests/v1/worker/test_attn_utils.py

+5 -5

@@ -135,7 +135,7 @@ def test_attention_checks_preserve_global_and_target_scoped_support():

135 _TargetBackend,

136 ["target"],

137 spec,

138 - 0, # type: ignore[arg-type]

138 + 0,

139 )

140 target_group.metadata_builders = [

141 _FakeMetadataBuilder(AttentionCGSupport.ALWAYS) # type: ignore[list-item]

@@ -144,7 +144,7 @@ def test_attention_checks_preserve_global_and_target_scoped_support():

144 _DraftBackend,

145 ["draft"],

146 spec,

147 - 0, # type: ignore[arg-type]

147 + 0,

148 )

149 draft_group.metadata_builders = [

150 _FakeMetadataBuilder(AttentionCGSupport.UNIFORM_BATCH) # type: ignore[list-item]

@@ -152,14 +152,14 @@ def test_attention_checks_preserve_global_and_target_scoped_support():

152 groups = [[target_group, draft_group]]

153

154 # The runner-wide execution mode must still honor the drafter's limit.

155 - unfiltered = get_attn_cg_support(groups, None) # type: ignore[arg-type]

155 + unfiltered = get_attn_cg_support(groups, None)

156 assert unfiltered.min_cg_support == AttentionCGSupport.UNIFORM_BATCH

157 assert unfiltered.min_cg_attn_backend == "_DraftBackend"

158

159 # Adaptive verification validates only the target's varlen graphs.

160 target_only = get_attn_cg_support(

161 groups,

162 - None, # type: ignore[arg-type]

162 + None,

163 checked_layer_names={"target"},

164 )

165 assert target_only.min_cg_support == AttentionCGSupport.ALWAYS

@@ -176,7 +176,7 @@ def test_attention_checks_preserve_global_and_target_scoped_support():

176 draft_group.layer_names.append("target")

177 target_with_shared_group = get_attn_cg_support(

178 groups,

179 - None, # type: ignore[arg-type]

179 + None,

180 checked_layer_names={"target"},

181 )

182 assert target_with_shared_group.min_cg_support == AttentionCGSupport.UNIFORM_BATCH

tests/v1/worker/test_gpu_kv_connector.py

+4 -6

@@ -39,9 +39,7 @@ def _make_connector(

39 kv_role="kv_consumer",

40 kv_buffer_device="cpu",

41 )

42 - connector = ActiveKVConnector( # type: ignore[arg-type]

43 - SimpleNamespace(kv_transfer_config=kv_config), {}

44 - )

42 + connector = ActiveKVConnector(SimpleNamespace(kv_transfer_config=kv_config), {})

43 events.clear()

44 return connector

45

@@ -66,7 +64,7 @@ def test_load_start_phase(

64 request_indices = torch.tensor([3, 1])

65 request_ids = ["first", "second"]

66 attn_metadata = {"layer": object()}

69 - connector.pre_forward( # type: ignore[arg-type]

67 + connector.pre_forward(

68 output,

69 request_state_indices=request_indices,

70 request_ids=request_ids,

@@ -85,7 +83,7 @@ def test_load_start_phase(

83 assert kwargs["attn_metadata"] is attn_metadata

84

85 # A subsequent step without a forward must not reuse the prior batch.

88 - connector.no_forward(_scheduler_output(False)) # type: ignore[arg-type]

86 + connector.no_forward(_scheduler_output(False))

87 assert connector.kv_connector.start_load_kv.call_count == 2

88 assert connector.kv_connector.start_load_kv.call_args.kwargs == {}

89

@@ -94,6 +92,6 @@ def test_no_forward_starts_deferred_load_once(monkeypatch: pytest.MonkeyPatch):

92 events: list[str] = []

93 connector = _make_connector(monkeypatch, events)

94

97 - connector.no_forward(_scheduler_output(False)) # type: ignore[arg-type]

95 + connector.no_forward(_scheduler_output(False))

96

97 assert events == ["handle", "bind", "start", "wait", "clear"]

tests/v1/worker/test_gpu_worker_weight_transfer.py

+1 -1

@@ -75,7 +75,7 @@ def _make_worker(engine: _RecordingEngine | None) -> Worker:

75 def test_reload_weights_sets_current_config():

76 worker = _make_worker(None)

77 model_runner = _RecordingModelRunner()

78 - worker.model_runner = model_runner # type: ignore[assignment]

78 + worker.model_runner = model_runner

79

80 Worker.reload_weights(worker)

81

tests/v1/worker/test_late_interaction_runner.py

+1 -1

@@ -142,7 +142,7 @@ def test_invalid_query_uses_raises():

142 mode=LATE_INTERACTION_MODE_CACHE_QUERY,

143 query_key="query-bad",

144 )

145 - bad_meta.query_uses = "bad-int" # type: ignore[assignment]

145 + bad_meta.query_uses = "bad-int"

146 bad_query_params = _make_pooling_params(bad_meta)

147

148 with pytest.raises(ValueError, match="must be an integer value"):

Disable mypy `arg-type` and `assignment` checks in tests (#59428)