Open RutraNickers opened 1 month ago
This problem is a base problem how PyTorch works. When running in GPU-mode, the whole segments (windows) that have to be processed HAVE TO fit into the VRAM. You can try to set smaller segment-sizes or experiment with other settings the model offers to lower the maximum memory consumption. If that does not work, you need to run CPU mode or get a card with more RAM.
I'm using the openCL version with a 1GB VRAM Radeon RX580, so definitely on the lowers of the low VRAM range. I was trying to use the Ensemble Mode, but it didn't run. I have 16gb of RAM though, so it would be lovely if it could use my RAM instead, but that's a PyTorch problem, isn't it? Could you guys create a low_vram mode akin to what ComfyUI did? Some nodes there also have this thing called tiling, where they process small chunks, one at a time, to limit how much VRAM is being used at each moment. While I can run in CPU Only mode, it is way too slow compared to what I can do in ComfyUI using DirectML and their -low_vram argument.
The error thrown is pasted bellow: