chore: update TRTLLM-Gen GEMM cubins

[bobboli]

📌 Description

Update the TRTLLM-Gen dense GEMM artifact pin to the newly published multi-architecture package:

• artifact: 7b1fc253cd6237950e76310873f4acf4d97a3904/gemm-b738138-25754e6/
• manifest SHA-256: ca9d4f956f3fb63bff3066db88fa7ccf08b00f4b0b2751cc14ba72454fd01638
• kernels: 186 cubins (93 sm100f + 93 sm107a)

This brings in the TRTLLM-Gen dense GEMM fixes from revision b738138 while retaining the existing GEMM config hash 25754e6.

🔍 Related Issues

N/A

🚀 Pull Request Checklist

✅ Pre-commit Checks

• Commit-time pre-commit hooks passed.
• pre-commit run --all-files was not run for this pin-only change.

🧪 Tests

• The public artifact and manifest URLs return HTTP 200.
• The downloaded checksums.txt SHA-256 matches the new pin and contains 201 entries.
• Full local test suite was not run.

Reviewer Notes

The artifact was published from cubin-publishing main pipeline 65421116, job 418937481; public publish job 418937489 completed successfully.

Summary by CodeRabbit

• Bug Fixes
  • Updated the TensorRT-LLM generated GEMM artifact reference.
  • Updated its verification checksum to ensure the correct artifact is retrieved and validated.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 39ed8e04-2173-4727-859d-9d4fbc87235e

📥 Commits

Reviewing files that changed from the base of the PR and between 2cc51dc and afc8caa.

📒 Files selected for processing (1)

• flashinfer/artifacts.py

Included review availability: Your plan provides up to 8 included reviews per hour; 7 remain after this review.

---

📝 Walkthrough

Walkthrough

The TRTLLM generated GEMM artifact path pin and its corresponding SHA256 checksum were updated in flashinfer/artifacts.py.

Changes

TRTLLM GEMM artifact

Layer / File(s)	Summary
Update artifact pin and checksum flashinfer/artifacts.py	ArtifactPath.TRTLLM_GEN_GEMM now uses the new artifact path pin. CheckSumHash.TRTLLM_GEN_GEMM now uses the matching SHA256 checksum.

Estimated code review effort: 1 (Trivial) | ~2 minutes

Merge Risk: ⚪ Minimal · up to afc8c

This PR updates the pinned TRTLLM-Gen GEMM package and matching manifest checksum without changing retrieval or integrity-checking behavior. The change is localized and merge-ready after normal checks and review, with no actionable merge-blocking risk remaining.

Suggested reviewers: jimmyzho, iwakurarein, jiahanc

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely identifies the main change: updating the TRTLLM-Gen GEMM cubin artifacts.
Description check	✅ Passed	The description follows the repository template and documents the artifact pin, checksum, validation steps, test status, and reviewer notes. It clearly states that the full test suite and all-file pre…
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 1…
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

Full details: Description check

Explanation

The description follows the repository template and documents the artifact pin, checksum, validation steps, test status, and reviewer notes. It clearly states that the full test suite and all-file pre-commit run were not performed.

Full details: Docstring Coverage

Explanation

No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 1 files.

✨ Finishing Touches 💡 1

🛠️ Fix failing CI checks 💡

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[jiahanc]

/bot run tests/gemm

[jiahanc]

LGTM

[flashinfer-bot]

GitLab MR !1368 has been created, and the CI pipeline #65443666 is currently running. I'll report back once the pipeline job completes.

[flashinfer-bot]

[FAILED] Pipeline #65443666 — 9/16 executed test jobs passed

Compared with nightly #65302610.

Unit Tests

GPU	CUDA 12.9	CUDA 13.0	Notes
B300	❔ Failed	❔ Failed	—
GB200	❌ New	❔ Failed	New: tests.gemm.test_mm_fp8.py (30 failures; CUDA 12.9)
GB300	❌ New	❔ Failed	New: tests.gemm.test_mm_fp8.py (30 failures; CUDA 12.9)
H100	✅ Pass	✅ Pass	—
RTX Pro 6000 Blackwell	✅ Pass	⚠️ Infra	Infrastructure: CI infrastructure failure (1 job; CUDA 13.0)

✅ Pass · 🟡 Old failure · ❌ New failure · ⏱ Test timeout · ⚠️ Infrastructure · ❔ Unknown or unclassified · — Not run

Multi-GPU and Multi-Node Tests — 6/6 passed

GPU	CUDA 12.9	CUDA 13.0	Notes
B300 (multi-GPU)	✅ Pass	✅ Pass	—
GB200 (multi-node)	✅ Pass	✅ Pass	—
GB300 (multi-node)	✅ Pass	✅ Pass	—

Failure details

New relative to nightly (attribution uncertain)

• tests.gemm.test_mm_fp8.py — 60 failures on GB200 / CUDA 12.9, GB300 / CUDA 12.9
  • not executed due to timeout

Timeouts, infrastructure, or incomplete jobs

• Infrastructure: CI infrastructure failure — RTX Pro 6000 Blackwell / CUDA 13.0
  • Jobs: unit_test_rtx-pro-6000-blackwell: [cu130]

[bobboli]

Follow-up on the reported tests/gemm/test_mm_fp8.py failures:

The PR head was based on an older main and did not contain #4848, which filters incompatible SM107 low-latency GEMM tactics on Blackwell. The merged result (b2e36efc) does contain #4848.

I re-ran the full test file against the actual merge commit, using the updated GEMM artifact (7b1fc253.../gemm-b738138-25754e6/) on B300 with the CUDA 12.9 CI image:

tests/gemm/test_mm_fp8.py
30 passed, 3 warnings in 49.50s

No additional fix or cubin rollback is needed for this issue.