If I train from a base model (InternLM-XComposer2.5), how can I obtain model files in the Hugging Face format?

[guoguo1314]

Hello, I'd like to try your training process. I used train.sh and set the number of epochs to one, then obtained pytorch_model.bin. Here's my question: What should I do to obtain model files in the Hugging Face format of MBZUAI/GeoPixel-7B

[AkashahS]

To merge Lora weights and get final model files in Hugging Face format, see Merging LoRA Weights.

[guoguo1314]

thanks a lot.
I've replaced InternLM-XComposer2.5 with another vision-language (VL) model. Therefore, the MBZUAI/GeoPixel-7B specified after --version in the command CUDA_VISIBLE_DEVICES=0 python merge_lora_weights_and_save_hf_model.py --version="MBZUAI/GeoPixel-7B" --weight="output/pytorch_model.bin" --save_path="GeoPixel-7B-finetuned" is no longer applicable. As a result, this model parameter in the command is invalid. Now, I need to obtain a model file similar to MBZUAI/GeoPixel-7B. (Currently, I'm using another VL model combined with SAM, and I've successfully trained it to generate a pytorch_model.bin file.)

I suspect that the pytorch_model.bin file (which is 22.7GB) might already be the final model. Perhaps it just needs to be split into three parts: pytorch_model-00001-of-00003.bin, pytorch_model-00002-of-00003.bin, and pytorch_model-00003-of-00003.bin (the total size of these three files is also 22.7 GB).