dpad same as dvpad for flash attention integration

This commit is contained in:
danyao12
2024-08-05 17:18:10 +00:00
parent 3d5b0755ef
commit 25db133926

View File

@@ -496,6 +496,7 @@ def get_bwd_dq_dk_dv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
cond = dtype in ['fp16', 'bf16']
cond &= bias in ['no', 'alibi']
cond &= dropout in ['no', 'dropout_wg32', 'dropout_wg16']
cond &= dpad == dvpad
if not cond:
continue
if receipt == 3: