ViT-Base

ViT-Base is the base-size Vision Transformer introduced by Google Research for image classification. Instead of scanning an image only with convolution filters, ViT divides it into fixed-size patches, turns those patches into a sequence of tokens, and processes them with a Transformer much like tokens in language. It is still an image classifier; ‘Transformer’ does not make it a text generator. TT-Metal’s matrix uses 224x224 images and lists n150 and n300 Wormhole systems, p150 Blackhole, and Wormhole QuietBox at DP=8, with batches from 8 to 64. These are alternate scales for the same ViT-Base architecture, ranging from one accelerator to replicated multi-device processing.