SegFormer

SegFormer is a semantic-segmentation architecture introduced by researchers from NVIDIA and academic collaborators. Semantic segmentation assigns a class to each image pixel, unlike classification, which chooses one label for the whole image. SegFormer combines a hierarchical Transformer encoder that captures features at several scales with a lightweight multilayer-perceptron decoder; an MLP is a stack of simple fully connected layers. It also avoids fixed positional encodings, a memorable simplification. TT-Metal lists 512x512 inputs at batch 1 on n150 Wormhole and at batch 2 with DP=2 on n300 Wormhole. The duplicate hardware rows collapse into one model architecture and show single- versus two-device data-parallel serving.