biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[MM] Pass prefix parameter to MMEncoderAttention (#33674)

Browse Source 
Signed-off-by: shen-shanshan <467638484@qq.com>
This commit is contained in:
Shanshan Shen
2026-02-03 22:47:41 +08:00
committed by GitHub
parent f3d8a34671
commit 5c4f2dd6ef
15 changed files with 58 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm/model_executor/models/aimv2.py
View File

@@ -127,7 +127,10 @@ class AIMv2Attention(nn.Module):
self.num_heads_per_partition = divide(self.num_heads, self.tp_size)
self.attn = MMEncoderAttention(
self.num_heads_per_partition, self.head_dim, self.scale
self.num_heads_per_partition,
self.head_dim,
self.scale,
prefix=prefix,
)
def forward(self, x: torch.Tensor) -> torch.Tensor:

5
vllm/model_executor/models/blip.py
View File

@@ -123,7 +123,10 @@ class BlipAttention(nn.Module):
self.num_heads_per_partition = divide(self.num_heads, self.tp_size)
self.attn = MMEncoderAttention(
self.num_heads_per_partition, self.head_dim, self.scale
self.num_heads_per_partition,
self.head_dim,
self.scale,
prefix=prefix,
)
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):

1
vllm/model_executor/models/glm4_1v.py
View File

@@ -296,6 +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,
)
self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True)

5
vllm/model_executor/models/glm4v.py
View File

@@ -136,7 +136,10 @@ class EVA2CLIPAttention(nn.Module):
)
self.attn = MMEncoderAttention(
self.num_heads_per_rank, self.head_dim, self.scale
self.num_heads_per_rank,
self.head_dim,
self.scale,
prefix=prefix,
)
self.output_dropout = torch.nn.Dropout(config.dropout_prob)

5
vllm/model_executor/models/idefics2_vision_model.py
View File

@@ -163,7 +163,10 @@ class Idefics2VisionAttention(nn.Module):
)
# Use unified MMEncoderAttention with Flash Attention support
self.attn = MMEncoderAttention(
self.num_heads_per_partition, self.head_dim, self.scale
self.num_heads_per_partition,
self.head_dim,
self.scale,
prefix=prefix,
)
def forward(

5
vllm/model_executor/models/intern_vit.py
View File

@@ -212,7 +212,10 @@ class InternParallelAttention(nn.Module):
)
self.attn = MMEncoderAttention(
self.num_heads_per_partition, self.head_dim, self.scale
self.num_heads_per_partition,
self.head_dim,
self.scale,
prefix=prefix,
)
def _apply_qk_norm(self, q: torch.Tensor, k: torch.Tensor):

14
vllm/model_executor/models/interns1_vit.py
View File

@@ -170,6 +170,7 @@ class InternSdpaAttention(nn.Module):
config: PretrainedConfig,
*,
num_dummy_heads: int = 0,
prefix: str = "",
) -> None:
super().__init__()
@@ -215,7 +216,12 @@ class InternSdpaAttention(nn.Module):
self.projection_layer = nn.Linear(self.dummy_dim, self.embed_dim)
# Use unified MMEncoderAttention with automatic backend selection
self.attn = MMEncoderAttention(self.num_heads, self.head_dim, self.scale)
self.attn = MMEncoderAttention(
self.num_heads,
self.head_dim,
self.scale,
prefix=prefix,
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""x shape: (B, N, C)"""
@@ -313,7 +319,11 @@ class InternS1VisionLayer(nn.Module):
num_dummy_heads: int,
prefix: str = "",
):
return InternSdpaAttention(config, num_dummy_heads=num_dummy_heads)
return InternSdpaAttention(
config,
num_dummy_heads=num_dummy_heads,
prefix=prefix,
)
def forward(
self,

5
vllm/model_executor/models/mllama4.py
View File

@@ -254,7 +254,10 @@ class Llama4VisionAttention(nn.Module):
self.scaling = self.head_dim**-0.5
self.attn = MMEncoderAttention(
self.num_local_heads, self.head_dim, self.scaling
self.num_local_heads,
self.head_dim,
self.scaling,
prefix=prefix,
)
if use_data_parallel:

6
vllm/model_executor/models/molmo.py
View File

@@ -231,7 +231,11 @@ class MultiHeadDotProductAttention(nn.Module):
self.scale = self.head_dim**-0.5
self.attn = MMEncoderAttention(
self.num_heads, self.head_dim, self.scale, num_kv_heads=self.num_kv_heads
self.num_heads,
self.head_dim,
self.scale,
num_kv_heads=self.num_kv_heads,
prefix=prefix,
)
def forward(

1
vllm/model_executor/models/molmo2.py
View File

@@ -611,6 +611,7 @@ class ImagePoolingAttention(nn.Module):
self.head_dim,
self.scale,
num_kv_heads=self.num_kv_heads,
prefix=prefix,
)
def forward_sdpa(

1
vllm/model_executor/models/qwen2_5_vl.py
View File

@@ -345,6 +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,
)
self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True)

1
vllm/model_executor/models/qwen2_vl.py
View File

@@ -319,6 +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,
)
self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True)

1
vllm/model_executor/models/qwen3_omni_moe_thinker.py
View File

@@ -194,6 +194,7 @@ class Qwen3OmniMoeAudioAttention(nn.Module):
num_heads=self.num_local_heads,
head_size=self.head_dim,
scale=self.scaling,
prefix=prefix,
)
def forward(

7
vllm/model_executor/models/step3_vl.py
View File

@@ -759,7 +759,12 @@ class Step3VisionAttention(nn.Module):
)
# Use unified MMEncoderAttention with automatic backend selection
self.attn = MMEncoderAttention(self.num_heads, self.head_dim, self.scale)
self.attn = MMEncoderAttention(
self.num_heads,
self.head_dim,
self.scale,
prefix=prefix,
)
def forward(
self,

7
vllm/model_executor/models/step_vl.py
View File

@@ -220,7 +220,12 @@ class PerceptionEncoderVisionAttention(nn.Module):
prefix=f"{prefix}.out_proj",
disable_tp=use_data_parallel,
)
self.attn = MMEncoderAttention(self.num_heads, self.head_dim, self.scale)
self.attn = MMEncoderAttention(
self.num_heads,
self.head_dim,
self.scale,
prefix=prefix,
)
self.rope = PerceptionEncoderRope2D(
dim=self.head_dim,
max_grid_height=max_grid_height,