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

[Model] Switch to Fused RMS norm in Qwen2.5_VL model. (#22184)

Browse Source 
Signed-off-by: kf <kuanfu.liu@embeddedllm.com>
Signed-off-by: tjtanaavllm <tunjian.tan@amd.com>
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
Co-authored-by: kf <kuanfu.liu@embeddedllm.com>
This commit is contained in:
vllmellm
2025-08-07 14:05:24 +08:00
committed by GitHub
parent 4d4297e8fe
commit cbc8457b26
1 changed files with 7 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
vllm/model_executor/models/qwen2_5_vl.py
View File

@@ -396,13 +396,13 @@ class Qwen2_5_VisionBlock(nn.Module):
max_seqlen: Optional[int] = None, # Only used for Flash Attention max_seqlen: Optional[int] = None, # Only used for Flash Attention
seqlens: Optional[list[int]] = None, # Only used for xFormers seqlens: Optional[list[int]] = None, # Only used for xFormers
) -> torch.Tensor: ) -> torch.Tensor:
x = x + self.attn(self.norm1(x), x_attn = self.attn(self.norm1(x),
cu_seqlens=cu_seqlens, cu_seqlens=cu_seqlens,
rotary_pos_emb=rotary_pos_emb, rotary_pos_emb=rotary_pos_emb,
max_seqlen=max_seqlen, max_seqlen=max_seqlen,
seqlens=seqlens) seqlens=seqlens)
x_fused_norm, residual = self.norm2(x, residual=x_attn)
x = x + self.mlp(self.norm2(x)) x = residual + self.mlp(x_fused_norm)
return x return x