XPU Weekly Report

Getting Started

1. Build the Docker Image

bash docker_build.sh

The image is based on intel/oneapi:2025.3.1-0-devel-ubuntu24.04 and installs Python, ccache, Miniforge, and other dependencies. The first build may take a while. Build logs are saved to compose.txt.

---

2. Create and Enter the Container

Start the container using docker-compose.yml:

# Start the container in the background
docker compose up -d

# Enter the container
docker compose exec node bash

Or start it directly with docker run:

docker run -it --rm \
  --privileged \
  --ipc=host \
  --network=host \
  --device /dev/dri:/dev/dri \
  -v /mnt:/mnt \
  -v $(pwd)/..:/workspace \
  -w /workspace/xpu_weekly_report \
  xpu-perf-weekly:latest \
  bash

Note: The container_name in docker-compose.yml is yupengzh-xpu-perf. If multiple users share the same machine, update name and container_name in docker-compose.yml accordingly.

---

3. Set Up the Environment (Inside the Container)

After entering the container, activate the conda environment and install all dependencies:

# Option 1: Run the full setup automatically (recommended for first-time use)
bash setup_baremetal.sh

# Option 2: Activate an already-set-up environment
source activate_env.sh

setup_baremetal.sh performs the following steps:

1. Install system packages (ccache, intel-ocloc)
2. Create / activate the conda environment xpu-perf-test (Python 3.12)
3. Install PyTorch XPU, numpy, and ninja
4. Build vllm-xpu-kernels, auto-round, sycl-tla, and oneDNN
5. Prepare the xpu-perf repository

---

4. Run Tests

bash run_test.sh [REPORT_DIR]

• REPORT_DIR: Optional. Specifies the output directory for reports. Defaults to reports/reports_<timestamp>.
• Results are saved under the specified REPORT_DIR after the tests complete.

---

5. Stop / Remove the Container

docker compose down