Use pytorch2+cu118 with ADA hardware for 50%+ speedup

StableDiffusionXL can now be made 50% faster on the RTX 4090 with the help of AI technology. The problem arises when using Pytorch2, as the upstream version pulls pytorch2.0.1 with cu117 which does not properly support newer GPUs like the RTX4090 or H100. This affects the performance of popular cloud GPU providers used for deep learning. To fix this, we should target cu118 until it is included in upstream pip packages. Upgrading to a CUDA version of at least 11.8 can be done through a docker base image or by upgrading the NVIDIA driver and CUDA version. By pulling pytorch with cu118, users can enjoy a 50% speedup on all ADA LOVELACE workloads. Contact us if you’re interested in building an AI app.

To top