Good news everyone!
Btw, would whisper-cpp be a better match for Debian than the openai-whisper implementation?I know very little about AI applications, but I see that whisper-cpp has a hipblas backend that can be enabled with `-DWHISPER_HIPBLAS=ON`. A quick review of the codebase suggests to me all dependencies required to package whisper-cpp with GPU acceleration are probably already in Debian.
It seems to work fairly well on my Radeon VII.
apt -y update apt -y upgrade apt -y install git wget ffmpeg hipcc libhipblas-dev librocblas-dev cmake build-essential wget https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3.bin?download=true -O ggml-large-v3.bin git clone https://github.com/ggerganov/whisper.cpp.git cd whisper.cpp git checkout 02b4c52c1289e05c8c04ff8370a4835b8ee99c86 CC=clang-15 CXX=clang++-15 cmake -S. -Bbuild -DWHISPER_HIPBLAS=ON -DAMDGPU_TARGETS="gfx803;gfx900;gfx906;gfx908;gfx90a;gfx1010;gfx1030" -DCMAKE_BUILD_TYPE=Release make -j16 -C build make samples ./build/bin/main -m ../ggml-large-v3.bin -f samples/jfk.wav --print-colors
Although, there seems to be a bug in whisper.cpp at that
particular commit. I'm seeing occasional errors in the
transcriptions output both when using hipblas and when using the
purely CPU implementation. When I try running on samples/gb1.wav,
the transcription seems to work properly until it reaches
00:02:02.200 at which point it just keeps repeating, "We will be
led into the darkness."
Sincerely,
Cory Bloor