bf16
Neural Weight Compression: lossless LLM weight compression decoded inside the CUDA matvec. BF16 31 % smaller and faster than cuBLAS; fp8 13 % smaller at native fp8 speed. Bit-identical weights.
Lossless AI model compression - ~34% smaller with bit-identical weights; the autopilot profiles your machine, picks the highest fidelity that runs, and streams models bigger than your RAM.
A Python binding of Berkeley SoftFloat
Fast BF16 PyTorch training operators for standard and low-rank Attention Residuals
