update-vllm-nightly

name	update-vllm-nightly
description	List available vLLM nightly builds and update config/models.yaml to pin a specific version. Use when user wants to browse nightlies, pin a specific vLLM version, or update vLLM nightly config.
user-invocable	true

Update vLLM Nightly Version

This skill lists available vLLM nightly builds and helps pin a specific version in config/models.yaml.

Step 1: Fetch Top 5 Docker Tags

curl -s 'https://hub.docker.com/v2/repositories/vllm/vllm-openai/tags?page_size=10&name=nightly-' | \
  jq -r '.results[] | select(.name | test("^nightly-[a-f0-9]{40}$")) | [(.name | ltrimstr("nightly-")), (.last_updated | split("T")[0])] | @tsv' | head -5

This outputs: {commit_hash}\t{push_date} for the 5 most recent nightly builds.

Step 2: Fetch Commit Details via gh CLI

For each of the 5 commit hashes, fetch the commit message using gh (authenticated, handles pagination):

gh api repos/vllm-project/vllm/commits/{COMMIT_HASH} --jq '[.sha[0:8], (.commit.author.date | split("T")[0]), (.commit.message | split("\n")[0] | .[0:60])] | @tsv'

Run these 5 requests in parallel. Each returns: {short_sha}\t{date}\t{message}

Step 3: Check cu130 Availability

Only check the top 5 commits (the ones most likely to be selected). Make these requests in parallel:

GET https://wheels.vllm.ai/{commit_hash}/cu130/

If returns 200, mark as "Y". If 404 or error, mark as "N". For commits not checked, mark as "?" (can check on demand if user selects).

Step 4: Display Results

Show a formatted table:

Recent vLLM Nightly Builds

#   COMMIT    DATE        CU130  MESSAGE
1   da6709c9  2025-01-10  Y      fix: memory leak in scheduler
2   72d9c316  2025-01-09  Y      feat: add flashinfer backend
3   abc12345  2025-01-08  N      refactor: cleanup unused imports
...

Current config: defaults.backends.vllm.backend_version = v0.13.0

Step 5: Interactive Selection

Use AskUserQuestion to ask the user:

Question 1: "Which version would you like to pin?"

• Options: Show top 4 commits as options (e.g., "1. da6709c9 - fix: memory leak")
• Include "Browse only (no update)" option

If user selects "Browse only", stop here.

Question 2: "Where should this version be applied?"

• "Global default" - Updates defaults.backends.vllm
• "Specific model" - Ask which model, then update that model's config

Step 6: Update Config

Read config/models.yaml and update the appropriate section. Use the commit hash directly (cu130 wheels are the default).

For Global Default:

defaults:
  backends:
    vllm:
      backend_version: {full_commit_hash}  # Update this line

For Model-Specific:

Add or update the model's profiles section:

models:
  - repo_id: org/model-name
    profiles:
      default:
        backend: vllm
        backend_version: {full_commit_hash}

Step 7: Confirm

After updating, show the user what was changed:

Updated config/models.yaml:
  defaults.backends.vllm.backend_version: da6709c9fe6965b7348692576ffadeee8439388e

To apply this change, rebuild images with:
  uv run python scripts/run_bench.py --model ~/models/... --rebuild