quantization_structure_report.md

Wan2.1 MXFP4 quantization structure report

• Source: /workspace/lj/models/Wan2.1-T2V-1.3B-Diffusers
• Quantized: /workspace/lj/models/Wan2.1-T2V-1.3B-Diffusers-MXFP4-AutoRoundFormat-rtn
• Mode: RTN (iters=0)
• Quantized shards: 1
• Quantized transformer size: 0.739 GiB
• Source transformer size: 5.286 GiB
• Approximate size reduction: 7.15x

quantization_config

{
  "act_bits": 4,
  "act_data_type": "mx_fp",
  "act_dynamic": true,
  "act_group_size": 32,
  "act_sym": true,
  "autoround_version": "0.14.2",
  "batch_size": 1,
  "bits": 4,
  "block_name_to_quantize": "blocks",
  "data_type": "mx_fp",
  "enable_quanted_input": false,
  "group_size": 32,
  "nsamples": 16,
  "packing_format": "auto_round:llm_compressor",
  "quant_method": "auto-round",
  "sym": true
}