DeepSeek/DeepSeek-VL2
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepSeek-VL2 synced 2025-01-22 12:25:32 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Update siglip_vit.py

Browse Source
This commit is contained in:
StevenLiuWen 2025-01-16 11:25:10 +08:00
parent 9dd177671e
commit ff23960c5c
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
deepseek_vl2/models/siglip_vit.py
View File

@ -139,7 +139,7 @@ class Attention(nn.Module):
if not self.qk_norm: if not self.qk_norm:
if self.head_dim % 32 == 0 and is_flash_attn_2_available(): if self.head_dim % 32 == 0 and is_flash_attn_2_available():
# flashattn的head_dim必须是32的倍数，SigLIP-SO400M无法使用flashattn # flashattn must have head_dim as a multiple of 32
x = flash_attn_qkvpacked_func(qkv, dropout_p=self.attn_drop.p if self.training else 0., x = flash_attn_qkvpacked_func(qkv, dropout_p=self.attn_drop.p if self.training else 0.,
deterministic=self.deterministic) deterministic=self.deterministic)
else: else: