[LLMs] LLaMA 기반 챗봇 언어모델, Vicuna 다운로드 및 사용법

지난 포스트에서 Meta에서 공개한 LLMs 모델인 LLaMA의 다운로드 방법에 대해 알아보았습니다.
이번 포스트에서는 LLaMA를 Finetuning 해 ChatGPT에 버금가는(90%) 챗봇 성능을 보여준
Vicuna 모델의 다운로드 및 실행 방법을 소개합니다.

Relative Response Quality Assessed by GPT-4 *

 

Vicuna

: An Open-Source Chatbot Impressing GPT-4 with 90%* ChatGPT Quality

학습 방법 등 모델에 대한 자세한 설명은 아래 Post를 참고하시기 바랍니다.
이번 포스트에서는 설치 및 실행 방법을 위주로 소개합니다.

Vicuna (generated by stable diffusion 2.1)

[ Github / Post / Demo ]

Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90%* ChatGPT Quality

 

Vicuna 13B Demo, 참고 - https://github.com/lm-sys/FastChat

 

1. Install

conda env에서 해당 라이브러리 설치를 추천드립니다.

pip3 install fschat

 

2. Model Weights

1) 우선 지난 포스트에서 설명한 것과 같이 LLaMA model license를 통해 LLaMA weights을 다운로드합니다.

2) 그 후 Hugginface format으로 변환하는 과정을 거칩니다. [참고]

아래와 같이 파일을 만들어 코드를 복붙한 뒤,

convert_llama_weights_to_hf.py [참고] - 기존 코드에서 LlamaTokenizerFast -> LlamaTokenizer로 수정

# Copyright 2022 EleutherAI and The HuggingFace Inc. team. All rights reserved.
#
# 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.
import argparse
import gc
import json
import math
import os
import shutil
import warnings

import torch

from transformers import LlamaConfig, LlamaForCausalLM, LlamaTokenizer


try:
    from transformers import LlamaTokenizer
except ImportError as e:
    warnings.warn(e)
    warnings.warn(
        "The converted tokenizer will be the `slow` tokenizer. To use the fast, update your `tokenizers` library and re-run the tokenizer conversion"
    )
    LlamaTokenizerFast = None

"""
Sample usage:

```
python src/transformers/models/llama/convert_llama_weights_to_hf.py \
    --input_dir /path/to/downloaded/llama/weights --model_size 7B --output_dir /output/path
```

Thereafter, models can be loaded via:

```py
from transformers import LlamaForCausalLM, LlamaTokenizer

model = LlamaForCausalLM.from_pretrained("/output/path")
tokenizer = LlamaTokenizer.from_pretrained("/output/path")
```

Important note: you need to be able to host the whole model in RAM to execute this script (even if the biggest versions
come in several checkpoints they each contain a part of each weight of the model, so we need to load them all in RAM).
"""

INTERMEDIATE_SIZE_MAP = {
    "7B": 11008,
    # "13B": 13824,
    # "30B": 17920,
    # "65B": 22016,
}
NUM_SHARDS = {
    "7B": 1,
    # "13B": 2,
    # "30B": 4,
    # "65B": 8,
}


def compute_intermediate_size(n):
    return int(math.ceil(n * 8 / 3) + 255) // 256 * 256


def read_json(path):
    with open(path, "r") as f:
        return json.load(f)


def write_json(text, path):
    with open(path, "w") as f:
        json.dump(text, f)


def write_model(model_path, input_base_path, model_size):
    os.makedirs(model_path, exist_ok=True)
    tmp_model_path = os.path.join(model_path, "tmp")
    os.makedirs(tmp_model_path, exist_ok=True)

    params = read_json(os.path.join(input_base_path, "params.json"))
    num_shards = NUM_SHARDS[model_size]
    n_layers = params["n_layers"]
    n_heads = params["n_heads"]
    n_heads_per_shard = n_heads // num_shards
    dim = params["dim"]
    dims_per_head = dim // n_heads
    base = 10000.0
    inv_freq = 1.0 / (base ** (torch.arange(0, dims_per_head, 2).float() / dims_per_head))

    # permute for sliced rotary
    def permute(w):
        return w.view(n_heads, dim // n_heads // 2, 2, dim).transpose(1, 2).reshape(dim, dim)

    print(f"Fetching all parameters from the checkpoint at {input_base_path}.")
    # Load weights
    if model_size == "7B":
        # Not sharded
        # (The sharded implementation would also work, but this is simpler.)
        loaded = torch.load(os.path.join(input_base_path, "consolidated.00.pth"), map_location="cpu")
    else:
        # Sharded
        loaded = [
            torch.load(os.path.join(input_base_path, f"consolidated.{i:02d}.pth"), map_location="cpu")
            for i in range(num_shards)
        ]
    param_count = 0
    index_dict = {"weight_map": {}}
    for layer_i in range(n_layers):
        filename = f"pytorch_model-{layer_i + 1}-of-{n_layers + 1}.bin"
        if model_size == "7B":
            # Unsharded
            state_dict = {
                f"model.layers.{layer_i}.self_attn.q_proj.weight": permute(
                    loaded[f"layers.{layer_i}.attention.wq.weight"]
                ),
                f"model.layers.{layer_i}.self_attn.k_proj.weight": permute(
                    loaded[f"layers.{layer_i}.attention.wk.weight"]
                ),
                f"model.layers.{layer_i}.self_attn.v_proj.weight": loaded[f"layers.{layer_i}.attention.wv.weight"],
                f"model.layers.{layer_i}.self_attn.o_proj.weight": loaded[f"layers.{layer_i}.attention.wo.weight"],
                f"model.layers.{layer_i}.mlp.gate_proj.weight": loaded[f"layers.{layer_i}.feed_forward.w1.weight"],
                f"model.layers.{layer_i}.mlp.down_proj.weight": loaded[f"layers.{layer_i}.feed_forward.w2.weight"],
                f"model.layers.{layer_i}.mlp.up_proj.weight": loaded[f"layers.{layer_i}.feed_forward.w3.weight"],
                f"model.layers.{layer_i}.input_layernorm.weight": loaded[f"layers.{layer_i}.attention_norm.weight"],
                f"model.layers.{layer_i}.post_attention_layernorm.weight": loaded[f"layers.{layer_i}.ffn_norm.weight"],
            }
        else:
            # Sharded
            # Note that in the 13B checkpoint, not cloning the two following weights will result in the checkpoint
            # becoming 37GB instead of 26GB for some reason.
            state_dict = {
                f"model.layers.{layer_i}.input_layernorm.weight": loaded[0][
                    f"layers.{layer_i}.attention_norm.weight"
                ].clone(),
                f"model.layers.{layer_i}.post_attention_layernorm.weight": loaded[0][
                    f"layers.{layer_i}.ffn_norm.weight"
                ].clone(),
            }
            state_dict[f"model.layers.{layer_i}.self_attn.q_proj.weight"] = permute(
                torch.cat(
                    [
                        loaded[i][f"layers.{layer_i}.attention.wq.weight"].view(n_heads_per_shard, dims_per_head, dim)
                        for i in range(num_shards)
                    ],
                    dim=0,
                ).reshape(dim, dim)
            )
            state_dict[f"model.layers.{layer_i}.self_attn.k_proj.weight"] = permute(
                torch.cat(
                    [
                        loaded[i][f"layers.{layer_i}.attention.wk.weight"].view(n_heads_per_shard, dims_per_head, dim)
                        for i in range(num_shards)
                    ],
                    dim=0,
                ).reshape(dim, dim)
            )
            state_dict[f"model.layers.{layer_i}.self_attn.v_proj.weight"] = torch.cat(
                [
                    loaded[i][f"layers.{layer_i}.attention.wv.weight"].view(n_heads_per_shard, dims_per_head, dim)
                    for i in range(num_shards)
                ],
                dim=0,
            ).reshape(dim, dim)

            state_dict[f"model.layers.{layer_i}.self_attn.o_proj.weight"] = torch.cat(
                [loaded[i][f"layers.{layer_i}.attention.wo.weight"] for i in range(num_shards)], dim=1
            )
            state_dict[f"model.layers.{layer_i}.mlp.gate_proj.weight"] = torch.cat(
                [loaded[i][f"layers.{layer_i}.feed_forward.w1.weight"] for i in range(num_shards)], dim=0
            )
            state_dict[f"model.layers.{layer_i}.mlp.down_proj.weight"] = torch.cat(
                [loaded[i][f"layers.{layer_i}.feed_forward.w2.weight"] for i in range(num_shards)], dim=1
            )
            state_dict[f"model.layers.{layer_i}.mlp.up_proj.weight"] = torch.cat(
                [loaded[i][f"layers.{layer_i}.feed_forward.w3.weight"] for i in range(num_shards)], dim=0
            )

        state_dict[f"model.layers.{layer_i}.self_attn.rotary_emb.inv_freq"] = inv_freq
        for k, v in state_dict.items():
            index_dict["weight_map"][k] = filename
            param_count += v.numel()
        torch.save(state_dict, os.path.join(tmp_model_path, filename))

    filename = f"pytorch_model-{n_layers + 1}-of-{n_layers + 1}.bin"
    if model_size == "7B":
        # Unsharded
        state_dict = {
            "model.embed_tokens.weight": loaded["tok_embeddings.weight"],
            "model.norm.weight": loaded["norm.weight"],
            "lm_head.weight": loaded["output.weight"],
        }
    else:
        state_dict = {
            "model.norm.weight": loaded[0]["norm.weight"],
            "model.embed_tokens.weight": torch.cat(
                [loaded[i]["tok_embeddings.weight"] for i in range(num_shards)], dim=1
            ),
            "lm_head.weight": torch.cat([loaded[i]["output.weight"] for i in range(num_shards)], dim=0),
        }

    for k, v in state_dict.items():
        index_dict["weight_map"][k] = filename
        param_count += v.numel()
    torch.save(state_dict, os.path.join(tmp_model_path, filename))

    # Write configs
    index_dict["metadata"] = {"total_size": param_count * 2}
    write_json(index_dict, os.path.join(tmp_model_path, "pytorch_model.bin.index.json"))

    config = LlamaConfig(
        hidden_size=dim,
        intermediate_size=compute_intermediate_size(dim),
        num_attention_heads=params["n_heads"],
        num_hidden_layers=params["n_layers"],
        rms_norm_eps=params["norm_eps"],
    )
    config.save_pretrained(tmp_model_path)

    # Make space so we can load the model properly now.
    del state_dict
    del loaded
    gc.collect()

    print("Loading the checkpoint in a Llama model.")
    model = LlamaForCausalLM.from_pretrained(tmp_model_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
    # Avoid saving this as part of the config.
    del model.config._name_or_path

    print("Saving in the Transformers format.")
    model.save_pretrained(model_path)
    shutil.rmtree(tmp_model_path)


def write_tokenizer(tokenizer_path, input_tokenizer_path):
    # Initialize the tokenizer based on the `spm` model
    tokenizer_class = LlamaTokenizer if LlamaTokenizer is None else LlamaTokenizer
    print(f"Saving a {tokenizer_class.__name__} to {tokenizer_path}.")
    tokenizer = tokenizer_class(input_tokenizer_path)
    tokenizer.save_pretrained(tokenizer_path)


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--input_dir",
        help="Location of LLaMA weights, which contains tokenizer.model and model folders",
    )
    parser.add_argument(
        "--model_size",
        # choices=["7B", "13B", "30B", "65B", "tokenizer_only"],
        choices= ["7B"]
    )
    parser.add_argument(
        "--output_dir",
        help="Location to write HF model and tokenizer",
    )
    args = parser.parse_args()
    if args.model_size != "tokenizer_only":
        write_model(
            model_path=args.output_dir,
            input_base_path=os.path.join(args.input_dir, args.model_size),
            model_size=args.model_size,
        )
    spm_path = os.path.join(args.input_dir, "tokenizer.model")
    write_tokenizer(args.output_dir, spm_path)


if __name__ == "__main__":
    main()

 

아래와 같은 명령을 실행합니다. 현재 명령어에서는 7B 예시를 보여주었지만 다른 더 큰 모델(13B 등)도 가능합니다.

python convert_llama_weights_to_hf.py \
	--input_dir /data/LLaMA/ckpt_7B \
    --model_size 7B --output_dir /data/LLaMA/HF

이때 /data/LLaMA/ckpt_7B는 아래와 같은 상태여야 합니다. 

위의 명령어를 실행하면 아래와 같이 hugginface format으로 변환을 완료할 수 있습니다.

Fetching all parameters from the checkpoint at /data/LLaMA/ckpt_7B/7B.
Loading the checkpoint in a Llama model.
Loading checkpoint shards: 100%|███████████████████████████████████████████████████████████████████████████████████████████| 33/33 [00:05<00:00,  5.86it/s]
Saving in the Transformers format.
Saving a LlamaTokenizer to /data/LLaMA/HF/.

 

3) Vicuna weights

python3 -m fastchat.model.apply_delta \
	--base-model-path /data/LLaMA/HF \
    --target-model-path /data/LLaMA/vicuna-7b \
    --delta-path lmsys/vicuna-7b-delta-v1.1

위와 같은 명령을 실행하면 아래와 같이 vicuna weigths을 얻을 수 있습니다.

Loading the delta weights from lmsys/vicuna-7b-delta-v1.1
Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:03<00:00,  1.89s/it]
Loading the base model from /data/LLaMA/HF
Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:03<00:00,  1.79s/it]
Applying the delta
Applying delta: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████| 323/323 [00:01<00:00, 179.63it/s]
Saving the target model to /data/LLaMA/vicuna-7b

 

 

3. Vicuna 실행

python3 -m fastchat.serve.cli --model-path /data/LLaMA/vicuna-7b

USER: 라는 글자가 뜨면 원하는 질문을 던지면 됩니다.

"Give me a 3day travel plan for Seoul" 이라는 질문을 하자 아래와 같이 답을 합니다.