riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-08-02 19:21:31 +06:00
Code Issues Actions 3 Packages Projects Releases Wiki Activity

[AttentionMaskConverter] fix sdpa unmask unattended (#28369)

Browse Source 
fix tensor device
This commit is contained in:
zspo 2024-01-08 20:33:44 +08:00 committed by GitHub
parent 98dba52ccd
commit 87a6cf41d0
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
src/transformers/modeling_attn_mask_utils.py
View File

@ -234,8 +234,8 @@ class AttentionMaskConverter:
# Get the index of the first non-zero value for every sample in the batch.
# In the above example, indices = [[2], [0], [1]]]
tmp = torch.arange(attention_mask.shape[1], 0, -1)
indices = torch.argmax(attention_mask.cpu() * tmp, 1, keepdim=True)
tmp = torch.arange(attention_mask.shape[1], 0, -1, device=attention_mask.device)
indices = torch.argmax(attention_mask * tmp, 1, keepdim=True)
# Find the batch indexes that have unattended tokens on the leftmost side (e.g. [0, 0, 1, 1, 1]), for which the first rows of the
# expanded mask will be completely unattended.