"""CPU check that GR00T's Cosmos-Reason2-2B pieces load offline from $G/hf-home exactly as groot_node.sh will. Usage: HF_HOME=$G/hf-home HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 $G/envs/groot-py312/bin/python check_reason2_offline.py """ import json import os import sys from pathlib import Path import torch sys.path.insert(0, str(Path(__file__).resolve().parent)) import hb_offline_fix # noqa: E402,F401 from transformers import AutoConfig, Qwen3VLForConditionalGeneration, Qwen3VLProcessor assert os.environ.get("HF_HUB_OFFLINE") == "1", "run with HF_HUB_OFFLINE=1" name = "nvidia/Cosmos-Reason2-2B" torch.set_num_threads(8) proc = Qwen3VLProcessor.from_pretrained(name, trust_remote_code=True) cfg = AutoConfig.from_pretrained(name, trust_remote_code=True) ids = proc.tokenizer("pick up the trash and drop it in the bin")["input_ids"] model = Qwen3VLForConditionalGeneration.from_pretrained(name, torch_dtype=torch.bfloat16) n = sum(p.numel() for p in model.parameters()) print("REASON2_OFFLINE_OK", json.dumps({"processor": type(proc).__name__, "tokens": len(ids), "model_type": cfg.model_type, "params_M": round(n / 1e6, 1), "llm_layers": len(model.model.language_model.layers)}), flush=True)