examples/quantization-verification.json

{
  "format": "cosmos3-sdnq-verification-v1",
  "source": "models/Cosmos3-Super-Image2Video-4Step/transformer",
  "checkpoint": "checkpoints/Cosmos3-Super-Image2Video-4Step-SDNQ-INT4-G64-BF16IO/transformer",
  "selection_report": "benchmarks/i2v/verification-selection.json",
  "weights_dtype": "int4",
  "group_size": 64,
  "selected_tensors": [
    "embed_tokens.weight",
    "layers.0.self_attn.to_q.weight",
    "layers.0.mlp.up_proj.weight",
    "layers.21.self_attn.to_k.weight",
    "layers.21.mlp_moe_gen.down_proj.weight",
    "layers.42.self_attn.add_v_proj.weight",
    "layers.42.mlp.gate_proj.weight",
    "layers.63.self_attn.to_out.weight",
    "layers.63.mlp.down_proj.weight",
    "norm.weight",
    "proj_in.weight",
    "proj_out.weight"
  ],
  "packing_exact": true,
  "global_relative_mse": 0.013438441064997239,
  "core_global_relative_mse": 0.013438441064997239,
  "elapsed_seconds": 1.4211450000002515,
  "tensors": [
    {
      "name": "embed_tokens.weight",
      "quantized": false,
      "exact": true
    },
    {
      "name": "layers.0.self_attn.to_q.weight",
      "quantized": true,
      "packing_exact": true,
      "relative_mse": 0.0119432761408363
    },
    {
      "name": "layers.0.mlp.up_proj.weight",
      "quantized": true,
      "packing_exact": true,
      "relative_mse": 0.012443561128219444
    },
    {
      "name": "layers.21.self_attn.to_k.weight",
      "quantized": true,
      "packing_exact": true,
      "relative_mse": 0.013138661031115623
    },
    {
      "name": "layers.21.mlp_moe_gen.down_proj.weight",
      "quantized": true,
      "packing_exact": true,
      "relative_mse": 0.015179319676087252
    },
    {
      "name": "layers.42.self_attn.add_v_proj.weight",
      "quantized": true,
      "packing_exact": true,
      "relative_mse": 0.01455430271461897
    },
    {
      "name": "layers.42.mlp.gate_proj.weight",
      "quantized": true,
      "packing_exact": true,
      "relative_mse": 0.012624818630985208
    },
    {
      "name": "layers.63.self_attn.to_out.weight",
      "quantized": true,
      "packing_exact": true,
      "relative_mse": 0.011737540160858885
    },
    {
      "name": "layers.63.mlp.down_proj.weight",
      "quantized": true,
      "packing_exact": true,
      "relative_mse": 0.013330119643253325
    },
    {
      "name": "norm.weight",
      "quantized": false,
      "exact": true
    },
    {
      "name": "proj_in.weight",
      "quantized": false,
      "exact": true
    },
    {
      "name": "proj_out.weight",
      "quantized": false,
      "exact": true
    }
  ]
}