llm

BigDL-LLM

bigdl-llm is a library for running LLM (large language model) on Intel XPU (from Laptop to GPU to Cloud) using INT4 with very low latency¹ (for any PyTorch model).

It is built on top of the excellent work of llama.cpp, gptq, ggml, llama-cpp-python, bitsandbytes, qlora, gptq_for_llama, chatglm.cpp, redpajama.cpp, gptneox.cpp, bloomz.cpp, etc.

Demos

See the optimized performance of chatglm2-6b and llama-2-13b-chat models on 12th Gen Intel Core CPU and Intel Arc GPU below.

12th Gen Intel Core CPU		Intel Arc GPU
chatglm2-6b	llama-2-13b-chat	chatglm2-6b	llama-2-13b-chat

Verified models

Over 20 models have been optimized/verified on bigdl-llm, including LLaMA/LLaMA2, ChatGLM/ChatGLM2, Mistral, Falcon, MPT, Dolly, StarCoder, Whisper, Baichuan, InternLM, QWen, Aquila, MOSS, and more; see the complete list below.

Model	CPU Example	GPU Example
LLaMA (such as Vicuna, Guanaco, Koala, Baize, WizardLM, etc.)	link1, link2	link
LLaMA 2	link1, link2	link1, link2-low GPU memory example
ChatGLM	link
ChatGLM2	link	link
ChatGLM3	link	link
Mistral	link	link
Falcon	link	link
MPT	link	link
Dolly-v1	link	link
Dolly-v2	link	link
Replit Code	link	link
RedPajama	link1, link2
Phoenix	link1, link2
StarCoder	link1, link2	link
Baichuan	link	link
Baichuan2	link	link
InternLM	link	link
Qwen	link	link
Qwen1.5	link	link
Qwen-VL	link	link
Aquila	link	link
Aquila2	link	link
MOSS	link
Whisper	link	link
Phi-1_5	link	link
Flan-t5	link	link
LLaVA	link	link
CodeLlama	link	link
Skywork	link
InternLM-XComposer	link
WizardCoder-Python	link
CodeShell	link
Fuyu	link
Distil-Whisper	link	link
Yi	link	link
BlueLM	link	link
Mamba	link	link
SOLAR	link	link
Phixtral	link	link
InternLM2	link	link
RWKV4		link
RWKV5		link
Bark	link	link
SpeechT5		link
DeepSeek-MoE	link
Ziya-Coding-34B-v1.0	link
Phi-2	link	link
Yuan2	link	link
DeciLM-7B	link	link
Deepseek	link	link

Working with bigdl-llm

Table of Contents

• BigDL-LLM
  • Demos
  • Verified models
  • Working with bigdl-llm
    • Install
      • CPU
      • GPU
    • Run Model
      • 1. Hugging Face transformers API
        • CPU INT4
        • GPU INT4
        • More Low-Bit Support
      • 2. Native INT4 model
      • 3. LangChain API
      • 4. CLI Tool
  • bigdl-llm API Doc
  • bigdl-llm Dependency

Install

CPU

You may install bigdl-llm on Intel CPU as follows:

pip install --pre --upgrade bigdl-llm[all]

Note: bigdl-llm has been tested on Python 3.9

GPU

You may install bigdl-llm on Intel GPU as follows:

# below command will install intel_extension_for_pytorch==2.0.110+xpu as default
# you can install specific ipex/torch version for your need
pip install --pre --upgrade bigdl-llm[xpu] -f https://developer.intel.com/ipex-whl-stable-xpu

Note: bigdl-llm has been tested on Python 3.9

Run Model

You may run the models using bigdl-llm through one of the following APIs:

1. Hugging Face transformers API
2. Native INT4 Model
3. LangChain API
4. CLI (command line interface) Tool

1. Hugging Face transformers API

You may run any Hugging Face Transformers model as follows:

CPU INT4

You may apply INT4 optimizations to any Hugging Face Transformers model on Intel CPU as follows.

#load Hugging Face Transformers model with INT4 optimizations
from bigdl.llm.transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained('/path/to/model/', load_in_4bit=True)

#run the optimized model on Intel CPU
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path)
input_ids = tokenizer.encode(input_str, ...)
output_ids = model.generate(input_ids, ...)
output = tokenizer.batch_decode(output_ids)

See the complete examples here.

GPU INT4

You may apply INT4 optimizations to any Hugging Face Transformers model on Intel GPU as follows.

#load Hugging Face Transformers model with INT4 optimizations
from bigdl.llm.transformers import AutoModelForCausalLM
import intel_extension_for_pytorch
model = AutoModelForCausalLM.from_pretrained('/path/to/model/', load_in_4bit=True)

#run the optimized model on Intel GPU
model = model.to('xpu')

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path)
input_ids = tokenizer.encode(input_str, ...).to('xpu')
output_ids = model.generate(input_ids, ...)
output = tokenizer.batch_decode(output_ids.cpu())

See the complete examples here.

More Low-Bit Support

• Save and load

  After the model is optimized using bigdl-llm, you may save and load the model as follows:

  model.save_low_bit(model_path)
  new_model = AutoModelForCausalLM.load_low_bit(model_path)

  See the complete example here.

• Additonal data types

  In addition to INT4, You may apply other low bit optimizations (such as INT8, INT5, NF4, etc.) as follows:

  model = AutoModelForCausalLM.from_pretrained('/path/to/model/', load_in_low_bit="sym_int8")

  See the complete example here.

2. Native INT4 model

You may also convert Hugging Face Transformers models into native INT4 model format for maximum performance as follows.

Notes: Currently only llama/bloom/gptneox/starcoder/chatglm model families are supported; for other models, you may use the Hugging Face transformers model format as described above).

#convert the model
from bigdl.llm import llm_convert
bigdl_llm_path = llm_convert(model='/path/to/model/',
        outfile='/path/to/output/', outtype='int4', model_family="llama")

#load the converted model
#switch to ChatGLMForCausalLM/GptneoxForCausalLM/BloomForCausalLM/StarcoderForCausalLM to load other models
from bigdl.llm.transformers import LlamaForCausalLM
llm = LlamaForCausalLM.from_pretrained("/path/to/output/model.bin", native=True, ...)
  
#run the converted model
input_ids = llm.tokenize(prompt)
output_ids = llm.generate(input_ids, ...)
output = llm.batch_decode(output_ids)

See the complete example here.

3. LangChain API

You may run the models using the LangChain API in bigdl-llm.

• Using Hugging Face transformers model

  You may run any Hugging Face Transformers model (with INT4 optimiztions applied) using the LangChain API as follows:

  from bigdl.llm.langchain.llms import TransformersLLM
  from bigdl.llm.langchain.embeddings import TransformersEmbeddings
  from langchain.chains.question_answering import load_qa_chain
  
  embeddings = TransformersEmbeddings.from_model_id(model_id=model_path)
  bigdl_llm = TransformersLLM.from_model_id(model_id=model_path, ...)
  
  doc_chain = load_qa_chain(bigdl_llm, ...)
  output = doc_chain.run(...)

  See the examples here.

• Using native INT4 model

  You may also convert Hugging Face Transformers models into native INT4 format, and then run the converted models using the LangChain API as follows.

  Notes:* Currently only llama/bloom/gptneox/starcoder/chatglm model families are supported; for other models, you may use the Hugging Face transformers model format as described above).

  from bigdl.llm.langchain.llms import LlamaLLM
  from bigdl.llm.langchain.embeddings import LlamaEmbeddings
  from langchain.chains.question_answering import load_qa_chain
  
  #switch to ChatGLMEmbeddings/GptneoxEmbeddings/BloomEmbeddings/StarcoderEmbeddings to load other models
  embeddings = LlamaEmbeddings(model_path='/path/to/converted/model.bin')
  #switch to ChatGLMLLM/GptneoxLLM/BloomLLM/StarcoderLLM to load other models
  bigdl_llm = LlamaLLM(model_path='/path/to/converted/model.bin')
  
  doc_chain = load_qa_chain(bigdl_llm, ...)
  doc_chain.run(...)

  See the examples here.

4. CLI Tool

Note: Currently bigdl-llm CLI supports LLaMA (e.g., vicuna), GPT-NeoX (e.g., redpajama), BLOOM (e.g., pheonix) and GPT2 (e.g., starcoder) model architecture; for other models, you may use the Hugging Face transformers or LangChain APIs.

• Convert model

  You may convert the downloaded model into native INT4 format using llm-convert.

  #convert PyTorch (fp16 or fp32) model; 
  #llama/bloom/gptneox/starcoder model family is currently supported
  llm-convert "/path/to/model/" --model-format pth --model-family "bloom" --outfile "/path/to/output/"
  
  #convert GPTQ-4bit model
  #only llama model family is currently supported
  llm-convert "/path/to/model/" --model-format gptq --model-family "llama" --outfile "/path/to/output/"

• Run model

  You may run the converted model using llm-cli or llm-chat (built on top of main.cpp in llama.cpp)

  #help
  #llama/bloom/gptneox/starcoder model family is currently supported
  llm-cli -x gptneox -h
  
  #text completion
  #llama/bloom/gptneox/starcoder model family is currently supported
  llm-cli -t 16 -x gptneox -m "/path/to/output/model.bin" -p 'Once upon a time,'
  
  #chat mode
  #llama/gptneox model family is currently supported
  llm-chat -m "/path/to/output/model.bin" -x llama

bigdl-llm API Doc

See the inital bigdl-llm API Doc here.

bigdl-llm Dependency

The native code/lib in bigdl-llm has been built using the following tools. Note that lower LIBC version on your Linux system may be incompatible with bigdl-llm.

Model family	Platform	Compiler	GLIBC
llama	Linux	GCC 11.2.1	2.17
llama	Windows	MSVC 19.36.32532.0
llama	Windows	GCC 13.1.0
gptneox	Linux	GCC 11.2.1	2.17
gptneox	Windows	MSVC 19.36.32532.0
gptneox	Windows	GCC 13.1.0
bloom	Linux	GCC 11.2.1	2.29
bloom	Windows	MSVC 19.36.32532.0
bloom	Windows	GCC 13.1.0
starcoder	Linux	GCC 11.2.1	2.29
starcoder	Windows	MSVC 19.36.32532.0
starcoder	Windows	GCC 13.1.0

Footnotes

1. Performance varies by use, configuration and other factors. bigdl-llm may not optimize to the same degree for non-Intel products. Learn more at www.Intel.com/PerformanceIndex. ↩