submission 513508

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

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 = hidden_states.shape[0]
    E = config["n_routed_experts"] + config["n_shared_experts"]
    topk = config["total_top_k"]
    d_expert = config["d_expert"]

    # Estimate tokens per expert for block_m selection
    m_per_expert = (M * topk) // E

    # Heuristic: try smaller block_m for sparse cases, larger for dense
    if m_per_expert <= 4:
        block_m = 32
    elif m_per_expert <= 32:
        block_m = 32
    elif m_per_expert <= 64:
        block_m = 64
    else:
        block_m = 128

    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,
        block_size_M=block_m,
        hidden_pad=hidden_pad,
        intermediate_pad=intermediate_pad,
    )

    return output