quantization_config.json

{
  "quantization_method": "selective_moe_roundnearest",
  "architecture": "fused_3d",
  "num_experts": 128,
  "coding_expert_ids": [
    21,
    27,
    31,
    43,
    59,
    66,
    71,
    113,
    126
  ],
  "coding_bits": 8,
  "other_bits": 4,
  "group_size": 64,
  "attention_bits": null
}