ipex-llm/python/llm/example/NPU/HF-Transformers-AutoModels/LLM/CPP_Examples

C++ Example of running LLM on Intel NPU using IPEX-LLM (Experimental)

In this directory, you will find a C++ example on how to run LLM models on Intel NPUs using IPEX-LLM (leveraging Intel NPU Acceleration Library). See the table blow for verified models.

Verified Models

Model	Model Link
Qwen2	Qwen/Qwen2-7B-Instruct, Qwen/Qwen2-1.5B-Instruct
Qwen2.5	Qwen/Qwen2.5-7B-Instruct
Llama2	meta-llama/Llama-2-7b-chat-hf
Llama3	meta-llama/Meta-Llama-3-8B-Instruct
MiniCPM	openbmb/MiniCPM-1B-sft-bf16, openbmb/MiniCPM-2B-sft-bf16

0. Requirements

To run this C++ example with IPEX-LLM on Intel NPUs, make sure to install the newest driver version of Intel NPU. Go to https://www.intel.com/content/www/us/en/download/794734/intel-npu-driver-windows.html to download and unzip the driver. Then go to Device Manager, find Neural Processors -> Intel(R) AI Boost. Right click and select Update Driver -> Browse my computer for drivers. And then manually select the unzipped driver folder to install.

1. Install

1.1 Installation on Windows

We suggest using conda to manage environment:

conda create -n llm python=3.10
conda activate llm

:: install ipex-llm with 'npu' option
pip install --pre --upgrade ipex-llm[npu]

:: [optional] for Llama-3.2-1B-Instruct & Llama-3.2-3B-Instruct
pip install transformers==4.45.0 accelerate==0.33.0

2. Convert Model

We provide a convert script under current directory, by running it, you can obtain the whole weights and configuration files which are required to run C++ example.

:: to convert Qwen2.5-7b-Instruct
python convert.py --repo-id-or-model-path Qwen/Qwen2.5-7B-Instruct --save-directory <converted_model_path>

Arguments info:

• --repo-id-or-model-path REPO_ID_OR_MODEL_PATH: argument defining the huggingface repo id for the model (e.g. Qwen/Qwen2.5-7B-Instruct) to be downloaded, or the path to the huggingface checkpoint folder.
• --save-directory SAVE_DIRECTORY: argument defining the path to save converted model. If it is a non-existing path, the original pretrained model specified by REPO_ID_OR_MODEL_PATH will be loaded, and the converted model will be saved into SAVE_DIRECTORY.
• --max-context-len MAX_CONTEXT_LEN: Defines the maximum sequence length for both input and output tokens. It is default to be 1024.
• --max-prompt-len MAX_PROMPT_LEN: Defines the maximum number of tokens that the input prompt can contain. It is default to be 960.
• --disable-transpose-value-cache: Disable the optimization of transposing value cache.

3. Build C++ Example llm-npu-cli

You can run below cmake script in cmd to build llm-npu-cli, don't forget to replace below conda env dir with your own path.

:: under current directory
:: please replace below conda env dir with your own path
set CONDA_ENV_DIR=C:\Users\arda\miniforge3\envs\llm\Lib\site-packages
mkdir build
cd build
cmake ..
cmake --build . --config Release -j
cd Release

4. Run llm-npu-cli

With built llm-npu-cli, you can run the example with specified paramaters. For example,

llm-npu-cli.exe -m <converted_model_path> -n 64 "AI是什么?"

Arguments info:

• -m : argument defining the path of saved converted model.
• -n : argument defining how many tokens will be generated.
• Last argument is your input prompt.

5. Sample Output

Qwen/Qwen2.5-7B-Instruct

Input:
<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
AI是什么?<|im_end|>
<|im_start|>assistant


Prefill 22 tokens cost xxxx ms.
Output:
AI是"人工智能"的缩写，是英文"Artificial Intelligence"的翻译。它是研究如何使计算机也具有智能的一种技术和理论。简而言之，人工智能就是让计算机能够模仿人智能行为的一项技术。

Decode 46 tokens cost xxxx ms (avg xx.xx ms each token).

Troubleshooting

Program crash with Chinese prompt

If you run CPP examples on Windows and find that your program raises below error when accepting Chinese prompts, you can open Region->Administrative->Change System locale.., check Beta: Use Unicode UTF-8 for worldwide language support option and then restart your computer.

thread '<unnamed>' panicked at src\lib.rs:151:91:
called `Result::unwrap()` on an `Err` value: Utf8Error { valid_up_to: 77, error_len: Some(1) }

For detailed instructions on how to do this, see this issue.