Skip to content

Fix flash/sage varlen prep under torch.compile with dynamic shapes#14568

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
ShivamShrirao wants to merge 4 commits into
base:main
Choose a base branch
from
+10 −11
Open

Fix flash/sage varlen prep under torch.compile with dynamic shapes #14568

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
e74d38e
Fix flash/sage varlen prep under torch.compile with dynamic shapes
ShivamShrirao Aug 23, 2026
f3bfd56
Merge branch 'main' into varlen-compile-fix
ShivamShrirao Aug 24, 2026
2628666
Merge branch 'main' into varlen-compile-fix
ShivamShrirao Aug 24, 2026
669b8cc
Merge branch 'main' into varlen-compile-fix
ShivamShrirao Aug 27, 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
21 changes: 10 additions & 11 deletions src/diffusers/models/attention_dispatch.py
View file
Open in desktop
Original file line numberDiff line numberDiff line change
Expand Up@@ -599,13 +599,12 @@ def _prepare_for_flash_attn_or_sage_varlen_without_mask(
):
seqlens_q = torch.full((batch_size,), seq_len_q, dtype=torch.int32, device=device)
seqlens_k = torch.full((batch_size,), seq_len_kv, dtype=torch.int32, device=device)
cu_seqlens_q = torch.zeros(batch_size + 1, dtype=torch.int32, device=device)
cu_seqlens_k = torch.zeros(batch_size + 1, dtype=torch.int32, device=device)
cu_seqlens_q[1:] = torch.cumsum(seqlens_q, dim=0)
cu_seqlens_k[1:] = torch.cumsum(seqlens_k, dim=0)
max_seqlen_q = seqlens_q.max().item()
max_seqlen_k = seqlens_k.max().item()
return (seqlens_q, seqlens_k), (cu_seqlens_q, cu_seqlens_k), (max_seqlen_q, max_seqlen_k)
# Built with arange instead of cumsum(full(...)): inductor rewrites that pattern into
# `arange * fill_value`, which raises under dynamic shapes because the fill value is a
# symbolic sequence length. The lengths are uniform here, so arange is also cheaper.
cu_seqlens_q = torch.arange(0, (batch_size + 1) * seq_len_q, seq_len_q, dtype=torch.int32, device=device)

@sergereviewsergereviewBotAug 26, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Behavioral edge case worth noting: torch.arange with step=0 raises RuntimeError: step must be nonzero, so a zero-length query/key sequence now errors where the previous cumsum(zeros) path returned an all-zero cu_seqlens. I don't think any dispatcher path can reach here with seq_len == 0, so this is informational rather than blocking — just flagging it since the PR claims strict numerical/behavioral equivalence.

cu_seqlens_k = torch.arange(0, (batch_size + 1) * seq_len_kv, seq_len_kv, dtype=torch.int32, device=device)
return (seqlens_q, seqlens_k), (cu_seqlens_q, cu_seqlens_k), (seq_len_q, seq_len_kv)


def _prepare_for_flash_attn_or_sage_varlen_with_mask(
Expand All@@ -616,13 +615,13 @@ def _prepare_for_flash_attn_or_sage_varlen_with_mask(
):
seqlens_q = torch.full((batch_size,), seq_len_q, dtype=torch.int32, device=device)
seqlens_k = attn_mask.sum(dim=1, dtype=torch.int32)
cu_seqlens_q = torch.zeros(batch_size + 1, dtype=torch.int32, device=device)
# Queries are uniform, so arange (see the no-mask helper: cumsum(full(...)) breaks inductor
# under dynamic shapes). Keys are data-dependent and keep the cumsum.
cu_seqlens_q = torch.arange(0, (batch_size + 1) * seq_len_q, seq_len_q, dtype=torch.int32, device=device)
cu_seqlens_k = torch.zeros(batch_size + 1, dtype=torch.int32, device=device)
cu_seqlens_q[1:] = torch.cumsum(seqlens_q, dim=0)
cu_seqlens_k[1:] = torch.cumsum(seqlens_k, dim=0)
max_seqlen_q = seqlens_q.max().item()
max_seqlen_k = seqlens_k.max().item()
return (seqlens_q, seqlens_k), (cu_seqlens_q, cu_seqlens_k), (max_seqlen_q, max_seqlen_k)
return (seqlens_q, seqlens_k), (cu_seqlens_q, cu_seqlens_k), (seq_len_q, max_seqlen_k)
Comment thread
ShivamShrirao marked this conversation as resolved.


def _prepare_for_flash_attn_or_sage_varlen(
Expand Down
Loading