[assistant-librarian]

Updated all pinned TheRock workflow and source refs to 8b46a27 due to submodule bump

Baseline run ID: 36390205996

[therock-pr-bot]

🤖 Bump PR detected (author @assistant-librarian[bot]). All policy checks are auto-approved for automated dependency bumps.

[tony-davis]

CI review: flake / tracked / new breakdown

Verdict: hold. This bump adds new CI red beyond what's tracked in Project 231. MIOpen's hipBLASLt FP8/BFP8 GEMM tests fail deterministically on gfx942 (reproduced on rerun, and green in the nightlies on either side of this PR's base), and rocgdb-corefile breaks on a CLI flag mismatch. Code-owner review is also still required.

Per-failure disposition (rocm-libraries scope)

Job	Target / Arch	Signature	Disposition	Ticket
miopen shard 1, 2, 3	miopen (gfx94X-dcgpu)	test-gtest: all FP8/BFP8 HipBLASLtGEMMTest cases, error ~0.12-0.17 vs 0.07	new, caused by bump (rerun bit-identical)	#11116 (updated)
miopen shard 4	miopen (gfx94X-dcgpu)	infra-gpu: HW Exception ... GPU Hang	probable flake, not rerun	none
TensileLite coverage	tensilelite (Component CI)	coverage ratchet: SubtileGREmit.py -2.08%	not bump-related, green on develop now	none

TheRock-owned (not ours to fix)

Job	Target / Arch	Signature	Ticket
rocgdb-corefile	rocgdb (gfx94X-dcgpu)	unrecognized arguments: --gpu-corefile-tests	not filed (TheRock-owned)

rocm-systems-owned reds (visibility only)

Job	Target / Arch	Signature
rocprofiler-sdk	rocprofiler-sdk (gfx94X-dcgpu)	3 signal-handler integration tests fail
hip-tests (ROCR) shard 2	hip-tests (Windows gfx110X-all)	test-timeout: 120 min

What this bump actually brings in

TheRock 7440cb8 to 8b46a27 (14 commits). Submodules that moved:

Submodule	Repo	Commits	Notable changes
rocm-systems	ROCm/rocm-systems	52	clr batch-copy kernarg fix (ROCm/rocm-systems#11650, the only HIP runtime change and the lead suspect for the FP8 miss), new rocddi systems layer, rccl test and perf work, hip-tests test-level discovery
debug-tools/rocgdb/source	ROCm/rocgdb	7	adds --gpu-corefile-tests to test_rocgdb.py, gcore test renames
math-libs/libhipcxx	ROCm/libhipcxx	2	Windows headertest path shortening, test requirements file
third-party/sysdeps/common/mesa-fork	ROCm/mesa-fork	340	upstream Mesa sync (radv/radeonsi/panvk), not on the compute path
rocm-libraries	ROCm/rocm-libraries	5	TheRock's own pointer; this repo's CI builds from the PR source, so it doesn't matter here

TheRock itself also adds the --gpu-corefile-tests CI switch (ROCm/TheRock#8563), an ASAN-debug nightly for gfx90a, and a ci:run-multi-gpu opt-in label.

Writes made in this review

• Updated: [Issue] [MIOpen][gfx942] HipBLASLt FP8/BFP8 GEMM Numerical Failures + DBSync Regression on gfx94X-dcgpu #11116 (evidence, ci-red/bump-pr/code-regression, added to Project 231)
• Commented: [Issue] Linux : hipkernelprovider — hip-kernel-provider-external-integration_comprehensive_suite hangs, 30-min timeout #11529 (three green nightlies, suggested close)

[tony-davis]

Superseded by #12826 (TheRock c9bde5b), which includes everything here. The two reds from this one are clean there: MIOpen gfx94X FP8/BFP8 passes (I think it was a hipBLASLt race fixed by #12615, not the bump; details on #11116), and rocgdb-corefile passes now that rocgdb moved to f5e5294. I'd close this one and review #12826 instead.