Enum CuBlasLtEpilogue
- Namespace
- NVMathNet.Interop
- Assembly
- NVMath.Net.dll
cuBLASLt epilog operations applied after the matrix multiplication.
public enum CuBlasLtEpilogueFields
BGradA = 256-
Backward bias-gradient accumulation for A.
BGradB = 512-
Backward bias-gradient accumulation for B.
Bias = 4-
Add a per-row bias vector to the output.
DGelu = 192-
Backward GELU — compute gradient through GELU.
DGeluBGrad = 208-
Backward GELU with bias gradient.
Default = 1-
No epilog \u2014 output D = alphaAB + beta*C unchanged.
DreLu = 136-
Backward ReLU — compute gradient through ReLU.
DreLuBGrad = 152-
Backward ReLU with bias gradient.
Gelu = 32-
Apply GELU activation.
GeluAux = 160-
Apply GELU and save auxiliary data for backward pass.
GeluAuxBias = 164-
Apply GELU with bias and save auxiliary data.
GeluBias = 36-
Add per-row bias, then apply GELU.
Relu = 2-
Apply ReLU activation: D = max(0, alphaAB + beta*C).
ReluAux = 130-
Apply ReLU and save auxiliary data for backward pass.
ReluAuxBias = 134-
Apply ReLU with bias and save auxiliary data.
ReluBias = 6-
Add per-row bias, then apply ReLU.