Guidance on using lower precision ("quantization") for deep learning

[qualiaMachine]

Deep learning is very resource intensive. A common trick to reduce compute needs is to lower the precision of your data (e.g., from float32 to float16). Alternatively, there are options to use automatic mixed precision via torch.amp. We should provide some guidelines on the impact of lowering precision (e.g., during training vs. during inference). Provide references/evidence to support these guidelines.