[vLLM] Update vLLM pin#7539
Draft
quinnlp wants to merge 28 commits into
Draft
Conversation
The install-benchmarks action installed the latest main-branch wheel whenever no same-run wheel artifact was present, so direct workflow_dispatch runs on a branch that modified benchmark sources silently ran the stale main wheel instead of the branch code. Detect branch divergence via `gh api compare main...<sha>` (mirroring install-vllm). When the branch changed benchmarks/ and no same-run wheel is available, build from the checked-out source; if that build fails, fail loudly rather than falling back to the known-stale main wheel. Behavior on the non-diverged path (main/schedule/PR) is unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The install-benchmarks action installed the latest main-branch wheel whenever no same-run wheel artifact was present, so direct workflow_dispatch runs on a branch that modified benchmark sources silently ran the stale main wheel instead of the branch code. Detect branch divergence via `gh api compare main...<sha>` (mirroring install-vllm). When the branch changed benchmarks/ and no same-run wheel is available, skip the main wheel and build from the checked-out source; the build step has no continue-on-error, so a failing branch build fails the job loudly rather than falling back to the known-stale main wheel. Behavior on the non-diverged path (main/schedule/PR) is unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The ^benchmarks/ pattern uses only the ^ anchor, which behaves identically in basic and extended regex, so -E is not needed. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Piping `gh api --paginate` directly into `grep -q` lets grep exit on the first match and close the pipe, killing gh with SIGPIPE. Under pipefail that surfaces as a pipeline failure, so the `if` reads it as "no match" and the stale main wheel gets installed anyway. Capture the file list first, then grep the string. Also tidy the source-build notice message. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The vllm-xpu-kernels cutlass_grouped_gemm wrapper was replaced by cutlass_grouped_gemm_xe2, which adds a bias arg before output and no longer wraps rows_per_expert in a tensor internally. Update both call sites to the 9-arg signature and pass an int32 xpu tensor for rows_per_expert. This dispatches correctly on both PVC and BMG. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Replace the bespoke `gh api compare` detection in install-benchmarks with the shared check-path-changes action, per review discussion. To support this: - check-path-changes now picks its diff base by trigger: github.base_ref on pull_request runs, main otherwise (e.g. workflow_dispatch, where base_ref is unset). This keeps existing PR callers unchanged while supporting dispatch. - The detection step is moved after "Validate benchmarks (same run)" and gated on that failing, so the common PR path (same-run wheel present) skips the extra git diff entirely. - Direct callers (triton-benchmarks, vllm-benchmarks, third-party-benchmarks) now checkout with fetch-depth: 0 so the git diff against origin/main works. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
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
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.
Closes #7531.
Updates pinned vLLM commit to
4ec199b66a791070348f3baf847b3a873c48cdd9.Updates pinned vLLM XPU kernels commit to
d1dab45d00973eef321cb41b088ff35ff29d8179.vLLM benchmarks:
vLLM benchmarks BMG:
vLLM tests:
vLLM tests BMG:
vLLM benchmarks performance impact: TODO