| """ | |
| Usage: | |
| python3 apply_delta --base ~/model_weights/llama-7b --target ~/model_weights/shikra-7b --delta lmsys/shikra-7b-delta | |
| """ | |
| import argparse | |
| import torch | |
| from tqdm import tqdm | |
| from transformers import AutoTokenizer, AutoModelForCausalLM | |
| from shikra import ShikraLlamaForCausalLM | |
| def apply_delta(base_model_path, target_model_path, delta_path): | |
| print("Loading base model") | |
| base = AutoModelForCausalLM.from_pretrained(base_model_path, torch_dtype=torch.float16, low_cpu_mem_usage=True) | |
| print("Loading delta") | |
| delta = ShikraLlamaForCausalLM.from_pretrained(delta_path, torch_dtype=torch.float16, low_cpu_mem_usage=True) | |
| delta_tokenizer = AutoTokenizer.from_pretrained(delta_path) | |
| print("Applying delta") | |
| for name, param in tqdm(delta.state_dict().items(), desc="Applying delta"): | |
| if name not in base.state_dict(): | |
| assert name in ['model.mm_projector.weight', 'model.mm_projector.bias'], f'{name} not in base model' | |
| continue | |
| if param.data.shape == base.state_dict()[name].shape: | |
| param.data += base.state_dict()[name] | |
| else: | |
| assert name in ['model.embed_tokens.weight', 'lm_head.weight'], \ | |
| f'{name} dimension mismatch: {param.data.shape} vs {base.state_dict()[name].shape}' | |
| bparam = base.state_dict()[name] | |
| param.data[:bparam.shape[0], :bparam.shape[1]] += bparam | |
| print("Saving target model") | |
| delta.save_pretrained(target_model_path) | |
| delta_tokenizer.save_pretrained(target_model_path) | |
| if __name__ == "__main__": | |
| parser = argparse.ArgumentParser() | |
| parser.add_argument("--base", type=str, required=True) | |
| parser.add_argument("--target", type=str, required=True) | |
| parser.add_argument("--delta", type=str, required=True) | |
| args = parser.parse_args() | |
| apply_delta(args.base, args.target, args.delta) | |