import argparse
import os
import shutil

import torch
from internvl.model.internvl_chat import InternVLChatModel


parser = argparse.ArgumentParser()
parser.add_argument('base_model_path', type=str, help='Base InternVL checkpoint to keep projector and LLM from.')
parser.add_argument('vit_source_model_path', type=str, help='InternVL checkpoint to copy vision_model weights from.')
parser.add_argument('output_path', type=str, help='Output path for the merged checkpoint.')
args = parser.parse_args()


base_model = InternVLChatModel.from_pretrained(args.base_model_path, torch_dtype=torch.bfloat16)
vit_source_model = InternVLChatModel.from_pretrained(args.vit_source_model_path, torch_dtype=torch.bfloat16)

base_model.vision_model.load_state_dict(vit_source_model.vision_model.state_dict(), strict=True)
base_model.to(torch.bfloat16)

if os.path.exists(args.output_path):
    shutil.rmtree(args.output_path)
os.makedirs(args.output_path, exist_ok=True)

base_model.save_pretrained(args.output_path, save_config=False)

for name in os.listdir(args.base_model_path):
    src = os.path.join(args.base_model_path, name)
    dst = os.path.join(args.output_path, name)
    if name.startswith('model-') or name == 'model.safetensors' or name == 'model.safetensors.index.json':
        continue
    if os.path.isdir(src):
        shutil.copytree(src, dst, dirs_exist_ok=True)
    else:
        shutil.copy2(src, dst)

print(f'finished: {args.output_path}')