perf(hipblaslt): Updating Origami libraries for MXFP4 TN for gfx950

[smalekta]

Motivation

Updating Origami's Libraries for MXFP4 TN for gfx950. Adding 23 new MTs with Average uplift of 7-8%.

JIRA ID: AIHPBLAS-4019

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	PR does not contain code files — Unit Test auto-passed
🔎 pre-commit	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—
🤖 therock-pr-bot	✅ Pass	—

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

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

🙋 Wish to Override Policy?

[therock-pr-bot]

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

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop   #10016      +/-   ##
===========================================
- Coverage    69.18%   68.74%   -0.44%     
===========================================
  Files         2762     2740      -22     
  Lines       454772   451339    -3433     
  Branches     67092    66531     -561     
===========================================
- Hits        314619   310267    -4352     
- Misses      117276   118445    +1169     
+ Partials     22877    22627     -250     

Flag	Coverage Δ		*Carryforward flag
TensileLite-CPP	38.12% <ø> (+0.02%)	⬆️	Carriedforward from b63b947
TensileLite-Unit	75.90% <ø> (-0.01%)	⬇️	Carriedforward from b63b947
hipBLAS	90.62% <ø> (ø)		Carriedforward from b63b947
hipBLASLt	34.89% <ø> (ø)
hipCUB	83.47% <ø> (ø)		Carriedforward from b63b947
hipDNN	86.35% <ø> (-0.17%)	⬇️	Carriedforward from b63b947
hipFFT	47.34% <ø> (-0.40%)	⬇️	Carriedforward from b63b947
hipRAND	76.12% <ø> (ø)		Carriedforward from b63b947
hipSOLVER	69.18% <ø> (+0.15%)	⬆️	Carriedforward from b63b947
hipSPARSE	86.27% <ø> (ø)		Carriedforward from b63b947
rocBLAS	47.95% <ø> (-0.28%)	⬇️	Carriedforward from b63b947
rocFFT	47.61% <ø> (-4.52%)	⬇️	Carriedforward from b63b947
rocRAND	57.01% <ø> (-0.01%)	⬇️	Carriedforward from b63b947
rocSOLVER	76.92% <ø> (ø)		Carriedforward from b63b947
rocSPARSE	72.57% <ø> (+0.01%)	⬆️	Carriedforward from b63b947
rocThrust	91.60% <ø> (ø)		Carriedforward from b63b947

*This pull request uses carry forward flags. Click here to find out more.
see 132 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[geomin12]

bump here: #10338

[AmosLewis]

Gardener triage: failures are a live GitHub Actions outage — proceeding with an override merge

Agreed that the failures are unrelated, and I can confirm the cause is platform-wide rather than
anything in this PR. Summary of what I checked and why I am comfortable overriding.

GitHub Actions is in an active major outage

GitHub's status page currently reports Actions: major_outage under
Incident with Actions (opened 15:22 UTC, still investigating at
18:46 UTC). The latest update states that workflow runs "are still failing, and jobs may remain
queued for an extended period before starting or may time out".

All 17 red checks on head cc9d848 trace to this, and none is a genuine test failure:

Failing check(s)	Error	Cause
Windows gfx1151 Test hipblas, Test hipblaslt; Test miopen (gfx94X); Windows::release math-libs (gfx110X-all); Build RPM Packages	Failed to resolve action download info → Service Unavailable / Internal Server Error / Bad Gateway, plus HTTP request timed out after 00:01:40	GitHub outage — these died in Getting action download info, i.e. before any test code ran
Build Python Packages	Executing the custom container implementation failed. Please contact your self hosted runner administrator.	Runner/container failure; GitHub's update explicitly flags self-hosted runner registration errors. It had materialised wheels through gfx1151 at 17:06:28 and died one second later
Test Python × 6 (py3.10/3.11/3.12 × ubi10/ubuntu24.04)	Could not find a version that satisfies the requirement rocm==10.1.0.dev0+819aa1c...; --find-links=/gfx94X-dcgpu/index.html does not exist	Cascade — the wheels were never produced by the job above
Windows::release math-libs (gfx120X-all), (gfx1151)	The operation was canceled.	GitHub outage
Windows gfx1151 Test rocblas shard 1	Driver / GPU sanity check timed out after 3 minutes	Known infra issue ROCm/TheRock#6857
TheRock CI Summary	no error of its own	Aggregates the above
codecov/project/hipBLASLt	coverage -0.44%	Carryforward-flag artifact on a data-only PR

This is repo-wide, not specific to this PR: during the outage window 73 pull_request runs were left
with no conclusion, 13 were cancelled, and only 7 succeeded. Waiting has been making it worse rather
than better — the red count on this PR grew from 7 to 17 over two hours as the queued
TheRock Multi-Arch CI run crawled forward and hit the outage at each new stage. Since reruns cannot
reliably start while Actions is degraded, waiting for auto-merge to go green is not a viable path
today.

Why this is safe to override

hipBLASLt's authoritative gate is AMD-internal math-ci, which is a separate system from GitHub
Actions and is unaffected by the outage. It is green on the current commit:

• Math CI Summary — pass, "All math-ci jobs passed for on current commit"
• mci/rocm-libraries/precheckin(hipblaslt), preliminary, static-analysis, codecov,
  tensilelite-unit-codecov — all pass on build 8

On the TheRock side the relevant lanes are green too: Build (gfx950-dcgpu), Build (gfx94X-dcgpu),
all 6 Test hipblaslt shards on gfx94X, Test hipBLASLt HOST_ASAN | gfx90a, pre-commit, and
therock-pr-bot. The diff itself is 3 gfx950 Origami tuning YAMLs — pure data, with no path by which
it could break Windows gfx1151/gfx110X/gfx120X, miopen, RPM packaging, or Python wheel builds.

Downstream impact confirmed

#10390
(refactor(hipblaslt): convert gfx950 F4 library logic to dict format) modifies a strict superset of
this PR's 3 files — the same F4BS_MXA32_MXB32_BH_BiasSB_HAS_SAB_SAV YAMLs under Origami/,
Origami_nta4/, and Origami_ntb4/. It reformats exactly the data this PR updates, so #10016 needs
to land first or that conversion has to be redone. #10390 is one of a ~18-PR series converting gfx950
logic to dict format, so the ordering matters beyond a single PR.

Override rationale, recorded per the policy

• Known infra issue with tracking links: GitHub incident above, plus
  ROCm/TheRock#6857 for the gfx1151 sanity-check hang.
• Unrelated to the PR code change: 3 gfx950 tuning YAMLs; failures are Windows/miopen/packaging.
• No new failure pattern: every red is a platform error or a documented cascade from one.
• Rationale documented in thread: this comment.

One point I want to be transparent about: TheRock's Test (gfx950-dcgpu) lane shows skipping by
matrix configuration — not as a result of the outage — so the GitHub side never exercised gfx950 at
runtime on this PR. gfx950 coverage for this change rests on math-ci precheckin and the reviewers'
approval. As gardener I am signing off on the CI signal interpretation and the infra classification,
not on the gfx950 tuning numerics, which remain with @babakpst / @msujon-AMD who approved the data.

Merging now with squash, preserving the original commit message so JIRA ID: AIHPBLAS-4019 is
retained. I will watch post-merge actions on develop and will flag anything that appears.