flashinfer.fused_moe.CuTileNvfp4Config

class flashinfer.fused_moe.CuTileNvfp4Config

cuTile NVFP4-weight x NVFP4-activation backend.

Expert parallelism and fused shared experts are not supported.

__init__() → None

Methods

__init__()
prepare_weights(w1_fp4, w1_block_scale, ...)	Build the cutile_nvfp4 view from checkpoint NVFP4 weights.
supported(arch)