bfloat16 (bf16)

Also written: torch.bfloat16, bf16

A 16-bit number format for storing model weights, which halves memory compared with 32-bit floats.