Open WSINTRA opened 1 year ago
It works on my 2020 iMac (16 GB AMD Radeon Pro 5700 XT), using the ORIGINAL
attention implementation instead of SPLIT_EINSUM
.
@pcuenca would be nice if performance data of some Intel Macs (using AMD GPUs like Vega 5700xt and 6xxx series) can be added to readme.. would be interesting if Apple ends supporting new RDNA3 like 7900xt also much like Rdna2..
@oscarbg I believe the main focus of this repo is about Apple Silicon on both macOS and iOS/iPadOS. Happy to share a data point taken on my iMac:
iMac Retina 5K, 2020 Processor: 3.6 GHz 10-Core Intel Core i9 GPU: AMD Radeon Pro 5700 XT 16 GB
Takes 14.9s to run inference using ORIGINAL
attention with compute units CPU AND GPU
. Tested on Stable Diffusion 2 Base with 25 inference steps of the DPM-Solver++ scheduler.
Feel free to share more data in our Swift Core ML Diffusers repo :)
@pcuenca yes seems is about Apple Silicon, but wanted to have more data for comparison as it’s using CoreML, and that’s supported on Mac Intel ecosystem too.. Thanks for sharing!! For me issue can be closed..
It works on my 2020 iMac (16 GB AMD Radeon Pro 5700 XT), using the ORIGINAL attention implementation instead of SPLIT_EINSUM.
@pcuenca How did you do that? Would you please provide some instructions ? Thanks.
Will this also work on a Mac that has an intel CPU with an AMD GPU using metal ?