quantization_metadata.json

{
  "model_type": "HunyuanImage-3.0-Instruct",
  "quantization_method": "bitsandbytes_nf4",
  "load_in_4bit": true,
  "bnb_4bit_quant_type": "nf4",
  "bnb_4bit_use_double_quant": true,
  "bnb_4bit_compute_dtype": "torch.bfloat16",
  "expected_vram_gb": 45,
  "modules_kept_bf16": [
    "vae",
    "vision_model",
    "vision_aligner",
    "patch_embed",
    "final_layer",
    "time_embed",
    "time_embed_2",
    "timestep_emb",
    "attention_projections"
  ],
  "notes": "Instruct model with vision encoder kept at BF16 for image understanding quality.",
  "attention_layers_quantized": false
}