submission 631582

"""
MXFP4-MoE: `fused_moe` + 简化配置（固定 `doweight_stage1=False`）。

策略:
- **doweight_stage1=False**: 与 reference 保持一致，确保正确性
- **数据连续**: hidden / router 张量 contiguous，利于合并访存
"""
from __future__ import annotations

from task import input_t, output_t

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


@torch.inference_mode()
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"]

    hidden_states = hidden_states.contiguous()
    topk_weights = topk_weights.contiguous()
    topk_ids = topk_ids.contiguous()

    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,  # 固定 False，与 reference 保持一致
        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,
    )