Repository navigation
knowledge: daily intake batch 09 - #140
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: 1b1387cc0f
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".
| HWR and checkpoint mmap；另注入 unregistration failure，断言 retryable retention、backend drop 和 | ||
| distributed teardown ordering。^[PR #6486] ^[PR #6591] | ||
|
|
||
| ## DIFF-2y — batch-first sampling gate 的 provided 标志必须进入 RequestBatchSamplingParamsKey |
There was a problem hiding this comment.
DIFF-2y already identifies the Diffusers shard-manifest contract in rules-checkpoint-loading.md:171, so searches and Direct routing now return two unrelated rules for the same auditable ID. Give this batching rule a unique ID and update its routing entries accordingly.
AGENTS.md reference: knowledge/AGENTS.md:L16-L16
Useful? React with 👍 / 👎.
| - 禁止：Worker 自行分配或释放 Scheduler-owned logical blocks；把部分 row 安装、native append/apply 异常或 stale snapshot 当作成功；从 padding 猜异构 prefix 的逻辑布局；在清理或 wake refresh 后继续复用旧的 native metadata/buffer view。 | ||
| - 验收：CPU/mock 覆盖 rank-local config、请求/上下文 row mapping、重复/过期/冲突 generation、非法 block 与容量、append/apply rollback、幂等清理和 wake refresh；GPU/NPU 另验证非连续 BlockTables、32Q/8KV GQA、异构 prefix 与 piecewise attention。Hunyuan paged request execution 不证明 `denoise_step`、cross-request prefix reuse、AR KV、negative CFG、Ring 或 AllGather。^[PR #6102] ^[PR #6563] | ||
|
|
||
| ## DIFF-4w — 声明为 str 的 diffusion cache_backend 必须在 kwargs 边界把显式 None 规范成 none |
There was a problem hiding this comment.
DIFF-4w already labels the worker-title/log-prefix contract in rules-worker-observability.md:13, and the shared Direct table explicitly routes that ID to worker observability. Reusing it for cache normalization makes exact-ID routing ambiguous, so this rule needs a new unique ID plus corresponding route updates.
AGENTS.md reference: knowledge/AGENTS.md:L16-L16
Useful? React with 👍 / 👎.
| - 强制：NPU 路径必须调用 `self.gate_up_proj(x)`（或等价模块 `__call__`）得到 packed gate/up，再交给融合 SwiGLU 与既有 `down_proj`；保留 packed 一次投影与融合激活，但不得绕过 linear 模块的量化 dispatch。 | ||
| - 禁止：`F.linear(x, self.gate_up_proj.weight)` 直接读 weight，使 INT8 `matB` 落入不支持的普通 matmul；把该要求推广到 H3 DiT 或其他模型而未单独验证。 |
There was a problem hiding this comment.
For the same MiniMax-H3 Qwen3-VL NPU text-MLP trigger, existing MMH3-4a still requires F.linear(x, self.gate_up_proj.weight), while this new rule requires the module call and explicitly forbids that expression. A future implementation cannot satisfy both active contracts; merge the quantization-aware requirement into MMH3-4a or explicitly replace its obsolete text rather than adding a contradictory rule.
AGENTS.md reference: AGENTS.md:L7-L11
Useful? React with 👍 / 👎.
| - 触发：修改 `AsyncOmniEngine`、headless stage 启动、`StageConfigFactory` 公开入口，或 registry/deploy/CLI 到最终 stage 的 source selection。 | ||
| - 强制：标准与 headless 生产路径只调用 `resolve_omni_config()` 获得 `OmniConfigResolution`（structured config、runtime compatibility stages、strategy-derived LB、effective deploy path）；factory 的 create/legacy 方法是 resolver 与测试的下层 primitive，engine/entrypoint 不得绕过 resolver 直接选源。通用单 stage diffusion 仅在 registry 解析失败后作为 fallback；legacy `stage_configs_path`/`stage_configs` 必须显式报错。 | ||
| - 禁止：engine 侧二次合成 default diffusion stage 或重复 merge LoRA/attention/quant/cache/profiler；把 OmegaConf-compatible stages 宣传为稳定公共 ABI；静默恢复已删除的 stage-configs-path 行为。 | ||
| - 验收：同一 model+deploy+override 在标准与 headless 路径得到等价 topology 与 LB policy；覆盖 registry 命中、generic diffusion fallback、以及拒绝已删除 legacy 参数。^[PR #5140] |
There was a problem hiding this comment.
This rule cites PR #5140, but the page's frontmatter sources: remains unchanged; the same omission affects newly cited PRs #7078, #7195, #7041, #7071, #7191, #6876, and #7228 in the other edited pages. The knowledge schema requires page-level evidence in sources:, so prose-only footnotes lose structured provenance; add each PR to the corresponding page metadata.
AGENTS.md reference: AGENTS.md:L7-L10
Useful? React with 👍 / 👎.
| - 禁止：通过 `VLLM_USE_FLASHINFER_MOE_FP16` 或其他 process-wide 环境变量选择 backend；把默认值写成覆盖用户/deploy 显式选择；只改 deploy YAML 却不覆盖 legacy 与 typed builder 的最终 engine args。 | ||
| - 验收：以 Qwen3-Omni default 和一个显式 backend control，分别经 `build_legacy_engine_args_dict` 与 `build_engine_args_dict_from_omni_stage_config` 断言最终 `moe_backend`；默认两路径均为 `triton`，显式值不变。helper 单测只能补充分支覆盖，不能代替 builder-level assertion。PR 只提供 compile/Ruff/diff 检查及因缺少 `torch` 未运行的 targeted pytest，故不构成 GPU 稳定性、性能或音频质量证据。^[PR #7019] | ||
|
|
||
| ## QOMNI-1g — 共享 code predictor 的 RoPE theta 必须先读 rope_parameters |
There was a problem hiding this comment.
The commit edits eight current rule pages without updating any nearest _index.md; concretely, models/qwen-omni/_index.md:63 still omits code-predictor/RoPE-layout work, and this page's Direct table at line 17 ends at QOMNI-1f. Because reviews stop after the compact route, QOMNI-1g will not be discovered for its stated trigger; update the owner indexes and affected quick maps for the complete batch.
AGENTS.md reference: AGENTS.md:L23-L24
Useful? React with 👍 / 👎.
Human curation of the bot's batch, addressing the Codex review: - `DIFF-2y` -> `DIFF-2ac` and `DIFF-4w` -> `DIFF-4ac`: both IDs already head unrelated rules on sibling diffusion pages (checkpoint-loading, worker-observability); exact-ID routing must stay unambiguous. - MMH3-4e contradicted MMH3-4a on the same NPU text-MLP trigger (`F.linear` on the raw weight vs. the module call). Fold the quantization-dispatch requirement from PR #6876 into MMH3-4a and drop the duplicate section. - Add each cited PR to its page's `sources:` frontmatter. - Route the new rules: Direct table rows for `DIFF-2ae`, `MCPMO-2e`, `MMH3-4d` (plus the existing `MMH3-4a`/`4b`), `QOMNI-1g`, and the nearest `_index.md` rows for every touched page. Validators: check_knowledge_tree 0 errors, check_wiki_lint 0 errors, test_knowledge_source green. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018MAFmwP3u5qAMDgHfM6LFi
|
Human curation pushed as 03b0b62 on top of the bot's commit, addressing every review thread:
Validators: check_knowledge_tree 0 errors, check_wiki_lint 0 errors, test_knowledge_source green. The failing @codex review |
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
The batch was distilled on the SDK pinned before #142 and against a main that predates #140/#145, so it reused two IDs those batches landed and its audit ran against the retired baseline. Merge main in and curate: - VOMNI-CFG-1s (CLI-only negative alias) -> VOMNI-CFG-1t and MCPMO-2e (audio_embeds registration) -> MCPMO-2f: both IDs already head other rules on main; every ID now heads exactly one section tree-wide. - The configuration entry page crossed the 32 KiB split cap once main's VOMNI-CFG-1s and this batch's 1t both sat on it. Both engine-args rules move to the existing topic page rules-legacy-engine-args.md (the owner's engine-args topic), the entry page's Direct table links there, and the nearest _index.md row is extended. No inbound anchors referenced them. - Every cited PR added to its page's `sources:` frontmatter; Direct table rows for OMNI-CI-2g/2h, DIFF-1af, MCPMO-2f, VOMNI-CFG-1t and _index.md rows for benchmark, ci, configuration and diffusion. The review gate also noted that the SDK deduplicates heading IDs within a single section, so a section repeating its own ID in a nested heading is not rejected; that is an SDK change and ships separately. Validators: check_knowledge_tree 0 errors, check_wiki_lint 0 errors, test_thin_mcp_server and test_knowledge_source green. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018MAFmwP3u5qAMDgHfM6LFi
Daily knowledge-intake batch from omni-reviewbot: executable rules distilled from merged PRs and copilot bugfix runs, append-only per the knowledge contract. Merging this PR is the human promotion gate — review the rules like any other knowledge edit.
Proposed rules
knowledge/repos/vllm-omni/components/diffusion/rules-component-lifecycle.md→DIFF-2y(sources: PR #7078)knowledge/repos/vllm-omni/components/diffusion/rules-system-runtime.md→DIFF-4w(sources: PR #7041)knowledge/repos/vllm-omni/components/diffusion/rules-lora.md→DIFF-2ae(sources: PR #7195)knowledge/repos/vllm-omni/models/minimax-h3/rules-vae-ops.md→MMH3-4d(sources: PR #7191)knowledge/repos/vllm-omni/models/minimax-h3/rules.md→MMH3-4e(sources: PR #6876)knowledge/repos/vllm-omni/models/minicpm-o-4-5/rules.md→MCPMO-2e(sources: PR #7071)knowledge/repos/vllm-omni/components/configuration/rules-parallel-topology.md→CONF-4f(sources: PR #5140)knowledge/repos/vllm-omni/models/qwen-omni/rules.md→QOMNI-1g(sources: PR #7228)Source events
Public SDK curation result
{ "batch_id": "sha256:5b0a4b8c946decd23172a58464b265fb774ffffc089007a62431c2a008f438cd", "success": true, "attempted": 8, "applied": 8, "accepted_indexes": [ 0, 1, 2, 3, 4, 5, 6, 7 ], "rejected_indexes": [], "updated_document_ids": [ "knowledge/repos/vllm-omni/components/configuration/rules-parallel-topology.md", "knowledge/repos/vllm-omni/components/diffusion/rules-component-lifecycle.md", "knowledge/repos/vllm-omni/components/diffusion/rules-lora.md", "knowledge/repos/vllm-omni/components/diffusion/rules-system-runtime.md", "knowledge/repos/vllm-omni/models/minicpm-o-4-5/rules.md", "knowledge/repos/vllm-omni/models/minimax-h3/rules-vae-ops.md", "knowledge/repos/vllm-omni/models/minimax-h3/rules.md", "knowledge/repos/vllm-omni/models/qwen-omni/rules.md" ], "updated_on": "2026-09-09", "validators": [ { "validator_id": "knowledge/tools/check_knowledge_tree.py", "passed": true, "status": "passed", "returncode": 0, "output": "提醒：文件接近拆分线：general/review/guides/review-execution-contract.md (183 个非空行，21548 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/ci/rules.md (235 个非空行，26002 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/configuration/rules.md (174 个非空行，30362 bytes)\n提醒：目录有 15 个普通页面，应该考虑分类：repos/vllm-omni/components/configuration\n提醒：文件接近拆分线：repos/vllm-omni/components/diffusion/rules-attention.md (179 个非空行，22696 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/diffusion/rules-checkpoint-loading.md (206 个非空行，25632 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/diffusion/rules-output-lifecycle.md (142 个非空行，19794 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/diffusion/rules-system-runtime.md (179 个非空行，24707 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/diffusion/rules.md (172 个非空行，27349 bytes)\n提醒：目录有 22 个普通页面，应该考虑分类：repos/vllm-omni/components/diffusion\n提醒：文件接近拆分线：repos/vllm-omni/components/model-executor/rules-bridge-batch.md (134 个非空行，18619 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/model-executor/rules.md (193 个非空行，31389 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/scheduler/rules.md (299 个非空行，32496 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/serving/rules-engine-lifecycle.md (246 个非空行，32271 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/components/serving/rules.md (196 个非空行，31197 bytes)\n提醒：目录有 12 个普通页面，应该考虑分类：repos/vllm-omni/components/serving\n提醒：文件接近拆分线：repos/vllm-omni/models/hunyuan-image3/rules.md (131 个非空行，24302 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/models/minicpm-o-4-5/rules.md (252 个非空行，29421 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/models/minimax-h3/rules-cache-task.md (190 个非空行，25247 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/models/minimax-h3/rules-deployment.md (236 个非空行，27707 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/models/minimax-h3/rules.md (187 个非空行，32635 bytes)\n提醒：文件接近拆分线：repos/vllm-omni/models/qwen3-tts/rules.md (190 个非空行，20981 bytes)\n知识目录检查通过：0 个错误，22 个提醒" }, { "validator_id": "knowledge/tools/check_wiki_lint.py", "passed": true, "status": "passed", "returncode": 0, "output": "wiki lint 通过：0 个错误，0 个提醒（共 387 页）" } ], "rolled_back": false }The vllm-omni release audit runs in CI on this PR as usual.