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

[vedaanta]

Before submitting

• I agree to license this contribution under the terms of LICENSE.txt.
• I ran pre-commit run and committed any formatting changes.
• I added GitHub labels: one cat-*, one or more mod-*, and one orig-*.

Affected area

FROST engine capabilities (python/cudnn/sdpa/fwd/engines.py) + the frost README's notch example.

Summary

Retire the bottom_right_with_swa capability notch — every row now serves the conjunction.

The notch guarded exactly one kernel gap: the mxfp8 row kept it off with a note to flip it "once that plumbing is fixed" — the plumbing being the _reshape_sf SF-size mismatch (B read from the bound tensor's shape, misread for flat F8_128x4 bindings), fixed in #606's mxfp8 commit. With that gone, all six bottom_right rows (SM100 f16/fp8/mxfp8, SM80, SM120 f16/fp8) serve BR + SWA and the flag differentiates nothing, so instead of flipping the last row on, this removes the field, its mismatch() rule, and the five constant-True spec lines.

The frost README's "notch" design example now points at bottom_right_padded_seq_q — a live notch (on for SM80/SM120, off for the three SM100 rows, whose kernels anchor the BR diagonal at the global S_q). The two probe tests pinning BR+SWA acceptance (test_probe_accepts_bottom_right_with_swa, test_sm120_probe_accepts_bottom_right_with_swa) are unchanged — they assert behavior, not the flag.

Validation

B200, cuDNN 9.26 nightly: graph-analyzer probe suite + full test_sdpa_mxfp8_fwd_L0 + test_sdpa_mxfp8_bwd_L0 sweeps — the mxfp8 engine serves the newly admitted BR+SWA config with numerics green against the fp32 reference (routing count 2, up from 1).

🤖 Generated with Claude Code

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: d6f801c9-3b56-4c1c-bfc4-8ddfec253c5b

📥 Commits

Reviewing files that changed from the base of the PR and between 491805e and 6c396f5.

📒 Files selected for processing (1)

• python/cudnn/sdpa/fwd/engines.py

Included review availability: Your plan includes up to 12 reviews per rolling hour; 7 remain after this review.

---

📝 Walkthrough

Walkthrough

The SM100 MXFP8 engine now enables bottom-right causal attention together with sliding-window attention. The previous eligibility restriction and its SF-size mismatch explanation were removed.

Changes

SM100 MXFP8 attention support

Layer / File(s)	Summary
Enable combined attention mode python/cudnn/sdpa/fwd/engines.py	The SM100 MXFP8 capability row now supports bottom-right causal attention with sliding-window attention. The previous exclusion was removed.

Estimated code review effort: 1 (Trivial) | ~5 minutes

Merge Risk: ⚪ Minimal · up to 6c396

This localized change enables bottom-right and sliding-window support for the SM100 MXFP8 path, with the supplied validation reporting 259 passing tests and no failures; no actionable merge-blocking risk remains beyond normal checks and review.

Suggested reviewers: aneureka

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly and concisely describes retiring the bottom-right-with-SWA capability notch.
Description check	✅ Passed	The description explains the change, rationale, affected area, and validation results, but omits several template headings and exact test commands.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}

[vedaanta]

@cudnn-ci-bot run frost

[vedaanta]

@cudnn-ci-bot run frost

[vedaanta]

@cudnn-ci-bot run frost

[vedaanta]

@cudnn-ci-bot run frost

[vedaanta]

@cudnn-ci-bot run frost