squall/flash-attention
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #60 from 201419/patch-1

Browse Source 
fix typo in function mha_fwd
This commit is contained in:
Tri Dao 2022-10-17 09:38:48 -07:00 committed by GitHub
commit 1d0b41be3b
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
csrc/flash_attn/fmha_api.cpp
View File

@ -216,7 +216,7 @@ mha_fwd(const at::Tensor &q, // total_q x num_heads x head_size, total_q
TORCH_CHECK(k.stride(-1) == 1);
TORCH_CHECK(v.stride(-1) == 1);
TORCH_CHECK(out.stride(-1) == 1);
TORCH_CHECK(cu_seqlens_k.is_contiguous());
TORCH_CHECK(cu_seqlens_q.is_contiguous());
TORCH_CHECK(cu_seqlens_k.is_contiguous());
const auto sizes = q.sizes();