FP8 Quantization

Technique converting model weights or activations to 8-bit floating-point format to reduce memory usage and accelerate inference