[object Object][object Object][object Object]undefined[object Object]

• Description: Performs backpropagation of [object Object].

• Formula: R: batch × sequence

  H: hidden

  K: topK

  gradY: (R, H)

  expandedRowIdx: (R × K)

  expandedXOptional: (R × K, H) or (activeNum, H) or (expertNum, expertCapacity, H)

  scalesOptional: (R, K)

  expertIdxOptional: (R, K)

  biasOptional: (E, H)

  i : 0 ~ R × K - 1

  j : 0 ~ H

  (1) [object Object] is a null pointer:

  $$gradExpandedXOut[expandedRowIdx[i]][j] = gradY[i / K][j]$$

  (2) [object Object] is not a null pointer, and [object Object] is a null pointer:

  $$gradExpandedXOut[expandedRowIdx[i]][j] = gradY[i / K][j] × scalesOptional[i]$$ $$gradScalesOut[i] = sum(expandedXOptional[expandedRowIdx[i]][j] × gradY[i / K][j])$$

  (3) [object Object] and [object Object] are not null pointers:

  $$gradExpandedXOut[expandedRowIdx[i]][j] = gradY[i / K][j] × scalesOptional[i]$$ $$gradScalesOut[i] = sum((expandedXOptional[expandedRowIdx[i]][j] + biasOptional[expertIdxOptional[i]][j]) × gradY[i / K][j])$$

[object Object]

Each operator consists of . You must first call the [object Object] API to obtain the required workspace size and the executor that contains the operator computation flow, and then call the [object Object] API to execute the computation.

• [object Object]
• [object Object]

[object Object]

• Parameters:

  • [object Object] (aclTensor*, computation input): [object Object] on the device, derivative of the forward output y of [object Object]. It must be a 2D tensor with shape (R, H). The data type can be FLOAT16, BFLOAT16, or FLOAT32. The must be ND. Non-contiguous input is supported.
  • [object Object] (aclTensor*, computation input): [object Object] on the device, indicating the token index sorted by expert sequence. It must be a 1D tensor with shape (R × K). When [object Object] is set to a null pointer, [object Object] must be 1. When [object Object] is 0, the value range is [0, R* K – 1] and no duplicate index is allowed. When [object Object] is 1, the value range is [–1, expertNum * expertCapacity – 1] and no duplicate index except –1 is allowed. The data type is INT32. The must be ND. Non-contiguous input is supported.
  • [object Object] (aclTensor*, optional computation input): [object Object] on the device, indicating the features extended based on [object Object]. When [object Object] is not a null pointer, this parameter cannot be a null pointer either. When [object Object] is 0, it must be a 2D tensor. When [object Object] is greater than 0 and less than R × K, the shape is (activeNum, H). Otherwise, the shape is (R × K, H). When [object Object] is 1, it must be a 3D tensor with shape (expertNum, expertCapacity, H). The data type is the same as that of [object Object]. The supported data types are FLOAT16, BFLOAT16, and FLOAT32. The must be ND. Non-contiguous input is supported.
  • [object Object] (aclTensor*, optional computation input): [object Object] on the device, indicating the feature scaling. It must be a 2D tensor with shape (R, K). The data type can be FLOAT16, BFLOAT16, or FLOAT32. The must be ND. Non-contiguous input is supported.
    • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object] and [object Object]Atlas A3 training series products/Atlas A3 inference series products[object Object]: The data type must be the same as that of [object Object].
  • [object Object] (aclTensor*, optional computation input): [object Object] on the device, indicating the index of an expert that processes a feature. When [object Object] is a non-null pointer, the value cannot be a null pointer. It must be a 2D tensor with shape (R, K) and value range [0, E - 1], E ≥ 1, allowing duplicate indices. The data type is the same as that of [object Object] and can be INT32. The must be ND. Non-contiguous tensors are supported.
  • [object Object] (aclTensor*, optional computation input): [object Object] on the device, indicating the feature bias. It must be a 2D tensor with shape (E, H). The data type is the same as that of [object Object] and can be FLOAT16, BFLOAT16, or FLOAT32. The must be ND. Non-contiguous input is supported.
  • [object Object] (int64_t, computation input): int64 data type, indicating different scenarios. The value can be 0 or 1. 0 indicates the dropless scenario, where [object Object] and [object Object] are not verified. 1 indicates the drop scenario, where [object Object] and [object Object] need to be verified. The values exceeding or below [object Object] will be processed accordingly.
  • [object Object] (int64_t, computation input): int64 data type, indicating the maximum number of rows in [object Object]. When [object Object] is 0, this parameter takes effect only when [object Object] is greater than 0 and less than R × K. When [object Object] is 1, this parameter does not take effect.
  • [object Object] (int64_t, computation input): int64 data type, indicating the number of experts. When [object Object] is 0, this parameter does not take effect. When [object Object] is 1 and [object Object] is a non-null pointer, expertNum must be equal to E. When [object Object] is a null pointer, [object Object] must be greater than 0. Otherwise, an error is reported.
  • [object Object] (int64_t, computation input): int64 data type, indicating the number of rows that can be processed by each expert. When [object Object] is 0, this parameter does not take effect. When [object Object] is 1, [object Object] must be greater than 0. Otherwise, an error is reported.
  • [object Object] (aclTensor*, computation output): [object Object] on the device, indicating the derivative of the forward input [object Object] of [object Object]. When [object Object] is 0, it must be a 2D tensor. When [object Object] is greater than 0 and less than R × K, the shape is (activeNum, H). Otherwise, the shape is (R × K, H). When [object Object] is 1, it must be a 3D tensor. The shape is (expertNum, expertCapacity, H). The data type is the same as that of [object Object]. The data type can be FLOAT16, BFLOAT16, or FLOAT32. The must be ND. Non-contiguous output is not supported.
  • [object Object] (aclTensor*, computation output): [object Object] on the device, indicating the derivative of the forward input scales of [object Object]. The output is valid only when [object Object] is not a null pointer. The output must be a 2D tensor with shape (R, K). The data type is the same as that of [object Object] and can be FLOAT16, BFLOAT16, or FLOAT32. The must be ND. Non-contiguous input is not supported.
  • [object Object] (uint64_t*, output): size of the workspace to be allocated on the device.
  • [object Object] (aclOpExecutor**, output): operator executor, covering the operator computation process.

• Returns:

  [object Object] status code. For details, see .

  [object Object]

[object Object]

• Parameters:

  • [object Object] (void*, input): address of the workspace to be allocated on the device.
  • [object Object] (uint64_t, input): size of the workspace to be allocated on the device, which is obtained by calling [object Object].
  • [object Object] (aclOpExecutor*, input): operator executor, covering the operator computation process.
  • [object Object] (aclrtStream, input): stream for executing the task.

• Returns:

  [object Object] status code. For details, see .

[object Object]

• Deterministic computing:
  • [object Object] defaults to a deterministic implementation.

[object Object]

The following example is for reference only. For details, see .