[MM Encoder]: Migrate legacy ViT MultiHeadAttention to new MMEncoderAttention interface (#30684)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>

vllm/model_executor/models/glm4v.py

@@ -19,7 +19,7 @@ from transformers import BatchFeature, PreTrainedTokenizer, TensorType
 from transformers.image_utils import ImageInput
 from transformers.tokenization_utils_base import TextInput
 
-from vllm.attention.layer import MultiHeadAttention
+from vllm.attention.layers.mm_encoder_attention import MMEncoderAttention
 from vllm.config import VllmConfig
 from vllm.config.multimodal import BaseDummyOptions
 from vllm.distributed import get_tensor_model_parallel_world_size
@@ -135,7 +135,7 @@ class EVA2CLIPAttention(nn.Module):
             prefix=f"{prefix}.dense",
         )
 
-        self.attn = MultiHeadAttention(
+        self.attn = MMEncoderAttention(
             self.num_heads_per_rank, self.head_dim, self.scale
         )
         self.output_dropout = torch.nn.Dropout(config.dropout_prob)