quantize_config.json

{
  "quant_method": "channel_wise_int8",
  "bits": 8,
  "format": "per_output_channel_scale",
  "weight_dtype": "I8",
  "scale_dtype": "BF16",
  "description": "Each weight tensor 'foo' is stored as I8 values in 'foo' with a BF16 per-output-channel scale in 'foo_scale'. Dequantize: weight_bf16 = weight_i8 * scale_bf16."
}