Running Protein Folding Tools on EC2 Instance

Yes, the AWS Deep Learning PyTorch AMI should support running AlphaFold with GPU acceleration on your g6e instance, but there might be a few issues preventing CUDA from initializing properly.

The Deep Learning OSS AMI GPU PyTorch comes pre-configured with NVIDIA drivers, CUDA toolkit, and other necessary components for GPU computing. For g6e instances, the AMI includes NVIDIA Driver and CUDA 12.8 stack with installation directories at /usr/local/cuda-12.8/.

Here are some steps to troubleshoot your issue:

1. Verify your GPU is detected by running nvidia-smi in your terminal. This should display information about your GPU.

2. Check if your CUDA environment variables are properly set:

export PATH=/usr/local/cuda/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH

3. Ensure your Python environment has CUDA-compatible libraries. If you're using JAX with AlphaFold, you might need to install the CUDA-compatible version:

pip install -U "jax[cuda12]"

(Replace cuda12 with your specific CUDA version)

4. Verify that the CUDA versions in nvcc --version and nvidia-smi match.

For running protein folding specifically, you might want to consider the Quantori solution on AWS, which provides a user interface for running protein folding analyses with AlphaFold2. This solution allows you to select the ML algorithm, EC2 instance type, and database configuration through a simple interface.
Sources
Run protein folding on AWS with Quantori | AWS HPC Blog
No CUDA available on EC2 instance g5.xlarge | AWS re:Post
Resolve NVIDIA GPU and GPU driver issues in Amazon EC2 | AWS re:Post
AWS Deep Learning OSS AMI GPU PyTorch 2.7 (Amazon Linux 2023) - AWS Deep Learning AMIs