flashinfer.fused_moe.QuantFormat¶

class flashinfer.fused_moe.QuantFormat(value)¶

Numeric format of one MMA operand or of the MoE layer output.

Axes describe the format consumed by the MMA, not the dtype of the tensor that crosses the Python API. A CUTLASS NVFP4 runner that takes BF16 activations and quantizes in-kernel is still (NVFP4, NVFP4).