#include <iostream>
#include <vector>
#include <getopt.h>
#include <unistd.h>
#include <dirent.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <sys/wait.h>
#include <dlfcn.h>
#include <cstdio>
#include <string>
#include <iomanip>
#include <string.h>
#include <list>
#include <algorithm>
#include <stdarg.h>
#include <fstream>
#include <sstream>
#include "acl/acl.h"
#include "aclnn/aclnn_base.h"
#include "acl/acl_base.h"
#include "acl/acl_rt.h"
#include "sec_common.h"
#include "utils.h"
#include "../fuzzlib/secodeFuzz.h"
#include "aclnnop/aclnn_advance_step.h"

using namespace std;

static const char* GetDtypeName(aclDataType dt) {
    switch (dt) {
        case ACL_FLOAT: return "FLOAT";
        case ACL_FLOAT16: return "FLOAT16";
        case ACL_INT8: return "INT8";
        case ACL_INT32: return "INT32";
        case ACL_UINT8: return "UINT8";
        case ACL_INT16: return "INT16";
        case ACL_UINT16: return "UINT16";
        case ACL_UINT32: return "UINT32";
        case ACL_INT64: return "INT64";
        case ACL_UINT64: return "UINT64";
        case ACL_DOUBLE: return "DOUBLE";
        case ACL_BOOL: return "BOOL";
        case ACL_BF16: return "BFLOAT16";
        case ACL_COMPLEX64: return "COMPLEX64";
        case ACL_COMPLEX128: return "COMPLEX128";
        default: return "UNKNOWN";
    }
}

static std::string ShapeToStr(const std::vector<int64_t>& s) {
    std::string r = "[";
    for (size_t k = 0; k < s.size(); k++) {
        if (k) r += ",";
        r += std::to_string(s[k]);
    }
    return r + "]";
}

static std::string IntArrayToStr(const std::vector<int64_t>& s) {
    std::string r = "[";
    for (size_t k = 0; k < s.size(); k++) {
        if (k) r += ",";
        r += std::to_string(s[k]);
    }
    return r + "]";
}

static std::string FloatArrayToStr(const std::vector<float>& s) {
    std::string r = "[";
    for (size_t k = 0; k < s.size(); k++) {
        if (k) r += ",";
        r += std::to_string(s[k]);
    }
    return r + "]";
}

struct ScenarioResult {
    int scenarioId;
    int retCode;
    std::string desc;
    std::string violation;
    std::string category;
    std::string errorType;
    std::string aclErrMsg;
};
static std::vector<ScenarioResult> g_allResults;
static int g_stdoutFd = -1;

void RedirectStdoutToScenarioFile(int scenarioId) {
    const std::string baseDir = "./opfuzz/scenario_logs";
    mkdir(baseDir.c_str(), 0755);
    std::string logFile = baseDir + "/scenario_" + std::to_string(scenarioId) + ".log";
    fflush(stdout);
    g_stdoutFd = dup(fileno(stdout));
    freopen(logFile.c_str(), "w", stdout);
    setvbuf(stdout, NULL, _IONBF, 0);
}

void RestoreStdout() {
    if (g_stdoutFd < 0) return;
    fflush(stdout);
    dup2(g_stdoutFd, fileno(stdout));
    close(g_stdoutFd);
    g_stdoutFd = -1;
    clearerr(stdout);
}

void GenerateFinalSummary() {
    const std::string baseDir = "./opfuzz/scenario_logs";
    std::string summaryFile = baseDir + "/summary.txt";
    std::ofstream sf(summaryFile);

    int paramErr = 0, innerErr = 0, runtimeErr = 0, otherErr = 0, unexpectedOk = 0, ezPass = 0;
    for (const auto& r : g_allResults) {
        if (r.errorType == "UNEXPECTED_SUCCESS") unexpectedOk++;
        else if (r.errorType == "PARAM_ERROR") paramErr++;
        else if (r.errorType == "INNER_ERROR") innerErr++;
        else if (r.errorType == "RUNTIME_ERROR") runtimeErr++;
        else otherErr++;
        if (r.aclErrMsg.find("should") != std::string::npos ||
            r.aclErrMsg.find("must") != std::string::npos ||
            r.aclErrMsg.find("expected") != std::string::npos ||
            r.aclErrMsg.find("supported") != std::string::npos) {
            ezPass++;
        }
    }
    int total = (int)g_allResults.size();
    int detected = paramErr + innerErr + runtimeErr + otherErr;
    int rate = (total > 0) ? detected * 100 / total : 0;
    int ezRate = (total > 0) ? ezPass * 100 / total : 0;

    sf << "==============================================" << std::endl;
    sf << "  ERROR SCENARIO TEST SUMMARY: aclnnAdvanceStep" << std::endl;
    sf << "==============================================" << std::endl;
    sf << "  Total Scenarios:      " << total << std::endl;
    sf << "  Error Detected:       " << detected << " / " << total << " (" << rate << "%)" << std::endl;
    sf << "    PARAM_ERROR:        " << paramErr << std::endl;
    sf << "    INNER_ERROR:        " << innerErr << std::endl;
    sf << "    RUNTIME_ERROR:      " << runtimeErr << std::endl;
    sf << "    OTHER_ERROR:        " << otherErr << std::endl;
    sf << "  Unexpected Success:   " << unexpectedOk << std::endl;
    sf << "  EZ Compliance:        " << ezPass << " / " << total << " (" << ezRate << "% have corrective hints)" << std::endl;
    sf << "==============================================" << std::endl;
    sf << std::endl;

    sf << std::string(120, '-') << std::endl;
    sf << "  Id | RetCode   | Type          | EZ  | desc" << std::endl;
    sf << std::string(120, '-') << std::endl;
    for (const auto& r : g_allResults) {
        bool ez = (r.aclErrMsg.find("should") != std::string::npos ||
                   r.aclErrMsg.find("must") != std::string::npos ||
                   r.aclErrMsg.find("expected") != std::string::npos);
        sf << "  " << std::setw(2) << r.scenarioId
           << " | " << std::setw(9) << r.retCode
           << " | " << std::setw(13) << r.errorType
           << " | " << (ez ? "OK " : "NO ")
           << " | " << r.desc << std::endl;
    }
    sf << std::string(120, '-') << std::endl;
    sf << std::endl;

    for (const auto& r : g_allResults) {
        std::string logPath = baseDir + "/scenario_" + std::to_string(r.scenarioId) + ".log";
        sf << std::endl << "===== Scenario " << r.scenarioId << " (" << r.errorType << ") ====" << std::endl;
        std::ifstream lf(logPath);
        if (lf.is_open()) {
            std::string line;
            while (std::getline(lf, line)) sf << line << std::endl;
            lf.close();
        } else {
            sf << "  (log file not found)" << std::endl;
        }
    }
    sf.close();

    printf("\n====== FINAL SUMMARY ======\n");
    {
        std::string catCmd = "cat " + summaryFile + " 2>/dev/null";
        system(catCmd.c_str());
    }
    printf("====== END SUMMARY ======\n\n");

    WriteDTLog("aclnnAdvanceStep", "FINAL: total=%d detected=%d param=%d inner=%d runtime=%d unexpected=%d rate=%d%%\n",
               total, detected, paramErr, innerErr, runtimeErr, unexpectedOk, rate);
}

int aclnnAdvanceStep_Normal(int fuzz_num)
{
    auto ret = Init(deviceId, &context, &stream);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);
    WriteDTLog("aclnnAdvanceStep", "aclnnAdvanceStep error scenario test start\n");
    aclrtSetOpExecuteTimeOut(1091);
    DT_Set_TimeOut_Second(3000);
    DT_FUZZ_START(0, fuzz_num, "aclnnAdvanceStep", 0)
    {
        int i = 0;
        void* inputTokensDeviceAddr = nullptr;
        aclTensor* inputTokens = nullptr;
        void* sampledTokenIdsDeviceAddr = nullptr;
        aclTensor* sampledTokenIds = nullptr;
        void* inputPositionsDeviceAddr = nullptr;
        aclTensor* inputPositions = nullptr;
        void* seqLensDeviceAddr = nullptr;
        aclTensor* seqLens = nullptr;
        void* slotMappingDeviceAddr = nullptr;
        aclTensor* slotMapping = nullptr;
        void* blockTablesDeviceAddr = nullptr;
        aclTensor* blockTables = nullptr;
        uint64_t workspaceSize = 0;
        aclOpExecutor* executor;
        void* workspaceAddr = nullptr;
        int scenario = 0;
        int caseRet = -1;
        std::string scenarioDesc, violationDetail, errorCategory, savedErrMsg;
        int expectedErrorCode = 0;
        std::string expectedErrorName, errorTypeStr, errorTargetStr, verifyRequirement;

        LOG_PRINT("aclnnAdvanceStep error test %d\n", fuzzi);

        RedirectStdoutToScenarioFile(fuzzi);

        try
        {
            scenario = *(int *)DTSetGetNumberRange(&g_Element[i++], 0, 0, 42, "i");

            // === Base config ===
            std::vector<int64_t> inputTokensShape = {2};
            std::vector<int64_t> sampledTokenIdsShape = {1, 1};
            std::vector<int64_t> inputPositionsShape = {2};
            std::vector<int64_t> seqLensShape = {2};
            std::vector<int64_t> slotMappingShape = {2};
            std::vector<int64_t> blockTablesShape = {2, 4};

            aclDataType inputTokensDtype = ACL_INT64;
            aclDataType sampledTokenIdsDtype = ACL_INT64;
            aclDataType inputPositionsDtype = ACL_INT64;
            aclDataType seqLensDtype = ACL_INT64;
            aclDataType slotMappingDtype = ACL_INT64;
            aclDataType blockTablesDtype = ACL_INT64;
            aclFormat tensorFormat = ACL_FORMAT_ND;
            bool nullInputTokens = false, nullSampledTokenIds = false, nullInputPositions = false, nullSeqLens = false, nullSlotMapping = false, nullBlockTables = false;

            // === Error scenario injection ===
            if (scenario == 0) {
                // [BASE_CHECK] base config sanity check: no violation injected, expect success
                scenarioDesc = "base config sanity check: no violation injected, expect success";
                violationDetail = "base config sanity check: no violation injected, expect success";
                errorCategory = "BASE";
                expectedErrorCode = 0;
                expectedErrorName = "ACL_SUCCESS";
                errorTypeStr = "BASE_CHECK";
                errorTargetStr = "";
                verifyRequirement = "基线自检:base配置应执行成功(ret==0);若失败则基线污染,所有异常场景报错不可信,需重构造base";
                LOG_PRINT("[BASE_CHECK] base config sanity check: no violation injected, expect success\n");
            }
            else if (scenario == 1) {
                // [A1_inputTokens_ACL_INT8] inputTokens dtype=ACL_INT8 not in supported list [INT64]
                scenarioDesc = "inputTokens dtype=ACL_INT8 not in supported list [INT64]";
                violationDetail = "inputTokens=ACL_INT8";
                errorCategory = "A1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "inputTokens";
                verifyRequirement = "返回码应为161001或161002,错误信息应包含ACL_INT8不支持或invalid dtype相关提示";
                inputTokensDtype = ACL_INT8;  // override
                LOG_PRINT("[A1_inputTokens_ACL_INT8] inputTokens dtype=ACL_INT8 not in supported list [INT64]\n");
            }
            else if (scenario == 2) {
                // [A1_inputTokens_ACL_INT32] inputTokens dtype=ACL_INT32 not in supported list [INT64]
                scenarioDesc = "inputTokens dtype=ACL_INT32 not in supported list [INT64]";
                violationDetail = "inputTokens=ACL_INT32";
                errorCategory = "A1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "inputTokens";
                verifyRequirement = "返回码应为161001或161002,错误信息应包含ACL_INT32不支持或invalid dtype相关提示";
                inputTokensDtype = ACL_INT32;  // override
                LOG_PRINT("[A1_inputTokens_ACL_INT32] inputTokens dtype=ACL_INT32 not in supported list [INT64]\n");
            }
            else if (scenario == 3) {
                // [A3_inputTokens_ACL_FLOAT16] inputTokens must be ACL_INT64, set to ACL_FLOAT16
                scenarioDesc = "inputTokens must be ACL_INT64, set to ACL_FLOAT16";
                violationDetail = "inputTokens=ACL_FLOAT16";
                errorCategory = "A3";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "inputTokens";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示ACL_FLOAT16不支持";
                inputTokensDtype = ACL_FLOAT16;  // override
                LOG_PRINT("[A3_inputTokens_ACL_FLOAT16] inputTokens must be ACL_INT64, set to ACL_FLOAT16\n");
            }
            else if (scenario == 4) {
                // [B1_inputTokens_dim_above_1] inputTokens dim=2D exceeds max=1D
                scenarioDesc = "inputTokens dim=2D exceeds max=1D";
                violationDetail = "inputTokens dim=2D exceeds max=1D";
                errorCategory = "B1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DIM_ILLEGAL";
                errorTargetStr = "inputTokens";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示维度数超限";
                inputTokensShape = {1, 1};  // override
                LOG_PRINT("[B1_inputTokens_dim_above_1] inputTokens dim=2D exceeds max=1D\n");
            }
            else if (scenario == 5) {
                // [C1_inputTokens_NCHW] inputTokens format=ACL_FORMAT_NCHW (only ND supported)
                scenarioDesc = "inputTokens format=ACL_FORMAT_NCHW (only ND supported)";
                violationDetail = "inputTokens format=ACL_FORMAT_NCHW (only ND supported)";
                errorCategory = "C1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "FORMAT_UNSUPPORTED";
                errorTargetStr = "inputTokens";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示format不支持";
                tensorFormat = ACL_FORMAT_NCHW;
                LOG_PRINT("[C1_inputTokens_NCHW] inputTokens format=ACL_FORMAT_NCHW (only ND supported)\n");
            }
            else if (scenario == 6) {
                // [D3_inputTokens_nullptr] inputTokens tensor is nullptr (input pointer is null)
                scenarioDesc = "inputTokens tensor is nullptr (input pointer is null)";
                violationDetail = "inputTokens tensor is nullptr (input pointer is null)";
                errorCategory = "D3";
                expectedErrorCode = 161001;
                expectedErrorName = "ACLNN_ERR_PARAM_NULLPTR";
                errorTypeStr = "NULLPTR_INPUT";
                errorTargetStr = "inputTokens";
                verifyRequirement = "返回码应为161001,错误信息应包含null或nullptr相关提示";
                nullInputTokens = true;
                LOG_PRINT("[D3_inputTokens_nullptr] inputTokens tensor is nullptr (input pointer is null)\n");
            }
            else if (scenario == 7) {
                // [A1_sampledTokenIds_ACL_INT8] sampledTokenIds dtype=ACL_INT8 not in supported list [INT64]
                scenarioDesc = "sampledTokenIds dtype=ACL_INT8 not in supported list [INT64]";
                violationDetail = "sampledTokenIds=ACL_INT8";
                errorCategory = "A1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "sampledTokenIds";
                verifyRequirement = "返回码应为161001或161002,错误信息应包含ACL_INT8不支持或invalid dtype相关提示";
                sampledTokenIdsDtype = ACL_INT8;  // override
                LOG_PRINT("[A1_sampledTokenIds_ACL_INT8] sampledTokenIds dtype=ACL_INT8 not in supported list [INT64]\n");
            }
            else if (scenario == 8) {
                // [A1_sampledTokenIds_ACL_INT32] sampledTokenIds dtype=ACL_INT32 not in supported list [INT64]
                scenarioDesc = "sampledTokenIds dtype=ACL_INT32 not in supported list [INT64]";
                violationDetail = "sampledTokenIds=ACL_INT32";
                errorCategory = "A1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "sampledTokenIds";
                verifyRequirement = "返回码应为161001或161002,错误信息应包含ACL_INT32不支持或invalid dtype相关提示";
                sampledTokenIdsDtype = ACL_INT32;  // override
                LOG_PRINT("[A1_sampledTokenIds_ACL_INT32] sampledTokenIds dtype=ACL_INT32 not in supported list [INT64]\n");
            }
            else if (scenario == 9) {
                // [A3_sampledTokenIds_ACL_FLOAT16] sampledTokenIds must be ACL_INT64, set to ACL_FLOAT16
                scenarioDesc = "sampledTokenIds must be ACL_INT64, set to ACL_FLOAT16";
                violationDetail = "sampledTokenIds=ACL_FLOAT16";
                errorCategory = "A3";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "sampledTokenIds";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示ACL_FLOAT16不支持";
                sampledTokenIdsDtype = ACL_FLOAT16;  // override
                LOG_PRINT("[A3_sampledTokenIds_ACL_FLOAT16] sampledTokenIds must be ACL_INT64, set to ACL_FLOAT16\n");
            }
            else if (scenario == 10) {
                // [B1_sampledTokenIds_dim_above_2] sampledTokenIds dim=3D exceeds max=2D
                scenarioDesc = "sampledTokenIds dim=3D exceeds max=2D";
                violationDetail = "sampledTokenIds dim=3D exceeds max=2D";
                errorCategory = "B1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DIM_ILLEGAL";
                errorTargetStr = "sampledTokenIds";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示维度数超限";
                sampledTokenIdsShape = {1, 1, 1};  // override
                LOG_PRINT("[B1_sampledTokenIds_dim_above_2] sampledTokenIds dim=3D exceeds max=2D\n");
            }
            else if (scenario == 11) {
                // [B1_sampledTokenIds_dim_below_2] sampledTokenIds dim=1D below min=2D
                scenarioDesc = "sampledTokenIds dim=1D below min=2D";
                violationDetail = "sampledTokenIds dim=1D below min=2D";
                errorCategory = "B1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DIM_ILLEGAL";
                errorTargetStr = "sampledTokenIds";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示维度数不足";
                sampledTokenIdsShape = {1};  // override
                LOG_PRINT("[B1_sampledTokenIds_dim_below_2] sampledTokenIds dim=1D below min=2D\n");
            }
            else if (scenario == 12) {
                // [B4_sampledTokenIds_zero_dim] sampledTokenIds shape has zero dimension: [0, 4]
                scenarioDesc = "sampledTokenIds shape has zero dimension: [0, 4]";
                violationDetail = "sampledTokenIds shape has zero dimension: [0, 4]";
                errorCategory = "B4";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "SHAPE_ILLEGAL";
                errorTargetStr = "sampledTokenIds";
                verifyRequirement = "空tensor可能合法也可能非法,需根据md文档确认;若非法应返回161001或161002";
                sampledTokenIdsShape = {0, 4};  // override
                LOG_PRINT("[B4_sampledTokenIds_zero_dim] sampledTokenIds shape has zero dimension: [0, 4]\n");
            }
            else if (scenario == 13) {
                // [C1_sampledTokenIds_NCHW] sampledTokenIds format=ACL_FORMAT_NCHW (only ND supported)
                scenarioDesc = "sampledTokenIds format=ACL_FORMAT_NCHW (only ND supported)";
                violationDetail = "sampledTokenIds format=ACL_FORMAT_NCHW (only ND supported)";
                errorCategory = "C1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "FORMAT_UNSUPPORTED";
                errorTargetStr = "sampledTokenIds";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示format不支持";
                tensorFormat = ACL_FORMAT_NCHW;
                LOG_PRINT("[C1_sampledTokenIds_NCHW] sampledTokenIds format=ACL_FORMAT_NCHW (only ND supported)\n");
            }
            else if (scenario == 14) {
                // [D3_sampledTokenIds_nullptr] sampledTokenIds tensor is nullptr (input pointer is null)
                scenarioDesc = "sampledTokenIds tensor is nullptr (input pointer is null)";
                violationDetail = "sampledTokenIds tensor is nullptr (input pointer is null)";
                errorCategory = "D3";
                expectedErrorCode = 161001;
                expectedErrorName = "ACLNN_ERR_PARAM_NULLPTR";
                errorTypeStr = "NULLPTR_INPUT";
                errorTargetStr = "sampledTokenIds";
                verifyRequirement = "返回码应为161001,错误信息应包含null或nullptr相关提示";
                nullSampledTokenIds = true;
                LOG_PRINT("[D3_sampledTokenIds_nullptr] sampledTokenIds tensor is nullptr (input pointer is null)\n");
            }
            else if (scenario == 15) {
                // [A1_inputPositions_ACL_INT8] inputPositions dtype=ACL_INT8 not in supported list [INT64]
                scenarioDesc = "inputPositions dtype=ACL_INT8 not in supported list [INT64]";
                violationDetail = "inputPositions=ACL_INT8";
                errorCategory = "A1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "inputPositions";
                verifyRequirement = "返回码应为161001或161002,错误信息应包含ACL_INT8不支持或invalid dtype相关提示";
                inputPositionsDtype = ACL_INT8;  // override
                LOG_PRINT("[A1_inputPositions_ACL_INT8] inputPositions dtype=ACL_INT8 not in supported list [INT64]\n");
            }
            else if (scenario == 16) {
                // [A1_inputPositions_ACL_INT32] inputPositions dtype=ACL_INT32 not in supported list [INT64]
                scenarioDesc = "inputPositions dtype=ACL_INT32 not in supported list [INT64]";
                violationDetail = "inputPositions=ACL_INT32";
                errorCategory = "A1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "inputPositions";
                verifyRequirement = "返回码应为161001或161002,错误信息应包含ACL_INT32不支持或invalid dtype相关提示";
                inputPositionsDtype = ACL_INT32;  // override
                LOG_PRINT("[A1_inputPositions_ACL_INT32] inputPositions dtype=ACL_INT32 not in supported list [INT64]\n");
            }
            else if (scenario == 17) {
                // [A3_inputPositions_ACL_FLOAT16] inputPositions must be ACL_INT64, set to ACL_FLOAT16
                scenarioDesc = "inputPositions must be ACL_INT64, set to ACL_FLOAT16";
                violationDetail = "inputPositions=ACL_FLOAT16";
                errorCategory = "A3";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "inputPositions";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示ACL_FLOAT16不支持";
                inputPositionsDtype = ACL_FLOAT16;  // override
                LOG_PRINT("[A3_inputPositions_ACL_FLOAT16] inputPositions must be ACL_INT64, set to ACL_FLOAT16\n");
            }
            else if (scenario == 18) {
                // [B1_inputPositions_dim_above_1] inputPositions dim=2D exceeds max=1D
                scenarioDesc = "inputPositions dim=2D exceeds max=1D";
                violationDetail = "inputPositions dim=2D exceeds max=1D";
                errorCategory = "B1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DIM_ILLEGAL";
                errorTargetStr = "inputPositions";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示维度数超限";
                inputPositionsShape = {1, 1};  // override
                LOG_PRINT("[B1_inputPositions_dim_above_1] inputPositions dim=2D exceeds max=1D\n");
            }
            else if (scenario == 19) {
                // [C1_inputPositions_NCHW] inputPositions format=ACL_FORMAT_NCHW (only ND supported)
                scenarioDesc = "inputPositions format=ACL_FORMAT_NCHW (only ND supported)";
                violationDetail = "inputPositions format=ACL_FORMAT_NCHW (only ND supported)";
                errorCategory = "C1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "FORMAT_UNSUPPORTED";
                errorTargetStr = "inputPositions";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示format不支持";
                tensorFormat = ACL_FORMAT_NCHW;
                LOG_PRINT("[C1_inputPositions_NCHW] inputPositions format=ACL_FORMAT_NCHW (only ND supported)\n");
            }
            else if (scenario == 20) {
                // [D3_inputPositions_nullptr] inputPositions tensor is nullptr (input pointer is null)
                scenarioDesc = "inputPositions tensor is nullptr (input pointer is null)";
                violationDetail = "inputPositions tensor is nullptr (input pointer is null)";
                errorCategory = "D3";
                expectedErrorCode = 161001;
                expectedErrorName = "ACLNN_ERR_PARAM_NULLPTR";
                errorTypeStr = "NULLPTR_INPUT";
                errorTargetStr = "inputPositions";
                verifyRequirement = "返回码应为161001,错误信息应包含null或nullptr相关提示";
                nullInputPositions = true;
                LOG_PRINT("[D3_inputPositions_nullptr] inputPositions tensor is nullptr (input pointer is null)\n");
            }
            else if (scenario == 21) {
                // [A1_seqLens_ACL_INT8] seqLens dtype=ACL_INT8 not in supported list [INT64]
                scenarioDesc = "seqLens dtype=ACL_INT8 not in supported list [INT64]";
                violationDetail = "seqLens=ACL_INT8";
                errorCategory = "A1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "seqLens";
                verifyRequirement = "返回码应为161001或161002,错误信息应包含ACL_INT8不支持或invalid dtype相关提示";
                seqLensDtype = ACL_INT8;  // override
                LOG_PRINT("[A1_seqLens_ACL_INT8] seqLens dtype=ACL_INT8 not in supported list [INT64]\n");
            }
            else if (scenario == 22) {
                // [A1_seqLens_ACL_INT32] seqLens dtype=ACL_INT32 not in supported list [INT64]
                scenarioDesc = "seqLens dtype=ACL_INT32 not in supported list [INT64]";
                violationDetail = "seqLens=ACL_INT32";
                errorCategory = "A1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "seqLens";
                verifyRequirement = "返回码应为161001或161002,错误信息应包含ACL_INT32不支持或invalid dtype相关提示";
                seqLensDtype = ACL_INT32;  // override
                LOG_PRINT("[A1_seqLens_ACL_INT32] seqLens dtype=ACL_INT32 not in supported list [INT64]\n");
            }
            else if (scenario == 23) {
                // [A3_seqLens_ACL_FLOAT16] seqLens must be ACL_INT64, set to ACL_FLOAT16
                scenarioDesc = "seqLens must be ACL_INT64, set to ACL_FLOAT16";
                violationDetail = "seqLens=ACL_FLOAT16";
                errorCategory = "A3";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "seqLens";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示ACL_FLOAT16不支持";
                seqLensDtype = ACL_FLOAT16;  // override
                LOG_PRINT("[A3_seqLens_ACL_FLOAT16] seqLens must be ACL_INT64, set to ACL_FLOAT16\n");
            }
            else if (scenario == 24) {
                // [B1_seqLens_dim_above_1] seqLens dim=2D exceeds max=1D
                scenarioDesc = "seqLens dim=2D exceeds max=1D";
                violationDetail = "seqLens dim=2D exceeds max=1D";
                errorCategory = "B1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DIM_ILLEGAL";
                errorTargetStr = "seqLens";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示维度数超限";
                seqLensShape = {1, 1};  // override
                LOG_PRINT("[B1_seqLens_dim_above_1] seqLens dim=2D exceeds max=1D\n");
            }
            else if (scenario == 25) {
                // [C1_seqLens_NCHW] seqLens format=ACL_FORMAT_NCHW (only ND supported)
                scenarioDesc = "seqLens format=ACL_FORMAT_NCHW (only ND supported)";
                violationDetail = "seqLens format=ACL_FORMAT_NCHW (only ND supported)";
                errorCategory = "C1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "FORMAT_UNSUPPORTED";
                errorTargetStr = "seqLens";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示format不支持";
                tensorFormat = ACL_FORMAT_NCHW;
                LOG_PRINT("[C1_seqLens_NCHW] seqLens format=ACL_FORMAT_NCHW (only ND supported)\n");
            }
            else if (scenario == 26) {
                // [D3_seqLens_nullptr] seqLens tensor is nullptr (input pointer is null)
                scenarioDesc = "seqLens tensor is nullptr (input pointer is null)";
                violationDetail = "seqLens tensor is nullptr (input pointer is null)";
                errorCategory = "D3";
                expectedErrorCode = 161001;
                expectedErrorName = "ACLNN_ERR_PARAM_NULLPTR";
                errorTypeStr = "NULLPTR_INPUT";
                errorTargetStr = "seqLens";
                verifyRequirement = "返回码应为161001,错误信息应包含null或nullptr相关提示";
                nullSeqLens = true;
                LOG_PRINT("[D3_seqLens_nullptr] seqLens tensor is nullptr (input pointer is null)\n");
            }
            else if (scenario == 27) {
                // [A1_slotMapping_ACL_INT8] slotMapping dtype=ACL_INT8 not in supported list [INT64]
                scenarioDesc = "slotMapping dtype=ACL_INT8 not in supported list [INT64]";
                violationDetail = "slotMapping=ACL_INT8";
                errorCategory = "A1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "slotMapping";
                verifyRequirement = "返回码应为161001或161002,错误信息应包含ACL_INT8不支持或invalid dtype相关提示";
                slotMappingDtype = ACL_INT8;  // override
                LOG_PRINT("[A1_slotMapping_ACL_INT8] slotMapping dtype=ACL_INT8 not in supported list [INT64]\n");
            }
            else if (scenario == 28) {
                // [A1_slotMapping_ACL_INT32] slotMapping dtype=ACL_INT32 not in supported list [INT64]
                scenarioDesc = "slotMapping dtype=ACL_INT32 not in supported list [INT64]";
                violationDetail = "slotMapping=ACL_INT32";
                errorCategory = "A1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "slotMapping";
                verifyRequirement = "返回码应为161001或161002,错误信息应包含ACL_INT32不支持或invalid dtype相关提示";
                slotMappingDtype = ACL_INT32;  // override
                LOG_PRINT("[A1_slotMapping_ACL_INT32] slotMapping dtype=ACL_INT32 not in supported list [INT64]\n");
            }
            else if (scenario == 29) {
                // [A3_slotMapping_ACL_FLOAT16] slotMapping must be ACL_INT64, set to ACL_FLOAT16
                scenarioDesc = "slotMapping must be ACL_INT64, set to ACL_FLOAT16";
                violationDetail = "slotMapping=ACL_FLOAT16";
                errorCategory = "A3";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "slotMapping";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示ACL_FLOAT16不支持";
                slotMappingDtype = ACL_FLOAT16;  // override
                LOG_PRINT("[A3_slotMapping_ACL_FLOAT16] slotMapping must be ACL_INT64, set to ACL_FLOAT16\n");
            }
            else if (scenario == 30) {
                // [B1_slotMapping_dim_above_1] slotMapping dim=2D exceeds max=1D
                scenarioDesc = "slotMapping dim=2D exceeds max=1D";
                violationDetail = "slotMapping dim=2D exceeds max=1D";
                errorCategory = "B1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DIM_ILLEGAL";
                errorTargetStr = "slotMapping";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示维度数超限";
                slotMappingShape = {1, 1};  // override
                LOG_PRINT("[B1_slotMapping_dim_above_1] slotMapping dim=2D exceeds max=1D\n");
            }
            else if (scenario == 31) {
                // [C1_slotMapping_NCHW] slotMapping format=ACL_FORMAT_NCHW (only ND supported)
                scenarioDesc = "slotMapping format=ACL_FORMAT_NCHW (only ND supported)";
                violationDetail = "slotMapping format=ACL_FORMAT_NCHW (only ND supported)";
                errorCategory = "C1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "FORMAT_UNSUPPORTED";
                errorTargetStr = "slotMapping";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示format不支持";
                tensorFormat = ACL_FORMAT_NCHW;
                LOG_PRINT("[C1_slotMapping_NCHW] slotMapping format=ACL_FORMAT_NCHW (only ND supported)\n");
            }
            else if (scenario == 32) {
                // [D3_slotMapping_nullptr] slotMapping tensor is nullptr (input pointer is null)
                scenarioDesc = "slotMapping tensor is nullptr (input pointer is null)";
                violationDetail = "slotMapping tensor is nullptr (input pointer is null)";
                errorCategory = "D3";
                expectedErrorCode = 161001;
                expectedErrorName = "ACLNN_ERR_PARAM_NULLPTR";
                errorTypeStr = "NULLPTR_INPUT";
                errorTargetStr = "slotMapping";
                verifyRequirement = "返回码应为161001,错误信息应包含null或nullptr相关提示";
                nullSlotMapping = true;
                LOG_PRINT("[D3_slotMapping_nullptr] slotMapping tensor is nullptr (input pointer is null)\n");
            }
            else if (scenario == 33) {
                // [A1_blockTables_ACL_INT8] blockTables dtype=ACL_INT8 not in supported list [INT64]
                scenarioDesc = "blockTables dtype=ACL_INT8 not in supported list [INT64]";
                violationDetail = "blockTables=ACL_INT8";
                errorCategory = "A1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "blockTables";
                verifyRequirement = "返回码应为161001或161002,错误信息应包含ACL_INT8不支持或invalid dtype相关提示";
                blockTablesDtype = ACL_INT8;  // override
                LOG_PRINT("[A1_blockTables_ACL_INT8] blockTables dtype=ACL_INT8 not in supported list [INT64]\n");
            }
            else if (scenario == 34) {
                // [A1_blockTables_ACL_INT32] blockTables dtype=ACL_INT32 not in supported list [INT64]
                scenarioDesc = "blockTables dtype=ACL_INT32 not in supported list [INT64]";
                violationDetail = "blockTables=ACL_INT32";
                errorCategory = "A1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "blockTables";
                verifyRequirement = "返回码应为161001或161002,错误信息应包含ACL_INT32不支持或invalid dtype相关提示";
                blockTablesDtype = ACL_INT32;  // override
                LOG_PRINT("[A1_blockTables_ACL_INT32] blockTables dtype=ACL_INT32 not in supported list [INT64]\n");
            }
            else if (scenario == 35) {
                // [A3_blockTables_ACL_FLOAT16] blockTables must be ACL_INT64, set to ACL_FLOAT16
                scenarioDesc = "blockTables must be ACL_INT64, set to ACL_FLOAT16";
                violationDetail = "blockTables=ACL_FLOAT16";
                errorCategory = "A3";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_UNSUPPORTED";
                errorTargetStr = "blockTables";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示ACL_FLOAT16不支持";
                blockTablesDtype = ACL_FLOAT16;  // override
                LOG_PRINT("[A3_blockTables_ACL_FLOAT16] blockTables must be ACL_INT64, set to ACL_FLOAT16\n");
            }
            else if (scenario == 36) {
                // [B1_blockTables_dim_above_2] blockTables dim=3D exceeds max=2D
                scenarioDesc = "blockTables dim=3D exceeds max=2D";
                violationDetail = "blockTables dim=3D exceeds max=2D";
                errorCategory = "B1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DIM_ILLEGAL";
                errorTargetStr = "blockTables";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示维度数超限";
                blockTablesShape = {1, 1, 1};  // override
                LOG_PRINT("[B1_blockTables_dim_above_2] blockTables dim=3D exceeds max=2D\n");
            }
            else if (scenario == 37) {
                // [B1_blockTables_dim_below_2] blockTables dim=1D below min=2D
                scenarioDesc = "blockTables dim=1D below min=2D";
                violationDetail = "blockTables dim=1D below min=2D";
                errorCategory = "B1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DIM_ILLEGAL";
                errorTargetStr = "blockTables";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示维度数不足";
                blockTablesShape = {1};  // override
                LOG_PRINT("[B1_blockTables_dim_below_2] blockTables dim=1D below min=2D\n");
            }
            else if (scenario == 38) {
                // [B4_blockTables_zero_dim] blockTables shape has zero dimension: [0, 4]
                scenarioDesc = "blockTables shape has zero dimension: [0, 4]";
                violationDetail = "blockTables shape has zero dimension: [0, 4]";
                errorCategory = "B4";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "SHAPE_ILLEGAL";
                errorTargetStr = "blockTables";
                verifyRequirement = "空tensor可能合法也可能非法,需根据md文档确认;若非法应返回161001或161002";
                blockTablesShape = {0, 4};  // override
                LOG_PRINT("[B4_blockTables_zero_dim] blockTables shape has zero dimension: [0, 4]\n");
            }
            else if (scenario == 39) {
                // [C1_blockTables_NCHW] blockTables format=ACL_FORMAT_NCHW (only ND supported)
                scenarioDesc = "blockTables format=ACL_FORMAT_NCHW (only ND supported)";
                violationDetail = "blockTables format=ACL_FORMAT_NCHW (only ND supported)";
                errorCategory = "C1";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "FORMAT_UNSUPPORTED";
                errorTargetStr = "blockTables";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示format不支持";
                tensorFormat = ACL_FORMAT_NCHW;
                LOG_PRINT("[C1_blockTables_NCHW] blockTables format=ACL_FORMAT_NCHW (only ND supported)\n");
            }
            else if (scenario == 40) {
                // [D3_blockTables_nullptr] blockTables tensor is nullptr (input pointer is null)
                scenarioDesc = "blockTables tensor is nullptr (input pointer is null)";
                violationDetail = "blockTables tensor is nullptr (input pointer is null)";
                errorCategory = "D3";
                expectedErrorCode = 161001;
                expectedErrorName = "ACLNN_ERR_PARAM_NULLPTR";
                errorTypeStr = "NULLPTR_INPUT";
                errorTargetStr = "blockTables";
                verifyRequirement = "返回码应为161001,错误信息应包含null或nullptr相关提示";
                nullBlockTables = true;
                LOG_PRINT("[D3_blockTables_nullptr] blockTables tensor is nullptr (input pointer is null)\n");
            }
            else if (scenario == 41) {
                // [A2_sampledTokenIds_vs_inputTokens_dtype_mismatch] inputTokens=ACL_INT64 but sampledTokenIds=ACL_INT8 (pair requires same)
                scenarioDesc = "inputTokens=ACL_INT64 but sampledTokenIds=ACL_INT8 (pair requires same)";
                violationDetail = "sampledTokenIds=ACL_INT8";
                errorCategory = "A2";
                expectedErrorCode = 161002;
                expectedErrorName = "ACLNN_ERR_PARAM_INVALID";
                errorTypeStr = "DTYPE_MISMATCH";
                errorTargetStr = "sampledTokenIds+inputTokens";
                verifyRequirement = "返回码应为161001或161002,错误信息应提示dtype不匹配";
                sampledTokenIdsDtype = ACL_INT8;  // override
                LOG_PRINT("[A2_sampledTokenIds_vs_inputTokens_dtype_mismatch] inputTokens=ACL_INT64 but sampledTokenIds=ACL_INT8 (pair requires same)\n");
            }
            else if (scenario == 42) {
                // [B2_sampledTokenIds_vs_inputTokens_mismatch] sampledTokenIds.shape=[5] mismatch with inputTokens.shape=[2]
                scenarioDesc = "sampledTokenIds.shape=[5] mismatch with inputTokens.shape=[2]";
                violationDetail = "sampledTokenIds.shape=[5] mismatch with inputTokens.shape=[2]";
                errorCategory = "B2";
                expectedErrorCode = 561002;
                expectedErrorName = "ACLNN_ERR_INNER_SHAPE_MISMATCH";
                errorTypeStr = "SHAPE_MISMATCH";
                errorTargetStr = "sampledTokenIds+inputTokens";
                verifyRequirement = "返回码可能为561001或561002,错误信息应提示shape不匹配";
                sampledTokenIdsShape = {5};  // override
                LOG_PRINT("[B2_sampledTokenIds_vs_inputTokens_mismatch] sampledTokenIds.shape=[5] mismatch with inputTokens.shape=[2]\n");
            }
            else {
                scenarioDesc = "OUT_OF_RANGE";
                violationDetail = "scenario index exceeds total";
                errorCategory = "SKIP";
                LOG_PRINT("Scenario %d out of range (0-42), skipping\n", scenario);
            }

            // === Create tensors ===
            if (!nullInputTokens) {
                long long inputTokens_sz = GetShapeSize(inputTokensShape);
                if (inputTokens_sz > 0 && inputTokens_sz <= 1073741824) {
                    if (inputTokensDtype == ACL_FLOAT) {
                        std::vector<float> itH(inputTokens_sz, 1);
                        ret = CreateAclTensor(itH, inputTokensShape, &inputTokensDeviceAddr, inputTokensDtype, &inputTokens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputTokensDtype == ACL_FLOAT16) {
                        std::vector<int16_t> itH(inputTokens_sz, 1);
                        ret = CreateAclTensor(itH, inputTokensShape, &inputTokensDeviceAddr, inputTokensDtype, &inputTokens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputTokensDtype == ACL_INT8) {
                        std::vector<int8_t> itH(inputTokens_sz, 1);
                        ret = CreateAclTensor(itH, inputTokensShape, &inputTokensDeviceAddr, inputTokensDtype, &inputTokens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputTokensDtype == ACL_INT32) {
                        std::vector<int32_t> itH(inputTokens_sz, 1);
                        ret = CreateAclTensor(itH, inputTokensShape, &inputTokensDeviceAddr, inputTokensDtype, &inputTokens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputTokensDtype == ACL_UINT8) {
                        std::vector<uint8_t> itH(inputTokens_sz, 1);
                        ret = CreateAclTensor(itH, inputTokensShape, &inputTokensDeviceAddr, inputTokensDtype, &inputTokens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputTokensDtype == ACL_INT16) {
                        std::vector<int16_t> itH(inputTokens_sz, 1);
                        ret = CreateAclTensor(itH, inputTokensShape, &inputTokensDeviceAddr, inputTokensDtype, &inputTokens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputTokensDtype == ACL_UINT16) {
                        std::vector<uint16_t> itH(inputTokens_sz, 1);
                        ret = CreateAclTensor(itH, inputTokensShape, &inputTokensDeviceAddr, inputTokensDtype, &inputTokens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputTokensDtype == ACL_UINT32) {
                        std::vector<uint32_t> itH(inputTokens_sz, 1);
                        ret = CreateAclTensor(itH, inputTokensShape, &inputTokensDeviceAddr, inputTokensDtype, &inputTokens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputTokensDtype == ACL_INT64) {
                        std::vector<int64_t> itH(inputTokens_sz, 1);
                        ret = CreateAclTensor(itH, inputTokensShape, &inputTokensDeviceAddr, inputTokensDtype, &inputTokens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputTokensDtype == ACL_UINT64) {
                        std::vector<uint64_t> itH(inputTokens_sz, 1);
                        ret = CreateAclTensor(itH, inputTokensShape, &inputTokensDeviceAddr, inputTokensDtype, &inputTokens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputTokensDtype == ACL_DOUBLE) {
                        std::vector<double> itH(inputTokens_sz, 1);
                        ret = CreateAclTensor(itH, inputTokensShape, &inputTokensDeviceAddr, inputTokensDtype, &inputTokens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputTokensDtype == ACL_BOOL) {
                        std::vector<int8_t> itH(inputTokens_sz, 1);
                        ret = CreateAclTensor(itH, inputTokensShape, &inputTokensDeviceAddr, inputTokensDtype, &inputTokens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputTokensDtype == ACL_BF16) {
                        std::vector<int16_t> itH(inputTokens_sz, 1);
                        ret = CreateAclTensor(itH, inputTokensShape, &inputTokensDeviceAddr, inputTokensDtype, &inputTokens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputTokensDtype == ACL_COMPLEX64) {
                        std::vector<float> itH(inputTokens_sz, 1);
                        ret = CreateAclTensor(itH, inputTokensShape, &inputTokensDeviceAddr, inputTokensDtype, &inputTokens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputTokensDtype == ACL_COMPLEX128) {
                        std::vector<double> itH(inputTokens_sz, 1);
                        ret = CreateAclTensor(itH, inputTokensShape, &inputTokensDeviceAddr, inputTokensDtype, &inputTokens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else {
                        std::vector<float> itH(inputTokens_sz, 1);
                        ret = CreateAclTensor(itH, inputTokensShape, &inputTokensDeviceAddr, inputTokensDtype, &inputTokens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    }
                }
            } else {
                inputTokens = nullptr;
            }

            if (!nullSampledTokenIds) {
                long long sampledTokenIds_sz = GetShapeSize(sampledTokenIdsShape);
                if (sampledTokenIds_sz > 0 && sampledTokenIds_sz <= 1073741824) {
                    if (sampledTokenIdsDtype == ACL_FLOAT) {
                        std::vector<float> stiH(sampledTokenIds_sz, 1);
                        ret = CreateAclTensor(stiH, sampledTokenIdsShape, &sampledTokenIdsDeviceAddr, sampledTokenIdsDtype, &sampledTokenIds);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (sampledTokenIdsDtype == ACL_FLOAT16) {
                        std::vector<int16_t> stiH(sampledTokenIds_sz, 1);
                        ret = CreateAclTensor(stiH, sampledTokenIdsShape, &sampledTokenIdsDeviceAddr, sampledTokenIdsDtype, &sampledTokenIds);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (sampledTokenIdsDtype == ACL_INT8) {
                        std::vector<int8_t> stiH(sampledTokenIds_sz, 1);
                        ret = CreateAclTensor(stiH, sampledTokenIdsShape, &sampledTokenIdsDeviceAddr, sampledTokenIdsDtype, &sampledTokenIds);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (sampledTokenIdsDtype == ACL_INT32) {
                        std::vector<int32_t> stiH(sampledTokenIds_sz, 1);
                        ret = CreateAclTensor(stiH, sampledTokenIdsShape, &sampledTokenIdsDeviceAddr, sampledTokenIdsDtype, &sampledTokenIds);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (sampledTokenIdsDtype == ACL_UINT8) {
                        std::vector<uint8_t> stiH(sampledTokenIds_sz, 1);
                        ret = CreateAclTensor(stiH, sampledTokenIdsShape, &sampledTokenIdsDeviceAddr, sampledTokenIdsDtype, &sampledTokenIds);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (sampledTokenIdsDtype == ACL_INT16) {
                        std::vector<int16_t> stiH(sampledTokenIds_sz, 1);
                        ret = CreateAclTensor(stiH, sampledTokenIdsShape, &sampledTokenIdsDeviceAddr, sampledTokenIdsDtype, &sampledTokenIds);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (sampledTokenIdsDtype == ACL_UINT16) {
                        std::vector<uint16_t> stiH(sampledTokenIds_sz, 1);
                        ret = CreateAclTensor(stiH, sampledTokenIdsShape, &sampledTokenIdsDeviceAddr, sampledTokenIdsDtype, &sampledTokenIds);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (sampledTokenIdsDtype == ACL_UINT32) {
                        std::vector<uint32_t> stiH(sampledTokenIds_sz, 1);
                        ret = CreateAclTensor(stiH, sampledTokenIdsShape, &sampledTokenIdsDeviceAddr, sampledTokenIdsDtype, &sampledTokenIds);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (sampledTokenIdsDtype == ACL_INT64) {
                        std::vector<int64_t> stiH(sampledTokenIds_sz, 1);
                        ret = CreateAclTensor(stiH, sampledTokenIdsShape, &sampledTokenIdsDeviceAddr, sampledTokenIdsDtype, &sampledTokenIds);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (sampledTokenIdsDtype == ACL_UINT64) {
                        std::vector<uint64_t> stiH(sampledTokenIds_sz, 1);
                        ret = CreateAclTensor(stiH, sampledTokenIdsShape, &sampledTokenIdsDeviceAddr, sampledTokenIdsDtype, &sampledTokenIds);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (sampledTokenIdsDtype == ACL_DOUBLE) {
                        std::vector<double> stiH(sampledTokenIds_sz, 1);
                        ret = CreateAclTensor(stiH, sampledTokenIdsShape, &sampledTokenIdsDeviceAddr, sampledTokenIdsDtype, &sampledTokenIds);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (sampledTokenIdsDtype == ACL_BOOL) {
                        std::vector<int8_t> stiH(sampledTokenIds_sz, 1);
                        ret = CreateAclTensor(stiH, sampledTokenIdsShape, &sampledTokenIdsDeviceAddr, sampledTokenIdsDtype, &sampledTokenIds);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (sampledTokenIdsDtype == ACL_BF16) {
                        std::vector<int16_t> stiH(sampledTokenIds_sz, 1);
                        ret = CreateAclTensor(stiH, sampledTokenIdsShape, &sampledTokenIdsDeviceAddr, sampledTokenIdsDtype, &sampledTokenIds);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (sampledTokenIdsDtype == ACL_COMPLEX64) {
                        std::vector<float> stiH(sampledTokenIds_sz, 1);
                        ret = CreateAclTensor(stiH, sampledTokenIdsShape, &sampledTokenIdsDeviceAddr, sampledTokenIdsDtype, &sampledTokenIds);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (sampledTokenIdsDtype == ACL_COMPLEX128) {
                        std::vector<double> stiH(sampledTokenIds_sz, 1);
                        ret = CreateAclTensor(stiH, sampledTokenIdsShape, &sampledTokenIdsDeviceAddr, sampledTokenIdsDtype, &sampledTokenIds);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else {
                        std::vector<float> stiH(sampledTokenIds_sz, 1);
                        ret = CreateAclTensor(stiH, sampledTokenIdsShape, &sampledTokenIdsDeviceAddr, sampledTokenIdsDtype, &sampledTokenIds);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    }
                }
            } else {
                sampledTokenIds = nullptr;
            }

            if (!nullInputPositions) {
                long long inputPositions_sz = GetShapeSize(inputPositionsShape);
                if (inputPositions_sz > 0 && inputPositions_sz <= 1073741824) {
                    if (inputPositionsDtype == ACL_FLOAT) {
                        std::vector<float> ipH(inputPositions_sz, 1);
                        ret = CreateAclTensor(ipH, inputPositionsShape, &inputPositionsDeviceAddr, inputPositionsDtype, &inputPositions);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputPositionsDtype == ACL_FLOAT16) {
                        std::vector<int16_t> ipH(inputPositions_sz, 1);
                        ret = CreateAclTensor(ipH, inputPositionsShape, &inputPositionsDeviceAddr, inputPositionsDtype, &inputPositions);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputPositionsDtype == ACL_INT8) {
                        std::vector<int8_t> ipH(inputPositions_sz, 1);
                        ret = CreateAclTensor(ipH, inputPositionsShape, &inputPositionsDeviceAddr, inputPositionsDtype, &inputPositions);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputPositionsDtype == ACL_INT32) {
                        std::vector<int32_t> ipH(inputPositions_sz, 1);
                        ret = CreateAclTensor(ipH, inputPositionsShape, &inputPositionsDeviceAddr, inputPositionsDtype, &inputPositions);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputPositionsDtype == ACL_UINT8) {
                        std::vector<uint8_t> ipH(inputPositions_sz, 1);
                        ret = CreateAclTensor(ipH, inputPositionsShape, &inputPositionsDeviceAddr, inputPositionsDtype, &inputPositions);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputPositionsDtype == ACL_INT16) {
                        std::vector<int16_t> ipH(inputPositions_sz, 1);
                        ret = CreateAclTensor(ipH, inputPositionsShape, &inputPositionsDeviceAddr, inputPositionsDtype, &inputPositions);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputPositionsDtype == ACL_UINT16) {
                        std::vector<uint16_t> ipH(inputPositions_sz, 1);
                        ret = CreateAclTensor(ipH, inputPositionsShape, &inputPositionsDeviceAddr, inputPositionsDtype, &inputPositions);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputPositionsDtype == ACL_UINT32) {
                        std::vector<uint32_t> ipH(inputPositions_sz, 1);
                        ret = CreateAclTensor(ipH, inputPositionsShape, &inputPositionsDeviceAddr, inputPositionsDtype, &inputPositions);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputPositionsDtype == ACL_INT64) {
                        std::vector<int64_t> ipH(inputPositions_sz, 1);
                        ret = CreateAclTensor(ipH, inputPositionsShape, &inputPositionsDeviceAddr, inputPositionsDtype, &inputPositions);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputPositionsDtype == ACL_UINT64) {
                        std::vector<uint64_t> ipH(inputPositions_sz, 1);
                        ret = CreateAclTensor(ipH, inputPositionsShape, &inputPositionsDeviceAddr, inputPositionsDtype, &inputPositions);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputPositionsDtype == ACL_DOUBLE) {
                        std::vector<double> ipH(inputPositions_sz, 1);
                        ret = CreateAclTensor(ipH, inputPositionsShape, &inputPositionsDeviceAddr, inputPositionsDtype, &inputPositions);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputPositionsDtype == ACL_BOOL) {
                        std::vector<int8_t> ipH(inputPositions_sz, 1);
                        ret = CreateAclTensor(ipH, inputPositionsShape, &inputPositionsDeviceAddr, inputPositionsDtype, &inputPositions);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputPositionsDtype == ACL_BF16) {
                        std::vector<int16_t> ipH(inputPositions_sz, 1);
                        ret = CreateAclTensor(ipH, inputPositionsShape, &inputPositionsDeviceAddr, inputPositionsDtype, &inputPositions);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputPositionsDtype == ACL_COMPLEX64) {
                        std::vector<float> ipH(inputPositions_sz, 1);
                        ret = CreateAclTensor(ipH, inputPositionsShape, &inputPositionsDeviceAddr, inputPositionsDtype, &inputPositions);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (inputPositionsDtype == ACL_COMPLEX128) {
                        std::vector<double> ipH(inputPositions_sz, 1);
                        ret = CreateAclTensor(ipH, inputPositionsShape, &inputPositionsDeviceAddr, inputPositionsDtype, &inputPositions);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else {
                        std::vector<float> ipH(inputPositions_sz, 1);
                        ret = CreateAclTensor(ipH, inputPositionsShape, &inputPositionsDeviceAddr, inputPositionsDtype, &inputPositions);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    }
                }
            } else {
                inputPositions = nullptr;
            }

            if (!nullSeqLens) {
                long long seqLens_sz = GetShapeSize(seqLensShape);
                if (seqLens_sz > 0 && seqLens_sz <= 1073741824) {
                    if (seqLensDtype == ACL_FLOAT) {
                        std::vector<float> slH(seqLens_sz, 1);
                        ret = CreateAclTensor(slH, seqLensShape, &seqLensDeviceAddr, seqLensDtype, &seqLens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (seqLensDtype == ACL_FLOAT16) {
                        std::vector<int16_t> slH(seqLens_sz, 1);
                        ret = CreateAclTensor(slH, seqLensShape, &seqLensDeviceAddr, seqLensDtype, &seqLens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (seqLensDtype == ACL_INT8) {
                        std::vector<int8_t> slH(seqLens_sz, 1);
                        ret = CreateAclTensor(slH, seqLensShape, &seqLensDeviceAddr, seqLensDtype, &seqLens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (seqLensDtype == ACL_INT32) {
                        std::vector<int32_t> slH(seqLens_sz, 1);
                        ret = CreateAclTensor(slH, seqLensShape, &seqLensDeviceAddr, seqLensDtype, &seqLens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (seqLensDtype == ACL_UINT8) {
                        std::vector<uint8_t> slH(seqLens_sz, 1);
                        ret = CreateAclTensor(slH, seqLensShape, &seqLensDeviceAddr, seqLensDtype, &seqLens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (seqLensDtype == ACL_INT16) {
                        std::vector<int16_t> slH(seqLens_sz, 1);
                        ret = CreateAclTensor(slH, seqLensShape, &seqLensDeviceAddr, seqLensDtype, &seqLens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (seqLensDtype == ACL_UINT16) {
                        std::vector<uint16_t> slH(seqLens_sz, 1);
                        ret = CreateAclTensor(slH, seqLensShape, &seqLensDeviceAddr, seqLensDtype, &seqLens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (seqLensDtype == ACL_UINT32) {
                        std::vector<uint32_t> slH(seqLens_sz, 1);
                        ret = CreateAclTensor(slH, seqLensShape, &seqLensDeviceAddr, seqLensDtype, &seqLens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (seqLensDtype == ACL_INT64) {
                        std::vector<int64_t> slH(seqLens_sz, 1);
                        ret = CreateAclTensor(slH, seqLensShape, &seqLensDeviceAddr, seqLensDtype, &seqLens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (seqLensDtype == ACL_UINT64) {
                        std::vector<uint64_t> slH(seqLens_sz, 1);
                        ret = CreateAclTensor(slH, seqLensShape, &seqLensDeviceAddr, seqLensDtype, &seqLens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (seqLensDtype == ACL_DOUBLE) {
                        std::vector<double> slH(seqLens_sz, 1);
                        ret = CreateAclTensor(slH, seqLensShape, &seqLensDeviceAddr, seqLensDtype, &seqLens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (seqLensDtype == ACL_BOOL) {
                        std::vector<int8_t> slH(seqLens_sz, 1);
                        ret = CreateAclTensor(slH, seqLensShape, &seqLensDeviceAddr, seqLensDtype, &seqLens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (seqLensDtype == ACL_BF16) {
                        std::vector<int16_t> slH(seqLens_sz, 1);
                        ret = CreateAclTensor(slH, seqLensShape, &seqLensDeviceAddr, seqLensDtype, &seqLens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (seqLensDtype == ACL_COMPLEX64) {
                        std::vector<float> slH(seqLens_sz, 1);
                        ret = CreateAclTensor(slH, seqLensShape, &seqLensDeviceAddr, seqLensDtype, &seqLens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (seqLensDtype == ACL_COMPLEX128) {
                        std::vector<double> slH(seqLens_sz, 1);
                        ret = CreateAclTensor(slH, seqLensShape, &seqLensDeviceAddr, seqLensDtype, &seqLens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else {
                        std::vector<float> slH(seqLens_sz, 1);
                        ret = CreateAclTensor(slH, seqLensShape, &seqLensDeviceAddr, seqLensDtype, &seqLens);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    }
                }
            } else {
                seqLens = nullptr;
            }

            if (!nullSlotMapping) {
                long long slotMapping_sz = GetShapeSize(slotMappingShape);
                if (slotMapping_sz > 0 && slotMapping_sz <= 1073741824) {
                    if (slotMappingDtype == ACL_FLOAT) {
                        std::vector<float> smH(slotMapping_sz, 1);
                        ret = CreateAclTensor(smH, slotMappingShape, &slotMappingDeviceAddr, slotMappingDtype, &slotMapping);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (slotMappingDtype == ACL_FLOAT16) {
                        std::vector<int16_t> smH(slotMapping_sz, 1);
                        ret = CreateAclTensor(smH, slotMappingShape, &slotMappingDeviceAddr, slotMappingDtype, &slotMapping);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (slotMappingDtype == ACL_INT8) {
                        std::vector<int8_t> smH(slotMapping_sz, 1);
                        ret = CreateAclTensor(smH, slotMappingShape, &slotMappingDeviceAddr, slotMappingDtype, &slotMapping);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (slotMappingDtype == ACL_INT32) {
                        std::vector<int32_t> smH(slotMapping_sz, 1);
                        ret = CreateAclTensor(smH, slotMappingShape, &slotMappingDeviceAddr, slotMappingDtype, &slotMapping);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (slotMappingDtype == ACL_UINT8) {
                        std::vector<uint8_t> smH(slotMapping_sz, 1);
                        ret = CreateAclTensor(smH, slotMappingShape, &slotMappingDeviceAddr, slotMappingDtype, &slotMapping);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (slotMappingDtype == ACL_INT16) {
                        std::vector<int16_t> smH(slotMapping_sz, 1);
                        ret = CreateAclTensor(smH, slotMappingShape, &slotMappingDeviceAddr, slotMappingDtype, &slotMapping);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (slotMappingDtype == ACL_UINT16) {
                        std::vector<uint16_t> smH(slotMapping_sz, 1);
                        ret = CreateAclTensor(smH, slotMappingShape, &slotMappingDeviceAddr, slotMappingDtype, &slotMapping);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (slotMappingDtype == ACL_UINT32) {
                        std::vector<uint32_t> smH(slotMapping_sz, 1);
                        ret = CreateAclTensor(smH, slotMappingShape, &slotMappingDeviceAddr, slotMappingDtype, &slotMapping);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (slotMappingDtype == ACL_INT64) {
                        std::vector<int64_t> smH(slotMapping_sz, 1);
                        ret = CreateAclTensor(smH, slotMappingShape, &slotMappingDeviceAddr, slotMappingDtype, &slotMapping);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (slotMappingDtype == ACL_UINT64) {
                        std::vector<uint64_t> smH(slotMapping_sz, 1);
                        ret = CreateAclTensor(smH, slotMappingShape, &slotMappingDeviceAddr, slotMappingDtype, &slotMapping);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (slotMappingDtype == ACL_DOUBLE) {
                        std::vector<double> smH(slotMapping_sz, 1);
                        ret = CreateAclTensor(smH, slotMappingShape, &slotMappingDeviceAddr, slotMappingDtype, &slotMapping);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (slotMappingDtype == ACL_BOOL) {
                        std::vector<int8_t> smH(slotMapping_sz, 1);
                        ret = CreateAclTensor(smH, slotMappingShape, &slotMappingDeviceAddr, slotMappingDtype, &slotMapping);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (slotMappingDtype == ACL_BF16) {
                        std::vector<int16_t> smH(slotMapping_sz, 1);
                        ret = CreateAclTensor(smH, slotMappingShape, &slotMappingDeviceAddr, slotMappingDtype, &slotMapping);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (slotMappingDtype == ACL_COMPLEX64) {
                        std::vector<float> smH(slotMapping_sz, 1);
                        ret = CreateAclTensor(smH, slotMappingShape, &slotMappingDeviceAddr, slotMappingDtype, &slotMapping);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (slotMappingDtype == ACL_COMPLEX128) {
                        std::vector<double> smH(slotMapping_sz, 1);
                        ret = CreateAclTensor(smH, slotMappingShape, &slotMappingDeviceAddr, slotMappingDtype, &slotMapping);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else {
                        std::vector<float> smH(slotMapping_sz, 1);
                        ret = CreateAclTensor(smH, slotMappingShape, &slotMappingDeviceAddr, slotMappingDtype, &slotMapping);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    }
                }
            } else {
                slotMapping = nullptr;
            }

            if (!nullBlockTables) {
                long long blockTables_sz = GetShapeSize(blockTablesShape);
                if (blockTables_sz > 0 && blockTables_sz <= 1073741824) {
                    if (blockTablesDtype == ACL_FLOAT) {
                        std::vector<float> btH(blockTables_sz, 1);
                        ret = CreateAclTensor(btH, blockTablesShape, &blockTablesDeviceAddr, blockTablesDtype, &blockTables);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (blockTablesDtype == ACL_FLOAT16) {
                        std::vector<int16_t> btH(blockTables_sz, 1);
                        ret = CreateAclTensor(btH, blockTablesShape, &blockTablesDeviceAddr, blockTablesDtype, &blockTables);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (blockTablesDtype == ACL_INT8) {
                        std::vector<int8_t> btH(blockTables_sz, 1);
                        ret = CreateAclTensor(btH, blockTablesShape, &blockTablesDeviceAddr, blockTablesDtype, &blockTables);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (blockTablesDtype == ACL_INT32) {
                        std::vector<int32_t> btH(blockTables_sz, 1);
                        ret = CreateAclTensor(btH, blockTablesShape, &blockTablesDeviceAddr, blockTablesDtype, &blockTables);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (blockTablesDtype == ACL_UINT8) {
                        std::vector<uint8_t> btH(blockTables_sz, 1);
                        ret = CreateAclTensor(btH, blockTablesShape, &blockTablesDeviceAddr, blockTablesDtype, &blockTables);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (blockTablesDtype == ACL_INT16) {
                        std::vector<int16_t> btH(blockTables_sz, 1);
                        ret = CreateAclTensor(btH, blockTablesShape, &blockTablesDeviceAddr, blockTablesDtype, &blockTables);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (blockTablesDtype == ACL_UINT16) {
                        std::vector<uint16_t> btH(blockTables_sz, 1);
                        ret = CreateAclTensor(btH, blockTablesShape, &blockTablesDeviceAddr, blockTablesDtype, &blockTables);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (blockTablesDtype == ACL_UINT32) {
                        std::vector<uint32_t> btH(blockTables_sz, 1);
                        ret = CreateAclTensor(btH, blockTablesShape, &blockTablesDeviceAddr, blockTablesDtype, &blockTables);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (blockTablesDtype == ACL_INT64) {
                        std::vector<int64_t> btH(blockTables_sz, 1);
                        ret = CreateAclTensor(btH, blockTablesShape, &blockTablesDeviceAddr, blockTablesDtype, &blockTables);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (blockTablesDtype == ACL_UINT64) {
                        std::vector<uint64_t> btH(blockTables_sz, 1);
                        ret = CreateAclTensor(btH, blockTablesShape, &blockTablesDeviceAddr, blockTablesDtype, &blockTables);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (blockTablesDtype == ACL_DOUBLE) {
                        std::vector<double> btH(blockTables_sz, 1);
                        ret = CreateAclTensor(btH, blockTablesShape, &blockTablesDeviceAddr, blockTablesDtype, &blockTables);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (blockTablesDtype == ACL_BOOL) {
                        std::vector<int8_t> btH(blockTables_sz, 1);
                        ret = CreateAclTensor(btH, blockTablesShape, &blockTablesDeviceAddr, blockTablesDtype, &blockTables);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (blockTablesDtype == ACL_BF16) {
                        std::vector<int16_t> btH(blockTables_sz, 1);
                        ret = CreateAclTensor(btH, blockTablesShape, &blockTablesDeviceAddr, blockTablesDtype, &blockTables);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (blockTablesDtype == ACL_COMPLEX64) {
                        std::vector<float> btH(blockTables_sz, 1);
                        ret = CreateAclTensor(btH, blockTablesShape, &blockTablesDeviceAddr, blockTablesDtype, &blockTables);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else if (blockTablesDtype == ACL_COMPLEX128) {
                        std::vector<double> btH(blockTables_sz, 1);
                        ret = CreateAclTensor(btH, blockTablesShape, &blockTablesDeviceAddr, blockTablesDtype, &blockTables);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    } else {
                        std::vector<float> btH(blockTables_sz, 1);
                        ret = CreateAclTensor(btH, blockTablesShape, &blockTablesDeviceAddr, blockTablesDtype, &blockTables);
                        CHECK_RET(ret == ACL_SUCCESS, goto ERROR);
                    }
                }
            } else {
                blockTables = nullptr;
            }

            printf("[FUZZ_LOG] plog_file: aclnnAdvanceStep_error_plog_case_%d.log\n", scenario);

            printf("[FUZZ_PARAM] inputTokens=%s, inputTokensShape=%s, sampledTokenIds=%s, sampledTokenIdsShape=%s, inputPositions=%s, inputPositionsShape=%s, seqLens=%s, seqLensShape=%s, slotMapping=%s, slotMappingShape=%s, blockTables=%s, blockTablesShape=%s, format=ND\n", GetDtypeName(inputTokensDtype), ShapeToStr(inputTokensShape).c_str(), GetDtypeName(sampledTokenIdsDtype), ShapeToStr(sampledTokenIdsShape).c_str(), GetDtypeName(inputPositionsDtype), ShapeToStr(inputPositionsShape).c_str(), GetDtypeName(seqLensDtype), ShapeToStr(seqLensShape).c_str(), GetDtypeName(slotMappingDtype), ShapeToStr(slotMappingShape).c_str(), GetDtypeName(blockTablesDtype), ShapeToStr(blockTablesShape).c_str());

            printf("[FUZZ_ERROR_POINT] case_%d: errorType=%s, errorTarget=%s, description=%s\n", scenario, errorTypeStr.c_str(), errorTargetStr.c_str(), scenarioDesc.c_str());

            // === Call GetWorkspaceSize ===
            int64_t vNumSeqs = (inputTokens != nullptr) ? (int64_t)inputTokensShape[0] : 2;
            int64_t vNumQueries = (sampledTokenIds != nullptr) ? (int64_t)sampledTokenIdsShape[0] : 1;
            int64_t vBlockSize = 2;
            ret = aclnnAdvanceStepGetWorkspaceSize(inputTokens, sampledTokenIds, inputPositions, seqLens, slotMapping, blockTables, vNumSeqs, vNumQueries, vBlockSize, &workspaceSize, &executor);
            caseRet = (int)ret;
            LOG_PRINT("aclnnAdvanceStepGetWorkspaceSize ret=%d, scenario=%d\n", ret, scenario);
            const char* rawErrMsg = aclGetRecentErrMsg();
            savedErrMsg = (rawErrMsg && rawErrMsg[0] != '\0') ? std::string(rawErrMsg) : "";
            LOG_PRINT("aclGetRecentErrMsg: %s\n", savedErrMsg.c_str());

            // === Error code verification (维测要求) ===
            printf("[FUZZ_VERIFY] case_%d: expected_code=%d, actual_code=%d, expected_name=%s, match=%s\n",
                       scenario, expectedErrorCode, caseRet, expectedErrorName.c_str(),
                       (caseRet == expectedErrorCode || (expectedErrorCode == 161001 && caseRet == 161002) || (expectedErrorCode == 161002 && caseRet == 161001)) ? "YES" : "NO");
            printf("[FUZZ_VERIFY] verify_requirement: %s\n", verifyRequirement.c_str());

            if (ret == 0) {
                if (workspaceSize > 8589934592ULL) {
                    LOG_PRINT("workspaceSize %llu exceeds 8G cap, skip phase-2 execution\n", (unsigned long long)workspaceSize);
                } else {
                if (workspaceSize > 0) {
                    ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
                    if (ret != ACL_SUCCESS) { caseRet = (int)ret; LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); goto ERROR; }
                }
                ret = aclnnAdvanceStep(workspaceAddr, workspaceSize, executor, stream);
                    if (ret != ACL_SUCCESS) { caseRet = (int)ret; LOG_PRINT("aclnnAdvanceStep execute failed : %d\n", ret); goto ERROR; }
                caseRet = (int)ret;
                LOG_PRINT("aclnnAdvanceStep execute success : %d\n", ret);
                ret = aclrtSynchronizeStream(stream);
                    if (ret != ACL_SUCCESS) { caseRet = (int)ret; LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); goto ERROR; }
                }
            }

        }
        catch (std::exception& e)
        {
            printf("catch exception %s\n", e.what());
        }

    ERROR:
        ret = aclrtSynchronizeStream(stream);
        printf("end \n");
        CheckAndLog(ret);

        fflush(stdout);
        RestoreStdout();

        {
            std::string fromPath = "./opfuzz/scenario_logs/scenario_" + std::to_string(fuzzi) + ".log";
            std::string toPath = "./opfuzz/scenario_logs/scenario_" + std::to_string(scenario) + ".log";
            rename(fromPath.c_str(), toPath.c_str());
        }

        std::string aclErrMsg = savedErrMsg.empty() ? "(no aclnn error msg)" : savedErrMsg;
        std::string errType;
        if (caseRet == -1) errType = "SETUP_ERROR";
        else if (expectedErrorCode == 0) errType = (caseRet == 0) ? "BASE_CHECK_OK" : "BASE_INVALID";
        else if (caseRet == 0) errType = "UNEXPECTED_SUCCESS";
        else if (caseRet == 161001 || caseRet == 161002) errType = "PARAM_ERROR";
        else if (caseRet > 561001 && caseRet < 561119) errType = "INNER_ERROR";
        else if (caseRet > 507000 && caseRet < 507901) errType = "RUNTIME_ERROR";
        else errType = "OTHER_ERROR";

        bool hasHint = (aclErrMsg.find("should") != std::string::npos ||
                        aclErrMsg.find("must") != std::string::npos ||
                        aclErrMsg.find("expected") != std::string::npos ||
                        aclErrMsg.find("supported") != std::string::npos ||
                        aclErrMsg.find("range") != std::string::npos ||
                        aclErrMsg.find("invalid") != std::string::npos ||
                        aclErrMsg.find("mismatch") != std::string::npos);
        std::string ezVerdict = hasHint ? "PASS" : "FAIL";

        printf("== Scenario %3d | ret=%-8d | %-16s | EZ:%-4s | %s ==\n", scenario, caseRet, errType.c_str(), ezVerdict.c_str(), scenarioDesc.c_str());
        printf("   aclnn error msg: %s\n", aclErrMsg.c_str());

        ScenarioResult sr;
        sr.scenarioId = scenario;
        sr.retCode = caseRet;
        sr.desc = scenarioDesc;
        sr.violation = violationDetail;
        sr.category = errorCategory;
        sr.errorType = errType;
        sr.aclErrMsg = aclErrMsg;
        g_allResults.push_back(sr);

        pid_t pid = getpid();
        std::string cmd = "grep -rinE 'ai(vec|core) error' /root/ascend/log/debug/plog/ ";
        printf("system return value %d\n", CheckResult(pid));

        if (inputTokens != nullptr) { aclDestroyTensor(inputTokens); inputTokens = nullptr; }
        if (sampledTokenIds != nullptr) { aclDestroyTensor(sampledTokenIds); sampledTokenIds = nullptr; }
        if (inputPositions != nullptr) { aclDestroyTensor(inputPositions); inputPositions = nullptr; }
        if (seqLens != nullptr) { aclDestroyTensor(seqLens); seqLens = nullptr; }
        if (slotMapping != nullptr) { aclDestroyTensor(slotMapping); slotMapping = nullptr; }
        if (blockTables != nullptr) { aclDestroyTensor(blockTables); blockTables = nullptr; }
        if (inputTokensDeviceAddr != nullptr) { aclrtFree(inputTokensDeviceAddr); inputTokensDeviceAddr = nullptr; }
        if (sampledTokenIdsDeviceAddr != nullptr) { aclrtFree(sampledTokenIdsDeviceAddr); sampledTokenIdsDeviceAddr = nullptr; }
        if (inputPositionsDeviceAddr != nullptr) { aclrtFree(inputPositionsDeviceAddr); inputPositionsDeviceAddr = nullptr; }
        if (seqLensDeviceAddr != nullptr) { aclrtFree(seqLensDeviceAddr); seqLensDeviceAddr = nullptr; }
        if (slotMappingDeviceAddr != nullptr) { aclrtFree(slotMappingDeviceAddr); slotMappingDeviceAddr = nullptr; }
        if (blockTablesDeviceAddr != nullptr) { aclrtFree(blockTablesDeviceAddr); blockTablesDeviceAddr = nullptr; }
        if (workspaceSize > 0) { aclrtFree(workspaceAddr); }
    }
    DT_FUZZ_END()

    GenerateFinalSummary();

    WriteDTLog("aclnnAdvanceStep", "aclnnAdvanceStep op fuzz total number:%u, para error number %u, execute error number %u \n",
               fuzz_num, aclnn_INVALID_Para, aclnn_Inner_Error + Runtime_Error);
    WriteDTLog("aclnnAdvanceStep", "aclnnAdvanceStep op fuzz end\n");
    aclrtDestroyStream(stream);
    aclrtDestroyContext(context);
    aclrtResetDevice(deviceId);
    aclFinalize();
    return 0;
}

extern "C" int aclnnAdvanceStep_error_fuzz(int argc, char *argv[]);
int aclnnAdvanceStep_error_fuzz(int argc, char *argv[])
{
    int fuzz_num = MAXFUZZNUMBER;
    if (argc < 2) {
        cout << "[ERROR] input arg num must be bigger than 1 " << endl;
        return -1;
    } else if (argc > 2 && argv[2] != nullptr) {
        fuzz_num = atoi(argv[2]);
    }
    if (argc == 4) { deviceId = atoi(argv[3]); }
    printf("fuzz_num %d \n", fuzz_num);
    printf("aclnnAdvanceStep_Normal :%d \n", aclnnAdvanceStep_Normal(fuzz_num));
    return 0;
}