[AnaghaRaoAMD]

Summary

Closes the gfx950 forward-attention top-left causal gap in rocKE: the unified kernels silently served top-left as bottom-right (and were unconditionally causal), their dispatch candidates over-claimed a mask they did not implement, and dense mishandled cross-length / ragged shapes. The mask is now the cuDNN/hipDNN band (left = sliding_window, right = right_bound, alignment = diagonal_alignment) computed exactly or rejected with a reason. Part of epic AICK-2500.

JIRA ID : AICK-2505

Risk Assessment

High risk. On gfx950 the default request mask (NO_MASK) now means full attention in the unified family instead of being served by a causal kernel, and AttentionRequest gains fields, a band capability feature and new rejections; the kernel changes are large. Default-off IR is pinned byte-identical by the existing platform representative golden (re-run after merging develop), and the new variants' lowered IR is pinned by 26 added cases in that golden. gfx950 hardware parity is green, but no gfx942 / gfx1250 hardware run has been done for the dispatch changes that apply to those arches, and the new bodies are not yet mirrored in the C++ engine (see Known gaps).

ASIC Coverage

Kernel changes are gfx950-only, so gfx950 hardware must pass (done locally, see below). The dispatch-layer changes are arch-wide: SLIDING_WINDOW ordinal handling, the new request fields and band feature, and gating of the gfx942 dense-pipe and gfx942/gfx1250 tiled paths. gfx942 and gfx1250 keep their legacy NO_MASK (causal) body by design, and non-causal bands are rejected there. That behaviour is covered by CPU tests, but a gfx942 GPU-lane run is required before merge to confirm no routing regression.

Testing Summary

• gfx950 hardware parity (MI355X): unified 2D / 3D / scalar kernels against an explicit PyTorch-style reference for top-left, full attention, non-causal and two-sided windows, lookahead, decode, sinks, softcap, D64/D128/D256, GQA, the tiled-lever matrix, and exactly-zero empty rows; dense band matrix through both the spec and the dispatch path, including persistent and ragged cross-attention.
• Default-off IR byte-identity at llvm20, llvm22 and llvm23 via the existing platform harness golden (rocke_representative_ir_sha256.json: 32 attention_dense/gfx950 and 10 attention/gfx950 cases). Scalar and tiled window / QQ-bias / D256 variants that golden does not cover were compared once against the base commit (31 cases x 3 flavors, 0 differences); no separate library golden is added.
• Dispatch wiring tests: distinct kernel names / cache keys for every unified candidate, band normalization, capability features, off-gfx950 gates.
• Empty-row behaviour cross-checked against PyTorch 2.10 SDPA (fully masked rows return 0).

Testing Checklist

• rocKE library CPU lane - python library/tests/run_all.py - Status: Passed (2314 passed; 3 failures, all pre-existing and identical on clean develop: test_tiled_2d_shipped_geometries_compile_and_fit_budget and the gfx942 dense llvm23 golden)
• rocKE library GPU lane - python library/tests/run_all.py --gpu-only - ASICs: gfx950 - Status: Passed (536 passed, on the branch merged with develop)
• rocKE installed golden gate - python platform/tests/rocke_installed_golden_test.py - Status: Passed (llvm20, llvm22, llvm23), re-run after merging develop (including the fp32 softmax-scale and dispatcher-restructure changes)
• Pre-commit hooks - git commit - Status: Passed
• Review fixes validated - GPU lane 544 passed (gfx950); CPU lane 2317 passed with the same 3 failures as clean develop; platform golden gate PASS at llvm20, llvm22, llvm23 with the 26 added cases
• rocKE library GPU lane on gfx942 - python library/tests/run_all.py --gpu-only - ASICs: gfx942 - Status: Not run
• PR CI - GitHub PR checks - Status: Pending

Technical Changes

• Review follow-up: top-left is no longer folded into bottom-right on the unified family when the request's seqlen_q == seqlen_k (batch maxima there; the kernels derive the diagonal per sequence); dense keeps that shortcut. Dense q-blocks entirely past the keys no longer read other batches' tokens (start tile clamped when rows can be empty; the dense band matrix now runs at batch=2). The new variants' IR is pinned in the platform golden, and a GPU case covers request -> dispatch -> launch on a mixed-length cohort.

• Merged develop (fix(rocke): Expose more knobs of SDPA kernels via updated dispatcher interface gfx942, gfx950  #12719 dispatcher restructure, fix(rocke): apply gfx950 dense softmax scale in fp32 #13122 fp32 softmax scale): the dense band moved into gfx950_dense.py (_dense_band, _is_ragged), the unified AttentionSpec fields into gfx950_unified.py, and right_bound is classified as a dense problem field. Under fix(rocke): apply gfx950 dense softmax scale in fp32 #13122's finite mask sentinel a row with no visible key is no longer NaN but averages V, so the dense kernels now zero such a row at the epilogue when its running max is still at the sentinel.

• Unified 2D tiled, 3D split-KV and scalar kernels: new causal_top_left and right_bound problem/spec fields. The window anchor stays d - k < W and the upper compare becomes k <= d + R (skipped when unbounded); tile bounds, the sliding-window tile end, the peel split and the mask_limit threshold all follow R. R = 0 lowers byte-identically.

• Dense kernel (default and persistent grids): right_bound, non-causal windows, lookahead, ragged cross-attention, and a zero-output guard for rows with no visible key (previously stale data from exp2(-inf - -inf)), emitted only when such rows can exist.

• Dispatch: AttentionRequest.right_bound / diagonal_alignment, a band capability feature, SLIDING_WINDOW read as the band trio (rejected when it names no bound), distinct kernel names (_tl, _rbu / _rb<R>, _sw<W>) and cache keys; unsupported combinations (e.g. ALiBi/QQ-bias with a right bound, dense bottom-right window at sq != sk) are rejected with a reason.

• Parity harness: ref_paged_attn band reference and --set topleft|nomask|band; gfx950 benchmarks that relied on the default mask now pin bottom-right causal.

• Known gaps (follow-ups):

  • C++ mirror (tracked follow-up): the causal_top_left / right_bound bodies of the tiled 2D / 3D and unified (scalar) kernels are not yet mirrored in platform/cpp/instances/, and there are no library/tests/parity/*_emit pairs for them (e.g. attention_tiled_2d_ctx_prologue.cpp still hard-codes context_len = seq_len - cur_batch_q_len). The default mask is unchanged, so Python/C++ byte-identity still holds; until the mirror lands the engines diverge on the new variants, whose Python IR is pinned only by the platform golden cases. Dense has no C++ mirror.
  • On arches with no top-left kernel (gfx942 / gfx1250) the legacy equal-length shortcut remains: top-left is served by the bottom-right body, correct only when every sequence has q_len == kv_len. gfx942 / gfx1250 tiled kernels remain causal-only; dense window/lookahead at bottom-right alignment with sq != sk; pre-existing 2D tiled ALiBi + sliding-window NaN at bottom-right.

🤖 Generated with Claude Code

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🔎 pre-commit	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—
🤖 therock-pr-bot	✅ Pass	—

🎉 All checks passed! This PR is ready for review.

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[therock-pr-bot]

🎉 All checks passed! This PR is ready for review.

[dlangbe]

the new mask bodies have no spec→IR coverage. KERNEL_AUTHORING.md "Compliance & hard invariants" (1) requires library attention emission changes to be mirrored in the C++ engine (platform/cpp/instances/ carries the unified and tiled 2D/3D bodies), with .py/.c emit pairs per config. this PR touches nothing under platform/ or library/tests/parity/:

• grep -rn "causal_top_left\|right_bound" platform/ library/tests/parity/ returns nothing.
• attention_tiled_2d_ctx_prologue.cpp:334 still hard-codes context_len = seq_len - cur_batch_q_len.

the platform golden only holds default-mask cases, and deeab72 dropped the library golden. so the _tl / _rbu / _rb<R> / non-causal-window IR on 2D, 3D, scalar and dense is pinned by nothing. the next refactor of any mask site can move it, and only a GPU lane would notice. separately, the risk assessment still says default-off IR is "pinned byte-identical by a new golden", which no longer matches the testing section.

either add the C++ mirror plus emit-pair configs (tl, rbu, rb, left-only window) to gfx950_attention_tiled_{2d,3d}_emit and attention_unified_emit, or at minimum add those variants as representative-IR golden cases and list the C++ mirror as an explicit follow-up in the PR body so the divergence between the two engines is tracked.

[therock-pr-bot]

Pre-commit check failed

⛔ pre-commit failed

Please run locally:

• python -m pip install pre-commit
• pre-commit install
• pre-commit run --all-files --show-diff-on-failure

This repo uses .pre-commit-config.yaml.

[cgmillette]

PR #13125 — top-left causal and attention band coverage on gfx950 (AnaghaRaoAMD, AICK-2505) — 🟢 ACCEPT: thorough; gfx942 lane first

Prior item	Status
dlangbe: new mask IR unpinned; C++ mirror required	🟡 PARTIAL — 26 golden cases match ×3 flavors (ran); C++ mirror deferred, no ticket
dlangbe: unified equal-length top-left fold	🟢 FIXED — fold only for dense; gfx950 TL → _tl (ran)
dlangbe: dense empty q-blocks read next batch	🟢 FIXED — start tile clamped + rows zeroed; band matrix at batch=2
dlangbe: stale NO_MASK test comment	🟢 FIXED — scoped to gfx942/gfx1250
yraparti: one normalized band	🟢 FIXED — normalize_band → AttentionBand
yraparti: band index truth tables	🟢 FIXED — unified + dense truth tables (GPU)
yraparti: NO_MASK arch-dependent / silently causal	🟡 PARTIAL — gfx950 exact; gfx942/gfx1250 legacy causal (documented; #13211)
yraparti: RAG_KBOUND missing in window loops	⚪ Unreachable — spec rejects ragged + window (ran); now commented
yraparti: tuning band validity / cache keys	🟡 PARTIAL — identity/keys + gfx942 decline tested; window/tile validity untested
yraparti: d256 names, sw suffix, loose inputs (×8)	🟢 FIXED — suffixes present; bad alignment/window/mixed vocab rejected (ran)
Bots / CI	🟡 Pre-commit green; TheRock green at a7b95c3a, head pending; gfx942 GPU lane unrun
Our checks	🟢 Merge clean; default-off IR identical; no PR regressions (ran)

| # | Sev | file:line | Finding |
| yraparti: d256 names, sw suffix, loose inputs (×8) | 🟢 FIXED — suffixes present; bad alignment/window/mixed vocab rejected (ran) |
| Bots / CI | 🟡 Pre-commit green; TheRock green at a7b95c3a, head pending; gfx942 GPU lane unrun |
| Our checks | 🟢 Merge clean; default-off IR identical; no PR regressions (ran) |

#	Sev	file:line	Finding
N1	🟡	dispatch/attention/common.py:720-729	SLIDING_WINDOW+W, default right: gfx950 now non-causal top-left, gfx942/1250 reject — not in Risk (ran)

Utility: 🟢 Positive — exact cuDNN band semantics on gfx950; rejects instead of mis-masking.

---

Legend: 🔴 blocking · 🟠 should-fix · 🟡 minor nit · 🟢 info/good

[assistant-librarian]

TheRock Submodule Bump Activity

Newest first

• 2026-10-19 — Included in TheRock via ROCm/TheRock#9348.