[chore] upgrade flashinfer to 0.6.14 - #2135
Merged
Merged
Conversation
vLLM 0.26.0's own metadata asks for flashinfer-python==0.6.14; we were holding all three flashinfer packages at 0.6.13 because flashinfer-cubin appeared to have no 0.6.14 release, and flashinfer hard-errors when cubin's version differs from its own. cubin 0.6.14 does exist -- it is just not on PyPI. PyPI stops at 0.6.13, but flashinfer publishes the pure-python wheel on its own index. So this adds a `flashinfer` index (https://flashinfer.ai/whl/) and sources flashinfer-cubin from it, the same way flashinfer-jit-cache already comes from the cu130 index. flashinfer-python stays on PyPI, which does carry 0.6.14. All three now move together at 0.6.14, satisfying the version-equality check. The >=0.6.13 floor still holds, so the MNNVL allreduce+RMS fusion fix is kept. Lock also picks up two transitive moves: cuda-tile 1.3.0 -> 1.5.0 and the removal of nvidia-cuda-tileiras. Testing: on a B200 worker, `test_new_inference_generation.py -m vllm` -- 19 passed. Verified the env under test resolved flashinfer-python 0.6.14, flashinfer-cubin 0.6.14 and flashinfer-jit-cache 0.6.14+cu130, and that `import flashinfer` succeeds (the check that would fail on a version mismatch). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01W7RfwMcdvBsVWS5W61wJEZ
SumanthRH
marked this pull request as ready for review
September 2, 2026 00:34
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request updates the flashinfer packages (flashinfer-python, flashinfer-jit-cache, and flashinfer-cubin) to version 0.6.14 in pyproject.toml and configures a new CUDA-agnostic index to source flashinfer-cubin. Additionally, it modifies tests/backends/skyrl_train/gpu/utils.py to default LD_LIBRARY_PATH to an empty string if it is not present in the environment. The reviewer correctly pointed out that defaulting LD_LIBRARY_PATH to an empty string can lead to security risks or library resolution issues, and suggested conditionally setting the variable only when it exists in the parent environment.
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
This was referenced Sep 2, 2026
Merged
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Bumps flashinfer packages (
flashinfer-python,flashinfer-cubinandflashinfer-jit-cache) to 0.6.14.For
flashinfer-cubinwe need to use the wheel on flashinfer's index instead of the wheel on PyPINote
Medium Risk
Touches the pinned GPU inference dependency graph (flashinfer + lockfile) used by vLLM on Linux; install/resolution regressions are the main risk, with no production training code changes beyond a small test env fix.
Overview
Aligns the Linux x86_64 flashinfer stack (
flashinfer-python,flashinfer-jit-cache,flashinfer-cubin) with 0.6.14, matching what vLLM 0.26.0 expects instead of the previous 0.6.13 pin. Override comments inpyproject.tomlare updated to reflect that rationale (including the ≥0.6.13 MNNVL fusion requirement).Because flashinfer-cubin is not published on PyPI beyond 0.6.13, the PR adds an explicit
flashinferuv index (https://flashinfer.ai/whl/) and wiresflashinfer-cubinthrough[tool.uv.sources]while jit-cache stays on the cu130 index.uv.lockis refreshed accordingly (including transitive bumps such ascuda-tile1.5.0 tied to the new flashinfer release).GPU test Ray init only forwards
LD_LIBRARY_PATHinto the runtime env when it is already set, avoiding propagating a missing/Nonevalue.Reviewed by Cursor Bugbot for commit 9b86de8. Bugbot is set up for automated code reviews on this repo. Configure here.