submission 701458

#!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, dtypes
from aiter.fused_moe import fused_moe
import aiter.fused_moe as _fm

# ── EXP-147: Fix S3 injection with kn1_small (EXP-146 discovery) ──
# EXP-146 found that current aiter defaults S3 (bs=512, E=257, d=256) to cktile kernel
# (305µs). Previously this was handled implicitly by aiter CSV; now it picks cktile.
# Fix: explicitly inject S3 with kn1_small (64x32x32x128_1x1) = 244µs from baseline.
# kn1_32 was tried in EXP-146 and gave 289µs (splitk=4+larger tile worse for sparse E=257).
#
# All other shapes unchanged from EXP-133 (verified best).


def _preload_and_inject():
    try:
        cu = _fm.get_cu_count() if hasattr(_fm, 'get_cu_count') else 256
        act = ActivationType.Silu
        dtype_t = torch.bfloat16
        q_a_t = dtypes.fp4x2
        q_w_t = dtypes.fp4x2
        q_type_t = QuantType.per_1x32

        _fm.get_2stage_cfgs(
            16, 7168, 256, 257, 9,
            dtype_t, q_a_t, q_w_t, q_type_t,
            True, act, False,
            0, 0, True
        )

        cfg = _fm.cfg_2stages
        if cfg is None:
            return

        act_s = str(act)
        dtype_s = str(dtype_t)
        q_a = str(q_a_t)
        q_w = str(q_w_t)
        q_type = str(q_type_t)

        kn1_small = 'moe_ck2stages_gemm1_64x32x32x128_1x1_MulABScaleShuffled_v3_Nswizzle0_Quant3_MulRoutedWeight0_silu_FP4X2_FP4X2_B16'
        kn2_small = 'moe_ck2stages_gemm2_64x32x32x128_1x1_MulABScaleExpertWeightShuffled_v1_Nswizzle0_Quant3_MulRoutedWeight1_FP4X2_FP4X2_B16'
        kn1_32 = 'moe_ck2stages_gemm1_256x32x128x128_1x4_MulABScaleShuffled_v3_Nswizzle0_Quant3_MulRoutedWeight0_silu_FP4X2_FP4X2_B16'
        kn2_32 = 'moe_ck2stages_gemm2_256x32x128x128_1x4_MulABScaleExpertWeightShuffled_v1_Nswizzle0_Quant3_MulRoutedWeight1_FP4X2_FP4X2_B16'
        kn1_64 = 'moe_ck2stages_gemm1_256x64x128x128_1x4_MulABScaleShuffled_v3_Nswizzle0_Quant3_MulRoutedWeight0_silu_FP4X2_FP4X2_B16'
        kn2_fly64_reduce = 'flydsl_moe2_afp4_wfp4_bf16_t64x256x256_reduce'
        kn2_fly64_atomic = 'flydsl_moe2_afp4_wfp4_bf16_t64x256x256_atomic'
        kn2_fly32_atomic_tn128 = 'flydsl_moe2_afp4_wfp4_bf16_t32x128x256_atomic'

        def make_entry(block_m, ksplit, kn1, kn2):
            return {
                'block_m': block_m, 'ksplit': ksplit,
                'kernelName1': kn1, 'kernelName2': kn2,
                'run_1stage': 0,
                'us': 0.0, 'us1': 0.0, 'us2': 0.0,
                'err1': '0', 'err2': '0',
                'tflops': 0.0, 'bw': 0.0, '_tag': float('nan'),
            }

        # Shape 1 (bs=16, E=257, d=256): CK 256x32 ksplit=4
        cfg[(cu, 16, 7168, 256, 257, 9, act_s, dtype_s, q_a, q_w, q_type, True, False)] = \
            make_entry(32, 4, kn1_32, kn2_32)
        # Shape 2 (bs=128, E=257, d=256): CK 256x32 ksplit=4
        cfg[(cu, 128, 7168, 256, 257, 9, act_s, dtype_s, q_a, q_w, q_type, True, False)] = \
            make_entry(32, 4, kn1_32, kn2_32)
        # Shape 3 (bs=512, E=257, d=256): CK 64x32 ksplit=1 (kn1_small) — EXP-147 FIX
        # Current aiter defaults to cktile (305µs). kn1_small (64x32x32x128_1x1) = 244µs.
        # 16 tokens/expert → need small tiles. kn1_32 tried in EXP-146 gave 289µs (worse).
        cfg[(cu, 512, 7168, 256, 257, 9, act_s, dtype_s, q_a, q_w, q_type, True, False)] = \
            make_entry(32, 1, kn1_small, kn2_small)
        # Shape 4 (bs=16, E=33, d=512): CK 256x32 ksplit=2
        cfg[(cu, 16, 7168, 512, 33, 9, act_s, dtype_s, q_a, q_w, q_type, True, False)] = \
            make_entry(32, 2, kn1_32, kn2_32)
        # Shape 5 (bs=128, E=33, d=512): CK 256x32 ksplit=2
        cfg[(cu, 128, 7168, 512, 33, 9, act_s, dtype_s, q_a, q_w, q_type, True, False)] = \
            make_entry(32, 2, kn1_32, kn2_32)
        # Shape 6 (bs=512, E=33, d=512): CK gemm1 256x64 + FlyDSL gemm2 t64 reduce
        cfg[(cu, 512, 7168, 512, 33, 9, act_s, dtype_s, q_a, q_w, q_type, True, False)] = \
            make_entry(64, 0, kn1_64, kn2_fly64_reduce)
        # Shape 7 (bs=512, E=33, d=2048): CK gemm1 256x64 + FlyDSL gemm2 t32 atomic tn128
        cfg[(cu, 512, 7168, 2048, 33, 9, act_s, dtype_s, q_a, q_w, q_type, True, False)] = \
            make_entry(64, 0, kn1_64, kn2_fly32_atomic_tn128)

        if hasattr(_fm.get_2stage_cfgs, 'cache_clear'):
            _fm.get_2stage_cfgs.cache_clear()
        for fn_name in ['get_block_size_M', 'use_nt', 'get_ksplit']:
            fn = getattr(_fm, fn_name, None)
            if fn and hasattr(fn, 'cache_clear'):
                fn.cache_clear()

    except Exception as e:
        import traceback; traceback.print_exc()

_preload_and_inject()


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"]

    try:
        bsm = 32 if hidden_states.shape[0] == 128 and gate_up_weight_shuffled.shape[0] == 33 else None
    except Exception:
        bsm = None

    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,
        block_size_M=bsm,
        hidden_pad=hidden_pad, intermediate_pad=intermediate_pad,
    )