| 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
- Read README.md and docs/DESIGN.md.
- Do not fork upstream perf JSON here — link or copy configs under
vllm-omni/benchmark_results/.
- Every comparison must record: commit SHAs,
vllm + vllm_omni versions, GPU SKU/count, serve args, workload dimensions.
- 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
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
Templates
See references/templates.md for index.md, SUMMARY.md, and PR body templates.
