flashinfer.fused_moe.CudnnFrostBf16Config

class flashinfer.fused_moe.CudnnFrostBf16Config

Frost BF16 MoE on SM107 and SM120, with packed precomputed routing.

Register prepared weights under cudnn_frost_bf16. The canonical cutlass_bf16 view is also accepted without conversion. Supported model geometries follow the backend’s measured BF16 shortlist.

__init__() → None

Methods

__init__()
prepare_activations(hidden_states_bf16, *[, ...])	Return contiguous BF16 activations and no block scales.
prepare_weights(w1_bf16, w2_bf16, *, ...[, ...])	Build the cutlass_bf16 canonical BF16 weight view.
supported(arch)