Mixed Precision Computing

Gradient Accumulation in Mixed Precision

Technique where gradients calculated in FP16 are accumulated in an FP32 buffer before weight update, preventing precision loss during aggregation over multiple mini-batches.

← Wstecz