| """Put the full gated nvidia/Cosmos-Reason2-2B (pinned 9ce19a1) into the shared HF hub cache $G/hf-home/hub. |
| |
| GR00T N1.7 builds its backbone with Qwen3VLForConditionalGeneration.from_pretrained("nvidia/Cosmos-Reason2-2B") and |
| its processor with Qwen3VLProcessor.from_pretrained(same name) before loading the GR00T checkpoint, so every node needs |
| the whole repo resolvable by name. groot_node.sh runs with HF_HUB_OFFLINE=1, so nodes read this cache and never need |
| the token. refs/main is pinned to the verified revision. |
| |
| Usage (token from stdin, kept in-process only): |
| $G/envs/groot-py312/bin/python cache_cosmos_reason2.py $G/hf-home/hub < /root/.hb-secrets/hf_token |
| """ |
| import hashlib |
| import json |
| import os |
| import sys |
| import time |
|
|
| from huggingface_hub import HfApi, snapshot_download |
|
|
| REPO = "nvidia/Cosmos-Reason2-2B" |
| REV = "9ce19a195e423419c349abfc86fd07178b230561" |
| cache = os.path.abspath(sys.argv[1]) |
| token = sys.stdin.readline().strip() |
| if not token.startswith("hf_"): |
| raise SystemExit("token format rejected") |
| api = HfApi(endpoint="https://huggingface.co", token=token) |
| info = api.model_info(REPO, revision=REV, files_metadata=True) |
| assert info.sha == REV, info.sha |
| t0 = time.monotonic() |
| for attempt in range(1, 21): |
| try: |
| path = snapshot_download(REPO, revision=REV, cache_dir=cache, token=token, max_workers=8) |
| break |
| except Exception as error: |
| if any(k in str(error).lower() for k in ("gated", "401", "403")): |
| raise |
| print(f"retry {attempt}: {type(error).__name__}: {str(error)[:200]}", flush=True) |
| time.sleep(10) |
| else: |
| raise SystemExit("download did not finish") |
| token = None |
| count, size = 0, 0 |
| for s in info.siblings: |
| f = os.path.join(path, s.rfilename) |
| assert os.path.isfile(f), s.rfilename |
| if s.size is not None: |
| assert os.path.getsize(f) == s.size, s.rfilename |
| if s.lfs is not None: |
| h = hashlib.sha256() |
| with open(f, "rb") as stream: |
| for chunk in iter(lambda: stream.read(1 << 22), b""): |
| h.update(chunk) |
| assert h.hexdigest() == s.lfs.sha256, s.rfilename |
| count += 1 |
| size += s.size or 0 |
| refs = os.path.join(cache, "models--" + REPO.replace("/", "--"), "refs") |
| os.makedirs(refs, exist_ok=True) |
| with open(os.path.join(refs, "main"), "w") as f: |
| f.write(REV) |
| print("CACHE_OK", json.dumps({"repo": REPO, "revision": REV, "snapshot": path, "files": count, "bytes": size, |
| "seconds": round(time.monotonic() - t0, 1)}), flush=True) |
|
|
