submission 722647

#!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

# Revert TF32 settings - MXFP4 on CDNA3 needs strict FP4 paths. TF32 conversion slows this down.
torch.backends.cuda.matmul.allow_tf32 = False
torch.backends.cudnn.allow_tf32 = False

# We can pre-calculate the padding instead of doing dictionary lookups in the hot loop
_CFG_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

    M = hidden_states.shape[0]
    
    # 1. Minimal dict lookup overhead.
    if M not in _CFG_CACHE:
        _CFG_CACHE[M] = (
            config["d_hidden_pad"] - config["d_hidden"],
            config["d_expert_pad"] - config["d_expert"]
        )
    hidden_pad, intermediate_pad = _CFG_CACHE[M]

    # 2. Direct pipeline call.
    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,
        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,
    )