[CI] Fix mypy multimodels tests folder

tests/models/multimodal/conftest.py

@@ -10,4 +10,4 @@
 # longer re-exports the typing aliases, so executing that module raises NameError.
 # TODO: Remove once the checkpoint fixes its imports.
 if not hasattr(F, "List"):
-    F.List = list
+    F.List = list  # type: ignore[attr-defined]

tests/models/multimodal/generation/test_common.py

@@ -697,7 +697,7 @@ def _granite4_vision_vllm_to_hf_output(vllm_output, model):
         img_idx_to_prompt=lambda idx: "(<image>./</image>)\n",
         max_model_len=4096,
         max_num_seqs=2,
-        get_stop_token_ids=lambda tok: [tok.eos_id, tok.eot_id],
+        get_stop_token_ids=lambda tok: [getattr(tok, "eos_id"), getattr(tok, "eot_id")],  # noqa: B009
         hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
         patch_hf_runner=model_utils.minicpmv_25_patch_hf_runner,
     ),
@@ -927,7 +927,7 @@ def _granite4_vision_vllm_to_hf_output(vllm_output, model):
         auto_cls=AutoModelForImageTextToText,
         vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
         marks=multi_gpu_marks(num_gpus=2),
-        **COMMON_BROADCAST_SETTINGS,  # type: ignore
+        **COMMON_BROADCAST_SETTINGS,
     ),
     "llava_next-broadcast": VLMTestInfo(
         models=["llava-hf/llava-v1.6-mistral-7b-hf"],
@@ -936,7 +936,7 @@ def _granite4_vision_vllm_to_hf_output(vllm_output, model):
         auto_cls=AutoModelForImageTextToText,
         vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
         marks=multi_gpu_marks(num_gpus=2),
-        **COMMON_BROADCAST_SETTINGS,  # type: ignore
+        **COMMON_BROADCAST_SETTINGS,
     ),
     ### Custom input edge-cases for specific models
     "intern_vl-diff-patches": VLMTestInfo(

tests/models/multimodal/generation/test_granite_speech.py

@@ -169,7 +169,7 @@ def test_models(
         hf_runner,
         vllm_runner,
         [
-            ([HF_AUDIO_PROMPT], [audio]),
+            ([HF_AUDIO_PROMPT], [(audio, int(sr))]),
         ],
         model,
         dtype=dtype,

tests/models/multimodal/generation/test_maverick.py

@@ -523,7 +523,7 @@ def save_weights_to_safetensors(
             print(f"Saved shard {i + 1}/{len(shards)}: {filename}")
 
     # Create index file
-    index_data = {
+    index_data: dict[str, Any] = {
         "metadata": {
             "total_size": sum(
                 tensor.numel() * tensor.element_size() for tensor in weights.values()

tests/models/multimodal/generation/test_mm_prefix_lm.py

@@ -50,7 +50,7 @@ def _get_hf_prefill_hidden(hf_model: HfRunner, image: Any):
     inputs = hf_model.get_inputs([PROMPT], images=[image])[0]
     with torch.no_grad():
         outputs = hf_model.model.model(
-            **hf_model.wrap_device(inputs),
+            **hf_model.wrap_device(dict(inputs)),
             use_cache=False,
         )
     return outputs.last_hidden_state[0].detach().float().cpu()

tests/models/multimodal/generation/test_nemotron_parse_35.py

@@ -35,7 +35,7 @@ def generate_hf_logprobs_from_logits(
     outputs = []
     for model_inputs in inputs:
         generated = hf_model.model.generate(
-            **hf_model.wrap_device(model_inputs),
+            **hf_model.wrap_device(dict(model_inputs)),
             do_sample=False,
             max_new_tokens=max_tokens,
             output_logits=True,
@@ -51,7 +51,9 @@ def generate_hf_logprobs_from_logits(
                 }
             )
         token_ids = generated.sequences[0, -len(logprobs) :].tolist()
-        outputs.append((token_ids, hf_model.tokenizer.decode(token_ids), logprobs))
+        decoded = hf_model.tokenizer.decode(token_ids)
+        assert isinstance(decoded, str)
+        outputs.append((token_ids, decoded, logprobs))
     return outputs
 
 

tests/models/multimodal/generation/test_nemotron_parse_35_mtp.py

@@ -23,10 +23,11 @@
 
 
 def test_parse_mtp_config_routes_only_causal_prefix_architecture() -> None:
-    config = PreTrainedConfig(
-        architectures=["RadioMBartPrefixForCausalLM"],
-        decoder=PreTrainedConfig(causal_vision_prefix=True, vocab_size=8, d_model=4),
-    )
+    config = PreTrainedConfig(architectures=["RadioMBartPrefixForCausalLM"])
+    config.decoder = PreTrainedConfig()
+    config.decoder.causal_vision_prefix = True
+    config.decoder.vocab_size = 8
+    config.decoder.d_model = 4
 
     draft = SpeculativeConfig.hf_config_override(config)
 

tests/models/multimodal/generation/test_qwen2_vl.py

@@ -126,6 +126,7 @@ def batch_make_image_embeddings(
     image_processor = processor.image_processor
 
     mm_config = llm.get_llm().model_config.multimodal_config
+    assert mm_config is not None
     mm_kwargs = mm_config.merge_mm_processor_kwargs({})
     preprocess_result = image_processor.preprocess(
         images=images, return_tensors="pt", **mm_kwargs
@@ -211,6 +212,7 @@ def batch_make_video_embeddings(
     video_processor = processor.video_processor
 
     mm_config = llm.get_llm().model_config.multimodal_config
+    assert mm_config is not None
     mm_kwargs = mm_config.merge_mm_processor_kwargs({})
     preprocess_result = video_processor.preprocess(
         videos=videos, return_tensors="pt", **mm_kwargs

tests/models/multimodal/generation/test_ultravox.py

@@ -24,7 +24,7 @@
 
 MULTI_AUDIO_PROMPT = "Describe each of the audios above."
 
-AudioTuple = tuple[np.ndarray, int]
+AudioTuple = tuple[np.ndarray, float]
 
 VLLM_PLACEHOLDER = "<|audio|>"
 HF_PLACEHOLDER = "<|audio|>"
@@ -92,7 +92,7 @@ def _get_prompt(audio_count, question, placeholder):
 
 def run_multi_audio_test(
     vllm_runner: type[VllmRunner],
-    prompts_and_audios: list[tuple[str, list[AudioTuple]]],
+    prompts_and_audios: list[tuple[str | list[int], list[AudioTuple]]],
     model: str,
     *,
     dtype: str,

tests/models/multimodal/generation/test_vit_cudagraph.py

@@ -34,7 +34,7 @@ class VitCudagraphTestConfig:
 
 def params_with_marks(
     configs: dict[str, VitCudagraphTestConfig],
-) -> list[pytest.param]:
+) -> list[object]:
     return [
         pytest.param(model_id, marks=cfg.marks) for model_id, cfg in configs.items()
     ]
@@ -520,6 +520,7 @@ def test_vit_cudagraph_video(model_id, vllm_runner, video_assets):
             "baby_reading": config.video_prompt,  # type: ignore[typeddict-item]
         }
     )
+    sampled_vids: list
     if config.needs_video_metadata:
         sampled_vids = [
             sample_frames_with_video_metadata(
@@ -576,7 +577,7 @@ def _check_eonly_encoder_outputs(worker, batches):
         for batch in batches:
             kwargs = {key: value.to(runner.device) for key, value in batch.items()}
             actual = manager.execute(kwargs)
-            expected = runner.model.embed_multimodal(**kwargs)
+            expected = getattr(runner.model, "embed_multimodal")(**kwargs)  # noqa: B009
             assert len(actual) == len(expected)
             for output, reference in zip(actual, expected):
                 assert torch.isfinite(output).all()

tests/models/multimodal/generation/test_voxtral.py

@@ -58,7 +58,7 @@ def test_models_with_multiple_audios(
     vllm_prompt = _get_prompt(audio_assets, MULTI_AUDIO_PROMPT)
     run_multi_audio_test(
         vllm_runner,
-        [(vllm_prompt, [a.audio_and_sample_rate for a in audio_assets])],  # type: ignore[list-item]
+        [(vllm_prompt, [a.audio_and_sample_rate for a in audio_assets])],
         MODEL_NAME,
         dtype=dtype,
         max_tokens=max_tokens,

tests/models/multimodal/generation/test_voxtral_realtime.py

@@ -138,7 +138,7 @@ async def async_engine(request):
 def _build_inputs(
     audio_assets: list[AudioAsset], tokenizer: MistralTokenizer
 ) -> tuple[list[dict], list[SamplingParams]]:
-    audio_config = tokenizer.instruct_tokenizer.tokenizer.audio
+    audio_config = getattr(tokenizer.instruct_tokenizer.tokenizer, "audio")  # noqa: B009
     inputs = []
     sampling_params = []
     for audio_asset in audio_assets:

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

@@ -224,7 +224,7 @@ def build_embedding_inputs_from_test_info(
     images = [asset.pil_image for asset in image_assets]
     embeds = test_info.convert_assets_to_embeddings(image_assets)
     if test_info.dtype != "auto":
-        dtype = getattr(torch, test_info.dtype)  # type: ignore
+        dtype = getattr(torch, test_info.dtype)
         embeds = [e.to(dtype=dtype) for e in embeds]
     assert len(images) == len(model_prompts)
 

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

@@ -22,6 +22,7 @@
     GenerationMixin,
 )
 from transformers.masking_utils import create_causal_mask
+from transformers.utils import ModelOutput
 from transformers.video_utils import VideoMetadata
 
 from vllm.logprobs import SampleLogprobs
@@ -105,7 +106,9 @@ def _llava_vllm_to_hf_output(
     hf_output_str = output_str[1:] if output_str[0] == " " else output_str
 
     if hf_output_ids[-1] == eos_token_id:
-        hf_output_str = hf_output_str + tokenizer.decode(eos_token_id)
+        decoded_eos = tokenizer.decode(eos_token_id)
+        assert isinstance(decoded_eos, str)
+        hf_output_str = hf_output_str + decoded_eos
 
     return hf_output_ids, hf_output_str, out_logprobs
 
@@ -137,7 +140,9 @@ def llava_onevision_vllm_to_hf_output(
 
     hf_output_str = output_str
     if hf_output_ids[-1] == eos_token_id:
-        hf_output_str = hf_output_str + tokenizer.decode(eos_token_id)
+        decoded_eos = tokenizer.decode(eos_token_id)
+        assert isinstance(decoded_eos, str)
+        hf_output_str = hf_output_str + decoded_eos
 
     return hf_output_ids, hf_output_str, out_logprobs
 
@@ -179,7 +184,9 @@ def paligemma_vllm_to_hf_output(vllm_output: RunnerOutput, model: str) -> Runner
     hf_output_str = output_str
 
     if hf_output_ids[-1] == eos_token_id:
-        hf_output_str = hf_output_str + tokenizer.decode(eos_token_id)
+        decoded_eos = tokenizer.decode(eos_token_id)
+        assert isinstance(decoded_eos, str)
+        hf_output_str = hf_output_str + decoded_eos
 
     return hf_output_ids, hf_output_str, out_logprobs
 
@@ -286,7 +293,9 @@ def gemma3_vllm_to_hf_output(vllm_output: RunnerOutput, model: str) -> RunnerOut
 
     hf_output_str = output_str
     if hf_output_ids[-1] == eos_token_id:
-        hf_output_str = hf_output_str + tokenizer.decode(eos_token_id)
+        decoded_eos = tokenizer.decode(eos_token_id)
+        assert isinstance(decoded_eos, str)
+        hf_output_str = hf_output_str + decoded_eos
 
     return hf_output_ids, hf_output_str, out_logprobs
 
@@ -572,11 +581,10 @@ def patched_forward(
         # Prepare attention mask
         attention_mask = create_causal_mask(
             config=self.config,
-            input_embeds=inputs_embeds,
+            inputs_embeds=inputs_embeds,
             attention_mask=attention_mask,
             past_key_values=past_key_values,
             position_ids=position_ids,
-            cache_position=cache_position,
         )
 
         # Initialize and collect hidden states
@@ -800,7 +808,7 @@ def _internvl_generate(
     input_ids: torch.FloatTensor,
     attention_mask: torch.LongTensor | None = None,
     **generate_kwargs,
-) -> torch.LongTensor:
+) -> torch.Tensor | ModelOutput:
     """Generate method for InternVL2 model without fixed use_cache."""
     assert self.img_context_token_id is not None
     target_dtype = next(self.parameters()).dtype
@@ -828,7 +836,7 @@ def _internvl_generate(
     if not isinstance(self.language_model, GenerationMixin):
         pytest.skip("HF impl is not compatible with current transformers")
 
-    outputs = self.language_model.generate(
+    outputs = getattr(self.language_model, "generate")(  # noqa: B009
         **forward_kwargs,
         **generate_kwargs,
     )
@@ -1207,7 +1215,7 @@ def patched_generate(*args, **kwargs):
         return output
 
     hf_model.get_inputs = patched_get_inputs  # type: ignore[method-assign]
-    hf_model.model.generate = patched_generate  # type: ignore[method-assign]
+    hf_model.model.generate = patched_generate
     return hf_model
 
 

tests/models/multimodal/generation_ppl_test/ppl_utils.py

@@ -17,7 +17,7 @@
 PPL_TOL = 0.01
 
 
-@torch.inference_mode
+@torch.inference_mode()
 def vqa_ppl_test(
     hf_runner,
     vllm_runner,

tests/models/multimodal/pooling/test_dse_qwen2_vl.py

@@ -180,7 +180,7 @@ def test_models_text(
         hf_runner,
         vllm_runner,
         input_texts,
-        input_images,  # type: ignore
+        input_images,
         embed_texts,
         model,
         dtype=dtype,

tests/models/multimodal/pooling/test_embedding_gemma2.py

@@ -9,6 +9,8 @@
 import pytest
 import torch
 
+from vllm.multimodal.video import EmbeddingGemma2VideoBackend
+
 MODEL = os.environ.get("EG2_MODEL_PATH")
 needs_model = pytest.mark.skipif(not MODEL, reason="EG2_MODEL_PATH not set")
 TEXT_COS, MM_COS = 0.999, 0.999
@@ -348,6 +350,7 @@ def test_video_loader_real_video(vllm_runner, st_model):
 
     video_path = "/projects/gemma4-vllm/videos/test_intel.mp4"
     backend_cls = VIDEO_LOADER_REGISTRY.name2class["embedding_gemma2"]
+    assert issubclass(backend_cls, EmbeddingGemma2VideoBackend)
     with open(video_path, "rb") as f:
         data = f.read()
 
@@ -396,6 +399,7 @@ def test_video_loader_duplicate_expansion_e2e(vllm_runner, st_model):
     from vllm.multimodal.video import VIDEO_LOADER_REGISTRY
 
     backend_cls = VIDEO_LOADER_REGISTRY.name2class["embedding_gemma2"]
+    assert issubclass(backend_cls, EmbeddingGemma2VideoBackend)
     # 4s clip at 0.5 fps = 2 unique frames
     lowfps_path = "/tmp/test_lowfps_05.mp4"
     _ensure_synthetic_video(lowfps_path, duration=4.0, fps=0.5, width=64, height=48)
@@ -436,6 +440,7 @@ def test_video_loader_indices():
     )
 
     backend_cls = VIDEO_LOADER_REGISTRY.name2class["embedding_gemma2"]
+    assert issubclass(backend_cls, EmbeddingGemma2VideoBackend)
     hf_proc = EmbeddingGemma2VideoProcessor()
 
     # Case 1: clip shorter than 32s (e.g. 5s @ 24fps = 120 frames)
@@ -521,6 +526,7 @@ def test_video_loader_duplicate_expansion():
     )
 
     backend_cls = VIDEO_LOADER_REGISTRY.name2class["embedding_gemma2"]
+    assert issubclass(backend_cls, EmbeddingGemma2VideoBackend)
     hf_proc = EmbeddingGemma2VideoProcessor()
 
     # 4s clip @ 0.5fps = 2 frames total. Requested target fps = 1.0 -> 4 frames.

tests/models/multimodal/pooling/test_intern_vit.py

@@ -49,14 +49,14 @@ def _patched_init(self, *args, **kwargs):
         _orig_init(self, *args, **kwargs)
         self.all_tied_weights_keys = {}
 
-    PreTrainedModel.__init__ = _patched_init
+    PreTrainedModel.__init__ = _patched_init  # type: ignore[method-assign]
 
     try:
         hf_model = AutoModel.from_pretrained(
             model, dtype=torch_dtype, trust_remote_code=True
         ).to(DEVICE_TYPE)
     finally:
-        PreTrainedModel.__init__ = _orig_init
+        PreTrainedModel.__init__ = _orig_init  # type: ignore[method-assign]
     hf_outputs_per_image = [
         hf_model(pixel_value.to(DEVICE_TYPE)).last_hidden_state
         for pixel_value in pixel_values

tests/models/multimodal/pooling/test_llava_next.py

@@ -126,7 +126,7 @@ def test_models_text(
         hf_runner,
         vllm_runner,
         input_texts,
-        input_images,  # type: ignore
+        input_images,
         model,
         dtype=dtype,
     )

tests/models/multimodal/pooling/test_phi3v.py

@@ -129,7 +129,7 @@ def test_models_text(
         hf_runner,
         vllm_runner,
         input_texts,
-        input_images,  # type: ignore
+        input_images,
         model,
         dtype=dtype,
     )
@@ -192,7 +192,7 @@ def test_models_image_special_tokens_processing(
         return_tensors="pt",
     )
 
-    image_token_id = hf_processor.get_special_image_token_id()
+    image_token_id = getattr(hf_processor, "get_special_image_token_id")()  # noqa: B009
     hf_prompt_token_ids = [
         image_token_id if token_id < 0 else token_id
         for token_id in hf_inputs["input_ids"][0].tolist()