frost(sdpa): drop the redundant zero-fill of the never-read THD sinks dummy (SM100)

[vedaanta]

Problem

A per-execute fill kernel on the SM100 THD execute hot path zeroes the dummy sinks buffer — scratch whose contents provably do not matter: the sinks slot is always part of the kernel ABI, but CFG.HAS_SINK is a compile-time fold, so when the graph declares no sink the kernel never reads the buffer (and execute() enforces has_sink <=> sinks is not None, so the dummy only exists in the never-read case).

The fill dates to the original FROST landing (#476) as belt-and-braces; AGENTS.md Rule 1 bans adapter-side fills on the execute hot path.

Fix

Bind the carved/allocated dummy without the zero_() / torch.zeros fill (one hunk in _execute_thd). Split out of #543 so it can land independently; #543 keeps the matching O-descriptor no-fill and touches the same else-branch for stream-binding — whichever lands second has a trivial one-hunk rebase.

Verification

What	Where	Result
THD fp16/graph suites (incl. the sink-less THD cases that bind the deliberately unfilled dummy, and the #522 cu_seq_len forms)	SM100 (B200-class)	180 passed

🤖 Generated with Claude Code

Summary by CodeRabbit

• Performance
  • Reduced unnecessary initialization overhead in SM100 THD execution when sink support is disabled.
• Bug Fixes
  • Ensured the unused no-sink buffer is safely provided without requiring explicit zero-filling.

[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: ecdc0bc9-b121-4638-87ef-bd3e4081a4cc

📥 Commits

Reviewing files that changed from the base of the PR and between a7b4ca2 and f050b57.

📒 Files selected for processing (1)

• python/cudnn/sdpa/fwd/api_dsl.py

---

📝 Walkthrough

Walkthrough

SM100 THD execution now obtains an uninitialized ABI-only dummy buffer when no sinks are provided. The buffer comes from workspace or torch.empty instead of zero-filled storage.

Changes

SM100 THD dummy buffer

Layer / File(s)	Summary
No-sink dummy buffer allocation python/cudnn/sdpa/fwd/api_dsl.py	The no-sink path uses uninitialized workspace or torch.empty storage for the ABI-only dummy buffer.

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

Mergeability Score: ⚪ Minimal · up to f050b

This change removes an unnecessary zero-fill from the SM100 THD execute path while retaining the required dummy buffer binding; no actionable merge-blocking risk remains after normal checks and review.

Suggested reviewers: aneureka

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The description explains the problem, fix, and verification, but omits the required checklist, affected area, related issues, and API or compatibility sections.	Add the template sections and complete the checklist; state the affected area, related issues, API or compatibility impact, and exact test commands and results.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies the SM100 THD change and the removal of the redundant zero-fill.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
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.

✨ 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

[cudnn-ci-bot]

🚀 Running mirror pipeline

Branch: cudnn-gh/pr-573-f050b57
Pipeline: 62478929
Targets: frost