1547727 : napari with nnInteracktive on maxwell

Created: 2025-10-28T06:53:39Z - current status: new

Anonymized Summary: A user is experiencing issues while trying to run napari-nninteractive for interactive ML-based segmentation on the Maxwell cluster. The problems vary depending on the node type:

  • Display Nodes: Errors related to loading cudnn*.so and segmentation failures.
  • Compute Nodes (e.g., max-wng064): Issues with 3D rendering.

The user is using the same Python environment (/data/dust/user/[USER]/mamba_envs/nnInteractive/) on both node types and launches the program with napari -w napari-nninteractive.

Possible Solution: 1. Display Nodes: Since the issue involves GPU-related errors (e.g., cudnn*.so), ensure that the environment is correctly configured for GPU acceleration. The user might need to load the appropriate CUDA modules or verify that the environment variables (e.g., LD_LIBRARY_PATH) are set correctly. Additionally, the user could try running the application inside a FastX session, as some graphical applications require hardware-accelerated GL (similar to Avizo/Amira).

  1. Compute Nodes: For 3D rendering issues, the user might need to set the environment variable LIBGL_ALWAYS_INDIRECT=1 to force indirect rendering, as some applications (like hexrd) require this to work properly over SSH.

  2. Environment Check: Verify that the Python environment (/data/dust/user/[USER]/mamba_envs/nnInteractive/) has all the required dependencies installed, especially those related to GPU acceleration (e.g., cudnn, cudatoolkit). The user could also try reinstalling the environment or checking for conflicts with other loaded modules.

  3. Node-Specific Issues: If the issues persist, the user could try running the application on a different display node or compute node to isolate whether the problem is node-specific.

References: - Using Amira or Avizo on Maxwell - Troubleshooting - Maxwell Display Nodes