From 9fb27dd3b3530cbf958a24ee2dcfc907acbf5455 Mon Sep 17 00:00:00 2001 From: Shanshan Shen <467638484@qq.com> Date: Wed, 4 Feb 2026 12:07:30 +0800 Subject: [PATCH] [MM] Align the prefix of MMEncoderAttention with Attention (#33750) Signed-off-by: shen-shanshan <467638484@qq.com> --- vllm/model_executor/models/aimv2.py | 2 +- vllm/model_executor/models/blip.py | 2 +- vllm/model_executor/models/glm4_1v.py | 2 +- vllm/model_executor/models/glm4v.py | 2 +- vllm/model_executor/models/glm_ocr.py | 1 + vllm/model_executor/models/idefics2_vision_model.py | 2 +- vllm/model_executor/models/intern_vit.py | 2 +- vllm/model_executor/models/interns1_vit.py | 2 +- vllm/model_executor/models/mllama4.py | 2 +- vllm/model_executor/models/molmo.py | 2 +- vllm/model_executor/models/molmo2.py | 2 +- vllm/model_executor/models/openpangu_vl.py | 1 + vllm/model_executor/models/qwen2_5_vl.py | 2 +- vllm/model_executor/models/qwen2_vl.py | 2 +- vllm/model_executor/models/qwen3_omni_moe_thinker.py | 2 +- vllm/model_executor/models/step3_vl.py | 2 +- vllm/model_executor/models/step_vl.py | 2 +- 17 files changed, 17 insertions(+), 15 deletions(-) diff --git a/vllm/model_executor/models/aimv2.py b/vllm/model_executor/models/aimv2.py index d6716c8e58..63cb9c96e2 100644 --- a/vllm/model_executor/models/aimv2.py +++ b/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: diff --git a/vllm/model_executor/models/blip.py b/vllm/model_executor/models/blip.py index ad8f7c1af5..73b0b8af93 100644 --- a/vllm/model_executor/models/blip.py +++ b/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): diff --git a/vllm/model_executor/models/glm4_1v.py b/vllm/model_executor/models/glm4_1v.py index b2886e85a3..5333042cbf 100644 --- a/vllm/model_executor/models/glm4_1v.py +++ b/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) diff --git a/vllm/model_executor/models/glm4v.py b/vllm/model_executor/models/glm4v.py index 8bad386fa8..56504029dc 100644 --- a/vllm/model_executor/models/glm4v.py +++ b/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) diff --git a/vllm/model_executor/models/glm_ocr.py b/vllm/model_executor/models/glm_ocr.py index 90c6baacbd..d037431403 100644 --- a/vllm/model_executor/models/glm_ocr.py +++ b/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) diff --git a/vllm/model_executor/models/idefics2_vision_model.py b/vllm/model_executor/models/idefics2_vision_model.py index d6f93a9d49..b90afbe5ab 100644 --- a/vllm/model_executor/models/idefics2_vision_model.py +++ b/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( diff --git a/vllm/model_executor/models/intern_vit.py b/vllm/model_executor/models/intern_vit.py index 8cacfe06ea..8161473641 100644 --- a/vllm/model_executor/models/intern_vit.py +++ b/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): diff --git a/vllm/model_executor/models/interns1_vit.py b/vllm/model_executor/models/interns1_vit.py index 421e0ffd4d..533f0681c1 100644 --- a/vllm/model_executor/models/interns1_vit.py +++ b/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: diff --git a/vllm/model_executor/models/mllama4.py b/vllm/model_executor/models/mllama4.py index 52fdeddf4b..58f63597a7 100644 --- a/vllm/model_executor/models/mllama4.py +++ b/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: diff --git a/vllm/model_executor/models/molmo.py b/vllm/model_executor/models/molmo.py index b1330d92d7..1ee1776569 100644 --- a/vllm/model_executor/models/molmo.py +++ b/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( diff --git a/vllm/model_executor/models/molmo2.py b/vllm/model_executor/models/molmo2.py index f9664f32e4..9d996a93b0 100644 --- a/vllm/model_executor/models/molmo2.py +++ b/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( diff --git a/vllm/model_executor/models/openpangu_vl.py b/vllm/model_executor/models/openpangu_vl.py index 239ef81d3e..d7df2cbb4c 100644 --- a/vllm/model_executor/models/openpangu_vl.py +++ b/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) diff --git a/vllm/model_executor/models/qwen2_5_vl.py b/vllm/model_executor/models/qwen2_5_vl.py index c06beb97fa..c2c52fa662 100644 --- a/vllm/model_executor/models/qwen2_5_vl.py +++ b/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) diff --git a/vllm/model_executor/models/qwen2_vl.py b/vllm/model_executor/models/qwen2_vl.py index 6169e72df4..d911fb1dd9 100644 --- a/vllm/model_executor/models/qwen2_vl.py +++ b/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) diff --git a/vllm/model_executor/models/qwen3_omni_moe_thinker.py b/vllm/model_executor/models/qwen3_omni_moe_thinker.py index 96294158c7..9500ce2e2b 100755 --- a/vllm/model_executor/models/qwen3_omni_moe_thinker.py +++ b/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( diff --git a/vllm/model_executor/models/step3_vl.py b/vllm/model_executor/models/step3_vl.py index fe2bb1ac6a..f3993348b3 100644 --- a/vllm/model_executor/models/step3_vl.py +++ b/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( diff --git a/vllm/model_executor/models/step_vl.py b/vllm/model_executor/models/step_vl.py index 31b266a7e7..4669771f4b 100644 --- a/vllm/model_executor/models/step_vl.py +++ b/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,