DLRMBenchmark

DLRM v2 benchmark for single-node AMD MI3xx GPUs.

This repository showcases DLRM v2 training on a single AMD node (8x MI3xx). The model is derived from the MLPerf DLRM v2 repo and targets MI300X, MI325X, and MI350X accelerators. We use the ROCm PyTorch Training Docker, which comes pre-installed with FBGEMM and TorchRec. This codebase is not meant to highlight peak achievable performance.

Installation

Requirements

• Single-node w/ 8x MI300X/325X/350X
• Docker / Torch wheels for ROCm

Setup

1. Clone repository:

   git clone https://github.com/AMD-AGI/DLRMBenchmark.git

2. Update permissions to 777 for all shell scripts
3. Container with ROCm, PyTorch, FBGEMM and torchrec installed are available at https://hub.docker.com/r/rocm/primus/. Pull the container:

   docker pull rocm/primus:v26.1

4. Launch container. Ensure all required paths including codebase are mounted (similar to /home_dir/).

   docker run -d \
     --ipc=host \
     -v /dev/shm:/dev/shm \
     -v /home_dir/:/home_dir/ \
     -e USER=$user -e UID=$uid -e GID=$gid \
     --device=/dev/kfd \
     --device=/dev/dri \
     --device=/dev/infiniband \
     --ulimit memlock=-1:-1 \
     --shm-size 32G \
     --cap-add=SYS_PTRACE \
     --security-opt seccomp=unconfined \
     --group-add video \
     --network=host \
     --name dlrm_demo \
     -it rocm/primus:v26.1 \
     tail -f /dev/null

Model Training

1. Start interactive shell session within container:

   docker exec -it dlrm_demo bash

2. Modify training configuration (if required) in training_config.sh
3. Launch training using ./launch_training_single_node.sh. Check for training progress in the ./training_logs folder. Upon completion, the final row in the results.csv file shows the mean rec/s.

Contributing

We welcome contributions! Please read CONTRIBUTING.md for development setup, branch strategy, coding standards, and the pull request process.

For bugs and feature requests, open a GitHub Issue.

---

Security

To report a security vulnerability, do not open a public GitHub issue. See SECURITY.md for our responsible disclosure policy.

---

Contact

For questions, issues, or contributions, please reach out to the maintainers:

• Tharun Adithya Srikrishnan — @tsrikris

See CODEOWNERS for the full ownership list.

---

License

This project is licensed under the MIT License.