MCTensor: A High-Precision Deep Learning Library with Multi-Component Floating-Point
This provides a solution for deep learning practitioners needing high-precision arithmetic with optimized performance, though it is incremental as it builds on existing PyTorch frameworks.
The authors tackled the problem of achieving high-precision deep learning training by introducing MCTensor, a library that matches or outperforms PyTorch's float32 or float64 precision using float16, as demonstrated in evaluations across various tasks.
In this paper, we introduce MCTensor, a library based on PyTorch for providing general-purpose and high-precision arithmetic for DL training. MCTensor is used in the same way as PyTorch Tensor: we implement multiple basic, matrix-level computation operators and NN modules for MCTensor with identical PyTorch interface. Our algorithms achieve high precision computation and also benefits from heavily-optimized PyTorch floating-point arithmetic. We evaluate MCTensor arithmetic against PyTorch native arithmetic for a series of tasks, where models using MCTensor in float16 would match or outperform the PyTorch model with float32 or float64 precision.