Skip to content

Handle zero-causal PrefixLM FA3 pass as no-op #13

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Zane12518 wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Handle zero-causal PrefixLM FA3 pass as no-op #13

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
35a92ef
Handle zero-causal PrefixLM FA3 pass as no-op
Zane12518 May 27, 2026
428dfc5
Avoid zero tensors for zero-causal FA3 no-op
Zane12518 May 29, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
46 changes: 26 additions & 20 deletions models/flash_attention_prefixlm_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -120,12 +120,15 @@ def flash_attn_varlen_prefixlm_compileop(
max_seqlen_q=max_seqlen_prefix_int, max_seqlen_k=max_seqlen_prefix_int,
causal=is_causal)
# Fwd pass 2 (causal)
_, softmax_lse_causal = _custom_flash_attn_forward(
out_=out, q=q, k=k, v=v,
cu_seqlens_q=cu_seqlens_shifted, cu_seqlens_k=cu_seqlens,
seqused_q=causal_lens,
max_seqlen_q=max_seqlen_causal_int, max_seqlen_k=max_seqlen_all_int,
causal=True)
if max_seqlen_causal_int > 0:
_, softmax_lse_causal = _custom_flash_attn_forward(
out_=out, q=q, k=k, v=v,
cu_seqlens_q=cu_seqlens_shifted, cu_seqlens_k=cu_seqlens,
seqused_q=causal_lens,
max_seqlen_q=max_seqlen_causal_int, max_seqlen_k=max_seqlen_all_int,
causal=True)
else:
softmax_lse_causal = torch.empty_like(softmax_lse_bidir)

out[total_seqlen_int:] = 0 # Zero padding
return out, softmax_lse_bidir, softmax_lse_causal
Expand Down Expand Up @@ -181,7 +184,6 @@ def flash_attn_varlen_prefixlm_bwd_compileop(
# Buffers
dq = torch.empty_like(q)
dk1, dv1 = torch.zeros_like(k), torch.zeros_like(v) # Zero-fill in advance
dk2, dv2 = torch.empty_like(k), torch.empty_like(v)
# Bwd pass 1 (bidirectional)
_flash_attn_backward(
dout=dout, q=q, k=k, v=v, out=out,
Expand All @@ -194,22 +196,26 @@ def flash_attn_varlen_prefixlm_bwd_compileop(
dv=dv1,
is_causal=is_causal)
# Bwd pass 2 (causal)
_flash_attn_backward(
dout=dout, q=q, k=k, v=v, out=out,
softmax_lse=softmax_lse_causal,
cu_seqlens_q=cu_seqlens_shifted, cu_seqlens_k=cu_seqlens,
max_seqlen_q=max_seqlen_causal_int, max_seqlen_k=max_seqlen_all_int,
sequed_q=causal_lens,
dq=dq,
dk=dk2,
dv=dv2,
is_causal=True)
if max_seqlen_causal_int > 0:
dk2, dv2 = torch.empty_like(k), torch.empty_like(v)
_flash_attn_backward(
dout=dout, q=q, k=k, v=v, out=out,
softmax_lse=softmax_lse_causal,
cu_seqlens_q=cu_seqlens_shifted, cu_seqlens_k=cu_seqlens,
max_seqlen_q=max_seqlen_causal_int, max_seqlen_k=max_seqlen_all_int,
sequed_q=causal_lens,
dq=dq,
dk=dk2,
dv=dv2,
is_causal=True)
dk2[total_seqlen_int:] = 0
dv2[total_seqlen_int:] = 0
dk1.add_(dk2)
dv1.add_(dv2)

# Zero padding grads
dq[total_seqlen_int:] = 0
dk2[total_seqlen_int:] = 0
dv2[total_seqlen_int:] = 0
return dq, dk1.add_(dk2), dv1.add_(dv2)
return dq, dk1, dv1


@torch.library.register_fake("flash_attn::flash_attn_varlen_prefixlm_bwd_compileop")
Expand Down