---
title: '57786'
pageLayout: default
slug: thoughts/inference/vllm-prs/57786
permalink: https://aarnphm.xyz/thoughts/inference/vllm-prs/57786.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
# [#57786](https://github.com/aarnphm/aarnphm.github.io/issues/57786) \[Docs\] Add model recipes skill

[vllm-project/vllm#57786](https://github.com/vllm-project/vllm/pull/57786) · esmeetu · MERGED · created 2026-09-20T11:54:05Z · merged 2026-10-06T12:40:10Z

## Purpose

Add a `model-recipes` skill for finding how to run a model on specific hardware for a given task. It documents direct catalog queries, variant-specific JSON, task filters, and hardware/strategy command selection, including multi-node and Omni cases. Recipe creation and updates link to `vllm-project/recipes`’s existing `add-recipe` skill so authoring rules stay with their owning repository.

Duplicate check: searched open PRs for `recipe skill`, `model-recipes`, and `"recipes" "skill" in:title`. No matching skill PR was found. Related [#57307](https://github.com/aarnphm/aarnphm.github.io/issues/57307) implements runtime tuning/sweep tooling under `tools/recipes`, rather than agent guidance for catalog lookup. No issue is linked to this change.

AI assistance was used to write and test this skill. This draft is for human line-by-line review and rerunning the relevant checks before marking it ready.

## Test Plan

- `.venv/bin/pre-commit run --files .agents/skills/model-recipes/SKILL.md .agents/skills/model-recipes/agents/openai.yaml`
- `git diff --cached --check` before committing.
- Ran the skill-creator `quick_validate.py` against `.agents/skills/model-recipes` using `.venv/bin/python`.
- Executed the documented `curl`/`jq` lookup against `https://recipes.vllm.ai/models.json` and fetched `https://recipes.vllm.ai/Qwen/Qwen3.8-27B.json`.
- Exercised 10 read-only request scenarios: model-family lookup, exact FP8 checkpoint, provider listing, H100 selection, four-node topology, Omni text-to-video, missing model, recipe creation/update routing, and an unrelated job-run request. Checked four-node output against the recipes checkout’s command generator.

## Test Result

All applicable pre-commit hooks, skill validation, and whitespace checks passed. The current lookup examples return Qwen3.8-27B and its variants. Scenario checks resolved the intended checkpoint/hardware, distinguished generated commands from verified configurations, and routed authoring to the recipes repository. Four-node synthesis produced TP=32 and distinct ranks 0–3 for 4 × 8 H100.

Browser interaction testing could not complete because the browser session was unavailable. No GPU jobs or model evaluations were run; this change adds agent documentation and UI metadata only, with no runtime or model-output changes.

- [x] Human submitter has reviewed every changed line and rerun the relevant checks.

---

## comments

**sfeng33** (2026-09-30T02:48:06Z):

Probably worth one sentence saying this is lookup only: don’t launch `vllm serve` and don’t edit the recipe’s flags or parallelism. The openai.yaml default prompt (“find a recipe for running my model”) makes that ambiguous.

**benchislett** (2026-09-30T03:00:03Z):

Also worth a clear disclaimer that recipes are a validated starting point and not a perfect recipe: recipes are not always tuned for best performance and following a recipe is not a substitute for performance tuning

**esmeetu** (2026-09-30T03:42:39Z):

@benchislett @sfeng33 Thanks for the great review! just fixed all above comments. PTAL!

**esmeetu** (2026-10-02T00:53:46Z):

Thanks @benchislett @faradawn, all addressed.

**esmeetu** (2026-10-03T01:29:59Z):

/ci run

**github-actions** (2026-10-03T01:30:18Z):

✅ Triggered [Buildkite CI #92728](https://buildkite.com/vllm/ci/builds/92728) for commit `1fee2b7dc277`.

**vllm-agent** (2026-10-03T02:05:02Z):

<!-- ci-selector-shadow -->

### CI selector (shadow): 0 test steps (0 jobs) instead of 0 (0 jobs)

Shadow mode: this changes nothing about what CI runs. It shows what the evidence-based selector would pick for this PR, next to today’s rules. [How it works](https://github.com/vllm-project/ci-infra/tree/main/buildkite/ci_selector).

**Feedback welcome:** reply here if it would skip a step this change needs, or runs something unrelated.

| steps (jobs)           | Today’s rules | Selector | Would skip | Would add |
| ---------------------- | ------------- | -------- | ---------- | --------- |
| NVIDIA, CPU and others | 0 (0)         | 0 (0)    | 0 (0)      | 0 (0)     |
| AMD mirrors            | 0 (0)         | 0 (0)    | 0 (0)      | 0 (0)     |

<details><summary>Selector would run (0)</summary>

none

</details>

<details><summary>Would skip (today's rules run them) (0)</summary>

none

</details>

<details><summary>Would add (today's rules do not run them) (0)</summary>

none

</details>

<details><summary>AMD mirrors: would skip (0)</summary>

none

</details>

<details><summary>AMD mirrors: would add (0)</summary>

none

</details>

<sub>4 changed files · base `09e0ce3d82` · head `377edc3993` · Python record: build 93039 at `1e5d0ea888` · kernel record: table [`1e5d0ea`](https://github.com/aarnphm/aarnphm.github.io/commit/1e5d0ea888) (build 93039), map [`1e5d0ea`](https://github.com/aarnphm/aarnphm.github.io/commit/1e5d0ea888) · not counted: 10 build steps, 5 A100 steps the generator no longer emits, 0 optional steps the selector would also run</sub>

**esmeetu** (2026-10-03T06:01:18Z):

/ci retry

**github-actions** (2026-10-03T06:01:36Z):

✅ [Buildkite CI #92728](https://buildkite.com/vllm/ci/builds/92728) failed during CI setup, so its test failure set is incomplete. Use `/ci run` for the new commit.

**esmeetu** (2026-10-03T06:02:12Z):

/ci run

**github-actions** (2026-10-03T06:02:29Z):

❌ This PR is 5 commits behind upstream `main`. Your branch must contain every commit currently on upstream `main`. No new CI build was started. Merge or rebase onto the latest `main`, then rerun `/ci run`. To test this branch at your own risk, use `/ci run --allow-stale`.

<!-- vllm-ci-command:5966177913 -->

**esmeetu** (2026-10-06T11:13:46Z):

/ci run

**github-actions** (2026-10-06T11:14:07Z):

✅ Triggered [Buildkite CI #93078](https://buildkite.com/vllm/ci/builds/93078) for commit `377edc3993b3`.

