quantization_config.json

{
  "recipe": "nvfp4_awq",
  "weight_only": true,
  "quant_lmhead": true,
  "modelopt_artifact": "transformer/modelopt_quantized.pt",
  "loader": "Cosmos3OmniTransformer.from_config + modelopt.restore_from_modelopt_state + safetensors load_state_dict",
  "exclusions": [
    "embed_tokens",
    "*norm*",
    "time_embedder",
    "proj_in",
    "proj_out",
    "audio_*",
    "action_*"
  ],
  "scale_layout": {
    "weight_scale_suffixes": [
      "_amax",
      "_scale"
    ],
    "global_scale_suffixes": [],
    "n_weight": 802,
    "n_scale": 1010,
    "granularity": "per-block-16"
  }
}