手把手教你:LLama2原始权重转HF模型

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 -eread -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 == "" ]]; thenMODEL_SIZE="7B,13B,70B,7B-chat,13B-chat,70B-chat"
fiecho "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)fifor m in ${MODEL_SIZE//,/ }
doif [[ $m == "7B" ]]; thenSHARD=0MODEL_PATH="llama-2-7b"elif [[ $m == "7B-chat" ]]; thenSHARD=0MODEL_PATH="llama-2-7b-chat"elif [[ $m == "13B" ]]; thenSHARD=1MODEL_PATH="llama-2-13b"elif [[ $m == "13B-chat" ]]; thenSHARD=1MODEL_PATH="llama-2-13b-chat"elif [[ $m == "70B" ]]; thenSHARD=7MODEL_PATH="llama-2-70b"elif [[ $m == "70B-chat" ]]; thenSHARD=7MODEL_PATH="llama-2-70b-chat"fiecho "Downloading ${MODEL_PATH}"mkdir -p ${TARGET_FOLDER}"/${MODEL_PATH}"for s in $(seq -f "0%g" 0 ${SHARD})dowget ${PRESIGNED_URL/'*'/"${MODEL_PATH}/consolidated.${s}.pth"} -O ${TARGET_FOLDER}"/${MODEL_PATH}/consolidated.${s}.pth"donewget --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_shardsdim = params["dim"]dims_per_head = dim // n_headsbase = 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 = 16384else:max_position_embeddings = 2048# 初始化tokenizertokenizer_class = LlamaTokenizer if LlamaTokenizerFast is None else LlamaTokenizerFastif 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 / MQAnum_local_key_value_heads = n_heads_per_shard // num_key_value_headskey_value_dim = dim // num_key_value_headselse:  # compatibility with other checkpointsnum_key_value_heads = n_headsnum_local_key_value_heads = n_heads_per_shardkey_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 = 0index_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:# Unshardedstate_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_freqfor k, v in state_dict.items():index_dict["weight_map"][k] = filenameparam_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] = filenameparam_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 1multiple_of = params["multiple_of"] if "multiple_of" in params else 256config = 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_dictdel loadedgc.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_pathmodel.config.torch_dtype = torch.float16print("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` modeltokenizer_class = LlamaTokenizer if LlamaTokenizerFast is None else LlamaTokenizerFastprint(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.rhkb.cn/news/185715.html

如若内容造成侵权/违法违规/事实不符,请联系长河编程网进行投诉反馈email:809451989@qq.com,一经查实,立即删除!

相关文章

Android sqlite 使用简介

进行Android应用开发时经常会用到数据库。Android系统支持sqlite数据库,在app开发过程中很容易通过SQLiteOpenHelper使用数据库,SQLiteOpenHelper依赖于Context对象,但是基于uiatomator1.0和Java程序等无法获取Context的应用如何使用数据库呢…

MongoDB副本集特点验证

MongoDB副本集特点验证 mogodb副本集概述副本集搭建副本集结构验证结果源码地址 mogodb副本集概述 MongoDB副本集是将数据同步在多个服务器的过程。 复制提供了数据的冗余备份,并在多个服务器上存储数据副本,提高了数据的可用性, 并可以保证…

【MongoDB】索引 - 复合索引

一、准备工作 这里准备一些学生数据 db.students.insertMany([{ _id: 1, name: "张三", age: 20, class: { id: 1, name: "1班" }},{ _id: 2, name: "李四", age: 22, class: { id: 2, name: "2班" }},{ _id: 3, name: "王五…

经验模态分解(Empirical Mode Decomposition,EMD)(附代码)

代码原理 EMD(Empirical Mode Decomposition),也称为经验模态分解,是一种将非线性和非平稳信号分解成多个本征模态函数(Intrinsic Mode Functions,简称IMF)的方法。 EMD的基本原理是通过一系列…

【Mysql】增删改查(基础版)

我使用的工具是Data Grip (SQLyog Naivact 都行) 使用Data Grip创建student表,具体步骤如下(熟悉Data Grip或者使用SQLyog,Naivact可以跳过) https://blog.csdn.net/m0_67930426/article/details/13429…

zookeeper:服务器有几种状态?

四种: looking(选举中)、leading(leader)、following( follower)、 observer(观察者角色)

字节流操作

for i in range(100):ai.to_bytes(2,byteorderbig)print(i,a,end )if i%40:print() 字节流 a5678 先把5678转换为二进制就变成 0001_0110_0010_1110拆分两个字节,高字节在前,低字节在后 hig_byte 0001_0110 对应的16进制 0x16 little_byte 0010_11…

掌动智能:UI自动化测试工具产品功能和优势

UI自动化测试工具是一种软件工具,用于模拟用户与应用程序的交互,检查应用程序的用户界面是否按预期工作。这些工具允许开发人员编写测试脚本,以模拟用户操作,例如点击按钮、输入文本、导航菜单等,然后验证应用程序的响…

在linux上脱离hadoop安装hbase-2.5.6集群

一、软件版本 1.1、jdk1.8 1.2、hbase 2.5.6 1.3、zookeeper 3.8.1 二、计算节点 准备三台服务器 192.168.42.139 node1 192.168.42.140 node2 192.168.42.141 node3三、配置环境 1、每台服务器都配置jdk环境变量 [rootnode1 data]# javac -version javac 1.8.0_3912、每…

lvgl 转换和使用新字体

一、背景 如果lvgl 提供的默认字体不符合我们的显示要求,我们可以在网上下载开源字体,或者利用系统自带(注意版权问题)的字体文件转换lvgl 能识别和调用的字体。 或者为了压缩存储空间,某些字体我们只需要个别字符&…

SpringBoot测试类启动web环境-下篇

一、响应状态 1.MockMvcResultMatchers 说明:模拟结果匹配。 package com.forever;import org.junit.jupiter.api.Test; import org.springframework.beans.factory.annotation.Autowired; import org.springframework.boot.test.autoconfigure.web.servlet.AutoC…

TCP网络编程

一)TCP Socket介绍: 1)TCP和UDP有着很大的不同,TCP想要进行网络通信的话首先需要通信双方建立连接以后然后才可以进行通信,TCP进行网络编程的方式和文件中的读写字节流类似,是以字节为单位的流进行传输 2)针对于TCP的套接字来说,J…

LangChain之关于RetrievalQA input_variables 的定义与使用

最近在使用LangChain来做一个LLMs和KBs结合的小Demo玩玩,也就是RAG(Retrieval Augmented Generation)。 这部分的内容其实在LangChain的官网已经给出了流程图。 我这里就直接偷懒了,准备对Webui的项目进行复刻练习,那么…

【QEMU-tap-windows-Xshell】QEMU 创建 aarch64虚拟机(附有QEMU免费资源)

“从零开始:在Windows上创建aarch64(ARM64)虚拟机” 前言 aarch64(ARM64)架构是一种现代的、基于 ARM 技术的计算架构,具有诸多优点,如低功耗、高性能和广泛应用等。为了在 Windows 平台上体验…

(欧拉)openEuler系统添加网卡文件配置流程、(欧拉)openEuler系统手动配置ipv6地址流程、(欧拉)openEuler系统网络管理说明

文章目录 系统说明openEuler23.03系统手动配置ip流程修改名称生成网卡配置文件【openEuler23.03系统添加网卡文件配置流程】手动指定ip添加ipv6地址修改配置文件信息和名称删除创建的网卡信息重启网卡生效并测试 openEuler23.03系统网络管理说明 系统说明 我这用云上最小化安装…

【Python大数据笔记_day05_Hive基础操作】

一.SQL,Hive和MapReduce的关系 用户在hive上编写sql语句,hive把sql语句转化为MapReduce程序去执行 二.Hive架构映射流程 用户接口: 包括CLI、JDBC/ODBC、WebGUI,CLI(command line interface)为shell命令行;Hive中的Thrift服务器允许外部客户端…

Python+Selenium+Unittest 之selenium12--WebDriver操作方法2-鼠标操作1(ActionChains类简介)

在我们平时的使用过程中,会使用鼠标去进行很多操作,比如鼠标左键点击、双击、鼠标右键点击,鼠标指针悬浮、拖拽等操作。在selenium中,我们也可以去实现常用的这些鼠标操作,这时候就需要用到selenium中的ActionChains类…

SQL第三次上机作业

1.查询与王利就读同一专业学生的借书证号和姓名 USE TSGL GO SELECT Lno,Rname FROM Reader WHERE Dept(SELECT DeptFROM ReaderWHERE Rname王利) and Rname ! 王利2.查询比希望出版社出版的所有图书价格都高的图书信息 SELECT * FROM Book WHERE Price>(SELECT MAX(Price…

5G-DFS最新动态-产品不在需要走FCC官方测试

添加图片注释,不超过 140 字(可选) 最近,FCC公布了最新版本的PAG(Product Acceptance Group)清单,即388624 D02 Pre-Approval Guidance List v18r04。这个清单的主要改变是将带有雷达侦测功能的…

AVL树详解

目录 AVL树的概念 旋转的介绍 单旋转 双旋转 旋转演示 具体实现 通过高度判断的实现 通过平衡因子判断的实现 AVL树的概念 AVL树是一种自平衡的平衡二叉查找树,它是一种高效的数据结构,可以在插入和删除节点时保持树的平衡,从而保证…