| name | cookbook-write-narrative |
| description | Write optional Zhihu draft for vllm-omni-cookbook model folders. Use after retro tables exist in index.md, for Chinese republishing of performance ledgers. |
Write Zhihu Draft (Optional)
Only after index.md retro tables are complete. The ledger stays authoritative; the Zhihu draft explains how and why for a Chinese audience.
Two layers
| File | Audience | Content |
|---|
index.md | Maintainers, quick lookup | Tracks, tables, commands, links |
{model}-performance-zhihu.md | Zhihu / 中文 | Paste-friendly narrative |
Do not duplicate full PR lists from index.md — summarize and link instead.
Zhihu draft rules
Reference: diffusion/wan2.2/wan22-i2v-performance-zhihu.md
<!--
标题建议:...
话题:#vLLM #扩散模型 ...
-->
# {中文标题}
**模型:** ... **环境:** ... **指标:** ...
- Full Chinese prose; keep English proper nouns (vLLM-Omni, USP, DiT, TTFP)
- ASCII charts OK; no HTML/Mermaid
- Tables: author copies to 飞书/Excel → 知乎 for formatting
- 置顶评论 template with GitHub links to cookbook + vllm-omni JSON
- After publish: add live Zhihu URL to
index.md header and zhihu md footer
- Delete the HTML comment block at the top before publishing
Wire up index.md header
**Zhihu draft (中文):** [wan22-i2v-performance-zhihu.md](wan22-i2v-performance-zhihu.md)
Optimization narrative structure
- Map releases → PRs (search vllm-omni release notes + perf PRs)
- Classify by pipeline stage: API/IPC → preprocess → VAE → DiT → runtime
- Tie measured deltas to stages using profiler logs when available
- Separate CI H100 thresholds from local retro numbers — never conflate