TensorTonic
Problems

Implement Block-Wise FP8 Quantization

Compute one scale per block from that block's own absolute maximum, and scale each block into the [-fp8 max, fp8 max] range with saturation.