aclnnMoeTokenUnpermuteGrad
支持的产品型号
- Atlas A2训练系列产品/Atlas 800I A2推理产品。
接口原型
每个算子分为两段式接口,必须先调用“aclnnMoeTokenUnpermuteGradGetWorkspaceSize”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“aclnnMoeTokenUnpermuteGrad”接口执行计算。
aclnnStatus aclnnMoeTokenUnpermuteGradGetWorkspaceSize(const aclTensor *permutedTokens, const aclTensor *unpermutedTokensGrad, const aclTensor *sortedIndices, const aclTensor *probsOptional, bool paddedMode, const aclIntArray *restoreShape, aclTensor *permutedTokensGradOut, aclTensor *probsGradOut, uint64_t *workspaceSize, aclOpExecutor **executor)
aclnnStatus aclnnMoeTokenUnpermuteGrad(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream)
功能描述
- 算子功能:aclnnMoeTokenUnpermute的反向传播。
- 计算公式:
- probs非None:
- probs为None:
aclnnMoeTokenUnpermuteGradGetWorkspaceSize
参数说明:
- permutedTokens(aclTensor*,计算输入):Device侧的aclTensor,输入token,要求为一个维度为2D的Tensor,shape为 (tokens_num * topK_num, hidden_size),其中topK_num <= 512,数据类型支持BFLOAT16,数据格式要求为ND。支持非连续输入。
- unpermutedTokensGrad(aclTensor*,计算输入):Device侧的aclTensor,正向输出unpermutedTokens的梯度,要求为一个维度为2D的Tensor,shape为 (tokens_num, hidden_size),数据类型支持BFLOAT16,数据格式要求为ND。支持非连续输入。
- sortedIndices(aclTensor*,计算输入):Device侧的aclTensor,要求shape为一个1D的(tokens_num * topK_num,),数据类型支持INT32,数据格式要求为ND。索引取值范围[0, tokens_num * topK_num - 1]。支持非连续输入。
- probsOptional(aclTensor*,计算输入):Device侧的aclTensor,可选输入,要求shape为一个2D的(tokens_num, topK_num),数据类型支持BFLOAT16,数据格式要求为ND。当probs传时,topK_num等于probs第2维;当probs不传时,topK_num=1。支持非连续输入。
- paddedMode(bool, 计算输入):true表示开启paddedMode,false表示关闭paddedMode,paddedMode解释见restoreShape参数。目前仅支持false。
- restoreShape(aclIntArray*,计算输入):当paddedMode为true后生效,否则不会对其进行操作。当paddedMode为true以后,此为unpermutedTokens的shape。当前仅支持nullptr。
- permutedTokensGradOut(aclTensor*,计算输出):输入permutedTokens的梯度,要求是一个2D的Tensor,shape为(tokens_num * topK_num, hidden_size). 数据类型同permutedTokens,支持BFLOAT16,数据格式要求为ND。不支持非连续输出。
- probsGradOut(aclTensor*,计算输出):可选输出,输入probs的梯度,要求是一个2D的Tensor,shape为(tokens_num, topK_num). 数据类型同probsOptional,支持BFLOAT16,数据格式要求为ND。不支持非连续输出。
- workspaceSize(uint64_t*,出参):返回需要在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:
返回aclnnStatus状态码,具体参见aclnn返回码。
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。
aclnnMoeTokenUnpermuteGrad
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aclnnMoeTokenUnpermuteGradGetWorkspaceSize获取。
- executor(aclOpExecutor*,入参):op执行器,包含了算子计算流程。
- stream(const 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。
约束与限制
topK_num <= 512
调用示例
#include <iostream>
#include <vector>
#include "acl/acl.h"
#include "aclnnop/aclnn_moe_token_unpermute_grad.h"
#include <iostream>
#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 shapeSize = 1;
for (auto i : shape) {
shapeSize *= i;
}
return shapeSize;
}
void PrintOutResult(std::vector<int64_t> &shape, void **deviceAddr) {
auto size = GetShapeSize(shape);
std::vector<float> resultData(size, 0);
auto ret = aclrtMemcpy(
resultData.data(), resultData.size() * sizeof(resultData[0]), *deviceAddr,
size * sizeof(resultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
CHECK_RET(
ret == ACL_SUCCESS,
LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret);
return );
for (int64_t i = 0; i < size; i++) {
LOG_PRINT("mean result[%ld] is: %f\n", i, resultData[i]);
}
}
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AscendCL对外接口列表
// 根据自己的实际device填写deviceId
int32_t deviceId = 2;
aclrtStream stream;
auto ret = Init(deviceId, &stream);
CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("Init acl failed. ERROR: %d\n", ret);
return ret);
// 2. 构造输入与输出,需要根据API的接口自定义构造
std::vector<int64_t> permutedTokensShape = {3, 2};
std::vector<int64_t> unpermutedTokensGradShape = {1, 2};
std::vector<int64_t> probsShape = {1, 3};
std::vector<int64_t> sortedIndicesShape = {3};
std::vector<int64_t> permutedTokensGradShape = {3, 2};
std::vector<int64_t> probsGradShape = {1, 3};
void* permutedTokensDeviceAddr = nullptr;
void* unpermutedTokensGradDeviceAddr = nullptr;
void* probsDeviceAddr = nullptr;
void* sortedIndicesDeviceAddr = nullptr;
void* permutedTokensGradDeviceAddr = nullptr;
void* probsGradDeviceAddr = nullptr;
aclTensor* permutedTokens = nullptr;
aclTensor* unpermutedTokensGrad = nullptr;
aclTensor* probs = nullptr;
aclTensor* sortedIndices = nullptr;
bool paddedMode = false;
aclTensor *permutedTokensGrad = nullptr;
aclTensor *probsGrad = nullptr;
std::vector<float> permutedTokensHostData = {1, 1, 1, 1, 1, 1};
std::vector<float> unpermutedTokensGradHostData = {1, 1};
std::vector<float> probsHostData = {1, 1, 1};
std::vector<int> sortedIndicesHostData = {0, 1, 2};
std::vector<float> permutedTokensGradHostData = {0, 0, 0, 0, 0, 0};
std::vector<float> probsGradHostData = {0, 0, 0};
ret = CreateAclTensor(permutedTokensHostData, permutedTokensShape,
&permutedTokensDeviceAddr, aclDataType::ACL_BF16,
&permutedTokens);
CHECK_RET(ret == ACL_SUCCESS, return ret);
ret = CreateAclTensor(unpermutedTokensGradHostData, unpermutedTokensGradShape, &unpermutedTokensGradDeviceAddr,
aclDataType::ACL_BF16, &unpermutedTokensGrad);
CHECK_RET(ret == ACL_SUCCESS, return ret);
ret = CreateAclTensor(probsHostData, probsShape, &probsDeviceAddr,
aclDataType::ACL_BF16, &probs);
CHECK_RET(ret == ACL_SUCCESS, return ret);
ret = CreateAclTensor(sortedIndicesHostData, sortedIndicesShape, &sortedIndicesDeviceAddr,
aclDataType::ACL_INT32, &sortedIndices);
CHECK_RET(ret == ACL_SUCCESS, return ret);
ret = CreateAclTensor(permutedTokensGradHostData, permutedTokensGradShape, &permutedTokensGradDeviceAddr, aclDataType::ACL_BF16,
&permutedTokensGrad);
CHECK_RET(ret == ACL_SUCCESS, return ret);
ret = CreateAclTensor(probsGradHostData, probsGradShape, &probsGradDeviceAddr, aclDataType::ACL_BF16,
&probsGrad);
CHECK_RET(ret == ACL_SUCCESS, return ret);
// 3. 调用CANN算子库API,需要修改为具体的Api名称
uint64_t workspaceSize = 0;
aclOpExecutor *executor;
// 调用aclnnMoeTokenUnpermuteGrad第一段接口
ret = aclnnMoeTokenUnpermuteGradGetWorkspaceSize(permutedTokens, unpermutedTokensGrad, sortedIndices, probs, paddedMode, nullptr,
permutedTokensGrad, probsGrad, &workspaceSize, &executor);
CHECK_RET(
ret == ACL_SUCCESS,
LOG_PRINT("aclnnMoeTokenUnpermuteGradGetWorkspaceSize 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);
}
// 调用aclnnMoeTokenUnpermuteGrad第二段接口
ret = aclnnMoeTokenUnpermuteGrad(workspaceAddr, workspaceSize, executor, stream);
CHECK_RET(ret == ACL_SUCCESS,
LOG_PRINT("aclnnMoeTokenUnpermuteGrad failed. 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
PrintOutResult(permutedTokensGradShape, &permutedTokensGradDeviceAddr);
PrintOutResult(probsGradShape, &probsGradDeviceAddr);
// 6. 释放aclTensor和aclScalar,需要根据具体API的接口定义修改
aclDestroyTensor(permutedTokens);
aclDestroyTensor(unpermutedTokensGrad);
aclDestroyTensor(sortedIndices);
aclDestroyTensor(probs);
aclDestroyTensor(permutedTokensGrad);
aclDestroyTensor(probsGrad);
// 7. 释放device资源
aclrtFree(permutedTokensDeviceAddr);
aclrtFree(unpermutedTokensGradDeviceAddr);
aclrtFree(probsDeviceAddr);
aclrtFree(sortedIndicesDeviceAddr);
aclrtFree(permutedTokensGradDeviceAddr);
aclrtFree(probsGradDeviceAddr);
if (workspaceSize > 0) {
aclrtFree(workspaceAddr);
}
aclrtDestroyStream(stream);
aclrtResetDevice(deviceId);
aclFinalize();
return 0;
}