Qihoo360/tensornet

TensorNet was a vital piece of infrastructure for its era (circa 2018-2020), addressing the specific bottleneck of training deep learning models with billions of sparse features (like ad-click prediction). Its moat is built on its ability to handle embedding tables larger than memory across distributed nodes, a feat standard TensorFlow often struggled with. However, the project shows zero recent velocity (0.0/hr) and has essentially been superseded by more modern frameworks like Meta's TorchRec, ByteDance's Monolith, and Alibaba's DeepRec. While it remains a production-grade tool, the industry shift from TensorFlow to PyTorch for high-dimensional sparse training significantly weakens its long-term defensibility. It lacks the massive community lock-in of Horovod or the cutting-edge feature set of DeepSpeed. A score of 4 reflects its status as a functional but stagnant specialized tool that is likely being phased out in favor of newer, more active ecosystem standards.