Hi @bigjoedata, glad you’re finding Streamlit sharing to be a useful service!
I think your instincts are correct about pytorch being the culprit. I don’t have any firm evidence, but I feel like most of the resource issues that are being reported seem to have this use case in mind. Is it possible to do a CPU-only version of Torch? I know you are saying that the app does fallback to CPU, but I wonder if there is a CPU-only version, it could be lighter on RAM usage without loading all the GPU code, exception handling, etc.
Internally, we are deeply analyzing the performance characteristics of the currently deployed apps, so that we can figure out how to provide more resources to the community without dramatically increasing our costs. Part of that will be creating paid tiers, but hopefully we can also find a way to set users resources in such a way that apps that obviously don’t need many resources can be scaled down, apps like yours can be scaled up, and yet maintain the same relative overall costs to us.
Best,
Randy