AKG (Auto Kernel Generator) is an optimizer for operators in Deep Learning Networks, which provides the ability to automatically fuse ops with specific patterns.
Skill Claude CodeCodex
A small language and method for sketching how a computing operation should work before writing its implementation. It describes tensors, memory actions, loops, and optimization hints in a compact form.
Skill Claude CodeCodex
A workflow guide for KernelAgent, an assistant that creates and checks computing operations through a step-by-step ReAct process. It requires the user to confirm the interpretation, generated task, execution method, and result.
Skill Claude CodeCodex
A workflow for generating, checking, and optimizing GPU kernels from a task description or existing code. A kernel is a GPU function designed to process data in parallel.
Skill Claude CodeCodex needs its repo
A skill for combining performance results from several operators, which are reusable computation steps such as ReLU or sigmoid. It reads recorded measurements and produces a comparison of latency, throughput, memory use, and speedup.
Skill Claude CodeCodex
A PyPTO example of GELU, an activation function that smoothly scales neural-network values instead of simply keeping or discarding them. It applies the calculation element by element after flattening the input.
Skill Claude CodeCodex
An example of a PyPTO kernel for calculating cross-entropy loss, a common machine-learning measure of prediction error. It processes predictions and target classes in tiles, applies softmax, selects the target probabilities, and returns one scalar loss.
Skill Claude CodeCodex
An example of a two-dimensional matrix multiplication kernel using loop-based blocks and separate handling for leftover rows. Matrix multiplication combines rows and columns of number grids to produce a new grid.
Skill Claude CodeCodex
A PyPTO pattern for multiplying a matrix by a column vector when the vector length is over 65,535. It reshapes the vector into a row and sums element-by-element products.
Skill Claude CodeCodex
A PyPTO example of BatchNorm, a method that normalizes values using statistics calculated across a batch and spatial positions. It reshapes the input to three dimensions and processes channels in groups.
Skill Claude CodeCodex
A PyPTO example of LayerNorm, which normalizes each row of values using that row’s mean and variance. It flattens each input item to two dimensions, processes rows in groups, and restores the original shape.
Skill Claude CodeCodex
A PyPTO example that adds values along one dimension of a three-dimensional tensor while keeping that dimension in the output. Reduction means combining many values into one, such as with a sum.
Skill Claude CodeCodex
A quick reference for PyPTO's programming interface, including kernel decorators, tensors, data types, and tile configuration. PyPTO is a Python interface for describing tensor and kernel computations.
Skill Claude CodeCodex
A set of rules for implementing PyPTO kernels from PyTorch code. It starts by writing down the original operation's exact mathematical meaning before choosing an implementation or optimization.
Skill Claude CodeCodex
A rulebook for using PyPTO loops and views when view shapes must be known during compilation. A view is a different way of interpreting part of a tensor without changing its data.
Skill Claude CodeCodex
A set of PyPTO performance-tuning rules for tile sizes, loops, data reduction, and operations such as softmax, normalization, and loss calculations. A tile is a small block of tensor data processed together.
Skill Claude CodeCodex
A list of PyPTO coding pitfalls and corrected patterns for tensor arithmetic, clamping, loss functions, factory parameters, and large matrix multiplications. PyPTO is a tensor-programming system with rules that differ from ordinary Python arithmetic.
Skill Claude CodeCodex
A workflow for extracting a PyTorch or Triton operation from a code repository and packaging it as one self-contained KernelBench task file. KernelBench is a format for standardizing and testing machine-learning kernel tasks.
Skill Claude CodeCodex
A reference manual for the TileLang CUDA programming interface, including kernel compilation, launch dimensions, thread blocks, memory allocation, and utility functions. TileLang is a Python-like language for writing GPU kernels.
Skill Claude CodeCodex
A beginner-to-advanced guide to TileLang CUDA, covering its Python-like syntax, GPU kernels, thread blocks, and memory types. A kernel is a function compiled to run many operations in parallel on a GPU.
Skill Claude CodeCodex
A collection of runnable examples showing how to write CUDA kernels in TileLang, a Python-like language for GPU code, and use them with PyTorch, a machine-learning library.
Skill Claude CodeCodex
A guide to improving how TileLang CUDA kernels move data through GPU memory. It covers shared memory, registers, local storage, data layouts, combined memory access, and avoiding bank conflicts, where threads compete for the same memory bank.
Skill Claude CodeCodex
A reference guide for optimizing and debugging TileLang programs that run CUDA GPU kernels. TileLang is a programming language for describing GPU computations.
Skill Claude CodeCodex
A collection of standard TileLang code patterns and templates for common GPU computations. It explains how to write element-wise operations, reductions, matrix multiplication, and matrix–vector multiplication.
Skill Claude CodeCodex
A set of rules for synchronizing CUDA threads in TileLang, especially when they share memory or work together. It explains how to avoid deadlocks, where threads wait forever for one another.
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: