# Quantization Configuration for Checkpoint Merger
# 
# This file defines which layers to quantize and what precision to use.
# Patterns use glob-style syntax where * matches any characters.

# Regex patterns of layers to DISABLE quantization
# If a layer matches any pattern here, it will NOT be quantized
disable_list: [
  # Example: disable input/output projection layers
  # "*img_in*",
  # "*txt_in*",
  # "*final_layer*",
  
  # Example: disable specific block types
  # "*norm*",
  # "*time_in*",
]

# Per-layer dtype configuration
# Maps layer name patterns to quantization formats
# Layers are matched in order - first match wins
per_layer_dtype: {
  # Default: quantize all layers to FP8 E4M3
  "*": "fp8_e4m3fn",
  
  # Example: use different precision for specific layers
  # "*attn*": "fp8_e4m3fn",      # Attention layers
  # "*mlp*": "fp8_e4m3fn",        # MLP layers
  # "*qkv*": "fp8_e4m3fn",        # Q/K/V projections
}