[object Object]

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

• Description: Applies rotary positional encoding (RoPE) to a single input tensor [object Object].

• Formulas:

  $$q = \text{reshape}(x, [B, N, S, D//2, 2]) \cdot \text{transpose}(-1, -2) \cdot \text{reshape}([B, N, S, D])$$ $$q_{\text{embed}} = q \cdot \text{cos} + \text{RotateHalf}(q) \cdot \sin$$

  RotateHalf(q) indicates that the elements in the second half of the D dimension of q are moved to the first half and multiplied by –1, and the second half is replaced by the values of the first half.

  $$\text{RotateHalf}(q)_{\text{i}} = \begin{cases} -q_{i+D//2} & \text{if } i < D//2 \\ q_{i+D//2} & \text{otherwise} \end{cases}$$

[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]

• Returns:

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

  The first-phase API implements input parameter validation. The following error codes may be returned.

  [object Object]

[object Object]

• Parameters

  [object Object]

• Returns:

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

[object Object]

• Deterministic computation:
  • [object Object] defaults to a deterministic implementation.
• This API can be used in inference scenarios.
• The x, cos, sin, and out must be 4D tensors with the shape of (B, N, S, D).
• The S dimension of cos and sin can be 1 or the same as that of x. The N dimension must be 1.
• The D dimension of x, cos, and sin must be 64.

[object Object]

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