Recent searches
Search options
@danieldk@sigmoid.socialWe recently added support for running spaCy transformer models on Apple Silicon Mac GPUs.
Madeesh Kannan and I have written a blog post with some background and a lot of benchmarks:
https://explosion.ai/blog/metal-performance-shaders
Teasers: transformer pipelines run 4.7 times faster on the M1 Max GPU cores than on its AMX units and the MacBook Air M2 uses less power for GPU inference than CPU inference.
ExplosionFast transformer inference with Metal Performance Shaders · ExplosionWe are happy to introduce support for Metal Performance Shaders in Thinc PyTorch layers. This makes it possible to run spaCy transformer-based pipelines on GPU on Apple Silicon Macs and improves inference speed up to 4.7 times.