fix: cap SWA/chunked-local runtime admission to startup pool-sizing bound#1659
Conversation
…ound Signed-off-by: AlpinDale <alpindale@gmail.com>
![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: 87d8d64015
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".
| # `sum(reservations) <= pool` <=> `sum(peak_real_held) <= pool`. | ||
| # Drift between the two would re-introduce the deadlock from | ||
| # issue #39734 or, worse, mid-prefill OOM. | ||
| num_required_blocks = min(num_required_blocks, self._max_admission_blocks_per_request) |
There was a problem hiding this comment.
For running chunked-local requests, this cap is applied before the fast-path subtracts len(req_to_blocks), but req_to_blocks includes historical null placeholders from skipped chunks. Once the slot list grows past the cap, get_num_blocks_to_allocate can return 0 even when a new real KV block is still needed in the current chunk (no new skipped block freed in that step). In Scheduler._schedule_running this underestimation bypasses the free-block guard in KVCacheManager.allocate_slots, and allocate_new_blocks can then raise ValueError when BlockPool.get_new_blocks is called with insufficient free blocks.
Useful? React with 👍 / 👎.
1 participant
No description provided.