TorchSparse++ is a high-performance computing library for efficient 3D sparse convolution. It offers significant performance improvement over TorchSparse++ by overlapping computation with memory access. It also searches for the best execution strategy for sparse workloads within a large design space through auto-tuning.
This project introduce PockEngine: a tiny, sparse and efficient engine to enable fine-tuning on various edge devices. PockEngine supports sparse backpropagation: it prunes the backward graph and sparsely updates the model with measured memory saving and latency reduction while maintaining the model quality.
We enable LLMs to work on infinite-length texts without compromising efficiency and performance.
LongLoRA takes advantage of shifted sparse attention to greatly reduce the finetuning cost of long context LLMs.
We present FlatFormer, an efficient ViT architecture for large-scale point cloud analysis.
Vision transformer on high-resolution images can learn richer visual representation. However, the improved performance comes at the cost of huge computation complexity. Thus, we present SparseViT, which accelerates high-resolution visual processing by skipping less important regions during computation.
In MCUNetV3, we enable on-device training under 256KB SRAM and 1MB Flash, using less than 1/1000 memory of PyTorch while matching the accuracy on the visual wake words application. It enables the model to adapt to newly collected sensor data and users can enjoy customized services without uploading the data to the cloud thus protecting privacy.