Loading EXLA is not proof that an Elixir workload used an NVIDIA GPU.

I documented how I validated Obscura, Nx, and EXLA on a Linux Tesla T4 through Lightning AI, including an NVSHMEM ABI mismatch, CUDA backend proof, warm inference latency, and the BEAM process’s GPU memory allocation.

https://hfiguera.github.io/obscura/blog/running-obscura-on-nvidia-exla/