# FLUX Quantization Config: Transformer Blocks Only
#
# Quantize only double and single transformer blocks,
# leave input/output projections in higher precision.

disable_list: [
  # Disable input projections
  "*img_in*",
  "*txt_in*",
  "*time_in*",
  "*vector_in*",
  "*guidance_in*",
  
  # Disable output layers
  "*final_layer*",

  # Disable positional embeddings
  "*pe_embedder*",

  "*modulation*",
  "*txt_mod*",
  "*img_mod*",
]

per_layer_dtype: {
  "*": "float4_e2m1fn_x2",
}