Documentation: Installing DGL with ROCm 6.4 Using Docker

1. Objective

To set up a GPU-accelerated deep learning environment using:

• ROCm 6.4
• PyTorch (ROCm build)
• DGL (Deep Graph Library)
• Docker containerization

This approach avoids dependency conflicts and ensures compatibility with AMD GPUs.

---

2. Environment Overview

Host System

• Linux distribution: Ubuntu 22.04 / 24.04 (adjust as appropriate)
• ROCm installed: 6.4.x
• GPU: AMD GPU compatible with ROCm
• Docker: Installed and verified

Container Base Image

• Repository: rocm/dgl
• Example tag used: dgl-2.4_rocm6.4_ubuntu22.04_py3.10_pytorch_release_2.4.1

---

3. Docker Installation Verification

After installing Docker, verification was performed using:

sudo docker run hello-world

Successful execution confirmed:

• Docker daemon is running
• User has permission to execute Docker commands

---

4. Pulling the ROCm DGL Image

The official AMD DGL image was pulled from Docker Hub:

sudo docker pull rocm/dgl:dgl-2.4_rocm6.4_ubuntu22.04_py3.10_pytorch_release_2.4.1

This image includes:

• ROCm 6.4 runtime
• PyTorch 2.4.1 (ROCm build)
• DGL preinstalled
• Python 3.10
• Ubuntu 22.04 base

---

5. Running the Container with GPU Access

To enable ROCm GPU access inside Docker, the container was launched with the following command:

sudo docker run -it \
  --cap-add=SYS_PTRACE \
  --security-opt seccomp=unconfined \
  --device=/dev/kfd \
  --device=/dev/dri \
  --group-add video \
  --ipc=host \
  --shm-size 8G \
  rocm/dgl:dgl-2.4_rocm6.4_ubuntu22.04_py3.10_pytorch_release_2.4.1

Explanation of Key Flags

Flag	Purpose
--device=/dev/kfd	Exposes AMD compute device
--device=/dev/dri	Exposes GPU rendering interface
--group-add video	Grants GPU access permissions
--ipc=host	Improves shared memory handling
--shm-size 8G	Prevents DataLoader memory issues
--cap-add=SYS_PTRACE	Required for debugging

---

6. Verification Inside Container

6.1 Verify ROCm Detection

rocminfo | head

Confirms GPU is visible.

---

6.2 Verify PyTorch ROCm Backend

python3 -c "import torch; print(torch.__version__, torch.version.hip)"

Expected:

• Correct PyTorch version
• HIP version not None

Check GPU availability:

python3 -c "import torch; print(torch.cuda.is_available())"

Expected output:

True

(Note: ROCm uses torch.cuda namespace.)

---

6.3 Verify DGL Installation

export DGLBACKEND=pytorch

python3 - << 'EOF'
import dgl
print("DGL version:", dgl.__version__)
EOF

Successful import confirms:

• DGL properly installed
• Linked correctly with PyTorch backend

---

7. Rationale for Using Docker

Using Docker provides:

1. Version isolation (no system contamination)
2. Guaranteed compatibility (prevalidated by AMD)
3. Reproducibility across machines
4. Simplified dependency management

This avoids:

• Pip version conflicts
• ROCm wheel mismatch errors
• Manual compilation complexity

---

8. Optional: Running with Project Mount

To use local project files:

sudo docker run -it \
  --device=/dev/kfd \
  --device=/dev/dri \
  --group-add video \
  --ipc=host \
  --shm-size 8G \
  -v $(pwd):/workspace \
  rocm/dgl:dgl-2.4_rocm6.4_ubuntu22.04_py3.10_pytorch_release_2.4.1

This mounts the current directory into /workspace inside the container.

---

9. Conclusion

The Docker-based ROCm installation successfully provides:

• Stable PyTorch + ROCm 6.4 integration
• Working DGL framework
• GPU acceleration inside container
• Controlled and reproducible environment

This setup is recommended for research workflows requiring ROCm compatibility and graph-based deep learning.

---

Q2: Will you run experiments on multiple machines (requiring reproducibility documentation)? Q3: Do you plan to benchmark GPU utilization to validate performance inside Docker?