diff --git a/vllm/model_executor/specialized_models/deepseek_v3_2_nvfp4/layer.py b/vllm/model_executor/specialized_models/deepseek_v3_2_nvfp4/layer.py
index 9ddbb3f3909..70a7b31362f 100644
--- a/vllm/model_executor/specialized_models/deepseek_v3_2_nvfp4/layer.py
+++ b/vllm/model_executor/specialized_models/deepseek_v3_2_nvfp4/layer.py
@@ -421,7 +421,7 @@ class DeepseekV32DecoderLayer(nn.Module):
             )
             return out
 
-        shared_experts.forward = _fused_forward
+        shared_experts.forward = _fused_forward  # type: ignore[method-assign]
 
 
 class DeepseekV32MLAAttention(nn.Module):
diff --git a/vllm/model_executor/specialized_models/deepseek_v3_2_nvfp4/mtp.py b/vllm/model_executor/specialized_models/deepseek_v3_2_nvfp4/mtp.py
index 8aa7e7a7fae..1b61f3620da 100644
--- a/vllm/model_executor/specialized_models/deepseek_v3_2_nvfp4/mtp.py
+++ b/vllm/model_executor/specialized_models/deepseek_v3_2_nvfp4/mtp.py
@@ -43,6 +43,7 @@ class DeepSeekMultiTokenPredictorLayer(DeepSeekMultiTokenPredictorLayerBase):
     def __init__(self, vllm_config: VllmConfig, prefix: str) -> None:
         nn.Module.__init__(self)
 
+        assert vllm_config.speculative_config is not None
         config = vllm_config.speculative_config.draft_model_config.hf_config
         quant_config = vllm_config.quant_config
 
diff --git a/vllm/model_executor/specialized_models/deepseek_v3_2_nvfp4/sparse_indexer.py b/vllm/model_executor/specialized_models/deepseek_v3_2_nvfp4/sparse_indexer.py
index e901a5cdaaf..cd127d11899 100644
--- a/vllm/model_executor/specialized_models/deepseek_v3_2_nvfp4/sparse_indexer.py
+++ b/vllm/model_executor/specialized_models/deepseek_v3_2_nvfp4/sparse_indexer.py
@@ -48,6 +48,7 @@ def sparse_attn_indexer(
         _ = torch.empty(max_logits_elems, dtype=torch.uint8, device=q_fp8.device)
         return None
 
+    assert isinstance(attn_metadata, dict)
     attn_metadata = attn_metadata[k_cache_prefix]
     assert isinstance(attn_metadata, DeepseekV32IndexerMetadata)
     has_decode = attn_metadata.num_decodes > 0