手把手教你:LLama2原始权重转HF模型

news2024/11/25 1:05:21

LLama2是meta最新开源的语言大模型,训练数据集2万亿token,上下文长度由llama的2048扩展到4096,可以理解和生成更长的文本,包括7B、13B和70B三个模型,在各种基准集的测试上表现突出,该模型可用于研究和商业用途。

LLama2模型权重和tokenizer下载需要申请访问。

申请链接:https://ai.meta.com/resources/models-and-libraries/llama-downloads/

由于下载的原始LLama2模型权重文件不能直接调用huggingface的transformers库进行使用,如果要使用huggingface transformer训练LLaMA2,需要使用额外的转换脚本。

转换脚本:https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/convert_llama_weights_to_hf.py

现在huggingface上已发布了llama的hf版本,可以直接使用。

现在介绍LLama2模型的原始权重获取和转换脚本。

LLama2模型原始权重获取

在MetaAI申请通过后将会在邮件中提及到PRESIGNED_URL,运行download.sh,按照提示输入即可。

set -e

read -p "Enter the URL from email: " PRESIGNED_URL 
echo ""
read -p "Enter the list of models to download without spaces (7B,13B,70B,7B-chat,13B-chat,70B-chat), or press Enter for all: " MODEL_SIZE  
TARGET_FOLDER="../target/file"             # where all files should end up
mkdir -p ${TARGET_FOLDER}

if [[ $MODEL_SIZE == "" ]]; then
    MODEL_SIZE="7B,13B,70B,7B-chat,13B-chat,70B-chat"
fi

echo "Downloading LICENSE and Acceptable Usage Policy"
wget --continue ${PRESIGNED_URL/'*'/"LICENSE"} -O ${TARGET_FOLDER}"/LICENSE"
wget --continue ${PRESIGNED_URL/'*'/"USE_POLICY.md"} -O ${TARGET_FOLDER}"/USE_POLICY.md"

echo "Downloading tokenizer"
wget --continue ${PRESIGNED_URL/'*'/"tokenizer.model"} -O ${TARGET_FOLDER}"/tokenizer.model"
wget --continue ${PRESIGNED_URL/'*'/"tokenizer_checklist.chk"} -O ${TARGET_FOLDER}"/tokenizer_checklist.chk"
CPU_ARCH=$(uname -m)
  if [ "$CPU_ARCH" = "arm64" ]; then
    (cd ${TARGET_FOLDER} && md5 tokenizer_checklist.chk)
  else
    (cd ${TARGET_FOLDER} && md5sum -c tokenizer_checklist.chk)
  fi

for m in ${MODEL_SIZE//,/ }
do
    if [[ $m == "7B" ]]; then
        SHARD=0
        MODEL_PATH="llama-2-7b"
    elif [[ $m == "7B-chat" ]]; then
        SHARD=0
        MODEL_PATH="llama-2-7b-chat"
    elif [[ $m == "13B" ]]; then
        SHARD=1
        MODEL_PATH="llama-2-13b"
    elif [[ $m == "13B-chat" ]]; then
        SHARD=1
        MODEL_PATH="llama-2-13b-chat"
    elif [[ $m == "70B" ]]; then
        SHARD=7
        MODEL_PATH="llama-2-70b"
    elif [[ $m == "70B-chat" ]]; then
        SHARD=7
        MODEL_PATH="llama-2-70b-chat"
    fi

    echo "Downloading ${MODEL_PATH}"
    mkdir -p ${TARGET_FOLDER}"/${MODEL_PATH}"

    for s in $(seq -f "0%g" 0 ${SHARD})
    do
        wget ${PRESIGNED_URL/'*'/"${MODEL_PATH}/consolidated.${s}.pth"} -O ${TARGET_FOLDER}"/${MODEL_PATH}/consolidated.${s}.pth"
    done

    wget --continue ${PRESIGNED_URL/'*'/"${MODEL_PATH}/params.json"} -O ${TARGET_FOLDER}"/${MODEL_PATH}/params.json"
    wget --continue ${PRESIGNED_URL/'*'/"${MODEL_PATH}/checklist.chk"} -O ${TARGET_FOLDER}"/${MODEL_PATH}/checklist.chk"
    echo "Checking checksums"
    if [ "$CPU_ARCH" = "arm64" ]; then
      (cd ${TARGET_FOLDER}"/${MODEL_PATH}" && md5 checklist.chk)
    else
      (cd ${TARGET_FOLDER}"/${MODEL_PATH}" && md5sum -c checklist.chk)
    fi
done

运行download.sh:

sh download.sh

代码注释

# 导入包
import argparse
import gc
import json
import os
import shutil
import warnings
import torch
from transformers import LlamaConfig, LlamaForCausalLM, LlamaTokenizer

# 判断LlamaTokenizerFast是否可用,LlamaTokenizerFast可以加速tokenization
try:
    from transformers import LlamaTokenizerFast
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

# 不同版本的LLama模型的分片数目
NUM_SHARDS = {
    "7B": 1,
    "7Bf": 1,
    "13B": 2,
    "13Bf": 2,
    "34B": 4,
    "30B": 4,
    "65B": 8,
    "70B": 8,
    "70Bf": 8,
}

# 计算中间层大小,优化计算效率
def compute_intermediate_size(n, ffn_dim_multiplier=1, multiple_of=256):
    return multiple_of * ((int(ffn_dim_multiplier * int(8 * n / 3)) + multiple_of - 1) // multiple_of)

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

# 写入json文件
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, tokenizer_path=None, safe_serialization=True):
    
    # 检查参数文件路径
    if not os.path.isfile(os.path.join(input_base_path, "params.json")):
        input_base_path = os.path.join(input_base_path, model_size)

    # 创建模型临时保存目录
    os.makedirs(model_path, exist_ok=True)
    tmp_model_path = os.path.join(model_path, "tmp")
    os.makedirs(tmp_model_path, exist_ok=True)

    # 读取参数
    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 = params.get("rope_theta", 10000.0)
    inv_freq = 1.0 / (base ** (torch.arange(0, dims_per_head, 2).float() / dims_per_head))
    if base > 10000.0:
        max_position_embeddings = 16384
    else:
        max_position_embeddings = 2048

    # 初始化tokenizer
    tokenizer_class = LlamaTokenizer if LlamaTokenizerFast is None else LlamaTokenizerFast
    if tokenizer_path is not None:
        tokenizer = tokenizer_class(tokenizer_path)
        tokenizer.save_pretrained(model_path)
    vocab_size = tokenizer.vocab_size if tokenizer_path is not None else 32000

    # 处理键值对头信息
    if "n_kv_heads" in params:
        num_key_value_heads = params["n_kv_heads"]  # for GQA / MQA
        num_local_key_value_heads = n_heads_per_shard // num_key_value_heads
        key_value_dim = dim // num_key_value_heads
    else:  # compatibility with other checkpoints
        num_key_value_heads = n_heads
        num_local_key_value_heads = n_heads_per_shard
        key_value_dim = dim

    # 张量变换
    def permute(w, n_heads=n_heads, dim1=dim, dim2=dim):
        return w.view(n_heads, dim1 // n_heads // 2, 2, dim2).transpose(1, 2).reshape(dim1, dim2)

    print(f"Fetching all parameters from the checkpoint at {input_base_path}.")
    # 加载权重
    if num_shards == 1:
        loaded = torch.load(os.path.join(input_base_path, "consolidated.00.pth"), map_location="cpu")
    else:
        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": {}}
    
    # 处理每一层的原始权重,并转化为bin文件
    for layer_i in range(n_layers):
        filename = f"pytorch_model-{layer_i + 1}-of-{n_layers + 1}.bin"
        if num_shards == 1:
            # 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 attention.w{q,k,v,o}, feed_fordward.w[1,2,3], attention_norm.weight and ffn_norm.weight share
            # the same storage object, saving attention_norm and ffn_norm will save other weights too, which is
            # redundant as other weights will be stitched from multiple shards. To avoid that, they are cloned.

            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(
                            num_local_key_value_heads, dims_per_head, dim
                        )
                        for i in range(num_shards)
                    ],
                    dim=0,
                ).reshape(key_value_dim, dim),
                num_key_value_heads,
                key_value_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(
                        num_local_key_value_heads, dims_per_head, dim
                    )
                    for i in range(num_shards)
                ],
                dim=0,
            ).reshape(key_value_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))

    # 处理最后一层权重,并保存
    filename = f"pytorch_model-{n_layers + 1}-of-{n_layers + 1}.bin"
    if num_shards == 1:
        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))

    # 写入配置文件
    index_dict["metadata"] = {"total_size": param_count * 2}
    write_json(index_dict, os.path.join(tmp_model_path, "pytorch_model.bin.index.json"))
    ffn_dim_multiplier = params["ffn_dim_multiplier"] if "ffn_dim_multiplier" in params else 1
    multiple_of = params["multiple_of"] if "multiple_of" in params else 256
    config = LlamaConfig(
        hidden_size=dim,
        intermediate_size=compute_intermediate_size(dim, ffn_dim_multiplier, multiple_of),
        num_attention_heads=params["n_heads"],
        num_hidden_layers=params["n_layers"],
        rms_norm_eps=params["norm_eps"],
        num_key_value_heads=num_key_value_heads,
        vocab_size=vocab_size,
        rope_theta=base,
        max_position_embeddings=max_position_embeddings,
    )
    config.save_pretrained(tmp_model_path)

    # 释放内存空间,以便正确加载模型
    del state_dict
    del loaded
    gc.collect()

    print("Loading the checkpoint in a Llama model.")
    # 从临时文件中加载模型
    model = LlamaForCausalLM.from_pretrained(tmp_model_path, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True)
    
    # 避免将此作为配置的一部分保存
    del model.config._name_or_path
    model.config.torch_dtype = torch.float16
    print("Saving in the Transformers format.")
    # 保存LLama模型到指定的路径
    model.save_pretrained(model_path, safe_serialization=safe_serialization)
    # 删除临时文件中的所有内容
    shutil.rmtree(tmp_model_path)

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


def main():
    
    # 参数处理
    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", "7Bf", "13B", "13Bf", "30B", "34B", "65B", "70B", "70Bf", "tokenizer_only"],
        help="'f' models correspond to the finetuned versions, and are specific to the Llama2 official release. For more details on Llama2, checkout the original repo: https://huggingface.co/meta-llama",
    )
    parser.add_argument(
        "--output_dir",
        help="Location to write HF model and tokenizer",
    )
    parser.add_argument("--safe_serialization", type=bool, help="Whether or not to save using `safetensors`.")
    args = parser.parse_args()
    
    spm_path = os.path.join(args.input_dir, "tokenizer.model")
    
    # 判断转换的对象
    if args.model_size != "tokenizer_only":
        write_model(
            model_path=args.output_dir,
            input_base_path=args.input_dir,
            model_size=args.model_size,
            safe_serialization=args.safe_serialization,
            tokenizer_path=spm_path,
        )
    else:
        write_tokenizer(args.output_dir, spm_path)


if __name__ == "__main__":
    main()

脚本运行

python convert_llama_weights_to_hf.py --input_dir raw-llama2-7b --output_dir llama2_7b_hf

raw-llama2-7b文件夹内容:

llama2_7b_hf转换文件内容:

本文来自互联网用户投稿,该文观点仅代表作者本人,不代表本站立场。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如若转载,请注明出处:http://www.coloradmin.cn/o/1189903.html

如若内容造成侵权/违法违规/事实不符,请联系多彩编程网进行投诉反馈,一经查实,立即删除!

相关文章

Head First Java 第二版

不管你的程序有多大,一定都会有一个main()来作为程序的起点。Java是强类型语言。float f23.5f 如果不加上f,就会被Java当做double处理。对于任意一个Java虚拟机来说,所有的引用大小都一样,但是不同的Java虚拟机可能会以不同的方…

什么是前台、中台、和后台?

前台:即包括与用户直接交互的界面,如:web页、app;也包括服务端各种实时响应用户请求的业务逻辑,如:商品查询、订单系统等。 后台:面向内部运营人员的管理系统、配置系统,如&#xf…

2023.11.9 IDEA 配置 Lombok

目录 什么是 Lombok 如何使用 Lombok Lombok 的 Data 注解 什么是 Lombok Lombok 是一个 Java 库,能自动插入编译器并构建工具,简化 Java 开发它通过注解实现这一目的,可用来帮助开发人员消除 Java 的冗长代码,尤其是对于简单…

通付盾Web3专题 | SharkTeam:Web3安全实践与创新

在Web3领域,安全漏洞、黑客攻击已愈发成为用户和投资者重点关注的领域。如何保障加密资产的安全,Web3黑暗森林中又有哪些新的攻击模式产生,SharkTeam将从一线进行分享和讨论。 我们先来看一下2023年1月到8月的安全事件数量和损失的数据统计。…

【Python自学笔记】python os.getcwd文件目录找不对关于“None”在VSCode里面的奇葩报错

写小组项目的时候需要按照路径读入数据表,数据库和图片列表显示到html,按ChatGPT的答案写了python os.getcwd(),结果迁移到同组同学的电脑上总是报错。 经过一番查询,在CSDN上发现一个完美解决问题的好帖,特此存下链接…

Java数据的基本(原始)类型和引用类型的特点差别

本文作为“Java数据类型”一文的补充https://blog.csdn.net/cnds123/article/details/110517272 Java的数据类型可以分为基本类型(primitive types)和引用类型(reference types)两大类。在实际编程中,要根据需求选择合…

React路由与导航

目录 前言: 什么是React路由? 导航和页面切换 路由参数和动态路由 路由守卫和权限控制 总结 前言: React是一个流行的JavaScript库,用于构建用户界面。在使用React开发Web应用程序时,路由和导航是必不可少的功能…

理解MySQL的日志 Redo、Undo

理解MySQL的Redo日志和Undo日志 1、MySQL 日志文件解决的问题2、redo 日志2.1、redo log 的组成2.2、redo log 刷盘策略2.3、MySQL 的 redo log解决了哪些问题 3、undo 日志3.1、undo 日志作用3.2、undo log 的类型3.3、undo log 的生命周期3.4、事务回滚相关的几个隐藏字段 1、…

【Mysql】where 条件子句之逻辑运算符

逻辑运算符 and &&or ||not ! student表 一.查询分数在80 - 90之间 and写法 &&写法 区间(between ....and......) 二.查询分数不为88 !写法 not写法 三.查询分数大于88或者年龄小于22 满足其中一个条件即可 or写法 ||写法

操作系统 day08(进程通信)

进程通信的概念 进程间通信是指两个进程之间产生数据交互进程通信需要操作系统的支持,由于进程是分配系统资源(包括内存地址)的单位,因此各进程拥有的内存地址空间相互独立。同时为了保证安全,一个进程不能直接访问另…

django安装和rest接口写法

django安装 确保已经安装了Python。命令行中输入python --version来检查Python的版本。 安装Django。你可以在命令行中使用以下命令来安装Django: pip install django创建一个新的Django项目。在命令行中,进入你想要创建项目的目录,并运行以…

SpringCloud-Gateway无法使用Feign服务(2021.X版本)

Spring Cloud Gateway 2021.x版本,无法使用Feign调用其他服务接口。 问题原因: 在官网的 issue 里面找到了相关的问题。 How to call another micro-service on GatewayFilterFactory ? Issue #1090 spring-cloud/spring-cloud-gateway GitHubHel…

python编程复习系列——week2(Input Output (2))

文章目录 一、多行代码语句二、Escape序列三、字符串格式四、数值运算课后作业 一、多行代码语句 🥞使用反斜杠\来表示在下一行中继续使用一条语句。 subject_code "CSCI111" subject_mark 80 subject_grade "D" result "Subject re…

SOLIDWORKS --电磁仿真篇

什么是 SIMULIA? 基于3DEXPERIENCE平台的品牌 多学科多领域的协同仿真与分析优化 三大核心仿真领域 结构仿真 流体仿真 SIMULIA电磁仿真是什么? 完备的求解技术,支持从静场、低频到高频、光波的电磁仿真,支持全波仿真、混合仿真、多物理场仿真和场路…

支持C#的开源免费、新手友好的数据结构与算法入门教程 - Hello算法

前言 前段时间完成了C#经典十大排序算法(完结)然后有很多小伙伴问想要系统化的学习数据结构和算法,不知道该怎么入门,有无好的教程推荐的。今天给大家推荐一个支持C#的开源免费、新手友好的数据结构与算法入门教程:He…

Python语言:经典例题分析讲解

题1: 通过观察我们可以得出以下结论: 代码实现: """ (3)输入整数n,输出n行的字符图案。如n5时输出以下图案:* *** ***** ******* *********""""" for…

多测师肖sir_高级金牌讲师_ui自动化po框架

ui自动化po框架 一、po框架 1、基本介绍(1)po是page object 的缩写 (2)业务流程与页面元素操作分类的模式, (3)提高测试用例的可维护性、可读性 二、自动化测试框架分层如下: 结构…

MATLAB|不给糖果就捣蛋

目录 扫一扫关注公众号 效果图 代码 绘制南瓜 绘制无脸男小鬼 其中绘制风车代码: 其中 EllipsePlotter类函数代码如下 属性 (properties) 方法 (methods) 扫一扫关注公众号 效果图 代码 绘制南瓜 clc;clear;close all; [X,Y,Z]sphere(200); R1(-(1-mod(0:…

Flink(三)【运行时架构】

前言 今天学习 Flink 的一些原理性的东西,比较偏概念,但是十分重要。有人觉得上来框框敲代码才能学到东西,那是狗屁不通的道理(虽然我以前也这么认为)。个人认为,学习 JavaEE那些框架,你上来就敲…

​软考-高级-系统架构设计师教程(清华第2版)【第1章-绪论-思维导图】​

软考-高级-系统架构设计师教程(清华第2版)【第1章-绪论-思维导图】 课本里章节里所有蓝色字体的思维导图