昇腾社区首页
中文
注册

aclnnMoeTokenPermuteWithRoutingMap

支持的产品型号

  • Atlas A2 训练系列产品/Atlas 800I A2 推理产品/A200I A2 Box 异构组件
  • Atlas A3 训练系列产品/Atlas A3 推理系列产品

功能说明

  • 算子功能:MoE的permute计算,将token和expert的标签作为routingMap传入,根据routingMaps将tokens和可选probsOptional广播后排序
  • 计算公式: tokens_num 为routingMap的第0维大小,expert_num为routingMap的第1维大小 dropAndPad为falseexpertIndex=arrange(tokens_num).expand(expert_num,1)expertIndex=arrange(tokens\_num).expand(expert\_num,-1) sortedIndicesFirst=expertIndex.maskedselect(routingMap.T)sortedIndicesFirst=expertIndex.maskedselect(routingMap.T) sortedIndicesOut=argSort(sortedIndices)sortedIndicesOut=argSort(sortedIndices) 当rangeOptional[0]<=sortedIndices[i]<rangeOptional[1]时topK=numOutTokens//tokens_numtopK = numOutTokens // tokens\_num outToken=topKtokens_numoutToken = topK * tokens\_num permuteTokens[sortedIndices[i]range[0]]=tokens[i//topK]permuteTokens[sortedIndices[i]-range[0]]=tokens[i//topK] permuteProbsOutOptional=probsOptional.T.maskedselect(routingMap.T)permuteProbsOutOptional=probsOptional.T.maskedselect(routingMap.T) dropAndPad为truecapacity=numOutTokens//expert_numcapacity = numOutTokens // expert\_num outToken=capacityexpert_numoutToken = capacity * expert\_num sortedIndicesOut=argsort(routingMap.T,dim=1)[:,:capacity]sortedIndicesOut = argsort(routingMap.T,dim=-1)[:, :capacity] permutedTokensOut=tokens.indexselect(0,sortedindices)permutedTokensOut = tokens.index_select(0, sorted_indices) 如果probs不是nonerobs_T_1D=probsOptional.T.view(1)robs\_T\_1D = probsOptional.T.view(-1) indices_dim0=arange(num_experts)indices\_dim0 = arange(num\_experts) indices_dim1=sortedindices.view(expert_num,capacity)indices\_dim1 = sorted_indices.view(expert\_num, capacity) indices_1D=(indicesdim0tokens_num+indices_dim1).view(1)indices\_1D = (indices_dim0 * tokens\_num + indices\_dim1).view(-1) permuteProbsOutOptional=probs_T_1D.indexselect(0,indices1D)permuteProbsOutOptional = probs\_T\_1D.index_select(0, indices_1D)

函数原型

每个算子分为两段式接口,必须先调用 “aclnnMoeTokenPermuteWithRoutingMapGetWorkspaceSize”接口获取计算所需workspace大小以及包含了算子计算流程的执行器,再调用“aclnnMoeTokenPermuteWithRoutingMap”接口执行计算。

  • aclnnStatus aclnnMoeTokenPermuteWithRoutingMapGetWorkspaceSize(const aclTensor *tokens, const aclTensor *routingMap, const aclTensor *probsOptional, int64_t numOutTokens, bool dropAndPad, aclTensor *permuteTokensOut, aclTensor *permuteProbsOutOptional, aclTensor *sortedIndicesOut, uint64_t *workspaceSize, aclOpExecutor **executor)
  • aclnnStatus aclnnMoeTokenPermuteWithRoutingMap(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream)

aclnnMoeTokenPermuteWithRoutingMapGetWorkspaceSize

  • 参数说明:

    • tokens(aclTensor *,计算输入):Device侧的aclTensor,输入token,公式中的tokens,要求为一个维度为2D的Tensor,shape为 (tokens_num, hidden_size),数据类型支持BFLOAT16,FLOAT16,FLOAT,数据格式要求为ND。支持非连续的Tensor
    • routingMap(aclTensor *,计算输入):Device侧的aclTensor,公式中的routingMap,代表token到expert的映射关系,要求shape为一个2D的(tokens_num,experts_num),数据类型支持INT8、BOOL。当数据类型为INT8,取值支持0、1,当数据类型为bool,取值支持true、false,数据格式要求为ND。支持非连续的Tensor。非droppad模式要求每行中包含topK个true 或 1。
    • probsOptional(aclTensor *,计算输入):Device侧的aclTensor,可选输入probsOptional,公式中的probsOptional,要求元素个数与routingMap相同,当probsOptional为空时,可选输出permuteProbsOutOptional为空,数据类型支持BFLOAT16,FLOAT16,FLOAT。数据格式要求为ND。支持非连续的Tensor
    • numOutTokens(int64_t,计算输入):公式中的numOutTokens,用于计算公式中topK 和capacity 的有效输出token数。
    • dropAndPad(bool,计算输入):公式中的dropAndPad,表示是否开启dropAndPad模式。
    • permutedTokensOut(aclTensor *,计算输出):Device侧的aclTensor,公式中的permutedTokensOut,根据indices进行扩展并排序筛选过的tokens,要求是一个2D的Tensor,shape为(outToken , hidden_size),即公式中的outToken。数据类型同tokens,数据格式要求为ND。支持非连续的Tensor
    • sortedIndicesOut(aclTensor *,计算输出):Device侧的aclTensor,公式中的sortedIndicesOut,permute_tokens和tokens的映射关系, 要求是一个1D的Tensor,Shape为(outToken),即公式中的outToken,数据类型支持INT32,数据格式要求为ND。支持非连续的Tensor
    • permuteProbsOutOptional(aclTensor *,计算输出):Device侧的aclTensor,公式中的permuteProbsOutOptional,根据indices进行排序并筛选过的probsOptional,Shape为(outToken),即公式中的outToken,数据类型同probsOptional,数据格式要求为ND。支持非连续的Tensor
    • workspaceSize(uint64_t *,出参):返回用户需要在Device侧申请的workspace大小。
    • executor(aclOpExecutor **,出参):返回op执行器,包含了算子计算流程。
  • 返回值:

    aclnnStatus:返回状态码,具体参见aclnn返回码

    第一段接口完成入参校验,出现以下场景时报错:
    161001(ACLNN_ERR_PARAM_NULLPTR): 1. 输入和输出的Tensor是空指针。
    161002(ACLNN_ERR_PARAM_INVALID): 1. 输入和输出的数据类型不在支持的范围内。
                                     2. 输入输出的shape不符合要求 
    561002(ACLNN_ERR_INNER_TILING_ERROR): 1. topkNum > 512

aclnnMoeTokenPermuteWithRoutingMap

  • 参数说明:

    • workspace(void*,入参):在Device侧申请的workspace内存地址。
    • workspaceSize(uint64_t,入参):在Device侧申请的workspace大小,由第一段接口aclnnMoeTokenUnpermuteWithEpGradGetWorkspaceSize获取。
    • executor(aclOpExecutor*,入参):op执行器,包含了算子计算流程。
    • stream(aclrtStream,入参):指定执行任务的AscendCL stream流。
  • 返回值:

    返回aclnnStatus状态码,具体参见aclnn返回码

约束说明

tokens_num和experts_num要求小于16777215,pad模式为false时routingMap 中 每行为1或true的个数固定且小于512

调用示例

示例代码如下,仅供参考,具体编译和执行过程请参考编译与运行样例

#include "acl/acl.h"
#include "aclnnop/aclnn_moe_token_permute_with_routing_map.h"
#include <iostream>
#include <vector>

#define CHECK_RET(cond, return_expr) \
  do {                               \
    if (!(cond)) {                   \
      return_expr;                   \
    }                                \
  } while (0)
#define LOG_PRINT(message, ...)     \
  do {                              \
    printf(message, ##__VA_ARGS__); \
  } while (0)
int64_t GetShapeSize(const std::vector<int64_t>& shape) {
    int64_t shape_size = 1;
    for (auto i : shape) {
        shape_size *= i;
    }
    return shape_size;
}
int Init(int32_t deviceId, aclrtStream* stream) {
    // 固定写法,AscendCL初始化
    auto ret = aclInit(nullptr);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclInit failed. ERROR: %d\n", ret); return ret);
    ret = aclrtSetDevice(deviceId);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
    ret = aclrtCreateStream(stream);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
    return 0;
}
template <typename T>
int CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
                    aclDataType dataType, aclTensor** tensor) {
    auto size = GetShapeSize(shape) * sizeof(T);
    // 调用aclrtMalloc申请device侧内存
    auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); return ret);
    // 调用aclrtMemcpy将host侧数据拷贝到device侧内存上
    ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMemcpy failed. ERROR: %d\n", ret); return ret);
    // 计算连续tensor的strides
    std::vector<int64_t> strides(shape.size(), 1);
    for (int64_t i = shape.size() - 2; i >= 0; i--) {
        strides[i] = shape[i + 1] * strides[i + 1];
    }
    // 调用aclCreateTensor接口创建aclTensor
    *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
                              shape.data(), shape.size(), *deviceAddr);
    return 0;
}
int main() {
    // 1. 固定写法,device/stream初始化, 参考acl对外接口列表
    // 根据自己的实际device填写deviceId
    int32_t deviceId = 0;
    aclrtStream stream;
    auto ret = Init(deviceId, &stream);
    // check根据自己的需要处理
    CHECK_RET(ret == 0, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);
    // 2. 构造输入与输出,需要根据API的接口定义构造
    std::vector<int64_t> xShape = {3, 4};
    std::vector<int64_t> idxShape = {3, 2};
    std::vector<int64_t> expandedXOutShape = {6, 4};
    std::vector<int64_t> idxOutShape = {6};
    void* xDeviceAddr = nullptr;
    void* indicesDeviceAddr = nullptr;
    void* expandedXOutDeviceAddr = nullptr;
    void* sortedIndicesOutDeviceAddr = nullptr;
    aclTensor* x = nullptr;
    aclTensor* indices = nullptr;
    int64_t numTokenOut = 6;
    bool padMode = false;

    aclTensor* expandedXOut = nullptr;
    aclTensor* sortedIndicesOut = nullptr;
    std::vector<float> xHostData = {0.1, 0.1, 0.1, 0.1, 0.2, 0.2, 0.2, 0.2, 0.3, 0.3, 0.3, 0.3};
    std::vector<int> indicesHostData = {1, 1, 1, 1, 1, 1};
    std::vector<float> expandedXOutHostData = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
    std::vector<int> sortedIndicesOutHostData = {0, 0, 0, 0, 0, 0};
    // 创建self aclTensor
    ret = CreateAclTensor(xHostData, xShape, &xDeviceAddr, aclDataType::ACL_BF16, &x);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    ret = CreateAclTensor(indicesHostData, idxShape, &indicesDeviceAddr, aclDataType::ACL_BOOL, &indices);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // 创建out aclTensor
    ret = CreateAclTensor(expandedXOutHostData, expandedXOutShape, &expandedXOutDeviceAddr, aclDataType::ACL_BF16, &expandedXOut);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    ret = CreateAclTensor(sortedIndicesOutHostData, idxOutShape, &sortedIndicesOutDeviceAddr, aclDataType::ACL_INT32, &sortedIndicesOut);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // 3. 调用CANN算子库API,需要修改为具体的API
    uint64_t workspaceSize = 0;
    aclOpExecutor* executor;
    // 调用aclnnMoeTokenPermuteWithRoutingMap第一段接口
    ret = aclnnMoeTokenPermuteWithRoutingMapGetWorkspaceSize(x, indices, nullptr, numTokenOut, padMode, expandedXOut, nullptr, sortedIndicesOut, &workspaceSize, &executor);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnMoeTokenPermuteWithRoutingMapGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
    // 根据第一段接口计算出的workspaceSize申请device内存
    void* workspaceAddr = nullptr;
    if (workspaceSize > 0) {
        ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
        CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret); return ret;);
    }
    ret = aclnnMoeTokenPermuteWithRoutingMap(workspaceAddr, workspaceSize, executor, stream);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnMoeTokenPermuteWithRoutingMapfailed. ERROR: %d\n", ret); return ret);
    // 4. 固定写法,同步等待任务执行结束
    ret = aclrtSynchronizeStream(stream);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);
    // 5. 获取输出的值,将device侧内存上的结果拷贝至host侧,需要根据具体API的接口定义修改
    auto expandedXSize = GetShapeSize(expandedXOutShape);
    std::vector<float> expandedXData(expandedXSize, 0);
    ret = aclrtMemcpy(expandedXData.data(), expandedXData.size() * sizeof(expandedXData[0]), expandedXOutDeviceAddr, expandedXSize * sizeof(float),
                      ACL_MEMCPY_DEVICE_TO_HOST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
    for (int64_t i = 0; i < expandedXSize; i++) {
        LOG_PRINT("expandedXData[%ld] is: %f\n", i, expandedXData[i]);
    }
    auto sortedIndicesSize = GetShapeSize(idxOutShape);
    std::vector<int> sortedIndicesData(sortedIndicesSize, 0);
    ret = aclrtMemcpy(sortedIndicesData.data(), sortedIndicesData.size() * sizeof(sortedIndicesData[0]), sortedIndicesOutDeviceAddr, sortedIndicesSize * sizeof(int32_t),
                      ACL_MEMCPY_DEVICE_TO_HOST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
    for (int64_t i = 0; i < sortedIndicesSize; i++) {
        LOG_PRINT("sortedIndicesData[%ld] is: %d\n", i, sortedIndicesData[i]);
    }
    // 6. 释放aclTensor和aclScalar,需要根据具体API的接口定义修改
    aclDestroyTensor(x);
    aclDestroyTensor(indices);
    aclDestroyTensor(expandedXOut);
    aclDestroyTensor(sortedIndicesOut);

    // 7. 释放device资源,需要根据具体API的接口定义修改
    aclrtFree(xDeviceAddr);
    aclrtFree(indicesDeviceAddr);
    aclrtFree(expandedXOutDeviceAddr);
    aclrtFree(sortedIndicesOutDeviceAddr);
    if (workspaceSize > 0) {
      aclrtFree(workspaceAddr);
    }
    aclrtDestroyStream(stream);
    aclrtResetDevice(deviceId);
    aclFinalize();
    return 0;
}