submission 528025

"""
MoE-MXFP4 submission generated by contestctl.
# Config: PerShapeBlockM
# Notes: Keep mixed-run winners on cases 4/5/6; revert case 3 to default.
"""
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"]

    # Per-shape block_size_M dispatch
    bs = hidden_states.shape[0]
    d_expert = config["d_expert"]
    n_routed_experts = config["n_routed_experts"]
    block_size_M = None
    if bs == 128 and d_expert == 512 and n_routed_experts == 32:
        block_size_M = 32  # Case 4 winner candidate from mixed run
    elif bs == 512 and d_expert == 512 and n_routed_experts == 32:
        block_size_M = 128  # Case 5 winner candidate from mixed run
    elif bs == 512 and d_expert == 2048 and n_routed_experts == 32:
        block_size_M = 64  # Case 6 winner candidate from mixed run

    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,
        block_size_M=block_size_M,
    )

    return output