submission 527260

import os
from pathlib import Path

import pandas as pd
import torch
from aiter import ActivationType, QuantType
from aiter.fused_moe import fused_moe
from task import input_t, output_t


def _bootstrap_cfg() -> str:
    out_path = Path("/tmp/josu_cfg_hybrid_under150_fmoe.csv")
    if out_path.exists() and out_path.stat().st_size > 0:
        return str(out_path)

    source_paths = [
        Path("/home/runner/aiter/aiter/configs/tuned_fmoe.csv"),
        Path(
            "/home/runner/aiter/aiter/configs/model_configs/"
            "a8w8_blockscale_tuned_fmoe_qwen3_235b.csv"
        ),
        Path("/home/runner/aiter/aiter/configs/model_configs/dsv3_fp4_tuned_fmoe.csv"),
    ]
    frames = [pd.read_csv(path) for path in source_paths if path.exists()]
    merged = pd.concat(frames, ignore_index=True)

    untuned_path = Path("/home/runner/aiter/aiter/configs/untuned_fmoe.csv")
    keys = pd.read_csv(untuned_path, nrows=0).columns.tolist()
    if "cu_num" not in keys:
        keys.append("cu_num")

    defaults = {column: "" for column in merged.columns}
    defaults.update(
        {
            "cu_num": 256,
            "act_type": "ActivationType.Silu",
            "dtype": "torch.bfloat16",
            "q_dtype_a": "torch.float4_e2m1fn_x2",
            "q_dtype_w": "torch.float4_e2m1fn_x2",
            "q_type": "QuantType.per_1x32",
            "use_g1u1": 1,
            "doweight_stage1": 0,
            "block_m": 32,
            "ksplit": 0,
            "us1": 0.0,
            "kernelName1": "",
            "err1": "0.0%",
            "us2": 0.0,
            "kernelName2": "",
            "err2": "0.0%",
            "us": 0.0,
            "run_1stage": 0,
            "tflops": 0.0,
            "bw": 0.0,
            "_tag": "",
        }
    )

    online_tuned_rows = [
        {
            "token": 16,
            "model_dim": 7168,
            "inter_dim": 512,
            "expert": 33,
            "topk": 9,
            "block_m": 32,
            "ksplit": 0,
            "us": 47.3336,
            "us1": 0.0,
            "kernelName1": "moe_ck2stages_gemm1_64x32x32x128_1x1_MulABScaleShuffled_v3_Nswizzle0_Quant3_MulRoutedWeight0_silu_FP4X2_FP4X2_B16",
            "err1": "0.0%",
            "us2": 0.0,
            "kernelName2": "moe_ck2stages_gemm2_64x32x32x128_1x1_MulABScaleExpertWeightShuffled_v1_Nswizzle0_Quant3_MulRoutedWeight1_FP4X2_FP4X2_B16",
            "err2": "2.9%",
            "run_1stage": 0,
            "tflops": 66.99,
            "bw": 7683.25,
        },
        {
            "token": 128,
            "model_dim": 7168,
            "inter_dim": 512,
            "expert": 33,
            "topk": 9,
            "block_m": 32,
            "ksplit": 0,
            "us": 58.5681,
            "us1": 0.0,
            "kernelName1": "moe_ck2stages_gemm1_64x32x32x128_1x1_MulABScaleShuffled_v3_Nswizzle0_Quant3_MulRoutedWeight0_silu_FP4X2_FP4X2_B16",
            "err1": "0.0%",
            "us2": 0.0,
            "kernelName2": "moe_ck2stages_gemm2_64x32x32x128_1x1_MulABScaleExpertWeightShuffled_v1_Nswizzle0_Quant3_MulRoutedWeight1_FP4X2_FP4X2_B16",
            "err2": "2.9%",
            "run_1stage": 0,
            "tflops": 433.12,
            "bw": 6250.57,
        },
        {
            "token": 512,
            "model_dim": 7168,
            "inter_dim": 512,
            "expert": 33,
            "topk": 9,
            "block_m": 32,
            "ksplit": 0,
            "us": 129.6056,
            "us1": 0.0,
            "kernelName1": "moe_ck2stages_gemm1_256x32x128x128_1x4_MulABScaleShuffled_v3_Nswizzle0_Quant3_MulRoutedWeight0_silu_FP4X2_FP4X2_B16",
            "err1": "0.0%",
            "us2": 0.0,
            "kernelName2": "moe_ck2stages_gemm2_64x32x32x128_1x1_MulABScaleExpertWeightShuffled_v1_Nswizzle0_Quant3_MulRoutedWeight1_FP4X2_FP4X2_B16",
            "err2": "2.9%",
            "run_1stage": 0,
            "tflops": 782.9,
            "bw": 2888.31,
        },
        {
            "token": 512,
            "model_dim": 7168,
            "inter_dim": 2048,
            "expert": 33,
            "topk": 9,
            "block_m": 128,
            "ksplit": 0,
            "us": 275.9788,
            "us1": 0.0,
            "kernelName1": "moe_ck2stages_gemm1_256x128x128x128_1x4_MulABScaleShuffled_v3_Nswizzle0_Quant3_MulRoutedWeight0_silu_FP4X2_FP4X2_B16",
            "err1": "0.0%",
            "us2": 0.0,
            "kernelName2": "moe_ck2stages_gemm2_256x128x128x128_1x4_MulABScaleExpertWeightShuffled_v3_Nswizzle0_Quant3_MulRoutedWeight1_FP4X2_FP4X2_B16",
            "err2": "7.1%",
            "run_1stage": 0,
            "tflops": 1470.67,
            "bw": 5305.97,
        },
    ]

    k2_overrides = [
        (16, 7168, 256, 257, 9),
        (128, 7168, 256, 257, 9),
        (16, 7168, 512, 33, 9),
        (128, 7168, 512, 33, 9),
    ]

    rows = []
    for row in online_tuned_rows:
        seeded = defaults.copy()
        seeded.update(row)
        rows.append(seeded)

    for token, model_dim, inter_dim, expert, topk in k2_overrides:
        mask = (
            (merged["cu_num"] == 256)
            & (merged["token"] == token)
            & (merged["model_dim"] == model_dim)
            & (merged["inter_dim"] == inter_dim)
            & (merged["expert"] == expert)
            & (merged["topk"] == topk)
            & (merged["act_type"] == "ActivationType.Silu")
            & (merged["dtype"] == "torch.bfloat16")
            & (merged["q_dtype_a"] == "torch.float4_e2m1fn_x2")
            & (merged["q_dtype_w"] == "torch.float4_e2m1fn_x2")
            & (merged["q_type"] == "QuantType.per_1x32")
            & (merged["use_g1u1"] == 1)
            & (merged["doweight_stage1"] == 0)
        )
        if mask.any():
            row = merged.loc[mask].sort_values("us").iloc[0].to_dict()
        else:
            row = defaults.copy()
            row.update(
                {
                    "token": token,
                    "model_dim": model_dim,
                    "inter_dim": inter_dim,
                    "expert": expert,
                    "topk": topk,
                }
            )

        row["ksplit"] = 2
        row["us"] = -1.0
        rows.append(row)

    merged = pd.concat([merged, pd.DataFrame(rows)], ignore_index=True)
    dedup_keys = keys + (["_tag"] if "_tag" in merged.columns else [])
    merged = (
        merged.sort_values("us")
        .drop_duplicates(subset=dedup_keys, keep="first")
        .reset_index(drop=True)
    )

    out_path.parent.mkdir(parents=True, exist_ok=True)
    merged.to_csv(out_path, index=False)
    return str(out_path)


os.environ["AITER_CONFIG_FMOE"] = _bootstrap_cfg()


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"]

    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,
    )