How can I obtain .gguf file if I use my onw dataset fine-tuned on clip?

[cr42yhc1]

Hi, Thank you for your contribution.
I have fine-tuned a CLIP model using my own dataset and currently have the checkpoint in PyTorch format with a .pth extension. I am looking to run this model in C++. Could you guide me on how to convert this .pth file into either the Hugging Face standard format or the gguf format? Any assistance or resources would be greatly appreciated.

[monatis]

Hi, how did you finetune it? Currently this repo supports converting CLIP models from OpenAI and OpenCLIP in Huggingface format. So you need to first convert it to HF format, then you can follow the guide here.

[cr42yhc1]

Hi，thank you for your repply.

For fine-tuning code, you can refer to this openai/CLIP#83 and CSDN. If you need it, I can also provide my training code.

Is safetensors the HF format you mentioned? I tried to convert a .pth file to a .safetensors file using the following code, but your convert_hf_to_gguf.py script seems to require other related files. I attempted to use the related files from Huggingface, but encountered other issues. Maybe I made some mistakes during the operation.

import safetensors.torch
import torch
import safetensors

def convert(root):
    pth = torch.load(root, map_location='cpu', weights_only=True)
    safetensors.torch.save_file(pth, 'woce.safetensors')

root = './epoch_29.pth'
convert(root)