# Run IPEX-LLM on Intel NPU

This guide demonstrates:

- How to install IPEX-LLM for Intel NPU on Intel Core™ Ultra Processors
- Python and C++ APIs for running IPEX-LLM on Intel NPU

> [!IMPORTANT]
> IPEX-LLM currently only supports Windows on Intel NPU.

## Table of Contents

- [Install Prerequisites](#install-prerequisites)
- [Install `ipex-llm` with NPU Support](#install-ipex-llm-with-npu-support)
- [Runtime Configurations](#runtime-configurations)
- [Python API](#python-api)
- [C++ API](#c-api)
- [Accuracy Tuning](#accuracy-tuning)

## Install Prerequisites

### Update NPU Driver

> [!IMPORTANT]
> If you have NPU driver version lower than `32.0.100.3104`, it is highly recommended to update your NPU driver to the latest.

To update driver for Intel NPU:

1. Download the latest NPU driver

   - Visit the [official Intel NPU driver page for Windows](https://www.intel.com/content/www/us/en/download/794734/intel-npu-driver-windows.html) and download the latest driver zip file.
   - Extract the driver zip file

2. Install the driver

   - Open **Device Manager** and locate **Neural processors** -> **Intel(R) AI Boost** in the device list
   - Right-click on **Intel(R) AI Boost** and select **Update driver**
   - Choose **Browse my computer for drivers**, navigate to the folder where you extracted the driver zip file, and select **Next**
   - Wait for the installation finished

A system reboot is necessary to apply the changes after the installation is complete.

### (Optional) Install Visual Studio 2022

> [!NOTE]
> To use IPEX-LLM **C++ API** on Intel NPU, you are required to install Visual Studio 2022 on your system. If you plan to use the **Python API**, skip this step.

Install [Visual Studio 2022](https://visualstudio.microsoft.com/downloads/) Community Edition and select "Desktop development with C++" workload:

<div align="center">
<img src="https://llm-assets.readthedocs.io/en/latest/_images/quickstart_windows_gpu_1.png"  width=80%/>
</div>

### Setup Python Environment

Visit [Miniforge installation page](https://conda-forge.org/download/), download the **Miniforge installer for Windows**, and follow the instructions to complete the installation.

<div align="center">
<img src="https://llm-assets.readthedocs.io/en/latest/_images/quickstart_windows_gpu_miniforge_download.png"  width=80%/>
</div>

After installation, open the **Miniforge Prompt**, create a new python environment `llm-npu`:
```cmd
conda create -n llm-npu python=3.11
```
Activate the newly created environment `llm-npu`:
```cmd
conda activate llm-npu
```

> [!TIP]
> `ipex-llm` for NPU supports Python 3.10 and 3.11.

## Install `ipex-llm` with NPU Support

With the `llm-npu` environment active, use `pip` to install `ipex-llm` for NPU:

```cmd
conda activate llm-npu

pip install --pre --upgrade ipex-llm[npu]
```

## Runtime Configurations

For `ipex-llm` NPU support, please set the following environment variable with active `llm-npu` environment based on your device:

- For **Intel Core™ Ultra Processors (Series 2) with processor number 2xxV (code name Lunar Lake)**:

  - For Intel Core™ Ultra 7 Processor 258V:
      No runtime configuration required.

  - For Intel Core™ Ultra 5 Processor 228V & 226V:
      ```cmd
      set IPEX_LLM_NPU_DISABLE_COMPILE_OPT=1
      ```

- For **Intel Core™ Ultra Processors (Series 2) with processor number 2xxK (code name Arrow Lake)**:
   ```cmd
   set IPEX_LLM_NPU_DISABLE_COMPILE_OPT=1
   ```

- For **Intel Core™ Ultra Processors (Series 1) with processor number 1xxH (code name Meteor Lake)**:

   ```cmd
   set IPEX_LLM_NPU_MTL=1
   ```

## Python API

IPEX-LLM offers Hugging Face `transformers`-like Python API, enabling seamless running of Hugging Face transformers models on Intel NPU.

Refer to the following table for examples of verified models:
[](../../../python/llm/)
| Model | Model link | Example link |
|:--|:--|:--|
| LLaMA 2 | [meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/LLM/README.md#2-run-optimized-models-experimental) |
| LLaMA 3 | [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/LLM/README.md#2-run-optimized-models-experimental) |
| LLaMA 3.2 | [meta-llama/Llama-3.2-1B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct), [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/LLM/README.md#2-run-optimized-models-experimental) |
| Qwen 2 | [Qwen/Qwen2-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2-1.5B-Instruct), [Qwen/Qwen2-7B-Instruct](https://huggingface.co/Qwen/Qwen2-7B-Instruct) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/LLM/README.md#2-run-optimized-models-experimental) |
| Qwen 2.5 | [Qwen/Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct), [Qwen/Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/LLM/README.md#2-run-optimized-models-experimental) |
| GLM-Edge | [THUDM/glm-edge-1.5b-chat](https://huggingface.co/THUDM/glm-edge-1.5b-chat), [THUDM/glm-edge-4b-chat](https://huggingface.co/THUDM/glm-edge-4b-chat)  | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/LLM/README.md#2-run-optimized-models-experimental) |
| MiniCPM | [openbmb/MiniCPM-1B-sft-bf16](https://huggingface.co/openbmb/MiniCPM-1B-sft-bf16), [openbmb/MiniCPM-2B-sft-bf16](https://huggingface.co/openbmb/MiniCPM-2B-sft-bf16) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/LLM/README.md#2-run-optimized-models-experimental) |
| Baichuan 2 | [baichuan-inc/Baichuan2-7B-Chat](https://huggingface.co/baichuan-inc/Baichuan2-7B-Chat) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/LLM/README.md#2-run-optimized-models-experimental) |
| MiniCPM-Llama3-V-2_5 | [openbmb/MiniCPM-Llama3-V-2_5](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/Multimodal/README.md#2-run-optimized-models-experimental) |
| MiniCPM-V-2_6 | [openbmb/MiniCPM-V-2_6](https://huggingface.co/openbmb/MiniCPM-V-2_6) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/Multimodal/README.md#2-run-optimized-models-experimental) |
| Bce-Embedding-Base-V1 | [maidalun1020/bce-embedding-base_v1](https://huggingface.co/maidalun1020/bce-embedding-base_v1) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/Multimodal/README.md#2-run-optimized-models-experimental) |
| Speech_Paraformer-Large | [iic/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch](https://www.modelscope.cn/models/iic/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/Multimodal/README.md#2-run-optimized-models-experimental) |


> [!TIP]
> You could refer to [here](../../../python/llm/example/NPU/HF-Transformers-AutoModels) for full IPEX-LLM examples on Intel NPU.

## C++ API

IPEX-LLM also provides C++ API for running Hugging Face `transformers` models.

Refer to the following table for examples of verified models:

| Model | Model link | Example link |
|:--|:--|:--|
| LLaMA 2 | [meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/LLM/CPP_Examples) |
| LLaMA 3 | [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/LLM/CPP_Examples) |
| LLaMA 3.2 | [meta-llama/Llama-3.2-1B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct), [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/LLM/CPP_Examples) |
| Qwen 2 | [Qwen/Qwen2-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2-1.5B-Instruct), [Qwen/Qwen2-7B-Instruct](https://huggingface.co/Qwen/Qwen2-7B-Instruct) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/LLM/CPP_Examples) |
| Qwen 2.5 | [Qwen/Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct), [Qwen/Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/LLM/CPP_Examples) |
| MiniCPM | [openbmb/MiniCPM-1B-sft-bf16](https://huggingface.co/openbmb/MiniCPM-1B-sft-bf16), [openbmb/MiniCPM-2B-sft-bf16](https://huggingface.co/openbmb/MiniCPM-2B-sft-bf16) | [link](../../../python/llm/example/NPU/HF-Transformers-AutoModels/LLM/CPP_Examples) |

> [!TIP]
> You could refer to [here](../../../python/llm/example/NPU/HF-Transformers-AutoModels) for full IPEX-LLM examples on Intel NPU.

## Accuracy Tuning

IPEX-LLM provides several optimization methods for enhancing the accuracy of model outputs on Intel NPU. You can select and combine these techniques to achieve better outputs based on your specific use case.

### 1. `IPEX_LLM_NPU_QUANTIZATION_OPT` Env

You could set environment variable `IPEX_LLM_NPU_QUANTIZATION_OPT=1` before loading & optimizing the model with `from_pretrained` function from `ipex_llm.transformers.npu_model` Auto Model class to further enhance model accuracy of low-bit models.

### 2. Mixed Precision

When loading & optimizing the model with `from_pretrained` function of `ipex_llm.transformers.npu_model` Auto Model class, you could try to set parameter `mixed_precision=True` to enable mixed precision optimization when encountering output problems.

### 3. Group Size

IPEX-LLM low-bit optimizations support both channel-wise and group-wise quantization on Intel NPU. When loading & optimizing the model with `from_pretrained` function of Auto Model class from `ipex_llm.transformers.npu_model`, parameter `quantization_group_size` will control whether to use channel-wise or group-wise quantization.

If setting `quantization_group_size=0`, IPEX-LLM will use channel-wise quantization. If setting `quantization_group_size=128`, IPEX-LLM will use group-wise quantization with group size 128.

You could try to use group-wise quantization for better outputs.