Unpin liger-kernel 0.8.1

[qgallouedec]

Requires #6741 (GPU CI on L40S).

0.8.1 was excluded because it needs bf16 gemm (compute capability >= 8.0) and our T4 runner is sm_75 see #6516. The L40S runner is sm_89, so the constraint is gone.

• liger-kernel>=0.8.0,!=0.8.1 → liger-kernel>=0.8.0 (liger + dev extras)
• Drop the SAPO IndexPutBackward0 warning filter, fixed upstream in 0.8.1 (Avoid torch.compile graph break in SAPO loss linkedin/Liger-Kernel#1274)

EDIT: after second thought, we shouldn't merge this now: trl[liger] is published, so unpinning breaks users on pre-Ampere GPUs (#6516) even though our CI moved to an L40S. PEP 508 markers can't scope a dependency by GPU arch. The upstream guard (linkedin/Liger-Kernel#1316) merged ~24h after 0.8.1 shipped, so it lands in 0.8.2 (not to be confused with #1274 (SAPO fix), which is in 0.8.1.)
When 0.8.2 ships: bump both extras to liger-kernel>=0.8.2, drop the SAPO IndexPutBackward0 filterwarnings entry.

---

Note

Low Risk
Packaging and test-warning cleanup only; no runtime training logic changes.

Overview
Allows liger-kernel 0.8.1 again by changing liger-kernel>=0.8.0,!=0.8.1 to liger-kernel>=0.8.0 in the liger and dev optional dependencies. That exclusion existed because 0.8.1 needs bf16 GEMM (compute capability ≥ 8.0), which the old T4 CI runner could not satisfy; GPU CI on L40S removes that constraint.

Removes the pytest filterwarnings entry that ignored Liger SAPO IndexPutBackward0 under torch.compile, since that behavior was fixed upstream in 0.8.1.

^{Reviewed by Cursor Bugbot for commit 545bf8f. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 545bf8f0de

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[albertvillanova]

Thanks, @qgallouedec.

I agree with your EDIT: this shouldn't be merged, and I'd go further and close it.

Some context on why the constraint has its current shape: the !=0.8.1 in #6518 was chosen over <0.8.1 precisely for this situation. It excludes exactly the one broken release while leaving every future release allowed, so no TRL-side change is needed for 0.8.2 to be picked up; the current pin already permits it. That makes this PR unnecessary rather than merely early.

Three additional reasons beyond the GPU-arch point:

1. The end state in this diff is >=0.8.0, which permanently re-permits 0.8.1. Even once 0.8.2 exists, a bare >=0.8.0 lets a resolver land on the broken release (old lockfiles, --no-deps installs, constrained resolutions). The right target is >=0.8.2, as your EDIT says.
2. Timing makes it worse than it looks. Because 0.8.1 is currently the newest release, unpinning today doesn't just allow the broken version: every fresh pip install trl[liger] would resolve to exactly it on pre-Ampere hardware (#6516).
3. The filterwarnings removal is coupled to the same bump. With !=0.8.1 in force, CI resolves 0.8.0, which predates #1274, so dropping the ignore entry brings the IndexPutBackward0 noise back. The existing comment already records this (Remove once: liger-kernel >= 0.8.2 is required).

And as you note, PEP 508 markers can't express "0.8.1 is fine iff compute capability >= 8.0", so there's no way to keep CI on L40S free of the constraint while protecting published trl[liger] users. A runtime guard in TRL would just duplicate the upstream fix, which I don't think is worth it.

So in conclusion, I would close this PR. Thanks for catching this before it went in: the GPU CI move to L40S (#6741) is genuinely the right change, but it just doesn't unblock this one.

[albertvillanova]

For the record, the follow-up work is tracked in two issues, so nothing is lost when this PR is closed:

• Bump liger-kernel to >=0.8.2 once released (drop the !=0.8.1 exclusion) #6767: bump liger-kernel>=0.8.0,!=0.8.1 → liger-kernel>=0.8.2 in both the liger and dev extras, once 0.8.2 ships with Guard the FLCE addmm fast path on compute capability 8.0 or higher linkedin/Liger-Kernel#1316.
• Remove SAPO IndexPutBackward0 warning filter once liger-kernel >= 0.8.2 is required #6435: remove the "ignore:Error detected in IndexPutBackward0:UserWarning" filter. Its upstream fix (Avoid torch.compile graph break in SAPO loss linkedin/Liger-Kernel#1274) is in 0.8.1, which we exclude, so 0.8.2 is the first non-excluded release that contains it.

The two are coupled and should land in the same PR; #6767 carries the full checklist.