LM Resiliency Documentation

LM Resiliency provides GEMINI checkpoint recovery and SCOUT failure localization for distributed LLM training.

Start here

  • API Guide — public APIs, framework integration, recovery, and manager hooks.
  • Compatibility — supported and tested Python/framework versions.
  • GEMINI — checkpoint tiers, replication, persistence, and recovery semantics.
  • SCOUT — failure coverage, replay, consensus, and localization contracts.
  • Torchrun Resiliency — fixed-size replacement architecture, coordinator protocol, fencing, and restart context.
  • Validation — qualification evidence, environments, and known boundaries.
  • Release Process — artifact validation and publication workflow.

Documentation versions

The hosted site publishes immutable static trees by source revision:

  • /dev/ tracks the current main branch.
  • /<version>/ contains documentation built from the successfully validated v<version> release revision.
  • /latest/ mirrors the most recently published stable release documentation.

Before the first stable documentation tree is published, the site root redirects to /dev/. Once a stable release is published through the repository release workflow, the root redirects to /latest/.

Publication workflows stage the complete versioned static tree on gh-pages. A dedicated trusted workflow then uploads the current staged tree as a GitHub Pages artifact and deploys it through GitHub's Pages deployment API. Repository Pages settings therefore need to use GitHub Actions as the publishing source rather than building directly from the gh-pages branch.

The repository Markdown remains the source of truth for documentation changes. CI builds the site with warnings treated as errors before changes can satisfy the required CI gate. A weekly workflow checks external links online with bounded retries, complementing the deterministic offline link and anchor check that runs on every pull request.