Nunchux AI's training-free VC-Attention fixes value outliers and FP32 softmax bottleneck, speeding video DiT attention up to 3.58×.