Llama.cpp 能 CPU & GPU 环境混合推理，这里记录一下在 windows10 平台上运行 Qwen-1.8B 的过程，显卡是 1660Ti 。

准备模型

• 安装conda
• Tun模式(管理员权限)

conda create -n llamaConvert python=3.10 git -c conda-forge
conda activate llamaConvert
cd D:\llama
git clone --depth=1 https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
python -m pip install -r requirements.txt
pip install tiktoken

python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='Qwen/Qwen-1_8B-Chat', local_dir=r'D:\qwen', ignore_patterns=['*.h5', '*.ot', '*.msgpack', '*.safetensors'])"
cd D:\qwen
D:\aria2\aria2c.exe --all-proxy='http://127.0.0.1:7890' -o 'model-00001-of-00002.safetensors' "https://huggingface.co/Qwen/Qwen-1_8B-Chat/resolve/main/model-00001-of-00002.safetensors?download=true"
D:\aria2\aria2c.exe --all-proxy='http://127.0.0.1:7890' -o 'model-00002-of-00002.safetensors' "https://huggingface.co/Qwen/Qwen-1_8B-Chat/resolve/main/model-00002-of-00002.safetensors?download=true"

cd D:\llama\llama.cpp
python convert-hf-to-gguf.py D:\qwen
# Model successfully exported to 'D:\qwen\ggml-model-f16.gguf'

运行模型

• 下载 llama-b1732-bin-win-cublas-cu12.2.0-x64.zip
• 提取文件到 D:\llama

conda create -n llamaCpp libcublas cuda-toolkit git -c nvidia -c conda-forge
conda activate llamaCpp
cd D:\llama ; .\main.exe ## 检查能否正确运行
cd D:\llama ; .\quantize.exe --help ## 自己决定量化方式
.\quantize.exe D:\qwen\ggml-model-f16.gguf .\qwen-1_8-f16.gguf COPY
.\server.exe -m .\qwen-1_8-f16.gguf -c 4096 --n-gpu-layers 50 ## 调节 n-gpu-layers 平衡 CPU & GPU

• 访问 http://127.0.0.1:8080 选择 Completion 进行测试

微调模型

• h-corpus数据集
• 官方微调教程

附加 Yi-6B-Chat

Yi-6B是零一万物开源的双语语言模型，经过3T多语种语料库的训练，在语言理解、常识推理、阅读理解等方面有一定潜力。

cd D:\models\01yi
D:\aria2\aria2c.exe --all-proxy='http://127.0.0.1:7890' -o 'model-00001-of-00003.safetensors' "https://huggingface.co/01-ai/Yi-6B-Chat/resolve/main/model-00001-of-00003.safetensors?download=true"
D:\aria2\aria2c.exe --all-proxy='http://127.0.0.1:7890' -o 'model-00002-of-00003.safetensors' "https://huggingface.co/01-ai/Yi-6B-Chat/resolve/main/model-00002-of-00003.safetensors?download=true"
D:\aria2\aria2c.exe --all-proxy='http://127.0.0.1:7890' -o 'model-00003-of-00003.safetensors'  https://huggingface.co/01-ai/Yi-6B-Chat/resolve/main/model-00003-of-00003.safetensors?download=true
conda activate llamaConvert
python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='01-ai/Yi-6B-Chat', local_dir=r'D:\models\01yi', ignore_patterns=['*.h5', '*.ot', '*.msgpack', '*.safetensors'])"

conda activate llamaConvert
cd D:\llama\llama.cpp
python convert.py D:\models\01yi
# Wrote D:\models\01yi\ggml-model-f16.gguf
conda activate llamaCpp
cd D:\llama ; .\quantize.exe --help
.\quantize.exe D:\models\01yi\ggml-model-f16.gguf .\01yi-6b-Q4_K_M.gguf Q4_K_M
.\server.exe -m .\01yi-6b-Q4_K_M.gguf -c 4096 --n-gpu-layers 50

附加 百川2

cd D:\models\baichuan
D:\aria2\aria2c.exe --all-proxy='http://127.0.0.1:7890' -o 'pytorch_model.bin' "https://huggingface.co/baichuan-inc/Baichuan2-7B-Chat/resolve/main/pytorch_model.bin?download=true"
conda activate llamaConvert
python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='baichuan-inc/Baichuan2-7B-Chat', local_dir=r'D:\models\baichuan', ignore_patterns=['*.h5', '*.bin', '*.ot', '*.msgpack', '*.safetensors'])"
cd D:\llama\llama.cpp
python convert.py D:\models\baichuan
# Wrote D:\models\baichuan\ggml-model-f16.gguf
conda activate llamaCpp
cd D:\llama ; .\quantize.exe --help
.\quantize.exe D:\models\baichuan\ggml-model-f16.gguf .\baichuan-7b-Q3_K_M.gguf Q3_K_M
.\server.exe -m .\baichuan-7b-Q3_K_M.gguf -c 2048 --n-gpu-layers 30

附加 tigerbot-13b

tigerbot-13b 在 chinese-llm-benchmark 上排名靠前。

cd D:\models\tigerbot
D:\aria2\aria2c.exe --all-proxy='http://127.0.0.1:7890' -o 'pytorch_model-00001-of-00003.bin' --max-download-limit=6M "https://huggingface.co/TigerResearch/tigerbot-13b-chat-v5/resolve/main/pytorch_model-00001-of-00003.bin?download=true"
D:\aria2\aria2c.exe --all-proxy='http://127.0.0.1:7890' -o 'pytorch_model-00002-of-00003.bin' --max-download-limit=6M "https://huggingface.co/TigerResearch/tigerbot-13b-chat-v5/resolve/main/pytorch_model-00002-of-00003.bin?download=true"
D:\aria2\aria2c.exe --all-proxy='http://127.0.0.1:7890' -o 'pytorch_model-00003-of-00003.bin' --max-download-limit=6M "https://huggingface.co/TigerResearch/tigerbot-13b-chat-v5/resolve/main/pytorch_model-00003-of-00003.bin?download=true"
conda activate llamaConvert
python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='TigerResearch/tigerbot-13b-chat-v5', local_dir=r'D:\models\tigerbot', ignore_patterns=['*.h5', '*.bin', '*.ot', '*.msgpack', '*.safetensors'])"
cd D:\llama\llama.cpp
python convert.py D:\models\tigerbot --padvocab
cd D:\llama ; .\quantize.exe --help
.\quantize.exe D:\models\tigerbot\ggml-model-f16.gguf D:\models\tigerbot-13B-Chat-Q4_K_M.gguf Q4_K_M
.\server.exe -m D:\models\tigerbot-13B-Chat-Q4_K_M.gguf -c 4096

感觉 6G 显存下，比较好用的是 Yi-6B-Chat-Q4_K_M
tigerbot-13b 在 R5 5600H 上推理速度 4.6 tokens/s，CPU 使用率 60%，频率 3.5GHz，应该是内存带宽瓶颈

附加 在 Colab 上量化

• llm2gguf.ipynb
• 量化后的结果

安装 llama.cpp

!git clone --depth=1 https://github.com/ggerganov/llama.cpp.git
%cd /content/llama.cpp
!LLAMA_CUDA=1 make -j

计算 imatrix

%cd /content
!wget -O transient.txt.gz https://huggingface.co/datasets/Limour/b-corpus/resolve/main/00-preview/00-transient.txt.gz?download=true
!gunzip transient.txt.gz
!mkdir -p /content/CausalLM-14B-GGUF
!wget -O /content/CausalLM-14B-GGUF/causallm_14b.Q8_0.gguf  https://huggingface.co/TheBloke/CausalLM-14B-GGUF/resolve/main/causallm_14b.Q8_0.gguf?download=true
!/content/llama.cpp/imatrix -m /content/CausalLM-14B-GGUF/causallm_14b.Q8_0.gguf -f /content/transient.txt -ngl 36

登录拥抱脸

from google.colab import userdata
from huggingface_hub import login
# login(token=os.environ.get("HF_TOKEN"), write_permission=True)
login(token=userdata.get('HF_TOKEN'), write_permission=True)
# from huggingface_hub import notebook_login
# notebook_login()

(跳过) 转换模型

%cd llama.cpp
!python -m pip install -r requirements.txt
!pip install tiktoken
from huggingface_hub import snapshot_download
!mkdir -p ~/CausalLM
snapshot_download(repo_id='CausalLM/7B', local_dir=r'/content/CausalLM', ignore_patterns=['*.h5', '*.ot', '*.msgpack', '*.safetensors'])
!python convert.py --vocab-type bpe --pad-vocab --outtype f16 /content/CausalLM 

量化模型

!/content/llama.cpp/quantize --allow-requantize --imatrix /content/imatrix.dat /content/CausalLM-14B-GGUF/causallm_14b.Q8_0.gguf /content/CausalLM-14B-GGUF/causallm_14b.IQ3_XS.gguf IQ3_XS

上传模型

from huggingface_hub import HfApi
api = HfApi()
api.upload_file(
    path_or_fileobj="/content/CausalLM-14B-GGUF/causallm_14b.IQ3_XS.gguf",
    path_in_repo="causallm_14b.IQ3_XS.gguf",
    repo_id="Limour/CausalLM-14B-GGUF"
)