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 guide to implementing and optimizing operations that process tensor elements independently, such as addition, multiplication, activation functions, and mathematical functions. It includes vectorized access and operation fusion patterns.
Skill Claude CodeCodex
A collection of complete examples showing how to call Triton CUDA kernels from PyTorch. The examples include vector addition, matrix multiplication, layer normalization, and softmax.
Skill Claude CodeCodex
A guide to choosing the grid and block layout used to launch Triton CUDA kernels. It explains one-, two-, and three-dimensional grids, block sizes, and handling very large inputs.
Skill Claude CodeCodex
A guide to speeding up matrix multiplication, including ordinary, batched, and linear-layer operations, on CUDA GPUs. It covers tiling, shared-memory caching, and Tensor Cores, specialized GPU units for matrix calculations.
Skill Claude CodeCodex
A guide to improving how Triton CUDA kernels read and arrange data in GPU memory. It covers shared memory, continuous access, avoiding bank conflicts, and data layouts.
Skill Claude CodeCodex
A general guide to optimizing Triton kernels on CUDA GPUs. It covers block size, warps, pipeline stages, memory access, operation fusion, occupancy, numerical stability, and API limits.
Skill Claude CodeCodex
A set of standard Triton CUDA programming patterns for vector operations, reductions, and matrix multiplication. It provides reusable structures for indexing, masked memory access, aggregation, and tiled computation.
Skill Claude CodeCodex
A guide to reducing many values into one result on a GPU, such as a sum, average, maximum, softmax, or layer normalization. It covers block-level reduction and numerical-stability techniques.
Skill Claude CodeOpenCode needs its repo
An environment setup procedure for AKG Agents that checks the local tools and language-model access, detects hardware and software details, and stores the results in a cache.
Skill Claude CodeOpenCode
An algorithm-sketch design tool for computational operators. It produces high-level pseudocode and implementation guidance for several programming and hardware languages.
Skill Claude CodeOpenCode
A code-generation tool for computational operators, with support for several hardware-focused programming languages and frameworks.
Skill Claude CodeOpenCode
A verification tool for generated computational-operator code. It first checks the code without running expensive tests, then compares its outputs with the framework implementation.
Skill Claude CodeOpenCode
A task-extraction tool that turns source code or a natural-language operator description into a self-contained Python task file for later generation or testing.
Skill Claude CodeOpenCode
A robustness-testing tool for computational operators, which are small pieces of code that perform tensor calculations.
Skill OpenCode
A performance-optimization verification tool for checking whether optimized code is still correct and faster in its target environment.
Skill Claude CodeOpenCode
A workflow for searching for better implementations of model operations with methods such as adaptive search or evolution. It runs the search in the background and checks its progress until results are available.
Skill Claude CodeOpenCode
A guide for turning useful lessons from a completed conversation into a reusable OpenCode skill. It describes when to save the lesson and how to structure the skill file.
Skill OpenCode
A guide to loading and adapting sparsely quantized models in vLLM-MindSpore. Sparse quantization stores model weights in a smaller numerical format while preserving a sparse structure, and W8A8SC is one such format.
Skill OpenCode
A guide to improving the processing that follows model inference in vLLM-Ascend, a text-generation system for Huawei Ascend hardware. It covers both PyTorch-based and Triton-based sampling paths.
Skill OpenCode
A guide to speeding up the processing that happens after vLLM-MindSpore produces model results and before it selects the next tokens. vLLM-MindSpore is a text-generation system adapted for Huawei Ascend hardware.
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: