quantization_metadata.json

{
  "source_repo": "nvidia/LocateAnything-3B",
  "source_revision": "7a81d810571dc5f244b2f0b6868128f24b1cbd85",
  "policy_class": "transformers:AutoModel",
  "quantization": {
    "scheme": "nf4",
    "backend": "bitsandbytes",
    "compute_dtype": "bfloat16",
    "min_params_to_quantize": 4000000,
    "rule": "Linear modules with >=4_000_000 weight elements rewritten to bnb.nn.Linear4bit; smaller heads kept in compute_dtype (bfloat16).",
    "runtime_status": "loader-backed (install_prequantized_linears)"
  },
  "dropped_state_entries": []
}