[ThruptiRajLakshmanaGowda]

Summary

Routes hipDNN SDPA forward requests to the unified attention kernels (AICK-2613) and the dense attention kernels (AICK-2614): for each family an admission gate that never admits a request and then drops an operand, and a binding that launches admitted requests; plus a fix for one gfx942 kernel bug the unified binding's GPU test found (the other two it found are fixed on develop by #14929, which this PR now reuses). Supersedes #14624, which is folded in here.

The translation, band comparison, value checks and operand views both families use live in dispatch/sdpa/bind_shared.py.

Admission (dispatch/sdpa/bind_unified.py)

• to_attention_request (bind_shared.py) translates an SdpaRequest (from feat(rocke): add hipDNN SDPA request contract and reference oracle #13596) into the AttentionRequest the unified candidates already consume: batch, heads, sequence lengths, head dims, causal/window bounds (sliding_window = left_bound: since fix(hipdnn): follow cuDNN semantics for the SDPA left diagonal-band bound #14483 hipDNN's left_bound = L keeps the L keys ending at the diagonal, the count sliding_window uses), sinks and LSE.
• Every feature in SDPA_FEATURES is classified once. SDPA_UNIFIED_FEATURES is what the unified launch path carries. Every other feature is declined with a reason that starts with its name: kv_heads_differ, dtype_mixed, compute_dtype, scale_tensor, bias, bias_broadcast, alibi, sinks (no reference oracle yet), dropout, paged, block_mask, fp8, fp8_s_o, max_sum_exp. Declines for kernel features not implemented yet (scale_tensor, bias, bias_broadcast, alibi, sinks, dropout, paged, block_mask, fp8, fp8_s_o) are tagged NOT_YET_IMPLEMENTED:; the ABI limits are not.
• Value checks: mma_core_mode must equal Q's dtype, implementation="composite" is declined, padding_mask with no seq_len tensor is declined, a pinned algorithm, spec_id, tuning_id or tuning_knobs is declined (dispatch_pin:), because the unified launch takes no pin, and a non-finite or non-positive scale is declined (scale:): the gfx942 4-warp bodies mask with -1e30 before scaling, so a scale <= 0 would bring masked keys back.
• Band check: the request is declined when the band the unified problem is built with (attention_band, what attention_problem hands every unified candidate — both exported from dispatch.attention for callers outside the package, rather than aliasing a private name) keeps different keys from the requested one. On gfx942 and gfx1250 full attention comes back bottom-right causal, so it is declined except at decode; this relaxes once those arches gain the full band (feat(rocke): attention band coverage on gfx942 #13211). Outside the band arches (gfx1151, gfx1201, gfx90a, gfx11/gfx12 generics) the problem is always bottom-right causal, so top-left bands with s_q != s_kv are declined there. Bands are compared as per-row kept-key intervals, O(s_q).
• Admission counts only the unified route candidates (unified_2d, unified_3d, d256_decode, d256_gfx950, dense_pipe), named from dispatch.attention.UNIFIED_ALGORITHMS. Both the tiled 2D and split-KV (3D) paths write LSE on gfx942/gfx950 (_reject_lse in kernels/common/attention_unified.py gates each on its own arch/feature rule); admission trusts each candidate's own gate rather than a separate, narrower algorithm allowlist, so an LSE request reaches whichever path the registry actually admits it to.
• The module adds no shape, dtype or head-size rule of its own; the candidates' admits() decide.

Family entry points (dispatch/sdpa/__init__.py)

registered_sdpa_combos / sdpa_sweep_space / dispatch_sdpa_all keep #13596's SDPA_REGISTRY delegation and add the dispatch.attention combos of the family an SDPA gate admits the request to: the unified tuning variants when the unified gate admits it, and the dense candidates the dense gate names. A request both gates decline (bias, dropout, block mask, paged, ...) gets nothing. These results describe attention kernels for ranking and triage: a result's bind_torch() takes attention-format operands and the attention default scale, so an SDPA request is launched through sdpa_unified_bind / sdpa_dense_bind, which dispatch.sdpa exports.

Binding (sdpa_unified_bind(req, tensors) -> TorchBinding)

Launches an admitted request through run_unified_attention_torch with no tuning spec and backend="auto": both the tiled 2D and split-KV (3D) paths can write LSE on gfx942/gfx950, so the launch defers to the kernel's own path selection rather than forcing 3D whenever LSE is requested. The unified kernels take packed [total_q, H, D] Q/O and a paged K/V cache, so the binding adapts the hipDNN operands:

• Lengths per batch come from seq_len_q / seq_len_kv, else the ragged offsets (element start = offset * ragged_offset_multiplier, as hipDNN reads them), else the full extent.
• Q is gathered into a packed buffer, or used in place when it is already packed BSHD at full length or contiguous ragged and its base is 16-byte aligned (the kernels declare align=16 Q/O pointers). O is handled the same way and scattered back otherwise.
• Dtypes: every tensor (operands, stats, seq_len_*, ragged offsets) must have the dtype its request descriptor names (bind_shared.check_dtype); nothing is truncated or cast on the way in.
• K/V are copied into a paged cache behind an identity block table exactly ceil(S_kv / block) wide, which the scalar fallback requires.
• LSE is scattered into the request's (B, H_q, S_q, 1) stats. Rows past a batch's query length get O = 0 and LSE = -inf, as hipDNN's reference writes them.
• Scale is the request's value; hipDNN's default is 1.0, not the kernel's 1/sqrt(d).

Lengths are snapshotted and buffers allocated once at bind time, so every launch hands the kernel the same tensors (the 2D path replays graphs keyed on them). launch(*, stream=None) runs pack, kernel and unpack on Torch's current stream; it accepts that stream unset or named by handle and refuses any other (bind_shared.launch_stream), because the copies and the workspace belong to it. The paged ABI is checked with validate_tuning_attention_tensors, because the route path checks no shapes itself.

Dense (dispatch/sdpa/bind_dense.py, AICK-2614)

sdpa_dense_admits / sdpa_dense_candidates / sdpa_dense_bind route a request to the gfx942 / gfx950 dense candidates (grid, qb-minor grid, persistent, persistent wide-DMA) with the same rules as the unified side:

• Every feature is classified once. SDPA_DENSE_FEATURES is what the dense launch carries: GQA, every layout (operands that are not contiguous [B, S, H, D] are copied through buffers allocated once at bind time), top-left and bottom-right causal, windows, lookahead bands and LSE. Declined with a reason: kv_heads_differ, d_v_differs (one head size), dtype_mixed, compute_dtype, scale_tensor, bias, bias_broadcast, alibi, sinks (no reference oracle yet), padded and ragged (the dense candidates take one length per batch), dropout, paged, block_mask, fp8, fp8_s_o, max_sum_exp; and the same value checks as unified.
• The dense candidates are opt-in, so each is probed with its own algorithm / spec_id pinned (opt_in_probe); their own admits() decide every shape, dtype, head-size and band rule (tile multiples on gfx942, bottom-right only on the gfx950 grid body, windows anchored top-left).
• A candidate counts only when the band in the spec it builds (causal, sliding_window, right_bound, causal_bottom_right) keeps the keys the request asks for. The gfx950 kernel's scale range [2^-64, 16] is checked here, read from the kernel module, because an AttentionRequest does not carry the scale.
• sdpa_dense_bind launches the first serving candidate in registry order (the grid body on gfx950) at its default spec through bind_dense_attention_torch, with the request's scale, under the same dtype, alignment and stream rules as the unified binding. LSE is written in place into a contiguous FP32 (B, H_q, S_q, 1) stats tensor, else through a buffer. Rows with no visible key come back O = 0, LSE = -inf from the kernel.
• gfx942 full attention with s_q != s_kv stays declined by the attention layer's own NO_MASK gap (feat(rocke): attention band coverage on gfx942 #13211), not by this module.

gfx942 4-warp kernel fix

build_gfx942_4warp_gqa (the D128 sliding-window cohort and the D256 lean body) assumed callers over-allocate Q and the block table, which hipDNN's exactly sized operands don't. On MI325X the unified binding's GPU test found three bugs there. Two of them, the unbounded Q load of a 128-row q-block and the last KV tile reading block-table entries past ceil(klen / BS), are fixed on develop by #14929; this PR takes #14929's clamps unchanged and adds no clamp of its own. It fixes the third:

Bug	Symptom	Fix
Rows that see no key (bottom-right, qlen > klen) weren't special-cased	O = mean of masked V, or NaN with GQA head-fold (x_mha_padded_br_win_none_gt, x_gqa4_ragged_br_win_none_gt)	write O = 0 in both bodies, as the default builder and gfx950 do

Against develop the kernel diff is that one select in each of the two bodies, so this intentionally changes the IR of exactly the two kernels: only the attention_d128_swa and attention_d256 shards of the representative IR golden are re-blessed (--write … --only attention_d128_swa,attention_d256, llvm20/22/23/24, re-recorded on top of #14929 and #15024), and nothing else drifts. MI325X timings of the cohort are unchanged within noise (relative to a 5-round alternating median, 200 attempts each; measured before the #14929 merge):

Shape	Δ
D128 bf16 GQA 32/8, window 4096, bs16, 2 x 8k	-0.03%
D128 fp16 GQA 32/8, window 4096, bs16, 2 x 8k	-0.39%
D256 bf16 GQA 16/2, causal, bs32, 2 x 4k	+0.18%

Tests

• tests/sdpa/test_unified_sdpa_wiring.py: every feature classified exactly once, one decline case per declined feature, value checks including dispatch pins, band checks on gfx942, gfx950 and four non-band arches, the interval comparison against a dense mask, a 128k band check under 64 MiB, layouts, padded, ragged, LSE path.
• tests/sdpa/test_unified_sdpa_case_table.py: for every cases.py case on gfx942 and gfx950, an admitted request reports only carried features and a unified candidate's own admits() accepts it; a declined one has a reason. gfx950 admits 405 cases, gfx942 283 (up from 261/202: the N8 fix below reaches the LSE cases a narrower algorithm allowlist used to drop).
• tests/sdpa/test_unified_sdpa_numeric.py (GPU, gfx942/gfx950): binds every admitted case, all tiers, launches it twice, and compares with sdpa_reference on the valid rows, fully masked rows included. --ignored by rocke_library_sdpa_pytest, enrolled in _ROCKE_LIBRARY_GPU_TEST_RELPATHS.
• tests/sdpa/test_unified_sdpa_binding.py (CPU, needs torch): replaces the kernel with sdpa_reference over exactly the packed and paged tensors the binding produces, across every layout and length mode; sdpa_unified_bind refuses bad seq_len_* values before launch, including a float32 or int64 seq_len_q, ragged offsets whose dtype is not the request's ragged_offset dtype, and FP16 stats against an FP32 descriptor; misaligned BSHD and ragged Q/O go through the workspace; launch runs on the current stream (unset, 0 or named), refuses another stream and an unknown keyword.
• tests/sdpa/test_dense_sdpa_wiring.py, test_dense_sdpa_case_table.py, test_dense_sdpa_binding.py (CPU, needs torch) and test_dense_sdpa_numeric.py (GPU): the same four layers for the dense binding. The case-table test also checks, independently of the gate, that each named candidate's spec keeps exactly the requested keys. The GPU test is --ignored by rocke_library_sdpa_pytest and enrolled in the GPU lane.
• test_coverage_source_of_truth.py: the unified and dense twins, every SDPA admit is a candidate admit.
• tests/sdpa/cases.py: 21 tile-multiple dense_tile_* cases (S_q a multiple of 256: the lengths gfx942's dense kernel takes), including one with fully masked rows and LSE.
• tests/sdpa/request_bridge.py, tests/sdpa/binding_harness.py: build the SdpaRequest and the binding's tensors for any case. The bridge lifts the case table's FlashAttention window W to hipDNN's left_bound = W + 1, so every case keeps the same keys; the oracles convert back with hipdnn_left_to_window.
• tests/sdpa/test_sdpa_family_entry_points.py: the entry points return exactly the gate-admitted families' combos on gfx942 and gfx950 (pinned on named candidates, e.g. gfx942 full attention reaches dense but not the unified tuning variants, gfx950 padded reaches unified but not dense — not just re-derived from the same _serves logic the entry points use), both the tiled 2D and split-KV unified paths with LSE, and nothing for bias, dropout, block-mask or paged requests.
• tests/sdpa/test_unified_sdpa_wiring.py::test_sliding_window_keeps_exactly_left_bound_keys: to_attention_request(...).sliding_window keeps the L keys the reference oracle keeps, at left_bound = 1, 4, 63, 64 — a CPU pin that fails if a + 1 returns to the translation.
• tests/sdpa/test_unified_sdpa_wiring.py (torch-free, so CI runs it): the scale decline (0, -1, NaN, ±inf) on gfx942 and gfx950; seq_len_q / seq_len_kv bind-time validation (negative, too large, wrong count, past the ragged lengths) and __all__ resolution for the dispatch/sdpa modules.
• tests/dispatch/attention/test_attention.py::TestAttentionPublicSurface (run by rocke_library_attention_pytest): every dispatch.attention.__all__ name resolves and star-imports; __all__ equals the listed public surface, so a new name fails until it is added on purpose; no underscored name or submodule is exported; every exported *_ALGORITHM(S) names a registered candidate.

Results

On this branch merged with develop at 10805993013 (including #13596, #14483, #14929 and #15024). Admission: 283 unified / 32 dense on gfx942, 405 unified / 192 dense on gfx950 — unified counts up from 202 (gfx942) / 261 (gfx950) after this round's fix to the unified LSE admission, which was wrongly narrowed to split-KV only (dense is unchanged); see "Admission" and the entry-points test above.

• gfx942 (MI325X) / gfx950 (MI355X) at cf2b94adaae, one process each: unified and dense numeric suites (every admitted case, all tiers), both binding tests, both wiring tests and the entry points: 676 / 958 passed, 0 failed, 0 skipped.

• Host at cf2b94adaae (CPU torch 2.15): tests/sdpa + tests/dispatch/attention 6509 passed, 257 skipped (the skips are the GPU suites). Earlier, after the perf(rocke): runtime heads, window and batch for gfx942/gfx950 tiled attention #15024 merge: tests/sdpa 5676 passed, 4 skipped (unified wiring 122, dense wiring 78, unified case table 1433, dense case table 1434, entry points 19, unified binding 74); tests/dispatch/attention 805 passed, 253 skipped; coverage source of truth 11 passed; test_sdpa_seqlen.py + test_strided_kv_decode.py + test_attention_tiled_launch_signature.py 168 passed, 12 skipped; platform test_sdpa_reference.py 46 passed.

• IR golden parity OK at llvm20/22/23/24 (rocke_ir_parity_harness.py --check).

• pre-commit clean.

Not in this PR

• Paged KV, dropout, fp8, block masks, tensor scales, additive bias, ALiBi, sinks; per-batch lengths on the dense path.
• Choosing between the unified and dense families for one request, and tuning which dense variant runs (the binding takes the first in registry order).
• The unified binding copies all of K and V into its paged cache on every launch (an extra K/V read and write per call, most visible at decode). Viewing contiguous BSHD K/V in place when s_kv % block == 0 would avoid the copy, but it drops the zero-filled tail the copy gives rows past each sequence's length, so it needs a GPU numeric check first.

New public API

dispatch.attention gains, for callers outside the package (SDPA's gates, in particular) that must not depend on a private symbol:

• attention_band(req) -> AttentionBand / attention_problem(req) -> UnifiedAttentionProblem: the single home for each (no more _attention_band/_problem alongside a public wrapper).
• The route algorithm names UNIFIED_2D_ALGORITHM, UNIFIED_3D_ALGORITHM, D256_DECODE_ALGORITHM, D256_GFX950_ALGORITHM, GFX942_DENSE_PIPE_ALGORITHM, and the set UNIFIED_ALGORITHMS built from them.
• dense_scale_range(arch) -> tuple[float, float] | None: the dense kernel's scale range where its module declares one (gfx950: [2^-64, 16]), so the dense gate does not read a private _MIN_SCALE / _MAX_SCALE.
• DENSE_ALGORITHMS (alongside UNIFIED_ALGORITHMS): every algorithm a routed family's gate must account for so a new candidate cannot go unreachable by omission; bind_dense.py's _DENSE_ALGORITHMS and the attention route-coverage test both build from it now instead of re-listing the three dense algorithm constants.

dispatch.sdpa exports sdpa_unified_admits / sdpa_unified_bind and sdpa_dense_admits / sdpa_dense_bind, the launch path for an SDPA request.

rocke.runtime.torch_interop gains torch_dtype and device_tensor (empty_workspace and resolve_stream already existed) — the torch launch-integration glue both bind_unified.py and bind_dense.py use instead of importing torch directly in dispatch/sdpa/.

See KERNEL_AUTHORING.md Appendix §H ("Route a family over dispatch.attention") for the binding-layer checklist this package's gates follow.

JIRA ID : AICK-2613, AICK-2614

🤖 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]

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.

[therock-pr-bot]

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

[AnaghaRaoAMD]

The PR body says 11/11 passed, but tests/sdpa/test_unified_sdpa_wiring.py now defines ~30 tests. Please refresh the count from the current head.

[cgmillette]

Automated overlap notice (rocKE merge monitor)

Hi @ThruptiRajLakshmanaGowda, this PR and #14929 each add the same bounds fix to build_gfx942_4warp_gqa, clamping Q rows to the sequence's last token and K/V pages to its last page, in two different forms. Row N13 of our 4 November follow-up review of this PR flagged it, and your reply on N13 already takes the resolution below; this notice records it and adds rebase steps checked on the merge.

The two copies, here at b7923d9dd5 and in #14929 at 3f8ecb6c60, in library/kernels/gfx942/attention_tiled_2d.py (paths under dnn-providers/hip-kernel-provider/rocke/):

• K/V pages: here last_blk (:6353-6356), applied in phys_key (:6360); fix(rocke): bound Q and K/V reads in the gfx942 4-warp GQA D128 body #14929's k_last_page (:6355), applied with b.smin in its phys_key (:6376), plus a refusal of a non-power-of-two block_size (:6194-6196) for its shift.
• Q rows: here the q_row(tok) helper (:6373-6376), called at the Q load in every KV tile (:6554, :6560); fix(rocke): bound Q and K/V reads in the gfx942 4-warp GQA D128 body #14929 clamps the row once before the loop (:6356-6372) and its Q load reads it (:6559).
• The golden: both re-record the same four 4warp_gqa_fold entries (one per LLVM flavor) in platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json.

Both clamp to the same row and page wherever the kernel reads, so dropping yours loses nothing. Your no-key-row O = 0 (:6732-6736, and :6153-6157 in the lean body) is not in #14929 and stays.

Owner: #14929, the bounds fix itself. Neither PR puts the clamps in a shared home, so the PR that lands first owns them, and #14929 is approved and first in the merge plan (cgmillette's priority). Resolution: this PR drops its two clamps, keeps its no-key-row O = 0, and re-records attention_d128_swa when it rebases after #14929 lands.

Suggested order: #14929 first, then this PR: the merge plan (develop 49a01e717d) lands #14929 in its first group and does not include this PR yet. In that order #14929 does nothing for this overlap, and this PR makes the changes below.

Rebase, after #14929 lands (checked with git merge-tree of this PR's b7923d9dd5 onto develop 49a01e717d merged with #14929's 3f8ecb6c60: 2 files conflict, both from #14929, since this PR merges cleanly with develop alone; platform/CMakeLists.txt merges cleanly):

1. library/kernels/gfx942/attention_tiled_2d.py, 2 conflicts: take fix(rocke): bound Q and K/V reads in the gfx942 4-warp GQA D128 body #14929's side in both (its clamp block with its phys_key, then its one-line Q load). Then delete your q_row(tok) helper: the merge keeps it outside the conflicts, where it rebinds q_row, fix(rocke): bound Q and K/V reads in the gfx942 4-warp GQA D128 body #14929's clamped row, to the function, and the kernel build fails with AttributeError: 'function' object has no attribute 'type' (4 of fix(rocke): bound Q and K/V reads in the gfx942 4-warp GQA D128 body #14929's TestD128BodyBoundsReads tests fail that way). Your no-key-row hunks merge cleanly; keep them.
2. platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json: neither side's hashes fit the merged builder. Re-record from platform/ with PYTHONPATH=python:../library python tests/instances/rocke_ir_parity_harness.py --write tests/golden/rocke_representative_ir_sha256 --only attention_d128_swa. Each flavor must print 1 ok, 0 failures: --write records a build error as an expected failure instead of stopping. attention_d256.json needs nothing: your entries merge cleanly and match a re-record on the merge.

Resolved this way, the shard re-records at all four flavors, and test_d256_gfx942_fast_routing.py, test_gfx942_4wgqa_bounds_numeric.py and test_gfx942_gqa_head_fold_numeric.py pass on the merge (27 passed; the 28 cases that need a gfx942 GPU skipped).

If this PR lands first (not the plan's order), take #14929's text now in place of your clamps: its :6194-6196 after BS = spec.block_size, its :6346-6378 for your :6349-6363, and its Q load (:6559) for your :6549-6561; delete your q_row helper, keep the no-key-row O = 0, and re-record attention_d128_swa. Checked: your builder is then the merge result above, git merge-file of #14929's builder onto it is clean, and #14929's rebase keeps develop's attention_d128_swa hashes in the one conflict left.

If you think this PR should own it instead, reply here and on #14929, and the reviews will follow the new decision.

Automated overlap check: one owner per shared piece of code, coordinated on the PRs. Nothing else changed.

[cgmillette]

Automated review: Model - qwen3.6:35b-a3b-coding-mtp-q4_K_M, effort: max

PR #13115 — route SDPA to unified and dense attention kernels (ThruptiRajLakshmanaGowda, AICK-2613, AICK-2614) — 🟠 Follow-up: fixes hold; 14929 rebase pending

Prior item	Status
Earlier 🟢 rows: arai713, AnaghaRaoAMD, dlangbe, Copilot, CongMa13 ×2, yraparti (10-31); our N1–N8; 14483 left bound	🟢 FIXED, unchanged • every SDPA and attention test passes with torch on the merge (ran)
AnaghaRaoAMD (11-03) ×6: old oracle, decline table, literals, torch, private names, §H	🟢 FIXED 9c489ef4, 9db85cee, b7ed4698 • unchanged by b7923d9d (read) • 6 threads answered; AnaghaRaoAMD resolves
AnaghaRaoAMD (11-03) bind_unified.py:395: K/V copied every launch	🟡 DEFERRED · AnaghaRaoAMD decides • unchanged; in the body's Not in this PR (read) • author's reason holds for padded K/V only (read)
yraparti (11-04) MEDIUM ×2: torch import; seq_len negative tests	🟢 FIXED 9db85cee, b7ed4698, b7923d9d • seq_len refusals now run torch-free in CI: N12 (ran) • left: ragged-offset checks still untested (ran)
yraparti (11-04) MEDIUM §S.4: edited __all__ has no name test	🟢 FIXED b7923d9d • an outdated or duplicate name fails TestAttentionPublicSurface (ran) • passed in both gfx94X shards (read); thread open, yraparti resolves
Our 11-03 N9: two claims with no failing test	🟢 FIXED b7ed4698, b7923d9d • the old + 1 now fails the left_bound test 4/4 (ran) • both PR numbers dropped (read)
Our N10: 9c489ef4's new names have second homes	🟢 FIXED b7ed4698 • left: open 12778, 13128, 13700, 14440, 14449 add _problem uses; the second to land ports (ran)
Our N11: body omits the new public API; names private helpers	🟡 OPEN • empty_workspace, resolve_stream predate this PR, not "gains" (read) • Tests section predates this round's tests (read)
Our N12: seq_len check copies develop's rule; tests never in CI	🟢 FIXED b7923d9d • without validate_lengths, 6 no-torch wiring cases fail (ran) • all pass in both gfx94X shards (read); float lengths: N14
Our N13: same 4-warp bounds fix as open 14929	🟠 OPEN • 14929 now 330c35b8; still conflicts in the builder and attention_d128_swa (ran) • agreed: after 14929 lands, rebase, keep O = 0, re-record; author acts
Bots / CI	🟡 running, none failed • running: Windows gfx110X test shards; 41 ✓, 0 ✗ • gfx94X rocKE shards ✓, every new test passed; math-libs ✓ after the hipCCL revert (read) • merge: BLOCKED; no approval; 8/29 threads open
Our checks	🟢 GREEN • vs 39fbe7b9: clean, 8 behind; same provider tree as the checks' merge (ran) • build + ctest 46 ✓ / 1 skip; byte identity, parity, golden ×4; dg 0 changed (ran) • tests 4142 ✓, 25 ✗ on develop too (issue 14652's cause); installed 3856 ✓; torch 5609 + 1045 ✓ (ran) • gfx1151 GPU: 55 cases match (ran); no gfx942/gfx950 run yet (the body's open box) • compile 29 ✓, 0 spills; no ABI change; rules 1–3 clean (ran)
Code quality	🟡 Minor • N9 and N12 fixed; mutants fail CI-run tests (ran) • float seq_len truncated before the dtype check: N14 (ran) • 🟡 seq_len cases in three files; new bind_* __all__ unchecked; fu9's 🟡s unchanged (read)

#	Sev	file:line	Finding
N14	🟡	dispatch/sdpa/bind_unified.py:245	_host_lengths casts each length to int before validate_lengths sees it • impact: a float32 seq_len_q [1.5, 1.0] binds as [1, 1], not refused; the docstring's dtype check never runs (ran) • scope: only a caller whose tensor contradicts its int32 request; in the PR since 10-28, missed in our earlier rounds • fix: hand validate_lengths the values with their dtype (.numpy()), not host_ints; add a float case

Utility: 🟢 Positive — routes hipDNN SDPA requests to unified and dense attention kernels.

---

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

Follow-up review: first pass by a local qwen3.6 model, verified by the orchestrator (Claude Opus 5.5). Checks: delta since d19c06a6, merge with develop 39fbe7b9, golden ×4 and saved probes, each reply's claims by source mutation, SDPA tests with torch, binding launches on a GPU, overlap with 14929, CI at head, code quality. Flags whether a closer look is needed; not an approval or rejection.

[ThruptiRajLakshmanaGowda]

Follow-up on the 11-04 23:10Z review, pushed as 33013ea:

• N14: _host_lengths now hands validate_lengths the host array with its dtype, so a float seq_len_* tensor is refused instead of truncated. Added a float32 test.
• N11: the body no longer lists empty_workspace and resolve_stream as gains, and the Tests section covers this round's tests.
• Unchecked __all__ in the new bind_* modules: added a torch-free test that resolves every __all__ name in dispatch/sdpa/{bind_shared,bind_unified,bind_dense,hipdnn,common}.
• N13: still waiting on fix(rocke): bound Q and K/V reads in the gfx942 4-warp GQA D128 body #14929 to land (it's open). After that I'll rebase, drop the clamps and q_row, keep the no-key-row O = 0, and re-record attention_d128_swa.

tests/sdpa: 5558 passed, 11 skipped.

[ThruptiRajLakshmanaGowda]

#14929 landed, so I merged develop (head below) and resolved N13 as agreed:

• build_gfx942_4warp_gqa: took fix(rocke): bound Q and K/V reads in the gfx942 4-warp GQA D128 body #14929's clamps (k_last_page, the hoisted Q row) and deleted my q_row helper.
• Kept the no-key-row O = 0 in both the lean and 4-warp bodies.
• Re-recorded attention_d128_swa at llvm20/22/23/24 (each printed 1 ok, 0 failures). The d128_swa and d256 shards check clean; the only drift lines are other families, which a single-shard check reports as "new".
• CPU: tests/sdpa plus the gfx942 4wgqa-bounds, head-fold and d256-routing tests give 5585 passed, 39 skipped (the skips need a gfx942 GPU).

I haven't re-run the gfx942 / gfx950 GPU suites on the merged head yet.

[ThruptiRajLakshmanaGowda]

GPU re-run on the merged head 6daf7e6 (single process each: tests/sdpa plus the gfx942 4wgqa-bounds, head-fold and d256-routing tests):

• gfx942 (MI325X): 5962 passed, 3 skipped, 0 failures.
• gfx950: 6216 passed, 31 skipped, 0 failures.

[yraparti]

rocKE checklist re-review of the fixes since d19c06a6d342

ROCm/rocm-libraries#13115 · 6daf7e6b96b0 · model gemini-4-argon, fallback gpt-5.5

PASS · graded 238/238 · BLOCKER 0, HIGH 0, MEDIUM 0, LOW 0, NEEDS-INFO 25

The diff was truncated before grading. Items that needed a missing hunk are NEEDS-INFO.

Since the last review (d19c06a6d342)

Resolved 1 · still open 0 · new 0

• ✅ §S.4 was FAIL, now PASS — had asked: Add a CI-run test that iterates dispatch.attention.__all__ and asserts each listed name resolves from the module, e.g. via getattr(dispatch.attention, name) or star-import coverage.

Findings

• No FAIL rows.

Needs info

• §A.1 Unresolved since previous review: dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6017 and :6372 modify the gfx942 4-warp attention builders (_build_gfx942_4warp_gqa_lean and build_gfx942_4warp_gqa) without a C++ update in the PR diff, and known_gaps.md and cpp/instances/ are not in the diff to verify whether the 4-warp GQA body has a C++ twin or is Python-only.
• §C.9 In dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:5862-5865, 6017, 6067, 6159-6161, 6372-6394, 6576, and 6749-6751, the Q-row/key clamping and no_key zeroing fixes were applied across both the 4-warp lean and 4-warp GQA bodies. — no diff line supports this verdict
• §E.9 New post-review diff content includes changes outside the SDPA/rocKE scope, such as projects/composablekernel/include/ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp:1824 and projects/hipblaslt/clients/common/include/testing_matmul.hpp:1238. — no diff line supports this verdict
• §E.10 Not resolved since the previous review: the diff still does not include the requested merge-simulation record listing open PRs that touch the same helpers/tests/goldens, agreed land order, and affected-suite reruns on current develop plus those PRs.
• §E.14 New post-review changes include omitted deletion hunks, so the diff provided is insufficient to verify that removed upstream code, comments, or KNOWN LIMITATION notes were intentionally ported after the rebase/port.
• §F.6 Still unresolved since previous review: local pre-commit execution status or CI lint check output is not visible in the diff.
• §F.7 Still unresolved since previous review: GitHub review-thread state and ticket IDs for any deferred review comments are not included in the provided PR context.
• §G.3 In the new commits, attention_d128_swa.json is updated across all four LLVM flavors (dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8) and attention_d256.json is updated (dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8), but the tail of attention_d256.json is still truncated in the diff and local non-GPU gate sweep execution results cannot be verified from the diff alone.

Section notes

• §D: Re-reviewed items 1-11 against current head 6daf7e6. No new issues introduced by recent commits.
• §E: Re-review of Structure & hygiene against the current diff.
• §G: Re-reviewed items 1-17 of Section G against head 6daf7e6.
• §H: Trace of device-side masks/bounds: (1) Q row load in _build_gfx942_4warp_gqa_lean and build_gfx942_4warp_gqa visits a TOKBLK (128-row) compile-time Q block; extent comes from runtime cu_seqlens_q (cu_q_stop, qlen); indices are clamped via _clamp_to_last(..., q_last) (q_last = cu_q_stop - 1, attention_tiled_2d.py:6067,6373,6382,6388) and O stores are guarded by q_inseq < qlen (attention_tiled_2d.py:6162,6752) — agree. (2) KV page/key lookup in phys_key visits BN-key tiles (BN=32 or 64); extent comes from runtime seqused_k (klen); lean clamps key to k_last (attention_tiled_2d.py:6017) and 4-warp GQA clamps page to k_last_page = (klen - 1) >> BS_LOG2 (attention_tiled_2d.py:6372,6393), while trailing rows of the last page are zero-initialized in k_cache/v_cache (bind_unified.py:352-358) and rows with no visible keys (context_off + q_inseq < 0) have O zeroed (attention_tiled_2d.py:6159-6160,6749-6750), though the in-loop score…
• §J: The diff includes Markdown changes in dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md. Shapes | Formula / Op | Operands | Inner dims | Result vs Declared | |---|---|---|---| | None | — | — | No tensor equations in KERNEL_AUTHORING.md | Symbols | Symbol | Defined at | Domain | |---|---|---| | SDPA_<FAMILY>_FEATURES | KERNEL_AUTHORING.md:609 | Set of carried SDPA feature names per family | | UNIFIED_ALGORITHMS | KERNEL_AUTHORING.md:639 (dispatch/attention/__init__.py:85) | Frozenset of unified route algorithm strings | | DENSE_ALGORITHMS | KERNEL_AUTHORING.md:641 (dispatch/attention/candidate.py:51) | Frozenset of dense route algorithm strings | Novelty | Claim | Symbol / Path | Status | |---|---|---| | Routed family SDPA package | library/dispatch/sdpa/ (bind_unified.py, bind_dense.py, bind_shared.py) | Added in this PR | | Exported route algorithm sets | dispatch.attention.UNIFIED_ALGORITHMS…
• §K: Identity table: no knob axes, declared defaults, canonicalization rules, or knob-driven builder semantics are added or modified in this PR (unified_rules.py only renames _problem to attention_problem, and the attention_tiled_2d.py edits fix unconditional bounds clamping in the 4-warp GQA emitter regardless of knobs); config_key and TUNING_ID_VERSION are unchanged.

Graded with the rocKE PR checklist, one model pass per section. FAIL rows that sit on a diff line are inline comments.

§A graded 14/14

#	Verdict	Evidence
1	NEEDS-INFO	Unresolved since previous review: dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6017 and :6372 modify the gfx942 4-warp attention builders (_build_gfx942_4warp_gqa_lean and build_gfx942_4warp_gqa) without a C++ update in the PR diff, and known_gaps.md and cpp/instances/ are not in the diff to verify whether the 4-warp GQA body has a C++ twin or is Python-only.
2	N/A	The PR does not modify any core lowering ops or IR op-handlers under core/lower_* or core/ir.
3	PASS	New dispatch and routing logic is confined to dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:1 and dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:1 (plus public exports in…
4	N/A	No C++ kernel twin or C++ IR emitter code is modified in this PR.
5	PASS	Both dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8 and dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8 re-bless only the…
6	N/A	No C++ twin is touched and the PR does not claim cross-engine parity from a golden update.
7	PASS	The bounds-clamping and no-key fixes in dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6017, :6067, :6159, :6372, and :6749 apply only to the 4warp_gqa builders, and the golden diffs in…
8	N/A	No rocKE C++ files are modified or added in this PR.
9	N/A	No shared op-handler or lowerer arch divergence fix is included in this PR.
10	N/A	No new kernel IR feature flag is introduced; the PR only adds host-side SDPA dispatch bindings and fixes bounds clamping in…
11	N/A	No new IR op is added in this PR.
12	N/A	core/backend.py and ROCKE_BACKEND handling are not touched in this PR.
13	N/A	LLVM .ll emission and debug-location handling are not modified in this PR.
14	N/A	No helper files under helpers/ or platform/cpp/helpers/ are modified in this PR.

§B graded 6/6

#	Verdict	Evidence
1	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:174-189 classifies implementation-shortcut declines with NOT_YET_IMPLEMENTED: separately from ABI/kernel limits, and…
2	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:211-214 probes each dense candidate with opt_in_probe, checks candidate.admits(probe), and builds the candidate spec only after admits succeeds…
3	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:219-226 declines requests when no unified candidate covers them, and dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:432-434 and…
4	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/hipdnn.py:220-222 rejects a node missing its 'type' key instead of defaulting to 'SdpaAttributes', and dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:243-248 and 260-265…
5	PASS	dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:49-57 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_numeric.py:43-51 run GPU numeric verification over every case admitted by sdpa_unified_admits…
6	N/A	The diff adds SDPA dispatch admission/binding logic and kernel bug fixes, without adding or modifying an is_valid_spec or…

§C graded 10/10

#	Verdict	Evidence
1	PASS	Shared predicates, decline reasons, and translation logic across unified and dense SDPA bindings are factored into dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:44-244, torch_dtype and device_tensor are consolidated in…
2	PASS	The SDPA bindings route to existing public entrypoints (bind_dense_attention_torch in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:342-344 and run_unified_attention_torch in…
3	N/A	No new kernel builder functions or kernel_name() definitions are introduced in this PR.
4	PASS	Arch-independent SDPA request/binding logic lives in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:1-326, while gfx942-specific kernel fixes stay in…
5	PASS	All touched files in the PR are directly scoped to SDPA dispatch/binding (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/__init__.py:1-152), SDPA tests (dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:869-871…
6	PASS	Shared SDPA request, band, decline-table, and operand abstractions are cleanly factored into dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:44-326, and single-source public attention-layer accessors (attention_band…
7	N/A	This PR is a feature routing and bug-fix change rather than a pure no-op cleanup.
8	PASS	The PR reuses existing develop helpers including opt_in_probe, empty_workspace, and resolve_stream (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:31-32), validate_lengths…
9	NEEDS-INFO	In dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:5862-5865, 6017, 6067, 6159-6161, 6372-6394, 6576, and 6749-6751, the Q-row/key clamping and no_key zeroing fixes were applied across both the 4-warp lean and 4-warp GQA bodies. — no diff line supports this verdict
10	PASS	Every feature in SDPA_FEATURES is explicitly partitioned into carried vs declined sets in both dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:84-130 and…

§D graded 21/21

#	Verdict	Evidence
1	N/A	This PR does not add or modify any KernelSpec knobs or environment variable reads.
2	PASS	Coverage limits and algorithm constants are imported from kernel/attention modules rather than transcribed (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:38 imports UNIFIED_ALGORITHMS and UNIFIED_HEAD_SIZES…
3	PASS	All candidate admission checks go through candidate.admits(probe) or ATTENTION_ROUTE_REGISTRY.supported(...) and never call _supports directly (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:212…
4	PASS	Selection and admission are pure functions of SdpaRequest taking req.arch directly without any device probes (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:72…
5	N/A	This PR does not introduce or modify any KernelSpec identity, spec_hash, compile_key, or kernel_name() methods.
6	N/A	This PR does not introduce or modify cohort routing, spec-builders, or kernel cache keys.
7	N/A	This PR does not add hand-pinned heuristic overrides.
8	N/A	This PR does not define or modify any kernel cohort predicates.
9	N/A	No new KernelCandidate or architecture is registered; the PR routes SdpaRequest onto existing ATTENTION_ROUTE_REGISTRY candidates.
10	PASS	Every SdpaRequest field and feature is translated into AttentionRequest and routed to the unified and dense bindings with dispatcher-level wiring, entry-point, and numeric tests (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:44…
11	PASS	Existing opt-in dense candidates remain opt-in and are queried explicitly via opt_in_probe(attn, candidate) in the SDPA dense gate (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:211).
12	PASS	Routing priority order is preserved and tested: dense SDPA asserts the first gfx950 variant in registry order is chosen at dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_wiring.py:312, and unified SDPA asserts returned candidates are sorted…
13	PASS	Pinned SDPA requests are explicitly declined instead of falling back to auto at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:231, and tested for known and unknown pins at…
14	N/A	This PR does not modify compiled-binary cache keys, selection-key logging, or any candidate kernarg ABI version.
15	N/A	No new KernelCandidate or arch registration module is added; SDPA routes through existing attention candidates.
16	PASS	registered_sdpa_combos, sdpa_sweep_space, and dispatch_sdpa_all short-circuit on _request_errors(req) before delegating at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/init.py:119, with coverage in…
17	N/A	No new KernelCandidate is registered in this PR, so per-candidate CPU routing and family fixture invariants do not apply.
18	N/A	This PR does not change registry ratchets, route-only exemptions, or candidate binding requirements.
19	PASS	The unified SDPA binding accurately documents and uses run_unified_attention_torch with no tuning spec (backend="auto") at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:440, while dense SDPA launches through…
20	N/A	The selection/routing changes are confined to the Python library dispatch layer (library/dispatch/), which has no C++ twin.
21	PASS	Band and scale mismatches are rejected in the chooser/admission gates rather than hacking the emitter, at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:151 and…

§E graded 17/17

#	Verdict	Evidence
1	PASS	Under platform, the current diff adds installed-test wiring in dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:867 and torch runtime helpers in dnn-providers/hip-kernel-provider/rocke/platform/python/rocke/runtime/torch_interop.py:96, with no platform…
2	PASS	Library SDPA binding code imports torch glue from rocke.runtime.torch_interop rather than importing torch directly, e.g. dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:32 and…
3	PASS	Torch-dependent tests gate torch with pytest.importorskip, e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_binding.py:22, and runtime torch access stays behind helper functions such as…
4	PASS	New dispatch code imports downward to kernels/platform helpers, e.g. dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:25 imports kernels.common._sdpa_seqlen and…
5	PASS	The platform installed-test entry uses the existing relative-test variable in dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:867 and adds relative ignore paths in dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:869.
6	PASS	Emitter-tree changes add explicit IR behavior rather than deleting unused handles, e.g. dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6154 and…
7	PASS	The routed-family process documentation moves with the SDPA routing change in dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:588.
8	PASS	New docs use descriptive references such as the routed-family section in dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:590 rather than code comments containing tracker IDs, reviewer handles, or PR numbers.
9	NEEDS-INFO	New post-review diff content includes changes outside the SDPA/rocKE scope, such as projects/composablekernel/include/ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp:1824 and projects/hipblaslt/clients/common/include/testing_matmul.hpp:1238. — no diff line supports this verdict
10	NEEDS-INFO	Not resolved since the previous review: the diff still does not include the requested merge-simulation record listing open PRs that touch the same helpers/tests/goldens, agreed land order, and affected-suite reruns on current develop plus those PRs.
11	PASS	The rocKE additions are deliverable docs, source, tests, and installed-test wiring, e.g. dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:588, dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:1, and…
12	PASS	Both gfx942 and gfx950 are explicitly exercised by SDPA tests, e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_wiring.py:37 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_wiring.py:26.
13	PASS	The public rename away from private attention helpers updates definitions and callers in the same diff, e.g. dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention/common.py:753 defines attention_band and…
14	NEEDS-INFO	New post-review changes include omitted deletion hunks, so the diff provided is insufficient to verify that removed upstream code, comments, or KNOWN LIMITATION notes were intentionally ported after the rebase/port.
15	PASS	New helper parameters are used in their bodies, e.g. dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:240 uses req_tensor/tensors/key and dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:249 uses…
16	N/A	This item is a meta-instruction for reviewer grading rather than a property of the PR code.
17	PASS	no absolute throughput, bandwidth, or latency figure in added lines

§F graded 11/11

#	Verdict	Evidence
1	PASS	Branch users/tlakshma/hack/sdpa/st2-unified matches the recognized users/<user>/... prefix, and PR changes are in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/__init__.py:1.
2	PASS	title feat(hip-kernel-provider): route SDPA to unified and dense attention kernels matches type(scope): summary and is 76 characters
3	PASS	description is at least 30 characters and contains an issue reference
4	PASS	test file dnn-providers/hip-kernel-provider/rocke/library/tests/dispatch/attention/test_additive_registration.py is in the diff
5	PASS	All changed files in the diff (such as dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:1 and dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:1) contain only code, docs, CMake, and test assets; no keys…
6	NEEDS-INFO	Still unresolved since previous review: local pre-commit execution status or CI lint check output is not visible in the diff.
7	NEEDS-INFO	Still unresolved since previous review: GitHub review-thread state and ticket IDs for any deferred review comments are not included in the provided PR context.
8	N/A	The PR targets Base: develop rather than a stacked feature branch, so the stacked-PR bot-skipping case does not apply.
9	N/A	There is no indication that the PR was artificially split or trimmed for a size cap (in fact, #14624 was folded into this PR).
10	PASS	The added and updated tests are substantive wiring, CPU data-movement, case-table, and GPU numeric tests rather than throwaway stubs, e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_binding.py:82 and…
11	PASS	The PR metadata header in this prompt shows the updated title (feat(hip-kernel-provider): route SDPA to unified and dense attention kernels) and full multi-section body (covering dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:1 and…

§G graded 34/34

#	Verdict	Evidence
1	PASS	Correctness is verified against an independent mathematical reference oracle rather than relying solely on IR byte-identity (dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:61 and…
2	PASS	On-GPU numeric tests against the reference oracle are included and enrolled in the GPU test suite (dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:1007 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:61).
3	NEEDS-INFO	In the new commits, attention_d128_swa.json is updated across all four LLVM flavors (dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8) and attention_d256.json is updated (dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8), but the tail of attention_d256.json is still…
4	PASS	New admission, binding, and length-validation rules have falsifiable unit tests checking exact decline prefixes, non-integer dtype rejection, data movement, and family entry points…
5	PASS	In test_an_out_of_range_block_table_physical_id_is_rejected_before_launch (dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:83), the docstring explicitly documents why the test enters at validate_tuning_attention_tensors…
6	PASS	Prose claims in KERNEL_AUTHORING.md, docstrings, and code comments (including the 4-warp GQA clamping comments in dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6363) match the adjacent code and map to concrete tests in…
7	PASS	KERNEL_AUTHORING.md specifies the reference oracle and four test layers as a first-class gate (dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:624), and the CPU and GPU binding tests validate against sdpa_reference…
8	PASS	Bindings and test harnesses perform strict dtype checks and loud-reject or skip unsupported dtypes/arches (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:267…
9	N/A	This PR does not introduce or claim any LDS size or occupancy numbers.
10	N/A	This PR is a functional routing and bug-fix change, not a performance optimization PR.
11	PASS	Kernel reachability through the SDPA routing gates and family entry points is verified in dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_wiring.py:480…
12	N/A	The changes in this PR are in Python-only dispatch layers and a gfx942 kernel builder…
13	N/A	No new LLVM intrinsic overload classes or mangled symbols are added or recommended.
14	N/A	No hardware bottleneck classification claims are made in this PR.
15	N/A	The PR diff is provided directly in the review prompt rather than inspected from a local git checkout.
16	PASS	Integration with existing attention dispatch symbols (ATTENTION_ROUTE_REGISTRY, UNIFIED_ALGORITHMS, DENSE_ALGORITHMS, attention_band, attention_problem, dense_scale_range, opt_in_probe) is verified by wiring and coverage tests…
17	N/A	No standalone CLI or benchmark scripts are shipped in this PR; new files are library modules and pytest suites enrolled in CMake.
18	N/A	This PR does not add statistical runtime guards or back-pressure detectors of the kind this item targets.
19	NEEDS-INFO	Fixtures include shapes/inputs that can express the named failures, including bad per-batch lengths in dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_binding.py:139 and an out-of-range block-table physical id in dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:77. — no diff line supports this verdict
20	N/A	This PR does not add a claim-ban detector or regex sweep.
21	N/A	This PR does not introduce performance measurement or profiling deliverable code.
22	PASS	The SDPA numeric GPU tests are excluded from the host SDPA pytest shard but explicitly enrolled in the GPU CTest lane at dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:1004 and dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:1005, and…
23	N/A	This PR does not make speedup or routing A/B performance claims.
24	NEEDS-INFO	The diff shows re-blessed attention golden hashes, but the GPU validation hunk for the exact moved 4-warp GQA configurations is omitted; an untruncated diff or cited test lines proving coverage for each moved golden case are needed.
25	N/A	This PR is a functional routing and bounds/keyless-row correctness change, not a codegen micro-optimization report.
26	N/A	This PR does not make campaign-branch benchmark speedup claims.
27	N/A	This PR does not ship a performance knob or A/B toggle.
28	N/A	This PR does not present counterfactual performance tables.
29	N/A	This PR does not swap guarded math primitives for fast primitives.
30	N/A	This PR makes no register-spill or scratch-reduction claims.
31	N/A	This PR does not claim a performance win over a baseline defect.
32	N/A	The PR makes no timing, throughput, spill, or other measured performance-number claim requiring reproducible provenance.
33	N/A	The rocKE diff exposes and reuses attention routing helpers and adds SDPA gates, but does not modify an existing scalar…
34	PASS	The optional stats/LSE path is tested with the flag present on padded/keyless rows, asserting O is zero and stats is -inf at dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_binding.py:130 and…

§H graded 5/5

#	Verdict	Evidence
1	PASS	sdpa_unified_bind allocates k_cache/v_cache sized to slots = batch * self.pages at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:352-358 and validates the launch tensors against plan.problem via…
2	PASS	_LaunchPlan builds an identity block_table in range(slots) at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:361-363, clamps logical page indices to k_last_page on device at…
3	NEEDS-INFO	Resolved since previous review: the new commits enforce that BS = spec.block_size is a power of two (if BS & (BS - 1): raise NotImplementedError(...)) and divides BN (if BN % BS: raise NotImplementedError(...)) in build_gfx942_4warp_gqa at dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6203-6219, and compute k_last_page via b.lshr(..., b.const_i32(BS_LOG2)) at…
4	NEEDS-INFO	In build_gfx942_4warp_gqa, the new commits clamp page to k_last_page in phys_key at dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6372 and 6391-6395 and zero O when context_off + q_inseq < 0 at dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6159-6160 and 6749-6750, while bind_unified.py zeroes k_cache/v_cache at…
5	NEEDS-INFO	Resolved since previous review: seq_len_q and seq_len_kv are host-validated via _host_lengths and validate_lengths (preserving dtype to reject non-integers) at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:243-265 and dnn-providers/hip-kernel-provider/rocke/library/kernels/common/_sdpa_seqlen.py:45, operand shapes/strides/ragged offsets are validated in Operand.__init__ at…

§I graded 6/6

#	Verdict	Evidence
1	N/A	This PR wires hipDNN SDPA forward requests to existing attention candidates and fixes gfx942 kernel edge cases without introducing…
2	PASS	Non-contiguous or strided layouts are repaired via copy buffers (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:276-284), unsupported features, bands, and non-power-of-two/non-divisible block sizes are loudly rejected…
3	PASS	Curated request and kernel case tables (dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/cases.py:1361-1439) are combined with generated parameter sweeps and explicit algorithm-partition coverage checks…
4	N/A	This PR does not add or modify performance routing comparison tables or tuning reports.
5	N/A	This PR does not introduce baked tuning search results or static tuning snapshots.
6	N/A	This PR does not add any new DSL ops or hardware-instruction-named primitives.

§J graded 7/7

#	Verdict	Evidence
1	N/A	The only Markdown file in the PR (dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md) contains no mathematical…
2	N/A	The KERNEL_AUTHORING.md diff introduces no mathematical formula symbols, decay/log-decay domains, or log bases.
3	PASS	dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:591-642 links to existing sections [E](#e-new-operator-with-a-hipdnn-graph-node), [F](#f-new-operand-or-output-on-an-existing-operator), and [A](#a-new-kernel-family) while stating only the delta for…
4	PASS	The novelty claims in dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:321 and dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:638-641 name library/dispatch/sdpa/, bind_unified.py, bind_dense.py…
5	PASS	dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md makes no scaling/complexity claims, and the O(s_q) interval comparison claim in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:94-95 states the aggregate complexity across…
6	N/A	The KERNEL_AUTHORING.md diff does not introduce any model configuration tables or tuned band-edge constants.
7	N/A	This PR edits the existing dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md and does not add a new design or…

§K graded 13/13

#	Verdict	Evidence
1	N/A	This PR adds no multi-configuration candidate (SDPA_REGISTRY remains empty and SDPA delegates to existing attention candidates…
2	N/A	This PR does not add or modify any tunable kernel spec fields, KnobAxis declarations, or field classifications.
3	N/A	This PR does not change config_key payloads, canonical-knob rules, declared defaults, or knob meanings (unified_rules.py only…
4	N/A	This PR does not modify KnobSpace.defaults() or KnobSpace.recorded() in any KnobSpace subclass.
5	N/A	This PR does not touch knob drop/refuse rules, inert(), base_value(), or KNOWN_WRONG_KNOBS.
6	N/A	This PR does not touch KnobSpace.validate or library/dispatch/attention/tuning_specs.py.
7	N/A	This PR does not modify KnobSpace.base(req) or _tuning_problem (it only renames _problem to attention_problem in…
8	N/A	This PR does not modify normalize_knobs or request-boundary knob scalar coercion.
9	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:231 (value_rejects) explicitly declines any pinned algorithm, spec_id, tuning_id, or tuning_knobs rather than adopting canonicalized ids or falling back to auto, tested in…
10	N/A	This PR does not modify KnobSpace.find or tuning_id resolution logic.
11	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/__init__.py:119 (registered_sdpa_combos, sdpa_sweep_space, dispatch_sdpa_all) delegates to SDPA_REGISTRY and registered_attention_combos / dispatch_attention_all filtered by _serves(req)…
12	N/A	This PR does not reorder any production stack, change accept_default, or edit a tuning variant's base spec.
13	PASS	The benchmark changes in dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/attention_combo_sweep.py:44 and dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx950/attention/decode/decode_dispatch_diagnose.py:13 only rename _problem to…

§L graded 10/10

#	Verdict	Evidence
1	N/A	This diff does not add or modify a rocKE gfx target in arch specs or ISA backend registries.
2	N/A	This diff does not add or modify an MMA op_id, MMA fragment-info row, or MMA lowering layout map.
3	N/A	This diff does not add or change MMA atom selection, MMA catalog fields, lane maps, or intrinsic-name emission.
4	N/A	This diff does not introduce or change wave-size handling or lane math derived from wave size.
5	PASS	The arch-specific codegen change is confined to the gfx942 4-warp kernel and masks fully masked rows without leaking gfx950 recipes at dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6157.
6	N/A	This diff does not move pipeline vocabulary onto ArchTarget/core arch data, mutate ArchTarget from policies, or import CK Tile…
7	N/A	This diff only routes or declines SDPA dtype features and does not change rocKE logical dtype, bit-packing, scalar lowering, or…
8	N/A	This diff does not touch scaled-MMA scale contracts, retired scaled WMMA ids, or MfmaAtom.emit_scaled lowering.
9	N/A	This diff does not add a new rocKE architecture target.
10	N/A	This diff does not touch intrinsic_arch_domain..json, gen_arch_domain.py, or LLVM_FLAVORS.

§M graded 10/10

#	Verdict	Evidence
1	NEEDS-INFO	dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:5865, :6159, and :6749 use b.cmp_lt and b.select for device-time predicates rather than Python control flow. — no diff line supports this verdict
2	N/A	The PR diff does not touch legalize_for_target or target legalization logic.
3	N/A	The PR diff does not change print_ir, ck.dsl.ir/v1 serialization, wire versions, attrs, SSA ids, or parameter ordering.
4	N/A	The PR diff does not change portable IR schemas, importers, magic division constants, recipe VM code, or recorder coverage.
5	N/A	The PR diff does not touch AsyncTileLoader, async DRAM-to-LDS loading, LDS async layouts, or VMEM waits.
6	NEEDS-INFO	dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6017, :6067, :6382, and :6393 clamp the Q row, key, and KV page indices before loading, avoiding loads from out-of-bounds addresses. — no diff line supports this verdict
7	N/A	The PR diff does not add or modify buffer-resource word 3 handling or TemporalHint.STREAMING lowering.
8	N/A	The PR diff does not add LLVM intrinsics or change backend selection precedence.
9	N/A	The PR diff does not touch public C entry points, cpp/include/rocke/ headers, or C++ codegen launch behavior.
10	N/A	The PR diff does not emit !dbg metadata or change debug-info consumers.

§N graded 18/18

#	Verdict	Evidence
1	N/A	The diff does not add or promote a helper under platform/python/rocke/helpers/.
2	NEEDS-INFO	dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6576 uses the shared q_desc.offset TensorView mechanism for Q addressing after the clamped row/head selection, rather than hand-expanded Q offsets. — no diff line supports this verdict
3	N/A	The diff does not touch SoftwarePipeline, sync_before_issue, or overlap_vmcnt behavior.
4	N/A	The diff does not add or modify fused kernels, EpilogueOp, fusion schedules, or fusion validation paths.
5	N/A	The diff does not add or modify a public API under rocke.helpers.tiling.
6	N/A	The diff does not add or modify front-door/toolbox tiling helpers or LayoutStyle classes.
7	N/A	The diff does not touch tiling layout overrides, accumulator c_desc derivation, or MMA soundness gates.
8	N/A	The diff does not touch helpers/tiling LDS allocation, lds_store/lds_read, contraction padding, or wide-load tiling emit rules.
9	N/A	The diff does not use or modify N-D tiling axis roles, at_index/squeeze behavior, FlowStage labeling, or AB-swap routing.
10	N/A	The diff does not modify tiling emit, LDS staging, or MMA issue code in helpers/tiling.
11	N/A	The diff does not touch tiling MMA traits, TileMma target selection, lds_swizzle policy, or derive_sched_group_counts.
12	N/A	The PR diff makes no helpers/tiling LDS bank-conflict or coalescing optimization claim requiring simulator, counter, or ASM-backed…
13	N/A	The diff does not construct WarpDistributionEncoding, change make_tile_desc factoring, re-export RegisterMapper, or alter tiling…
14	N/A	The diff does not use or modify helpers/tiling verbs or add a new helpers/tiling LDS pipeline variant.
15	N/A	The diff does not touch helpers/tiling descriptor allocation or transformed-range bounds checks after swizzle, pad, or XOR.
16	N/A	The diff does not use or modify optimize_layout, recommend(), or evaluate_transform ranking behavior.
17	N/A	The diff does not modify tiling_recorder.py, round-trip analysis, or tiling visualization emission behavior.
18	N/A	The diff does not touch InterleavedStyle, helpers/tiling/kernels demos, or tiling numeric gates for interleaved layouts.

§O graded 16/16

#	Verdict	Evidence
1	N/A	The PR does not add or modify IR address spaces, required attributes, or core/verify.py.
2	N/A	The PR does not modify scf.for structure or terminator emission in core IR/lowerers.
3	N/A	The PR does not touch IRBuilder, intrinsic declaration tables, or metadata slot numbering.
4	N/A	The PR does not modify IR types or SmemType serialization/parsing.
5	N/A	The PR does not modify op purity tables or core/passes.py / cpp/core/passes.cpp.
6	N/A	The PR does not modify constant folding in Python or C++ passes.
7	N/A	The PR does not modify kernel attributes, codegen policy, or cluster dimension checks.
8	N/A	The PR does not set or modify scheduler strategies or CodegenPolicy.
9	N/A	The PR does not modify waitcnt encoding or barrier lowering.
10	N/A	The PR does not modify lowerer attribute validation or pointer parameter lowering.
11	PASS	dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_numeric.py:30 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:36 resolve the device architecture via…
12	N/A	The PR does not add hardcoded width/size caps in core/codegen or silent arch or "gfx950" defaults.
13	N/A	The PR does not touch any public C structs or enums under cpp/include/rocke/.
14	N/A	The PR does not touch C++ sources or add C++ compiler flags in platform/CMakeLists.txt.
15	N/A	The PR does not modify C++ engine fallback mechanisms or lowerer entry points.
16	N/A	The PR does not touch the C++ codegen library or gfx1250 builders/lowerers.

§P graded 14/14

#	Verdict	Evidence
1	N/A	The visible diff does not add or modify platform instance builder signatures or spec REQUIRED_FIELDS.
2	N/A	The visible diff does not touch platform kernarg signatures, manifest args_signature, runtime argument packing, or GEMM/conv ABI…
3	N/A	The visible diff does not touch GEMM, implicit-GEMM, or block-scaled GEMM launch geometry.
4	N/A	The visible diff does not modify LDS allocation gates or C++ validators.
5	N/A	The visible diff does not touch WMMA/RDNA GEMM, GemmMultiD, or GemmMultiAbd admission paths.
6	N/A	The visible diff does not touch Stream-K, MoE sorting, fused MoE, or split-K caller-zeroed workspaces in rocKE.
7	N/A	The visible diff does not touch Stream-K rejection rules, persistent GEMM, GroupedGemmLauncher, or SUPPORT_MATRIX.md.
8	N/A	The visible diff does not touch quantization helpers, MX GEMM, or MX scale decoding.
9	N/A	The visible diff does not add or modify numeric emulation or low-precision decode paths requiring edge-case emulation tests.
10	N/A	The visible diff does not add a manifest kind or touch helpers/manifest.py, run_manifest, or manifest allocate/verify consumers.
11	PASS	The new SDPA bindings allocate and retain launch buffers once in the bound plan and launch on resolve_stream: see dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:341-352 and…
12	N/A	The visible diff does not touch COMGR or HIP library loading logic.
13	N/A	The visible diff adds SDPA routing and binding logic rather than refactoring shared host-side grid, tile-count, workspace…
14	N/A	The visible diff does not touch examples/, helpers/ placement, or add new modules under platform instances/common family glue.

§Q graded 7/7

#	Verdict	Evidence
1	NEEDS-INFO	Visible SDPA tests are registered in installed CTest entries at dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:867 and GPU numeric entries at dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:1004, but the current-head hunks for additional new tests are omitted and the diff does not show ROCKE_ENGINE_test_categories_external.yaml entries or the CI lane that runs them.
2	N/A	No new kernel family or variant is added; the PR routes SDPA to existing attention candidates and edits existing gfx942 attention…
3	PASS	The representative golden tripwire is updated for the changed attention kernels at dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:12 and…
4	PASS	The affected representative goldens are updated across repeated LLVM-flavor blocks, including dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8 and…
5	NEEDS-INFO	Resolved for the visible SDPA tests: GPU numeric tests derive the active architecture and skip when unavailable at dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:26 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_numeric.py:24, and torch-free length validation remains covered at…
6	N/A	This PR does not refactor a platform validation gate or differential harness with per-configuration timeout, hash fallback, or…
7	N/A	This PR does not change ROCKE_BACKEND, ROCKE_CPP_STRICT, binding-present fallback behavior, or engine rejection contracts.

§R graded 11/11

#	Verdict	Evidence
1	N/A	The diff routes SDPA over existing attention candidates and changes correctness/address guards in an existing gfx942 kernel, but…
2	NEEDS-INFO	The changed KV page clamp and Q-row hoist in dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6364 and dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6374 need a runbook §6/§8 mapping, matching §13 attention checklist entry, and gfx942 arch recipe/ArchTarget rationale; those are not provided in the diff.
3	NEEDS-INFO	The global-load address changes at dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6063 and dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6364 need utilities-skill evidence for prefetch/data-load or ISA inspection, especially because dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6368 makes a measured…
4	NEEDS-INFO	The intended primitive/address-computation change is stated at dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6374, but the diff does not provide before/after probe_occupancy.py, probe_regs.py, or probe_intrinsic_counts.py/probe_isa_inspect.py evidence for the claimed movement.
5	NEEDS-INFO	Changes since the previous review do not resolve the prior NEEDS-INFO: the current diff still changes the Q load address at dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6063 and the block-table/KV page address at dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6364, but no…
6	N/A	The rocKE diff does not change an LDS layout, MMA layout, tiling pipeline identity/relabel/reorder, or bank-conflict lever.
7	N/A	The rocKE diff makes no bottleneck claim such as memory-bound, LDS-bound, barrier-bound, or latency-bound.
8	N/A	The rocKE diff does not add, remove, or modify waits, barriers, sched_barrier, sched_group_barrier, iglp_opt, or scheduling hints.
9	N/A	The rocKE diff is not presented as a performance-regression fix and does not add a shape-trading performance heuristic or dispatch…
10	NEEDS-INFO	The new comments make measured-performance claims at dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6368 and dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6374, but the diff does not include runbook §2.3 benchmark-hygiene evidence for those claims.
11	N/A	The documentation added is an SDPA routed-family authoring checklist, not a new optimization failure mode, heuristic, case study…

§S graded 8/8

#	Verdict	Evidence
1	NEEDS-INFO	The review-response change resolves part of the prior gap by adding an all public-surface test in dnn-providers/hip-kernel-provider/rocke/library/tests/dispatch/attention/test_attention.py:250 and SDPA host/GPU lane entries in dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:867 and dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:1004, but the visible diff still does not show the…
2	PASS	Existing behavior changes keep updated tests: missing hipDNN node type now raises in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/hipdnn.py:217 and is covered by dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_sdpa_from_hipdnn.py:454…
3	N/A	The visible rocKE diff does not remove or rename a public API; underscore-prefixed _problem and _validate_lengths migrations are…
4	PASS	The prior FAIL is resolved: dispatch.attention.all is edited in dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention/init.py:561, and the new public-surface test imports and resolves every listed name in…
5	PASS	API tests assert behavior and negative contracts: torch_dtype resolves names and rejects an invalid dtype in dnn-providers/hip-kernel-provider/rocke/platform/tests/test_rocke.py:7658, device_tensor verifies values/dtype/device forwarding in…
6	PASS	CPU-checkable SDPA contracts are kept in the host SDPA pytest lane while GPU numeric files are explicitly ignored there in dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:867, GPU numeric SDPA tests are separately added to the GPU pytest file set in…
7	N/A	The added and changed public APIs in this section are Python dispatch/runtime helpers or SDPA binding-layer APIs without a C++…
8	NEEDS-INFO	The prior NEEDS-INFO is not resolved: the visible diff does not contain a PR test-run statement showing the README entry points from rocke/platform were run, nor GPU result arch names for the new GPU SDPA coverage.

[cgmillette]

Automated review: Model - qwen3.6:35b-a3b-coding-mtp-q4_K_M, effort: max

PR #13115 — route SDPA to unified and dense attention kernels (ThruptiRajLakshmanaGowda, AICK-2613, AICK-2614) — 🟡 Follow-up: 14929 merge holds; nits remain

Prior item	Status
Earlier 🟢 rows: arai713, AnaghaRaoAMD ×6, dlangbe, Copilot, CongMa13 ×2, yraparti (10-31); our N1–N10, N12; 14483 left bound	🟢 FIXED, unchanged • SDPA, attention and golden checks pass on the merge (ran)
AnaghaRaoAMD (11-03) bind_unified.py:396: K/V copied every launch	🟡 DEFERRED · AnaghaRaoAMD decides • unchanged; in the body's Not in this PR (read)
yraparti (11-04) ×3: torch import; seq_len tests; __all__ test	🟢 FIXED 9db85cee, b7ed4698, b7923d9d, 33013ead • each bind_* __all__ now tested; a bad name fails it (ran) • left: ragged-offset count and range checks untested (ran) • §S.4 thread open, no in-thread reply; yraparti resolves
Our N11: body omits the new public API; names private helpers	🟡 OPEN • fixed: torch_interop "gains"; Tests section (read) • left: kernel section still claims 14929's two clamps (read) • left: GPU box open; host counts predate the merge (read)
Our N13: same 4-warp bounds fix as 14929	🟢 FIXED 6daf7e6b • builder vs develop: only the no-key-row O = 0, both bodies (read) • 14929's 8 bounds tests ✓; both golden shards match ×4 (ran) • left: O = 0 has no CPU test like 14929's (ran)
Our N14: _host_lengths cast lengths to int first	🟢 FIXED 33013ead • float32 seq_len_q now refused at bind, not truncated (ran) • left: new test calls _lengths; passes with the old read (ran) • fix: a float32 tensor through sdpa_unified_bind; drop the wiring copy
Bots / CI	🟡 running, none failed • gfx94X rocKE shards ✓; new tests and 14929's passed (read) • running: 4 Windows gfx110X test jobs; 39 ✓, 0 ✗ • merge: BLOCKED; no approval; 8/29 threads open
Our checks	🟢 GREEN • vs 027419e6: clean, 4 behind; same provider tree as the checks' merge (ran) • build + ctest 46 ✓ / 1 skip; byte identity, parity, golden ×4; dg 0 changed (ran) • tests 4148 ✓, 25 ✗ on develop too; installed 3862 ✓; torch 5615 + 1045 ✓ (ran) • gfx1151 GPU: 55 cases match; compile 29 ✓, 0 spills; rules 1–3 clean (ran)
Code quality	🟡 Minor • 14929's clamps reused; no second clamp or q_row (read) • float seq_len refused; its new test cannot fail: N14 (ran) • offsets still truncated: N15; __all__ test repeats develop's list; earlier 🟡s unchanged (ran)

#	Sev	file:line	Finding
N15	🟡	dispatch/sdpa/bind_shared.py:251	host_ints still casts ragged offsets with int(): N14's other caller • impact: float32 offsets [0.5, 1.5, 2.5, 3.5] bind like int32 [0, 1, 2, 3]; the request says int32 or int64 (ran) • fix: Operand refuses offsets not of desc.ragged_offset.dtype; add a float case

Utility: 🟢 Positive — routes hipDNN SDPA to unified and dense kernels; reuses 14929's clamps.

---

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

Follow-up review: first pass by a local qwen3.6 model, verified by the orchestrator (Claude Opus 5.5). Checks: delta since b7923d9d, merge with develop 027419e6, 14929's bounds tests and both golden shards, reply claims by source mutation, SDPA tests with torch, binding launches on a GPU, CI at head, code quality. Flags whether a closer look is needed; not an approval or rejection.

[Copilot]

🟡 Changes recommended

Critical SDPA binding defects break request semantics, ignore caller streams, and pass misaligned pointers to kernels, risking incorrect results and data races.

7 open issues

• Preserve the SDPA contract in returned dispatch results · New
• Honor the requested stream for the full launch sequence · New
• Reject non-integer ragged offsets · New
• Route misaligned Q/O views through the workspace · New
• Reject non-positive scales or apply masking after scaling · New
• Check pointer alignment before using Q/O in place · New
• Honor the requested stream instead of discarding launch arguments · New

2 resolved since last review

• Preserve registry priority when ordering matching attention routes
• Remove tracking IDs from caller-visible errors, tests, and docs

_{Review effort: Max}

[ThruptiRajLakshmanaGowda]

@cgmillette thanks for the follow-up review. Addressed in head 7ad42faf6aa (also merges develop at 10805993013, after #15024: develop's runtime-head signatures kept, no-key-row O = 0 kept, attention_d128_swa/attention_d256 golden shards re-recorded, IR parity OK at llvm20/22/23/24):

• N15 → 8ad171d67b6: Operand refuses ragged offsets whose dtype is not desc.ragged_offset.dtype before reading them. New test_ragged_offsets_of_another_dtype_are_refused (float32 [0.5, 4.5, 8.5] and int64 against an int32 ragged_offset) fails on the old int() read.
• N14 leftover → 7ad42faf6aa: test_float_seq_len_is_refused_not_truncated now pushes a float32 seq_len_q ([1.5, 1.0], truncates to the valid [1, 1]) through sdpa_unified_bind and asserts the integer-dtype refusal; it fails with the old read. The _lengths copy in the wiring tests is dropped.
• N11 leftover → PR description: the kernel section now credits fix(rocke): bound Q and K/V reads in the gfx942 4-warp GQA D128 body #14929 for the Q and block-table clamps (reused unchanged) and claims only the no-key-row O = 0; host counts are re-run after this merge. The GPU checkbox is unchanged.

[cgmillette]

lgtm - please post fixes to follow:

[cgmillette]

Automated review: Model - qwen3.6:35b-a3b-coding-mtp-q4_K_M, effort: max

PR #13115 — route SDPA to unified and dense attention kernels (ThruptiRajLakshmanaGowda, AICK-2613, AICK-2614) — 🔴 Follow-up: under-aligned Q/O give wrong O

Prior item	Status
Earlier 🟢 rows: arai713, AnaghaRaoAMD ×6, dlangbe, Copilot (earlier), CongMa13 ×2, yraparti (10-31, 11-04 ×3); our N1–N10, N12, N13; 14483 left bound	🟢 FIXED, unchanged • SDPA, attention and golden checks pass on the merge (ran) • AnaghaRaoAMD's 6 threads answered 11-04, still open: the reviewer resolves
AnaghaRaoAMD (11-03) bind_unified.py:396: K/V copied every launch	🟡 DEFERRED · AnaghaRaoAMD decides • reason: a view loses the zero-filled tail; needs a GPU check: sound • in Not in this PR and KERNEL_AUTHORING §H (read)
yraparti (11-04) §S.4: dispatch.attention.__all__ test	🟢 FIXED 6a5761c5 • adding or dropping a name fails the new pin (ran) • CI ran it: both gfx94X shards PASSED (read) • answered 17:15Z; thread open: yraparti resolves
Copilot (11-05) bind_unified.py:345, bind_shared.py:326: in-place Q/O not alignment-checked	🔴 OPEN • gfx1151, Q/O at a 2-byte offset: wrong O on 6 of 8 admitted cases (ran) • aligned twins match; copying misaligned views fixes all, SDPA tests 5617 ✓ (ran) • no reply; fix: in place only if data_ptr() % 16 == 0, as _DensePlan
Copilot (11-05) bind_unified.py:186: non-positive scale admitted	🟠 OPEN • gfx942 4-warp bodies mask at -1e30 before scaling (read) • gfx1151's kernel is right at 0 and -0.5 (ran) • no reply; fix: decline non-finite or ≤ 0 scales, like dense's _scale_reject
Copilot (11-05) bind_unified.py:445, bind_dense.py:348: stream discarded	🟠 OPEN: N16 • reproduced on both bindings (ran) • no reply
Copilot (11-05) sdpa/__init__.py:146: raw attention DispatchResults	🟡 OPEN • each result binds an AttentionRequest, which has no scale (ran) • no caller launches them yet (read) • no reply; fix: export the SDPA binders; say so in the docstring
Our N14 leftover, N15; Copilot (11-05) bind_shared.py:253: float lengths and offsets truncated	🟢 FIXED 7ad42faf, 8ad171d6 • float32 seq_len_q and offsets refused at bind (ran) • each new test fails without its fix (ran) • left: torch-only module, CI skips it; Copilot thread unanswered (read)
Our N11: body omits the new public API; names private helpers	🟡 OPEN • fixed: kernel section credits 14929; host counts after 15024 (read) • left: dense_scale_range missing from New public API (read) • left: GPU box says no GPU; the 02:47Z comment reports runs (read)
Bots / CI	🟡 running, none failed • gfx94X rocKE shards 91/91 ✓ ×2; SDPA binding tests skip, no torch (read) • queued: 4 Windows gfx110X test jobs; 40 ✓, 0 ✗ • merge: BLOCKED; no approval; 15/36 threads open
Our checks	🟢 GREEN • vs 05fa3eb7: clean, 15 behind; provider tree = checks' merge (ran) • build + ctest 48 ✓ / 1 skip; byte identity, parity, golden ×4; dg 0 changed (ran) • goldens differ from develop only in 3 gfx942 4-warp entries ×4 (ran) • tests 4177 ✓; installed 3864 ✓; torch 5617 + 1051 ✓ (ran) • gfx1151 GPU: 55 cases match; compile 29 ✓, 0 spills; rules 1–3 clean (ran)
Code quality	🟠 Should-fix • N14, N15 fixed; each new test fails its mutant (ran) • launch() drops stream: N16; dtype rules differ per tensor: N17 (ran) • ragged test fixture copied 3×; earlier 🟡s unchanged (read)

#	Sev	file:line	Finding
N16	🟠	dispatch/sdpa/bind_unified.py:444, dispatch/sdpa/bind_dense.py:347	launch(**_kw) drops the caller's stream, and any other keyword • impact: runs on the current stream, no error; a launch(stream=s) caller races on s (ran) • fix: launch(*, stream=None); refuse a stream it cannot honour; one test each
N17	🟡	dispatch/sdpa/bind_shared.py:292	Each tensor gets its own dtype rule: offsets exact, seq_len any integer, stats none • impact: int64 seq_len_q and fp16 stats bind; LSE silently cast to fp16 (ran) • fix: one check in Operand: tensor vs desc.dtype, offsets vs desc.ragged_offset.dtype

Utility: 🟢 Positive — routes hipDNN SDPA to unified and dense kernels; fix alignment first.

---

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

Follow-up review: first pass by a local qwen3.6 model, verified by the orchestrator (Claude Opus 5.5). Checks: delta since 6daf7e6b, merge with develop 05fa3eb7, golden diff vs develop, reply claims by mutation, SDPA tests with torch, binding launches on a GPU (stream, alignment, scale), CI at head, code quality. Flags whether a closer look is needed; not an approval or rejection.

[ThruptiRajLakshmanaGowda]

Follow-up on the 11-05 22:49Z review (Copilot's 6 remaining threads + N16), pushed as 0d916301ba3:

• Alignment (bind_shared.py:326, bind_unified.py:345): Operand.packed_view now requires data_ptr() % 16 == 0 on top of contiguity, falling back to the workspace copy otherwise -- as _DensePlan already does for its BSHD views. New test_misaligned_packed_bshd_q_and_o_fall_back_to_a_workspace_copy.
• Scale (bind_unified.py:186): new _scale_reject declines a non-finite or non-positive scale at admission (the gfx942/gfx950 tiled 2D bodies mask with -1e30 before scaling, so scale <= 0 would let masked keys back in). New parametrized test_non_positive_or_non_finite_scale_is_declined.
• N16 / stream (bind_unified.py:445, bind_dense.py:348): both launch()s now take an explicit stream keyword and raise NotImplementedError for anything but the current stream, instead of discarding it through **_kw. It can't be honored piecemeal: pack()/unpack() always run on the ambient current stream while the bound workspace buffers are allocated once, at bind time, against whatever was current then. New test_launch_rejects_a_stream_other_than_the_current_one in both binding test files.
• N15 leftover (bind_shared.py:253): this was already fixed in 8ad171d67b6; replied in-thread with that SHA since the thread had no reply yet.

All of tests/sdpa + tests/dispatch/attention pass on CPU (6491 passed, 257 skipped -- the GPU numeric suites skip without hardware); pre-commit run is clean on the changed files.

Left open, by design, for your call rather than guessed at:

• sdpa/__init__.py:146 (Copilot): dispatch_sdpa_all's raw attention DispatchResults would mis-execute if someone calls .bind_torch() on them directly (wrong operand layout, wrong default scale) instead of going through sdpa_unified_bind/sdpa_dense_bind. Fixing this properly means either wrapping each combo's bind_torch with an SDPA-aware adapter that reuses the specific candidate+spec (not just the bindings' single production route), or just documenting/raising loudly against direct use -- both are real options and no caller in this repo exercises .bind_torch() on these results today, so I didn't want to guess which you want.
• N17 (dtype rule per tensor, bind_shared.py:292): noted but not changed this pass -- unifying the per-tensor dtype check in Operand risks touching _DensePlan's separate stats/fp32 handling; flagging rather than risking a behavior change without a design nod.
• AnaghaRaoAMD's K/V-copy-per-launch thread (bind_unified.py:396) stays as previously agreed: "Not in this PR" (KERNEL_AUTHORING.md §H).

[cgmillette]

Automated review: Model - qwen3.6:35b-a3b-coding-mtp-q4_K_M, effort: max

PR #13115 — route SDPA to unified and dense attention kernels (ThruptiRajLakshmanaGowda, AICK-2613, AICK-2614) — 🟠 Follow-up: stream refusal, CI test gaps

Prior item	Status
Earlier 🟢 rows: arai713, AnaghaRaoAMD ×6, dlangbe, Copilot (earlier; 11-05 int(x) offsets), CongMa13 ×2, yraparti (10-31, 11-04 ×4); our N1–N10, N12–N15; 14483 left bound	🟢 FIXED, unchanged • delta reopens none; SDPA, attention and golden checks pass (ran) • int(x) thread reply cites 0d916301; the fix is 8ad171d6 (read)
Open threads: 15 of 36	🟡 OPEN: 11 resolvable • fixes hold: AnaghaRaoAMD 6, yraparti 1, Copilot 4 (ran, read) • Copilot's: the author or a maintainer resolves • wait: both stream threads (N16), K/V copy, sdpa/__init__.py:146
AnaghaRaoAMD (11-03) bind_unified.py:396: K/V copied every launch	🟡 DEFERRED · AnaghaRaoAMD decides • reason: a view loses the zero-filled tail; needs a GPU check: sound • unchanged: Not in this PR, KERNEL_AUTHORING §H (read)
Copilot (11-05) bind_unified.py:345, bind_shared.py:326: in-place Q/O not alignment-checked	🟢 FIXED 0d916301 • gfx1151, 2-byte offset: 3 of 3 cases match; 7ad42faf wrong O (ran) • new test catches the BSHD check only; ragged half: N18 (ran)
Copilot (11-05) bind_unified.py:186: non-positive scale admitted	🟢 FIXED 0d916301 • _scale_reject declines ≤ 0 and non-finite at every entry point (read) • its 5 cases catch the removal, but only with torch: N18 (ran)
Copilot (11-05) bind_unified.py:445, bind_dense.py:348: stream discarded; our N16	🟠 OPEN, narrowed 0d916301 • fixed: a foreign stream now raises; each new test catches its mutant (ran) • left: refuses the named current stream (graph capture's form); ignores unknown keywords (ran) • fix: one bind_shared helper over resolve_stream; keyword-only launch(*, stream=None)
Copilot (11-05) sdpa/__init__.py:146: raw attention DispatchResults	🟡 OPEN · cgmillette decides • latent: no caller binds these results (read) • author offers: SDPA adapter, or document and raise (06:37Z) • our pick: document it and export the SDPA binders, rather than raise
Our N17 (11-05): each tensor gets its own dtype rule	🟡 OPEN · cgmillette decides • author asks for a design nod (06:37Z) • our nod: one check in Operand; _DensePlan's fp32 stats path stays (read) • unchanged: int64 seq_len_q, fp16 stats still bind (ran)
Our N11: body omits new public API; out of date	🟡 OPEN • left: dense_scale_range unlisted; GPU box still unchecked (read) • last gfx942/gfx950 run at 6daf7e6b, before the 15024 merge (read) • new: body lacks the alignment, scale and stream rules (read)
cgmillette (11-05): approval	🟢 APPROVED 7ad42faf • older head; still counts on develop • predates 0d916301's three fixes, reviewed above
Bots / CI	🟢 Green • 44 ✓, 53 skipped, 0 ✗, none running (10:36Z) • gfx94X rocKE shards 100/100 ✓ ×2; SDPA binding tests skip, no torch (read) • merge: BLOCKED on 15 open threads
Our checks	🟢 GREEN • vs e61df496: clean, 41 behind; provider tree = checks' merge (ran) • build + ctest 51 ✓ / 1 skip; byte identity, parity, golden ×4; dg 0 changed (ran) • tests 4177 ✓; installed 3864 ✓; with torch: SDPA + attention 6654 ✓ (ran) • compile 29 ✓, 0 spills; rules 1–3 clean (ran)
Code quality	🟠 Should-fix • N16 half fixed: the named current stream is refused (ran) • CI skips the unified scale tests; ragged alignment untested: N18 (ran) • in-place rule, finite check, test helpers written twice; docstring blames gfx950 too (read)

#	Sev	file:line	Finding
N18	🟠	tests/sdpa/test_unified_sdpa_binding.py:270, dispatch/sdpa/bind_shared.py:348	Unified scale decline untested in CI; ragged alignment check untested • impact: CI's view passes 5497 of 5497 with the decline removed; nothing fails without the ragged check (ran) • fix: scale tests into test_unified_sdpa_wiring.py, beside dense's; a ragged twin of the misaligned test

Utility: 🟢 Positive — three defects fixed; stream check and CI tests need one more pass

---

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

Follow-up review: first pass by a local qwen3.6 model, verified by the orchestrator (Claude Opus 5.5). Checks: delta since 7ad42faf, merge with develop e61df496, the saved alignment (GPU) and stream probes at both heads, mutants of the new tests with and without torch, a side-stream launch on a GPU, SDPA and attention tests with torch, CI at head, code quality. Flags whether a closer look is needed; not an approval or rejection.

[CongMa13]

I reviewed this PR again. The following review threads are still unresolved, or comments have no author reply:

• feat(hip-kernel-provider): route SDPA to unified and dense attention kernels #13115 (comment)
• feat(hip-kernel-provider): route SDPA to unified and dense attention kernels #13115 (comment)
• feat(hip-kernel-provider): route SDPA to unified and dense attention kernels #13115 (comment)

For reassessment, the PR author should reply to top-level feedback or state that thread feedback is addressed. Before approval, top-level comments need explicit resolution from their author or that author's approval of the PR.

[yraparti]

rocKE checklist re-review of the fixes since 6daf7e6b96b0

ROCm/rocm-libraries#13115 · 0d916301ba32 · model gemini-4-argon, fallback gpt-5.5

PASS · graded 238/238 · BLOCKER 0, HIGH 0, MEDIUM 0, LOW 0, NEEDS-INFO 12

The diff was truncated before grading. Items that needed a missing hunk are NEEDS-INFO.

Since the last review (6daf7e6b96b0)

Resolved 0 · still open 0 · new 0

Findings

• No FAIL rows.

Needs info

• §A.1 Unresolved since previous review: dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6185 and :6784 modify _build_gfx942_4warp_gqa_lean and build_gfx942_4warp_gqa without a C++ update in the PR diff, and known_gaps.md and cpp/instances/gfx942/ are not in the PR diff to verify whether the 4-warp GQA body has a C++ twin or is documented as Python-only.
• §D.14 The kernarg layout change for tiled attention (head counts and sliding window becoming kernargs) bumps ATTENTION_ABI_VERSION to v2 at dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention/common.py:80 and strided_decode's abi_version to v2 at dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention/strided_decode.py:222. — no diff line supports this verdict
• §E.10 Not resolved since the previous review: the PR description and diff (which touches shared files such as dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:894 and dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8) still do not provide a merge-simulation record listing open PRs touching the same tables/tests/goldens and suite reruns.
• §F.6 Still unresolved since previous review: local pre-commit execution status or CI lint check output is not visible in the diff.
• §F.7 Still unresolved since previous review: Although the new commits add review-response fixes (e.g. stream rejection in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:364, 16-byte alignment check in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:340, and positive scale check in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:166), the GitHub review-thread list and any deferred-item ticket IDs are not included in the provided PR context.
• §G.3 In the updated head, attention_d128_swa.json is updated across all four LLVM flavors (dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8) and attention_d256.json is updated (dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8), but the tail of attention_d256.json at line 98 is still truncated in the diff and local non-GPU gate sweep execution results cannot be verified from the diff alone.
• §G.19 The new commits and current diff provide explicit fixtures and inline checks for the named failures: misaligned contiguous BSHD views at dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_binding.py:122 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_binding.py:150, bad per-batch sequence lengths at dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_binding.py:168 and dnn-providers/hip-kernel-provider/rocke/library/tests/test_sdpa_seqlen.py:172, non-int32 ragged offsets at dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_binding.py:241, and an out-of-range block-table physical id…
• §G.24 The diff updates golden hashes in dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8 and dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8 (including attention_d256/gfx942/4warp_gqa_d192 at line 16), while the SDPA unified gate admits UNIFIED_HEAD_SIZES (64, 128, 256); confirmation of GPU validation coverage for every moved golden configuration (especially D192 4warp_gqa) is needed.

Section notes

• §A: Re-evaluated against current head 0d91630. Item 1 remains NEEDS-INFO because the PR modifies _build_gfx942_4warp_gqa_lean and build_gfx942_4warp_gqa in attention_tiled_2d.py without touching C++ or known_gaps.md in the PR diff.
• §B: Re-reviewed against head 0d91630. New commits further strengthen trust-boundary rejections (ragged offset dtype check, scale positivity check, non-current stream rejection, and 16-byte alignment check on packed views).
• §E: Re-review of head 0d91630. The full PR diff is now clean of non-rocKE project changes (resolving item 9) and updated golden hashes are present for the gfx942 4-warp fix (resolving item 14).
• §G: Re-review of items 18-34 against head 0d91630.
• §H: Trace of device-side bounds/masks touched in this PR: 1. O epilogue store in build_gfx942_4warp_gqa (attention_tiled_2d.py:6187-6191, 6786-6790): extent source is runtime qlen (and context_off + q_inseq >= 0 for rows with no visible key); visited range is q_inseq = qbase + wq + c across the CTA's Q tile; guarded by scf_if(q_inseq < qlen) and select(context_off + q_inseq < 0, 0.0, ...), agreeing with the valid row range. 2. k_cache/v_cache and block_table in bind_unified.py:376-389: extent source is slots = batch * pages where pages = max(1, ceil(s_kv / block)), cu_seqlens_q from q_len, and seqused_k from kv_len; validated before launch via validate_tuning_attention_tensors (bind_unified.py:466-467).
• §J: Shapes: | Equation / Op | Operands | Inner Dims | Result vs Declared | |---|---|---|---| | (none in KERNEL_AUTHORING.md) | — | — | — | Symbols: | Symbol | Defined at | Domain | |---|---|---| | AttentionRequest | dispatch/attention/common.py | Attention dispatch request | | SdpaRequest | dispatch/sdpa/common.py | hipDNN SDPA request contract | | ATTENTION_ROUTE_REGISTRY | dispatch/attention/__init__.py | Route candidate registry | | UNIFIED_ALGORITHMS | dispatch/attention/__init__.py:85 | frozenset of unified algorithm names | | DENSE_ALGORITHMS | dispatch/attention/candidate.py:51 | frozenset of dense algorithm names | | SDPA_<FAMILY>_FEATURES | dispatch/sdpa/bind_{unified,dense}.py | frozenset of carried SDPA feature tags | Novelty: | Claim | Symbol / Path | Status | |---|---|---| | Routed SDPA family & checklist H | library/dispatch/sdpa/, bind_unified.py, bind_dense.py | Added in this PR | | Unified route algorithm set |…

Graded with the rocKE PR checklist, one model pass per section. FAIL rows that sit on a diff line are inline comments.

§A graded 14/14

#	Verdict	Evidence
1	NEEDS-INFO	Unresolved since previous review: dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6185 and :6784 modify _build_gfx942_4warp_gqa_lean and build_gfx942_4warp_gqa without a C++ update in the PR diff, and known_gaps.md and cpp/instances/gfx942/ are not in the PR diff to verify whether the 4-warp GQA body has a C++ twin or is documented as Python-only.
2	N/A	The PR does not modify any core lowering ops or IR op-handlers under core/lower_* or core/ir.
3	PASS	New dispatch and routing logic is confined to dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:1 and dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:1 (plus public exports in…
4	N/A	No C++ kernel twin or C++ IR emitter code is modified in this PR.
5	PASS	Both dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8 and dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8 re-bless only the…
6	N/A	No C++ twin is touched in the PR diff and the PR does not claim cross-engine parity from a golden update.
7	PASS	The no-key output zeroing fix in dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6185 and :6784 applies only to the 4warp_gqa builders, and the golden diffs in…
8	N/A	No rocKE C++ files are modified or added in the PR diff.
9	N/A	No shared op-handler or lowerer arch divergence fix is included in this PR.
10	N/A	No new kernel IR feature flag is introduced in this PR; it adds host-side SDPA dispatch bindings and fixes a bug in existing…
11	N/A	No new IR op is added in this PR.
12	N/A	core/backend.py and ROCKE_BACKEND handling are not touched in the PR diff.
13	N/A	LLVM .ll emission and debug-location handling are not modified in this PR.
14	N/A	No helper files under helpers/ or platform/cpp/helpers/ are modified in the PR diff.

§B graded 6/6

#	Verdict	Evidence
1	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:176-191 classifies implementation-shortcut declines with NOT_YET_IMPLEMENTED: separately from ABI/kernel limits, and…
2	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:211-214 probes each dense candidate with opt_in_probe, checks candidate.admits(probe), and builds the candidate spec only after admits succeeds…
3	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:240-250 declines requests when no unified candidate covers them, and dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:459-461 and…
4	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/hipdnn.py:220-222 rejects a node missing its 'type' key instead of defaulting; dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:292-297 rejects ragged offsets whose dtype…
5	PASS	dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:49-57 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_numeric.py:43-51 run GPU numeric verification over every case admitted by sdpa_unified_admits…
6	N/A	The diff adds SDPA dispatch admission/binding logic and kernel bug fixes without adding or modifying an is_valid_spec or…

§C graded 10/10

#	Verdict	Evidence
1	PASS	Shared predicates, decline reasons, and translation logic across unified and dense SDPA bindings are factored into dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:46-246, and torch_dtype / device_tensor are consolidated in…
2	PASS	The SDPA bindings route to existing public entrypoints (bind_dense_attention_torch in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:345-347 and run_unified_attention_torch in…
3	N/A	No new kernel builder functions or kernel_name() definitions are introduced in this PR.
4	PASS	Arch-independent SDPA request/binding logic lives in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:1-348, while gfx942-specific kernel fixes stay in…
5	PASS	All touched files in the PR diff are directly scoped to SDPA dispatch/binding (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/__init__.py:1-152), SDPA tests (dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:894-895…
6	PASS	Shared SDPA request, band, decline-table, and operand abstractions are cleanly factored into dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:46-348, and single-source public attention-layer accessors (attention_band…
7	N/A	This PR is a feature routing and bug-fix change rather than a pure no-op cleanup.
8	PASS	After rebasing onto develop (which included #14929's clamp fixes), the PR reuses existing helpers including opt_in_probe, empty_workspace, and resolve_stream (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:31-32)…
9	PASS	The previous NEEDS-INFO is resolved in the current diff: the no_key zeroing fix for rows where qlen > klen in bottom-right causal mode is applied across both the 4-warp lean and 4-warp GQA bodies in…
10	PASS	Every feature in SDPA_FEATURES is explicitly partitioned into carried vs declined sets in both dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:85-131 and…

§D graded 21/21

#	Verdict	Evidence
1	N/A	This PR does not add or modify any KernelSpec knobs or environment variable reads.
2	PASS	Coverage limits and algorithm constants are imported from kernel/attention modules rather than transcribed (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:39 imports UNIFIED_ALGORITHMS and UNIFIED_HEAD_SIZES…
3	PASS	All candidate admission checks go through candidate.admits(probe) or ATTENTION_ROUTE_REGISTRY.supported(...) and never call _supports directly (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:212…
4	PASS	Selection and admission are pure functions of SdpaRequest taking req.arch directly without any device probes (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:74…
5	N/A	This PR does not introduce or modify any KernelSpec identity, spec_hash, compile_key, or kernel_name() methods.
6	N/A	This PR does not introduce or modify cohort routing, spec-builders, or kernel cache keys.
7	N/A	This PR does not add hand-pinned heuristic overrides.
8	N/A	This PR does not define or modify any kernel cohort predicates.
9	N/A	No new KernelCandidate or architecture is registered; the PR routes SdpaRequest onto existing ATTENTION_ROUTE_REGISTRY candidates.
10	PASS	Every SdpaRequest field and feature is translated into AttentionRequest and routed to the unified and dense bindings with dispatcher-level wiring, entry-point, and numeric tests (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:46…
11	PASS	Existing opt-in dense candidates remain opt-in and are queried explicitly via opt_in_probe(attn, candidate) in the SDPA dense gate (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:211).
12	PASS	Routing priority order is preserved and tested: dense SDPA asserts the first gfx950 variant in registry order is chosen at dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_wiring.py:312, and unified SDPA asserts returned candidates are sorted…
13	PASS	Pinned SDPA requests are explicitly declined instead of falling back to auto at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:233, and tested for known and unknown pins at…
14	NEEDS-INFO	The kernarg layout change for tiled attention (head counts and sliding window becoming kernargs) bumps ATTENTION_ABI_VERSION to v2 at dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention/common.py:80 and strided_decode's abi_version to v2 at dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention/strided_decode.py:222. — no diff line supports this verdict
15	N/A	No new KernelCandidate or arch registration module is added; SDPA routes through existing attention candidates.
16	PASS	registered_sdpa_combos, sdpa_sweep_space, and dispatch_sdpa_all short-circuit on _request_errors(req) before delegating at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/init.py:123, with coverage in…
17	N/A	No new KernelCandidate is registered in this PR, so per-candidate CPU routing and family fixture invariants do not apply.
18	N/A	This PR does not change registry ratchets, route-only exemptions, or candidate binding requirements.
19	PASS	The unified SDPA binding accurately documents and uses run_unified_attention_torch with no tuning spec (backend="auto") at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:467, while dense SDPA launches through…
20	N/A	The selection/routing changes are confined to the Python library dispatch layer (library/dispatch/), which has no C++ twin.
21	PASS	Band and scale mismatches are rejected in the chooser/admission gates rather than hacking the emitter, at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:152 and…

§E graded 17/17

#	Verdict	Evidence
1	PASS	Platform changes remain strictly one-way with no library imports, e.g. dnn-providers/hip-kernel-provider/rocke/platform/python/rocke/runtime/torch_interop.py:96 and dnn-providers/hip-kernel-provider/rocke/platform/tests/test_rocke.py:7656.
2	PASS	Library SDPA bindings use rocke.runtime.torch_interop helpers rather than importing torch directly or querying devices with torch, e.g. dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:32 and…
3	PASS	Torch remains optional: tests gate it via pytest.importorskip in dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_binding.py:21 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_binding.py:22, while runtime…
4	PASS	Imports inside library/ flow downward from benchmarks to dispatch to kernels, e.g. dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/attention_combo_sweep.py:44 imports dispatch.attention.common and…
5	PASS	Installed CTest entries use relative paths via ${ROCKE_TEST_CTEST_REL} in dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:894 and dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:1031.
6	PASS	Emitter changes in dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6187 and dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6786 add explicit IR zero-select guards for masked rows rather than…
7	PASS	Authoring and routing documentation is updated alongside the code in dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:321 and dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:591.
8	PASS	Code comments and docstrings use descriptive references without Jira IDs, PR numbers, or user handles, e.g. dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:166 and dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:591.
9	PASS	Resolved since previous review: the current PR diff against develop is cleanly scoped to SDPA routing and its gfx942 4-warp fix (e.g. dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:1 and…
10	NEEDS-INFO	Not resolved since the previous review: the PR description and diff (which touches shared files such as dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:894 and dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8) still do not provide a merge-simulation record listing open PRs touching the same tables/tests/goldens and suite reruns.
11	PASS	Committed files consist only of deliverable source, tests, CMake test wiring, and required authoring documentation, e.g. dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:591, dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:1, and…
12	PASS	Both gfx942 and gfx950 are covered in the SDPA routing gates and test suites, e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_wiring.py:39 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_wiring.py:43.
13	PASS	The rename of _problem, _attention_band, and _validate_lengths to public helpers updates definitions and callers across the tree, e.g. dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention/common.py:758…
14	PASS	Resolved since previous review: the rebase onto develop cleanly preserves upstream code while updating only the intended lines and golden hashes in dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6185 and…
15	PASS	New commits explicitly validate the stream parameter in launch(*, stream=None, **_kw) rather than ignoring it in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:350 and…
16	N/A	This item is a meta-instruction for reviewer grading rather than a property of the PR code.
17	PASS	No absolute or relative performance claims appear in the PR description, comments, or documentation such as dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:591.

§F graded 11/11

#	Verdict	Evidence
1	PASS	Branch users/tlakshma/hack/sdpa/st2-unified matches the recognized users/<user>/... prefix, and PR changes are in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/__init__.py:1.
2	PASS	title feat(hip-kernel-provider): route SDPA to unified and dense attention kernels matches type(scope): summary and is 76 characters
3	PASS	description is at least 30 characters and contains an issue reference
4	PASS	test file dnn-providers/hip-kernel-provider/rocke/library/tests/dispatch/attention/test_additive_registration.py is in the diff
5	PASS	All changed files in the diff (such as dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:1 and dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:1) contain only code, docs, CMake, and test assets; no keys…
6	NEEDS-INFO	Still unresolved since previous review: local pre-commit execution status or CI lint check output is not visible in the diff.
7	NEEDS-INFO	Still unresolved since previous review: Although the new commits add review-response fixes (e.g. stream rejection in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:364, 16-byte alignment check in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:340, and positive scale check in…
8	N/A	The PR targets Base: develop rather than a stacked feature branch, so the stacked-PR bot-skipping case does not apply.
9	N/A	There is no indication that the PR was artificially split or trimmed for a size cap (in fact, #14624 was folded into this PR).
10	PASS	The added and updated tests are substantive wiring, CPU data-movement, case-table, and GPU numeric tests rather than throwaway stubs, e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_binding.py:82 and…
11	PASS	The PR metadata header in this prompt shows the updated title (feat(hip-kernel-provider): route SDPA to unified and dense attention kernels) and full multi-section body (covering dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:1 and…

§G graded 34/34

#	Verdict	Evidence
1	PASS	Correctness is verified against an independent mathematical reference oracle (sdpa_reference) rather than relying solely on IR byte-identity (dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:61 and…
2	PASS	On-GPU numeric tests against the reference oracle are included and enrolled in the GPU test suite (dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:1031 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:61).
3	NEEDS-INFO	In the updated head, attention_d128_swa.json is updated across all four LLVM flavors (dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8) and attention_d256.json is updated (dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8), but the tail of attention_d256.json at line 98 is…
4	PASS	New admission, binding, alignment, stream-rejection, scale-rejection, and length/offset-validation rules have falsifiable unit tests (dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_binding.py:122…
5	PASS	In test_an_out_of_range_block_table_physical_id_is_rejected_before_launch (dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:83), the docstring explicitly documents why the test enters at validate_tuning_attention_tensors…
6	PASS	Prose claims in KERNEL_AUTHORING.md, docstrings, and code comments (including the new alignment, scale, and stream rejection comments in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:322…
7	PASS	KERNEL_AUTHORING.md specifies the reference oracle and four test layers as a first-class gate (dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:624), and the CPU and GPU binding tests validate against sdpa_reference…
8	PASS	Bindings and test harnesses perform strict dtype checks on operands, lengths, and ragged offsets (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:293, dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:267…
9	N/A	This PR does not introduce or claim any LDS size or occupancy numbers.
10	N/A	This PR is a functional routing and bug-fix change, not a performance optimization PR.
11	PASS	Kernel reachability through the SDPA routing gates and family entry points is verified in dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_wiring.py:474…
12	N/A	The changes in this PR are in Python-only dispatch layers and a gfx942 4-warp GQA kernel builder…
13	N/A	No new LLVM intrinsic overload classes or mangled symbols are added or recommended.
14	N/A	No hardware bottleneck classification claims are made in this PR.
15	N/A	The PR diff is provided directly in the review prompt rather than inspected from a local git checkout.
16	PASS	Integration with existing attention dispatch symbols (ATTENTION_ROUTE_REGISTRY, UNIFIED_ALGORITHMS, DENSE_ALGORITHMS, attention_band, attention_problem, dense_scale_range, opt_in_probe) is verified by wiring and coverage tests…
17	N/A	No standalone CLI or benchmark scripts are shipped in this PR; new files are library modules and pytest suites enrolled in CMake.
18	N/A	This PR does not add statistical runtime guards or back-pressure detectors of the kind this item targets.
19	NEEDS-INFO	The new commits and current diff provide explicit fixtures and inline checks for the named failures: misaligned contiguous BSHD views at dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_binding.py:122 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_binding.py:150, bad per-batch sequence lengths at…
20	N/A	This PR does not add a claim-ban detector or regex sweep.
21	N/A	This PR does not introduce performance measurement or profiling deliverable code.
22	PASS	The SDPA numeric GPU tests are excluded from the host SDPA pytest shard at dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:894 and dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:895 and explicitly enrolled in the GPU CTest lane at…
23	N/A	This PR does not make speedup or routing A/B performance claims.
24	NEEDS-INFO	The diff updates golden hashes in dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8 and dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8 (including attention_d256/gfx942/4warp_gqa_d192 at line 16), while the SDPA unified gate admits UNIFIED_HEAD_SIZES (64, 128, 256)…
25	N/A	This PR is a functional routing and bounds/keyless-row correctness change, not a codegen micro-optimization report.
26	N/A	This PR does not make campaign-branch benchmark speedup claims.
27	N/A	This PR does not ship a performance knob or A/B toggle.
28	N/A	This PR does not present counterfactual performance tables.
29	N/A	This PR does not swap guarded math primitives for fast primitives.
30	N/A	This PR makes no register-spill or scratch-reduction claims.
31	N/A	This PR does not claim a performance win over a baseline defect.
32	N/A	The PR makes no timing, throughput, spill, or other measured performance-number claim requiring reproducible provenance.
33	N/A	The rocKE diff exposes and reuses attention routing helpers and adds SDPA gates, but does not modify an existing scalar…
34	PASS	The optional stats/LSE path is tested on padded/keyless rows, asserting O is zero and stats is -inf at dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_binding.py:157 and…

§H graded 5/5

#	Verdict	Evidence
1	PASS	sdpa_unified_bind allocates k_cache/v_cache sized to slots = batch * self.pages at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:376-383 and validates the launch tensors against plan.problem via…
2	PASS	_LaunchPlan builds an identity block_table in range(slots) at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:385-387, validates physical block IDs before launch via validate_tuning_attention_tensors at…
3	N/A	After rebasing onto develop (which already merged #14929), this PR's diff no longer adds or modifies any device-side div/mod or…
4	PASS	In build_gfx942_4warp_gqa, rows whose diagonal is left of key 0 (context_off + q_inseq < 0, where every score in the loop was masked) now explicitly select 0.0 for O before storing at…
5	NEEDS-INFO	Resolved since previous review: seq_len_q/seq_len_kv are host-validated via _host_lengths and validate_lengths at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:267-289 and dnn-providers/hip-kernel-provider/rocke/library/kernels/common/_sdpa_seqlen.py:45, operand shapes/strides/ragged offsets (including offset dtype) and 16-byte pointer alignment are validated in Operand at…

§I graded 6/6

#	Verdict	Evidence
1	N/A	This PR wires hipDNN SDPA forward requests to existing attention candidates and fixes gfx942 kernel edge cases without introducing…
2	PASS	Non-contiguous or unaligned (<16-byte) layouts are repaired via workspace copy buffers (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:276-284, dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:340-348), and…
3	PASS	Curated request and kernel case tables (dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/cases.py:1361-1439) are combined with generated parameter sweeps and explicit algorithm-partition coverage checks…
4	N/A	This PR does not add or modify performance routing comparison tables or tuning reports.
5	N/A	This PR does not introduce baked tuning search results or static tuning snapshots.
6	N/A	This PR does not add any new DSL ops or hardware-instruction-named primitives.

§J graded 7/7

#	Verdict	Evidence
1	N/A	The only Markdown file in the PR (dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md) contains no mathematical…
2	N/A	The KERNEL_AUTHORING.md diff introduces no mathematical formula symbols, decay/log-decay domains, or log bases.
3	PASS	dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:591-642 links to existing sections [E](#e-new-operator-with-a-hipdnn-graph-node), [F](#f-new-operand-or-output-on-an-existing-operator), and [A](#a-new-kernel-family) while stating only the delta for…
4	PASS	The novelty claims in dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:321 and dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:638-641 name library/dispatch/sdpa/, bind_unified.py, bind_dense.py…
5	PASS	dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md makes no scaling/complexity claims, and the O(s_q) interval comparison claim in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:96-97 states the aggregate complexity across…
6	N/A	The KERNEL_AUTHORING.md diff does not introduce any model configuration tables or tuned band-edge constants.
7	N/A	This PR edits the existing dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md and does not add a new design or…

§K graded 13/13

#	Verdict	Evidence
1	N/A	This PR adds no multi-configuration candidate (SDPA_REGISTRY remains empty and SDPA delegates to existing attention candidates…
2	N/A	This PR does not add or modify any tunable kernel spec fields, KnobAxis declarations, or field classifications.
3	N/A	This PR does not change config_key payloads, canonical-knob rules, declared defaults, or knob meanings (unified_rules.py only…
4	N/A	This PR does not modify KnobSpace.defaults() or KnobSpace.recorded() in any KnobSpace subclass.
5	N/A	This PR does not touch knob drop/refuse rules, inert(), base_value(), or KNOWN_WRONG_KNOBS.
6	N/A	This PR does not touch KnobSpace.validate or library/dispatch/attention/tuning_specs.py.
7	N/A	This PR does not modify KnobSpace.base(req) or _tuning_problem (it only renames _problem to attention_problem in…
8	N/A	This PR does not modify normalize_knobs or request-boundary knob scalar coercion.
9	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:220 (value_rejects) explicitly declines any pinned algorithm, spec_id, tuning_id, or tuning_knobs rather than adopting canonicalized ids or falling back to auto, tested in…
10	N/A	This PR does not modify KnobSpace.find or tuning_id resolution logic.
11	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/__init__.py:119 (registered_sdpa_combos, sdpa_sweep_space, dispatch_sdpa_all) delegates to SDPA_REGISTRY and registered_attention_combos / dispatch_attention_all filtered by _serves(req)…
12	N/A	This PR does not reorder any production stack, change accept_default, or edit a tuning variant's base spec.
13	PASS	The benchmark changes in dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/attention_combo_sweep.py:44 and dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx950/attention/decode/decode_dispatch_diagnose.py:13 only rename _problem to…

§L graded 10/10

#	Verdict	Evidence
1	N/A	This diff does not add or modify a rocKE gfx target in arch specs or ISA backend registries.
2	N/A	This diff does not add or modify an MMA op_id, MMA fragment-info row, or MMA lowering layout map.
3	N/A	This diff does not add or change MMA atom selection, MMA catalog fields, lane maps, or intrinsic-name emission.
4	N/A	This diff does not introduce or change wave-size handling or lane math derived from wave size.
5	PASS	The arch-specific codegen change is confined to the gfx942 4-warp kernel and masks fully masked rows without leaking gfx950 recipes at dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6187.
6	N/A	This diff does not move pipeline vocabulary onto ArchTarget/core arch data, mutate ArchTarget from policies, or import CK Tile…
7	N/A	This diff only routes or declines SDPA dtype features and does not change rocKE logical dtype, bit-packing, scalar lowering, or…
8	N/A	This diff does not touch scaled-MMA scale contracts, retired scaled WMMA ids, or MfmaAtom.emit_scaled lowering.
9	N/A	This diff does not add a new rocKE architecture target.
10	N/A	This diff does not touch intrinsic_arch_domain..json, gen_arch_domain.py, or LLVM_FLAVORS.

§M graded 10/10

#	Verdict	Evidence
1	PASS	Resolved in the current diff: dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6187 and dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6786 compute no_key with b.cmp_lt and select the…
2	N/A	The PR diff does not touch legalize_for_target or target legalization logic.
3	N/A	The PR diff does not change print_ir, ck.dsl.ir/v1 serialization, wire versions, attrs, SSA ids, or parameter ordering.
4	N/A	The PR diff does not change portable IR schemas, importers, magic division constants, recipe VM code, or recorder coverage.
5	N/A	The PR diff does not touch AsyncTileLoader, async DRAM-to-LDS loading, LDS async layouts, or VMEM waits.
6	PASS	Resolved in the current diff: the kernel changes in dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6189 and dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6788 only guard b.global_store…
7	N/A	The PR diff does not add or modify buffer-resource word 3 handling or TemporalHint.STREAMING lowering.
8	N/A	The PR diff does not add LLVM intrinsics or change backend selection precedence.
9	N/A	The PR diff does not touch public C entry points, cpp/include/rocke/ headers, or C++ codegen launch behavior.
10	N/A	The PR diff does not emit !dbg metadata or change debug-info consumers.

§N graded 18/18

#	Verdict	Evidence
1	N/A	The PR diff against develop does not add or modify any files under platform/python/rocke/helpers/.
2	N/A	In the current head diff against develop, the previously flagged Q-addressing clamp in attention_tiled_2d.py was dropped (reused…
3	N/A	The diff does not touch SoftwarePipeline, sync_before_issue, or overlap_vmcnt behavior.
4	N/A	The diff does not add or modify fused kernels, EpilogueOp, fusion schedules, or fusion validation paths.
5	N/A	The diff does not add or modify a public API under rocke.helpers.tiling.
6	N/A	The diff does not add or modify front-door or toolbox tiling helpers or LayoutStyle classes.
7	N/A	The diff does not touch tiling layout overrides, accumulator c_desc derivation, or MMA soundness gates.
8	N/A	The diff does not touch helpers/tiling LDS allocation, lds_store/lds_read, contraction padding, or wide-load tiling emit rules.
9	N/A	The diff does not use or modify N-D tiling axis roles, at_index/squeeze behavior, FlowStage labeling, or AB-swap routing.
10	N/A	The diff does not modify tiling emit, LDS staging, or MMA issue code in helpers/tiling.
11	N/A	The diff does not touch tiling MMA traits, TileMma target selection, lds_swizzle policy, or derive_sched_group_counts.
12	N/A	The PR diff makes no helpers/tiling LDS bank-conflict or coalescing optimization claim requiring simulator, counter, or ASM-backed…
13	N/A	The diff does not construct WarpDistributionEncoding, change make_tile_desc factoring, re-export RegisterMapper, or alter tiling…
14	N/A	The diff does not use or modify helpers/tiling verbs or add a new helpers/tiling LDS pipeline variant.
15	N/A	The diff does not touch helpers/tiling descriptor allocation or transformed-range bounds checks after swizzle, pad, or XOR.
16	N/A	The diff does not use or modify optimize_layout, recommend(), or evaluate_transform ranking behavior.
17	N/A	The diff does not modify tiling_recorder.py, round-trip analysis, or tiling visualization emission behavior.
18	N/A	The diff does not touch InterleavedStyle, helpers/tiling/kernels demos, or tiling numeric gates for interleaved layouts.

§O graded 16/16

#	Verdict	Evidence
1	N/A	The PR does not add or modify IR address spaces, required attributes, or core/verify.py.
2	N/A	The PR does not modify scf.for structure or terminator emission in core IR/lowerers.
3	N/A	The PR does not touch IRBuilder, intrinsic declaration tables, or metadata slot numbering.
4	N/A	The PR does not modify IR types or SmemType serialization/parsing.
5	N/A	The PR does not modify op purity tables or core/passes.py / cpp/core/passes.cpp.
6	N/A	The PR does not modify constant folding in Python or C++ passes.
7	N/A	The PR does not modify kernel attributes, codegen policy, or cluster dimension checks.
8	N/A	The PR does not set or modify scheduler strategies or CodegenPolicy.
9	N/A	The PR does not modify waitcnt encoding or barrier lowering.
10	N/A	The PR does not modify lowerer attribute validation or pointer parameter lowering.
11	PASS	dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_numeric.py:30 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:36 resolve the device architecture via…
12	N/A	The PR does not add hardcoded width/size caps in core/codegen or silent arch or "gfx950" defaults.
13	N/A	The PR does not touch any public C structs or enums under cpp/include/rocke/.
14	N/A	The PR does not touch C++ sources or add C++ compiler flags in platform/CMakeLists.txt.
15	N/A	The PR does not modify C++ engine fallback mechanisms or lowerer entry points.
16	N/A	The PR does not touch the C++ codegen library or gfx1250 builders/lowerers.

§P graded 14/14

#	Verdict	Evidence
1	N/A	The PR diff does not add or modify platform instance builder signatures or spec REQUIRED_FIELDS.
2	N/A	The PR diff does not touch platform kernarg signatures, manifest args_signature, runtime argument packing, or GEMM/conv ABI…
3	N/A	The PR diff does not touch GEMM, implicit-GEMM, or block-scaled GEMM launch geometry.
4	N/A	The PR diff does not modify LDS allocation gates or C++ validators.
5	N/A	The PR diff does not touch WMMA/RDNA GEMM, GemmMultiD, or GemmMultiAbd admission paths.
6	N/A	The PR diff does not touch Stream-K, MoE sorting, fused MoE, or split-K caller-zeroed workspaces in rocKE.
7	N/A	The PR diff does not touch Stream-K rejection rules, persistent GEMM, GroupedGemmLauncher, or SUPPORT_MATRIX.md.
8	N/A	The PR diff does not touch quantization helpers, MX GEMM, or MX scale decoding.
9	N/A	The PR diff does not add or modify numeric emulation or low-precision decode paths requiring edge-case emulation tests.
10	N/A	The PR diff does not add a manifest kind or touch helpers/manifest.py, run_manifest, or manifest allocate/verify consumers.
11	PASS	The SDPA bindings allocate and retain launch buffers once in the bound plan, launch on resolve_stream(0, device=device), and reject non-current streams to prevent cross-stream races on preallocated workspaces: see…
12	N/A	The PR diff does not touch COMGR or HIP library loading logic.
13	N/A	The PR diff adds SDPA routing and binding logic rather than refactoring shared host-side grid, tile-count, workspace, cache-key…
14	N/A	The PR diff does not touch examples/, helpers/ placement, or add new modules under platform instances/common family glue.

§Q graded 7/7

#	Verdict	Evidence
1	PASS	In the full PR diff, the new SDPA CPU tests under tests/library/tests/sdpa run via the existing rocke_library_sdpa_pytest CTest target, with the two new GPU numeric files excluded from the CPU suite at…
2	N/A	This PR routes SDPA requests onto existing attention candidates and fixes an existing gfx942 4-warp attention kernel without…
3	PASS	The representative IR golden tripwires are updated for the modified gfx942 4-warp attention kernels at dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8 and…
4	PASS	The representative golden files are updated across all four LLVM flavor blocks in dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8 through…
5	NEEDS-INFO	All new tests in the full PR diff are portable: GPU numeric suites dynamically query the active device arch and skip cleanly without torch or a supported GPU (dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:17, dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:33…
6	N/A	This PR does not refactor any platform validation gate or differential harness.
7	N/A	This PR does not modify engine rejection contracts or ROCKE_BACKEND / ROCKE_CPP_STRICT fallback behavior.

§R graded 11/11

#	Verdict	Evidence
1	N/A	The diff routes SDPA over existing attention candidates and adds a correctness zero-select for fully-masked rows in…
2	N/A	Changes since the previous review dropped the KV page clamp and Q-row hoist hunks after rebasing onto develop (#14929), leaving…
3	N/A	Changes since the previous review removed the global-load address clamp and Q-row hoist hunks from…
4	N/A	Changes since the previous review removed the hoisted Q-row / KV clamp hunks and their performance comments from…
5	N/A	Changes since the previous review removed the Q load and block-table/KV page address clamp changes from…
6	N/A	The diff does not change an LDS layout, MMA layout, tiling pipeline identity/relabel/reorder, or bank-conflict lever.
7	N/A	The diff makes no bottleneck claim such as memory-bound, LDS-bound, barrier-bound, or latency-bound.
8	N/A	The diff does not add, remove, or modify waits, barriers, sched_barrier, sched_group_barrier, iglp_opt, or scheduling hints.
9	N/A	The diff is not a performance-regression fix and does not add a shape-trading performance heuristic or dispatch rule.
10	N/A	Changes since the previous review removed the measured-performance comments from…
11	N/A	The documentation added in dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:591 is an SDPA routed-family authoring…

§S graded 8/8

#	Verdict	Evidence
1	NEEDS-INFO	The new commits since 6daf7e6 do not resolve the missing hunk: torch_dtype and device_tensor are added at dnn-providers/hip-kernel-provider/rocke/platform/python/rocke/runtime/torch_interop.py:96 and tested in dnn-providers/hip-kernel-provider/rocke/platform/tests/test_rocke.py:7656, and SDPA tests are registered in dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:892 and…
2	PASS	Existing API behavior changes update their tests in the same PR: from_hipdnn requiring 'type' in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/hipdnn.py:220 is tested at…
3	N/A	The diff does not remove or rename a public API; _problem, _attention_band, and _validate_lengths were private…
4	PASS	dispatch.attention.__all__ is updated in dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention/__init__.py:563 and verified by TestAttentionPublicSurface in…
5	PASS	Tests assert behavioral outputs and exact fail-fast exceptions: torch_dtype and device_tensor in dnn-providers/hip-kernel-provider/rocke/platform/tests/test_rocke.py:7660, non-positive/non-finite unified scale rejection in…
6	PASS	CPU-checkable contracts run in the host lane rocke_library_sdpa_pytest while ignoring the GPU numeric files in dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:892, GPU numeric tests are registered in the GPU test list at…
7	N/A	The public APIs added or modified in this PR are Python dispatch/routing and torch_interop helpers that are Python-only by…
8	NEEDS-INFO	The new commits do not resolve the prior NEEDS-INFO: the PR description is truncated at | Rows that see no key ... and does not show whether the README test entry points (python tests/test_rocke.py and python tests/run_all.py from rocke/platform/) were stated as run.

[ThruptiRajLakshmanaGowda]

@cgmillette @CongMa13: follow-up on the 11-06 10:43Z review, pushed as cf2b94adaae.

• N16 (stream): both launches are keyword-only, launch(*, stream=None), through one helper, bind_shared.launch_stream. The current stream is accepted unset, as 0, or named by handle (graph capture's form). Any other stream raises NotImplementedError, and an unknown keyword raises TypeError.
• N17 (dtype rule): bind_shared.check_dtype is the one rule, applied in Operand and to seq_len_*: every tensor must have its descriptor's dtype. int64 seq_len_q and fp16 stats against an fp32 descriptor are now refused. _DensePlan's fp32 stats path is unchanged.
• N18 (CI tests): the unified scale declines moved into the torch-free test_unified_sdpa_wiring.py (0, -1, NaN, ±inf on gfx942 and gfx950). Removing _scale_reject now fails 10 of them without torch. Added test_misaligned_ragged_q_and_o_fall_back_to_a_workspace_copy, which also checks that the aligned twin is used in place; it fails when the ragged alignment check is removed.
• sdpa/__init__.py:146: your pick. Documented, with no raise. dispatch.sdpa exports sdpa_{unified,dense}_{admits,bind}.
• Code quality:
  • The in-place rule (usable_in_place) and the finite-scale check (nonfinite_scale_reject) are each written once in bind_shared.
  • The strided and misaligned test tensors live once in binding_harness.strided_tensor.
  • The scale docstring now names only the gfx942 4-warp bodies (_online_softmax_rescale), the only ones that mask before scaling.
  • KERNEL_AUTHORING.md §H lists the binding rules.
• N11 (body): adds dense_scale_range and the dispatch.sdpa binder exports under New public API, and the alignment, dtype, scale and stream rules. The GPU box is now checked with this head's runs.

Each new test fails against its fix's mutant. Runs at cf2b94adaae:

• Host: tests/sdpa + tests/dispatch/attention 6509 passed, 257 skipped (GPU suites).
• GPU, one process each (unified and dense numeric, binding, wiring, entry points): gfx942 MI325X 676 passed, gfx950 MI355X 958 passed, 0 failed, 0 skipped.

Threads: the 7 fixed Copilot threads have replies and can be resolved. AnaghaRaoAMD's K/V-copy thread stays deferred ("Not in this PR", §H), with a follow-up reply asking them to resolve it or ask for it here.

[yraparti]

rocKE checklist re-review of the fixes since 0d916301ba32

ROCm/rocm-libraries#13115 · cf2b94adaae7 · model gemini-4-argon, fallback gpt-5.5

PASS · graded 238/238 · BLOCKER 0, HIGH 0, MEDIUM 0, LOW 0, NEEDS-INFO 11

The diff was truncated before grading. Items that needed a missing hunk are NEEDS-INFO.

Since the last review (0d916301ba32)

Resolved 0 · still open 0 · new 0

Findings

• No FAIL rows.

Needs info

• §A.1 Unresolved since previous review: dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6258 and :6857 modify _build_gfx942_4warp_gqa_lean and build_gfx942_4warp_gqa without a C++ update in the PR diff, and known_gaps.md and cpp/instances/gfx942/ are not in the PR diff to verify whether the 4-warp GQA body has a C++ twin or is documented as Python-only.
• §D.14 Previous NEEDS-INFO is not resolved: the visible diff does not include the cache-key use or ATTENTION_ABI_VERSION/abi_version hunks needed to verify compile-key identity and old-id refusal.
• §E.10 Not resolved since the previous review: the PR description and diff (which touches shared files such as dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:932 and dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8) still do not provide a merge-simulation record listing open PRs touching the same tables/tests/goldens and suite reruns.
• §F.6 Still unresolved since previous review: local pre-commit execution status or CI lint check output is not visible in the diff.
• §F.7 Still unresolved since previous review: although the commits include review-response fixes (e.g. launch_stream in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:276, usable_in_place 16-byte alignment check in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:323, and _scale_reject in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:163), the GitHub review-thread list and any deferred-item ticket IDs are not included in the provided PR context.
• §G.3 In the current head, attention_d128_swa.json (dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8) and attention_d256.json (dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8) are updated, but the tail of attention_d256.json at line 98 is still truncated in the diff and local non-GPU gate sweep execution results cannot be verified from the diff alone.
• §G.24 The new commits did not resolve the previous question: the diff updates golden hashes in dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8 and dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8 (including attention_d256/gfx942/4warp_gqa_d192 at dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:16) due to the shared keyless-row O store fix at dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6258 and…
• §H.5 Resolved since previous review: all launch inputs feeding bounds/masks are host-validated with negative tests — seq_len_q/seq_len_kv via _host_lengths and validate_lengths at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:267-290 and dnn-providers/hip-kernel-provider/rocke/library/kernels/common/_sdpa_seqlen.py:45 (tested at dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_binding.py:155-198, 317-330 and dnn-providers/hip-kernel-provider/rocke/library/tests/test_sdpa_seqlen.py:172-196), operand dtypes, shapes, strides, ragged offsets, and 16-byte alignment in check_dtype/usable_in_place/Operand at…

Section notes

• §C: Re-evaluated against cf2b94a; all review-response additions (dtype checks, alignment checks, stream checks) are factored cleanly into bind_shared.py and binding_harness.py.
• §F: Re-evaluated against head cf2b94a; items 6 and 7 remain NEEDS-INFO because pre-commit execution logs and GitHub review-thread states are not present in the diff context.
• §G: Re-grading items 18-34 of section G against head cf2b94a.
• §H: 1) gfx942 4-warp GQA epilogue empty-row mask (no_key = context_off + q_inseq < 0) in attention_tiled_2d.py: extent source is runtime qlen and klen (context_off = klen - qlen); compile-time range visited is the Q tile rows (q_inseq < qlen). When qlen > klen on bottom-right causal, rows with context_off + q_inseq < 0 have all K positions masked out in the main loop; select(no_key, 0.0, acc * recip) zeroes their output before storing, so they agree. 2) sdpa_unified_bind paged cache & block_table in bind_unified.py: extent source is slots = batch * pages where pages = max(1, ceil(s_kv / block)); compile-time/runtime visited range is seqused_k[b] <= s_kv across pages blocks per batch; k_cache/v_cache are zero-initialized to (slots, block, h, d), block_table is identity range(slots) shaped (batch, pages), and validate_tuning_attention_tensors validates shapes/bounds before launch, so they agree. 3) cu_seqlens_q / seqused_k from…
• §J: Shapes: | Equation | Operands | Inner dims | Result vs Declared | |---|---|---|---| | (none in Markdown diff) | - | - | - | Symbols: | Symbol | Defined at | Domain | |---|---|---| | SDPA_<FAMILY>_FEATURES | KERNEL_AUTHORING.md:609 | carried feature set per SDPA launch family | | UNIFIED_ALGORITHMS | KERNEL_AUTHORING.md:648 | unified attention route algorithm names | | DENSE_ALGORITHMS | KERNEL_AUTHORING.md:649 | dense attention route algorithm names | Novelty: | Claim | Symbol / Path | Status | |---|---|---| | Routed SDPA family over dispatch.attention | library/dispatch/sdpa/, bind_unified.py, bind_dense.py | Added in this PR | | Exported route algorithm sets | dispatch.attention.UNIFIED_ALGORITHMS, dispatch.attention.DENSE_ALGORITHMS | Added in this PR |
• §K: Identity table: no tuned knob, axis, default, canonicalization rule, or builder behavior behind a tuning knob is changed by this PR (unified_rules.py only renames _problem to attention_problem).

Graded with the rocKE PR checklist, one model pass per section. FAIL rows that sit on a diff line are inline comments.

§A graded 14/14

#	Verdict	Evidence
1	NEEDS-INFO	Unresolved since previous review: dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6258 and :6857 modify _build_gfx942_4warp_gqa_lean and build_gfx942_4warp_gqa without a C++ update in the PR diff, and known_gaps.md and cpp/instances/gfx942/ are not in the PR diff to verify whether the 4-warp GQA body has a C++ twin or is documented as Python-only.
2	N/A	The PR does not modify any core lowering ops or IR op-handlers under core/lower_* or core/ir.
3	PASS	New dispatch and routing logic is confined to dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:1 and dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:1 (plus public exports in…
4	N/A	No C++ kernel twin or C++ IR emitter code is modified in this PR.
5	PASS	Both dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8 and dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8 re-bless only the…
6	N/A	No C++ twin is touched in the PR diff and the PR does not claim cross-engine parity from a golden update.
7	PASS	The no-key output zeroing fix in dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6258 and :6857 applies only to the 4warp_gqa builders, and the golden diffs in…
8	N/A	No rocKE C++ files are modified or added in the PR diff.
9	N/A	No shared op-handler or lowerer arch divergence fix is included in this PR.
10	N/A	No new kernel IR feature flag is introduced in this PR; it adds host-side SDPA dispatch bindings and fixes a bug in existing…
11	N/A	No new IR op is added in this PR.
12	N/A	core/backend.py and ROCKE_BACKEND handling are not touched in the PR diff.
13	N/A	LLVM .ll emission and debug-location handling are not modified in this PR.
14	N/A	No helper files under helpers/ or platform/cpp/helpers/ are modified in the PR diff.

§B graded 6/6

#	Verdict	Evidence
1	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:176-191 classifies shortcut feature declines with NOT_YET_IMPLEMENTED: separately from ABI/kernel limits, and…
2	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:211-214 probes each dense candidate with opt_in_probe, checks candidate.admits(probe), and only then builds the candidate spec…
3	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:239-255 declines requests when no unified candidate covers them, and dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:459-461 plus…
4	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/hipdnn.py:217-222 rejects a node missing its type key instead of defaulting it; dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:291-297 rejects tensors whose dtype differs…
5	PASS	dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:49-67 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_numeric.py:43-61 run GPU numeric verification over every case admitted by the unified and dense…
6	N/A	The visible diff adds SDPA dispatch admission/binding logic and kernel bug fixes without adding or modifying an is_valid_spec or…

§C graded 10/10

#	Verdict	Evidence
1	PASS	Shared predicates, decline reasons, dtype checks (check_dtype), 16-byte alignment checks (usable_in_place), and stream validation (launch_stream) are factored into dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:46-329, test tensor…
2	PASS	The SDPA bindings route to existing public entrypoints (bind_dense_attention_torch in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:345-347 and run_unified_attention_torch in…
3	N/A	No new kernel builder functions or kernel_name() definitions are introduced in this PR.
4	PASS	Arch-independent SDPA request/binding logic lives in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:1-408, while gfx942-specific kernel fixes stay in…
5	PASS	All touched files in the PR diff are directly scoped to SDPA dispatch/binding (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/__init__.py:1-192), SDPA tests (dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:930-935…
6	PASS	Shared SDPA request, band, decline-table, stream, dtype, and operand abstractions are cleanly factored into dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:62-408, and single-source public attention-layer accessors (attention_band…
7	N/A	This PR is a feature routing and bug-fix change rather than a pure no-op cleanup.
8	PASS	After rebasing onto develop, the PR reuses existing helpers including opt_in_probe, empty_workspace, and resolve_stream (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:30-31…
9	PASS	The no_key zeroing fix for rows where qlen > klen in bottom-right causal mode is applied across both the 4-warp lean and 4-warp GQA bodies in dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6256-6260 and…
10	PASS	Every feature in SDPA_FEATURES is explicitly partitioned into carried vs declined sets in both dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:82-128 and…

§D graded 21/21

#	Verdict	Evidence
1	N/A	This PR does not add or modify any KernelSpec knobs or environment variable reads.
2	PASS	Coverage limits and algorithm constants are imported from kernel/attention modules rather than transcribed (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:35 imports UNIFIED_ALGORITHMS and UNIFIED_HEAD_SIZES…
3	PASS	All candidate admission checks go through candidate.admits(probe) or ATTENTION_ROUTE_REGISTRY.supported(...) and never call _supports directly (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:215…
4	PASS	Selection and admission are pure functions of SdpaRequest taking req.arch directly without any device probes (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:90…
5	N/A	This PR does not introduce or modify any KernelSpec identity, spec_hash, compile_key, or kernel_name() methods.
6	N/A	This PR does not introduce or modify cohort routing, spec-builders, or kernel cache keys.
7	N/A	This PR does not add hand-pinned heuristic overrides.
8	N/A	This PR does not define or modify any kernel cohort predicates.
9	N/A	No new KernelCandidate or architecture is registered; the PR routes SdpaRequest onto existing ATTENTION_ROUTE_REGISTRY candidates.
10	PASS	Every SdpaRequest field and feature is translated into AttentionRequest and routed to the unified and dense bindings with dispatcher-level wiring, entry-point, and numeric tests (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:62…
11	PASS	Existing opt-in dense candidates remain opt-in and are queried explicitly via opt_in_probe(attn, candidate) in the SDPA dense gate (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:214).
12	PASS	Dense SDPA asserts the first serving candidate is the first variant in registry order at dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_wiring.py:312, and unified SDPA asserts candidate priorities remain sorted at…
13	PASS	Pinned SDPA dispatch choices are declined before routing rather than retried as auto at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:234, with tests covering known and missing pins at…
14	NEEDS-INFO	Previous NEEDS-INFO is not resolved: the visible diff does not include the cache-key use or ATTENTION_ABI_VERSION/abi_version hunks needed to verify compile-key identity and old-id refusal.
15	N/A	No new KernelCandidate or arch registration module is added; SDPA routes through existing attention candidates.
16	PASS	The SDPA family entry points short-circuit on _request_errors(req) before delegating or adding attention combos at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/init.py:112, with family-entry coverage at…
17	N/A	No new KernelCandidate is registered in this PR, so per-candidate CPU routing and family fixture invariants do not apply.
18	N/A	This PR does not change registry ratchets, route-only exemptions, or candidate binding requirements.
19	PASS	The SDPA docs state the returned dispatch results are attention results and that SDPA requests must launch through sdpa_unified_bind or sdpa_dense_bind at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/init.py:17, while unified binding uses…
20	N/A	The visible selection/routing changes are in the Python SDPA dispatch/binding layer over existing attention candidates, not in a…
21	PASS	Band mismatches are rejected in the SDPA admission chooser at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:135, and dense candidates are accepted only when their selected spec keeps the requested keys at…

§E graded 17/17

#	Verdict	Evidence
1	PASS	Platform changes remain strictly one-way with no library imports, e.g. dnn-providers/hip-kernel-provider/rocke/platform/python/rocke/runtime/torch_interop.py:96 and dnn-providers/hip-kernel-provider/rocke/platform/tests/test_rocke.py:7658.
2	PASS	Library SDPA bindings use rocke.runtime.torch_interop helpers rather than importing torch directly or querying devices with torch, e.g. dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:31 and…
3	PASS	Torch remains optional: tests gate it via pytest.importorskip in dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_binding.py:21 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_binding.py:22, while runtime…
4	PASS	Imports inside library/ flow downward from benchmarks to dispatch to kernels, e.g. dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/attention_combo_sweep.py:44 imports dispatch.attention.common and…
5	PASS	Installed CTest entries use relative paths via ${ROCKE_TEST_CTEST_REL} in dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:932 and dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:1075.
6	PASS	Emitter changes in dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6258 and dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6857 add explicit IR zero-select guards for masked rows rather than…
7	PASS	Authoring and routing documentation is updated alongside the code in dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:321 and dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:591.
8	PASS	Code comments and docstrings use descriptive references without Jira IDs, PR numbers, or user handles, e.g. dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:167 and dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:591.
9	PASS	The current PR diff against develop is cleanly scoped to SDPA routing and its gfx942 4-warp fix, e.g. dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:1 and…
10	NEEDS-INFO	Not resolved since the previous review: the PR description and diff (which touches shared files such as dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:932 and dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8) still do not provide a merge-simulation record listing open PRs touching the same tables/tests/goldens and suite reruns.
11	PASS	Committed files consist only of deliverable source, tests, CMake test wiring, and required authoring documentation, e.g. dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:591, dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:1, and…
12	PASS	Both gfx942 and gfx950 are covered in the SDPA routing gates and test suites, e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_wiring.py:39 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_wiring.py:43.
13	PASS	The rename of _problem, _attention_band, and _validate_lengths to public helpers updates definitions and callers across the tree, e.g. dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention/common.py:758…
14	PASS	The rebase onto develop cleanly preserves upstream code while updating only the intended lines and golden hashes in dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6256 and…
15	PASS	Every parameter in the bindings is used and validated, including stream in launch(*, stream: Any = None) in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:350 and…
16	N/A	This item is a meta-instruction for reviewer grading rather than a property of the PR code.
17	PASS	No absolute or relative performance claims appear in the PR description, comments, or documentation such as dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:591.

§F graded 11/11

#	Verdict	Evidence
1	PASS	Branch users/tlakshma/hack/sdpa/st2-unified matches the recognized users/<user>/... prefix, and PR changes are in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/__init__.py:1.
2	PASS	title feat(hip-kernel-provider): route SDPA to unified and dense attention kernels matches type(scope): summary and is 76 characters
3	PASS	description is at least 30 characters and contains an issue reference
4	PASS	test file dnn-providers/hip-kernel-provider/rocke/library/tests/dispatch/attention/test_additive_registration.py is in the diff
5	PASS	All changed files in the diff (such as dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:1 and dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:1) contain only code, docs, CMake, and test assets; no keys…
6	NEEDS-INFO	Still unresolved since previous review: local pre-commit execution status or CI lint check output is not visible in the diff.
7	NEEDS-INFO	Still unresolved since previous review: although the commits include review-response fixes (e.g. launch_stream in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:276, usable_in_place 16-byte alignment check in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:323, and _scale_reject in…
8	N/A	The PR targets Base: develop rather than a stacked feature branch, so the stacked-PR bot-skipping case does not apply.
9	N/A	There is no indication that the PR was artificially split or trimmed for a size cap (in fact, #14624 was folded into this PR).
10	PASS	The added and updated tests are substantive wiring, CPU data-movement, case-table, and GPU numeric tests rather than throwaway stubs, e.g. dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_binding.py:82…
11	PASS	The PR metadata header in this prompt shows the updated title (feat(hip-kernel-provider): route SDPA to unified and dense attention kernels) and full multi-section body (covering dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:1 and…

§G graded 34/34

#	Verdict	Evidence
1	PASS	Correctness is verified against an independent mathematical reference oracle (sdpa_reference) rather than relying solely on IR byte-identity (dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:61 and…
2	PASS	On-GPU numeric tests against the reference oracle are included and enrolled in the GPU test suite (dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:1075, dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:61, and…
3	NEEDS-INFO	In the current head, attention_d128_swa.json (dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8) and attention_d256.json (dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8) are updated, but the tail of attention_d256.json at line 98 is still truncated in the diff and local…
4	PASS	New admission, binding, alignment, stream-rejection, scale-rejection, dtype-rejection, and length/offset-validation rules have falsifiable unit tests (dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_binding.py:106…
5	PASS	In test_an_out_of_range_block_table_physical_id_is_rejected_before_launch (dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:83), the docstring explicitly documents why the test enters at validate_tuning_attention_tensors…
6	PASS	Prose claims in KERNEL_AUTHORING.md (dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:591), docstrings, and code comments (including the alignment, scale, dtype, and stream rejection comments in…
7	PASS	KERNEL_AUTHORING.md specifies the reference oracle and four test layers as a first-class gate (dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:633), and the CPU and GPU binding tests validate against sdpa_reference…
8	PASS	Bindings and test harnesses perform strict dtype checks on operands, stats, lengths, and ragged offsets (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:305…
9	N/A	This PR does not introduce or claim any LDS size or occupancy numbers.
10	N/A	This PR is a functional routing and bug-fix change, not a performance optimization PR.
11	PASS	Kernel reachability through the SDPA routing gates and family entry points is verified in dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_wiring.py:493…
12	N/A	The changes in this PR are in Python-only dispatch layers and a gfx942 4-warp GQA kernel builder…
13	N/A	No new LLVM intrinsic overload classes or mangled symbols are added or recommended.
14	N/A	No hardware bottleneck classification claims are made in this PR.
15	N/A	The PR diff is provided directly in the review prompt rather than inspected from a local git checkout.
16	PASS	Integration with existing attention dispatch symbols (ATTENTION_ROUTE_REGISTRY, UNIFIED_ALGORITHMS, DENSE_ALGORITHMS, attention_band, attention_problem, dense_scale_range, opt_in_probe) is verified by wiring and coverage tests…
17	N/A	No standalone CLI or benchmark scripts are shipped in this PR; new files are library modules and pytest suites enrolled in CMake.
18	N/A	This PR does not add statistical runtime guards or back-pressure detectors of the kind this item targets.
19	PASS	The fixtures and inline checks explicitly construct and catch the named failure modes across the binding and wiring gates: misaligned contiguous BSHD/ragged views at dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/binding_harness.py:17…
20	N/A	This PR does not add a claim-ban detector or regex sweep.
21	N/A	This PR does not introduce performance measurement or profiling deliverable code.
22	PASS	The SDPA numeric GPU tests are excluded from the host SDPA pytest shard at dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:933 and dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:934 and explicitly enrolled in the GPU CTest lane at…
23	N/A	This PR does not make speedup or routing A/B performance claims.
24	NEEDS-INFO	The new commits did not resolve the previous question: the diff updates golden hashes in dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8 and dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d256.json:8 (including attention_d256/gfx942/4warp_gqa_d192 at…
25	N/A	This PR is a functional routing and bounds/keyless-row correctness change, not a codegen micro-optimization report.
26	N/A	This PR does not make campaign-branch benchmark speedup claims.
27	N/A	This PR does not ship a performance knob or A/B toggle.
28	N/A	This PR does not present counterfactual performance tables.
29	N/A	This PR does not swap guarded math primitives for fast primitives.
30	N/A	This PR makes no register-spill or scratch-reduction claims.
31	N/A	This PR does not claim a performance win over a baseline defect.
32	N/A	The PR makes no timing, throughput, spill, or other measured performance-number claim requiring reproducible provenance.
33	N/A	The rocKE diff exposes and reuses attention routing helpers and adds SDPA gates, but does not modify an existing scalar…
34	PASS	The optional stats/LSE path is tested on padded/keyless rows, asserting O is zero and stats is -inf at dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_binding.py:144 and…

§H graded 5/5

#	Verdict	Evidence
1	PASS	sdpa_unified_bind allocates k_cache/v_cache sized to slots = batch * self.pages at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:376-382 and validates the launch tensors against plan.problem via…
2	PASS	_LaunchPlan builds an identity block_table in range(slots) at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:385-387, validates physical block IDs before launch via validate_tuning_attention_tensors at…
3	N/A	This PR's diff does not add or modify any device-side div/mod or shift/mask page-indexing logic (only epilogue zeroing for empty…
4	PASS	In build_gfx942_4warp_gqa, rows whose diagonal is left of key 0 (context_off + q_inseq < 0, where every score in the loop was masked) explicitly select 0.0 for O before storing at…
5	NEEDS-INFO	Resolved since previous review: all launch inputs feeding bounds/masks are host-validated with negative tests — seq_len_q/seq_len_kv via _host_lengths and validate_lengths at dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_unified.py:267-290 and dnn-providers/hip-kernel-provider/rocke/library/kernels/common/_sdpa_seqlen.py:45 (tested at…

§I graded 6/6

#	Verdict	Evidence
1	N/A	This PR wires hipDNN SDPA forward requests to existing attention candidates and fixes gfx942 kernel edge cases without introducing…
2	PASS	Non-contiguous or unaligned (<16-byte) layouts are repaired via workspace copy buffers (dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_dense.py:276-284, dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:323-328), and…
3	PASS	Curated request and kernel case tables (dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/cases.py:1361-1439) are combined with generated parameter sweeps and explicit algorithm-partition coverage checks…
4	N/A	This PR does not add or modify performance routing comparison tables or tuning reports.
5	N/A	This PR does not introduce baked tuning search results or static tuning snapshots.
6	N/A	This PR does not add any new DSL ops or hardware-instruction-named primitives.

§J graded 7/7

#	Verdict	Evidence
1	N/A	The only Markdown file in the PR (dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md) contains no mathematical…
2	N/A	The KERNEL_AUTHORING.md diff introduces no mathematical formula symbols, decay/log-decay domains, or log bases.
3	PASS	dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:591-651 links to existing sections [E](#e-new-operator-with-a-hipdnn-graph-node), [F](#f-new-operand-or-output-on-an-existing-operator), and [A](#a-new-kernel-family) while stating only the delta for…
4	PASS	The novelty claims in dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:321 and dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:647-650 name library/dispatch/sdpa/, bind_unified.py, bind_dense.py…
5	PASS	dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md makes no scaling/complexity claims, and the O(s_q) interval comparison claim in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:112-113 states the aggregate complexity…
6	N/A	The KERNEL_AUTHORING.md diff does not introduce any model configuration tables or tuned band-edge constants.
7	N/A	This PR edits the existing dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md and does not add a new design or…

§K graded 13/13

#	Verdict	Evidence
1	N/A	This PR adds no multi-configuration candidate (SDPA_REGISTRY remains empty and SDPA delegates to existing attention candidates…
2	N/A	This PR does not add or modify any tunable kernel spec fields, KnobAxis declarations, or field classifications.
3	N/A	This PR does not change config_key payloads, canonical-knob rules, declared defaults, or knob meanings (unified_rules.py only…
4	N/A	This PR does not modify KnobSpace.defaults() or KnobSpace.recorded() in any KnobSpace subclass.
5	N/A	This PR does not touch knob drop/refuse rules, inert(), base_value(), or KNOWN_WRONG_KNOBS.
6	N/A	This PR does not touch KnobSpace.validate or library/dispatch/attention/tuning_specs.py.
7	N/A	This PR does not modify KnobSpace.base(req) or _tuning_problem (it only renames _problem to attention_problem in…
8	N/A	This PR does not modify normalize_knobs or request-boundary knob scalar coercion.
9	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/bind_shared.py:244 (value_rejects) explicitly declines any pinned algorithm, spec_id, tuning_id, or tuning_knobs rather than adopting canonicalized ids or falling back to auto, tested in…
10	N/A	This PR does not modify KnobSpace.find or tuning_id resolution logic.
11	PASS	dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/__init__.py:130 (registered_sdpa_combos, sdpa_sweep_space, dispatch_sdpa_all) delegates to SDPA_REGISTRY and registered_attention_combos / dispatch_attention_all filtered by _serves(req)…
12	N/A	This PR does not reorder any production stack, change accept_default, or edit a tuning variant's base spec.
13	PASS	The benchmark changes in dnn-providers/hip-kernel-provider/rocke/library/benchmarks/common/attention_combo_sweep.py:44 and dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx950/attention/decode/decode_dispatch_diagnose.py:13 only rename _problem to…

§L graded 10/10

#	Verdict	Evidence
1	N/A	This diff does not add or modify a rocKE gfx target in arch specs or ISA backend registries.
2	N/A	This diff does not add or modify an MMA op_id, MMA fragment-info row, or MMA lowering layout map.
3	N/A	This diff does not add or change MMA atom selection, MMA catalog fields, lane maps, or intrinsic-name emission.
4	N/A	This diff does not introduce or change wave-size handling or lane math derived from wave size.
5	PASS	The arch-specific codegen change is confined to the gfx942 4-warp kernel and masks fully masked rows without leaking gfx950 recipes at dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6258.
6	N/A	This diff does not move pipeline vocabulary onto ArchTarget/core arch data, mutate ArchTarget from policies, or import CK Tile…
7	N/A	This diff only routes or declines SDPA dtype features and does not change rocKE logical dtype, bit-packing, scalar lowering, or…
8	N/A	This diff does not touch scaled-MMA scale contracts, retired scaled WMMA ids, or MfmaAtom.emit_scaled lowering.
9	N/A	This diff does not add a new rocKE architecture target.
10	N/A	This PR diff against base does not touch intrinsic_arch_domain..json, gen_arch_domain.py, or LLVM_FLAVORS.

§M graded 10/10

#	Verdict	Evidence
1	PASS	dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6258 and dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6857 compute no_key with b.cmp_lt and select the output with b.select and…
2	N/A	The PR diff does not touch legalize_for_target or target legalization logic.
3	N/A	The PR diff does not change print_ir, ck.dsl.ir/v1 serialization, wire versions, attrs, SSA ids, or parameter ordering.
4	N/A	The PR diff does not change portable IR schemas, importers, magic division constants, recipe VM code, or recorder coverage.
5	N/A	The PR diff does not touch AsyncTileLoader, async DRAM-to-LDS loading, LDS async layouts, or VMEM waits.
6	PASS	The kernel changes in dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6261 and dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tiled_2d.py:6860 only guard b.global_store inside…
7	N/A	The PR diff does not add or modify buffer-resource word 3 handling or TemporalHint.STREAMING lowering.
8	N/A	The PR diff does not add LLVM intrinsics or change backend selection precedence.
9	N/A	The PR diff does not touch public C entry points, cpp/include/rocke/ headers, or C++ codegen launch behavior.
10	N/A	The PR diff does not emit !dbg metadata or change debug-info consumers.

§N graded 18/18

#	Verdict	Evidence
1	N/A	The PR diff against develop does not add or modify any files under platform/python/rocke/helpers/.
2	N/A	The PR diff against develop only adds the no-key select guard in library/kernels/gfx942/attention_tiled_2d.py and does not add or…
3	N/A	The diff does not touch SoftwarePipeline, sync_before_issue, or overlap_vmcnt behavior.
4	N/A	The diff does not add or modify fused kernels, EpilogueOp, fusion schedules, or fusion validation paths.
5	N/A	The diff against develop does not add or modify a public API under rocke.helpers.tiling.
6	N/A	The diff against develop does not add or modify front-door or toolbox tiling helpers or LayoutStyle classes.
7	N/A	The diff against develop does not touch tiling layout overrides, accumulator c_desc derivation, or MMA soundness gates.
8	N/A	The diff against develop does not touch helpers/tiling LDS allocation, lds_store/lds_read, contraction padding, or wide-load…
9	N/A	The diff against develop does not use or modify N-D tiling axis roles, at_index/squeeze behavior, FlowStage labeling, or AB-swap…
10	N/A	The diff against develop does not modify tiling emit, LDS staging, or MMA issue code in helpers/tiling.
11	N/A	The diff against develop does not touch tiling MMA traits, TileMma target selection, lds_swizzle policy, or…
12	N/A	The PR diff makes no helpers/tiling LDS bank-conflict or coalescing optimization claim requiring simulator, counter, or ASM-backed…
13	N/A	The diff against develop does not construct WarpDistributionEncoding, change make_tile_desc factoring, re-export RegisterMapper…
14	N/A	The diff against develop does not use or modify helpers/tiling verbs or add a new helpers/tiling LDS pipeline variant.
15	N/A	The diff against develop does not touch helpers/tiling descriptor allocation or transformed-range bounds checks after swizzle…
16	N/A	The diff against develop does not use or modify optimize_layout, recommend(), or evaluate_transform ranking behavior.
17	N/A	The diff against develop does not modify tiling_recorder.py, round-trip analysis, or tiling visualization emission behavior.
18	N/A	The diff against develop does not touch InterleavedStyle, helpers/tiling/kernels demos, or tiling numeric gates for interleaved…

§O graded 16/16

#	Verdict	Evidence
1	N/A	The PR does not add or modify IR address spaces, required attributes, or core/verify.py.
2	N/A	The PR does not modify scf.for structure or terminator emission in core IR/lowerers.
3	N/A	The PR does not touch IRBuilder, intrinsic declaration tables, or metadata slot numbering.
4	N/A	The PR does not modify IR types or SmemType serialization/parsing.
5	N/A	The PR does not modify op purity tables or core/passes.py / cpp/core/passes.cpp.
6	N/A	The PR does not modify constant folding in Python or C++ passes.
7	N/A	The PR does not modify kernel attributes, codegen policy, or cluster dimension checks.
8	N/A	The PR does not set or modify scheduler strategies or CodegenPolicy.
9	N/A	The PR does not modify waitcnt encoding or barrier lowering.
10	N/A	The PR does not modify lowerer attribute validation or pointer parameter lowering.
11	PASS	dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_dense_sdpa_numeric.py:30 and dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:36 resolve the device architecture via…
12	N/A	The PR does not add hardcoded width/size caps in core/codegen or silent arch or "gfx950" defaults.
13	N/A	The PR does not touch any public C structs or enums under cpp/include/rocke/.
14	N/A	The PR does not touch C++ sources or add C++ compiler flags in platform/CMakeLists.txt.
15	N/A	The PR does not modify C++ engine fallback mechanisms or lowerer entry points.
16	N/A	The PR does not touch the C++ codegen library or gfx1250 builders/lowerers.

§P graded 14/14

#	Verdict	Evidence
1	N/A	The PR diff does not add or modify platform instance builder signatures or spec REQUIRED_FIELDS.
2	N/A	The PR diff does not touch platform kernarg signatures, manifest args_signature, runtime argument packing, or GEMM/conv ABI…
3	N/A	The PR diff does not touch GEMM, implicit-GEMM, or block-scaled GEMM launch geometry.
4	N/A	The PR diff does not modify LDS allocation gates or C++ validators.
5	N/A	The PR diff does not touch WMMA/RDNA GEMM, GemmMultiD, or GemmMultiAbd admission paths.
6	N/A	The PR diff does not touch Stream-K, MoE sorting, fused MoE, or split-K caller-zeroed workspaces in rocKE.
7	N/A	The PR diff does not touch Stream-K rejection rules, persistent GEMM, GroupedGemmLauncher, or SUPPORT_MATRIX.md.
8	N/A	The PR diff does not touch quantization helpers, MX GEMM, or MX scale decoding.
9	N/A	The PR diff does not add or modify numeric emulation or low-precision decode paths requiring edge-case emulation tests.
10	N/A	The PR diff does not add a manifest kind or touch helpers/manifest.py, run_manifest, or manifest allocate/verify consumers.
11	PASS	The SDPA bindings allocate and retain launch buffers once in the bound plan via empty_workspace/device_tensor, launch on resolve_stream(0, device=device), and reject non-current streams to prevent cross-stream races on preallocated workspaces: see…
12	N/A	The PR diff does not touch COMGR or HIP library loading logic.
13	N/A	The PR diff adds SDPA routing and binding logic rather than refactoring shared host-side grid, tile-count, workspace, cache-key…
14	N/A	The PR diff does not touch examples/, helpers/ placement, or add new modules under platform instances/common family glue.

§Q graded 7/7

#	Verdict	Evidence
1	PASS	The new SDPA CPU tests under tests/library/tests/sdpa run via the existing rocke_library_sdpa_pytest CTest target, with the two new GPU numeric files excluded from the CPU suite at dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:933-934 and…
2	N/A	This PR routes SDPA requests onto existing attention candidates and fixes an existing gfx942 4-warp attention kernel without…
3	PASS	The representative IR golden tripwires are updated for the modified gfx942 4-warp attention kernels at dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8 and…
4	PASS	The representative golden files are updated across all four LLVM flavor blocks in dnn-providers/hip-kernel-provider/rocke/platform/tests/golden/rocke_representative_ir_sha256/attention_d128_swa.json:8-50 and…
5	NEEDS-INFO	The full PR diff is now visible and all tests are portable: GPU numeric suites dynamically query the active device arch and skip cleanly without torch or a supported GPU (dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:17, dnn-providers/hip-kernel-provider/rocke/library/tests/sdpa/test_unified_sdpa_numeric.py:33…
6	N/A	This PR does not refactor any platform validation gate or differential harness.
7	N/A	This PR does not modify engine rejection contracts or ROCKE_BACKEND / ROCKE_CPP_STRICT fallback behavior.

§R graded 11/11

#	Verdict	Evidence
1	N/A	The diff routes SDPA over existing attention candidates and adds a correctness zero-select for fully-masked rows in…
2	N/A	The only kernel change is the correctness no_key zero-select at `dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/a…
3	N/A	The diff only adds the epilogue no_key correctness select in `dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/atte…
4	N/A	The diff adds only the no_key correctness guard in `dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_tile…
5	N/A	The diff does not add or change any global load or store address expression or access width (global_store at…
6	N/A	The diff does not change an LDS layout, MMA layout, tiling pipeline identity/relabel/reorder, or bank-conflict lever.
7	N/A	The diff makes no bottleneck claim such as memory-bound, LDS-bound, barrier-bound, or latency-bound.
8	N/A	The diff does not add, remove, or modify waits, barriers, sched_barrier, sched_group_barrier, iglp_opt, or scheduling hints.
9	N/A	The diff is not a performance-regression fix and does not add a shape-trading performance heuristic or dispatch rule.
10	N/A	The diff makes no codegen optimization, assembly-parity, occupancy, or benchmark-timing claims.
11	N/A	The documentation added in dnn-providers/hip-kernel-provider/rocke/KERNEL_AUTHORING.md:591 is an SDPA routed-family authoring…

§S graded 8/8

#	Verdict	Evidence
1	NEEDS-INFO	The new commits (rebase onto develop) do not resolve the missing hunk: torch_dtype and device_tensor are added at dnn-providers/hip-kernel-provider/rocke/platform/python/rocke/runtime/torch_interop.py:96 and tested in dnn-providers/hip-kernel-provider/rocke/platform/tests/test_rocke.py:7658, and SDPA tests are registered in dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:931 and…
2	PASS	Existing API behavior changes update their tests in the same PR: from_hipdnn requiring 'type' in dnn-providers/hip-kernel-provider/rocke/library/dispatch/sdpa/hipdnn.py:220 is tested at…
3	N/A	The diff does not remove or rename a public API; _problem, _attention_band, and _validate_lengths were private…
4	PASS	dispatch.attention.__all__ is updated in dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention/__init__.py:563 and verified by TestAttentionPublicSurface in…
5	PASS	Tests assert behavioral outputs and exact fail-fast exceptions: torch_dtype and device_tensor in dnn-providers/hip-kernel-provider/rocke/platform/tests/test_rocke.py:7662, non-positive/non-finite unified scale rejection in…
6	PASS	CPU-checkable contracts run in the host lane rocke_library_sdpa_pytest while ignoring the GPU numeric files in dnn-providers/hip-kernel-provider/rocke/platform/CMakeLists.txt:931, GPU numeric tests are registered in the GPU test list at…
7	N/A	The public APIs added or modified in this PR are Python dispatch/routing and torch_interop helpers that are Python-only by…
8	NEEDS-INFO	The new commits do not resolve the prior NEEDS-INFO: the PR description is truncated at LSE is written in place into a contiguous FP32 (B and does not show whether the README test entry points (python tests/test_rocke.py and python tests/run_all.py from rocke/platform/) were stated as run.

[cgmillette]

Automated review: Model - qwen3.6:35b-a3b-coding-mtp-q4_K_M, effort: high

PR #13115 — route SDPA to unified and dense attention kernels (ThruptiRajLakshmanaGowda, AICK-2613, AICK-2614) — 🟠 Follow-up: develop conflict, scale range

Prior item	Status
Earlier 🟢 rows: arai713, AnaghaRaoAMD ×5, dlangbe, Copilot (earlier; 11-05 alignment, scale, int(x) offsets), CongMa13 ×2, yraparti (10-31, 11-04 ×4); our N1–N10, N12–N15; 14483 left bound	🟢 FIXED, unchanged • delta reopens none at head; SDPA + attention tests pass (ran)
Open threads: 8 of 36	🟡 OPEN: 6 resolvable • fixes hold: AnaghaRaoAMD 5, yraparti 1 (read) • wait: _MIN_SCALE getattr (N20), K/V copy
AnaghaRaoAMD (11-03) bind_dense.py: getattr _MIN_SCALE silently disables the check	🟠 OPEN on the merge • holds at head; develop's 14908 moved the constants (ran) • the merge disables the check exactly as warned: N20
AnaghaRaoAMD (11-03) bind_unified.py:420: K/V copied every launch	🟡 DEFERRED · AnaghaRaoAMD decides • unchanged: Not in this PR, KERNEL_AUTHORING §H (read) • author asked them to resolve or ask for it (11-06)
Copilot (11-05) bind_unified.py:445, bind_dense.py:348: stream discarded; our N16	🟢 FIXED cf2b94ad • one launch_stream helper; named current stream accepted (ran) • other streams refused; unknown keywords raise TypeError (ran)
Copilot (11-05) sdpa/__init__.py:146: raw attention DispatchResults	🟢 FIXED cf2b94ad • documented as ranking results; SDPA binders exported (read)
Our N17 (11-05): each tensor gets its own dtype rule	🟢 FIXED cf2b94ad • one check_dtype: operands, stats, seq_len_*, offsets (read) • int64 seq_len_q, fp16 stats now refused; tests pass (ran)
Our N18: unified scale decline untested in CI; ragged alignment untested	🟢 FIXED cf2b94ad • 10 scale cases run in CI; removing _scale_reject fails all 10 without torch (ran) • ragged twin fails without the alignment check (ran)
Our N11: body omits new public API	🟢 FIXED cf2b94ad • lists dense_scale_range, binders, the four binding rules (read) • GPU box checked at this head (read)
CongMa13 (11-06): threads and our post unanswered	🟢 ANSWERED • author replied to all three (11-06) (read)
yraparti (11-07): checklist re-review at cf2b94ad	🟢 PASS • 0 FAIL rows; 11 NEEDS-INFO from a truncated diff (read)
cgmillette (11-05): approval	🟢 APPROVED 7ad42faf • older head; still counts on develop • predates 0d916301, cf2b94ad and the develop merge
AnaghaRaoAMD (11-06): approval	🟢 APPROVED 53ef210c • older head; still counts on develop • predates cf2b94ad's stream and dtype rules
Bots / CI	🟢 Green at head • 46 ✓, 52 skipped, 0 ✗, none running (read) • gfx94X SDPA lane runs the 10 new scale cases (read) • merge: dirty; run predates develop's 14908
Our checks	🟠 Conflict • merge vs 477ab5b9: 1 file conflicts, 34 behind (ran) • build, ctest, byte identity, goldens: not run, no merged tree • with torch: head 6672 ✓; merge (develop's side) 4 ✗ / 6718 ✓ (ran) • ABI: develop 12, no change here; rules 1–3 clean (ran)
Code quality	🟠 Should-fix • stream, dtype, in-place, finite-scale rules: one home each (read) • each new guard's removal fails a test (ran) • left: N20 reads a moved constant through a getattr default

#	Sev	file:line	Finding
N19	🟠	tests/dispatch/attention/test_gfx950_dense_wiring.py:777	Conflicts with develop's 14908, which moved these scale tests • impact: no merged tree, so no CI, build or golden run; 2 gfx942 SDPA tests fail on the merge (ran) • fix: take develop's file; move test_dense_scale_range_names_this_module to test_attention_dense_scale.py; pin 14908's three gfx942 dense candidates
N20	🟠	dispatch/attention/common.py:937	dense_scale_range returns None once develop is merged • impact: SDPA dense gate admits scales outside [2^-64, 16] on gfx942 and gfx950; the runner raises at launch (ran) • fix: return DENSE_MIN_SCALE/DENSE_MAX_SCALE (kernels.common.attention_dense_spec) for both arches, no getattr default; update the body's line

Utility: 🟢 Positive — stream, dtype and CI test gaps fixed; merging develop needs two ports

---

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

Follow-up review: first pass by qwen3.6 (local), verified by the orchestrator (Claude Opus 5.5). Checks: delta since 0d916301, merge with develop 477ab5b9 taking develop's side of the conflict, the saved stream probe and a new scale-range probe, mutants of the new tests with and without torch, SDPA and attention tests with torch at head and on the merge, CI at head, code quality. Flags whether a closer look is needed; not an approval or rejection.