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

• Description: Performs the backpropagation of [object Object].
• Formula:$$permuteTokenId, outIndex= sortedIndices.sort(dim=-1)$$ $$capacity = permutedTokenOutputGrad.size(0) / numExperts$$
  • When [object Object] is not set to [object Object]:
  $$probsGradOutOptional = zeros(tokens_num, numExperts)$$

  • When [object Object] is set to [object Object]:

    $$probsGradOutOptional [sortedIndices[i], i/capacity] = permutedProbsOutputGradOptional[i]$$

  • When [object Object] is set to [object Object]:

    $$probsGradOutOptional = maskedscatter(probsGradOutOptional,routingMap,permutedProbsOutputGradOptional)$$

  • If [object Object] is set to [object Object]:

    $$tokensGradout= zeros(restoreShapeOptional, dtype=permutedTokens.dtype, device=permutedTokens.device)$$ $$tokensGradout[permuteTokenId[i]] += permutedTokens[outIndex[i]]$$

[object Object]

Each operator has calls. First, [object Object] is called to obtain the workspace size required for computation and the executor that contains the operator computation process. Then, [object Object] is called to perform computation.

• [object Object]
• [object Object]

[object Object]

• Parameters:

  • [object Object] (aclTensor*, computation input): aclTensor on the device. It is the gradient of the forward output [object Object], and the value is a 2D tensor. In drop/pad-less mode, the shape must be 2D with size (tokens_num * topK_num, hidden_size). In drop/pad mode, the shape must be a 2D with size (experts_num * capacity, hidden_size). [object Object] indicates the number of experts selected for each token, and [object Object] indicates the number of tokens selected by each expert. The data type can be BFLOAT16, FLOAT16, or FLOAT32. The must be ND. are supported, but empty tensors are not supported.
  • [object Object] (aclTensor*, computation input): aclTensor on the device. This input is optional. If it is not passed, [object Object] does not need to be computed. In drop/pad-less mode, the shape must be 1D with size (tokens_num * topK_num). In drop/pad mode, the shape must be 1D with size (experts_num * capacity). [object Object] indicates the number of experts selected for each token, and [object Object] indicates the number of tokens selected by each expert. The data type can be BFLOAT16, FLOAT16, or FLOAT32. The must be ND. are supported.
  • [object Object] (aclTensor*, input): aclTensor on the device. In drop/pad-less mode, the value must be a 1D shape with size (tokens_num * topK_num,). The data type can be INT32. The must be ND. The index value range is [0, tokens_num * topK_num – 1]. In drop/pad mode, the value must be a 1D tensor with shape (experts_num * capacity). The data type can be INT32. The must be ND. The index value range is [0, experts_num * capacity – 1]. are supported.
  • [object Object] (aclTensor*, computation input): aclTensor on the device, which indicates the mapping between tokens and experts, that is, [object Object] in the formula. The value must be a 2D shape with size (tokens_num, experts_num). The data type can be INT8 or BOOL. If the data type is INT8, the value can be [object Object] or [object Object]. If the data type is BOOL, the value can be [object Object] or [object Object]. The must be ND. are supported. In drop/pad-less mode, each row must contain topK [object Object] or [object Object] values.
  • [object Object] (int64_t, computation input): number of experts involved in the computation.
  • [object Object] (int64_t, computation input): number of tokens involved in the computation.
  • [object Object] (bool, computation input): [object Object] indicates that [object Object] is enabled, and [object Object] indicates that [object Object] is disabled.
  • [object Object] (aclTensor*, output): gradient of the input [object Object]. The value must be a 2D tensor with shape (tokens_num, hidden_size). The data type is the same as that of [object Object]. The data type can be BFLOAT16, FLOAT16, or FLOAT32. The must be ND. are not supported.
  • [object Object] (aclTensor*, output): gradient of the input [object Object]. This is an optional output. The value must be a 2D tensor with shape (tokens_num, experts_num). The data type is the same as that of [object Object]. The data type can be BFLOAT16, FLOAT16, or FLOAT32. The must be ND. are not supported.
  • [object Object] (uint64_t*, output): size of the workspace to be allocated on the device.
  • [object Object] (aclOpExecutor**, output): operator executor, containing the operator computation process.

• Returns:

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

[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 the first-phase API [object Object].
  • [object Object] (aclOpExecutor*, input): operator executor, containing the operator computation process.
  • [object Object] (aclrtStream, input): stream for executing the task.

• Returns:

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

[object Object]

• Deterministic computation:
  • [object Object] defaults to deterministic implementation.
• Non-dropPaddedMode scenario: The value of [object Object] is less than or equal to [object Object].
• Mixed precision input is not supported. That is, [object Object], [object Object], [object Object], and [object Object] must have the same data type.

[object Object]

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