flashinfer.fused_moe.CuTileMxfp8Config

class flashinfer.fused_moe.CuTileMxfp8Config

cuTile MXFP8 weights and activations, with E8M0 scales per 32 values.

Inputs are BF16; both GEMM inputs are dynamically quantized to MXFP8. Expert parallelism and fused shared experts are not supported.

__init__() → None

Methods

__init__()
prepare_weights(w1_fp8, w1_scale, w2_fp8, ...)	Prepare E4M3 weights with logical [E, N, K/32] E8M0 scales.
supported(arch)