Skip to main content

Documentation Index

Fetch the complete documentation index at: https://fireworks.ai/docs/llms.txt

Use this file to discover all available pages before exploring further.

Early Access Feature. This page is part of the same private-preview external-bucket hot-load workflow for RL rollouts. Contact Fireworks to enable this path on your account before using non-FW_HOSTED storage.
Most readers should follow the linear workflow in RL Rollouts with Your Own Trainer. This page is only for advanced BYOT integrations or lower-level troubleshooting.
Use incremental snapshots between full snapshots to reduce upload and load time during RL training. Recommended cadence:
  • First snapshot: full.
  • Every 20th or 30th snapshot: full.
  • All other snapshots: incremental against the currently loaded snapshot.

​
Reference

  • Give every snapshot a new identity.
  • Point incremental_snapshot_metadata.previous_snapshot_identity at the previous snapshot in the chain.
  • Use ARC2 (arc_v2) for incremental_snapshot_metadata.compression_format.
  • Upload the incremental snapshot under the normal snapshot directory for the new identity.
  • If an incremental hot-load fails, fall back to a new full snapshot.

Full integration guide

The end-to-end BYOT flow.

Ledger & debugging

Inspect snapshot history and recover from a broken chain.