TorchSeg provides a fast, modular framework for implementing and training semantic segmentation algorithms in PyTorch. It includes pre-trained models (ResNet18/50/101) and supports popular segmentation architectures such as FCN, DFN, BiSeNet, PSPNet, and PSANet. The project offers distributed training capabilities for multi-GPU setups and benchmarks on datasets including PASCAL VOC 2012, Cityscapes, ADE20K, and CamVid.