submission 625722

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

import torch
from task import input_t, output_t

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

# Cache pad values per config id (they don't depend on M)
_PAD_CACHE = {}


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

    cfg_id = id(config)
    cached = _PAD_CACHE.get(cfg_id)
    if cached is None:
        hidden_pad = config["d_hidden_pad"] - config["d_hidden"]
        intermediate_pad = config["d_expert_pad"] - config["d_expert"]
        cached = (hidden_pad, intermediate_pad)
        _PAD_CACHE[cfg_id] = cached
    hidden_pad, intermediate_pad = cached

    return fused_moe(
        hidden_states,
        gate_up_weight_shuffled,
        down_weight_shuffled,
        topk_weights,
        topk_ids,
        activation=ActivationType.Silu,
        quant_type=QuantType.per_1x32,
        w1_scale=gate_up_weight_scale_shuffled,
        w2_scale=down_weight_scale_shuffled,
        hidden_pad=hidden_pad,
        intermediate_pad=intermediate_pad,
    )