A field report on validating the Obscura NLP library on an NVIDIA Tesla T4 GPU using Elixir, Nx, and EXLA on Lightning AI Studio. The author walks through a layered verification process: starting on CPU, switching to a T4, resolving an NVSHMEM ABI mismatch between the precompiled XLA archive and the Studio image, and confirming that compiled Nx operations actually execute on the GPU. Key steps include setting XLA_TARGET=cuda13, extracting a compatible NVSHMEM 3.3.24 package to fix the shared library failure, and verifying GPU execution via nvidia-smi and EXLA.Backend tensor results. The balanced and accurate Obscura profiles were both exercised, with warm inference times around 900ms and 1800ms respectively. The post is explicit about what the test proves and what it does not — it is a compatibility report, not a production benchmark.
