Modelconverter

Scripts for safetensors model conversion and editing, such as nvfp4 quantization (see to_format.py).

to_format.py

usage:

python to_format.py path/to/model.safetensors path/to/modeldef.sc types...

types can be the name of a pytorch type, or nvfp4/fp8_scaled for quant types.
Multiple types are supported if the .sc config supports it. More info here. Example .sc configs can be found in model_defs.

types

Supports loading regular pytorch types, comfy format nvfp4 and fp8_scaled.
Supports converting to the following types (and aliases):

• fp64 - double
• fp32 - fp, float, full
• fp16 - half
• bf16 - brain
• fp8_e4m3fn - fp8_e4m3, fp8_e4, fp8
• fp8_e5m2 - fp8_e5
• nvfp4 - fp4 (using comfy kitchen)
• fp8_scaled - scaled_fp8 (using comfy kitchen)

additional features

• SVD creation from precision losses use --svd [rank] to create a corrector lora. (Note: effectiveness depends on implementation.)

Credits

• comfy kitchen - Used for nvfp4 and fp8 scaled quantization and dequantization.