Skip to content

Latest commit

 

History

History
196 lines (131 loc) · 7.35 KB

bmg_quickstart.md

File metadata and controls

196 lines (131 loc) · 7.35 KB

Install and Use IPEX-LLM on Intel Arc B-Series GPU (code-named Battlemage)

This guide demonstrates how to install and use IPEX-LLM on the Intel Arc B-Series GPU (such as B580).

Note

Ensure your GPU driver and software environment meet the prerequisites before proceeding.

Table of Contents

  1. Linux
    1.1 Install Prerequisites
    1.2 Install IPEX-LLM (for PyTorch and HuggingFace)
    1.3 Install IPEX-LLM (for llama.cpp and Ollama)
  2. Windows
    2.1 Install Prerequisites
    2.2 Install IPEX-LLM (for PyTorch and HuggingFace)
    2.3 Install IPEX-LLM (for llama.cpp and Ollama)
  3. Use Cases
    3.1 PyTorch
    3.2 Ollama
    3.3 llama.cpp
    3.4 vLLM
  4. Troubleshooting
    4.1 RuntimeError: could not create an engine
    4.2 Connection timeout error when installing the intel-graphics PPA

1. Linux

1.1 Install Prerequisites

Note

Ensure that Resizable BAR is enabled in your system's BIOS before proceeding. This is essential for optimal GPU performance and to avoid potential issues such as Bus error (core dumped). For detailed steps, please refer to the official guidance here.

We recommend using Ubuntu 24.10 and kernel version 6.11 or above, as support for Battle Mage has been backported from kernel version 6.12 to version 6.11, which is included in Ubuntu 24.10, according to the official documentation here. However, since this version of Ubuntu does not include the latest compute and media-related packages, we offer the intel-graphics Personal Package Archive (PPA). The PPA provides early access to newer packages, along with additional tools and features such as EU debugging.

Use the following commands to install the intel-graphics PPA and the necessary compute and media packages:

sudo apt-get update
sudo apt-get install -y software-properties-common
sudo add-apt-repository -y ppa:kobuk-team/intel-graphics
sudo apt-get install -y libze-intel-gpu1 libze1 intel-ocloc intel-opencl-icd clinfo intel-gsc intel-media-va-driver-non-free libmfx1 libmfx-gen1 libvpl2 libvpl-tools libva-glx2 va-driver-all vainfo
sudo apt-get install -y intel-level-zero-gpu-raytracing  # Optional: Hardware ray tracing support

Setup Python Environment

Download and install Miniforge:

wget https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Linux-x86_64.sh
bash Miniforge3-Linux-x86_64.sh
source ~/.bashrc

Create and activate a Python environment:

conda create -n llm python=3.11
conda activate llm

1.2 Install IPEX-LLM

With the llm environment active, install the appropriate ipex-llm package based on your use case:

For PyTorch and HuggingFace:

Install the ipex-llm[xpu_2.6] package:

pip install --pre --upgrade ipex-llm[xpu_2.6] --extra-index-url https://download.pytorch.org/whl/xpu

For llama.cpp and Ollama:

You may use Ollama Portable Zip and llama.cpp Portable Zip.

2. Windows

2.1 Install Prerequisites

Update GPU Driver

If your driver version is lower than 32.0.101.6449/32.0.101.101.6256, update it from the Intel download page. After installation, reboot the system.

Setup Python Environment

Download and install Miniforge for Windows from the official page. After installation, create and activate a Python environment:

conda create -n llm python=3.11
conda activate llm

2.2 Install IPEX-LLM

With the llm environment active, install the appropriate ipex-llm package based on your use case:

For PyTorch and HuggingFace:

Install the ipex-llm[xpu_2.6] package:

pip install --pre --upgrade ipex-llm[xpu_2.6] --extra-index-url https://download.pytorch.org/whl/xpu

For llama.cpp and Ollama:

You may use Ollama Portable Zip and llama.cpp Portable Zip.

3. Use Cases

3.1 PyTorch

Run a Quick PyTorch Example:

  1. Activate the environment: 
    conda activate llm  # On Windows, use 'cmd'
  2. Run the code: 
    import torch
from ipex_llm.transformers import AutoModelForCausalLM

tensor_1 = torch.randn(1, 1, 40, 128).to('xpu')
tensor_2 = torch.randn(1, 1, 128, 40).to('xpu')
print(torch.matmul(tensor_1, tensor_2).size())
  3. Expected Output: 
    torch.Size([1, 1, 40, 40])

Tip

Please refer to here (Linux or Windows) regarding runtime configurations for PyTorch with IPEX-LLM on B-Series GPU.

For benchmarks and performance measurement, refer to the Benchmark Quickstart guide.

3.2 Ollama

To integrate and run with Ollama, follow the Ollama Quickstart guide.

3.3 llama.cpp

For instructions on how to run llama.cpp with IPEX-LLM, refer to the llama.cpp Quickstart guide.

3.4 vLLM

To set up and run vLLM, follow the vLLM Quickstart guide.

4. Troubleshooting

4.1 RuntimeError: could not create an engine

image

If you encounter a RuntimeError like the output shown above while working on Linux after running conda deactivate and then reactivating your environment using conda activate env, the issue is likely caused by the OCL_ICD_VENDORS environment variable.

To fix this on Linux, run the following command:

unset OCL_ICD_VENDORS

This will remove the conflicting environment variable and allow your program to function correctly.

Note: This issue only occurs on Linux systems. It does not affect Windows environments.

4.2 Connection timeout error when installing the intel-graphics PPA

While installting prerequisites on Linux, if you encounter a connection timeout error when adding the intel-graphics PPA, consider disabling IPv6 first through:

sudo sysctl -w net.ipv6.conf.all.disable_ipv6=1
sudo sysctl -w net.ipv6.conf.default.disable_ipv6=1
sudo sysctl -w net.ipv6.conf.lo.disable_ipv6=1

Afterward, disconnect and reconnect your network adapter before attempting the installation again.

Tip

  • The disabling of IPv6 by the above command is temporary and will be reverted after a system reboot.
  • You could find more information about this issue here.