fkl_python

flex: import numpy lazily — keep 'import fkl' zero-dep (fixes CI)

flex attention + block-sparse attention in fkl-python

fp8 e4m3 KV cache: compress_kv(fmt='fp8') + kv_layout='fp8'

as_device_view: torch fast path — call overhead 28us -> hidden under GPU time

flash_attention: prologue fusion (Q/K/V as Read IOps) + mma tensor-core path