SageAttention3: Microscaling FP4 Attention for Inference and An Exploration of 8-Bit Training | Read Paper on Bytez