submission 565944

#!POPCORN leaderboard amd-moe-mxfp4
#!POPCORN gpu MI355X

import inspect
from task import input_t, output_t
from aiter import ActivationType, QuantType
from aiter.fused_moe import fused_moe

# Known kernels from your logs (N=256 cases)
K1_SMALLM = "moe_ck2stages_gemm1_64x32x32x128_1x1_MulABScaleShuffled_v3_Nswizzle0_Quant3_MulRoutedWeight0_silu_FP4X2_FP4X2_B16"
K1_LARGEM = "moe_ck2stages_gemm1_256x32x128x128_1x4_MulABScaleShuffled_v3_Nswizzle0_Quant3_MulRoutedWeight0_silu_FP4X2_FP4X2_B16"
K2_COMMON = "moe_ck2stages_gemm2_64x32x32x128_1x1_MulABScaleExpertWeightShuffled_v1_Nswizzle0_Quant3_MulRoutedWeight1_FP4X2_FP4X2_B16"

# Cache signature once (avoid per-call overhead)
try:
    _SIG = inspect.signature(fused_moe)
    FUSED_PARAMS = set(_SIG.parameters.keys())
except Exception:
    FUSED_PARAMS = set()

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"]

    K = int(gate_up_weight_shuffled.shape[-1])  # 7168
    N = int(config["d_expert_pad"])             # 256/512/2048
    TOPK = int(topk_ids.shape[-1])              # 9
    M = int(hidden_states.shape[0])             # 16/128/512

    kwargs = dict(
        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,
    )

    # Keep your only proven-positive lever if supported (H1-style)
    if (K == 7168 and TOPK == 9 and N in (512, 2048) and M >= 256) and ("non_temporal_load" in FUSED_PARAMS):
        kwargs["non_temporal_load"] = True

    # Try kernel forcing ONLY for default family (N=512/2048)
    if (K == 7168 and TOPK == 9 and N in (512, 2048)):
        k1 = K1_SMALLM if M <= 32 else K1_LARGEM
        if "kernelName1" in FUSED_PARAMS:
            kwargs["kernelName1"] = k1
        if "kernelName2" in FUSED_PARAMS:
            kwargs["kernelName2"] = K2_COMMON

    # IMPORTANT: positional for first 5 args
    return fused_moe(
        hidden_states,
        gate_up_weight_shuffled,
        down_weight_shuffled,
        topk_weights,
        topk_ids,
        **kwargs,
    )