aclnnFlashAttentionScore

Applicable Products

Function

• Description: In training scenarios, the FlashAttention algorithm is used to implement self-attention computation.

• Formula:

  The forward computation formula for attention is as follows:

  $$attention\_out = Dropout(Softmax(Mask(scale*(pse+query*key^T), atten\_mask)), keep\_prob)*value$$

Function Prototype

Each operator is divided into a two-phase API. You must first call the "aclnnFlashAttentionScoreGetWorkspaceSize" API to obtain the workspace size required for computation and the executor that contains the operator computation process, and then call the "aclnnFlashAttentionScore" API to execute the computation.

aclnnStatus aclnnFlashAttentionScoreGetWorkspaceSize(
  const aclTensor   *query,
  const aclTensor   *key,
  const aclTensor   *value,
  const aclTensor   *realShiftOptional,
  const aclTensor   *dropMaskOptional,
  const aclTensor   *paddingMaskOptional,
  const aclTensor   *attenMaskOptional,
  const aclIntArray *prefixOptional,
  double             scaleValue,
  double             keepProb,
  int64_t            preTokens,
  int64_t            nextTokens,
  int64_t            headNum,
  char              *inputLayout,
  int64_t            innerPrecise,
  int64_t            sparseMode,
  const aclTensor   *softmaxMaxOut,
  const aclTensor   *softmaxSumOut,
  const aclTensor   *softmaxOutOut,
  const aclTensor   *attentionOutOut,
  uint64_t          *workspaceSize,
  aclOpExecutor    **executor)

aclnnStatus aclnnFlashAttentionScore(
  void              *workspace,
  uint64_t           workspaceSize,
  aclOpExecutor     *executor,
  const aclrtStream  stream)

aclnnFlashAttentionScoreGetWorkspaceSize

• Parameters

  Parameter	Input/Output	Description	Instruction	Data Type	Data Format	Dimension (Shape)	Non-contiguous Tensor
  query	Input	Query in the formula.	The data type of query must be consistent with that of key and value.	FLOAT16、BFLOAT16、FLOAT32	ND	[BNSD], [BSND], [BSH], [SBH]	√
  key	Input	Key in the formula.	Data type consistent with that of query and value.	FLOAT16, BFLOAT16, FLOAT32	ND	[BNSD], [BSND], [BSH], [SBH]	√
  value	Input	value in the formula.	The data type is the same as that of query/key.	FLOAT16, BFLOAT16, FLOAT32	ND	[BNSD], [BSND], [BSH], [SBH]	√
  realShiftOptional	Optional input	pse in the formula.	Data type consistency with query.	FLOAT16, BFLOAT16, FLOAT32	ND	[B,N,Sq,Skv], [B,N,1,Skv], [1,N,Sq,Skv], [B,N,1024,Skv], [1,N,1024,Skv]	√
  dropMaskOptional	Optional input	Dropout in the formula.	-	UINT8	ND	0, 1	√
  paddingMaskOptional	Optional input	Reserved parameter, not used currently.	-	-	-	-	-
  attenMaskOptional	Optional input	atten_mask in the formula.	A value of 1 indicates that the corresponding position does not participate in the computation, and a value of 0 indicates that it participates in the computation.	BOOL、UINT8	ND	[B,N,Sq,Skv]、[B,1,Sq,Skv]、[1,1,Sq,Skv]、[Sq,Skv]	√
  prefixOptional	Optional input	Prefix sparse computation scenario. N value of each batch.	-	INT64	ND	0, 1	-
  scaleValue	Optional input	scale in the formula represents the scaling factor.	-	DOUBLE	-	-	-
  keepProb	Optional input.	Represents the proportion of 1s in dropMaskOptional.	Value range: (0, 1].	DOUBLE	-	-	-
  preTokens	Optional input	Used for sparse computation. Represents the left boundary of the sliding window.	-	INT64	-	-	-
  nextTokens	Optional input	Specifies the right boundary of the sliding window for sparse computation.	-	INT64	-	-	-
  headNum	Input	Represents the number of heads on a single card, that is, the length of the N-axis of the input query.	-	INT64	-	-	-
  inputLayout	Input	Represents the data layout format of the input query, key, and value.	Supports BSH, SBH, BSND, and BNSD.	String	-	-	-
  innerPrecise	Optional input	Improves precision.	Defaults to 0.	INT64	-	-	-
  sparseMode	Optional input	Represents the sparse mode.	Supports configuration values of 0, 1, 2, 3, 4, 5, and 6.	INT64	-	-	-
  softmaxMaxOut	Output	Stores the intermediate maximum result of the Softmax computation, used for backward computation.	-	FLOAT	ND	[B,N,Sq,8]	√
  softmaxSumOut	Output	Intermediate sum result of the Softmax computation, used for backward computation.	-	FLOAT	ND	[B,N,Sq,8]	√
  softmaxOutOut	Output	Reserved parameter, not used yet.	-	-	-	-	-
  attentionOutOut	Output	Final output of the formula.	The data type and shape type are consistent with query.	FLOAT16, BFLOAT16, FLOAT32	ND	[BNSD], [BSND], [BSH], [SBH]	√
  workspaceSize	Output	Returns the workspace size to be applied on the Device side.	-	-	-	-	-
  executor	Output	Returns the op executor, which contains the operator computation flow.	-	-	-	-	-

• 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, realShiftOptional, dropMaskOptional, paddingMaskOptional, attenMaskOptional, softmaxMaxOut, softmaxSumOut, softmaxOutOut, or attentionOutOut is not within the supported range.
  		The data format of query, key, value, realShiftOptional, dropMaskOptional, paddingMaskOptional, attenMaskOptional, softmaxMaxOut, softmaxSumOut, softmaxOutOut, or attentionOutOut is not within the supported range.

aclnnFlashAttentionScore

• Parameters

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

• Return Value

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

Constraints

• Deterministic computation:

  • aclnnFlashAttentionScore defaults to a deterministic implementation.

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

• The constraints on the inputs query, key, and value are as follows:

  • The data layout can be interpreted from multiple dimensions, where B (Batch) represents the batch size of input samples, S (Seq-Length) represents the sequence length of input samples, H (Head-Size) represents the size of the hidden layer, N (Head-Num) represents the number of heads, and D (Head-Dim) represents the smallest unit dimension of the hidden layer, satisfying D = H/N.

  • B: The batchsize must be equal.

  • D: Head-Dim must satisfy (qD == kD && kD >= vD).

  • inputLayout must be consistent.

  • The data type must be consistent with that of realShiftOptional.

• The shapes of the input key and value must be identical except for D.

• The N of the input query and the N of key/value may differ, but they must be in a proportional relationship, i.e., Nq/Nkv must be a non-zero integer, with Nq in the value range of 1 to 256. When Nq/Nkv > 1, it is GQA (grouped-query attention); when Nkv = 1, it is MQA (multi-query attention). Unless otherwise specified in this document, N refers to Nq.

• For constraints on data shapes, taking BSND and BNSD of inputLayout as examples (in BSH and SBH, H = N*D):

  • B: The value range is 1 to 2M. When prefixOptional is present, B supports a maximum of 2K.

  • N: The value range is 1 to 256.

  • S: value range is 1 to 1M.

  • D: value range is 1 to 768.

• realShiftOptional: If Sq is greater than 1024, Sq and Skv are of equal length for each batch, and the scenario is a lower triangular mask with sparseMode 0, 2, or 3, alibi position encoding compression can be enabled. In this case, only the last 1024 rows of the original PSE need to be input for memory optimization, i.e., alibi_compress = ori_pse[:, :, -1024:, :], as detailed below:

  • When the parameter differs for each batch, the shape is BNHSkv (H=1024).

  • When the parameter is the same for each batch, the shape is 1NHSkv (H=1024).

  • If this parameter is not used, nullptr can be passed in.

• innerPrecise: Currently, 0 and 1 are reserved configuration values, and 2 enables invalid row computation. Its function is to avoid precision loss caused by entire-row masks during computation, but this configuration degrades performance. If the operator can determine that an invalid row scenario exists, it automatically enables invalid row computation, for example, when sparseMode is 3 and Sq > Skv.

• The constraints of sparseMode are as follows:

  • When the shapes of all attenMaskOptional are smaller than 2048 and identical, it is recommended to use the default mode to reduce memory usage.

  • When configured as 0 or 4, ensure that attenMaskOptional is consistent with the ranges of preTokens and nextTokens.

  • When set to 1, 2, 3, 5, or 6, the user-configured preTokens and nextTokens do not take effect.

  • When set to 1, 2, 3, 4, 5, or 6, the correct attenMaskOptional must be passed in; otherwise, the computation result will be incorrect. When attenMaskOptional is None, the sparseMode, preTokens, and nextTokens parameters do not take effect and the computation defaults to full computation.

  • It is recommended to pass 0 when no specific configuration is required.

  • For detailed descriptions of different sparse modes, see Sparse Mode Description.

• In some scenarios, if the computation load is too large, the operator execution may time out (aicore error, with errorStr: timeout or trap error). In this case, axis splitting is recommended. Note: The computation load here is affected by parameters such as B, S, N, and D; larger values result in a heavier computation load.

• In the prefixOptional sparse computation scenario (sparseMode = 5 or 6), when Sq > Skv, the N value range of prefix is [0, Skv]; when Sq <= Skv, the N value range of prefix is [Skv-Sq, Skv].

• In the band scenario, preTokens and nextTokens must have an intersection.

• When Sq in realShiftOptional is greater than 1024 and the shape is BNHS or 1NHS, Sq and Skv must have the same length.

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_flash_attention_score.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;
}

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, aclrtContext* context, 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 = aclrtCreateContext(context, deviceId);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateContext failed. ERROR: %d\n", ret); return ret);
  ret = aclrtSetCurrentContext(*context);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetCurrentContext 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 memory on the device side.
  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 data from the host side to the device side.
  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);

  // Compute 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_ND,
                            shape.data(), shape.size(), *deviceAddr);
  return 0;
}

int main() {
  // 1. (Boilerplate) Initialize the device/stream. Refer to the acl API manual.
  // Fill in the deviceId based on your actual device.
  int32_t deviceId = 0;
  aclrtContext context;
  aclrtStream stream;
  auto ret = Init(deviceId, &context, &stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);

  // 2. Construct the input and output. Customize the construction based on the API interface.
  std::vector<int64_t> qShape = {256, 1, 128};
  std::vector<int64_t> kShape = {256, 1, 128};
  std::vector<int64_t> vShape = {256, 1, 128};
  std::vector<int64_t> attenmaskShape = {256, 256};

  std::vector<int64_t> attentionOutShape = {256, 1, 128};
  std::vector<int64_t> softmaxMaxShape = {1, 1, 256, 8};
  std::vector<int64_t> softmaxSumShape = {1, 1, 256, 8};

  void* qDeviceAddr = nullptr;
  void* kDeviceAddr = nullptr;
  void* vDeviceAddr = nullptr;
  void* attenmaskDeviceAddr = nullptr;
  void* attentionOutDeviceAddr = nullptr;
  void* softmaxMaxDeviceAddr = nullptr;
  void* softmaxSumDeviceAddr = nullptr;

  aclTensor* q = nullptr;
  aclTensor* k = nullptr;
  aclTensor* v = nullptr;
  aclTensor* pse = nullptr;
  aclTensor* dropMask = nullptr;
  aclTensor* padding = nullptr;
  aclTensor* attenmask = nullptr;
  aclTensor* attentionOut = nullptr;
  aclTensor* softmaxMax = nullptr;
  aclTensor* softmaxSum = nullptr;
  aclTensor* softmaxOut = nullptr;

  std::vector<float> qHostData(32768, 1);
  std::vector<float> kHostData(32768, 1);
  std::vector<float> vHostData(32768, 1);
  std::vector<uint8_t> attenmaskHostData(65536, 0);
  std::vector<float> attentionOutHostData(32768, 0);
  std::vector<float> softmaxMaxHostData(2048, 3.0);
  std::vector<float> softmaxSumHostData(2048, 3.0);

  ret = CreateAclTensor(qHostData, qShape, &qDeviceAddr, aclDataType::ACL_FLOAT, &q);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(kHostData, kShape, &kDeviceAddr, aclDataType::ACL_FLOAT, &k);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(vHostData, vShape, &vDeviceAddr, aclDataType::ACL_FLOAT, &v);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(attenmaskHostData, attenmaskShape, &attenmaskDeviceAddr, aclDataType::ACL_UINT8, &attenmask);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(attentionOutHostData, attentionOutShape, &attentionOutDeviceAddr, aclDataType::ACL_FLOAT, &attentionOut);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(softmaxMaxHostData, softmaxMaxShape, &softmaxMaxDeviceAddr, aclDataType::ACL_FLOAT, &softmaxMax);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(softmaxSumHostData, softmaxSumShape, &softmaxSumDeviceAddr, aclDataType::ACL_FLOAT, &softmaxSum);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  std::vector<int64_t> prefixOp = {0};
  aclIntArray *prefix = aclCreateIntArray(prefixOp.data(), 1);
  double scaleValue = 0.088388;
  double keepProb = 1;
  int64_t preTokens = 65536;
  int64_t nextTokens = 65536;
  int64_t headNum = 1;
  int64_t innerPrecise = 0;
  int64_t sparseMode = 0;

  char layOut[5] = {'S', 'B', 'H', 0};

  // 3. Call the CANN operator library API. Modify the API name to the specific one.
  uint64_t workspaceSize = 0;
  aclOpExecutor* executor;

  // Call the first-phase API of aclnnFlashAttentionScore.
  ret = aclnnFlashAttentionScoreGetWorkspaceSize(
            q, k, v, pse, dropMask, padding, attenmask, prefix, scaleValue,
            keepProb, preTokens, nextTokens, headNum, layOut, innerPrecise,
            sparseMode, softmaxMax, softmaxSum, softmaxOut, attentionOut, &workspaceSize, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnFlashAttentionScoreGetWorkspaceSize failed. ERROR: %d.\n[ERROR msg]%s", ret, aclGetRecentErrMsg()); 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[ERROR msg]%s", ret, aclGetRecentErrMsg()); return ret);
  }

  // Call the second-phase API of aclnnFlashAttentionScore.
  ret = aclnnFlashAttentionScore(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnFlashAttentionScore failed. ERROR: %d.\n[ERROR msg]%s", ret, aclGetRecentErrMsg()); 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[ERROR msg]%s", ret, aclGetRecentErrMsg()); return ret);

  // 5. Obtain the output value and copy the result from the device memory to the host memory. Modify based on the specific API definition.
  PrintOutResult(attentionOutShape, &attentionOutDeviceAddr);
  PrintOutResult(softmaxMaxShape, &softmaxMaxDeviceAddr);
  PrintOutResult(softmaxSumShape, &softmaxSumDeviceAddr);

  // 6. Release aclTensor and aclScalar. Modify based on the specific API definition.
  aclDestroyTensor(q);
  aclDestroyTensor(k);
  aclDestroyTensor(v);
  aclDestroyTensor(attenmask);
  aclDestroyTensor(attentionOut);
  aclDestroyTensor(softmaxMax);
  aclDestroyTensor(softmaxSum);

  // 7. Release device resources.
  aclrtFree(qDeviceAddr);
  aclrtFree(kDeviceAddr);
  aclrtFree(vDeviceAddr);
  aclrtFree(attenmaskDeviceAddr);
  aclrtFree(attentionOutDeviceAddr);
  aclrtFree(softmaxMaxDeviceAddr);
  aclrtFree(softmaxSumDeviceAddr);
  if (workspaceSize > 0) {
    aclrtFree(workspaceAddr);
  }
  aclrtDestroyStream(stream);
  aclrtDestroyContext(context);
  aclrtResetDevice(deviceId);
  aclFinalize();

  return 0;
}