fix(rocke): name seqlen and tile when dense attention rejects a partial tile by yassinsolim · Pull Request #14335 · ROCm/rocm-libraries

yassinsolim · 2026-10-23T21:49:35Z

Replaces #13681 with the same commit (1b17bab). GitHub closed #13681 when its branch moved to users/ysoliman/rocke/ for the rocKE branch-naming rule, and a PR cannot be reopened once its branch is renamed. The review history, including the rocKE checklist reviews, stays on #13681.

Carved out of #12928, which is closed. The rest of that PR is already on develop (#12710 landed the nhead_k guard, #12719 replaced the factories this touched and routes _request_errors through every dense candidate).

Motivation

The dense _base_spec factories set ragged=True for any self-attention length that is not a tile multiple, so Sq = Skv = 384 or 1152 at block_m 256 is an ordinary graph taking the on-chip ragged path. The gfx942 predicate then declined with "gfx942 attention_dense: ragged not yet supported" and the shared spec validator with "ragged is not supported with sliding_window". Both read as a varlen problem and name nothing the caller can act on. Found while running the hipDNN kernel-integration runbook on gfx942.

Technical details

kernels/common/attention_dense_spec.py: new partial_tile_reason(spec) names the length and the tile, for example "seqlen_q 384 is not a multiple of block_m 256". The sliding-window branch of __post_init__ now says "sliding_window needs whole tiles: (the partial-tile path has no window mask)".
kernels/gfx942/attention_dense.py: supports_attention_dense says "gfx942 attention_dense needs whole tiles: (the partial-tile path is not yet supported)".
Varlen messages are unchanged.

What could break: anything matching the old "ragged not yet supported" or "ragged is not supported with sliding_window" text. A repo grep for those two strings finds only the two lines this PR rewrites, and that was not enough. One test matched on less than the whole string: tests/dispatch/attention/test_unified_mask_wiring.py asserted the substring "ragged" in the gfx950 window refusal, and this branch broke it. It is now test_partial_tile_window_is_declined_not_served and checks the length, the tile and the window, the same way the gfx942 and gfx950 wiring tests do. Anything outside this repo that keys on the word "ragged" has the same exposure.

Test plan and result

Host only, no GPU, Windows Python 3.12, from dnn-providers/hip-kernel-provider/rocke/library with library and platform/python on PYTHONPATH.

tests/dispatch/attention plus tests/test_attention_dense_gfx942.py at head 1b17babac3: 882 passed, 121 skipped, 3296 subtests passed. The merge base, develop at 28af48c, gives 882 passed, 121 skipped, 3293 subtests; the 3 extra subtests are the new 384 / 1152 / 1000 cases.
Negative control: with the two messages reverted, test_rejects_a_partial_tile_and_names_the_length, test_sliding_window_with_partial_tile_rejected_at_dispatch and the gfx942 supports marker all fail.
black 25.12.0 --check is clean on the changed files.

Refs #12920.

Submission checklist

Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

Merged develop 3f51737 and updated the three rocke attention wiring tests to match the new refusal message.

…clines a partial tile The dense _base_spec factories set ragged=True for any self-attention length that is not a tile multiple, so Sq = Skv = 384 or 1152 at block_m 256 is an ordinary graph that takes the on-chip ragged path. The gfx942 predicate then declined with "ragged not yet supported" and the shared spec validator with "ragged is not supported with sliding_window", both of which read as a varlen problem and name nothing the caller can act on. partial_tile_reason(spec) names the length and the tile. The gfx942 predicate now says "gfx942 attention_dense needs whole tiles: seqlen_q 384 is not a multiple of block_m 256 (the partial-tile path is not yet supported)" and the sliding-window check says "sliding_window needs whole tiles: <reason> (the partial-tile path has no window mask)". Varlen messages are unchanged. What could break: anything matching the old "ragged not yet supported" or "ragged is not supported with sliding_window" text. No other caller in the repo matches either string. Checked host only, no GPU: tests/dispatch/attention plus tests/test_attention_dense_gfx942.py give 752 passed, 75 skipped, 2600 subtests passed. With the two messages reverted, the three updated tests fail. black 25.12.0 --check is clean on the changed files.

…in the unified mask tests The sliding_window branch of AttentionDenseSpec.__post_init__ no longer says "ragged", so test_ragged_window_is_declined_not_served asserted text that this branch removed. It now checks the same three parts the gfx950 and gfx942 wiring tests check: the query length against block_m, the key length against block_n, and that the reason still names sliding_window and no longer says ragged.

therock-pr-bot · 2026-10-23T21:52:12Z

✅ 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 · 2026-10-23T21:52:15Z

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

Copilot

🟢 Approval recommended

The diagnostic-only changes preserve rejection behavior and include focused regression coverage, with no unresolved blocking issues.

No open issues 🎉

What changed in this PR

This PR makes rocKE dense-attention rejection messages actionable by naming sequence lengths and tile sizes instead of describing ordinary partial-tile inputs as “ragged.” Existing rejection conditions and varlen messages remain unchanged.

Changes:

Adds a shared helper that explains which sequence lengths are not tile multiples.
Updates gfx942 and sliding-window rejection messages to use that explanation.
Expands regression assertions for partial-tile diagnostics through dispatch and kernel support checks.

File	Description
`dnn-providers/hip-kernel-provider/rocke/library/tests/test_attention_dense_gfx942.py`	Checks the specific length-and-tile rejection message.
`dnn-providers/hip-kernel-provider/rocke/library/tests/dispatch/attention/test_unified_mask_wiring.py`	Verifies partial-tile window refusals identify both lengths and tiles.
`dnn-providers/hip-kernel-provider/rocke/library/tests/dispatch/attention/test_gfx950_dense_wiring.py`	Updates sliding-window diagnostic assertions.
`dnn-providers/hip-kernel-provider/rocke/library/tests/dispatch/attention/test_gfx942_dense_wiring.py`	Covers multiple partial-tile lengths and omits aligned dimensions from diagnostics.
`dnn-providers/hip-kernel-provider/rocke/library/kernels/gfx942/attention_dense.py`	Uses the shared explanation for partial-tile rejection.
`dnn-providers/hip-kernel-provider/rocke/library/kernels/common/attention_dense_spec.py`	Adds the diagnostic helper and updates sliding-window validation wording.

_{Review effort: Balanced}

cgmillette · 2026-10-29T09:30:35Z

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

PR #14335 — name seqlen and tile when dense attention rejects a partial tile (yassinsolim) — 🟢 First-round: no closer look needed — wording-only; every verdict matches develop; the red summaries are cancelled duplicate runs

Prior item	Status
Copilot (10-14 on #13681 at `938a7683`; 10-23 here at `1b17baba`): overviews only, no findings; "rejection conditions and varlen messages remain unchanged"	🟢 Confirmed — a probe over 376 cases gives the same accept/reject on develop and on the merge; only the two rewritten messages change (ran; see Our checks). This is the only review on this PR
yraparti (10-19 at `7d9983a5`, re-review 10-20 at `1b17baba`, on #13681) rocKE checklist §F.2 HIGH: move the branch onto a name with a `rocke` segment	🟢 FIXED — the head branch is now `users/ysoliman/rocke/dense-partial-tile-reason`. That rename closed #13681 and opened this PR with the same commit (read)
yraparti §F.3 MEDIUM: retitle as `type(scope): summary`, 10–80 characters, scope `[a-z0-9-]+`	🟢 FIXED — #13681 was retitled on 10-23, before the move here, from a 90-character title to today's `fix(rocke): name seqlen and tile …`, 76 characters, scope `rocke`. The checklist's NEEDS-INFO rows are verdicts it could not tie to a diff line, or context it lacked; none asks for a change (read)
Bots / CI	🟢 Policy bot pass. CI ran at `1b17baba` on 10-23, 409 commits behind today's develop. Every run that went to completion passed: TheRock Multi-Arch CI (29 jobs passed, none failed), hipDNN Superbuild (Linux and Windows), Component CI, pre-commit and clang-tidy; the ASAN workflow selected no jobs. hipkernelprovider ran on Linux gfx94X and Windows gfx110X, 2 shards each, 41/41 ctests, and the three changed dispatch wiring tests PASSED in every shard (`rocke_library_attention_pytest`). • The two red checks, Multi-Arch CI Summary and Multi-Arch CI ASAN Summary, come from duplicate runs that started with the PR and were cancelled within two seconds. Their summary jobs failed only because `ci-env` / `configure` and `setup` were cancelled; the surviving runs' summaries passed → unrelated to this PR. • The same commit's 10-20 run under #13681 passed on its second attempt. Attempt 1 hit a Windows `rocke_pytest` timeout in `test_hip_device_properties_layout` and a gfx94X packaging test, both outside this PR. • `test_attention_dense_gfx942.py` is in no CTest lane (pre-existing), so its one changed assertion runs only locally; `test_rejects_a_partial_tile_and_names_the_length` pins the same gfx942 message in CI (read)
Our checks	🟢 • Clean merge with develop `f22bed43` (409 behind, 4 ahead). Develop edited five of the six files since the base (#12714, #13802), and all five auto-merged. Native build and 33 ctests OK (1 skip, the arch-domain reproduction gate); C++/Python library parity passes at llvm20/22/23; differential golden OK at llvm20/22/23/24 and canonical, with 0 unblessed changes and 0 missing (the NEW configs are develop's, as no emitter changes here); the PR's four test files on the merge: 476 passed, 214 subtests (ran). • Only the wording changes. A probe ran 376 cases on develop and on the merge: 176 shared-spec constructions, 60 gfx942 predicate calls, and 140 pinned `admits` calls across all five `attention_dense` candidates on gfx942 and gfx950, through the spec factories and `_request_errors`. Accept/reject is identical in all 376, and the 132 changed messages all come from the two rewritten sites. A query-only, KV-only or two-length partial tile names exactly the failing length(s), and a hand-set `ragged=True` on whole tiles says so. The non-ragged messages (`seqlen_q must be a multiple of block_m=…`) are unchanged (ran); the varlen ones are untouched (read). • No semantic drift since the base.* gfx942 still has no partial-tile path on develop (`supports_attention_dense` rejects `ragged`, and its docstring still lists ragged as a later follow-up), and the shared validator still rejects a partial tile with a window on every arch, so both new messages hold. On develop the old strings occur only at the two lines rewritten here. The only rocKE tests on develop keyed on "ragged" are the three this PR updates and `test_attention_dense_bwd.py:260`, which checks the separate backward spec (read). • Rules 1–3 clean (ran); no links added (read)

Net-new: none.

Utility: 🟢 Positive — refusals name the failing length and tile, not "ragged"; verdicts unchanged.

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

First-round review: a local qwen3.6 model made the independent first pass. The orchestrator (Claude Opus 5.5) verified it. On a clean merge with today's develop, it ran the native build and ctests, the C++/Python library parity, the differential goldens and the PR's four test files, plus a probe that compares every accept/reject verdict and message with develop's. It also checked develop's changes since the PR's base for drift, read the predecessor #13681's reviews, and read the CI jobs at head, including the two cancelled duplicate runs. The orchestrator owns this merge-worthiness call. It flags whether a closer look is needed; it is not an approval or a rejection.

AnaghaRaoAMD

Merged into current develop (43e2cb7aa47), this branch fails 14 tests in the files it touches, while develop alone passes all 558. #14367 merged after this branch's base (28af48c) and made the dense block_m depend on the shape, so some shapes now build at 128. Because of that, the hardcoded block_m 256 expectations no longer hold, and gfx942 now accepts Sq = Skv = 384. CI here is green because it ran on the old base. The source change still applies cleanly and is still needed, since both old messages are still on develop. It needs a develop merge and test updates (details inline).

…eason

…l-tile tests Develop now builds these requests at block_m 128 instead of 256, so the hardcoded 256 and the Sq = Skv = 384 case no longer hold. The gfx942 test reads block_m and block_n from the default spec, the gfx950 tests parse them from the refusal message, and the 384 case is dropped since a 128-row tile accepts it.

cgmillette · 2026-11-04T16:14:17Z

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

PR #14335 — name seqlen and tile when dense attention rejects a partial tile (yassinsolim) — 🟠 Follow-up: no test pins the tile

Prior item	Status
AnaghaRaoAMD (11-04) 12 gfx950 window subtests fail since #14367	🟢 FIXED `a630f00a83` • all 12 pass on today's develop; no hard-coded 256 (ran) • tile now read from the refusal itself: N1 • AnaghaRaoAMD caught this before our checks; resolved by cgmillette
AnaghaRaoAMD (11-04) gfx950 dense `block_m 256` assert fails	🟢 FIXED `a630f00a83` • passes on today's develop; no hard-coded 256 (ran) • same refusal-read tile: N1 • resolved by cgmillette
AnaghaRaoAMD (11-04) gfx942 accepts 384; 1152/1000 hard-code 256	🟢 FIXED `a630f00a83` • 384 dropped; tile read from `_base_spec`; passes (ran) • 1152 = 9 × 128: breaks again at a 128-row pick (ran) • resolved by cgmillette
cgmillette (10-29), BrianHarrisonAMD (11-04) approvals	🟢 APPROVED `1b17babac3` • older head; still counts on develop • not the `a630f00a83` test rewrite; source unchanged (read)
Copilot (10-23), ours (10-29): wording only, verdicts unchanged	🟢 Confirmed `a630f00a83` • source patch identical to `1b17babac3` (read) • only the tests' hard-coded 256 broke when develop moved
yraparti (on #13681) §F.2 branch name, §F.3 title	🟢 FIXED • both still hold at head (read)
Bots / CI	🟢 Green • 46 ✓ · 51 skipped; 4 required ✓ (read) • 3 changed wiring files ran in all 4 rocKE shards: 170 ✓ (read) • merge: blocked on rocke-lib code-owner review; squash auto-merge armed
Our checks	🟢 Clean • merge vs today's `fccc13fa33` clean; PR tests 558 ✓, 228 subtests (ran) • `961dd76390` merge: build + 46 ctests, parity 2 ✓ ×4, goldens OK (ran) • installed: 3 of 4 files, 170 ✓; `test_attention_dense_gfx942.py` in no CTest entry, as on develop (ran) • rules 1–3 clean (ran); body and armed squash message still cite `1b17babac3` and 384 (read)
Code quality	🟠 Should-fix • a constant 256x64 tile passes all 558 tests: N1 (ran) • exported helper misreports a non-ragged spec: N2 (ran) • `test_attention_dense_gfx942.py:853` restates the imported `_BLOCK_M` (read) • tile parse copied at `test_unified_mask_wiring.py:382`; pairs come from the catalogue (read)

#	Sev	file:line	Finding
N1	🟠	`library/tests/dispatch/attention/test_gfx950_dense_wiring.py:634`	No test checks which tile the refusal names • impact: a helper always naming 256x64 passes all 558 (ran); here and `test_unified_mask_wiring.py:382` the tile comes from the refusal • fix: gfx950: expect the refused spec's tile (spy on `partial_tile_reason`); gfx942 `:220`: add 500 (128 rows), 1088 for 1152 (ran)
N2	🟡	`library/kernels/common/attention_dense_spec.py:69`	Exported helper says `ragged=True` for a non-ragged spec • impact: a whole-tile 2048 spec gets "ragged=True at seqlen_q 2048 …"; no test covers this text (ran) • fix: raise `ValueError` unless `spec.ragged` (breaks no test, ran); test a hand-set `ragged=True` on whole tiles

Utility: 🟢 Positive — refusals name the failing length and tile; tests follow develop's picks

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: merge with today's develop, build and ctest, library parity, differential goldens, changed tests (source and installed), refusal-tile mutation, CI shard logs, code quality. Flags whether a closer look is needed; not an approval or rejection.

… spec helpers Merges origin/develop. N1. #14335 rewrote the ragged + sliding_window reject: it now names the length and tile that leave a partial tile (partial_tile_reason). The C port still said "ragged is not supported with sliding_window", so on develop the vectors test failed, and a regenerate would have failed the C ctest. Ported partial_tile_reason and the new text; three cases cover its seqlen_kv-only, both and whole-tile branches (only seqlen_q was recorded). Vectors regenerated; the C ctest matches all four. N2. #14835 added rocke_spec_set_reasonf / ROCKE_SPEC_FAIL and rocke_floor_div to helpers/spec. The port's local reject() and py_floordiv() are gone: 58 rejects use ROCKE_SPEC_FAIL and the three floor divisions rocke_floor_div. rocke_spec_set_reasonf now marks a cut reason with "..." like rocke_spec_set_reason, as its header already promised ("same truncation"). Nit: the header's ZeroDivisionError note no longer quotes one CPython 3.12 text for both operators.

…al tile (#14335) Replaces #13681 with the same commit (1b17bab). GitHub closed #13681 when its branch moved to `users/ysoliman/rocke/` for the rocKE branch-naming rule, and a PR cannot be reopened once its branch is renamed. The review history, including the rocKE checklist reviews, stays on #13681. Carved out of #12928, which is closed. The rest of that PR is already on develop (#12710 landed the nhead_k guard, #12719 replaced the factories this touched and routes `_request_errors` through every dense candidate). ## Motivation The dense `_base_spec` factories set `ragged=True` for any self-attention length that is not a tile multiple, so Sq = Skv = 384 or 1152 at block_m 256 is an ordinary graph taking the on-chip ragged path. The gfx942 predicate then declined with "gfx942 attention_dense: ragged not yet supported" and the shared spec validator with "ragged is not supported with sliding_window". Both read as a varlen problem and name nothing the caller can act on. Found while running the hipDNN kernel-integration runbook on gfx942. ## Technical details - `kernels/common/attention_dense_spec.py`: new `partial_tile_reason(spec)` names the length and the tile, for example "seqlen_q 384 is not a multiple of block_m 256". The sliding-window branch of `__post_init__` now says "sliding_window needs whole tiles: <reason> (the partial-tile path has no window mask)". - `kernels/gfx942/attention_dense.py`: `supports_attention_dense` says "gfx942 attention_dense needs whole tiles: <reason> (the partial-tile path is not yet supported)". - Varlen messages are unchanged. What could break: anything matching the old "ragged not yet supported" or "ragged is not supported with sliding_window" text. A repo grep for those two strings finds only the two lines this PR rewrites, and that was not enough. One test matched on less than the whole string: `tests/dispatch/attention/test_unified_mask_wiring.py` asserted the substring "ragged" in the gfx950 window refusal, and this branch broke it. It is now `test_partial_tile_window_is_declined_not_served` and checks the length, the tile and the window, the same way the gfx942 and gfx950 wiring tests do. Anything outside this repo that keys on the word "ragged" has the same exposure. ## Test plan and result Host only, no GPU, Windows Python 3.12, from `dnn-providers/hip-kernel-provider/rocke/library` with `library` and `platform/python` on PYTHONPATH. - `tests/dispatch/attention` plus `tests/test_attention_dense_gfx942.py` at head `1b17babac3`: 882 passed, 121 skipped, 3296 subtests passed. The merge base, develop at 28af48c, gives 882 passed, 121 skipped, 3293 subtests; the 3 extra subtests are the new 384 / 1152 / 1000 cases. - Negative control: with the two messages reverted, `test_rejects_a_partial_tile_and_names_the_length`, `test_sliding_window_with_partial_tile_rejected_at_dispatch` and the gfx942 `supports` marker all fail. - black 25.12.0 `--check` is clean on the changed files. Refs #12920. ## Submission checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. Merged develop 3f51737 and updated the three rocke attention wiring tests to match the new refusal message.

assistant-librarian · 2026-11-06T12:54:31Z

TheRock Submodule Bump Activity

Newest first

2026-11-06 — Included in TheRock via ROCm/TheRock#9759.

yassinsolim and others added 4 commits October 14, 2026 12:19

Merge branch 'develop' into users/ysoliman/hqs-rocke942-tiles

142d3ad

Merge branch 'develop' into users/ysoliman/hqs-rocke942-tiles

7d9983a

yassinsolim requested a review from a team as a code owner October 23, 2026 21:49

yassinsolim added organization: ROCm project: hip-kernel-provider rocke-lib: attention Touches rocKE library attention/FMHA kernels labels Oct 23, 2026

yassinsolim requested review from a team and a balanced review from Copilot October 23, 2026 21:49

yassinsolim mentioned this pull request Oct 23, 2026

fix(rocke): name seqlen and tile when dense attention rejects a partial tile #13681

Closed

1 task done

Copilot started reviewing on behalf of yassinsolim October 23, 2026 21:50

View session

Copilot AI reviewed Oct 23, 2026

View reviewed changes

cgmillette approved these changes Oct 29, 2026

View reviewed changes

BrianHarrisonAMD approved these changes Nov 4, 2026

View reviewed changes

AnaghaRaoAMD reviewed Nov 4, 2026

View reviewed changes

yassinsolim added 2 commits November 4, 2026 00:12

Merge branch 'develop' into users/ysoliman/rocke/dense-partial-tile-r…

a556232

…eason

cgmillette enabled auto-merge (squash) November 4, 2026 15:55

AnaghaRaoAMD approved these changes Nov 4, 2026

View reviewed changes

cgmillette merged commit c6bca3b into develop Nov 4, 2026
99 checks passed

cgmillette deleted the users/ysoliman/rocke/dense-partial-tile-reason branch November 4, 2026 16:26

yassinsolim mentioned this pull request Nov 5, 2026

rocKE: check which tile the dense partial-tile refusal names #15136

Open

1 task done

Navigation Menu

Repository navigation

fix(rocke): name seqlen and tile when dense attention rejects a partial tile - #14335

fix(rocke): name seqlen and tile when dense attention rejects a partial tile#14335
cgmillette merged 6 commits into
developfrom
users/ysoliman/rocke/dense-partial-tile-reason

yassinsolim commented Oct 23, 2026 •

edited

Loading

Uh oh!

therock-pr-bot Bot commented Oct 23, 2026 •

edited

Loading

Uh oh!

therock-pr-bot Bot commented Oct 23, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

cgmillette commented Oct 29, 2026

Uh oh!

AnaghaRaoAMD left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

cgmillette commented Nov 4, 2026

Uh oh!

Uh oh!

assistant-librarian Bot commented Nov 6, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Conversation

yassinsolim commented Oct 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Technical details

Test plan and result

Submission checklist

Uh oh!

therock-pr-bot Bot commented Oct 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

✅ All Checks Passed — Ready for Review

Uh oh!

therock-pr-bot Bot commented Oct 23, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟢 Approval recommended

Uh oh!

cgmillette commented Oct 29, 2026

PR #14335 — name seqlen and tile when dense attention rejects a partial tile (yassinsolim) — 🟢 First-round: no closer look needed — wording-only; every verdict matches develop; the red summaries are cancelled duplicate runs

Uh oh!

AnaghaRaoAMD left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

cgmillette commented Nov 4, 2026

PR #14335 — name seqlen and tile when dense attention rejects a partial tile (yassinsolim) — 🟠 Follow-up: no test pins the tile

Uh oh!

Uh oh!

assistant-librarian Bot commented Nov 6, 2026

TheRock Submodule Bump Activity

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

yassinsolim commented Oct 23, 2026 •

edited

Loading

therock-pr-bot Bot commented Oct 23, 2026 •

edited

Loading