Flight Log / vllm-project/vllm #21146
[Core] Freeze gc during cuda graph capture to speed up init — a OpenAI Codex (OpenAI) flight by Misha Goin
Misha Goin (@mgoin) opened this OpenAI Codex (OpenAI) pull request in vllm-project/vllm on Jul 17, 2025; merged after rework in 6.1 d: +10,535 −15,819, 321 files, 9 reviews.
Flown by
- Lines
- +10,535 −15,819
- Files
- 321
- Commits
- 6
- Reviews
- 9 (1 ✕)
- To merge
- 6.1 d
- Task
- perf
Line and file counts are summed per commit from the AIDev dataset; GitHub's net diff can be smaller.
OpenAI Codex, working in vllm-project/vllm, took on a performance task: freeze the garbage collector during CUDA graph capture to shorten engine initialization. The change touched 321 files across six commits. Nine reviews were filed, including one change request and four review comments. The pull request was merged after rework, 146.21 hours after opening.
“While I agree that maybe we need to do `gc.collect` only once for the entire graphs (rather than once per graph), I think a more”
— from the review (source)
From the pull request description
## Summary Speed up cudagraph capture loops by calling `gc.freeze` before capture. This speeds up cudagraph capture a huge amount, especially for small models. Qwen3-0.6B goes from 35s to 2s. For the "proper" approach we should possible use https://github.com/pytorch/pytorch/pull/158193 in a future …
Excerpt (first 300 characters) — the full description is on GitHub.
View the pull request on GitHub →
This is your flight? Claim it — sign in with GitHub. Ownership is proven by the sign-in: the account that opened the PR is the only one that can claim it.
Flight timeline
- —committedMichael Goin
- —reviewed
- —committedmgoin
- —reviewed
- —reviewed
- —committedmgoin
- Jul 17, 2025labeledmgoin
- Jul 17, 2025labeledmergify[bot]
- Jul 17, 2025labeledmgoin
- Jul 17, 2025changes requestedWoosukKwon↗
- Jul 19, 2025labeledmergify[bot]
Read next
Cite this flight: AgentRiders Flight Log, vllm-project/vllm#21146, record updated Nov 4, 2026, https://agentriders.com/logs/vllm-project/vllm/21146/
Data from public GitHub activity; this page is not affiliated with GitHub or the repository.