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
An optimization pattern for dividing a 2D tensor by values broadcast across rows or columns. It keeps small dimensions together and controls how many processing blocks handle the larger dimension.
Skill Claude CodeCodex
An optimization pattern for 3D broadcast division, where one tensor's values are repeated across selected dimensions. It first expands the repeated values and then reshapes the work into a simpler 2D form.
Skill Claude CodeCodex
An optimization pattern for converting large arrays from int8 numbers to fp16 numbers on Ascend hardware. It splits the work into blocks and smaller tiles so processing can use the available on-chip memory.
Skill Claude CodeCodex
An optimization pattern that combines several slice operations and a concatenation into one kernel. Slicing selects part of a tensor, while concatenation joins those parts together.
Skill Claude CodeCodex
A tuning guide for creating small tensors with operations such as zeros, arange, full, and their variants on Ascend hardware. It shows how to use fewer processing blocks for small shapes.
Skill Claude CodeCodex
An optimization pattern for histogram counting, which records how often each value appears. It sorts the input first, then uses binary search to find each value's range.
Skill Claude CodeCodex
An optimization pattern for indexed assignment, which writes values into positions chosen by index arrays. It loads index data into fast on-chip memory so a loop can reuse it.
Skill Claude CodeCodex
An optimization guide for matrix multiplication where the shared dimension K is much larger than the output dimensions M and N. It splits the K dimension so multiple processor cores can calculate parts of the same output.
Skill Claude CodeCodex
An optimization guide for Triton matrix-multiplication kernels on Ascend processors. Matrix multiplication combines rows and columns of two number grids to produce a third grid.
Skill Claude CodeCodex
An optimization guide for Triton kernels that find maximum values across a very large dimension of a 2D array on an Ascend processor. It targets cases where the other dimension is very small.
Skill Claude CodeCodex
An optimization guide for Triton kernels that compute the maximum value across one dimension of a medium-sized 2D array on an Ascend processor.
Skill Claude CodeCodex
A Triton optimization guide for finding the maximum value in a very small array on Ascend hardware. It compares processing the whole input with one core against using several cores.
Skill Claude CodeCodex
A Triton optimization guide for minimum-value reduction when the reduced dimension is much larger than the other dimension. It uses multiple cores and atomic operations, which combine partial results safely, for extreme shapes such as 16 by 262,144.
Skill Claude CodeCodex
A specialized optimization note for very large one-dimensional minimum reductions in Triton on Ascend hardware.
Skill Claude CodeCodex
A Triton optimization guide for finding row minimums in a large two-dimensional array on Ascend hardware. It handles cases where the dimension being reduced contains hundreds of thousands of elements.
Skill Claude CodeCodex
A Triton optimization guide for finding the minimum value in a medium-sized one-dimensional array on Ascend hardware. It focuses on choosing a suitable amount of parallel work for inputs around 65,536 elements.
Skill Claude CodeCodex
A specialized optimization note for Triton mean-reduction kernels on Ascend hardware, focused on large two-dimensional arrays with a long reduction axis.
Skill Claude CodeCodex
A guide to optimizing medium-sized mean reductions, which calculate the average across one dimension of a tensor on an Ascend AI processor.
Skill Claude CodeCodex
A Triton Ascend guide for speeding up small product reductions, which multiply values along one dimension of a tensor on an Ascend AI processor.
Skill Claude CodeCodex
A guide to optimizing a fused operation that first transforms each value and then adds the results along a dimension.
Skill Claude CodeCodex
A guide to optimizing large sum reductions, which add values along one tensor dimension when the other dimension is very large.
Skill Claude CodeCodex
A Triton Ascend guide for optimizing the backward pass of Weighted SwiGLU, a neural-network operation that applies weights and computes gradients.
Skill Claude CodeCodex
A troubleshooting guide for common Triton Ascend kernel failures, including on-chip memory overflow and compiler errors caused by complex masks or addresses.
Skill Claude CodeCodex
A practical guide to optimizing fused Triton operations on Ascend AI processors, where several calculations are combined into one kernel.
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: