awesome-rocm-autodrive

awesome-rocm-autodrive-training is a collection of autonomous driving model training examples adapted and optimized for AMD GPUs using the ROCm platform.
This project provides an out-of-the-box (OOTB) training experience through ROCm-compatible MMCV, prebuilt Docker environments, and a wide range of real-world autonomous driving model examples.

---

🚀 Key Highlights

• ✅ Out-of-the-box training experience on AMD ROCm GPUs
• 🛠️ Patched MMCV with full ROCm compatibility and performance fixes
• 🐳 Ready-to-use Docker environment and image
• 🧩 Diverse training examples across major autonomous driving tasks
• ⚡ ROCm-specific performance optimizations in selected models

---

📁 Repository Structure

awesome-rocm-autodrive-training/
├── docker/          # Dockerfile and prebuilt ROCm training environment
├── examples/        # Training examples for various AD tasks
│   ├── backbone/    # Backbone networks (e.g., ResNet50, EfficientNet)
│   ├── 3d_detection/   # 2D/3D object detection (e.g., PointPillars)
│   ├── prediction/       # Prediction models (e.g., QCNet)
│   ├── bev/         # BEV perception models (e.g., BEVFormer)
│   ├── mapping/         # HD map construction models (e.g., MapTR)
│   ├── occupancy/   # Occupancy prediction (e.g., FlashOcc, SurroundOcc)
│   ├── end2end/         # End-to-end driving pipelines (e.g., UniAD)
├── mmcv/            # ROCm-adapted mmcv source code
├── tools/           # Utility scripts for benchmarking/tuning
└── README.md

---

🔧 Supported Models (Initial Release)

Model	Type	Repo Link	README for ROCm	Notes
ResNet-50	Backbone	https://github.com/amd-fuweiy/vision	example/backbone
EfficientNet-B7	Backbone	https://github.com/amd-fuweiy/vision		Currently have performance issue with DWConv
PointPillars	Point Cloud	https://github.com/Treemann/mmdetection3d	examples/detection/pointpillars	Need ROCm mmcv to get better performance
MapTR	Vector Prediction	https://github.com/aaab8b/MapTR	See modified Readme in git
FlashOCC	Scene Occupancy	https://github.com/mingjielu/FlashOCC	examples/occupancy/FlashOCC
Sparse4D	Sparse Detector	https://github.com/binding7012/Sparse4D	examples/detection/sparse4d
BEVFormer	Multi-view	https://github.com/jun-amd/BEVFormer	examples/detection/BEVFormer
PETR	3D Detection	https://github.com/aaab8b/PETR	See modified Readme in git
QCNet	Trajectory Prediction	https://github.com/aaab8b/QCNet	See modified Readme in git
SurroundOcc	Scene Occupancy	https://github.com/mingjielu/SurroundOcc	examples/occupancy/FlashOCC
UniAD	End to End	https://github.com/mingjielu/UniAD	examples/end2end/UniAD

---

🚀 Quick Start

1. Clone the repo:

git clone https://github.com/AMD-AGI/awesome-rocm-autodrive.git
cd awesome-rocm-autodrive

2. Build Docker Image

cd docker
docker build -t rocm-autodrive .

or directly pull the docker image provided by AMD:

docker pull amdagi/autodrive_training_rocm6.4:v1 

build mmcv (optional)

If you want to build mmcv separately, just

MMCV_WITH_OPS=1 pip install .

3. Launch Docker Container

docker run --rm -it --ipc=host --network=host \
  --device=/dev/kfd --device=/dev/dri \
  --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined \
  -v $PWD:/workspace \
  rocm-autodrive

4. Run an Example

git clone the repo link provided in README, repare dataset, and run as corresponding readme.

🔮 What's Next

We are actively expanding the coverage of this project. Upcoming efforts include:

• ✅ More Models: Support for popular/SOTA models such as DeepAccident, and end-to-end transformer-based planners.
• 🔍 Performance Benchmarking: Add ROCm vs CUDA training benchmark results across all models.
• 🧠 FP8 & Mixed Precision Training: Explore FP8/BF16 optimization on MI300X GPUs.
• 🤝 Upstream Collaboration: Work with upstream repos to upstream ROCm compatibility patches and improvements.

Have ideas or requests? Open an issue or start a discussion!

---

🤝 Contributing

We welcome contributions from the community!

To contribute:

1. Fork this repository.
2. Clone your fork and create a feature branch.
3. Make your changes and test them.
4. Submit a pull request with a clear description.

When adding a new model:

• ✅ Include training scripts and a minimal README.
• ⚙️ Note any ROCm-specific patches or performance tips.
• 🧪 Validate functionality on at least one ROCm GPU (MI300X, MI325X, or MI355X).
• 📢 Consider submitting upstream patches for long-term compatibility.

---

📝 License

This project is licensed under the MIT license. while some other examples in this library are with other licenses. See the full license text here: LICENSE