submission 655326

"""MoE v19: KSPLIT=4 for small d_expert, KSPLIT=2 for medium, none for large."""
import os
import sys
import torch
from task import input_t, output_t
from aiter import ActivationType, QuantType
from aiter.fused_moe import fused_moe

_first = True

def _setup():
    global _first
    if not _first:
        return
    _first = False
    os.environ["AITER_USE_OPUS_MOE_SORTING"] = "1"
    os.environ["AITER_USE_NT"] = "1"
    try:
        import aiter as _a
        r = os.path.dirname(os.path.dirname(_a.__file__))
        d = os.path.join(r, "aiter", "configs", "model_configs", "dsv3_fp4_tuned_fmoe.csv")
        b = os.path.join(r, "aiter", "configs", "tuned_fmoe.csv")
        p = [x for x in [b, d] if os.path.exists(x)]
        if p:
            os.environ["AITER_CONFIG_FMOE"] = ":".join(p)
    except:
        pass


def custom_kernel(data: input_t) -> output_t:
    _setup()
    (
        hidden_states, guw, dw, gus, ds,
        guw_sh, dw_sh, gus_sh, ds_sh,
        topk_weights, topk_ids, config,
    ) = data

    hp = config["d_hidden_pad"] - config["d_hidden"]
    ip = config["d_expert_pad"] - config["d_expert"]
    de = config["d_expert"]

    # Finer-grained KSPLIT: more aggressive for very small expert dims
    if de <= 256:
        os.environ["AITER_KSPLIT"] = "4"
    elif de <= 512:
        os.environ["AITER_KSPLIT"] = "2"
    else:
        os.environ.pop("AITER_KSPLIT", None)

    return fused_moe(
        hidden_states, guw_sh, dw_sh,
        topk_weights, topk_ids,
        expert_mask=None,
        activation=ActivationType.Silu,
        quant_type=QuantType.per_1x32,
        w1_scale=gus_sh, w2_scale=ds_sh,
        a1_scale=None, a2_scale=None,
        hidden_pad=hp, intermediate_pad=ip,
    )