sdnq_quantization_summary.json

{
  "source_model": "microsoft/Lens-Turbo",
  "method": "SDNQ uint4 static",
  "scope": "transformer only, excluding modulation linears",
  "ablation_fix": "Transformer block img_mod and txt_mod linears are left in bfloat16 because UINT4 quantization caused periodic grid artifacts and severe text degradation.",
  "config": {
    "weights_dtype": "uint4",
    "quantized_matmul_dtype": "int8",
    "group_size": 0,
    "use_static_quantization": true,
    "use_dynamic_quantization": false,
    "use_quantized_matmul": true,
    "use_svd": false,
    "use_hadamard": false,
    "quant_conv": false,
    "quant_embedding": false,
    "dequantize_fp32": false,
    "modules_to_not_convert": [
      "*.img_mod.*",
      "*.txt_mod.*"
    ],
    "modules_to_not_use_matmul": [],
    "quantization_device": "cuda",
    "return_device": "cuda"
  }
}