Poster
in
Workshop: Hardware-aware efficient training (HAET)

MCTensor: A High-Precision Deep Learning Library with Multi-Component Floating-Point

Tao Yu · Wentao Guo · Canal Li · Tiancheng Yuan · Christopher De Sa


Abstract:

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.

Chat is not available.