aclnnNsaSelectedAttentionGrad

Applicable Products

Function

• Description: Rearranges data of a selected size (selectedBlockSize) from key and value based on topkIndices, and then computes the backward output of attention in training scenarios.

• Formula:

  Rearranges data of selectedBlockCount blocks, each of size selectedBlockSize, from key and value based on the input topkIndices. The formula is as follows:

  $$selectedKey = Gather(key, topkIndices[i]),0<=i<selectedBlockCount \\ selectedValue = Gather(value, topkIndices[i]),0<=i<selectedBlockCount$$

  Then, the backward computation of the attention mechanism is performed. The formula is:

  $$dV=P^TdY$$ $$dQ=\frac{((dS)*K)}{\sqrt{d}}$$ $$dK=\frac{((dS)^T*Q)}{\sqrt{d}}$$

Function Prototype

Each operator uses a two-phase API. First call the "aclnnNsaSelectedAttentionGradGetWorkspaceSize" API to obtain the workspace size required for computation and the executor that contains the operator computation process, and then call the "aclnnNsaSelectedAttentionGrad" API to perform the computation.

aclnnStatus aclnnNsaSelectedAttentionGradGetWorkspaceSize(
  const aclTensor   *query,
  const aclTensor   *key,
  const aclTensor   *value,
  const aclTensor   *attentionOut,
  const aclTensor   *attentionOutGrad,
  const aclTensor   *softmaxMax,
  const aclTensor   *softmaxSum,
  const aclTensor   *topkIndices,
  const aclIntArray *actualSeqQLenOptional,
  const aclIntArray *actualSeqKvLenOptional,
  const aclTensor   *attenMaskOptional,
  double             scaleValue,
  int64_t            selectedBlockSize,
  int64_t            selectedBlockCount,
  int64_t            headNum,
  char              *inputLayout,
  int64_t            sparseMode,
  const aclTensor         *dqOut,
  const aclTensor         *dkOut,
  const aclTensor         *dvOut,
  uint64_t          *workspaceSize,
  aclOpExecutor    **executor)

aclnnStatus aclnnNsaSelectedAttentionGrad(
  void          *workspace,
  uint64_t       workspaceSize,
  aclOpExecutor *executor,
  aclrtStream    stream);

aclnnNsaSelectedAttentionGradGetWorkspaceSize

• Parameters

  Parameter	Input/Output	Description	Instruction	Data Type	Data Format	Dimension (Shape)	Non-contiguous Tensor
  query	Input	Query in the formula.	-	BFLOAT16, FLOAT16	ND	3	√
  key	Input	The key in the formula.	-	BFLOAT16, FLOAT16	ND	3-4	√
  value	Input	value in the formula.	-	BFLOAT16, FLOAT16	ND	3	√
  attentionOut	Input.	Final output of the attention forward computation.	-	BFLOAT16、FLOAT16	ND	3	√
  attentionOutGrad	Input	dY in the formula.	-	BFLOAT16, FLOAT16	ND	3	√
  softmaxMax	Input	Max intermediate result of the Softmax computation.	Used for backward computation.	FLOAT	ND	3	√
  softmaxSum	Input	Intermediate sum result of the softmax computation.	Used for backward computation.	FLOAT	ND	3	√
  topkIndices	Input	topk_indices in the formula.	-	INT32	ND	3	√
  attenMaskOptional	Input	atten_mask in the formula.	-	BOOL, UINT8	ND	2	√
  actualSeqQLenOptional	Input	The cumulative sum length of S for each batch of query.	-	INT64	ND	1	-
  actualSeqKvLenOptional	Input	Cumulative sum of S lengths for each batch of key/value.	-	INT64	ND	1	-
  scaleValue	Input	Scaling factor scale.	Generally D^-0.5.	DOUBLE	-	-	-
  headNum	Input	Number of heads.	-	INT64	-	-	-
  inputLayout	Input.	Data layout of query, key, and value.	Currently, only TND is supported.	String	-	-	-
  selectedBlockSize	Input	Length of each block.	-	INT64	-	-	-
  selectedBlockCount	Input	Number of select blocks.	-	INT64	-	-	-
  sparseMode	Input	Sparse mode.	Supports 0 or 2.	INT64	-	-	-
  dqOut	Output	dQ in the formula, representing the gradient of query.	-	FLOAT16, BFLOAT16, FLOAT32	ND	[BNSD], [BSND], [BSH], [SBH]	√
  dkOut	Output	In the formula, dK represents the gradient of keyIn.	-	FLOAT16、BFLOAT16、FLOAT32	ND	[BNSD]、[BSND]、[BSH]、[SBH]	√
  dvOut	Output	dV in the formula, representing the gradient of value.	-	FLOAT16, BFLOAT16, FLOAT32	ND	[BNSD], [BSND], [BSH], [SBH]	√
  workspaceSize	Output	Returns the workspace size to be applied for on the Device side.	-	-	-	-	-
  executor	Output	Returns the operator executor, which contains the operator computation process.	-	-	-	-	-

• Return Value

  aclnnStatus: return code. For details, see aclnn Return Code.

  The first-phase API performs input parameter validation and returns an error in the following scenarios:

  Return Value	Error Code	Description
  ACLNN_ERR_PARAM_NULLPTR	161001	The input parameter is a required input, output, or required attribute, and is a null pointer.
  ACLNN_ERR_PARAM_INVALID	161002	The data type of query, key, value, attentionOut, attentionOutGrad, pseShiftOptional, dropMaskOptional, paddingMaskOptional, attenMaskOptional, softmaxMaxOptional, softmaxSumOptional, softmaxInOptional, dqOut, dkOut, or dvOut is not within the supported range.
  		The data format of query, key, value, attentionOut, attentionOutGrad, pseShiftOptional, dropMaskOptional, paddingMaskOptional, attenMaskOptional, softmaxMaxOptional, softmaxSumOptional, softmaxInOptional, dqOut, dkOut, or dvOut is not within the supported range.

aclnnNsaSelectedAttentionGrad

• Parameters

  Parameter	Input/Output	Description
  workspace	Input	Workspace memory address applied for on the Device side.
  workspaceSize	Input	Workspace size applied for on the Device side, obtained from the first-phase API aclnnNsaSelectedAttentionGradGetWorkspaceSize.
  executor	Input.	Operator executor, which contains the operator computation flow.
  stream	Input.	Specifies the stream for task execution.

• Return Value

  aclnnStatus: return code. For details, see aclnn Return Code.

Constraints

• Deterministic computation:

  • aclnnNsaSelectedAttentionGrad defaults to a non-deterministic implementation. Deterministic computation can be enabled by calling aclrtCtxSetSysParamOpt.

• When this API is used with PyTorch, ensure that the versions of CANN-related packages and PyTorch-related packages match.

• The B (batchsize) of the input query, key, value, attentionOut, and attentionOutGrad must be equal.

• The N (numHead) of the input key and value must be consistent.

• The N (numHead) of the input query, attentionOut, and attentionOutGrad must be consistent.

• The D (HeadDim) of the input value, attentionOut, and attentionOutGrad must be consistent.

• The inputLayout of the input query, key, value, attentionOut, and attentionOutGrad must be consistent.

• For constraints on data shape, take TND of inputLayout as an example, where:

  • T1: Value range: 1 to 2M. T1 represents the sum of S across all batches of query.

  • T2: Value range: 1 to 2M. T2 represents the sum of S across all batches of key and value.

  • B: Value range: 1 to 2M.

  • N1: Value range: 1 to 128. Represents the headNum of query. N1 must be an integer multiple of N2.

  • N2: Value range: 1 to 128. Represents the headNum of key and value.

  • G: value range is 1 to 32. G = N1 / N2

  • S: value range is 1 to 128K. For key and value, S must be greater than or equal to selectedBlockSize * selectedBlockCount, and must be an integer multiple of selectedBlockSize.

  • D: value range is 192 or 128. Supports unequal D (HeadDim) for K and V.

  • selectedBlockSize supports ≤ 128 and must be an integer multiple of 16.

  • selectedBlockCount: supports [1, 128]. The total selected size selectedBlockCount * selectedBlockSize < 128*64 (8K).

• When Layout is TND, S2 of each batch must be greater than the total selected size selectedBlockCount * selectedBlockSize.

• Constraints on the shape of the softmaxMax and softmaxSum parameters: [T1, N1, 8].

• Constraints on the shape of the topkIndices parameter: [T1, N2, selectedBlockCount].

Example

The following provides example code for reference only. For details about compilation and running, see Compile and Run Samples.

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

#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;
}

int Init(int32_t deviceId, aclrtStream* stream) {
    // Boilerplate: resource initialization.
    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);
    // Call aclrtMalloc to apply for device-side memory.
    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);

    // Call aclrtMemcpy to copy host-side data to device-side memory.
    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);

    // Calculate the strides of a contiguous tensor.
    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];
    }

    // Call the aclCreateTensor API to create an aclTensor.
    *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_NCHW,
                            shape.data(), shape.size(), *deviceAddr);
    return 0;
}

int main() {
    // 1. (Boilerplate) Initialize the device/stream. See the AscendCL external API list.
    // Fill in the deviceId based on the actual device.
    int32_t deviceId = 0;
    aclrtStream stream;
    auto ret = Init(deviceId, &stream);
    // Handle the check based on your requirements.
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);

    // 2. Construct the inputs and outputs. Customize the construction based on the API interface.
    int64_t b = 1;
    int64_t s1 = 1;
    int64_t s2 = 1024;
    int64_t t1 = b * s1;
    int64_t t2 = b * s2;
    int64_t n1 = 1;
    int64_t n2 = 1;
    int64_t d = 192;

    int64_t sparseMode = 0;
    char inputLayout[5] = {'T', 'N', 'D', 0};
    double scaleValue = 1.0f;
    int64_t selectedBlockSize = 64;
    int64_t selectedBlockCount = 16;
    int32_t headNum = n1;

    std::vector<int64_t> queryShape = {t1, n1, d};
    std::vector<int64_t> keyShape = {t2, n2, d};
    std::vector<int64_t> valueShape = {t2, n2, d};
    std::vector<int64_t> attentionOutShape = {t1, n1, d};
    std::vector<int64_t> attentionOutGradShape = {t1, n1, d};
    std::vector<int64_t> softmaxMaxShape = {t1, n1, 8};
    std::vector<int64_t> softmaxSumShape = {t1, n1, 8};
    std::vector<int64_t> topkIndicesShape = {t1, n2, selectedBlockCount};
    std::vector<int64_t> actualSeqQLenOptionalShape = {b};
    std::vector<int64_t> actualSeqKvLenOptionalShape = {b};
    std::vector<int64_t> dqOutShape = {t1, n1, d};
    std::vector<int64_t> dkOutShape = {t2, n2, d};
    std::vector<int64_t> dvOutShape = {t2, n2, d};

    void* queryDeviceAddr = nullptr;
    void* keyDeviceAddr = nullptr;
    void* valueDeviceAddr = nullptr;
    void* attentionOutDeviceAddr = nullptr;
    void* attentionOutGradDeviceAddr = nullptr;
    void* softmaxMaxDeviceAddr = nullptr;
    void* softmaxSumDeviceAddr = nullptr;
    void* topkIndicesDeviceAddr = nullptr;
    void* dqOutDeviceAddr = nullptr;
    void* dkOutDeviceAddr = nullptr;
    void* dvOutDeviceAddr = nullptr;

    aclTensor* query = nullptr;
    aclTensor* key = nullptr;
    aclTensor* value = nullptr;
    aclTensor* attentionOut = nullptr;
    aclTensor* attentionOutGrad = nullptr;
    aclTensor* softmaxMax = nullptr;
    aclTensor* softmaxSum = nullptr;
    aclTensor* topkIndices = nullptr;
    aclTensor* dqOut = nullptr;
    aclTensor* dkOut = nullptr;
    aclTensor* dvOut = nullptr;

    std::vector<aclFloat16> queryHostData(GetShapeSize(queryShape), 2);
    std::vector<aclFloat16> keyHostData(GetShapeSize(keyShape), 2);
    std::vector<aclFloat16> valueHostData(GetShapeSize(valueShape), 2);
    std::vector<aclFloat16> attentionOutHostData(GetShapeSize(attentionOutShape), 2);
    std::vector<aclFloat16> attentionOutGradHostData(GetShapeSize(attentionOutGradShape), 2);
    std::vector<float> softmaxMaxHostData(GetShapeSize(softmaxMaxShape), 2);
    std::vector<float> softmaxSumHostData(GetShapeSize(softmaxSumShape), 2);
    std::vector<int32_t> topkIndicesHostData(GetShapeSize(topkIndicesShape), 1);
    std::vector<aclFloat16> dqOutHostData(GetShapeSize(dqOutShape), 2);
    std::vector<aclFloat16> dkOutHostData(GetShapeSize(dkOutShape), 2);
    std::vector<aclFloat16> dvOutHostData(GetShapeSize(dvOutShape), 2);

    for (int32_t i = 0; i < topkIndicesHostData.size(); i++) {
        topkIndicesHostData[i] = i;
    }

    // Create the query aclTensor.
    ret = CreateAclTensor(queryHostData, queryShape, &queryDeviceAddr, aclDataType::ACL_FLOAT16, &query);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the key aclTensor.
    ret = CreateAclTensor(keyHostData, keyShape, &keyDeviceAddr, aclDataType::ACL_FLOAT16, &key);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create value aclTensor.
    ret = CreateAclTensor(valueHostData, valueShape, &valueDeviceAddr, aclDataType::ACL_FLOAT16, &value);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create attentionOut aclTensor.
    ret = CreateAclTensor(attentionOutHostData, attentionOutShape, &attentionOutDeviceAddr, aclDataType::ACL_FLOAT16, &attentionOut);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create attentionOutGrad aclTensor.
    ret = CreateAclTensor(attentionOutGradHostData, attentionOutGradShape, &attentionOutGradDeviceAddr, aclDataType::ACL_FLOAT16, &attentionOutGrad);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create softmaxMax aclTensor.
    ret = CreateAclTensor(softmaxMaxHostData, softmaxMaxShape, &softmaxMaxDeviceAddr, aclDataType::ACL_FLOAT, &softmaxMax);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create softmaxSum aclTensor.
    ret = CreateAclTensor(softmaxSumHostData, softmaxSumShape, &softmaxSumDeviceAddr, aclDataType::ACL_FLOAT, &softmaxSum);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the topkIndices aclTensor.
    ret = CreateAclTensor(topkIndicesHostData, topkIndicesShape, &topkIndicesDeviceAddr, aclDataType::ACL_INT32, &topkIndices);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    int64_t tempQ[1] = {1};
    int64_t tempK[1] = {1024};
    aclIntArray* actualSeqQLenOptional = aclCreateIntArray(tempQ, static_cast<uint64_t>(1));
    aclIntArray* actualSeqKvLenOptional = aclCreateIntArray(tempK, static_cast<uint64_t>(1));
    // Create the dq aclTensor.
    ret = CreateAclTensor(dqOutHostData, dqOutShape, &dqOutDeviceAddr, aclDataType::ACL_FLOAT16, &dqOut);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the dk aclTensor.
    ret = CreateAclTensor(dkOutHostData, dkOutShape, &dkOutDeviceAddr, aclDataType::ACL_FLOAT16, &dkOut);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the dv aclTensor.
    ret = CreateAclTensor(dvOutHostData, dvOutShape, &dvOutDeviceAddr, aclDataType::ACL_FLOAT16, &dvOut);
    CHECK_RET(ret == ACL_SUCCESS, return ret);

    // Example of calling the aclnnNsaSelectedAttentionGrad API.
    // 3. Call the CANN operator library API. Replace with the specific API name.
    uint64_t workspaceSize = 0;
    aclOpExecutor* executor;
    // Call the first-phase API of aclnnNsaSelectedAttentionGrad.
    ret = aclnnNsaSelectedAttentionGradGetWorkspaceSize(query, key, value, attentionOut, attentionOutGrad, softmaxMax,
                                                        softmaxSum, topkIndices, actualSeqQLenOptional,
                                                        actualSeqKvLenOptional, nullptr, scaleValue, selectedBlockSize,
                                                        selectedBlockCount, headNum, inputLayout, sparseMode,
                                                        dqOut, dkOut, dvOut, &workspaceSize, &executor);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnNsaSelectedAttentionGradGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
    // Apply for device memory based on the workspaceSize calculated by the first-phase API.
    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);
    }
    // Call the second-phase API of aclnnNsaSelectedAttentionGrad.
    ret = aclnnNsaSelectedAttentionGrad(workspaceAddr, workspaceSize, executor, stream);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnNsaSelectedAttentionGrad failed. ERROR: %d\n", ret); return ret);

    // 4. (Boilerplate) Synchronize and wait for task execution to complete.
    ret = aclrtSynchronizeStream(stream);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);

    // 5. Obtain the output value and copy the result from the Device side memory to the Host side. Modify based on the specific API definition.
    auto dqSize = GetShapeSize(dqOutShape);
    std::vector<aclFloat16> dqResultData(dqSize, 0);
    ret = aclrtMemcpy(dqResultData.data(), dqResultData.size() * sizeof(dqResultData[0]), dqOutDeviceAddr,
                      dqSize * sizeof(dqResultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy out result dq from device to host failed. ERROR: %d\n", ret); return ret);
    for (int64_t i = 0; i < dqSize; i++) {
        LOG_PRINT("result dq[%ld] is: %f\n", i, dqResultData[i]);
    }

    auto dkSize = GetShapeSize(dkOutShape);
    std::vector<aclFloat16> dkResultData(dkSize, 0);
    ret = aclrtMemcpy(dkResultData.data(), dkResultData.size() * sizeof(dkResultData[0]), dkOutDeviceAddr,
                      dkSize * sizeof(dkResultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy out result dk from device to host failed. ERROR: %d\n", ret); return ret);
    for (int64_t i = 0; i < dkSize; i++) {
        LOG_PRINT("result dk[%ld] is: %f\n", i, dkResultData[i]);
    }

    auto dvSize = GetShapeSize(dvOutShape);
    std::vector<aclFloat16> dvResultData(dvSize, 0);
    ret = aclrtMemcpy(dvResultData.data(), dvResultData.size() * sizeof(dvResultData[0]), dvOutDeviceAddr,
                      dvSize * sizeof(dvResultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy out result dv from device to host failed. ERROR: %d\n", ret); return ret);
    for (int64_t i = 0; i < dvSize; i++) {
        LOG_PRINT("result dv[%ld] is: %f\n", i, dvResultData[i]);
    }

    // 6. Release aclTensor and aclScalar. Modify based on the specific API definition.
    aclDestroyTensor(query);
    aclDestroyTensor(key);
    aclDestroyTensor(value);
    aclDestroyTensor(attentionOut);
    aclDestroyTensor(attentionOutGrad);
    aclDestroyTensor(softmaxMax);
    aclDestroyTensor(softmaxSum);
    aclDestroyTensor(topkIndices);
    aclDestroyTensor(dqOut);
    aclDestroyTensor(dkOut);
    aclDestroyTensor(dvOut);
    aclDestroyIntArray(actualSeqQLenOptional);
    aclDestroyIntArray(actualSeqKvLenOptional);
    // 7. Release device resources. Modify based on the specific API definition.
    aclrtFree(queryDeviceAddr);
    aclrtFree(keyDeviceAddr);
    aclrtFree(valueDeviceAddr);
    aclrtFree(attentionOutDeviceAddr);
    aclrtFree(attentionOutGradDeviceAddr);
    aclrtFree(softmaxMaxDeviceAddr);
    aclrtFree(softmaxSumDeviceAddr);
    aclrtFree(topkIndicesDeviceAddr);
    aclrtFree(dqOutDeviceAddr);
    aclrtFree(dkOutDeviceAddr);
    aclrtFree(dvOutDeviceAddr);
    if (workspaceSize > 0) {
        aclrtFree(workspaceAddr);
    }
    aclrtDestroyStream(stream);
    aclrtResetDevice(deviceId);
    aclFinalize();
    return 0;
}