0xdti / triton-activations

Collection of neural network activation function kernels for Triton Language Compiler by OpenAI

Triton Activations

Expanded collection of Neural Network activation functions and other function kernels in Triton by OpenAI.

Usage

git clone github.com/simudt/triton-activations
cd triton-activations
python3 setup.py install
python3 examples.py

List of the activation functions in Triton Activations

Hyperbolic Tangent Activation Function Kernel

Applies the hyperbolic tangent activation function to transform input data. It squashes the input values to the range [-1, 1], providing non-linear behavior with outputs centered around zero.

Hard Tanh Activation Function Kernel

Applies the hard tanh activation function to transform input data. It is a piecewise linear approximation of the hyperbolic tangent function and maps input values to the range [-1, 1]. Unlike the hyperbolic tangent function, the hard tanh function has a flat region where the output saturates, resulting in a thresholding effect.

Rectified Linear Unit Activation Function Kernel

Uses the rectified linear unit (ReLU) activation function to process input data. It sets negative values to zero while passing positive values unchanged, resulting in a piecewise linear activation function.

ReLU 6 Activation Function Kernel

Similar to the Rectified Linear Unit (ReLU) kernel, this kernel also uses the ReLU activation function. However, it additionally clips the output values at 6, limiting them to the range [0, 6].

Leaky ReLU Activation Function Kernel

Computes the element-wise function leaky_relu(x) = max(x, alpha * x), where x is the input tensor and alpha is a given slope parameter.

Softplus Activation Function Kernel

Applies the softplus activation function to the input data. It produces smoothed and continuously differentiable outputs, with positive values increasing linearly and negative values converging to zero.

Softsign Activation Function Kernel

Transforms the input data using the softsign activation function. It maps the input to the range [-1, 1], allowing the output to be sensitive to small changes around zero.

Sigmoid Activation Function Kernel

Applies the sigmoid activation function to the input data. It squashes the input values between 0 and 1, interpreting them as probabilities.

Hard Sigmoid Activation Function Kernel

Approximates the sigmoid function with a piecewise linear function. It speeds up computation by sacrificing some accuracy, mapping input values to the range [0, 1].

Sigmoid-weighted Linear Unit Activation Function Kernel

Combines the sigmoid and linear activation functions. It multiplies the linear component by the sigmoid output, emphasizing or de-emphasizing the linear contribution based on the sigmoid value.

Hard SiLU Activation Function Kernel

Piecewise linear function that approximates the SiLU (sigmoid-weighted linear unit) activation function. It is designed to provide a non-linear activation while being computationally efficient.

Gaussian Error Linear Unit Activation Function Kernel

Applies a smooth approximation of the Gaussian cumulative distribution function to the input data. It introduces non-linearity while preserving certain properties of the Gaussian distribution.

Softmax Activation Function Kernel

Applies the softmax activation function, which converts a vector of real values into a probability distribution. It exponentiates and normalizes the input values, ensuring that the resulting outputs sum up to 1.