Open nathanweir opened 10 months ago
Ah this would be awesome, thanks for this PR!
@thecooltechguy This looks good to go on my end, but fair warning I know nothing about the API being called here - this just mimics what that generative-models
repo does and in my experience functions OK. I'm happy to tweak the console output or anything else if you'd like!
Hello!
I've had great success at setting
lowvram_mode = True
in the Generative Models app from Stability AI when first trying out SVD: https://github.com/Stability-AI/generative-models/blob/main/scripts/demo/streamlit_helpers.py#L61-L75. In my experience I can run SVDxt at less than 24GB vram use on my 4090 with this enabled, whereas without I go just over 24GB, triggering shared GPU ram use and greatly reducing performance.This PR introduces a simple toggle to trigger half precision when used.