FEC

Build From Source

1. Build a conda environment python=3.8

   conda create -n your_env python=3.8
   conda activate your_env
   

2. Install pytorch with CUDA 11.3 . See pytorch documentation

   conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
   export CUB_DIR=/usr/local/cuda-11.3/include/cub
   export CUDA_BIN_PATH=/usr/local/cuda-11.3/
   export CUDACXX=/usr/local/cuda-11.3/bin/nvcc
   

3. Install Requirements

   pip install -r requirements.txt
   

4. Then, install FBGEMM_GPU from source

   cd third_party/fbgemm/fbgemm_gpu/
   conda install scikit-build jinja2 ninja cmake hypothesis
   python setup.py install
   

5. Next, install FEC (FBGEMM_GPU included) from source (included in third_party folder of torchrec)

   python setup.py install develop
   

6. Run FEC

   Examples

   use FEC to train model WDL on 8 GPU workers with local batch size 8192

   cd example
   ./run.sh -s FEC -d criteo-tb -m WDL -b 8192 -w 8 -h -1
   

   Arguments

   • Systems
     • -s <Training Systems> DLRM training systems, one of {FEC, TorchRec, Parallax, FLECHE}
   • model
     • -s <Training Models> Training models, one of {WDL, DFM, DLRM, DCN}
   • batch size
     • -b <Training Batch Size> Training batch size, default setting: 8192
   • datasets
     • -b <Training datasets> Training datasets, one of {criteo, avazu, criteo-tb}
   • number of gpu workers
     • -w <number of gpu workers> the number of training gpu workers, default setting: 8
   • number of gpu workers
     • -w <number of gpu workers> the number of training gpu workers, default setting: 8
   • (Used in FEC only) number of hot entries
     • -h <number of hot entries> the number of hot entries use in FEC, -1 to use the predicted optimal number. default setting: -1