hf_quant_config.json

{
  "producer": {
    "name": "quantization-tool",
    "version": "1.0"
  },
  "quantization": {
    "quant_algo": "NVFP4",
    "kv_cache_quant_algo": null,
    "group_size": 16,
    "exclude_modules": [
      "lm_head",
      "model.embed_tokens",
      "vision_tower*",
      "multi_modal_projector*",
      "patch_merge_mlp*",
      "*block_sparse_moe.gate"
    ]
  }
}