transformer/config.json

{
  "_class_name": "Flux2Transformer2DModel",
  "_diffusers_version": "0.37.1",
  "_name_or_path": "/content/unquant_transformer",
  "attention_head_dim": 128,
  "axes_dims_rope": [
    32,
    32,
    32,
    32
  ],
  "eps": 1e-06,
  "guidance_embeds": false,
  "in_channels": 128,
  "joint_attention_dim": 12288,
  "mlp_ratio": 3.0,
  "num_attention_heads": 32,
  "num_layers": 8,
  "num_single_layers": 24,
  "out_channels": null,
  "patch_size": 1,
  "quantization_config": {
    "add_skip_keys": true,
    "dequantize_fp32": true,
    "dynamic_loss_threshold": 0.01,
    "group_size": 0,
    "is_integer": true,
    "is_training": false,
    "modules_dtype_dict": {
      "int8": [
        "context_embedder",
        "x_embedder",
        "proj_out",
        "norm_out"
      ]
    },
    "modules_quant_config": {},
    "modules_to_not_convert": [
      "double_stream_modulation_txt",
      "time_guidance_embed",
      "context_embedder",
      "norm_k",
      "norm_q",
      "proj_out",
      "x_embedder",
      "norm_added_k",
      ".proj_out",
      "norm_out",
      "single_stream_modulation",
      "double_stream_modulation_img",
      "norm_added_q"
    ],
    "non_blocking": false,
    "quant_conv": false,
    "quant_embedding": false,
    "quant_method": "sdnq",
    "quantization_device": "cuda",
    "quantized_matmul_dtype": null,
    "return_device": "cpu",
    "sdnq_version": "0.1.8",
    "svd_rank": 32,
    "svd_steps": 32,
    "use_dynamic_quantization": true,
    "use_grad_ckpt": true,
    "use_quantized_matmul": true,
    "use_quantized_matmul_conv": false,
    "use_static_quantization": true,
    "use_stochastic_rounding": false,
    "use_svd": true,
    "weights_dtype": "uint4"
  },
  "rope_theta": 2000,
  "timestep_guidance_channels": 256
}