Skip to content

Compile SDPA with bounded tensor-dependent batch sizes #1257

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

Merged
aw471 merged 5 commits into from
Oct 23, 2026
Merged

Compile SDPA with bounded tensor-dependent batch sizes #1257

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
f6b6bda
Avoid symbolic batch iteration in single-chunk attention
aw471 Oct 22, 2026
1d19ebd
Merge branch 'main' into compile-attention-symbolic-chunk-loop
aw471 Oct 23, 2026
cfef9e0
Inline attention chunk iteration and clarify compilation comment
aw471 Oct 23, 2026
85f220b
Shorten compilation comment and place it beside attention loop
aw471 Oct 23, 2026
765c509
Apply suggestion from @aw471
aw471 Oct 23, 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
8 changes: 6 additions & 2 deletions sdm/nn/attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -171,7 +171,7 @@ def forward(
key = key.transpose(-3, -2) # [B, Hkv, KV, C]
value = value.transpose(-3, -2) # [B, Hkv, KV, C]

# CUDA attention kernels accept at most 65,535 batch entries per call:
# CUDA attention kernels accept at most 65,535 batch entries per call.
outs = [
F.scaled_dot_product_attention(
query=query[i : i + 65_535],
Expand All @@ -185,7 +185,11 @@ def forward(
is_causal=is_causal,
scale=self.scale,
).transpose(-3, -2) # [B, Q, Hq, C]
for i in range(0, query.size(0), 65_535)
for i in (
(0,)
if query.size(0) <= 65_535
else range(0, query.size(0), 65_535)
)
]
out = outs[0] if len(outs) == 1 else torch.cat(outs)

Expand Down
Loading