`CUDA error: no kernel image is available for execution on the device` on a Pascal GPU (sm_61) in the v3 CUDA image

I'm running web-v3.0.0-beta-cuda-12.9.0 (version 3.0.0-beta.2) on a GTX 1080 Ti (Pascal, compute capability 6.1). The GPU is detected, but transcription fails with CUDA error: no kernel image is available for execution on the device and CUDA : ARCHS = 750,800,860,890,900,1000,1200. The same host runs latest-cuda-12.9.0 (2.4.13) against the same GPU with no problem at all.

Resolution: The whisper.cpp binary in the early v3 CUDA images was rebuilt without the Pascal architectures: the arch list in the error starts at sm_75 and no longer contains sm_60/sm_61, so no kernel matches a 1080 Ti. The v2 image still shipped a whisper.cpp built before that rebuild, which is why the identical GPU works there -- the two containers carry different binaries at the same path, ~/.local/share/ghost-story/whisper-cpp/<version>/linux-x64-cuda-12.9.0/bin/whisper-cli, with different sha256 sums even though the version directory name matches. Verify with docker exec <container> sh -lc 'sha256sum "$(find ~/.local/share/ghost-story/whisper-cpp -type f -name whisper-cli -print -quit)"' in both containers. Stopgap: copy whisper-cli from the working v2 CUDA container into the beta container at the same path -- that restores GPU transcription and alignment (confirmed, ~3 GB VRAM and ~62% GPU utilisation). The proper fix shipped in v3.0.0-beta.17, which restores Pascal support, and a related CPU-side issue was fixed in v3.0.0-beta.18; update to at least beta.18 rather than keeping the hand-copied binary.