Repository navigation
knowledge(vllm-omni): 2 change(s) from refine - #343
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: 801c3a1df5
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".
| # 数值精确性与平台可移植性规则 | ||
|
|
||
| ## VLLM-OMNI-PR8562-TRITON-HIP-RN — Triton libdevice 显式舍入算子必须按平台编译期分支 |
There was a problem hiding this comment.
This page combines a MOSS-specific _rotate/HIP contract with a Breeze-specific graph-oracle contract under a generic numeric-exactness topic because both came from PR #8562. The existing MOSS owner is knowledge/repos/vllm-omni/models/moss-tts/, and the existing Breeze platform topic is rules-breeze-rocm.md; owner-routed reviews can therefore miss these rules or load only part of the applicable contract. Move or merge each conclusion into its nearest existing owner rule rather than creating this cross-model page.
AGENTS.md reference: AGENTS.md:L12-L17
Useful? React with 👍 / 👎.
| - 触发：在 `vllm_omni/model_executor/models/` 下的 Triton kernel（如 `moss_tts/first_frame_special.py` 的 `_rotate`）调用 CUDA libdevice 内建（`add_rn`/`mul_rn` 等显式舍入算子），或修改这类 kernel 的跨平台数值路径（本例的修复对象是 AMD READY R2-01 暴露的 MOSS first-frame attention 失败）。 | ||
| - 强制：CUDA-only 的 libdevice 算子必须在 Triton 编译期用单一 `tl.constexpr` 按平台选择（PR 的形态：模块级 `_CUDA_RN_OPS = current_omni_platform.is_cuda()` 作为 constexpr 默认参数）：NVIDIA CUDA 保留显式 round-to-nearest 路径，非 CUDA 平台（含 HIP/ROCm）改走可移植的 FP32 乘加表达式（`value_fp32 * cosine + signed_partner * sine`）；与原始实现（reference attention）的既有数值对照不得因平台分支而放松。 | ||
| - 禁止：让 HIP 平台编译进 Triton HIP mapping 中不存在的 libdevice 调用（PR 记录 `add_rn` 与配对的显式 RN 算子在 HIP mapping 均无映射）；把平台判断做成 kernel 内运行时分支或多个互相不一致的常量。 | ||
| - 验收：审查者核对平台选择是单一编译期常量、CUDA 与可移植两种表达式在 kernel 内显式分支、既有数值对照未放松，并要求申请者分别给出两条平台路径的数值测试和修复后目标平台的 CI 通过证据。PR #8562 自身的证据不满足后两项：所示 diff 除 `_rotate` kernel 外还改了两个测试文件——`test_audio_embed_kernel.py` 把随机 reduction 断言放宽为默认 `assert_close` 并新增可精确表示的 bitwise 用例、`test_breeze_tts_2_graphs.py` 重写 Breeze graph oracle——但所示改动都在 CUDA 上运行，不触及 `_rotate` 的任何平台分支；PR body 明确保持 R2-01 `NonBlocking` 并把绿色 soak 推迟到后续 promote PR，证据未附修复后的 AMD CI 通过记录。^[PR #8562] |
There was a problem hiding this comment.
The acceptance clause persists PR #8562's changed-test inventory, body status, and missing AMD CI result. Those are transient review/result history rather than a reusable trigger, requirement, prohibition, or acceptance condition, and they will become stale if the deferred promotion later supplies the missing evidence. Keep the platform-test and target-CI requirements, retain the PR only as provenance, and drop the replay-style account of what that PR did not prove.
AGENTS.md reference: AGENTS.md:L12-L17
Useful? React with 👍 / 👎.
| sources: ["PR #8562"] | ||
| --- | ||
|
|
||
| # 数值精确性与平台可移植性规则 |
There was a problem hiding this comment.
The new rule page contains no relative Markdown links, while doc/knowledge/SCHEMA.md:94-96 requires every sedimented knowledge page to link at least two related pages. The ^[PR #8562] provenance markers are not links, so readers cannot navigate from this page to the existing MOSS, Breeze, or platform contracts; add at least two relevant owner-page links.
AGENTS.md reference: AGENTS.md:L7-L10
Useful? React with 👍 / 👎.
|
Knowledge quality gate · checked 2026-10-10 at Result: not passed. This PR will not be merged until the problems below are fixed; push a fix and it is checked again at the next daily pass. Problems:
|
Generated automatically by the Copilot knowledge service (changeset
vllm-omni-refine-25fee39dda30).repos/vllm-omni/components/model-executor/rules-numeric-exactness.mdrepos/vllm-omni/components/model-executor/rules-numeric-exactness.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.