diff --git a/tools/pre_commit/mypy.py b/tools/pre_commit/mypy.py
index 8a832746f4..fcdf6d50c5 100755
--- a/tools/pre_commit/mypy.py
+++ b/tools/pre_commit/mypy.py
@@ -112,8 +112,6 @@ SEPARATE_GROUPS = [
 ]
 
 EXCLUDE = [
-    r"vllm/model_executor/models/[hH]",
-    r"vllm/model_executor/models/[iI]",
     r"vllm/model_executor/models/[jJ]",
     r"vllm/model_executor/models/[kK]",
     r"vllm/model_executor/models/[lL]",
diff --git a/vllm/model_executor/models/chatglm.py b/vllm/model_executor/models/chatglm.py
index 4363188ff6..6e6b8ac8b6 100644
--- a/vllm/model_executor/models/chatglm.py
+++ b/vllm/model_executor/models/chatglm.py
@@ -430,7 +430,7 @@ class ChatGLMBaseModel(nn.Module):
 
 
 class ChatGLMForCausalLM(ChatGLMBaseModel, SupportsLoRA, SupportsPP, SupportsQuant):
-    packed_modules_mapping = {
+    packed_modules_mapping: dict[str, list[str]] = {
         "query_key_value": ["query_key_value"],
         "dense_h_to_4h": ["dense_h_to_4h"],
     }
diff --git a/vllm/model_executor/models/commandr.py b/vllm/model_executor/models/commandr.py
index 87b832f91e..3fe78f081e 100644
--- a/vllm/model_executor/models/commandr.py
+++ b/vllm/model_executor/models/commandr.py
@@ -417,7 +417,7 @@ class CohereForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsQuant):
         orig_to_new_substr={"_quantizer.": None},
         orig_to_new_prefix={"lm_head": None},
     )
-    packed_modules_mapping = {
+    packed_modules_mapping: dict[str, list[str]] = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],
     }
diff --git a/vllm/model_executor/models/gemma.py b/vllm/model_executor/models/gemma.py
index 983822c158..940ba0a728 100644
--- a/vllm/model_executor/models/gemma.py
+++ b/vllm/model_executor/models/gemma.py
@@ -339,7 +339,7 @@ class GemmaForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsQuant):
             ".up_proj": (".gate_up_proj", 1),
         }
     )
-    packed_modules_mapping = {
+    packed_modules_mapping: dict[str, list[str]] = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],
     }
diff --git a/vllm/model_executor/models/gemma4_mm.py b/vllm/model_executor/models/gemma4_mm.py
index 428ab9ec16..2a15dacb64 100644
--- a/vllm/model_executor/models/gemma4_mm.py
+++ b/vllm/model_executor/models/gemma4_mm.py
@@ -1019,7 +1019,7 @@ class Gemma4ForConditionalGeneration(
     mm_prefix_clamp_sliding_window: bool = True
     supports_tower_connector_lora = True
 
-    packed_modules_mapping = {
+    packed_modules_mapping: dict[str, list[str]] = {
         "qkv_proj": [
             "q_proj",
             "k_proj",
diff --git a/vllm/model_executor/models/granite.py b/vllm/model_executor/models/granite.py
index 7b8f42fdd1..55a90f5ff0 100644
--- a/vllm/model_executor/models/granite.py
+++ b/vllm/model_executor/models/granite.py
@@ -395,7 +395,7 @@ class GraniteModel(nn.Module):
 class GraniteForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsQuant):
     hf_to_vllm_mapper = GraniteModel.hf_to_vllm_mapper
     # LoRA specific attributes
-    packed_modules_mapping = {
+    packed_modules_mapping: dict[str, list[str]] = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],
     }
diff --git a/vllm/model_executor/models/granitemoehybrid.py b/vllm/model_executor/models/granitemoehybrid.py
index 52b3a72a74..810770aab4 100644
--- a/vllm/model_executor/models/granitemoehybrid.py
+++ b/vllm/model_executor/models/granitemoehybrid.py
@@ -587,7 +587,7 @@ class GraniteMoeHybridForCausalLM(
     SupportsQuant,
     SupportsMambaPrefixCaching,
 ):
-    packed_modules_mapping = {
+    packed_modules_mapping: dict[str, list[str]] = {
         "qkv_proj": [
             "q_proj",
             "k_proj",
diff --git a/vllm/model_executor/models/h2ovl.py b/vllm/model_executor/models/h2ovl.py
index 02e31e719c..5c4f984c13 100644
--- a/vllm/model_executor/models/h2ovl.py
+++ b/vllm/model_executor/models/h2ovl.py
@@ -114,6 +114,7 @@ class H2OVLMultiModalProcessor(BaseInternVLMultiModalProcessor[H2OVLProcessingIn
             if isinstance(images, ImageEmbeddingItems):
                 feature_size = images.get_feature_size(item_idx)
             else:
+                assert isinstance(images, ImageProcessorItems)
                 image_size = images.get_image_size(item_idx)
                 feature_size = self.info.get_num_image_tokens(
                     image_width=image_size.width,
diff --git a/vllm/model_executor/models/hrm_text.py b/vllm/model_executor/models/hrm_text.py
index 29825ece5a..db8d9c20c3 100644
--- a/vllm/model_executor/models/hrm_text.py
+++ b/vllm/model_executor/models/hrm_text.py
@@ -248,7 +248,7 @@ class HrmTextDecoderLayer(nn.Module):
         self,
         config: PretrainedConfig,
         layer_idx_in_stack: int,
-        stack_kind: str,
+        stack_kind: Literal["L", "H"],
         cache_config: CacheConfig | None = None,
         quant_config: QuantizationConfig | None = None,
         prefix: str = "",
@@ -312,7 +312,7 @@ class HrmTextStack(nn.Module):
     def __init__(
         self,
         config: PretrainedConfig,
-        stack_kind: str,
+        stack_kind: Literal["L", "H"],
         cache_config: CacheConfig | None = None,
         quant_config: QuantizationConfig | None = None,
         prefix: str = "",
diff --git a/vllm/model_executor/models/hy_v3.py b/vllm/model_executor/models/hy_v3.py
index 43293e8372..0b54c26f47 100644
--- a/vllm/model_executor/models/hy_v3.py
+++ b/vllm/model_executor/models/hy_v3.py
@@ -161,6 +161,7 @@ class HYV3MoEFused(nn.Module):
             prefix=f"{prefix}.gate",
         )
 
+        self.shared_mlp: HYV3FeedForward | None
         if config.num_shared_experts > 0:
             self.shared_mlp = HYV3FeedForward(
                 hidden_size=config.hidden_size,
@@ -467,28 +468,27 @@ class HYV3Model(nn.Module, MixtureOfExperts):
         # Set MoE hyperparameters
         self.num_expert_groups = 1
         self.moe_layers = []
-        example_layer = None
+        example_layer: HYV3MoEFused | None = None
         for layer in self.layers:
             if isinstance(layer, PPMissingLayer):
                 continue
 
             assert isinstance(layer, HYV3DecoderLayer)
             if layer.block_type == "moe":
+                assert isinstance(layer.mlp, HYV3MoEFused)
                 example_layer = layer.mlp
-                self.moe_layers.append(layer.mlp.experts)
+                self.moe_layers.append(example_layer.experts)
 
         if example_layer is None:
             self.num_moe_layers = 0
             raise RuntimeError("No MoE layer found in model.layers.")
 
         self.num_moe_layers = len(self.moe_layers)
-        self.num_logical_experts = getattr(example_layer, "n_logical_experts", None)
-        self.num_physical_experts = getattr(example_layer, "n_physical_experts", None)
-        self.num_local_physical_experts = getattr(
-            example_layer, "n_local_physical_experts", None
-        )
-        self.num_routed_experts = getattr(example_layer, "n_routed_experts", None)
-        self.num_redundant_experts = getattr(example_layer, "n_redundant_experts", None)
+        self.num_logical_experts = example_layer.n_logical_experts
+        self.num_physical_experts = example_layer.n_physical_experts
+        self.num_local_physical_experts = example_layer.n_local_physical_experts
+        self.num_routed_experts = example_layer.n_routed_experts
+        self.num_redundant_experts = example_layer.n_redundant_experts
 
     def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
         return self.embed_tokens(input_ids)
@@ -572,9 +572,10 @@ class HYV3Model(nn.Module, MixtureOfExperts):
                 continue
             if "scale" in name:
                 # Remapping the name of FP8 kv-scale.
-                name = maybe_remap_kv_scale_name(name, params_dict)
-                if name is None:
+                remapped_name = maybe_remap_kv_scale_name(name, params_dict)
+                if remapped_name is None:
                     continue
+                name = remapped_name
             is_found = False
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name:
diff --git a/vllm/model_executor/models/hy_v3_mtp.py b/vllm/model_executor/models/hy_v3_mtp.py
index 193280d7c8..aa9c2f35b5 100644
--- a/vllm/model_executor/models/hy_v3_mtp.py
+++ b/vllm/model_executor/models/hy_v3_mtp.py
@@ -24,7 +24,7 @@
 # limitations under the License.
 """Inference-only HY V3 MTP model compatible with HuggingFace weights."""
 
-from collections.abc import Iterable
+from collections.abc import Callable, Iterable
 
 import regex as re
 import torch
@@ -288,7 +288,15 @@ class HYV3MTP(nn.Module):
         num_kv_heads = getattr(
             self.config, "num_key_value_heads", self.config.num_attention_heads
         )
-        split_params_mapping = [
+        split_params_mapping: list[
+            tuple[
+                str,
+                str,
+                int,
+                list[tuple[str | int, int]],
+                Callable[[torch.Tensor], torch.Tensor] | None,
+            ]
+        ] = [
             (".gate_up_proj", ".gate_and_up_proj", 2, [(1, 1), (0, 1)], None),
             (
                 ".qkv_proj",
@@ -308,7 +316,7 @@ class HYV3MTP(nn.Module):
                 num_experts=self.config.num_experts,
             )
         else:
-            expert_params_mapping = {}
+            expert_params_mapping = []
 
         params_dict = dict(self.named_parameters())
 
@@ -354,9 +362,10 @@ class HYV3MTP(nn.Module):
             if name == "__skip__":
                 continue
             if "scale" in name:
-                name = maybe_remap_kv_scale_name(name, params_dict)
-                if name is None:
+                remapped_name = maybe_remap_kv_scale_name(name, params_dict)
+                if remapped_name is None:
                     continue
+                name = remapped_name
             is_found = False
 
             for param_name, weight_name, shard_id in stacked_params_mapping:
diff --git a/vllm/model_executor/models/hyperclovax_vision_v2.py b/vllm/model_executor/models/hyperclovax_vision_v2.py
index b1d64db9d9..96e718751d 100644
--- a/vllm/model_executor/models/hyperclovax_vision_v2.py
+++ b/vllm/model_executor/models/hyperclovax_vision_v2.py
@@ -12,7 +12,7 @@ Supports:
 
 from collections.abc import Iterable, Mapping, Sequence
 from functools import partial
-from typing import Annotated, Literal
+from typing import Annotated, Literal, TypedDict
 
 import torch
 import torch.nn as nn
@@ -122,6 +122,11 @@ class HCXVisionV2VideoEmbeddingInputs(TensorSchema):
 HCXVisionV2VideoInputs = HCXVisionV2VideoPixelInputs | HCXVisionV2VideoEmbeddingInputs
 
 
+class HCXVisionV2MultiModalInputs(TypedDict, total=False):
+    image: HCXVisionV2ImageInputs | None
+    video: HCXVisionV2VideoInputs | None
+
+
 class HCXVisionV2ProcessingInfo(BaseProcessingInfo):
     """Processing info for HyperCLOVAX V2 (32B Think model)."""
 
@@ -329,6 +334,7 @@ class HCXVisionV2MultiModalProcessor(
                 if grid_thw_elem is not None:
                     # Access .data to get the actual tensor from MultiModalFieldElem
                     grid_thw = grid_thw_elem.data
+                    assert isinstance(grid_thw, torch.Tensor)
                     # Qwen2.5-VL style calculation
                     h, w = grid_thw[1].item(), grid_thw[2].item()
                     num_tokens = (h * w) // (merge_size**2)
@@ -340,6 +346,7 @@ class HCXVisionV2MultiModalProcessor(
                 if grid_thw_elem is not None:
                     # Access .data to get the actual tensor from MultiModalFieldElem
                     grid_thw = grid_thw_elem.data
+                    assert isinstance(grid_thw, torch.Tensor)
                     t, h, w = grid_thw[0].item(), grid_thw[1].item(), grid_thw[2].item()
                     num_tokens = (t * h * w) // (merge_size**2)
                 else:
@@ -606,8 +613,10 @@ class HCXVisionV2ForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
         sizes = (grid_thw.prod(-1) // merge_size // merge_size).tolist()
         return video_embeds.split(sizes)
 
-    def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict:
-        modalities = {}
+    def _parse_and_validate_multimodal_inputs(
+        self, **kwargs: object
+    ) -> HCXVisionV2MultiModalInputs:
+        modalities: HCXVisionV2MultiModalInputs = {}
 
         for input_key in kwargs:
             if (
diff --git a/vllm/model_executor/models/idefics3.py b/vllm/model_executor/models/idefics3.py
index 9a55ac51d3..80e223e4c1 100644
--- a/vllm/model_executor/models/idefics3.py
+++ b/vllm/model_executor/models/idefics3.py
@@ -625,6 +625,7 @@ class Idefics3ForConditionalGeneration(
         if image_input["type"] == "image_embeds":
             return image_input["data"]
 
+        assert isinstance(image_input, Idefics3ImagePixelInputs)
         image_features = self._process_image_pixels(image_input)
         image_features = self.model.connector(image_features)
 
@@ -788,7 +789,7 @@ class Idefics3ForConditionalGeneration(
         path: str = "default",
     ) -> torch.Tensor:
         image_input = self._parse_and_validate_image_input(**mm_kwargs)
-        assert image_input is not None and image_input["type"] == "pixel_values"
+        assert isinstance(image_input, Idefics3ImagePixelInputs)
         image_features = self._process_image_pixels(image_input)
         image_features = self.model.connector(image_features)
         return image_features.flatten(0, 1)
diff --git a/vllm/model_executor/models/interfaces.py b/vllm/model_executor/models/interfaces.py
index b3102f567a..fd81bb502f 100644
--- a/vllm/model_executor/models/interfaces.py
+++ b/vllm/model_executor/models/interfaces.py
@@ -32,6 +32,7 @@ from typing_extensions import Self, TypeIs
 
 from vllm.logger import init_logger
 from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.models.interfaces_base import VllmModel
 from vllm.utils.collection_utils import common_prefix
 from vllm.utils.func_utils import supports_kw
 
@@ -46,8 +47,9 @@ if TYPE_CHECKING:
     from vllm.inputs import PromptType, TokensPrompt
     from vllm.lora.model_manager import LoRAModelManager
     from vllm.model_executor.layers.fused_moe import MoERunner
+    from vllm.model_executor.layers.logits_processor import LogitsProcessor
     from vllm.model_executor.layers.mamba.mamba_utils import MambaStateCopyFunc
-    from vllm.model_executor.models.interfaces_base import VllmModel
+    from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
     from vllm.model_executor.models.utils import WeightsMapper
     from vllm.multimodal.inputs import MultiModalFeatureSpec, MultiModalKwargsItem
     from vllm.multimodal.registry import _ProcessorFactories
@@ -128,14 +130,24 @@ class SupportsMultiModalEmbeddings(Protocol):
 
     supports_multimodal_embeddings: ClassVar[Literal[True]] = True
 
+    @overload
     def embed_input_ids(
         self,
         input_ids: Tensor,
-        multimodal_embeddings: MultiModalEmbeddings | None = None,
+        multimodal_embeddings: None = None,
         *,
         is_multimodal: Tensor | None = None,
     ) -> Tensor: ...
 
+    @overload
+    def embed_input_ids(
+        self,
+        input_ids: Tensor,
+        multimodal_embeddings: MultiModalEmbeddings,
+        *,
+        is_multimodal: Tensor,
+    ) -> Tensor: ...
+
 
 @runtime_checkable
 class SupportsMultiModal(SupportsMultiModalEmbeddings, Protocol):
@@ -243,26 +255,30 @@ class SupportsMultiModal(SupportsMultiModalEmbeddings, Protocol):
             torch.nn.Module: The core language model component.
         """
         # Cached
-        if self in _language_model_by_module:
-            return _language_model_by_module[self]
+        assert isinstance(self, nn.Module)
+        self_module = self
+        if self_module in _language_model_by_module:
+            return _language_model_by_module[self_module]
 
         if self._language_model_names:
-            mod = self
+            mod = self_module
             for attr in common_prefix(
                 [name.split(".") for name in self._language_model_names]
             ):
                 if attr:
                     mod = getattr(mod, attr)
 
-            if mod is not self and hasattr(mod, "embed_input_ids"):
-                _language_model_by_module[self] = mod
-                return mod
+            candidate: object = mod
+            if candidate is not self_module and isinstance(candidate, VllmModel):
+                _language_model_by_module[self_module] = candidate
+                return candidate
 
         # Fallback
-        for mod in self.children():
-            if hasattr(mod, "embed_input_ids"):
-                _language_model_by_module[self] = mod
-                return mod
+        for mod in self_module.children():
+            candidate = mod
+            if isinstance(candidate, VllmModel):
+                _language_model_by_module[self_module] = candidate
+                return candidate
 
         raise NotImplementedError(
             f"No language model found in {type(self).__name__}! "
@@ -289,7 +305,7 @@ class SupportsMultiModal(SupportsMultiModalEmbeddings, Protocol):
         """
         from .utils import StageMissingLayer, collect_children, no_init_weights
 
-        mm_config = vllm_config.model_config.multimodal_config
+        mm_config = vllm_config.model_config.get_multimodal_config()
 
         with collect_children(self, targets=targets) as children_names:  # noqa: SIM117
             with (
@@ -339,7 +355,7 @@ class SupportsMultiModal(SupportsMultiModalEmbeddings, Protocol):
         else:
             stage_name = "_".join([*modalities, "tower"])
 
-        mm_config = vllm_config.model_config.multimodal_config
+        mm_config = vllm_config.model_config.get_multimodal_config()
 
         with collect_children(self, targets=targets) as children_names:  # noqa: SIM117
             with (
@@ -361,7 +377,8 @@ class SupportsMultiModal(SupportsMultiModalEmbeddings, Protocol):
         offloader = get_offloader()
         if offloader.supports_tower_offload:
             for name in children_names:
-                offloader.wrap_modules(iter([attrgetter(name)(self)]), prefix=name)
+                modules = (module for module in [attrgetter(name)(self)])
+                offloader.wrap_modules(modules, prefix=name)
 
     @contextmanager
     def _mark_composite_model(
@@ -434,18 +451,6 @@ class SupportsMultiModal(SupportsMultiModalEmbeddings, Protocol):
             num_connector_tokens if isinstance(num_connector_tokens, int) else None,
         )
 
-    @overload
-    def embed_input_ids(self, input_ids: Tensor) -> Tensor: ...
-
-    @overload
-    def embed_input_ids(
-        self,
-        input_ids: Tensor,
-        multimodal_embeddings: MultiModalEmbeddings,
-        *,
-        is_multimodal: torch.Tensor,
-    ) -> Tensor: ...
-
     def _embed_text_input_ids(
         self,
         input_ids: Tensor,
@@ -465,6 +470,24 @@ class SupportsMultiModal(SupportsMultiModalEmbeddings, Protocol):
 
         return embed_input_ids(input_ids)
 
+    @overload
+    def embed_input_ids(
+        self,
+        input_ids: Tensor,
+        multimodal_embeddings: None = None,
+        *,
+        is_multimodal: Tensor | None = None,
+    ) -> Tensor: ...
+
+    @overload
+    def embed_input_ids(
+        self,
+        input_ids: Tensor,
+        multimodal_embeddings: MultiModalEmbeddings,
+        *,
+        is_multimodal: Tensor,
+    ) -> Tensor: ...
+
     def embed_input_ids(
         self,
         input_ids: Tensor,
@@ -671,7 +694,7 @@ class SupportsLoRA(Protocol):
     # The `embedding_module` and `embedding_padding_modules`
     # are empty by default.
     embedding_modules: ClassVar[dict[str, str]] = {}
-    packed_modules_mapping: ClassVar[dict[str, list[str]]] = {}
+    packed_modules_mapping: dict[str, list[str]] = {}
     # Module prefixes to skip during LoRA loading (e.g., ["mtp."] for MTP layers)
     lora_skip_prefixes: ClassVar[list[str]] = []
     lora_manager: "LoRAModelManager | None"
@@ -864,11 +887,11 @@ class HasInnerState(Protocol):
 
 
 @overload
-def has_inner_state(model: object) -> TypeIs[HasInnerState]: ...
+def has_inner_state(model: type[object]) -> TypeIs[type[HasInnerState]]: ...
 
 
 @overload
-def has_inner_state(model: type[object]) -> TypeIs[type[HasInnerState]]: ...
+def has_inner_state(model: object) -> TypeIs[HasInnerState]: ...
 
 
 def has_inner_state(
@@ -891,11 +914,11 @@ class IsAttentionFree(Protocol):
 
 
 @overload
-def is_attention_free(model: object) -> TypeIs[IsAttentionFree]: ...
+def is_attention_free(model: type[object]) -> TypeIs[type[IsAttentionFree]]: ...
 
 
 @overload
-def is_attention_free(model: type[object]) -> TypeIs[type[IsAttentionFree]]: ...
+def is_attention_free(model: object) -> TypeIs[IsAttentionFree]: ...
 
 
 def is_attention_free(
@@ -946,11 +969,11 @@ class IsHybrid(Protocol):
 
 
 @overload
-def is_hybrid(model: object) -> TypeIs[IsHybrid]: ...
+def is_hybrid(model: type[object]) -> TypeIs[type[IsHybrid]]: ...
 
 
 @overload
-def is_hybrid(model: type[object]) -> TypeIs[type[IsHybrid]]: ...
+def is_hybrid(model: object) -> TypeIs[IsHybrid]: ...
 
 
 def is_hybrid(
@@ -1025,7 +1048,7 @@ class MixtureOfExperts(Protocol):
         self.expert_weights = []
         for layer_idx, layer in enumerate(self.moe_layers):
             # Register the expert weights.
-            self.expert_weights.append(layer.get_expert_weights())
+            self.expert_weights.append(list(layer.get_expert_weights()))
             layer.set_eplb_state(
                 moe_layer_idx=layer_idx,
                 expert_load_view=expert_load_view,
@@ -1083,11 +1106,11 @@ class HasNoOps(Protocol):
 
 
 @overload
-def has_noops(model: object) -> TypeIs[HasNoOps]: ...
+def has_noops(model: type[object]) -> TypeIs[type[HasNoOps]]: ...
 
 
 @overload
-def has_noops(model: type[object]) -> TypeIs[type[HasNoOps]]: ...
+def has_noops(model: object) -> TypeIs[HasNoOps]: ...
 
 
 def has_noops(
@@ -1108,14 +1131,14 @@ class SupportsMambaPrefixCaching(Protocol):
 
 @overload
 def supports_mamba_prefix_caching(
-    model: object,
-) -> TypeIs[SupportsMambaPrefixCaching]: ...
+    model: type[object],
+) -> TypeIs[type[SupportsMambaPrefixCaching]]: ...
 
 
 @overload
 def supports_mamba_prefix_caching(
-    model: type[object],
-) -> TypeIs[type[SupportsMambaPrefixCaching]]: ...
+    model: object,
+) -> TypeIs[SupportsMambaPrefixCaching]: ...
 
 
 def supports_mamba_prefix_caching(
@@ -1136,11 +1159,11 @@ class SupportsReplaySSM(Protocol):
 
 
 @overload
-def supports_replayssm(model: object) -> TypeIs[SupportsReplaySSM]: ...
+def supports_replayssm(model: type[object]) -> TypeIs[type[SupportsReplaySSM]]: ...
 
 
 @overload
-def supports_replayssm(model: type[object]) -> TypeIs[type[SupportsReplaySSM]]: ...
+def supports_replayssm(model: object) -> TypeIs[SupportsReplaySSM]: ...
 
 
 def supports_replayssm(
@@ -1172,7 +1195,7 @@ class SupportsQuant:
     """The interface required for all models that support quantization."""
 
     hf_to_vllm_mapper: "WeightsMapper | None" = None
-    packed_modules_mapping: ClassVar[dict[str, list[str]]]
+    packed_modules_mapping: dict[str, list[str]]
     quant_config: QuantizationConfig | None = None
 
     def __new__(cls, *args, **kwargs) -> Self:
@@ -1504,6 +1527,9 @@ class LocalArgmaxMixin:
         ``self.draft_id_to_target_id`` (optional): nn.Parameter
     """
 
+    logits_processor: "LogitsProcessor"
+    lm_head: "ParallelLMHead"
+
     def get_top_tokens(self, hidden_states: torch.Tensor) -> torch.Tensor:
         """Vocab-parallel argmax with optional D2T remapping."""
         top = self.logits_processor.get_top_tokens(
diff --git a/vllm/model_executor/models/interfaces_base.py b/vllm/model_executor/models/interfaces_base.py
index 0c182a891c..e1a6b650c6 100644
--- a/vllm/model_executor/models/interfaces_base.py
+++ b/vllm/model_executor/models/interfaces_base.py
@@ -57,7 +57,7 @@ class VllmModel(Protocol[T_co]):
 
 
 def _check_vllm_model_init(model: type[object] | object) -> bool:
-    model_init = model.__init__
+    model_init = getattr(model, "__init__")  # noqa: B009
     return supports_kw(model_init, "vllm_config")
 
 
@@ -274,10 +274,11 @@ def get_attn_type(model: type[object] | object) -> AttnTypeStr:
 
 
 def get_score_type(model: type[object] | object) -> ScoreType:
-    score_types = set()
-    for m in model.__mro__:
-        score_type = getattr(m, "score_type", "bi-encoder")
+    score_types: set[ScoreType] = set()
+    model_cls = model if isinstance(model, type) else type(model)
+    for m in model_cls.__mro__:
+        score_type: ScoreType = getattr(m, "score_type", "bi-encoder")
         if score_type != "bi-encoder":
             score_types.add(score_type)
     assert len(score_types) < 2
-    return "bi-encoder" if not score_types else list(score_types)[0]
+    return "bi-encoder" if not score_types else next(iter(score_types))
diff --git a/vllm/model_executor/models/internlm2.py b/vllm/model_executor/models/internlm2.py
index 4996723b3c..29dd124b86 100644
--- a/vllm/model_executor/models/internlm2.py
+++ b/vllm/model_executor/models/internlm2.py
@@ -323,7 +323,7 @@ class InternLM2Model(nn.Module):
 
 class InternLM2ForCausalLM(nn.Module, SupportsPP, SupportsLoRA, SupportsQuant):
     hf_to_vllm_mapper = InternLM2Model.hf_to_vllm_mapper
-    packed_modules_mapping = {
+    packed_modules_mapping: dict[str, list[str]] = {
         "wqkv": ["wqkv"],
         "gate_up_proj": ["w1", "w3"],
     }
diff --git a/vllm/model_executor/models/interns1.py b/vllm/model_executor/models/interns1.py
index f8062cb751..0ed2723c32 100644
--- a/vllm/model_executor/models/interns1.py
+++ b/vllm/model_executor/models/interns1.py
@@ -7,7 +7,7 @@
 # Licensed under The MIT License [see LICENSE for details]
 # --------------------------------------------------------
 from collections.abc import Iterable, Mapping, Sequence
-from typing import Annotated, Literal, TypeAlias
+from typing import Annotated, Literal, TypeAlias, TypedDict
 
 import regex as re
 import torch
@@ -22,7 +22,7 @@ from transformers.models.internvl.video_processing_internvl import (
 )
 
 from vllm.config import VllmConfig
-from vllm.config.multimodal import BaseDummyOptions
+from vllm.config.multimodal import BaseDummyOptions, VideoDummyOptions
 from vllm.inputs import MultiModalDataDict
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.models.interns1_vit import InternS1VisionModel
@@ -148,6 +148,11 @@ class InternS1VideoEmbeddingInputs(TensorSchema):
 InternS1VideoInputs: TypeAlias = InternS1VideoPixelInputs | InternS1VideoEmbeddingInputs
 
 
+class InternS1MultiModalInputs(TypedDict, total=False):
+    images: InternS1ImageInputs | None
+    videos: InternS1VideoInputs | None
+
+
 def resolve_interns1_min_max_num(
     min_dynamic_patch: int,
     max_dynamic_patch: int,
@@ -312,6 +317,7 @@ class InternS1DummyInputsBuilder(BaseDummyInputsBuilder[InternS1ProcessingInfo])
 
         image_overrides = mm_options.get("image")
         video_overrides = mm_options.get("video")
+        assert video_overrides is None or isinstance(video_overrides, VideoDummyOptions)
 
         return {
             "image": self._get_dummy_images(
@@ -474,22 +480,24 @@ class InternS1MultiModalProcessor(BaseMultiModalProcessor[InternS1ProcessingInfo
         video_token = hf_processor.video_token
 
         tokenizer = self.info.get_tokenizer()
-        video_token_id = cached_encode(tokenizer, video_token, add_special_tokens=False)
-        assert len(video_token_id) == 1
-        video_token_id = video_token_id[0]
+        video_token_ids = cached_encode(
+            tokenizer, video_token, add_special_tokens=False
+        )
+        assert len(video_token_ids) == 1
+        video_token_id = video_token_ids[0]
 
         out_mm_data = out_mm_kwargs.get_data()
         if "video_num_patches" in out_mm_data:
-            video_num_patches = out_mm_data["video_num_patches"]
-            assert isinstance(video_num_patches, torch.Tensor)
-            video_num_patches = video_num_patches.tolist()
+            video_num_patches_tensor = out_mm_data["video_num_patches"]
+            assert isinstance(video_num_patches_tensor, torch.Tensor)
+            video_num_patches: list[int] = video_num_patches_tensor.tolist()
         else:
             video_num_patches = []
 
         if "image_num_patches" in out_mm_data:
-            image_num_patches = out_mm_data["image_num_patches"]
-            assert isinstance(image_num_patches, torch.Tensor)
-            image_num_patches = image_num_patches.tolist()
+            image_num_patches_tensor = out_mm_data["image_num_patches"]
+            assert isinstance(image_num_patches_tensor, torch.Tensor)
+            image_num_patches: list[int] = image_num_patches_tensor.tolist()
         else:
             image_num_patches = []
 
@@ -602,8 +610,8 @@ class InternS1ForConditionalGeneration(
                 prefix=maybe_prefix(prefix, "language_model"),
             )
 
-        self.img_context_token_id = None
-        self.video_context_token_id = None
+        self.img_context_token_id: int | None = None
+        self.video_context_token_id: int | None = None
 
         self.visual_token_mask = None
         self.make_empty_intermediate_tensors = (
@@ -756,8 +764,10 @@ class InternS1ForConditionalGeneration(
         ]
         return image_embeds.split(image_feature_sizes)
 
-    def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict:
-        modalities = {}
+    def _parse_and_validate_multimodal_inputs(
+        self, **kwargs: object
+    ) -> InternS1MultiModalInputs:
+        modalities: InternS1MultiModalInputs = {}
 
         # Preserve the order of modalities if there are multiple of them
         # from the order of kwargs.
@@ -789,12 +799,14 @@ class InternS1ForConditionalGeneration(
         for modality in modalities:
             if modality == "images":
                 image_input = modalities["images"]
-                image_embeddings = self._process_vision_input(image_input)
-                multimodal_embeddings += tuple(image_embeddings)
+                if image_input is not None:
+                    image_embeddings = self._process_vision_input(image_input)
+                    multimodal_embeddings += tuple(image_embeddings)
             if modality == "videos":
                 video_input = modalities["videos"]
-                video_embeddings = self._process_vision_input(video_input)
-                multimodal_embeddings += tuple(video_embeddings)
+                if video_input is not None:
+                    video_embeddings = self._process_vision_input(video_input)
+                    multimodal_embeddings += tuple(video_embeddings)
 
         return multimodal_embeddings
 
diff --git a/vllm/model_executor/models/interns1_pro.py b/vllm/model_executor/models/interns1_pro.py
index ee52fbd03e..40c5330ad3 100644
--- a/vllm/model_executor/models/interns1_pro.py
+++ b/vllm/model_executor/models/interns1_pro.py
@@ -26,7 +26,7 @@
 
 import functools
 from collections.abc import Iterable
-from typing import Any
+from typing import Any, ClassVar
 
 import torch
 from torch import nn
@@ -321,6 +321,14 @@ class InternS1ProMoeAttention(nn.Module):
             dual_chunk_attention_config=dual_chunk_attention_config,
         )
 
+        attn_kwargs: dict[str, Any] = (
+            {
+                "layer_idx": extract_layer_index(prefix),
+                "dual_chunk_attention_config": dual_chunk_attention_config,
+            }
+            if dual_chunk_attention_config
+            else {}
+        )
         self.attn = Attention(
             self.num_heads,
             self.head_dim,
@@ -329,12 +337,7 @@ class InternS1ProMoeAttention(nn.Module):
             cache_config=cache_config,
             quant_config=quant_config,
             prefix=f"{prefix}.attn",
-            **{
-                "layer_idx": extract_layer_index(prefix),
-                "dual_chunk_attention_config": dual_chunk_attention_config,
-            }
-            if dual_chunk_attention_config
-            else {},
+            **attn_kwargs,
         )
 
         self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
@@ -489,6 +492,9 @@ class InternS1ProMoeLLMForCausalLM(Qwen3MoeForCausalLM):
 
 
 class InternS1ProMoeMixtureOfExperts(MixtureOfExperts):
+    language_model: InternS1ProMoeLLMForCausalLM
+    num_local_physical_experts: int
+
     def update_physical_experts_metadata(
         self,
         num_physical_experts: int,
@@ -538,8 +544,8 @@ class InternS1ProMoeMixtureOfExperts(MixtureOfExperts):
 class InternS1ProForConditionalGeneration(
     Qwen3VLForConditionalGeneration, InternS1ProMoeMixtureOfExperts
 ):
-    is_3d_moe_weight: bool = True
-    packed_modules_mapping = {
+    is_3d_moe_weight: ClassVar[bool] = True
+    packed_modules_mapping: dict[str, list[str]] = {
         "qkv_proj": [
             "q_proj",
             "k_proj",
@@ -559,7 +565,7 @@ class InternS1ProForConditionalGeneration(
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super(Qwen3VLForConditionalGeneration, self).__init__()
         config: PretrainedConfig = vllm_config.model_config.hf_config
-        multimodal_config = vllm_config.model_config.multimodal_config
+        multimodal_config = vllm_config.model_config.get_multimodal_config()
 
         self.config = config
         self.multimodal_config = multimodal_config
@@ -601,8 +607,8 @@ class InternS1ProForConditionalGeneration(
         # Set MoE hyperparameters
         self.set_moe_parameters()
 
-    def get_frope_params_map(self) -> str:
-        mapper = {}
+    def get_frope_params_map(self) -> dict[str, str]:
+        mapper: dict[str, str] = {}
         for name, params in self.language_model.model.named_parameters():
             if "rotary_emb.sin_coef" in name:
                 mapper["language_model.model.rotary_emb.sin_coef"] = (
diff --git a/vllm/model_executor/models/interns2_mobius.py b/vllm/model_executor/models/interns2_mobius.py
index b8111412f0..4efaa5a08e 100644
--- a/vllm/model_executor/models/interns2_mobius.py
+++ b/vllm/model_executor/models/interns2_mobius.py
@@ -476,7 +476,7 @@ class InternS2MobiusForConditionalGeneration(InternS2PreviewForConditionalGenera
         if getattr(config, "image_token_index", None) is None:
             config.image_token_index = config.image_token_id
         quant_config = vllm_config.quant_config
-        multimodal_config = vllm_config.model_config.multimodal_config
+        multimodal_config = vllm_config.model_config.get_multimodal_config()
 
         self.config = config
         self.model_config = vllm_config.model_config
diff --git a/vllm/model_executor/models/internvl.py b/vllm/model_executor/models/internvl.py
index 54fa28cf5a..1170666be2 100644
--- a/vllm/model_executor/models/internvl.py
+++ b/vllm/model_executor/models/internvl.py
@@ -10,14 +10,14 @@
 from abc import abstractmethod
 from collections.abc import Iterable, Mapping, Sequence
 from functools import cached_property
-from typing import Annotated, Any, Literal, TypeAlias, TypeVar
+from typing import Annotated, Any, Literal, TypeAlias, TypedDict, TypeVar
 
 import torch
 import torch.nn as nn
 from transformers import BatchFeature, PretrainedConfig
 
 from vllm.config import VllmConfig
-from vllm.config.multimodal import BaseDummyOptions
+from vllm.config.multimodal import BaseDummyOptions, VideoDummyOptions
 from vllm.inputs import MultiModalDataDict
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.layers.quantization.auto_awq import AutoAWQConfig
@@ -128,6 +128,11 @@ class InternVLVideoEmbeddingInputs(TensorSchema):
 InternVLVideoInputs: TypeAlias = InternVLVideoPixelInputs | InternVLVideoEmbeddingInputs
 
 
+class InternVLMultiModalInputs(TypedDict, total=False):
+    images: InternVLImageInputs | None
+    videos: InternVLVideoInputs | None
+
+
 class BaseInternVLProcessingInfo(BaseProcessingInfo):
     """Basic image-only ProcessingInfo for InternVL-style models."""
 
@@ -286,6 +291,7 @@ class BaseInternVLMultiModalProcessor(BaseMultiModalProcessor[_I]):
             if isinstance(images, ImageEmbeddingItems):
                 feature_size = images.get_feature_size(item_idx)
             else:
+                assert isinstance(images, ImageProcessorItems)
                 image_size = images.get_image_size(item_idx)
                 feature_size = self.info.get_num_image_tokens(
                     image_width=image_size.width,
@@ -433,6 +439,9 @@ class InternVLDummyInputsBuilder(
             )
             num_videos = mm_counts.get("video", 0)
             video_overrides = mm_options.get("video")
+            assert video_overrides is None or isinstance(
+                video_overrides, VideoDummyOptions
+            )
             dummy_video = {
                 "video": self._get_dummy_videos(
                     width=image_size,
@@ -504,9 +513,9 @@ class InternVLMultiModalProcessor(
         tokenizer = self.info.get_tokenizer()
 
         if "video_num_patches" in out_mm_data:
-            video_num_patches = out_mm_data["video_num_patches"]
-            assert isinstance(video_num_patches, torch.Tensor)
-            video_num_patches = video_num_patches.tolist()
+            video_num_patches_tensor = out_mm_data["video_num_patches"]
+            assert isinstance(video_num_patches_tensor, torch.Tensor)
+            video_num_patches: list[int] = video_num_patches_tensor.tolist()
         else:
             video_num_patches = []
 
@@ -591,7 +600,7 @@ class InternVLChatModel(
 
         config = vllm_config.model_config.hf_config
         quant_config = vllm_config.quant_config
-        multimodal_config = vllm_config.model_config.multimodal_config
+        multimodal_config = vllm_config.model_config.get_multimodal_config()
 
         self.config = config
         self.multimodal_config = multimodal_config
@@ -621,15 +630,15 @@ class InternVLChatModel(
                 prefix=maybe_prefix(prefix, "language_model"),
             )
 
-        self.img_context_token_id = None
-        self.video_context_token_id = None
+        self.img_context_token_id: int | None = None
+        self.video_context_token_id: int | None = None
 
         self.make_empty_intermediate_tensors = (
             self.language_model.make_empty_intermediate_tensors
         )
 
     def _patch_quant_config(
-        self, config: PretrainedConfig, quant_config: QuantizationConfig
+        self, config: PretrainedConfig, quant_config: QuantizationConfig | None
     ):
         # the awq models from OpenGVLab missing `modules_to_not_convert`
         # patch the quant_config to add `modules_to_not_convert` back
@@ -743,7 +752,7 @@ class InternVLChatModel(
 
     def _parse_and_validate_video_input(
         self, **kwargs: object
-    ) -> InternVLVideoPixelInputs | None:
+    ) -> InternVLVideoInputs | None:
         pixel_values_flat_video = kwargs.pop("pixel_values_flat_video", None)
         video_num_patches = kwargs.pop("video_num_patches", None)
         video_embeds = kwargs.pop("image_embeds", None)
@@ -804,8 +813,10 @@ class InternVLChatModel(
         ]
         return image_embeds.split(image_feature_sizes)
 
-    def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict:
-        modalities = {}
+    def _parse_and_validate_multimodal_inputs(
+        self, **kwargs: object
+    ) -> InternVLMultiModalInputs:
+        modalities: InternVLMultiModalInputs = {}
 
         # Preserve the order of modalities if there are multiple of them
         # from the order of kwargs.
@@ -834,12 +845,14 @@ class InternVLChatModel(
         for modality in modalities:
             if modality == "images":
                 image_input = modalities["images"]
-                image_embeddings = self._process_vision_input(image_input)
-                multimodal_embeddings += tuple(image_embeddings)
+                if image_input is not None:
+                    image_embeddings = self._process_vision_input(image_input)
+                    multimodal_embeddings += tuple(image_embeddings)
             if modality == "videos":
                 video_input = modalities["videos"]
-                video_embeddings = self._process_vision_input(video_input)
-                multimodal_embeddings += tuple(video_embeddings)
+                if video_input is not None:
+                    video_embeddings = self._process_vision_input(video_input)
+                    multimodal_embeddings += tuple(video_embeddings)
 
         return multimodal_embeddings
 
diff --git a/vllm/model_executor/models/iquest_loopcoder.py b/vllm/model_executor/models/iquest_loopcoder.py
index d3b670658a..782bec0ed0 100644
--- a/vllm/model_executor/models/iquest_loopcoder.py
+++ b/vllm/model_executor/models/iquest_loopcoder.py
@@ -16,7 +16,7 @@
 
 from __future__ import annotations
 
-from collections.abc import Iterable
+from collections.abc import Callable, Iterable
 from dataclasses import replace
 from typing import Any
 
@@ -487,11 +487,14 @@ class IQuestLoopCoderModel(nn.Module):
                     continue
                 if name.endswith("scale"):
                     # Remapping the name of FP8 kv-scale.
-                    name = maybe_remap_kv_scale_name(name, params_dict)
-                    if name is None:
+                    remapped_name = maybe_remap_kv_scale_name(name, params_dict)
+                    if remapped_name is None:
                         continue
+                    name = remapped_name
                 param = params_dict[name]
-                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader: Callable[..., None] = getattr(
+                    param, "weight_loader", default_weight_loader
+                )
                 if weight_loader == default_weight_loader:
                     weight_loader(param, loaded_weight)
                 else:
@@ -518,9 +521,10 @@ class IQuestLoopCoderModel(nn.Module):
                 if name.endswith(".bias") and name not in params_dict:
                     continue
                 # Remapping the name of FP8 kv-scale.
-                name = maybe_remap_kv_scale_name(name, params_dict)
-                if name is None:
+                remapped_name = maybe_remap_kv_scale_name(name, params_dict)
+                if remapped_name is None:
                     continue
+                name = remapped_name
                 param = params_dict[name]
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)
diff --git a/vllm/model_executor/models/isaac.py b/vllm/model_executor/models/isaac.py
index 024477e8e7..1a97288d83 100644
--- a/vllm/model_executor/models/isaac.py
+++ b/vllm/model_executor/models/isaac.py
@@ -901,7 +901,11 @@ class IsaacForConditionalGeneration(
         for mm_feature in sorted(mm_features, key=lambda f: f.mm_position.offset):
             offset = mm_feature.mm_position.offset
             if mm_feature.modality == "image":
-                t, h, w = mm_feature.data["image_grid_thw"].data.tolist()
+                mm_data = mm_feature.data
+                assert mm_data is not None
+                grid_thw = mm_data["image_grid_thw"].data
+                assert isinstance(grid_thw, torch.Tensor)
+                t, h, w = grid_thw.tolist()
                 assert t == 1, f"Image must have 1 frame, got {t}"
                 yield offset, h // spatial_merge_size, w // spatial_merge_size
             else:
@@ -912,7 +916,7 @@ class IsaacForConditionalGeneration(
         input_tokens: list[int],
         mm_features: list[MultiModalFeatureSpec],
     ) -> tuple[torch.Tensor, int]:
-        llm_pos_ids_list = []
+        llm_pos_ids_list: list[np.ndarray] = []
         st = 0
         for offset, llm_grid_h, llm_grid_w in self.iter_mm_grid_hw(
             input_tokens, mm_features
diff --git a/vllm/model_executor/models/qwen3_vl.py b/vllm/model_executor/models/qwen3_vl.py
index 5b45ce0c9e..c36d78ce7a 100644
--- a/vllm/model_executor/models/qwen3_vl.py
+++ b/vllm/model_executor/models/qwen3_vl.py
@@ -1753,7 +1753,7 @@ class Qwen3VLForConditionalGeneration(
     SupportsEagle3,
     SupportsMultiModalPruning,
 ):
-    packed_modules_mapping = {
+    packed_modules_mapping: dict[str, list[str]] = {
         "qkv_proj": [
             "q_proj",
             "k_proj",