Skip to content

knowledge(vllm-omni): 1 change(s) from refine #414

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 25, 2026
Merged

knowledge(vllm-omni): 1 change(s) from refine #414

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
13 changes: 11 additions & 2 deletions knowledge/repos/vllm-omni/models/z-image/rules.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,10 +1,10 @@
---
title: "Z-Image TeaCache 校准合同"
created: 2026-10-09
updated: 2026-10-09
updated: 2026-10-25
type: rule
tags: [vllm-omni, models]
sources: ["PR #8408"]
sources: ["PR #8408", "PR #3283"]
confidence: high
---

Expand All @@ -22,3 +22,12 @@ confidence: high
- 强制：使用 Z-Image-Turbo 自身拟合系数，不借用 Qwen-Image 系数；现有拟合测量域为 input relative L1 0.024–0.336，threshold 默认 0.2。同 pipeline 类的其他 checkpoint 不自动获得校准保证。DataCollectionHook 初始化时解析并校验 extractor，未知模型在 forward 前失败。
- 禁止：把 polynomial 在测量域外的行为当作安全保证，或把 borrowed coefficients 的功能可跑说成可靠加速；从够步数模型的收益外推 distilled 短步数。
- 验收：测试系数与 Qwen 不同、默认 config/threshold 和 adapter 注册；更换 checkpoint/输入域/threshold 时重做同 workload 的质量与命中率校准，记录域外风险及 hardware/head。 ^[PR #8408]

## VLLM-OMNI-PR3283-ZIMAGE-SERVING — Z-Image-Turbo 在线服务只按已验证配置声明，加速收益不得外推

@chatgpt-codex-connector chatgpt-codex-connector Bot Oct 25, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P1 Badge Update the owner index to expose the new serving rules

When a task concerns Z-Image serving, quantization, or image editing, the nearest _index.md still describes this page only as “Z-Image TeaCache 校准合同” and lists no serving-related lookup scenario. This commit expands the page beyond that advertised scope without updating the index, leaving the new contract undiscoverable through its documented topic routing. Update the owner index to advertise the serving rules alongside TeaCache; root AGENTS.md explicitly requires synchronizing the nearest index for current rules pages.

AGENTS.md reference: AGENTS.md:L23-L24

Useful? React with 👍 / 👎.


- 触发：修改 Z-Image-Turbo 的 serving/recipe 文档、`--omni`、`--quantization`、`--cache-backend` 相关默认值或说明，或 `/v1/images/generations`、`/v1/images/edits` 的行为声明。
- 强制：以已验证面为基准——1x RTX 5880 Ada 48GB（CUDA cu130 栈、vLLM 0.19.0）上 `vllm serve --omni` 在线图像生成与图像编辑均验证通过，`--omni` 必需；未量化峰值 24987 MB，`--quantization fp8` 在全部测试组合把峰值压到约 17.9–19.8 GB，是主要时延杠杆（生成 ~9500ms→~5600ms，编辑 ~34500ms→~8200ms）。recipe 的示例请求显式设置 `num_inference_steps=9` 与 `guidance_scale=0`，recipe 声明 Z-Image-Turbo 为 distilled 少步数模型、默认不启用 CFG。
- 禁止：把 `num_inference_steps=9` 说成引擎或模型的默认步数——它只是 recipe 示例请求显式传入的参数；把 `tea_cache`/`cache_dit` 当作少步数生成的提速项——9 步生成实测 ~9300ms，与基线 ~9500ms 近持平，量级收益只出现在编辑路径（~34500ms→~11000ms）；把仅 `1024x1024`、单机单驱动测得的参考值外推成其他尺寸/硬件的保证。
- 验收：改动服务默认值、量化或 cache 路径时，文档须附同硬件的峰值显存与时延表，并保留“单机参考值，随驱动/预热/尺寸/payload 变化”的限定；涉及步数或 CFG 默认行为的声明须能指回已验证的示例或代码默认值；新增尺寸或硬件先验证再声明；TeaCache 系数校准约束另见 ZIMAGE-1a。 ^[PR #3283]

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