docs(skills): point vllm-bench install at vllm-project/vllm rust workspace - #888
Conversation
1ecf541 to
f519c9b
Compare
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1ecf541e77
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".
| outputs, | ||
| finished_requests: (!finished_requests.is_empty()).then_some(finished_requests), | ||
| scheduler_stats: Some(Box::new(scheduler_stats_from(&self.scheduler.load()))), | ||
| scheduler_stats: Some(self.stats_with_spec()), |
There was a problem hiding this comment.
When an abort arrives while a speculative GPU step is still running, handle_message removes the stream before that step is delivered; dispatch_step then drops the unknown request update and returns at outputs.is_empty() before this new stats_with_spec() call. If this is the last active request, the scheduler's cumulative acceptance increment is never sent to the frontend, leaving /metrics stale indefinitely. Send a stats-only batch or consume the snapshot for every received step rather than only steps with surviving request output.
Useful? React with 👍 / 👎.
| git clone https://github.com/vllm-project/vllm-bench.git | ||
| cd vllm-bench | ||
| ./install.sh | ||
| cargo install --path rust/src/bench # -> ~/.cargo/bin/vllm-bench |
There was a problem hiding this comment.
Following the documented sequence leaves the shell in <clone>/vllm/rust, so cargo install --path rust/src/bench resolves to the nonexistent <clone>/vllm/rust/rust/src/bench and the new installation workflow fails. From that working directory the path should be src/bench, or the command should first return to the clone root.
AGENTS.md reference: AGENTS.md:L137-L139
Useful? React with 👍 / 👎.
…space The standalone vllm-project/vllm-bench repo is archived and read-only; vllm-bench now lives as a crate at rust/src/bench in the main vLLM repo. Build with cargo build --release -p vllm-bench (or cargo install --path rust/src/bench); drop the frozen standalone-repo install routes. Signed-off-by: xiaguan <751080330@qq.com>
f519c9b to
5414e6e
Compare
|
Review comment addressed in 5414e6e: the install command now runs Scope note: this PR is now docs-only (skill install update). The stepped-bridge fix commit has been split off — tracked in #896 instead. |
xiaguan
changed the title
1 participant
Summary
The standalone
vllm-project/vllm-benchrepository is archived and read-only;vllm-benchis now developed as a crate in the vLLM Rust workspace (vllm-project/vllm→rust/src/bench). The oldcargo install --git/ prebuilt-binary routes still work but are frozen at the last standalone commit.Update the vllm-bench skill to the maintained path:
git clone --depth 1 https://github.com/vllm-project/vllm.git && cd vllm/rust && cargo build --release -p vllm-benchcargo install --path src/bench<vllm-repo>/rustVerified locally:
cargo build --release -p vllm-benchsucceeds andvllm-bench --versionruns.