Deploy RecBole on a GPU Cloud to Learn Recommendation Algorithms

RecBole overview

RecBole is a PyTorch‑based recommendation system library aimed at researchers. It implements 94 recommendation models covering general, sequential, context‑aware and knowledge‑based categories and provides a unified data file format that supports 44 benchmark datasets.

Generic and extensible data structures enable unified handling of diverse recommendation datasets.

Comprehensive collection of benchmark models and datasets.

GPU‑accelerated implementations use a series of optimization techniques to improve efficiency.

Supports a wide range of standard evaluation metrics for large‑scale benchmarking.

Running RecBole on SumW GPU platform (MLU accelerator)

The SumW GPU platform offers a pre‑built RecBole‑MLU Docker image. After launching a JupyterLab instance, the following steps install and run RecBole on the Cambricon MLU accelerator.

Extract the provided source package and enter the directory:

tar -zxvf RecBole_MLU_Ready.tar.gz
cd RecBole_mlu

Configure pip to use the Aliyun mirror (required on mainland China to avoid download stalls):

pip config set global.index-url https://mirrors.aliyun.com/pypi/simple/

Install dependencies while preventing the default PyTorch package from overwriting the MLU‑optimized torch:

# Install all non‑torch dependencies
cat requirements.txt | grep -v "torch" | xargs pip install

# Install missing visualization and logging packages
pip install colorlog==4.7.2 plotly tabulate tensorboard texttable colorama==0.4.4

# Install RecBole without dependency checks
pip install -e . --no-deps

Run a baseline model with a single command. The --use_gpu=True flag maps computation to the MLU card; --gpu_id=0 selects the first accelerator:

python run_recbole.py --model=BPR --dataset=ml-100k --use_gpu=True --gpu_id=0

If the console displays the large RecBole logo and the progress bar advances rapidly, the MLU accelerator is operating at full speed.