Steps to Reproduce
First of all, thank you very much to ovo and all contributors for developing and maintaining this excellent software .
While deploying and running the project using conda + WSL, I encountered several issues related to environment configuration and GPU usage. I am reporting them here for reference and discussion.
Environment
OS: Windows + WSL
Environment manager: Conda
CUDA: 12.8
GPU: NVIDIA 3060ti (properly detected in WSL)
Issues
- rfdiffusion.yml does not explicitly specify GPU versions of torch and dgl
In rfdiffusion.yml, GPU-enabled versions of torch and dgl are not explicitly specified.
In practice:
dgl 1.1.2 requires a CUDA 11.8 build
If the CPU version of dgl/torch is installed by default, the GPU may not be used correctly even when CUDA is available
It may be helpful to explicitly specify GPU versions in the environment file, for example:
torch==2.2.2+cu118
- proteinmpnn-fastrelax.yml should explicitly specify Python 3.12
- GPU version of rfdiffusion mainly uses CPU at runtime
After confirming that GPU-enabled dependencies are installed and running the GPU version of rfdiffusion, I observed that:
GPU utilization is relatively low
CPU utilization is high (observed via Windows Task Manager)
The root cause is currently unclear.
Thanks again to ovo and the maintainers for their great work!
Expected Behavior
No response
Actual Behavior
No response
Logs / Screenshots
Steps to Reproduce
First of all, thank you very much to ovo and all contributors for developing and maintaining this excellent software .
While deploying and running the project using conda + WSL, I encountered several issues related to environment configuration and GPU usage. I am reporting them here for reference and discussion.
Environment
OS: Windows + WSL
Environment manager: Conda
CUDA: 12.8
GPU: NVIDIA 3060ti (properly detected in WSL)
Issues
In rfdiffusion.yml, GPU-enabled versions of torch and dgl are not explicitly specified.
In practice:
dgl 1.1.2 requires a CUDA 11.8 build
If the CPU version of dgl/torch is installed by default, the GPU may not be used correctly even when CUDA is available
It may be helpful to explicitly specify GPU versions in the environment file, for example:
torch==2.2.2+cu118
After confirming that GPU-enabled dependencies are installed and running the GPU version of rfdiffusion, I observed that:
GPU utilization is relatively low
CPU utilization is high (observed via Windows Task Manager)
The root cause is currently unclear.
Thanks again to ovo and the maintainers for their great work!
Expected Behavior
No response
Actual Behavior
No response
Logs / Screenshots