[MM] Align the prefix of MMEncoderAttention with Attention (#33750)

Signed-off-by: shen-shanshan <467638484@qq.com>

vllm/model_executor/models/aimv2.py

@@ -130,7 +130,7 @@ class AIMv2Attention(nn.Module):
             self.num_heads_per_partition,
             self.head_dim,
             self.scale,
-            prefix=prefix,
+            prefix=f"{prefix}.attn",
         )
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:

vllm/model_executor/models/blip.py

@@ -126,7 +126,7 @@ class BlipAttention(nn.Module):
             self.num_heads_per_partition,
             self.head_dim,
             self.scale,
-            prefix=prefix,
+            prefix=f"{prefix}.attn",
         )
 
     def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):

vllm/model_executor/models/glm4_1v.py

@@ -296,7 +296,7 @@ class Glm4vVisionAttention(nn.Module):
             num_heads=self.num_attention_heads_per_partition,
             head_size=self.hidden_size_per_attention_head,
             scale=self.hidden_size_per_attention_head**-0.5,
-            prefix=prefix,
+            prefix=f"{prefix}.attn",
         )
 
         self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True)

vllm/model_executor/models/glm4v.py

@@ -139,7 +139,7 @@ class EVA2CLIPAttention(nn.Module):
             self.num_heads_per_rank,
             self.head_dim,
             self.scale,
-            prefix=prefix,
+            prefix=f"{prefix}.attn",
         )
         self.output_dropout = torch.nn.Dropout(config.dropout_prob)
 

vllm/model_executor/models/glm_ocr.py

@@ -137,6 +137,7 @@ class GlmOcrVisionAttention(nn.Module):
             num_heads=self.num_attention_heads_per_partition,
             head_size=self.hidden_size_per_attention_head,
             scale=self.hidden_size_per_attention_head**-0.5,
+            prefix=f"{prefix}.attn",
         )
         self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True)
 

vllm/model_executor/models/idefics2_vision_model.py

@@ -166,7 +166,7 @@ class Idefics2VisionAttention(nn.Module):
             self.num_heads_per_partition,
             self.head_dim,
             self.scale,
-            prefix=prefix,
+            prefix=f"{prefix}.attn",
         )
 
     def forward(

vllm/model_executor/models/intern_vit.py

@@ -215,7 +215,7 @@ class InternParallelAttention(nn.Module):
             self.num_heads_per_partition,
             self.head_dim,
             self.scale,
-            prefix=prefix,
+            prefix=f"{prefix}.attn",
         )
 
     def _apply_qk_norm(self, q: torch.Tensor, k: torch.Tensor):

vllm/model_executor/models/interns1_vit.py

@@ -220,7 +220,7 @@ class InternSdpaAttention(nn.Module):
             self.num_heads,
             self.head_dim,
             self.scale,
-            prefix=prefix,
+            prefix=f"{prefix}.attn",
         )
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:

vllm/model_executor/models/mllama4.py

@@ -257,7 +257,7 @@ class Llama4VisionAttention(nn.Module):
             self.num_local_heads,
             self.head_dim,
             self.scaling,
-            prefix=prefix,
+            prefix=f"{prefix}.attn",
         )
 
         if use_data_parallel:

vllm/model_executor/models/molmo.py

@@ -235,7 +235,7 @@ class MultiHeadDotProductAttention(nn.Module):
             self.head_dim,
             self.scale,
             num_kv_heads=self.num_kv_heads,
-            prefix=prefix,
+            prefix=f"{prefix}.attn",
         )
 
     def forward(

vllm/model_executor/models/molmo2.py

@@ -611,7 +611,7 @@ class ImagePoolingAttention(nn.Module):
                 self.head_dim,
                 self.scale,
                 num_kv_heads=self.num_kv_heads,
-                prefix=prefix,
+                prefix=f"{prefix}.attn",
             )
 
     def forward_sdpa(

vllm/model_executor/models/openpangu_vl.py

@@ -125,6 +125,7 @@ class OpenPanguVisionAttention(nn.Module):
             num_heads=self.num_attention_heads_per_partition,
             head_size=self.hidden_size_per_attention_head,
             scale=self.hidden_size_per_attention_head**-0.5,
+            prefix=f"{prefix}.attn",
         )
         self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True)
 

vllm/model_executor/models/qwen2_5_vl.py

@@ -345,7 +345,7 @@ class Qwen2_5_VisionAttention(nn.Module):
             num_heads=self.num_attention_heads_per_partition,
             head_size=self.hidden_size_per_attention_head,
             scale=self.hidden_size_per_attention_head**-0.5,
-            prefix=prefix,
+            prefix=f"{prefix}.attn",
         )
 
         self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True)

vllm/model_executor/models/qwen2_vl.py

@@ -319,7 +319,7 @@ class Qwen2VisionAttention(nn.Module):
             num_heads=self.num_attention_heads_per_partition,
             head_size=self.hidden_size_per_attention_head,
             scale=self.hidden_size_per_attention_head**-0.5,
-            prefix=prefix,
+            prefix=f"{prefix}.attn",
         )
 
         self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True)

vllm/model_executor/models/qwen3_omni_moe_thinker.py

@@ -194,7 +194,7 @@ class Qwen3OmniMoeAudioAttention(nn.Module):
             num_heads=self.num_local_heads,
             head_size=self.head_dim,
             scale=self.scaling,
-            prefix=prefix,
+            prefix=f"{prefix}.attn",
         )
 
     def forward(

vllm/model_executor/models/step3_vl.py

@@ -763,7 +763,7 @@ class Step3VisionAttention(nn.Module):
             self.num_heads,
             self.head_dim,
             self.scale,
-            prefix=prefix,
+            prefix=f"{prefix}.attn",
         )
 
     def forward(

vllm/model_executor/models/step_vl.py

@@ -224,7 +224,7 @@ class PerceptionEncoderVisionAttention(nn.Module):
             self.num_heads,
             self.head_dim,
             self.scale,
-            prefix=prefix,
+            prefix=f"{prefix}.attn",
         )
         self.rope = PerceptionEncoderRope2D(
             dim=self.head_dim,