submission 713329

#!POPCORN leaderboard amd-moe-mxfp4
#!POPCORN gpu MI355X
"""
Optimized MXFP4 MoE (Mixture of Experts) kernel for AMD Instinct MI355X.

Key optimizations over baseline:
1. Adaptive doweight_stage1 selection: apply routing weights in stage1 GEMM
   when beneficial (reduces memory bandwidth in the reduction stage).
2. Pre-sort topk_ids to improve expert data locality and reduce bank conflicts.
3. Use aiter's optimized fused_moe with tuned parameters.
4. Minimize Python-side overhead by avoiding redundant dict lookups.
5. Explore using expert_mask for better workload distribution when many
   experts have few tokens.

Benchmark reference (aiter fused_moe, us):
  bs=16,  E=257, d=7168, dexp=256,  top9: 152.7
  bs=128, E=257, d=7168, dexp=256,  top9: 239.0
  bs=512, E=257, d=7168, dexp=256,  top9: 336.5
  bs=16,  E=33,  d=7168, dexp=512,  top9: 106.2
  bs=128, E=33,  d=7168, dexp=512,  top9: 141.1
  bs=512, E=33,  d=7168, dexp=512,  top9: 225.0
  bs=512, E=33,  d=7168, dexp=2048, top9: 380.4
"""
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"]

    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