Important

bigdl-llm has now become ipex-llm (see the migration guide here); you may find the original BigDL project here.

---

💫 Intel® LLM library for PyTorch*

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.

Note

• It is built on top of the excellent work of llama.cpp, transformers, bitsandbytes, vLLM, qlora, AutoGPTQ, AutoAWQ, etc.
• It provides seamless integration with llama.cpp, Ollama, Text-Generation-WebUI, HuggingFace transformers, LangChain, LlamaIndex, DeepSpeed-AutoTP, vLLM, FastChat, Axolotl, HuggingFace PEFT, 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.

Latest Update 🔥

• [2024/07] We added extensive support for Large Multimodal Models, including StableDiffusion, Phi-3-Vision, Qwen-VL, and more.
• [2024/07] We added FP6 support on Intel GPU.
• [2024/06] We added experimental NPU support for Intel Core Ultra processors; see the examples here.
• [2024/06] We added extensive support of pipeline parallel inference, which makes it easy to run large-sized LLM using 2 or more Intel GPUs (such as Arc).
• [2024/06] We added support for running RAGFlow with ipex-llm on Intel GPU.
• [2024/05] You can now easily run ipex-llm inference, serving and finetuning using the Docker images.
• [2024/05] You can now install ipex-llm on Windows using just "one command".
• [2024/05] ipex-llm now supports Axolotl for LLM finetuning on Intel GPU; see the quickstart here.

More updates

• [2024/04] You can now run Open WebUI on Intel GPU using ipex-llm; see the quickstart here.
• [2024/04] You can now run Llama 3 on Intel GPU using llama.cpp and ollama with ipex-llm; see the quickstart here.
• [2024/04] ipex-llm now supports Llama 3 on both Intel GPU and CPU.
• [2024/04] ipex-llm now provides C++ interface, which can be used as an accelerated backend for running llama.cpp and ollama on Intel GPU.
• [2024/03] bigdl-llm has now become ipex-llm (see the migration guide here); you may find the original BigDL project here.
• [2024/02] ipex-llm now supports directly loading model from ModelScope (魔搭).
• [2024/02] ipex-llm added initial INT2 support (based on llama.cpp IQ2 mechanism), which makes it possible to run large-sized LLM (e.g., Mixtral-8x7B) on Intel GPU with 16GB VRAM.
• [2024/02] Users can now use ipex-llm through Text-Generation-WebUI GUI.
• [2024/02] ipex-llm now supports Self-Speculative Decoding, which in practice brings ~30% speedup for FP16 and BF16 inference latency on Intel GPU and CPU respectively.
• [2024/02] ipex-llm now supports a comprehensive list of LLM finetuning on Intel GPU (including LoRA, QLoRA, DPO, QA-LoRA and ReLoRA).
• [2024/01] Using ipex-llm QLoRA, we managed to finetune LLaMA2-7B in 21 minutes and LLaMA2-70B in 3.14 hours on 8 Intel Max 1550 GPU for Standford-Alpaca (see the blog here).
• [2023/12] ipex-llm now supports ReLoRA (see "ReLoRA: High-Rank Training Through Low-Rank Updates").
• [2023/12] ipex-llm now supports Mixtral-8x7B on both Intel GPU and CPU.
• [2023/12] ipex-llm now supports QA-LoRA (see "QA-LoRA: Quantization-Aware Low-Rank Adaptation of Large Language Models").
• [2023/12] ipex-llm now supports FP8 and FP4 inference on Intel GPU.
• [2023/11] Initial support for directly loading GGUF, AWQ and GPTQ models into ipex-llm is available.
• [2023/11] ipex-llm now supports vLLM continuous batching on both Intel GPU and CPU.
• [2023/10] ipex-llm now supports QLoRA finetuning on both Intel GPU and CPU.
• [2023/10] ipex-llm now supports FastChat serving on on both Intel CPU and GPU.
• [2023/09] ipex-llm now supports Intel GPU (including iGPU, Arc, Flex and MAX).
• [2023/09] ipex-llm tutorial is released.

ipex-llm Performance

See the Token Generation Speed on Intel Core Ultra and Intel Arc GPU below¹ (and refer to [2][3][4] for more details).

You may follow the Benchmarking Guide to run ipex-llm performance benchmark yourself.

ipex-llm Demo

See demos of running local LLMs on Intel Iris iGPU, Intel Core Ultra iGPU, single-card Arc GPU, or multi-card Arc GPUs using ipex-llm below.

Intel Iris iGPU	Intel Core Ultra iGPU	Intel Arc dGPU	2-Card Intel Arc dGPUs
llama.cpp (Phi-3-mini Q4_0)	Ollama (Mistral-7B Q4_K)	TextGeneration-WebUI (Llama3-8B FP8)	FastChat (QWen1.5-32B FP6)

Model Accuracy

Please see the Perplexity result below (tested on Wikitext dataset using the script here).

Perplexity	sym_int4	q4_k	fp6	fp8_e5m2	fp8_e4m3	fp16
Llama-2-7B-chat-hf	6.3638	6.2179	6.0924	6.1796	6.0980	6.0963
Mistral-7B-Instruct-v0.1	6.0025	5.9581	5.8930	5.8884	5.8820	5.8734
Qwen1.5-7B-chat	8.8652	8.8163	8.5573	8.8463	8.5304	8.6065

ipex-llm Quickstart

Docker

• GPU Inference in C++: running llama.cpp, ollama, OpenWebUI, etc., with ipex-llm on Intel GPU
• GPU Inference in Python : running HuggingFace transformers, LangChain, LlamaIndex, ModelScope, etc. with ipex-llm on Intel GPU
• vLLM on GPU: running vLLM serving with ipex-llm on Intel GPU
• vLLM on CPU: running vLLM serving with ipex-llm on Intel CPU
• FastChat on GPU: running FastChat serving with ipex-llm on Intel GPU
• VSCode on GPU: running and developing ipex-llm applications in Python using VSCode on Intel GPU

Use

• llama.cpp: running llama.cpp (using C++ interface of ipex-llm as an accelerated backend for llama.cpp) on Intel GPU
• Ollama: running ollama (using C++ interface of ipex-llm as an accelerated backend for ollama) on Intel GPU
• Llama 3 with llama.cpp and ollama: running Llama 3 on Intel GPU using llama.cpp and ollama with ipex-llm
• vLLM: running ipex-llm in vLLM on both Intel GPU and CPU
• FastChat: running ipex-llm in FastChat serving on on both Intel GPU and CPU
• Serving on multiple Intel GPUs: running ipex-llm serving on multiple Intel GPUs by leveraging DeepSpeed AutoTP and FastAPI
• Text-Generation-WebUI: running ipex-llm in oobabooga WebUI
• Axolotl: running ipex-llm in Axolotl for LLM finetuning
• Benchmarking: running (latency and throughput) benchmarks for ipex-llm on Intel CPU and GPU

Applications

• Local RAG: running LangChain-Chatchat (Knowledge Base QA using RAG pipeline) with ipex-llm
• Coding copilot: running Continue (coding copilot in VSCode) with ipex-llm
• Open WebUI: running Open WebUI with ipex-llm
• PrivateGPT: running PrivateGPT to interact with documents with ipex-llm
• Dify platform: running ipex-llm in Dify(production-ready LLM app development platform)
• RAGFlow: running RAGFlow (an open-source RAG engine) with ipex-llm

Install

• Windows GPU: installing ipex-llm on Windows with Intel GPU
• Linux GPU: installing ipex-llm on Linux with Intel GPU
• For more details, please refer to the full installation guide

Code Examples

• Low bit inference

  • INT4 inference: INT4 LLM inference on Intel GPU and CPU
  • FP8/FP6/FP4 inference: FP8, FP6 and FP4 LLM inference on Intel GPU
  • INT8 inference: INT8 LLM inference on Intel GPU and CPU
  • INT2 inference: INT2 LLM inference (based on llama.cpp IQ2 mechanism) on Intel GPU

• FP16/BF16 inference

  • FP16 LLM inference on Intel GPU, with possible self-speculative decoding optimization
  • BF16 LLM inference on Intel CPU, with possible self-speculative decoding optimization

• Distributed inference

  • Pipeline Parallel inference on Intel GPU
  • DeepSpeed AutoTP inference on Intel GPU

• Save and load

  • Low-bit models: saving and loading ipex-llm low-bit models (INT4/FP4/FP6/INT8/FP8/FP16/etc.)
  • GGUF: directly loading GGUF models into ipex-llm
  • AWQ: directly loading AWQ models into ipex-llm
  • GPTQ: directly loading GPTQ models into ipex-llm

• Finetuning

  • LLM finetuning on Intel GPU, including LoRA, QLoRA, DPO, QA-LoRA and ReLoRA
  • QLoRA finetuning on Intel CPU

• Integration with community libraries

  • HuggingFace transformers
  • Standard PyTorch model
  • LangChain
  • LlamaIndex
  • DeepSpeed-AutoTP
  • Axolotl
  • HuggingFace PEFT
  • HuggingFace TRL
  • AutoGen
  • ModeScope
• Tutorials

API Doc

• HuggingFace Transformers-style API (Auto Classes)
• API for arbitrary PyTorch Model

FAQ

• FAQ & Trouble Shooting

Verified Models

Over 50 models have been optimized/verified on ipex-llm, including LLaMA/LLaMA2, Mistral, Mixtral, Gemma, LLaVA, Whisper, ChatGLM2/ChatGLM3, Baichuan/Baichuan2, Qwen/Qwen-1.5, InternLM and more; see the list below.

Model	CPU Example	GPU Example
LLaMA (such as Vicuna, Guanaco, Koala, Baize, WizardLM, etc.)	link1, link2	link
LLaMA 2	link1, link2	link
LLaMA 3	link	link
ChatGLM	link
ChatGLM2	link	link
ChatGLM3	link	link
GLM-4	link	link
GLM-4V	link	link
Mistral	link	link
Mixtral	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
Qwen2	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
Phi-3	link	link
Phi-3-vision	link	link
Yuan2	link	link
Gemma	link	link
DeciLM-7B	link	link
Deepseek	link	link
StableLM	link	link
CodeGemma	link	link
Command-R/cohere	link	link
CodeGeeX2	link	link
MiniCPM	link	link

Get Support

• Please report a bug or raise a feature request by opening a Github Issue
• Please report a vulnerability by opening a draft GitHub Security Advisory

Footnotes

1. Performance varies by use, configuration and other factors. ipex-llm may not optimize to the same degree for non-Intel products. Learn more at www.Intel.com/PerformanceIndex. ↩ ↩²