xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 02:05:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix interns1-vit qk norm code path (#27480)

Browse Source 
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py 2025-10-25 01:43:45 +08:00 committed by GitHub
parent 0f67d4d962
commit acc78aeb88
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 3 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/model_executor/models/interns1_vit.py
View File

@ -217,16 +217,15 @@ class InternSdpaAttention(nn.Module):
self.attn = MultiHeadAttention(self.num_heads, self.head_dim, self.scale)
def forward(self, x: torch.Tensor) -> torch.Tensor:
B, N, C = x.shape
"""x shape: (B, N, C)"""
q = self.q_proj(x)
k = self.k_proj(x)
v = self.v_proj(x)
if self.qk_normalization:
B_, N_, H_, D_ = q.shape
q = self.q_norm(q.flatten(-2, -1)).view(B_, N_, H_, D_)
k = self.k_norm(k.flatten(-2, -1)).view(B_, N_, H_, D_)
q = self.q_norm(q)
k = self.k_norm(k)
# Use unified MultiHeadAttention with automatic backend selection
x = self.attn(q, k, v)