chore: swap vllm-router fork for upstream on the router stack - #3257
Draft
mikasenghaas wants to merge 2 commits into
Draft
chore: swap vllm-router fork for upstream on the router stack#3257mikasenghaas wants to merge 2 commits into
mikasenghaas wants to merge 2 commits into
Conversation
Consume vllm-router from PyPI (vllm-project/router, >=0.1.15) instead of the pinned wheel from the PrimeIntellect-ai/router fork (v0.1.26). Upstream now covers everything prime-rl invokes: the consistent_hash policy, --intra-node-data-parallel-size, --vllm-pd-disaggregation, and a typed /inference/v1/generate route that preserves extra request fields. Fork-only features are dropped: the routed-experts merge for P/D disaggregation, JWT auth, and per-run usage metrics. Single-replica router replay still works - the generate route passes return_routed_experts through and forwards response bytes verbatim. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The router subprocess shared the terminal's foreground process group, so Ctrl-C SIGINTed it directly: it died before the engine drained, and the watcher thread saw the exit before the finally block marked shutdown, logging a spurious ERROR and re-signalling the process. Start the router in its own session so only inference_local owns its lifecycle: the engine drains first, then the finally block terminates the router. The watcher now also logs a zero exit as a warning - that is an external shutdown racing ours (e.g. the rl launcher SIGTERMs the whole tree, router first), not a crash. Co-Authored-By: Claude Fable 5 <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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Stacks #3250 (replace the vllm-router fork with the upstream release) on top of #3255, so router performance — fork
consistent_hashvscache_awarevs upstream — can be compared within one stack using the per-engine metrics from #3240.Same two commits as #3250, rebased onto the stack:
vllm-routerfrom PyPI (vllm-project/router,>=0.1.15) instead of the pinned wheel from the PrimeIntellect-ai/router fork (v0.1.26); drop the[tool.uv.sources]wheel URLs and theexclude-newer-packageexemption.See #3250 for the full rationale and the
## Breakingnotes (P/D routed-experts merge, JWT auth, and per-run usage metrics are fork-only and dropped).Verification
uv syncresolvesvllm-router0.1.15 from PyPI on this base; the cherry-picked lock is coherent (no drift after re-sync).uv run inference --vllm.model Qwen/Qwen3-0.6Bon this branch: upstream router binary starts withpolicy: CacheAware, serves chat completions, and tears down cleanly on SIGINT with no spurious ERROR.