update-sglang-nightly

6's avatarfrom 6

List available SGLang nightly builds and update config/models.yaml to pin a specific version. Use when user wants to browse nightlies, pin a specific SGLang version, or update SGLang nightly config.

⭐1stars🔀0forks📁View on GitHub🕐Updated Jan 11, 2026
[Developer Tools]

When & Why to Use This Skill

The update-sglang-nightly skill automates the management of SGLang inference engine versions. It streamlines the process of fetching the latest CUDA 13 nightly builds from Docker Hub and provides an interactive interface to update the `config/models.yaml` file. By enabling precise version pinning for global defaults or specific models, it helps developers maintain stable, reproducible, and optimized LLM deployment and benchmarking environments.

Use Cases

  • Version Pinning for Reproducibility: Pin a specific SGLang nightly build in your configuration to ensure consistent performance results across different benchmarking sessions.
  • Nightly Build Exploration: Quickly list and browse the 10 most recent SGLang CUDA 13 nightly tags, including their release dates, without leaving the terminal or manually checking Docker Hub.
  • Model-Specific Configuration: Update the backend version for a specific LLM model in `models.yaml` to test new features or bug fixes available in a particular nightly release.
  • Automated Environment Setup: Streamline the workflow of keeping the SGLang serving engine up-to-date, replacing manual curl commands and YAML editing with a guided, error-free process.
nameupdate-sglang-nightly
descriptionList available SGLang nightly builds and update config/models.yaml to pin a specific version. Use when user wants to browse nightlies, pin a specific SGLang version, or update SGLang nightly config.
user-invocabletrue

Update SGLang Nightly Version

This skill lists available SGLang CUDA 13 nightly builds and helps pin a specific version in config/models.yaml.

Step 1: Fetch Recent CUDA 13 Nightly Tags

curl -s 'https://hub.docker.com/v2/repositories/lmsysorg/sglang/tags?page_size=20&name=nightly-dev-cu13' | \
  jq -r '.results[] | select(.name | startswith("nightly-dev-cu13-")) | [.name, (.last_updated | split("T")[0])] | @tsv' | head -10

This outputs: {tag}\t{push_date} for the 10 most recent CUDA 13 nightly builds.

Step 2: Display Results

Show a formatted table:

Recent SGLang CUDA 13 Nightly Builds

#   TAG                                      DATE
1   nightly-dev-cu13-20260110-8a1be0dc       2026-01-10
2   nightly-dev-cu13-20260109-abc12345       2026-01-09
3   nightly-dev-cu13-20260108-def67890       2026-01-08
...

Current config: defaults.backends.sglang.backend_version = nightly-dev-cu13-20260110-8a1be0dc

Step 3: Interactive Selection

Use AskUserQuestion to ask the user:

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

  • Options: Show top 4 tags as options (e.g., "1. nightly-dev-cu13-20260110-8a1be0dc")
  • 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.sglang
  • "Specific model" - Ask which model, then update that model's config

Step 4: Update Config

Read config/models.yaml and update the appropriate section.

For Global Default:

defaults:
  backends:
    sglang:
      backend_version: {selected_tag}  # Update this line

For Model-Specific:

Add or update the model's profiles section:

models:
  - repo_id: org/model-name
    profiles:
      default:
        backend: sglang
        backend_version: {selected_tag}

Step 5: Confirm

After updating, show the user what was changed:

Updated config/models.yaml:
  defaults.backends.sglang.backend_version: nightly-dev-cu13-20260110-8a1be0dc

To apply this change, run:
  uv run python scripts/run_bench.py --model ~/models/... --backend sglang

Notes

  • CUDA 13 nightlies use the tag pattern: nightly-dev-cu13-{YYYYMMDD}-{short_hash}
  • For stable releases, use tags like v0.5.7-cu130-runtime or latest-cu130-runtime
  • The short hash is derived from the SGLang commit (not the full SHA like vLLM)