Skip to content

frost(sdpa): retire the bottom_right_with_swa notch — every row serves BR + SWA #623

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
vedaanta merged 1 commit into from
Aug 17, 2026
Merged

frost(sdpa): retire the bottom_right_with_swa notch — every row serves BR + SWA #623

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
2 changes: 1 addition & 1 deletion python/cudnn/frost/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -506,7 +506,7 @@ expressible, with one discipline separating them:
- The box is pure data: per-axis fields on `Capabilities`. Covers most of the
surface; adding an engine is writing a row, not logic.
- A notch is a rule in `mismatch()` gated by a conjunction flag on the row
(e.g. `bottom_right_with_swa: bool`). The matcher encodes the SHAPE of the
(e.g. `bottom_right_padded_seq_q: bool`). The matcher encodes the SHAPE of the
interaction once; each engine's row supplies the VERDICT. When a future
kernel supports the conjunction, flip its flag -- never edit the matcher.
This is what keeps interaction checks from regressing into a per-engine
Expand Down
11 changes: 0 additions & 11 deletions python/cudnn/sdpa/fwd/engines.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -140,7 +140,6 @@ class Capabilities:

causal: bool = False
bottom_right: bool = False
bottom_right_with_swa: bool = False # kernel gap: BR diagonal excludes SWA
# Kernel gap (pre-existing): for a DENSE padded graph the BR diagonal is
# computed as seq_len_kv[b] - GLOBAL S_q, but cuDNN semantics for a dense
# padded graph carrying per-batch seq_len_q anchor it at
Expand Down Expand Up @@ -357,8 +356,6 @@ def mismatch(capabilities: Capabilities, facts: "ga.SdpaGraphFacts", knobs: Opti
return "bottom-right alignment requires a causal upper bound (plain or right-widened)"
if not capabilities.bottom_right:
return "graph uses bottom-right causal, which this kernel does not support"
if facts.window_left is not None and not capabilities.bottom_right_with_swa:
return "bottom-right causal combined with a sliding window is not supported"
if facts.padded and not facts.thd and facts.seq_q_t is not None and not capabilities.bottom_right_padded_seq_q:
return (
"bottom-right causal with a dense padding mask carrying per-batch seq_len_q is not "
Expand Down Expand Up @@ -410,7 +407,6 @@ def _sm100_spec(d: int, d_v: Optional[int] = None) -> EngineSpec:
dtypes=frozenset({cudnn.data_type.HALF, cudnn.data_type.BFLOAT16}),
causal=True,
bottom_right=True,
bottom_right_with_swa=True,
right_band_widening=True,
swa=True,
padded=True,
Expand Down Expand Up @@ -460,9 +456,6 @@ def _sm100_mxfp8_spec(d: int) -> EngineSpec:
is_mxfp8=True,
causal=True,
bottom_right=True,
# BR+SWA stays off this row: the one mhas graph it admits trips the
# mxfp8 executor's SF-size mismatch (pre-existing SF-layout issue,
# independent of the mask) — flip once that plumbing is fixed.
right_band_widening=True,
swa=True,
padded=True,
Expand Down Expand Up @@ -501,7 +494,6 @@ def _sm100_fp8_spec(d: int) -> EngineSpec:
is_fp8=True,
causal=True,
bottom_right=True,
bottom_right_with_swa=True,
right_band_widening=True,
swa=True,
padded=True,
Expand Down Expand Up @@ -556,7 +548,6 @@ def _sm80_spec() -> EngineSpec:
right_band_widening=True,
causal=True,
bottom_right=True,
bottom_right_with_swa=True,
bottom_right_padded_seq_q=True,
swa=True,
padded=True,
Expand Down Expand Up @@ -652,7 +643,6 @@ def _sm120_spec() -> EngineSpec:
dtypes=frozenset({cudnn.data_type.HALF, cudnn.data_type.BFLOAT16}),
causal=True,
bottom_right=True,
bottom_right_with_swa=True,
bottom_right_padded_seq_q=True,
swa=True,
right_band_widening=True,
Expand Down Expand Up @@ -987,7 +977,6 @@ def _sm120_fp8_spec() -> EngineSpec:
is_fp8=True,
causal=True,
bottom_right=True,
bottom_right_with_swa=True,
bottom_right_padded_seq_q=True,
swa=True,
padded=True,
Expand Down