submission 738808

#!POPCORN leaderboard amd-moe-mxfp4
#!POPCORN gpu MI355X

"""
MXFP4 MoE — DeepSeek-R1 style fused MoE on MI355X (CDNA4).

Uses CK's fused_moe which handles the full pipeline internally:
  moe_sorting → gate_up GEMM + SiLU → down GEMM → weighted reduction

All using native FP4 MFMA (10 PFLOPS) via pre-tuned ASM kernels.
CK auto-selects block_m, kernel variant, and split strategy per shape.
"""

import torch
from task import input_t, output_t

from aiter import ActivationType, QuantType
from aiter.fused_moe import fused_moe


def custom_kernel(data: input_t) -> output_t:
    (
        hidden_states,
        gate_up_weight, down_weight,
        gate_up_weight_scale, down_weight_scale,
        gate_up_weight_shuffled, down_weight_shuffled,
        gate_up_weight_scale_shuffled, down_weight_scale_shuffled,
        topk_weights, topk_ids,
        config,
    ) = data

    return fused_moe(
        hidden_states,
        gate_up_weight_shuffled,
        down_weight_shuffled,
        topk_weights,
        topk_ids,
        expert_mask=None,
        activation=ActivationType.Silu,
        quant_type=QuantType.per_1x32,
        doweight_stage1=False,
        w1_scale=gate_up_weight_scale_shuffled,
        w2_scale=down_weight_scale_shuffled,
        a1_scale=None,
        a2_scale=None,
        hidden_pad=config["d_hidden_pad"] - config["d_hidden"],
        intermediate_pad=config["d_expert_pad"] - config["d_expert"],
    )