vllm-omni-cookbook

name	vllm-omni-cookbook
description	Contribute release-by-release performance summaries to vllm-omni-cookbook: per-model improvement ledgers (index.md) and cross-model release tables (SUMMARY.md). Use when documenting what improved at vX.Y.Z, updating perf deltas, or adding a new model ledger.

vLLM-Omni Performance Cookbook — Contributor Guide

This repo summarizes model performance improvements at every even stable release — not deployment how-tos. vLLM-Omni owns recipes and benchmarks; here we publish what changed, by how much, and why (PR links).

Even releases only: add cookbook sections for v0.18.0, v0.20.0, v0.22.0, … Skip odd minors. Deltas compare to the previous even release.

Each {category}/{model}/index.md is the model's full improvement timeline. SUMMARY.md is the per-release cross-model snapshot.

Before you start

1. Read README.md and docs/DESIGN.md.
2. Do not fork upstream perf JSON here — link or copy configs under vllm-omni/benchmark_results/.
3. Every comparison must record: commit SHAs, vllm + vllm_omni versions, GPU SKU/count, serve args, workload dimensions.
4. Retro hardware on this cluster is NVIDIA H200 (internal logs may label it L20X — same SKU; document as H200 in the cookbook).

Choose a workflow

Goal	Skill
First-time model in the cookbook	cookbook-add-model
New even stable release (v0.22.0, …)	cookbook-add-release
Optional Zhihu draft	cookbook-write-narrative

Category → metrics

Category	Folder	Primary metrics
Omni	omni/	TTFT, TTFP, TPOT, RTF, E2EL, throughput
Diffusion	diffusion/	E2E latency (latency_mean), optional stage profiler

Repo layout

vllm-omni-cookbook/
├── SUMMARY.md                 # per-release cross-model snapshot
├── omni/{model}/index.md
└── diffusion/{model}/
    ├── index.md               # slim reference ledger
    └── *-performance-zhihu.md # optional Chinese draft

Canonical examples: diffusion/wan2.2/, omni/qwen3-omni/.

PR checklist

• Numbers come from a completed retro run or CI JSON (not estimates)
• Footnotes explain non-comparable rows (missing CLI, different GPU count, sync vs async API)
• SUMMARY.md updated if adding a release section
• Links to upstream recipe, perf JSON, and relevant PRs
• No duplicate full PR lists in both index.md and Zhihu draft

Templates

See references/templates.md for index.md, SUMMARY.md, and PR body templates.