前言

准备工作

Git

Python3.9

Cmake

下载模型

合并模型

部署模型

前言

想必有小伙伴也想跟我一样体验下部署大语言模型, 但碍于经济实力, 不过民间上出现了大量的量化模型, 我们平民也能体验体验啦~, 该模型可以在笔记本电脑上部署, 确保你电脑至少有16G运行内存

开原地址：GitHub - ymcui/Chinese-LLaMA-Alpaca: 中文LLaMA&Alpaca大语言模型+本地CPU部署 (Chinese LLaMA & Alpaca LLMs)

准备工作

最好是有代理, 不然你下载东西可能失败, 我为了下个模型花了一天时间, 痛哭~

我们需要先在电脑上安装以下环境：

Git
Python3.9（使用Anaconda3创建该环境）
Cmake（如果你电脑没有C和C++的编译环境还需要安装mingw）

Git

下载地址：Git - Downloading Package

下载好安装包后打开, 一直点下一步安装即可...

在cmd窗口输入以下如果有版本号显示说明已经安装成功

git -v

Python3.9

我这里使用Anaconda3来使用Python, Anaconda3是什么？

如果你熟悉docker, 那么你可以把docker的概念带过来, docker可以创建很多个容器, 每个容器的环境可能一样也可能不一样, Anaconda3也是一样的, 它可以创建很多个不同的Python版本, 互相不冲突, 想用哪个版本就切换到哪个版本...

Anaconda3下载地址：Anaconda | Anaconda Distribution

安装步骤参考：

等待安装好后一直点next, 直到点Finish关闭即可

在cmd窗口输入以下命令, 显示版本号则说明安装成功

conda -V

接下来我们在cmd窗口输入以下命令创建一个python3.9的环境

conda create --name py39 python=3.9 -y

--name后面的py39是环境名字, 可以自己任意起, 切换环境的时候需要它

python=3.9是指定python版本

添加-y后就不需要手动输入y去确认安装了

查看有哪些环境的命令：

conda info -e

激活/切换环境的命令：

conda activate py39

要使用哪个环境的话换成对应名字即可

进入环境后你就可以在这输入python相关的命令了, 如：

要退出环境的话输入：

conda deactivate

当我退出环境后再查看python版本的话会提示我不是内部或外部命令，也不是可运行的程序
或批处理文件。如：

Cmake

这是一个编译工具, 我们需要使用它去编译llama.cpp, 量化模型需要用到, 不量化模型个人电脑跑不起来, 觉得量化这个概念不理解的可以理解为压缩, 这种概念是不对的, 只是为了帮助你更好的理解.

在安装之前我们需要安装mingw, 避免编译时找不到编译环境, 按下win+r快捷键输入powershell

输入命令安装scoop, 这是一个包管理器, 我们使用它来下载安装mingw：

iex "& {$(irm get.scoop.sh)} -RunAsAdmin"

安装好后分别运行下面两个命令（添加库）：

scoop bucket add extras

scoop bucket add main

输入命令安装mingw

scoop install mingw

到这就已经安装好mingw了, 如果报错了请评论, 我看到了会回复

接下来安装Cmake

地址：Download | CMake

安装参考：

安装好后点Finish即可

下载模型

我们需要下载两个模型, 一个是原版的LLaMA模型, 一个是扩充了中文的模型, 后续会进行一个合并模型的操作

原版模型下载地址：https://ipfs.io/ipfs/Qmb9y5GCkTG7ZzbBWMu2BXwMkzyCKcUjtEKPpgdZ7GEFKm/

扩充了中文的模型下载：

建议在D盘上新建一个文件夹, 在里面进行下载操作, 如下：

在弹出的框中分别输入以下命令：

git lfs install

git clone https://huggingface.co/ziqingyang/chinese-alpaca-lora-7b

这里可能会因为网络问题一直失败......一直重试就行, 有别的问题请评论, 看到会回复

合并模型

终于写到这里了, 累~

在你下载了模型的目录内打开cmd窗口, 如下：

这里我先说下这图片中的两个目录里文件是啥吧

先是chinese-alpaca-lora-7b目录, 这个目录一般你下载下来就不用动了, 格式如下：

chinese-alpaca-lora-7b/
        - adapter_config.json
        - adapter_model.bin
        - special_tokens_map.json
        - tokenizer_config.json
        - tokenizer.model

然后是path_to_original_llama_root_dir目录, 这个文件夹需要创建, 保持一致的文件名, 目录内的格式如下：

path_to_original_llama_root_dir/

- 7B/        #这是一个名为7B的文件夹

- checklist.chk

- consolidated.00.pth

- params.json

- tokenizer_checklist.chk

- tokenizer.model

自行按照上面的格式存放

打开窗口后需要先激活python环境, 使用的就是前面装Anaconda3

# 不记得有哪些环境的先运行以下命令
conda info -e

# 然后激活你需要的环境  我的环境名是py39
conda activate py39

切换好后分别执行以下命令安装依赖库

pip install git+https://github.com/huggingface/transformers

pip install sentencepiece==0.1.97

pip install peft==0.2.0

执行命令安装成功后会有Successfully的字眼

接下来需要将原版模型转HF格式, 需要借助最新版🤗transformers提供的脚本convert_llama_weights_to_hf.py

在目录内新建一个convert_llama_weights_to_hf.py文件, 用记事本打开后把以下代码粘贴进去

# 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 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

"""
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 shared
        # (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 LlamaTokenizerFast is None else LlamaTokenizerFast
    print("Saving a {tokenizer_class} 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"],
    )
    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()

在cmd窗口执行命令：

python convert_llama_weights_to_hf.py --input_dir path_to_original_llama_root_dir --model_size 7B --output_dir path_to_original_llama_hf_dir

经过漫长的等待....

接下来合并输出PyTorch版本权重（.pth文件），使用merge_llama_with_chinese_lora.py脚本

在目录新建一个merge_llama_with_chinese_lora.py文件, 用记事本打开将以下代码粘贴进去

"""
Borrowed and modified from https://github.com/tloen/alpaca-lora
"""

import argparse
import os
import json
import gc

import torch
import transformers
import peft
from peft import PeftModel

parser = argparse.ArgumentParser()
parser.add_argument('--base_model',default=None,required=True,type=str,help="Please specify a base_model")
parser.add_argument('--lora_model',default=None,required=True,type=str,help="Please specify a lora_model")

# deprecated; the script infers the model size from the checkpoint
parser.add_argument('--model_size',default='7B',type=str,help="Size of the LLaMA model",choices=['7B','13B'])

parser.add_argument('--offload_dir',default=None,type=str,help="(Optional) Please specify a temp folder for offloading (useful for low-RAM machines). Default None (disable offload).")
parser.add_argument('--output_dir',default='./',type=str)
args = parser.parse_args()


assert (
    "LlamaTokenizer" in transformers._import_structure["models.llama"]
), "LLaMA is now in HuggingFace's main branch.\nPlease reinstall it: pip uninstall transformers && pip install git+https://github.com/huggingface/transformers.git"
from transformers import LlamaTokenizer, LlamaForCausalLM

BASE_MODEL = args.base_model
LORA_MODEL = args.lora_model
output_dir = args.output_dir

assert (
    BASE_MODEL
), "Please specify a BASE_MODEL in the script, e.g. 'decapoda-research/llama-7b-hf'"

tokenizer = LlamaTokenizer.from_pretrained(LORA_MODEL)
if args.offload_dir is not None:
    # Load with offloading, which is useful for low-RAM machines.
    # Note that if you have enough RAM, please use original method instead, as it is faster.
    base_model = LlamaForCausalLM.from_pretrained(
        BASE_MODEL,
        load_in_8bit=False,
        torch_dtype=torch.float16,
        offload_folder=args.offload_dir,
        offload_state_dict=True,
        low_cpu_mem_usage=True,
        device_map={"": "cpu"},
    )
else:
    # Original method without offloading
    base_model = LlamaForCausalLM.from_pretrained(
        BASE_MODEL,
        load_in_8bit=False,
        torch_dtype=torch.float16,
        device_map={"": "cpu"},
    )

base_model.resize_token_embeddings(len(tokenizer))
assert base_model.get_input_embeddings().weight.size(0) == len(tokenizer)
tokenizer.save_pretrained(output_dir)
print(f"Extended vocabulary size: {len(tokenizer)}")

first_weight = base_model.model.layers[0].self_attn.q_proj.weight
first_weight_old = first_weight.clone()

## infer the model size from the checkpoint
emb_to_model_size = {
    4096 : '7B',
    5120 : '13B',
    6656 : '30B',
    8192 : '65B',
}
embedding_size = base_model.get_input_embeddings().weight.size(1)
model_size = emb_to_model_size[embedding_size]
print(f"Loading LoRA for {model_size} model")

lora_model = PeftModel.from_pretrained(
    base_model,
    LORA_MODEL,
    device_map={"": "cpu"},
    torch_dtype=torch.float16,
)

assert torch.allclose(first_weight_old, first_weight)
# merge weights
print(f"Peft version: {peft.__version__}")
print(f"Merging model")
if peft.__version__ > '0.2.0':
    # merge weights - new merging method from peft
    lora_model = lora_model.merge_and_unload()
else:
    # merge weights
    for layer in lora_model.base_model.model.model.layers:
        if hasattr(layer.self_attn.q_proj,'merge_weights'):
            layer.self_attn.q_proj.merge_weights = True
        if hasattr(layer.self_attn.v_proj,'merge_weights'):
            layer.self_attn.v_proj.merge_weights = True
        if hasattr(layer.self_attn.k_proj,'merge_weights'):
            layer.self_attn.k_proj.merge_weights = True
        if hasattr(layer.self_attn.o_proj,'merge_weights'):
            layer.self_attn.o_proj.merge_weights = True
        if hasattr(layer.mlp.gate_proj,'merge_weights'):
            layer.mlp.gate_proj.merge_weights = True
        if hasattr(layer.mlp.down_proj,'merge_weights'):
            layer.mlp.down_proj.merge_weights = True
        if hasattr(layer.mlp.up_proj,'merge_weights'):
            layer.mlp.up_proj.merge_weights = True

lora_model.train(False)

# did we do anything?
assert not torch.allclose(first_weight_old, first_weight)

lora_model_sd = lora_model.state_dict()
del lora_model, base_model

num_shards_of_models = {'7B': 1, '13B': 2}
params_of_models = {
    '7B':
        {
        "dim": 4096,
        "multiple_of": 256,
        "n_heads": 32,
        "n_layers": 32,
        "norm_eps": 1e-06,
        "vocab_size": -1,
        },
    '13B':
        {
        "dim": 5120,
        "multiple_of": 256,
        "n_heads": 40,
        "n_layers": 40,
        "norm_eps": 1e-06,
        "vocab_size": -1,
        },
}

params = params_of_models[model_size]
num_shards = num_shards_of_models[model_size]


n_layers = params["n_layers"]
n_heads = params["n_heads"]
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))


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


def unpermute(w):
    return (
        w.view(n_heads, 2, dim // n_heads // 2, dim).transpose(1, 2).reshape(dim, dim)
    )


def translate_state_dict_key(k):
    k = k.replace("base_model.model.", "")
    if k == "model.embed_tokens.weight":
        return "tok_embeddings.weight"
    elif k == "model.norm.weight":
        return "norm.weight"
    elif k == "lm_head.weight":
        return "output.weight"
    elif k.startswith("model.layers."):
        layer = k.split(".")[2]
        if k.endswith(".self_attn.q_proj.weight"):
            return f"layers.{layer}.attention.wq.weight"
        elif k.endswith(".self_attn.k_proj.weight"):
            return f"layers.{layer}.attention.wk.weight"
        elif k.endswith(".self_attn.v_proj.weight"):
            return f"layers.{layer}.attention.wv.weight"
        elif k.endswith(".self_attn.o_proj.weight"):
            return f"layers.{layer}.attention.wo.weight"
        elif k.endswith(".mlp.gate_proj.weight"):
            return f"layers.{layer}.feed_forward.w1.weight"
        elif k.endswith(".mlp.down_proj.weight"):
            return f"layers.{layer}.feed_forward.w2.weight"
        elif k.endswith(".mlp.up_proj.weight"):
            return f"layers.{layer}.feed_forward.w3.weight"
        elif k.endswith(".input_layernorm.weight"):
            return f"layers.{layer}.attention_norm.weight"
        elif k.endswith(".post_attention_layernorm.weight"):
            return f"layers.{layer}.ffn_norm.weight"
        elif k.endswith("rotary_emb.inv_freq") or "lora" in k:
            return None
        else:
            print(layer, k)
            raise NotImplementedError
    else:
        print(k)
        raise NotImplementedError


def save_shards(lora_model_sd, num_shards: int):
    # Add the no_grad context manager
    with torch.no_grad():
        if num_shards == 1:
            new_state_dict = {}
            for k, v in lora_model_sd.items():
                new_k = translate_state_dict_key(k)
                if new_k is not None:
                    if "wq" in new_k or "wk" in new_k:
                        new_state_dict[new_k] = unpermute(v)
                    else:
                        new_state_dict[new_k] = v

            os.makedirs(output_dir, exist_ok=True)
            print(f"Saving shard 1 of {num_shards} into {output_dir}/consolidated.00.pth")
            torch.save(new_state_dict, output_dir + "/consolidated.00.pth")
            with open(output_dir + "/params.json", "w") as f:
                json.dump(params, f)
        else:
            new_state_dicts = [dict() for _ in range(num_shards)]
            for k in list(lora_model_sd.keys()):
                v = lora_model_sd[k]
                new_k = translate_state_dict_key(k)
                if new_k is not None:
                    if new_k=='tok_embeddings.weight':
                        print(f"Processing {new_k}")
                        assert v.size(1)%num_shards==0
                        splits = v.split(v.size(1)//num_shards,dim=1)
                    elif new_k=='output.weight':
                        print(f"Processing {new_k}")
                        splits = v.split(v.size(0)//num_shards,dim=0)

                    elif new_k=='norm.weight':
                        print(f"Processing {new_k}")
                        splits = [v] * num_shards
                    elif 'ffn_norm.weight' in new_k:
                        print(f"Processing {new_k}")
                        splits = [v] * num_shards
                    elif 'attention_norm.weight' in new_k:
                        print(f"Processing {new_k}")
                        splits = [v] * num_shards


                    elif 'w1.weight' in new_k:
                        print(f"Processing {new_k}")
                        splits = v.split(v.size(0)//num_shards,dim=0)
                    elif 'w2.weight' in new_k:
                        print(f"Processing {new_k}")
                        splits = v.split(v.size(1)//num_shards,dim=1)
                    elif 'w3.weight' in new_k:
                        print(f"Processing {new_k}")
                        splits = v.split(v.size(0)//num_shards,dim=0)


                    elif 'wo.weight' in new_k:
                        print(f"Processing {new_k}")
                        splits = v.split(v.size(1)//num_shards,dim=1)

                    elif 'wv.weight' in new_k:
                        print(f"Processing {new_k}")
                        splits = v.split(v.size(0)//num_shards,dim=0)

                    elif "wq.weight" in new_k or "wk.weight" in new_k:
                        print(f"Processing {new_k}")
                        v = unpermute(v)
                        splits = v.split(v.size(0)//num_shards,dim=0)
                    else:
                        print(f"Unexpected key {new_k}")
                        raise ValueError
                    for sd,split in zip(new_state_dicts,splits):
                        sd[new_k] = split.clone()
                        del split
                    del splits
                del lora_model_sd[k],v
                gc.collect()    # Effectively enforce garbage collection

            os.makedirs(output_dir, exist_ok=True)
            for i,new_state_dict in enumerate(new_state_dicts):
                print(f"Saving shard {i+1} of {num_shards} into {output_dir}/consolidated.0{i}.pth")
                torch.save(new_state_dict, output_dir + f"/consolidated.0{i}.pth")
            with open(output_dir + "/params.json", "w") as f:
                print(f"Saving params.json into {output_dir}/params.json")
                json.dump(params, f)


save_shards(lora_model_sd=lora_model_sd, num_shards=num_shards)

执行命令：

python merge_llama_with_chinese_lora.py --base_model path_to_original_llama_hf_dir --lora_model chinese-alpaca-lora-7b --output_dir path_to_output_dir

参数说明：

--base_model：存放HF格式的LLaMA模型权重和配置文件的目录（前面步骤中转的hf格式）
--lora_model：扩充了中文的模型目录
--output_dir：指定保存全量模型权重的目录，默认为./（合并出来的目录）
（可选）--offload_dir：对于低内存用户需要指定一个offload缓存路径

更详细的请看开原仓库：GitHub - ymcui/Chinese-LLaMA-Alpaca: 中文LLaMA&Alpaca大语言模型+本地CPU/GPU部署 (Chinese LLaMA & Alpaca LLMs)

到这里就已经合并好模型了, 目录：

接下来就准备部署吧

部署模型

我们需要先下载llama.cpp进行模型的量化, 输入以下命令:

git clone https://github.com/ggerganov/llama.cpp

目录如：

重点来了, 在窗口中分别输入以下命令进入llama.cpp, 然后下载依赖并编译

cd llama.cpp

cmake . -G "MinGW Makefiles"

make

走完以上步骤后你应该能在llama.cpp的bin目录内看到以下文件：

如果没有以上的文件, 那你应该是报错了, 基本上要么就是下载依赖的地方错, 要么就是编译的地方出错, 我在这里摸索了好久

接下来在llama.cpp内新建一个zh-models文件夹, 准备生成量化版本模型

zh-models的目录格式如下：

zh-models/

        - 7B/ #这是一个名为7B的文件夹
                - consolidated.00.pth
                - params.json
        - tokenizer.model

把path_to_output_dir文件夹内的consolidated.00.pth和params.json文件放入上面格式中的位置

把path_to_output_dir文件夹内的tokenizer.model文件放在跟7B文件夹同级的位置

接着在窗口中输入命令将上述.pth模型权重转换为ggml的FP16格式，生成文件路径为zh-models/7B/ggml-model-f16.bin

python convert-pth-to-ggml.py zh-models/7B/ 1

进一步对FP16模型进行4-bit量化，生成量化模型文件路径为zh-models/7B/ggml-model-q4_0.bin

D:\llama\llama.cpp\bin\quantize.exe ./zh-models/7B/ggml-model-f16.bin ./zh-models/7B/ggml-model-q4_0.bin 2

quantize.exe文件在bin目录内, 自行根据路径更改

到这就已经量化好了, 可以进行部署看看效果了, 部署的话如果你电脑配置好的可以选择部署f16的,否则就部署q4_0的....

D:\llama\llama.cpp\bin\main.exe -m zh-models/7B/ggml-model-q4_0.bin --color -f prompts/alpaca.txt -ins -c 2048 --temp 0.2 -n 256 --repeat_penalty 1.3

在提示符 > 之后输入你的prompt，cmd/ctrl+c中断输出，多行信息以\作为行尾

常用参数（更多参数请执行D:\llama\llama.cpp\bin\main.exe -h命令）：

-ins 启动类ChatGPT对话交流的运行模式
-f 指定prompt模板，alpaca模型请加载prompts/alpaca.txt
-c 控制上下文的长度，值越大越能参考更长的对话历史（默认：512）
-n 控制回复生成的最大长度（默认：128）
-b 控制batch size（默认：8），可适当增加
-t 控制线程数量（默认：4），可适当增加
--repeat_penalty 控制生成回复中对重复文本的惩罚力度
--temp 温度系数，值越低回复的随机性越小，反之越大
--top_p, top_k 控制解码采样的相关参数

想要部署f16的可以把命令中-m参数换成zh-models/7B/ggml-model-f16.bin即可

部署效果：