Verify a round-trip

The claim OpenWeights makes is that an upload followed by a download from a cold cache returns identical bytes. Here is how to check it yourself.

By hand

Upload a file, download it into a fresh directory with fresh caches, and compare SHA-256 hashes:

# 1. upload
HF_TOKEN=<your-key> HF_ENDPOINT=http://localhost:8080 \
  hf upload <owner>/<repo> ./model-dir

# 2. download from empty caches
HF_ENDPOINT=http://localhost:8080 \
HF_HOME=$(mktemp -d) HF_XET_CACHE=$(mktemp -d) \
  hf download <owner>/<repo> --local-dir ./out

# 3. compare
shasum -a 256 ./model-dir/weights.bin
shasum -a 256 ./out/weights.bin

The two hashes match.

If step 2 fails, the most common cause is that the xorb is not pinned yet. Wait and retry: see Choose an indexer.

With the bundled harness

The repository ships a script that does the whole thing, including waiting for the pin:

OPENWEIGHTS_API_KEY=<your-key> \
CAS_URL=http://localhost:8080 \
  bash tests/hf-roundtrip/standalone-roundtrip.sh

Use a key with the read + write scope (the console default). The harness both uploads and downloads with the same key, so it needs both scopes.

What it does:

1. Creates a temporary directory with a 4 MiB random binary (which exercises the Xet chunking path) and a small text file.
2. Records the SHA-256 of every file.
3. Runs hf upload against CAS_URL with HF_ENDPOINT pointed at it.
4. Runs hf download with HF_HOME and HF_XET_CACHE set to empty directories, retrying every 20 seconds until it succeeds.
5. Compares the hash of every downloaded file against the original and exits non-zero on any mismatch.

On success it prints OK: byte-identical round-trip followed by the hashes, and exits 0.

Variables it reads

The shipped Compose stack serves the CAS on http://localhost:8080 (the default). Set CAS_URL explicitly only if you run the dev override, which remaps it to http://localhost:28080.

What makes this hold

Each layer verifies the bytes it handles, so corruption cannot pass silently:

• The CAS recomputes each xorb's Merkle hash before writing to Sia, and rejects a mismatch with 400.
• The gateway hash-verifies every xorb before writing it into its disk cache.
• hf_xet verifies each reconstructed file against its content hash.