@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 9fb27dd3b3

Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 8, 22:45 UTC

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

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

Shanshan Shen committed Feb 4, 2026 at 12:07 UTC 9fb27dd3b3530cbf958a24ee2dcfc907acbf5455

17 files changed +17 -15

vllm/model_executor/models/aimv2.py

+1 -1

@@ -130,7 +130,7 @@ class AIMv2Attention(nn.Module):

130 self.num_heads_per_partition,

131 self.head_dim,

132 self.scale,

133 - prefix=prefix,

133 + prefix=f"{prefix}.attn",

134 )

135

136 def forward(self, x: torch.Tensor) -> torch.Tensor:

vllm/model_executor/models/blip.py

+1 -1

@@ -126,7 +126,7 @@ class BlipAttention(nn.Module):

126 self.num_heads_per_partition,

127 self.head_dim,

128 self.scale,

129 - prefix=prefix,

129 + prefix=f"{prefix}.attn",

130 )

131

132 def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):

vllm/model_executor/models/glm4_1v.py

+1 -1

@@ -296,7 +296,7 @@ class Glm4vVisionAttention(nn.Module):

296 num_heads=self.num_attention_heads_per_partition,

297 head_size=self.hidden_size_per_attention_head,

298 scale=self.hidden_size_per_attention_head**-0.5,

299 - prefix=prefix,

299 + prefix=f"{prefix}.attn",

300 )

301

302 self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True)

vllm/model_executor/models/glm4v.py

+1 -1

@@ -139,7 +139,7 @@ class EVA2CLIPAttention(nn.Module):

139 self.num_heads_per_rank,

140 self.head_dim,

141 self.scale,

142 - prefix=prefix,

142 + prefix=f"{prefix}.attn",

143 )

144 self.output_dropout = torch.nn.Dropout(config.dropout_prob)

145

vllm/model_executor/models/glm_ocr.py

@@ -137,6 +137,7 @@ class GlmOcrVisionAttention(nn.Module):

137 num_heads=self.num_attention_heads_per_partition,

138 head_size=self.hidden_size_per_attention_head,

139 scale=self.hidden_size_per_attention_head**-0.5,

140 + prefix=f"{prefix}.attn",

141 )

142 self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True)

143

vllm/model_executor/models/idefics2_vision_model.py

+1 -1

@@ -166,7 +166,7 @@ class Idefics2VisionAttention(nn.Module):

166 self.num_heads_per_partition,

167 self.head_dim,

168 self.scale,

169 - prefix=prefix,

169 + prefix=f"{prefix}.attn",

170 )

171

172 def forward(

vllm/model_executor/models/intern_vit.py

+1 -1

@@ -215,7 +215,7 @@ class InternParallelAttention(nn.Module):

215 self.num_heads_per_partition,

216 self.head_dim,

217 self.scale,

218 - prefix=prefix,

218 + prefix=f"{prefix}.attn",

219 )

220

221 def _apply_qk_norm(self, q: torch.Tensor, k: torch.Tensor):

vllm/model_executor/models/interns1_vit.py

+1 -1

@@ -220,7 +220,7 @@ class InternSdpaAttention(nn.Module):

220 self.num_heads,

221 self.head_dim,

222 self.scale,

223 - prefix=prefix,

223 + prefix=f"{prefix}.attn",

224 )

225

226 def forward(self, x: torch.Tensor) -> torch.Tensor:

vllm/model_executor/models/mllama4.py

+1 -1

@@ -257,7 +257,7 @@ class Llama4VisionAttention(nn.Module):

257 self.num_local_heads,

258 self.head_dim,

259 self.scaling,

260 - prefix=prefix,

260 + prefix=f"{prefix}.attn",

261 )

262

263 if use_data_parallel:

vllm/model_executor/models/molmo.py

+1 -1

@@ -235,7 +235,7 @@ class MultiHeadDotProductAttention(nn.Module):

235 self.head_dim,

236 self.scale,

237 num_kv_heads=self.num_kv_heads,

238 - prefix=prefix,

238 + prefix=f"{prefix}.attn",

239 )

240

241 def forward(

vllm/model_executor/models/molmo2.py

+1 -1

@@ -611,7 +611,7 @@ class ImagePoolingAttention(nn.Module):

611 self.head_dim,

612 self.scale,

613 num_kv_heads=self.num_kv_heads,

614 - prefix=prefix,

614 + prefix=f"{prefix}.attn",

615 )

616

617 def forward_sdpa(

vllm/model_executor/models/openpangu_vl.py

@@ -125,6 +125,7 @@ class OpenPanguVisionAttention(nn.Module):

125 num_heads=self.num_attention_heads_per_partition,

126 head_size=self.hidden_size_per_attention_head,

127 scale=self.hidden_size_per_attention_head**-0.5,

128 + prefix=f"{prefix}.attn",

129 )

130 self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True)

131

vllm/model_executor/models/qwen2_5_vl.py

+1 -1

@@ -345,7 +345,7 @@ class Qwen2_5_VisionAttention(nn.Module):

345 num_heads=self.num_attention_heads_per_partition,

346 head_size=self.hidden_size_per_attention_head,

347 scale=self.hidden_size_per_attention_head**-0.5,

348 - prefix=prefix,

348 + prefix=f"{prefix}.attn",

349 )

350

351 self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True)

vllm/model_executor/models/qwen2_vl.py

+1 -1

@@ -319,7 +319,7 @@ class Qwen2VisionAttention(nn.Module):

319 num_heads=self.num_attention_heads_per_partition,

320 head_size=self.hidden_size_per_attention_head,

321 scale=self.hidden_size_per_attention_head**-0.5,

322 - prefix=prefix,

322 + prefix=f"{prefix}.attn",

323 )

324

325 self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True)

vllm/model_executor/models/qwen3_omni_moe_thinker.py

+1 -1

@@ -194,7 +194,7 @@ class Qwen3OmniMoeAudioAttention(nn.Module):

194 num_heads=self.num_local_heads,

195 head_size=self.head_dim,

196 scale=self.scaling,

197 - prefix=prefix,

197 + prefix=f"{prefix}.attn",

198 )

199

200 def forward(

vllm/model_executor/models/step3_vl.py

+1 -1

@@ -763,7 +763,7 @@ class Step3VisionAttention(nn.Module):

763 self.num_heads,

764 self.head_dim,

765 self.scale,

766 - prefix=prefix,

766 + prefix=f"{prefix}.attn",

767 )

768

769 def forward(

vllm/model_executor/models/step_vl.py

+1 -1

@@ -224,7 +224,7 @@ class PerceptionEncoderVisionAttention(nn.Module):

224 self.num_heads,

225 self.head_dim,

226 self.scale,

227 - prefix=prefix,

227 + prefix=f"{prefix}.attn",

228 )

229 self.rope = PerceptionEncoderRope2D(

230 dim=self.head_dim,