Closed riccardopinosio closed 6 months ago
is there a plan to support GPU?
@sedletsky-f5 hi, there sure is, as onnxruntime should make that relatively painless. I haven't gotten around to it yet, wanted to add support for a couple more pipelines first, but could look at it earlier if you think it's important for you application
Today i started a new branch (gpu-support) which exposes all the options for the available providers. If you want to try a pre alpha version i think the remaining work is to select an appropriate gpu-enabled onnxruntime library from https://github.com/microsoft/onnxruntime/releases , then test this in an environment with the right OS dependencies/libraries. Let us know if you try this early.
Cuda is now fully working for me on the GPU branch, so now would be a good time to try it if you like. We'll release this imminently.
sorry, missed previous messages - yes, it is important for my project and I will test gpu-support
branch shortly. thanks a lot!
Hi @sedletsky-f5, this is now implemented in main (new release 0.1.0).
great news, @riccardopinosio ! I've exactly tested it on my env and GPU version works properly and performance is also much better than CPU version (about 5 times faster for running feature extraction pipeline).
now, I have some improvement suggestion - today users should decide what version to use and even specify the selected library with WithOnnxLibraryPath()
, but it can be done automatically using function gotch.CudaIfAvailable()
available in https://github.com/sugarme/gotch/
Since we rely on onnxruntime it should be fairly straightforward to support GPU for the pipelines.