fp8_quant#
Composable Triton JIT functions for FP8 (E4M3) fake quantization.
Counterpart of nvfp4_quant.py for per-tensor FP8. Used by the unified
flash-attention kernel’s softmax-P qdq and linear-attention state/factor qdq.
Composable Triton JIT functions for FP8 (E4M3) fake quantization.
Counterpart of nvfp4_quant.py for per-tensor FP8. Used by the unified
flash-attention kernel’s softmax-P qdq and linear-attention state/factor qdq.