flashinfer.fused_moe.cute_dsl_fused_moe_mxfp8_mxfp4

flashinfer.fused_moe.cute_dsl_fused_moe_mxfp8_mxfp4(x: Tensor, x_sf: Tensor, token_selected_experts: Tensor, token_final_scales: Tensor, w1_weight: Tensor, w1_weight_sf: Tensor, w1_alpha: Tensor, w2_weight: Tensor, w2_weight_sf: Tensor, w2_alpha: Tensor, num_experts: int, top_k: int, num_local_experts: int | None = None, local_expert_offset: int = 0, moe_output: Tensor | None = None, aux_stream: Stream | None = None, tactic: Tuple[Any, ...] | None = None, enable_pdl: bool = True, activation_type: int = 3, swiglu_alpha: float = 1.0, swiglu_beta: float = 0.0, swiglu_limit: float = 3.4028234663852886e+38) → Tensor

Run fused MoE with MXFP8 activations and packed MXFP4 weights.

No tensor conversion is performed. x must be E4M3, x_sf must be linear block-32 E8M0 bytes, packed weights must be uint8 E2M1 pairs, and weight scales must already use the block-32 MMA E8M0 layout. The output is BF16. Unlike the NVFP4 API, this interface has no fc2_input_scale.