These scripts require [llama.cpp](https://github.com/ggml-org/llama.cpp) and [Convert-to-Quant](https://github.com/silveroxides/convert_to_quant).

`hf`, the Hugging Face CLI command is also suggested, but not required.

For quantizations or conversions, try:
- `hf_to_nvfp4_gguf` (huggingface structure to GGUF)
- `gguf_to_nvfp4_gguf` (GGUF to GGUF)
- `safetensors_to_nvfp4_gguf` (ComfyUI style safetensors to GGUF (untested))
- `hf_to_nvfp4_safetensors` (huggingface structure to ComfyUI style safetensors)
- `hf_to_nvfp4_hf` (huggingface structure to NVFP4 Hugging Face structure)
- `gguf_to_nvfp4_hf` (GGUF to NVFP4 Hugging Face structure)

For model info, try:
- `hf_info` (huggingface structure)
- `gguf_info` (GGUF structure)
---

Example:

```bash
hf download coder3101/gemma-4-12B-it-heretic --local-dir gemma-4-12B-it-heretic
hf_to_nvfp4_gguf -o gemma-4-12B-it-heretic_nvfp4 gemma-4-12B-it-heretic
```

This would create a NVFP4 GGUF for Ollama/llama.cpp from `https://huggingface.co/coder3101/gemma-4-12B-it-heretic`

The same directory can be quantized to one ComfyUI NVFP4 safetensors file:

`hf_to_nvfp4_safetensors -o gemma-4-12B-it-heretic-NVFP4.safetensors gemma-4-12B-it-heretic`

Conversions between formats without quantization is available in the custom commands with `--no-quant`.

Currently the only supported modes are "Quantize to NVFP4" or "Copy without quantizing" (`--no-quant`). I may add other quantization options in the future.