[object Object]

This API will be deprecated in later versions. Do not use this API.

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

• Description: Implements the multi-rank parallel AllReduce function based on grouped matrix multiplication, supporting non-uniform matrix dimension sizes across multiple groups. The following four scenarios are supported based on the tensor count of $x$, $weight$, and $y$:

  • The number of tensors for [object Object], [object Object], and [object Object] equals the group count. That is, the tensors of each group are independent.
  • [object Object] has one tensor, while [object Object] and [object Object] have tensors equal to the group count. In this case, use the optional parameter [object Object] to define the row-wise grouping of [object Object]. For example, [object Object] indicates that the first 10 rows of [object Object] participate in the multiplication of the first group of matrices.
  • The number of tensors for [object Object] and [object Object] equals the group count, but [object Object] has only one tensor. In this case, products of each matrix group multiplication are stored contiguously within a single tensor.
  • [object Object] and [object Object] each have one tensor, while [object Object] has tensors equal to the group count. This is a hybrid configuration combining the preceding two cases.

• Formula:

  Non-quantization scenario:

  $$y_i=x_i\times weight_i + bias_i$$

[object Object]

Each operator has calls. First, [object Object] is called to obtain the input parameters and compute the required workspace size based on the process. Then, [object Object] is called to perform computation.

[object Object]

• Parameters

  [object Object]

• Return

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

  The first-phase API implements input parameter validation. The following errors may be thrown:

  [object Object]

[object Object]

• Parameters

  [object Object]

• Return

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

[object Object]

• Deterministic computation:

  • [object Object] defaults to a deterministic implementation.

• Constraints on data type combinations: The supported data type combinations for [object Object], [object Object], and [object Object] are as follows:

  • [object Object]: FLOAT16; [object Object]: FLOAT16; [object Object]: FLOAT16
  • [object Object]: BFLOAT16; [object Object]: BFLOAT16; [object Object]: FLOAT32

• Dimension constraints:

  • When [object Object] is set to [object Object], [object Object] and [object Object] both support 2D to 6D.
  • When [object Object] is set to [object Object], [object Object], or [object Object], both [object Object] and [object Object] support 2D.
  • Regardless of the value of [object Object], [object Object] supports 2D.

• Device quantity: 2, 4, or 8 devices can be deployed.

• Tensor dimension size requirements:

  • The size of the last dimension for each tensor in [object Object] and [object Object] must be less than 65536. (For [object Object], the last dimension refers to the K-axis when [object Object] is false, and the M-axis when it is true; for [object Object], the last dimension refers to the N-axis when [object Object] is false, and the K-axis when it is true.)
  • The size of each dimension for every tensor in [object Object] and [object Object], after 32-byte alignment, should be less than the maximum value of INT32 (2147483647).

[object Object]

The following is sample code and is for reference only. For the detailed compilation and execution process, see the compilation and execution example.