`CUDA error: no kernel image is available` on GTX 10-series (sm_61) after updating to the cuda-12.9.0 image

I updated to the latest image and transcription stopped working - the task fails with a CUDA error about no kernel image being available for execution. My card is a GTX 10-series (compute capability 6.1). The docs tell you to use the cuda-12.9.0 image, but switching to the 11.8.0 image made everything work again.

Resolution: The bundled whisper.cpp build had stopped emitting kernels for sm_61, so 10-series cards loaded a binary with no matching kernel image even though CUDA 12.9 itself still supports that architecture - it was a build-flag problem, not an NVIDIA support drop. The immediate workaround is to run the cuda-11.8.0 image tag instead of cuda-12.9.0; performance and functionality are otherwise unchanged for these cards. The real fix shipped in 2.14.19 and 3.0.0-beta.17, so pull one of those (or newer) and you can go back to the CUDA 12.9 image. If you're on beta.17 and now hit a SIGILL while loading the model, that's a separate AVX-512 build issue fixed in 2.14.20 / 3.0.0-beta.18.