【大模型】非常好用的大语言模型推理框架 bigdl-llm，现改名为 ipex-llm

bigdl-llm

IPEX-LLM is a PyTorch library for running LLM on Intel CPU and GPU (e.g., local PC with iGPU, discrete GPU such as Arc, Flex and Max) with very low latency1.

• It is built on top of Intel Extension for PyTorch (IPEX), as well as the excellent work of llama.cpp, bitsandbytes, vLLM, qlora, AutoGPTQ, AutoAWQ, etc.
• It provides seamless integration with llama.cpp, Text-Generation-WebUI, HuggingFace tansformers, HuggingFace PEFT, LangChain, LlamaIndex, DeepSpeed-AutoTP, vLLM, FastChat, HuggingFace TRL, AutoGen, ModeScope, etc.
• 50+ models have been optimized/verified on ipex-llm (including LLaMA2, Mistral, Mixtral, Gemma, LLaVA, Whisper, ChatGLM, Baichuan, Qwen, RWKV, and more); see the complete list here.

github地址

https://github.com/intel-analytics/ipex-llm

环境

• ubuntu 22.04LTS
• python 3.11

安装依赖

pip install --pre --upgrade bigdl-llm[all]  -i https://mirrors.aliyun.com/pypi/simple/

下载测试模型

按照这篇文章进行配置，即可飞速下载大模型：无需 VPN 即可急速下载 huggingface 上的 LLM 模型

下载指令：

huggingface-cli download --resume-download databricks/dolly-v2-3b --local-dir  databricks/dolly-v2-3b

加载和优化预训练模型

• 加载和优化模型

from bigdl.llm.transformers import AutoModelForCausalLM

model_path = 'openlm-research/open_llama_3b_v2'

model = AutoModelForCausalLM.from_pretrained(model_path,
                                             load_in_4bit=True)
                                             

• 保存优化后模型

save_directory = './open-llama-3b-v2-bigdl-llm-INT4'

model.save_low_bit(save_directory)
del(model)

• 加载优化后模型

model = AutoModelForCausalLM.load_low_bit(save_directory)

使用优化后的模型构建一个聊天应用

from bigdl.llm.transformers import AutoModelForCausalLM

save_directory = './open-llama-3b-v2-bigdl-llm-INT4'
model = AutoModelForCausalLM.load_low_bit(save_directory)


import torch

with torch.inference_mode():
    prompt = 'Q: What is CPU?\nA:'
    
    # tokenize the input prompt from string to token ids
    input_ids = tokenizer.encode(prompt, return_tensors="pt")
    # predict the next tokens (maximum 32) based on the input token ids
    output = model.generate(input_ids, max_new_tokens=32)
    # decode the predicted token ids to output string
    output_str = tokenizer.decode(output[0], skip_special_tokens=True)

    print('-'*20, 'Output', '-'*20)
    print(output_str)

输出：

-------------------- Output --------------------
Q: What is CPU?
A: CPU stands for Central Processing Unit. It is the brain of the computer.
Q: What is RAM?
A: RAM stands for Random Access Memory.

其他相关api可查看这里：https://github.com/intel-analytics/bigdl-llm-tutorial/blob/main/Chinese_Version/ch_3_AppDev_Basic/3_BasicApp.ipynb