Repository navigation
knowledge(vllm-omni): 1 change(s) from intake - #342
Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![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: c4e2c17e0e
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".
| - 触发：修改 async-chunk placeholder prewarm 的任一层：`StagePipelineConfig.async_chunk_prewarm_payload_func`（legacy 与 typed metadata 两条路径都要解析）、orchestrator 初始 add 的 `_attach_async_chunk_prewarm_payload`、scheduler `add_request` 对 `_async_chunk_prewarm.*` 键的弹出与 `OmniSchedulerOutput.pending_request_prewarms` 投递，或 GPU/NPU generation runner 的 `on_requests_added` / `run_idle_prefetch` 模型 hook。 | ||
| - 强制：payload 函数按 stage 显式 opt-in（MiniCPM-o 仅 Code2Wav 设置 `code2wav_prewarm_payload`）并在 stage client 暴露；orchestrator 只在初始 add 附带——streaming re-prewarm、session-owned 与 resumable 请求跳过——并把扁平 `{name: Tensor|scalar}` 展开为 `additional_information` 顶层 `_async_chunk_prewarm.<name>` 键使 tensor 能过序列化；函数失败或返回空时 placeholder 原样提交。scheduler 在 `add_request` 弹出这些键使其不进入 runner 的 per-request buffer，只对仍存活的 request id 经 `pending_request_prewarms` 恰好投递一次，finish/abort 丢弃未投递项。`run_idle_prefetch` 仅在「零 token step 且前一 step 也是零 token」时运行——有 chunk 调度、或 busy 后首个 idle step（输出可能仍在途）都不得运行；hook 异常只记日志（每个 hook 首次 warning），绝不传到 EngineCore；无 hook 的模型零影响。prefetch 在 forward 同一线程按排队 request 单 phase 执行，不自建线程/锁/side stream（CFM/HiFT CUDA graph 共享静态 buffer，`prepare_prompt` 设置进程全局默认 dtype）。chunk 0 必须仍自带 stage 所需全部输入。 |
There was a problem hiding this comment.
This rule combines configuration (StagePipelineConfig), Serving (orchestrator), Scheduler, Model Executor (GPU/NPU runners), and MiniCPM-o-specific Code2Wav behavior in one Scheduler page. Direct reviews route those paths to different owner/model entries, so changes confined to a runner, orchestrator, or Code2Wav may never load this rule. Split the conclusions into each nearest owner—especially moving the MiniCPM-o defaults and prefetch conditions to that model owner—while retaining only scheduler queue/delivery invariants here.
AGENTS.md reference: AGENTS.md:L12-L17
Useful? React with 👍 / 👎.
|
|
||
| # Async-chunk prewarm 规则 | ||
|
|
||
| ## VLLM-OMNI-PR8198-ASYNC-PREWARM — async-chunk prewarm payload 是四层同契约的提示，不是数据源 |
There was a problem hiding this comment.
The Scheduler owner entry states in components/scheduler/rules.md:12 that only SCHED-* identifiers are auditable, but this active rule is identified as VLLM-OMNI-PR8198-ASYNC-PREWARM. That prevents intake and review references from mapping the conclusion into the Scheduler rule namespace and makes its stable identity depend on the source PR; assign a SCHED-* ID and retain PR #8198 only as source metadata.
Useful? React with 👍 / 👎.
Generated automatically by the Copilot knowledge service (changeset
vllm-omni-intake-c9ef10a3152b).repos/vllm-omni/components/scheduler/rules-async-chunk-prewarm.mdGenerator:
zcode:GLM-5.3· judge:codex:gpt-6-sol:medium.The publisher merges it only if its local gate passes on the exact merge result.