Skip to content

Checkpoint transport update

Latest
Latest

Choose a tag to compare

View all tags
@isabella618033 isabella618033 released this 23 Apr 22:02
· 349 commits to master since this release
v0.1.1
876d6d8
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Hugging Face is now the primary transport for validator-published checkpoints.

At the end of each cycle:

  1. Validators upload the checkpoint to their configured Hugging Face repo.
  2. Validators commit the exact revision SHA on-chain.
  3. Miners pull that exact revision from Hugging Face.

The validator /get-checkpoint HTTP endpoint remains available as a compatibility and outage fallback. If Hugging Face is misconfigured on a validator, or if a miner's Hugging Face download fails, the miner will automatically fall back to that validator's HTTP endpoint.

What changed

  • Hugging Face is now the primary checkpoint transport.
  • Added explicit HF readiness checks so missing HF_TOKEN or an unset hf.checkpoint_repo is surfaced in logs at startup instead of silently skipping uploads.
  • Pinned huggingface_hub to a version compatible with transformers 5.4.0.
  • Added failure-mode tests covering:
    • missing tokens
    • missing repos
    • HF download errors
    • validator HTTP fallback
  • End-to-end HF upload/download against a real temporary repo passed with revision-pinned verification.

Validator actions

Step 0

Make sure your axon port is open.

Step 1

Add the following to your validator config.yaml:

hf:
  checkpoint_repo: <your-hugging-face-account-id>/cyc

Step 2

Set a write-scoped HF_TOKEN in your Docker .env file, or in your shell environment for bare-metal deployments.

Step 3

Update and restart:

git pull
docker compose up -d

If you're running Watchtower, the image upgrade is automatic. However, the HF config must still be added manually. Validators without this config will continue serving checkpoints over HTTP, and miners will fall back to that path.

Miner actions

Please pull from the master branch.

  • Hugging Face is tried first.
  • HTTP fallback is automatic.

Note that validators must update manually, so the improvement from this release may not be immediately visible. Some validators may still be unreachable for submission if their submission path is not open.

Expected improvement

Checkpoint distribution now scales off the validator's uplink and onto Hugging Face's CDN.

As validators enable the HF path, high-fanout validators should see reduced slowdown during checkpoint distribution, starting from the next cycle.

Assets 2
Loading