NVIDIA/TransformerEngine

A library for accelerating Transformer models on NVIDIA GPUs, including using 8-bit and 4-bit floating point (FP8 and FP4) precision on Hopper, Ada and Blackwell GPUs, to provide better performance with lower memory utilization in both training and inference.

语言：Python · GitHub 星标：3,600 · Fork：888

话题：cuda · deep-learning · fp4 · fp8 · gpu · jax · machine-learning · python

在 GitHub 上查看原仓库