submission 572714

import os
import torch
from typing import Dict
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

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

    M = config["bs"]
    topk = config["total_top_k"]
    E = config["n_routed_experts"] + config["n_shared_experts"]

    # Per-config split-K: use ksplit=4 for low CU utilization configs
    # ksplit=4 gives 31% speedup for small batches (bs<=128, E=33)
    # but hurts large batches (bs=512) due to split-K reduction overhead
    avg_m_per_expert = M * topk / E
    if avg_m_per_expert < 50:
        os.environ['AITER_KSPLIT'] = '4'
    else:
        os.environ['AITER_KSPLIT'] = '0'

    output = 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,
    )

    return output