Skip to content

knowledge(vllm-omni): 1 change(s) from refine #408

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
tzhouam merged 1 commit into from
Oct 24, 2026
Merged

knowledge(vllm-omni): 1 change(s) from refine #408

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions knowledge/repos/vllm-omni/models/minimax-h3/_index.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -172,3 +172,4 @@ CPU-only validation limit 见 [encoder state rules](rules-encoder-state.md)。

- [MiniMax-H3 VDNH3 checkpoint 与请求合同](rules-vdnh3.md)。
- [MiniMax-H3 audio VAE 确定性规则](rules-audio-vae-determinism.md)
- [MiniMax-H3 ComfyUI audio_mode 与 supplied-audio MV 规则](rules-comfyui-audio.md)
19 changes: 19 additions & 0 deletions knowledge/repos/vllm-omni/models/minimax-h3/rules-comfyui-audio.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,19 @@
---
title: "MiniMax-H3 ComfyUI audio_mode 与 supplied-audio MV 规则"
created: 2026-10-24
updated: 2026-10-24
type: rule
tags: [vllm-omni, models, diffusion]
sources: ["PR #9226"]
---

# MiniMax-H3 ComfyUI audio_mode 与 supplied-audio MV 规则

## VLLM-OMNI-PR9226-H3-AUDIO-LOCK — ComfyUI H3 audio_mode 必须默认省略并绑定单一驱动音频

- 触发：修改 `apps/ComfyUI-vLLM-Omni` 中 MiniMax-H3 Video Params 的 `audio_mode` 下拉、`_minimaxh3_params_builder` 的 extra_params 组装、Video References 的音频 reference 计数、H3 example workflow 模板或 supplied-audio Music Video workflow。
- 强制：`audio_mode` 下拉默认 `native`；`_minimaxh3_params_builder` 的实现合同是 pop `audio_mode` 时缺省即 `native`，任何不等于 `native` 的值都写入 `extra_params`（当前下拉唯一非默认选项是 `lock_source`，但转发条件是“非 native”，不是“等于 lock_source”），因此默认值不进入请求，既有 graph 与请求 payload 保持不变。`lock_source` 语义按本 PR shipped 文档：恰好的单一音频作为驱动音轨进入 target audio latent（crop/zero-pad 到 shot 时长并在采样期间保持固定），不计入 `<Audio N>` reference 计数；非恰好一个音频时的拒绝由 server 执行。向已发布节点新增 widget 后，所有 shipped H3 模板必须带新 widget 值（如 `"native"`）重存，因为 ComfyUI 按 position 恢复 widgets。MV 模板中每个 shot 是 Ref2VA：共享 character/style image 作 `image_1`（`<Picture 1>`）、同一 track 的连续 124 帧 slice 作 `audio_1`；Concatenate Video 的 `complete_audio` 不是直连整条原始 track，而是同一 supplied track 经 Trim Audio Duration 从相同起点裁剪到全部 shot 总时长（模板 3 shot 即 3 × 5.1667 s = 15.5 s）的输出，用于替换各 shot 生成的音频（要求 ComfyUI >= v0.36.0）。
- 禁止：让默认值改变既有请求 payload；把 `lock_source` 驱动音频当作普通 `<Audio 1>` reference tag 计数；新增 widget 却不重存使用该节点的 shipped 模板；在 shot 边界重新生成音频、留下音轨接缝，或以各 shot 生成的音频拼接充当最终音轨。不得把“客户端在请求发出前 fail-fast 拒绝非恰好一个音频”写成已存在的行为或验收项：所示 diff 未建立客户端校验，shipped 文档写明由 server 拒绝，`test_video_references.py` 的新增只检查 forwarding。
- 验收：builder/integration 测试断言默认 `native` 被省略、非 `native` 值进入 `extra_params` 并出现在发往 server 的请求中；`test_h3_music_video_workflow.py` 结构测试须保持每个 shot 均为 Ref2VA、由同一 track 的连续 124 帧（5.1667 s @ 24 fps，满足 H3 `17k+5` 帧约束）slice 驱动、shot 时长与 sampling/H3 设置跨 shot 共享，并断言 Concatenate Video 的 `complete_audio` 输入来自裁剪 supplied track 的 Trim Audio Duration 节点、其时长等于全部 shot 的总时长，而不是各 shot 生成的音频（CI markers `core_model and cpu`）。未来 PR 若声称 ComfyUI 客户端实现了 exactly-one-audio 拒绝，必须附带客户端拒绝测试，不得以 forwarding 测试充当。^[PR #9226]

<!-- kb:rule status=active since=v0.30.0 -->
Loading