int8/quant_config.json

{
  "bits": 8,
  "group_size": 64,
  "expert_gate_up_bits": 8,
  "expert_down_bits": 8,
  "quantized_linear_paths": [
    "layers.0.attn.gater",
    "layers.0.attn.wkv",
    "layers.0.attn.wo",
    "layers.0.attn.wq",
    "layers.0.ffn.w_in",
    "layers.0.ffn.w_out",
    "layers.1.attn.gater",
    "layers.1.attn.wkv",
    "layers.1.attn.wo",
    "layers.1.attn.wq",
    "layers.1.ffn.w_in",
    "layers.1.ffn.w_out",
    "layers.10.attn.gater",
    "layers.10.attn.wkv",
    "layers.10.attn.wo",
    "layers.10.attn.wq",
    "layers.11.attn.gater",
    "layers.11.attn.wkv",
    "layers.11.attn.wo",
    "layers.11.attn.wq",
    "layers.12.attn.gater",
    "layers.12.attn.wkv",
    "layers.12.attn.wo",
    "layers.12.attn.wq",
    "layers.13.attn.gater",
    "layers.13.attn.wkv",
    "layers.13.attn.wo",
    "layers.13.attn.wq",
    "layers.14.attn.gater",
    "layers.14.attn.wkv",
    "layers.14.attn.wo",
    "layers.14.attn.wq",
    "layers.15.attn.gater",
    "layers.15.attn.wkv",
    "layers.15.attn.wo",
    "layers.15.attn.wq",
    "layers.16.attn.gater",
    "layers.16.attn.wkv",
    "layers.16.attn.wo",
    "layers.16.attn.wq",
    "layers.17.attn.gater",
    "layers.17.attn.wkv",
    "layers.17.attn.wo",
    "layers.17.attn.wq",
    "layers.18.attn.gater",
    "layers.18.attn.wkv",
    "layers.18.attn.wo",
    "layers.18.attn.wq",
    "layers.19.attn.gater",
    "layers.19.attn.wkv",
    "layers.19.attn.wo",
    "layers.19.attn.wq",
    "layers.2.attn.gater",
    "layers.2.attn.wkv",
    "layers.2.attn.wo",
    "layers.2.attn.wq",
    "layers.2.ffn.w_in",
    "layers.2.ffn.w_out",
    "layers.20.attn.gater",
    "layers.20.attn.wkv",
    "layers.20.attn.wo",
    "layers.20.attn.wq",
    "layers.21.attn.gater",
    "layers.21.attn.wkv",
    "layers.21.attn.wo",
    "layers.21.attn.wq",
    "layers.22.attn.gater",
    "layers.22.attn.wkv",
    "layers.22.attn.wo",
    "layers.22.attn.wq",
    "layers.23.attn.gater",
    "layers.23.attn.wkv",
    "layers.23.attn.wo",
    "layers.23.attn.wq",
    "layers.24.attn.gater",
    "layers.24.attn.wkv",
    "layers.24.attn.wo",
    "layers.24.attn.wq",
    "layers.25.attn.gater",
    "layers.25.attn.wkv",
    "layers.25.attn.wo",
    "layers.25.attn.wq",
    "layers.26.attn.gater",
    "layers.26.attn.wkv",
    "layers.26.attn.wo",
    "layers.26.attn.wq",
    "layers.27.attn.gater",
    "layers.27.attn.wkv",
    "layers.27.attn.wo",
    "layers.27.attn.wq",
    "layers.27.ffn.w_in",
    "layers.27.ffn.w_out",
    "layers.3.attn.gater",
    "layers.3.attn.wkv",
    "layers.3.attn.wo",
    "layers.3.attn.wq",
    "layers.4.attn.gater",
    "layers.4.attn.wkv",
    "layers.4.attn.wo",
    "layers.4.attn.wq",
    "layers.5.attn.gater",
    "layers.5.attn.wkv",
    "layers.5.attn.wo",
    "layers.5.attn.wq",
    "layers.6.attn.gater",
    "layers.6.attn.wkv",
    "layers.6.attn.wo",
    "layers.6.attn.wq",
    "layers.7.attn.gater",
    "layers.7.attn.wkv",
    "layers.7.attn.wo",
    "layers.7.attn.wq",
    "layers.8.attn.gater",
    "layers.8.attn.wkv",
    "layers.8.attn.wo",
    "layers.8.attn.wq",
    "layers.9.attn.gater",
    "layers.9.attn.wkv",
    "layers.9.attn.wo",
    "layers.9.attn.wq",
    "lm_head"
  ],
  "quantized_expert_bases": [
    "layers.10.moe.experts",
    "layers.11.moe.experts",
    "layers.12.moe.experts",
    "layers.13.moe.experts",
    "layers.14.moe.experts",
    "layers.15.moe.experts",
    "layers.16.moe.experts",
    "layers.17.moe.experts",
    "layers.18.moe.experts",
    "layers.19.moe.experts",
    "layers.20.moe.experts",
    "layers.21.moe.experts",
    "layers.22.moe.experts",
    "layers.23.moe.experts",
    "layers.24.moe.experts",
    "layers.25.moe.experts",
    "layers.26.moe.experts",
    "layers.3.moe.experts",
    "layers.4.moe.experts",
    "layers.5.moe.experts",
    "layers.6.moe.experts",
    "layers.7.moe.experts",
    "layers.8.moe.experts",
    "layers.9.moe.experts"
  ],
  "expert_split_scheme": "split_gate_up_interleaved"
}