Model Parallelism

Offloading

Memory management strategy where data (weights, gradients, activations) are dynamically moved between fast GPU memory and slower but more extensive CPU memory.

← Terug