Miguel Palhas
a7c74d5a9f
The previous commit wired CUDA 13 against the bundled onnxruntime provider, which only accelerates diarization/embeddings — transcription still fell back to CPU with "GPU could not be activated". The GPU whisper-server is fetched at runtime and links CUDA 12 SONAMEs, so it needs cudaPackages_12 on LD_LIBRARY_PATH (plus libstdc++/libgomp, which it also could not resolve). Verified: whisper-server now reports "found 1 CUDA devices ... RTX 4060" and loads the model on the CUDA0 backend. Drops the closure 4.6 -> 3.7 GiB.