Low-Precision Training

Investigates methods for training large models in low-precision numerical formats, including 4-bit floating-point, to reduce memory and computation while preserving accuracy.

1 papers