frost(sdpa): port #585 LPT schedulers to the SM107 fp8 sibling; defaults must obey declared knob domains

[vedaanta]

Fixes #653.

Two changes that are one rule: a scheduler the served kernel cannot decode must be unreachable — whether requested or defaulted.

1. The #585 port (the bug fix)

#585 landed the LPT/LPT_L2 tile schedulers in prefill_d128_fp8_sm100.py, but the hunk-symmetric port to the SM107 sibling was never done. The adapter's auto causal sched policy picks SCHED_LPT_L2, the sibling's decode guard refuses, and every causal/masked per-tensor FP8 graph on cc10.7 failed plan build (29 of 33 suite cases). Ported position-for-position with the SM100 file: qh_per_kh/seqlen_kv threaded through the five warp groups and appended at all ten scheduler decode call sites.

2. Domain honesty end to end

• Both fp8 kernel files declare SUPPORTED_SCHED_POLICIES — decode truth at the source.
• The adapter gains _sm100_sched_domain(); check_support now admits explicit NATURAL/LPT/LPT_L2 requests (was NATURAL-only — under-declared while the lowering auto-picked LPT), and the causal defaulting heuristic clamps its pick into the served route's domain, falling back to NATURAL. A default may never be a value the route could not honor as an explicit request — the invariant that turns this bug class into a clean fallback.
• The three SM100-family engine rows declare the true request domain {NATURAL, LPT, LPT_L2} (previously {NATURAL} — dishonest in both directions). SM120 rows untouched.
• frost/README knob-channel section documents the rule.

3. Sibling-lockstep fences (test_sdpa_fp8_sibling_parity.py)

Third silent-drift incident for this kernel pair (#574 signature skew, #602 amax ABI, #585/#653). The fences: source-scan tests (no GPU/DSL — run in any CI) assert every decode call site carries the LPT args and the sched-domain constants match; module tests assert compile() signatures in lockstep, kernel domains == adapter table, and CFG divergence is exactly the intentional Rubin set {TILE_K_HW_BMM1, TILE_K_HW_BMM2, STAGES_KV}. Negative-tested: removing one call site's args turns the fence red.

Testing

• SM100 (cc10.0): 40 passed — 5 parity + 2 routing + 33 fp8 e2e (unchanged behavior; LPT paths exercised as before).
• Rubin (cc10.7, w2u1g board): 33/33 fp8 e2e — the suite that fails 29/33 at develop tip — plus 7/7 parity+routing. Causal/BR/SWA per-tensor FP8 now runs LPT-scheduled on sm107 silicon.

🤖 Generated with Claude Code

Summary by CodeRabbit

• New Features

  • Expanded scheduler support for SM100 and SM107 FP8 attention routes, including natural, LPT, and L2-grouped LPT policies.
  • Improved causal scheduling fallback when a selected policy is unavailable.
  • Added SM107 FP8 configuration support for dense processing and grouped-query attention.

• Bug Fixes

  • Prevented invalid route-specific defaults from causing undecodable plans.

• Tests

  • Added parity checks to keep sibling FP8 routes, scheduling options, and configurations synchronized.

[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: a7aaa40a-8d64-404e-a387-69de0ee27203

📥 Commits

Reviewing files that changed from the base of the PR and between f2638ba and f418fd8.

📒 Files selected for processing (6)

• python/cudnn/frost/README.md
• python/cudnn/sdpa/fwd/api_dsl.py
• python/cudnn/sdpa/fwd/engines.py
• python/cudnn/sdpa/fwd/kernels/prefill_d128_fp8_sm100.py
• python/cudnn/sdpa/fwd/kernels/prefill_d128_fp8_sm107.py
• test/python/sdpa/frost/test_sdpa_fp8_sibling_parity.py

Included review availability: Your plan provides up to 12 included reviews per hour; 11 remain after this review.

---

📝 Walkthrough

Walkthrough

SM100 and SM107 FP8 SDPA routes now support matching natural, LPT, and LPT_L2 scheduler domains. The SM107 kernel receives required decode parameters across its execution paths. GPU-independent tests enforce sibling parity and routing consistency.

Changes

FP8 scheduler parity

Layer / File(s)	Summary
Scheduler domains and engine capabilities python/cudnn/sdpa/fwd/api_dsl.py, python/cudnn/sdpa/fwd/engines.py	SM100 route validation and engine specifications now support natural, LPT, and LPT_L2 policies. Causal compilation falls back to natural scheduling when needed.
Sibling kernel declarations and decode propagation python/cudnn/sdpa/fwd/kernels/prefill_d128_fp8_sm100.py, python/cudnn/sdpa/fwd/kernels/prefill_d128_fp8_sm107.py	Both kernels declare matching scheduler domains. The SM107 kernel propagates qh_per_kh and seqlen_kv through tile decoding and warp-group dispatch paths.
Parity contract and validation python/cudnn/frost/README.md, test/python/sdpa/frost/test_sdpa_fp8_sibling_parity.py	The documented contract covers route-specific defaults and synchronized domains. Tests check decode arguments, scheduler domains, compile signatures, adapter routing, and approved configuration differences.

Estimated code review effort: 4 (Complex) | ~45 minutes

Merge Risk: ⚪ Minimal · up to f418f

The PR ports the scheduler behavior, aligns declared domains and defaults, and adds parity coverage; no actionable merge-blocking risk remains beyond normal checks and review.

Possibly related PRs

• NVIDIA/cudnn-frontend#648: Modifies the same FP8 SDPA API, engine declarations, and sibling kernels for THD/varlen support.
• NVIDIA/cudnn-frontend#651: Modifies the same FP8 SDPA areas for FP16 softmax precision.

Suggested reviewers: aneureka

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 75.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies the SM107 FP8 scheduler port and scheduler-domain default fix.
Description check	✅ Passed	The description explains the changes, rationale, related issue, API impact implicitly, and detailed test results, but omits template headings and checklist items.
Linked Issues check	✅ Passed	The PR ports LPT and LPT_L2 scheduling to SM107, threads required arguments, and reports successful cc10.7 causal, BR, and SWA validation for issue #653.
Out of Scope Changes check	✅ Passed	The adapter, engine, documentation, and sibling-parity test changes directly support scheduler-domain consistency and the SM107 fix.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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