Re: whisper-cpp with AMD GPUs (Was: llama-cpp with AMD GPUs)
[Cordell Bloor]
> It seems to work fairly well on my Radeon VII.
It worked on one of my test machines doo, using GeForce GT 755M and the
OpenCL backend. I built using
cmake -S. -Bbuild cmake . -DWHISPER_CLBLAST=ON
-DCMAKE_BUILD_TYPE=Release
I could not get the CUDA stuff working, not sure why.
I am also not sure how much the GPU is used, but it do print out these
when running, at least:
ggml_opencl: selecting platform: 'NVIDIA CUDA'
ggml_opencl: selecting device: 'NVIDIA GeForce GT 755M'
The clocktime spent transcribing the jfk.wav sample is 34.8s with OpenCL
support compiled in, and 41.2s using the CPU, so I guess it has some
effect (15.5% less time spent).
Perhaps someone should set up a project to transcribe all Debian videos
using Whisper, to provide searchable text for each Debconf presentation
and other talks.
Perhaps <URL: https://bugs.debian.org/1034091 > is better solved using
whisper.cpp?
--
Happy hacking
Petter Reinholdtsen
Reply to: