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

• API function: Adds dequant and quant operations before and after the Swish gating linear unit activation function to implement the DequantSwigluQuant computation of x. Compared with , this API has two types of new parameters: (1) Three input parameters used by the Ascend 950 chip are added: dstType, roundModeOptional, and activateDim. (2) On the Atlas A2 and Atlas A3 chips, four parameters are added for the variant SwiGLU used by GPT-OSS: swigluMode, clampLimit, gluAlpha, and gluBias. When this API is used on the Ascend 950 chip, default values need to be set for these four parameters. Select the appropriate API as required.

• The formula for swigluMode = 0 is as follows:

  $$dequantOut_i = Dequant(x_i)$$ $$swigluOut_i = Swiglu(dequantOut_i)=Swish(A_i)*B_i$$ $$out_i = Quant(swigluOut_i)$$

  where A[object Object]i[object Object] indicates the first half of dequantOut[object Object]i[object Object], and B[object Object]i[object Object] indicates the second half of dequantOut[object Object]i[object Object].

• The formula for swigluMode = 1 is as follows:

  $$dequantOut_i = Dequant(x_i)$$ $$x\_glu = x\_glu.clamp(min=None, max=clampLimit)$$ $$x\_linear = x\_linear.clamp(min=-clampLimit, max=clampLimit)$$ $$out\_glu = x\_glu * sigmoid(gluAlpha * x\_glu)$$ $$swigluOut_i = out\_glu * (x\_linear + gluBias)$$ $$out_i = Quant(swigluOut_i)$$

  x_glu indicates the even-indexed part of dequantOut[object Object]i[object Object], and x_linear indicates the odd-indexed part of dequantOut[object Object]i[object Object].

[object Object]

Each operator has [object Object]two-phase API calls[object Object]. You must call aclnnDequantSwigluQuantV2GetWorkspaceSize to obtain the workspace size required for computation and the executor that contains the operator computation process, and then call aclnnDequantSwigluQuantV2 to perform the 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:

  • The default deterministic implementation of aclnnDequantSwigluQuantV2 is used.

• The dimension of activateDim corresponding to the input x must be a multiple of 2, and the number of dimensions of x must be greater than 1.

• If the data type of the input x is INT32, weightScaleOptional cannot be null. If the data type of the input x is not INT32, weightScaleOptional cannot be input; a null pointer must be passed.

• If the data type of the input x is not INT32, activationScaleOptional cannot be input and a null pointer is passed.

• If the data type of the input x is not INT32, biasOptional cannot be input and a null pointer is passed.

• When the output data type of yOut is FLOAT4_E2M1 or FLOAT4_E1M2, the last dimension of yOut must be a multiple of 2.

• If the dimension corresponding to activateDim is not the last axis of x, the last axis of the output yOut cannot exceed 5120.

• The sum of all elements in groupIndexOptional cannot be greater than the product of the remaining axes of the input x except the last axis.

• The parts of the output yOut and scaleOut that exceed the sum of all elements in groupIndexOptional are not cleared. The memory of this part is junk data.

• [object Object]Atlas A2 training products/Atlas A2 inference products[object Object] and [object Object]Atlas A3 training products/Atlas A3 inference products[object Object]: When groupIndexOptional is input, the maximum size of the input tensor supported by the operator is limited. The last axis of x cannot exceed 7232.

[object Object]

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