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

Signed-off-by: Martin Hickey <martin.hickey@ie.ibm.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>

tests/compile/correctness_e2e/test_sequence_parallel.py

@@ -348,10 +348,14 @@ def test_tp_sp_generation(
     if use_inductor_graph_partition and not is_torch_equal_or_newer("2.9.0.dev"):
         pytest.skip("inductor graph partition is only available in PyTorch 2.9+")
 
-    # Skip FP8 SP-only test on sm89 (compute capability 8.9)
+    # Skip FP8 SP-only test on sm89 (compute capability 8.9).
+    # An unknown capability (None) is left to fail in the test body rather than
+    # being silently skipped here.
+    capability = current_platform.get_device_capability()
     if (
         "fp8" in model_id.lower()
-        and current_platform.get_device_capability() < (9, 0)
+        and capability is not None
+        and capability < (9, 0)
         and (not fuse_gemm_comms)
     ):
         pytest.skip("FP8 reduction support begins with sm90 capable devices.")

tests/compile/fullgraph/test_full_cudagraph.py

@@ -9,7 +9,7 @@ import pytest
 from tests.utils import wait_for_gpu_memory_to_clear
 from tests.v1.attention.utils import full_cg_backend_configs as backend_configs
 from vllm import LLM, SamplingParams
-from vllm.config import CompilationConfig
+from vllm.config import CompilationConfig, CUDAGraphMode
 from vllm.platforms import current_platform
 from vllm.utils.torch_utils import is_torch_equal_or_newer
 from vllm.v1.attention.backends.registry import AttentionBackendEnum
@@ -97,7 +97,9 @@ def llm_pair(request):
             trust_remote_code=True,
             max_model_len=1024,
             max_num_seqs=128,
-            compilation_config=CompilationConfig(cudagraph_mode="PIECEWISE"),
+            compilation_config=CompilationConfig(
+                cudagraph_mode=CUDAGraphMode.PIECEWISE
+            ),
             generation_config="vllm",
             seed=42,
         )

tests/compile/fullgraph/test_full_graph.py

@@ -75,7 +75,7 @@ def test_full_graph(
     monkeypatch: pytest.MonkeyPatch,
     model: str,
     model_kwargs: dict[str, Any],
-    compilation_mode: int,
+    compilation_mode: CompilationMode,
 ):
     if (
         "w8a8" in model
@@ -197,7 +197,7 @@ def test_custom_compile_config(
     ],
 )
 def test_fp8_kv_scale_compile(
-    compilation_mode: int,
+    compilation_mode: CompilationMode,
     model: str,
     backend: AttentionBackendEnum | None,
 ):
@@ -213,7 +213,9 @@ def test_fp8_kv_scale_compile(
     run_model(compilation_mode, model, **model_kwargs)
 
 
-def run_model(compile_config: int | CompilationConfig, model: str, **model_kwargs):
+def run_model(
+    compile_config: CompilationMode | CompilationConfig, model: str, **model_kwargs
+):
     compilation_config = (
         compile_config
         if isinstance(compile_config, CompilationConfig)

tests/compile/fusions_e2e/conftest.py

@@ -12,7 +12,9 @@ from vllm.config import CompilationConfig, CompilationMode, CUDAGraphMode
 from .common import FUSION_LOG_PATTERNS, AttentionBackendCase, Matches
 
 
-def run_model(compile_config: int | CompilationConfig, model: str, **model_kwargs):
+def run_model(
+    compile_config: CompilationMode | CompilationConfig, model: str, **model_kwargs
+):
     """Run a model with the given compilation config for E2E fusion tests."""
     compilation_config = (
         compile_config
@@ -54,7 +56,7 @@ def run_model(compile_config: int | CompilationConfig, model: str, **model_kwarg
     )
 
     # Fetch match table from each worker via RPC and sum across workers.
-    worker_tables = llm.llm_engine.engine_core.collective_rpc(
+    worker_tables: list[dict[str, int]] = llm.llm_engine.engine_core.collective_rpc(
         "get_compilation_match_table"
     )
     combined: defaultdict[str, int] = defaultdict(int)

tests/config/test_bailing_mtp_config.py

@@ -1,6 +1,8 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+from typing import get_args
+
 from transformers import PretrainedConfig
 
 from vllm.config.speculative import MTPModelTypes, SpeculativeConfig
@@ -34,7 +36,7 @@ def test_bailing_hybrid_mtp_hf_config_override():
     assert overridden.model_type == "bailing_hybrid_mtp"
     assert overridden.architectures == ["BailingMoeV25MTPModel"]
     assert overridden.n_predict == 1
-    assert "bailing_hybrid_mtp" in MTPModelTypes.__args__
+    assert "bailing_hybrid_mtp" in get_args(MTPModelTypes)
 
 
 def test_bailing_hybrid_mtp_model_arch_config():

tests/config/test_multimodal_config.py

@@ -15,13 +15,13 @@ from vllm.v1.attention.backends.registry import AttentionBackendEnum
 
 
 def test_mm_encoder_attn_backend_str_conversion():
-    config = MultiModalConfig(mm_encoder_attn_backend="FLASH_ATTN")
+    config = MultiModalConfig(mm_encoder_attn_backend="FLASH_ATTN")  # type: ignore[arg-type]
     assert config.mm_encoder_attn_backend == AttentionBackendEnum.FLASH_ATTN
 
 
 def test_mm_encoder_attn_backend_invalid():
     with pytest.raises(ValueError):
-        MultiModalConfig(mm_encoder_attn_backend="not_a_backend")
+        MultiModalConfig(mm_encoder_attn_backend="not_a_backend")  # type: ignore[arg-type]
 
 
 def test_mm_encoder_attn_backend_hash_updates():

tests/config/test_speculative_draft_max_position_embeddings.py

@@ -14,6 +14,7 @@ methods only.
 """
 
 import logging
+from typing import Literal
 
 import pytest
 from transformers import PretrainedConfig
@@ -77,7 +78,7 @@ def test_override_ignores_missing_attribute(vllm_caplog: pytest.LogCaptureFixtur
 @pytest.mark.cpu_test
 @pytest.mark.parametrize("method", ["eagle", "eagle3"])
 def test_eagle_draft_inherits_target_max_model_len(
-    method: str, vllm_caplog: pytest.LogCaptureFixture
+    method: Literal["eagle", "eagle3"], vllm_caplog: pytest.LogCaptureFixture
 ):
     target_model_config = ModelConfig(LLAMA3_TARGET)
     assert target_model_config.max_model_len > 2048

tests/entrypoints/generate/generative_scoring/test_generative_scoring.py

@@ -285,7 +285,7 @@ class TestPromptBuilding:
         )
 
         for i, exp in enumerate(expected):
-            assert engine_inputs[i]["prompt_token_ids"] == exp
+            assert engine_inputs[i]["prompt_token_ids"] == exp  # type: ignore[typeddict-item]
 
 
 class TestGeneration:

tests/entrypoints/tool_parsers/test_hermes_tool_parser.py

@@ -45,7 +45,7 @@ class ServerConfig(TypedDict, total=False):
     model: str
     arguments: list[str]
     model_arg: str
-    tool_parser: ToolParser
+    tool_parser: type[ToolParser]
 
 
 CONFIGS: dict[str, ServerConfig] = {

tests/entrypoints/weight_transfer/test_weight_transfer_llm.py

@@ -329,4 +329,5 @@ def test_weight_transfer_config_backend():
     )
 
     config = llm.llm_engine.vllm_config.weight_transfer_config
+    assert config is not None
     assert config.backend == "nccl"

tests/kernels/core/test_fused_qk_norm_rope.py

@@ -35,10 +35,12 @@ def _apply_qk_norm_rope(
 
     q_by_head = q.view(*q.shape[:-1], q.shape[-1] // head_dim, head_dim)
     q_by_head = q_norm.forward_native(q_by_head)
+    assert isinstance(q_by_head, torch.Tensor)
     q = q_by_head.view(q.shape)
 
     k_by_head = k.view(*k.shape[:-1], k.shape[-1] // head_dim, head_dim)
     k_by_head = k_norm.forward_native(k_by_head)
+    assert isinstance(k_by_head, torch.Tensor)
     k = k_by_head.view(k.shape)
 
     q, k = rope.forward_native(positions, q, k)

tests/kernels/core/test_vit_fp8_attn.py

@@ -34,10 +34,9 @@ NUM_HEADS = [16]
 @pytest.fixture
 def _fp8_attention():
     """Create FP8-enabled MMEncoderAttention via config."""
-    from types import SimpleNamespace
-    from unittest.mock import patch
+    from unittest.mock import MagicMock, patch
 
-    from vllm.config import VllmConfig, set_current_vllm_config
+    from vllm.config import ModelConfig, VllmConfig, set_current_vllm_config
     from vllm.config.multimodal import MultiModalConfig
 
     if not is_flashinfer_cudnn_fp8_prefill_attn_supported():
@@ -45,7 +44,7 @@ def _fp8_attention():
 
     mm_config = MultiModalConfig(mm_encoder_attn_dtype="fp8")
     vllm_config = VllmConfig()
-    vllm_config.model_config = SimpleNamespace(multimodal_config=mm_config)
+    vllm_config.model_config = MagicMock(spec=ModelConfig, multimodal_config=mm_config)
 
     # MMEncoderAttention reads torch.get_default_dtype() during init
     # to determine the output dtype. In real model loading this is bf16.

tests/kernels/core/test_vit_fp8_scaling.py

@@ -4,8 +4,7 @@
 
 import contextlib
 import json
-from types import SimpleNamespace
-from unittest.mock import patch
+from unittest.mock import MagicMock, patch
 
 import pytest
 import torch
@@ -48,7 +47,7 @@ def _build_attention(mm_config, attn_backend=None):
     invokes ``process_weights_after_loading`` to simulate the model loader's
     auto-scan. Yields ``None`` if FlashInfer cuDNN is not available.
     """
-    from vllm.config import VllmConfig, set_current_vllm_config
+    from vllm.config import ModelConfig, VllmConfig, set_current_vllm_config
     from vllm.model_executor.layers.attention.mm_encoder_attention import (
         MMEncoderAttention,
     )
@@ -65,7 +64,7 @@ def _build_attention(mm_config, attn_backend=None):
         return
 
     vllm_config = VllmConfig()
-    vllm_config.model_config = SimpleNamespace(multimodal_config=mm_config)
+    vllm_config.model_config = MagicMock(spec=ModelConfig, multimodal_config=mm_config)
 
     with (
         set_current_vllm_config(vllm_config),
@@ -212,7 +211,7 @@ def test_aiter_static_scales_loaded(tmp_path) -> None:
 
 def test_static_scales_missing_layer(tmp_path) -> None:
     """Verify error when requested layer is not in the scale file."""
-    from vllm.config import VllmConfig, set_current_vllm_config
+    from vllm.config import ModelConfig, VllmConfig, set_current_vllm_config
     from vllm.config.multimodal import MultiModalConfig
     from vllm.v1.attention.backends.registry import AttentionBackendEnum
 
@@ -228,7 +227,7 @@ def test_static_scales_missing_layer(tmp_path) -> None:
         mm_encoder_fp8_scale_path=str(scale_file),
     )
     vllm_config = VllmConfig()
-    vllm_config.model_config = SimpleNamespace(multimodal_config=mm_config)
+    vllm_config.model_config = MagicMock(spec=ModelConfig, multimodal_config=mm_config)
 
     from vllm.model_executor.layers.attention.mm_encoder_attention import (
         MMEncoderAttention,

tests/kernels/mamba/cpu/test_cpu_gdn_ops.py

@@ -12,6 +12,7 @@ import vllm._custom_ops as ops
 from vllm.model_executor.layers.mamba.ops.cpu import gdn_attention
 from vllm.platforms import current_platform
 from vllm.utils.torch_utils import set_random_seed
+from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadata
 
 if not current_platform.is_cpu():
     pytest.skip("skipping CPU-only tests", allow_module_level=True)
@@ -452,12 +453,17 @@ def test_spec_aware_mixed_routing_preserves_token_order(
 
     spec_indices = torch.tensor([0, 2])
     nonspec_indices = torch.tensor([1, 3])
-    metadata = types.SimpleNamespace(
+    metadata = GDNAttentionMetadata(
+        num_prefills=1,
+        num_prefill_tokens=0,
+        num_decodes=0,
+        num_decode_tokens=0,
+        num_spec_decodes=0,
+        num_spec_decode_tokens=0,
+        num_actual_tokens=0,
         spec_sequence_masks=torch.ones(1, dtype=torch.bool),
         spec_token_indx=spec_indices,
         non_spec_token_indx=nonspec_indices,
-        num_prefills=1,
-        num_decodes=0,
     )
     routed = []
 
@@ -501,13 +507,16 @@ def test_spec_aware_nonspec_materializes_state_indices(
     state_indices = block_table[:, 0]
     assert not state_indices.is_contiguous()
 
-    metadata = types.SimpleNamespace(
-        non_spec_state_indices_tensor=state_indices,
-        non_spec_query_start_loc=torch.tensor([0, 2, 4], dtype=torch.int32),
-        num_decodes=0,
-        num_decode_tokens=0,
+    metadata = GDNAttentionMetadata(
         num_prefills=2,
         num_prefill_tokens=4,
+        num_decodes=0,
+        num_decode_tokens=0,
+        num_spec_decodes=0,
+        num_spec_decode_tokens=0,
+        num_actual_tokens=4,
+        non_spec_state_indices_tensor=state_indices,
+        non_spec_query_start_loc=torch.tensor([0, 2, 4], dtype=torch.int32),
         has_initial_state=torch.tensor([False, False]),
     )
 
@@ -575,8 +584,14 @@ def test_spec_forward_prepares_native_conv_metadata(
     state_indices = block_table[:, 0]
     accepted_counts = torch.tensor([1, 4], dtype=torch.int32)
     assert not state_indices.is_contiguous()
-    metadata = types.SimpleNamespace(
+    metadata = GDNAttentionMetadata(
+        num_prefills=0,
+        num_prefill_tokens=0,
+        num_decodes=0,
+        num_decode_tokens=0,
         num_spec_decodes=2,
+        num_spec_decode_tokens=8,
+        num_actual_tokens=8,
         spec_state_indices_tensor=block_table,
         spec_query_start_loc=torch.tensor([0, 4, 8], dtype=torch.int32),
         num_accepted_tokens=accepted_counts,

tests/kernels/mamba/test_gdn_forward_core_split.py

@@ -174,7 +174,9 @@ def test_forward_core_split_matches_unified(
     batch = BatchSpec(seq_lens=seq_lens, query_lens=query_lens)
 
     builder = GDNAttentionMetadataBuilder(
-        kv_cache_spec=MambaSpec(
+        # GDNAttentionMetadataBuilder declares kv_cache_spec as AttentionSpec but
+        # asserts isinstance(kv_cache_spec, MambaSpec). MambaSpec is correct here.
+        kv_cache_spec=MambaSpec(  # type: ignore[arg-type]
             block_size=BLOCK_SIZE, shapes=((16, 64),), dtypes=(torch.float16,)
         ),
         layer_names=[PREFIX],
@@ -198,6 +200,7 @@ def test_forward_core_split_matches_unified(
     # Full-batch chunk metadata for the unified reference path, built the same
     # way the builder would for a non-split batch (backend-matched).
     cu_full = meta_split.non_spec_query_start_loc
+    assert cu_full is not None
     if builder.gdn_prefill_backend == "cutedsl":
         from vllm.model_executor.layers.mamba.ops.gdn_chunk_cutedsl import (
             prepare_metadata_cutedsl,
@@ -227,6 +230,7 @@ def test_forward_core_split_matches_unified(
     )
 
     # Size the state pools from the indices the builder actually produced.
+    assert meta_split.non_spec_state_indices_tensor is not None
     pool_size = int(meta_split.non_spec_state_indices_tensor.max().item()) + 1
     conv_state_shape, temporal_state_shape = (
         MambaStateShapeCalculator.gated_delta_net_state_shape(

tests/kernels/mamba/test_precopy_mamba_align.py

@@ -21,7 +21,9 @@ The kernel must also no-op when ``src_col < 0`` (fresh request) or
 
 from __future__ import annotations
 
+from collections.abc import Callable
 from types import SimpleNamespace
+from typing import Any
 
 import numpy as np
 import torch
@@ -31,6 +33,8 @@ from vllm.platforms import current_platform
 from vllm.v1.worker import mamba_utils as worker_mamba_utils
 from vllm.v1.worker.mamba_utils import precopy_mamba_align_fused_kernel
 
+_parametrize: Callable[..., Callable[[Any], Any]]
+
 try:
     import pytest
 
@@ -40,17 +44,18 @@ try:
     )
     _parametrize = pytest.mark.parametrize
 except ModuleNotFoundError:  # allow running directly as ``python <thisfile>``
-    pytest = None
 
     def _cuda_required(fn):
         return fn
 
-    def _parametrize(_name, _values):
+    def _no_parametrize(_name, _values):
         def _deco(fn):
             return fn
 
         return _deco
 
+    _parametrize = _no_parametrize
+
 
 NUM_LAYERS = 3
 CONV_WIDTH = 4  # conv_kernel - 1 + num_spec
@@ -308,8 +313,10 @@ def _make_preprocess_case(token_bias):
 def test_preprocess_fused_align_matches_scalar_bookkeeping(monkeypatch, token_bias):
     block_size = 4
     mamba_spec = SimpleNamespace(block_size=block_size, num_speculative_blocks=1)
-    cache_config = SimpleNamespace(enable_prefix_caching=True)
-    kv_cache_config = SimpleNamespace()
+    # Fakes stand in for real config/context objects. The fused path tests
+    # below only touch the attributes preprocess_mamba actually reads.
+    cache_config: Any = SimpleNamespace(enable_prefix_caching=True)
+    kv_cache_config: Any = SimpleNamespace()
     scalar_copy_calls = []
 
     def fake_collect(
@@ -340,7 +347,7 @@ def test_preprocess_fused_align_matches_scalar_bookkeeping(monkeypatch, token_bi
     scalar_case = _make_preprocess_case(token_bias)
     fused_case = _make_preprocess_case(token_bias)
 
-    scalar_copy_bufs = SimpleNamespace(
+    scalar_copy_bufs: Any = SimpleNamespace(
         mamba_group_ids=[0],
         mamba_spec=mamba_spec,
         offset=0,
@@ -357,8 +364,8 @@ def test_preprocess_fused_align_matches_scalar_bookkeeping(monkeypatch, token_bi
         copy_bufs=scalar_copy_bufs,
     )
 
-    ctx = _FakePrecopyContext(len(fused_case[1].req_ids))
-    fused_copy_bufs = SimpleNamespace(
+    ctx: Any = _FakePrecopyContext(len(fused_case[1].req_ids))
+    fused_copy_bufs: Any = SimpleNamespace(
         mamba_group_ids=[0],
         mamba_spec=mamba_spec,
         offset=0,

tests/models/language/generation/test_mistral.py

@@ -261,7 +261,7 @@ def test_mistral_function_calling(vllm_runner, model: str, dtype: str) -> None:
 
         model_output = outputs[0].outputs[0].text.strip()
         assert model_output.startswith(tool_parser.bot_token), model_output
-        parsed_message = tool_parser.extract_tool_calls(model_output, None)
+        parsed_message = tool_parser.extract_tool_calls(model_output, None)  # type: ignore[arg-type]
 
         assert parsed_message.tools_called
 
@@ -304,7 +304,7 @@ def test_mistral_function_call_nested_json():
 
     model_output = f"{parser.bot_token}get_current_weather{json.dumps(args_dict)}"
 
-    parsed = parser.extract_tool_calls(model_output, None)
+    parsed = parser.extract_tool_calls(model_output, None)  # type: ignore[arg-type]
 
     # Assertions: the tool call is detected and the full nested JSON is parsed
     # without truncation.
@@ -337,7 +337,7 @@ def test_mistral_function_call_nested_json():
         ]
     )
 
-    parsed = parser.extract_tool_calls(model_output, None)
+    parsed = parser.extract_tool_calls(model_output, None)  # type: ignore[arg-type]
 
     # Assertions: the tool call is detected and the full nested JSON is parsed
     # without truncation.

tests/models/quantization/test_bitsandbytes.py

@@ -323,7 +323,7 @@ def test_bitsandbytes_passes_revision_by_name():
             return_value=["/folder/model.safetensors"],
         ),
     ):
-        bnb.BitsAndBytesModelLoader._prepare_weights(fake_self, "org/model", "myrev")
+        bnb.BitsAndBytesModelLoader._prepare_weights(fake_self, "org/model", "myrev")  # type: ignore[arg-type]
 
     mock_idx.assert_called_once()
     assert mock_idx.call_args.kwargs.get("revision") == "myrev"

tests/plugins/bge_m3_sparse_plugin/bge_m3_sparse_processor/sparse_embeddings_processor.py

@@ -29,7 +29,7 @@ class BgeM3SparseEmbeddingsProcessor(
         self.online_requests: dict[str, SparseEmbeddingCompletionRequestMixin] = {}
         self.renderer: BaseRenderer = renderer
         self.default_pooling_params = {}
-        pooler_config: PoolerConfig = vllm_config.model_config.pooler_config
+        pooler_config: PoolerConfig | None = vllm_config.model_config.pooler_config
         if pooler_config is not None:
             for param in ["use_activation", "dimensions"]:
                 if getattr(pooler_config, param, None) is None:
@@ -91,11 +91,11 @@ class BgeM3SparseEmbeddingsProcessor(
     ) -> list[SparseEmbeddingTokenWeight]:
         token_ids = sparse_embedding.keys()
         token_weights = sparse_embedding.values()
-        tokens = [None] * len(token_ids)
+        tokens: Sequence[str | None] = [None] * len(token_ids)
 
         if return_tokens and self.renderer is not None:
             tokens = convert_ids_list_to_tokens(
-                self.renderer.get_tokenizer(), token_ids
+                self.renderer.get_tokenizer(), list(token_ids)
             )
         sparse_embedding_output: list[SparseEmbeddingTokenWeight] = []
         for token_id, weight, token in zip(token_ids, token_weights, tokens):

tests/plugins/prithvi_io_processor_plugin/prithvi_io_processor/prithvi_processor.py

@@ -170,7 +170,7 @@ def load_image(
         list of meta info for each image in *file_paths*
     """
 
-    imgs = []
+    imgs: list[np.ndarray] = []
     metas = []
     temporal_coords = []
     location_coords = []
@@ -209,11 +209,11 @@ def load_image(
         except Exception:
             logger.exception("Could not extract timestamp for %s", file)
 
-    imgs = np.stack(imgs, axis=0)  # num_frames, H, W, C
-    imgs = np.moveaxis(imgs, -1, 0).astype("float32")  # C, num_frames, H, W
-    imgs = np.expand_dims(imgs, axis=0)  # add batch di
+    stacked = np.stack(imgs, axis=0)  # num_frames, H, W, C
+    stacked = np.moveaxis(stacked, -1, 0).astype("float32")  # C, num_frames, H, W
+    stacked = np.expand_dims(stacked, axis=0)  # add batch di
 
-    return imgs, temporal_coords, location_coords, metas
+    return stacked, temporal_coords, location_coords, metas
 
 
 class PrithviMultimodalDataProcessor(IOProcessor[ImagePrompt, ImageRequestOutput]):
@@ -327,7 +327,7 @@ class PrithviMultimodalDataProcessor(IOProcessor[ImagePrompt, ImageRequestOutput
                 }
             )
 
-        return prompts
+        return prompts  # type: ignore[return-value]
 
     def post_process(
         self,

tests/plugins/vllm_add_dummy_platform/vllm_add_dummy_platform/dummy_platform.py

@@ -6,6 +6,8 @@ from vllm.platforms.interface import Platform, PlatformEnum
 
 if TYPE_CHECKING:
     from vllm.config import VllmConfig
+    from vllm.v1.attention.backends.registry import AttentionBackendEnum
+    from vllm.v1.attention.selector import AttentionSelectorConfig
 else:
     VllmConfig = None
 
@@ -20,16 +22,11 @@ class DummyPlatform(Platform):
     def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
         vllm_config.compilation_config.custom_ops = ["all"]
 
+    @classmethod
     def get_attn_backend_cls(
-        self,
-        backend_name,
-        head_size,
-        dtype,
-        kv_cache_dtype,
-        block_size,
-        use_mla,
-        has_sink,
-        use_sparse,
-        use_mm_prefix,
-    ):
+        cls,
+        selected_backend: "AttentionBackendEnum",
+        attn_selector_config: "AttentionSelectorConfig",
+        num_heads: int | None = None,
+    ) -> str:
         return "vllm_add_dummy_platform.dummy_attention_backend.DummyAttentionBackend"  # noqa E501

tests/plugins/vllm_add_dummy_stat_logger/dummy_stat_logger/dummy_stat_logger.py

@@ -17,7 +17,9 @@ class DummyStatLogger(StatLoggerBase):
         self.logged = False
         self.engine_initialized = False
 
-    def record(self, scheduler_stats, iteration_stats, mm_cache_stats, engine_idx):
+    def record(
+        self, scheduler_stats, iteration_stats, mm_cache_stats=None, engine_idx=0
+    ):
         self.recorded.append(
             (scheduler_stats, iteration_stats, mm_cache_stats, engine_idx)
         )

tests/plugins_tests/gguf/test_gguf_plugin_multimodal.py

@@ -92,7 +92,7 @@ def run_multimodal_gguf_test(
     num_logprobs: int,
 ):
     # Load images at runtime (inside subprocess) to avoid pickle issues
-    images = [ImageAsset(name).pil_image for name in model.image_names]
+    images = [ImageAsset(name).pil_image for name in model.image_names]  # type: ignore[arg-type]
     size_factors = [0.25, 0.5, 1.0]
     inputs_per_image = [
         (

tests/plugins_tests/lora_resolvers/test_hf_hub_resolver.py

@@ -31,6 +31,7 @@ async def test_hf_resolver_with_direct_path():
     assert hf_resolver is not None
 
     lora_request = await hf_resolver.resolve_lora(LORA_REPO_MODEL_NAME, LORA_REPO)
+    assert lora_request is not None
     assert lora_request.lora_name == LORA_REPO
     assert REPO_DOWNLOAD_DIR in lora_request.lora_path
     assert "adapter_config.json" in os.listdir(lora_request.lora_path)

tests/spec_decode/test_custom_proposer.py

@@ -30,6 +30,7 @@ class DummyDraftProposer:
         Args:
             vllm_config: vLLM configuration containing model and speculative settings.
         """
+        assert vllm_config.speculative_config is not None
         self.num_speculative_tokens = (
             vllm_config.speculative_config.num_speculative_tokens
         )

tests/transformers_utils/test_processor.py

@@ -44,7 +44,7 @@ class _ProcWithUnpack:
 
 def test_get_processor_kwargs_from_processor_unpack_path_returns_full_union():
     proc = _ProcWithUnpack()
-    keys = get_processor_kwargs_keys(get_processor_kwargs_type(proc))
+    keys = get_processor_kwargs_keys(get_processor_kwargs_type(proc))  # type: ignore[arg-type]
     _assert_has_all_expected(keys)
 
 
@@ -63,5 +63,5 @@ def test_get_processor_kwargs_from_processor_module_scan_returns_full_union():
     assert hasattr(mod, "_FakeProcessorKwargs")
 
     proc = _ProcWithoutUnpack()
-    keys = get_processor_kwargs_keys(get_processor_kwargs_type(proc))
+    keys = get_processor_kwargs_keys(get_processor_kwargs_type(proc))  # type: ignore[arg-type]
     _assert_has_all_expected(keys)

tests/v1/distributed/test_async_llm_dp.py

@@ -12,13 +12,14 @@ import pytest
 
 from vllm import SamplingParams
 from vllm.config import VllmConfig
+from vllm.config.parallel import DataParallelBackend
 from vllm.engine.arg_utils import AsyncEngineArgs
 from vllm.inputs import PromptType
 from vllm.outputs import RequestOutput
 from vllm.platforms import current_platform
 from vllm.sampling_params import RequestOutputKind
 from vllm.v1.engine.async_llm import AsyncLLM
-from vllm.v1.engine.core_client import DPAsyncMPClient
+from vllm.v1.engine.core_client import DPLBAsyncMPClient
 from vllm.v1.metrics.loggers import StatLoggerBase
 from vllm.v1.metrics.stats import IterationStats, MultiModalCacheStats, SchedulerStats
 
@@ -82,7 +83,7 @@ async def generate(
 async def test_load(
     model: str,
     output_kind: RequestOutputKind,
-    data_parallel_backend: str,
+    data_parallel_backend: DataParallelBackend,
     async_scheduling: bool,
 ):
     if async_scheduling and data_parallel_backend == "ray":
@@ -162,7 +163,8 @@ async def test_load(
         assert not engine.output_processor.has_unfinished_requests()
 
         # testing internals here which may break
-        core_client: DPAsyncMPClient = engine.engine_core
+        core_client = engine.engine_core
+        assert isinstance(core_client, DPLBAsyncMPClient)
         # the engines only synchronize stopping every N steps so
         # allow a small amount of time here.
         for _ in range(10):

tests/v1/distributed/test_eagle_dp.py

@@ -61,7 +61,7 @@ async def test_run_eagle_dp(monkeypatch: pytest.MonkeyPatch, attn_backend: str):
         data_parallel_backend="mp",  # ray takes more time
         trust_remote_code=True,
         max_model_len=16384,
-        attention_config={"backend": attn_backend},
+        attention_config={"backend": attn_backend},  # type: ignore[arg-type]
     )
 
     eagle_engine_args = replace(

tests/v1/shutdown/test_processor_error.py

@@ -9,7 +9,7 @@ import pytest
 from tests.v1.shutdown.utils import SHUTDOWN_TEST_TIMEOUT_SEC
 from vllm import SamplingParams
 from vllm.engine.arg_utils import AsyncEngineArgs
-from vllm.inputs import TokensPrompt
+from vllm.inputs import ExplicitEncoderDecoderPrompt
 from vllm.sampling_params import RequestOutputKind
 from vllm.v1.engine.async_llm import AsyncLLM
 from vllm.v1.engine.exceptions import EngineGenerateError
@@ -29,9 +29,14 @@ async def test_async_llm_processor_error(model: str) -> None:
     async_llm = AsyncLLM.from_engine_args(engine_args)
 
     async def generate(request_id: str):
-        # [] is not allowed and will raise a ValueError in Processor.
+        # An encoder/decoder prompt is rejected by the Processor for a
+        # decoder only model.
         generator = async_llm.generate(
-            TokensPrompt([]), request_id=request_id, sampling_params=SamplingParams()
+            ExplicitEncoderDecoderPrompt(
+                encoder_prompt="Hello my name is", decoder_prompt=None
+            ),
+            request_id=request_id,
+            sampling_params=SamplingParams(),
         )
         try:
             async for _ in generator:

tools/pre_commit/mypy.py

@@ -21,34 +21,55 @@ import sys
 
 import regex as re
 
+# Paths verified clean under follow_imports="silent". Matched before
+# SEPARATE_GROUPS, so these files join the default group and are checked at the
+# stricter setting even while a parent directory remains in SEPARATE_GROUPS.
+#
+# Fixing a directory means moving it from SEPARATE_GROUPS to here. Without that
+# move the fixes are not enforced because "tests" claims every file below it.
+SILENT_GROUPS = [
+    "tests/compile/correctness_e2e",
+    "tests/compile/fullgraph",
+    "tests/compile/fusions_e2e",
+    "tests/config",
+    "tests/entrypoints/generate",
+    "tests/entrypoints/tool_parsers",
+    "tests/entrypoints/weight_transfer",
+    "tests/kernels/core",
+    "tests/kernels/mamba",
+    "tests/models/language",
+    "tests/models/quantization",
+    "tests/plugins/bge_m3_sparse_plugin",
+    "tests/plugins/prithvi_io_processor_plugin",
+    "tests/plugins/vllm_add_dummy_platform",
+    "tests/plugins/vllm_add_dummy_stat_logger",
+    "tests/plugins_tests/gguf",
+    "tests/plugins_tests/lora_resolvers",
+    "tests/spec_decode",
+    "tests/transformers_utils",
+    "tests/v1/distributed",
+    "tests/v1/shutdown",
+]
+
 # After fixing errors resulting from changing follow_imports
-# from "skip" to "silent", remove its directory from SEPARATE_GROUPS.
+# from "skip" to "silent", move its directory to SILENT_GROUPS.
 SEPARATE_GROUPS = [
     "tests",
     "tests/benchmarks",
-    "tests/compile/correctness_e2e",
-    "tests/config",
     "tests/compile",
-    "tests/compile/fullgraph",
-    "tests/compile/fusions_e2e",
     "tests/compile/passes",
     "tests/distributed",
     "tests/entrypoints/anthropic",
-    "tests/entrypoints/generate",
     "tests/entrypoints/llm",
     "tests/entrypoints/multimodal",
     "tests/entrypoints/openai",
     "tests/entrypoints/pooling",
     "tests/entrypoints/serve",
     "tests/entrypoints/speech_to_text",
-    "tests/entrypoints/tool_parsers",
     "tests/entrypoints/unit_tests",
-    "tests/entrypoints/weight_transfer",
     "tests/kernels",
     "tests/kernels/attention",
-    "tests/kernels/core",
     "tests/kernels/helion",
-    "tests/kernels/mamba",
     "tests/kernels/moe",
     "tests/kernels/quantization",
     "tests/lora",
@@ -57,34 +78,23 @@ SEPARATE_GROUPS = [
     "tests/model_executor/model_loader",
     "tests/models",
     "tests/models/test_initialization.py",
-    "tests/models/language",
     "tests/models/multimodal",
-    "tests/models/quantization",
     "tests/multimodal",
     "tests/parser",
-    "tests/plugins_tests/gguf",
-    "tests/plugins_tests/lora_resolvers",
-    "tests/plugins/bge_m3_sparse_plugin",
-    "tests/plugins/prithvi_io_processor_plugin",
-    "tests/plugins/vllm_add_dummy_platform",
-    "tests/plugins/vllm_add_dummy_stat_logger",
     "tests/plugins_tests",
     "tests/quantization",
     "tests/reasoning",
     "tests/renderers",
     "tests/samplers",
-    "tests/spec_decode",
     "tests/tokenizers_",
     "tests/tool_parsers",
     "tests/tool_use",
-    "tests/transformers_utils",
     "tests/utils_",
     "tests/v1",
     "tests/v1/attention",
     "tests/v1/core",
     "tests/v1/cudagraph",
     "tests/v1/determinism",
-    "tests/v1/distributed",
     "tests/v1/e2e",
     "tests/v1/ec_connector",
     "tests/v1/engine",
@@ -94,7 +104,6 @@ SEPARATE_GROUPS = [
     "tests/v1/logits_processors",
     "tests/v1/metrics",
     "tests/v1/sample",
-    "tests/v1/shutdown",
     "tests/v1/simple_kv_offload",
     "tests/v1/spec_decode",
     "tests/v1/streaming_input",
@@ -142,14 +151,22 @@ def group_files(changed_files: list[str]) -> dict[str, list[str]]:
         A dictionary mapping file group names to lists of changed files.
     """
     exclude_pattern = re.compile(f"^{'|'.join(EXCLUDE)}.*")
-    file_groups = {"": []}
+    silent_pattern = re.compile(f"^({'|'.join(SILENT_GROUPS)}).*")
+    file_groups: dict[str, list[str]] = {"": []}
     file_groups.update({k: [] for k in SEPARATE_GROUPS})
+    # Longest path first so a sub-directory is not shadowed by its parent
+    separate_groups = sorted(SEPARATE_GROUPS, key=len, reverse=True)
     for changed_file in changed_files:
         # Skip files which should be ignored completely
         if exclude_pattern.match(changed_file):
             continue
+        # Already-fixed paths go in the default group, which runs at the
+        # stricter follow_imports setting from pyproject.toml
+        if silent_pattern.match(changed_file):
+            file_groups[""].append(changed_file)
+            continue
         # Group files by mypy call
-        for directory in SEPARATE_GROUPS:
+        for directory in separate_groups:
             if re.match(f"^{directory}.*", changed_file):
                 file_groups[directory].append(changed_file)
                 break