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

Fix dv = torch::empty_like(k) for mha_bwd_varlen as well

Browse Source
This commit is contained in:
Tri Dao 2024-02-10 01:03:00 -08:00
parent a190df011c
commit d9a5cb291c
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/flash_api.cpp
View File

@ -1069,7 +1069,7 @@ mha_varlen_bwd(const at::Tensor &dout, // total_q x num_heads, x head_size
TORCH_CHECK(dv.stride(-1) == 1, "dv must have contiguous last dimension");
CHECK_SHAPE(dv, total_k, num_heads_k, head_size);
} else {
dv = torch::empty_like(k);
dv = torch::empty_like(v);
}
at::Tensor dout_padded;