flashinfer.fused_moe.CudnnFrostMxfp8Mxfp4Config

class flashinfer.fused_moe.CudnnFrostMxfp8Mxfp4Config

Frost MXFP8-activation / MXFP4-weight MoE on SM107.

Requires packed precomputed routing and H/I divisible by 128. Register prepared weights under cudnn_frost_mxfp8_mxfp4; the canonical cutlass_mxfp8_mxfp4 view is also accepted without conversion. Pass the layer’s quant config to prepare_activations.

__init__() → None

Methods

__init__()
prepare_activations(hidden_states_bf16, *[, ...])	Quantize BF16 activations to MXFP8 for MXFP4×MXFP8.
prepare_weights(w1_bf16, w2_bf16, *, ...[, ...])	Quantize canonical BF16 weights into CUTLASS MXFP4.
supported(arch)