flashinfer.fused_moe.CudnnFrostNvfp4Config

class flashinfer.fused_moe.CudnnFrostNvfp4Config

Frost NVFP4 MoE on SM107, with packed precomputed routing.

Register prepared weights under cudnn_frost_nvfp4. The canonical cutlass_nvfp4 view is also accepted without conversion. H/I must be multiples of 128. Pass the layer’s quant config to prepare_activations.

__init__() → None

Methods

__init__()
prepare_activations(hidden_states_bf16, *[, ...])	Prepare activations for NVFP4×NVFP4, MXFP4×MXFP8, or MXFP4×BF16.
prepare_weights(w1_bf16, w2_bf16, *, ...[, ...])	Quantize canonical BF16 weights into CUTLASS-swizzled NVFP4.
supported(arch)