EXLA crashed with CUDA 12.4 and cuDNN 9.11

[msluszniak]

When I set XLA_TARGET=cuda12 and run the code of my project I get the following logs and then segfault:

WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1767784508.583888   12926 se_gpu_pjrt_client.cc:1101] Using BFC allocator.
I0000 00:00:1767784508.583950   12926 gpu_helpers.cc:136] XLA backend allocating 4714043801 bytes on device 0 for BFCAllocator.
I0000 00:00:1767784508.583991   12926 gpu_helpers.cc:177] XLA backend will use up to 18856175206 bytes on device 0 for CollectiveBFCAllocator.
I0000 00:00:1767784508.590265   12926 cuda_dnn.cc:529] Loaded cuDNN version 91100
Segmentation fault

I use Nx on version 0.10 and EXLA on 0.10. CUDA and cuDNN as in the title.

[polvalente]

This CUDA version is a bit old, I suggest trying 12.8.
Otherwise, you should get more information on the Segfault with lldb: https://dockyard.com/blog/2024/05/22/debugging-elixir-nifs-with-lldb