[Python] HOOMD-blue GPU error: "invalid device ordinal" in WSL2 despite CUDA and GPU detection

I am trying to run HOOMD-blue 5.2.0 with GPU support inside WSL2 (Windows Subsystem for Linux), but I keep getting the following error:

RuntimeError: CUDA Error: invalid device ordinal before /hoomd/Autotuner.h:497

My setup:

• 
  WSL2 with Ubuntu 22.04
  
  
• 
  NVIDIA GPU: RTX 4060 with latest drivers
  
  
• 
  nvidia-smi works inside WSL and detects the GPU
  
  
• 
  HOOMD installed using:
  
  micromamba install hoomd=5.2.0
• 
  CUDA Toolkit version: 12.2 (also tried 11.8)
  

Test script:


import hoomd
device = hoomd.device.GPU() # Also tried device_id=0 explicitly


Calling GPU.get_available_devices() returns:

['[0]NVIDIA GeForce RTX 2060 30 SM_7.5 @ 1.68 GHz, 6143 MiB DRAM']


So the GPU is detected. But as soon as I initialize the device, I get the "invalid device ordinal" error.

Things I’ve tried:

• 
  Setting CUDA_VISIBLE_DEVICES=0
  
  
• 
  Explicitly passing device_id=0
  
  
• 
  Downgrading/upgrading HOOMD and CUDA versions
  
  
• 
  Installing with both conda and micromamba
  
  
• 
  Running minimal examples with just hoomd.device.GPU()
  

None of these worked. The only configuration that runs is when I switch to hoomd.device.CPU(), but I need GPU acceleration.

Questions:

• 
  Is HOOMD-blue fully compatible with GPU under WSL2?
  
  
• 
  If so, how can I fix the invalid device ordinal error (which seems to appear when neighbor lists are used)?
  
  
• 
  If not, is there a known workaround (besides switching to CPU or using native Windows/Linux)?
  

Any insights from users who have successfully run HOOMD with GPU inside WSL2 would be much appreciated!

Continue reading...