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 -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：