submission 514027

#!POPCORN leaderboard amd-moe-mxfp4
#!POPCORN gpu MI355X
import torch
from task import input_t, output_t
from utils import make_match_reference

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


def custom_kernel(data: input_t) -> output_t:
    """
    Optimized MXFP4 MoE forward pass for MI355X.

    Uses a single fused_moe call with combined routed + shared experts,
    matching the reference kernel's approach for numerical consistency.
    Pre-shuffled weights are used for hardware-optimal CDNA4 tensor core
    utilization.
    """
    (
        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

    hidden_pad = config["d_hidden_pad"] - config["d_hidden"]
    inter_pad  = config["d_expert_pad"] - config["d_expert"]

    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=hidden_pad,
        intermediate_pad=inter_pad,
    )


solution = custom_kernel

from reference import ref_kernel
check_implementation = make_match_reference(ref_kernel, rtol=5e-2, atol=5e-2)