submission 740357

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

import os, math
os.environ['PYTORCH_ROCM_ARCH'] = 'gfx950'

import torch, torch.nn as nn
import aiter
from aiter import ActivationType, QuantType, dtypes
from aiter.fused_moe import fused_moe
from task import input_t, output_t


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

    # Mark weights as shuffled — this triggers the optimized CK dispatch path
    # with tuned kernel selection from dsv3_fp4_tuned_fmoe.csv
    if not getattr(gate_up_weight_shuffled, 'is_shuffled', False):
        gate_up_weight_shuffled.is_shuffled = True
    if not getattr(down_weight_shuffled, 'is_shuffled', False):
        down_weight_shuffled.is_shuffled = True

    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=intermediate_pad,
    )