Is there a way to UN-gguf a model?

[ColumbusAI]

I'm curious if this functionality could be added. This way, users wouldn't have to keep multiple quants of a model as they can try GGML/GGUF quant, then un-do and compare to other quantization types to see which works best for their setup.

[leafspark]

You can dequantize GGUFs, but precision is lost and can't be fully recovered. For example, bf16 (safetensors) -> bf16 (GGUF) -> Q4_K_M (GGUF) -> Q6_K (GGUF) loses precision. To try multiple quantization types, you would normally keep the bf16 (GGUF), and quantize from that for maximum precision.

I hacked together a quick Python script to test dequantizing GGUF models, although it doesn't recover precision:

import json

import numpy as np
import torch
from safetensors.torch import save_file

import gguf


def gguf_to_safetensors(gguf_path, safetensors_path, dequantize=False, metadata_path=None):
    """
    Converts a GGUF file to SafeTensors format.

    Args:
        gguf_path: Path to the GGUF file.
        safetensors_path: Path to save the SafeTensors file.
        dequantize: If True, dequantize tensors to float32.
        metadata_path: Optional path to save metadata as a JSON file.
    """

    reader = gguf.GGUFReader(gguf_path)
    tensors = {}
    metadata = {}

    for tensor in reader.tensors:
        if dequantize and tensor.tensor_type != gguf.GGMLQuantizationType.F32:
            dequantized_data = tensor.data.astype(np.float32)
            tensors[tensor.name] = torch.from_numpy(dequantized_data.reshape(tuple(reversed(tensor.shape)))) # Reshape
        else:
            tensors[tensor.name] = torch.from_numpy(np.array(tensor.data).reshape(tuple(reversed(tensor.shape)))) # Reshape

    for field_name, field in reader.fields.items():
        if field.data:
            metadata[field_name] = field.parts[field.data[0]].tolist()

    save_file(tensors, safetensors_path)

    decoded_metadata = {}
    for key, value in metadata.items():
        if isinstance(value, list) and all(isinstance(item, int) for item in value):
            decoded_value = ""
            for item in value:
                if 48 <= item <= 57:  # Check if it's an ASCII code for a number
                    decoded_value += str(item - 48)  # Convert to the actual number
                elif 32 <= item <= 126:  # Check if it's a printable ASCII character
                    decoded_value += chr(item)
                else:
                    decoded_value += str(item) # Keep as it is if not a number or printable char.
            decoded_metadata[key] = decoded_value
        else:
            decoded_metadata[key] = value

    if metadata_path:
        with open(metadata_path, "w") as f:
            json.dump(decoded_metadata, f, indent=4)

# Example usage
gguf_file = "quantized.gguf"
safetensors_file = "dequantized_gguf.safetensors"
metadata_file = "config.gguf.json"

gguf_to_safetensors(gguf_file, safetensors_file, dequantize=True, metadata_path=metadata_file)