transformer/quant_config.json

{
  "group_size": 32,
  "bits": 4,
  "scope": "attn|feed_forward",
  "note": "attn+ffn Linears quantized; x_embedder/time_caption_embed/norm_out/AdaLN kept bf16"
}