llava/model/make_delta.py

# Copyright (c) 2025 NVIDIA CORPORATION.
# Licensed under the MIT license.

# Adapted from https://github.com/NVlabs/VILA/tree/main under the Apache 2.0 license.
# LICENSE is in incl_licenses directory.

# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# SPDX-License-Identifier: Apache-2.0

# This file is modified from https://github.com/haotian-liu/LLaVA/

"""
Usage:
python3 -m llava.model.make_delta --base ~/model_weights/llama-7b --target ~/model_weights/llava-7b --delta ~/model_weights/llava-7b-delta --hub-repo-id liuhaotian/llava-7b-delta
"""
import argparse

import torch
from tqdm import tqdm
from transformers import AutoModelForCausalLM, AutoTokenizer

from llava.model.utils import auto_upgrade


def make_delta(base_model_path, target_model_path, delta_path, hub_repo_id):
    print("Loading base model")
    base = AutoModelForCausalLM.from_pretrained(base_model_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)

    print("Loading target model")
    auto_upgrade(target_model_path)
    target = AutoModelForCausalLM.from_pretrained(target_model_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)

    print("Calculating delta")
    for name, param in tqdm(target.state_dict().items(), desc="Calculating 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 delta")
    if hub_repo_id:
        kwargs = {"push_to_hub": True, "repo_id": hub_repo_id}
    else:
        kwargs = {}
    target.save_pretrained(delta_path, **kwargs)
    target_tokenizer = AutoTokenizer.from_pretrained(target_model_path)
    target_tokenizer.save_pretrained(delta_path, **kwargs)


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--base-model-path", type=str, required=True)
    parser.add_argument("--target-model-path", type=str, required=True)
    parser.add_argument("--delta-path", type=str, required=True)
    parser.add_argument("--hub-repo-id", type=str, default=None)
    args = parser.parse_args()

    make_delta(args.base_model_path, args.target_model_path, args.delta_path, args.hub_repo_id)