"""Sequential, resumable, pinned model downloads into /mnt/localssd/models (one at a time: no tokenizer/download races).""" import json, os, sys os.environ["HF_HOME"] = "/mnt/localssd/hf_cache" from huggingface_hub import HfApi, snapshot_download spec = json.load(open("/mnt/localssd/seqroute1m/scripts/models_requested.json")) only = set(sys.argv[1:]) api = HfApi() for m in spec: if only and m["id"] not in only: continue dest = f"/mnt/localssd/models/{m['id']}" if os.path.exists(dest + "/.download_complete"): continue repo = m["repo"] try: sha = api.model_info(repo).sha snapshot_download(repo, revision=sha, local_dir=dest, max_workers=8, allow_patterns=["*.json", "*.safetensors", "*.model", "*.txt", "tokenizer*", "*.jinja", "*.py", "tekken*"], ignore_patterns=["consolidated*", "original/*", "metal/*", "*.gguf", "*.pth", "*.bin"]) open(dest + "/.download_complete", "w").write(f"{repo}@{sha}\n") print("done", m["id"], repo, sha, flush=True) except Exception as e: print("FAILED", m["id"], repo, repr(e)[:200], flush=True)