Guides Metal backend kernel development, operator fusion, runtime scheduling and performance diagnosis on Apple GPUs.
alibaba/MNN
Guides coding agents through MNN inference engine work: adding operators and models, fixing bugs, and optimizing CPU, GPU and NPU backends.
- 13 skills
- 16.4k stars
The 5 most installed
Ordered by installs on skills.sh- metal-optimize1 installs
- vulkan-optimize1 installs
Helps optimize Vulkan kernels, add new Vulkan features and verify results across Adreno, Mali and Apple GPUs.
Walks through adding a new operator to MNN, from schema and shape calculation to backend code and unit tests.
Routes MNN correctness and regression bugs by symptom, covering memory, quantization, threading, GPU and cache issues.
Points to the right CPU backend guide for ARM, x86_64 or RISC-V, whether optimizing speed or writing kernels.
The other 8 skills
Guides writing or porting CPU kernels with SIMD and assembly, registering them and checking correctness across instruction sets.
Diagnoses why an operator or LLM run is slow on CPU and identifies which layer is the bottleneck.
Helps optimize, refactor and test the Hexagon DSP backend, including profiling on device and crash diagnosis.
Helps optimize OpenCL kernels, add new OpenCL features and verify results on Android devices.
Diagnoses QNN NPU errors, accuracy problems and performance issues, and guides adding new operators.
After a nontrivial task, captures lessons learned from debugging and failed attempts for later reuse.
Walks through adding a new LLM to MNN, from downloading and analyzing the model to export and alignment tests.
Runs MNN tests and CI on host or real devices, and benchmarks LLMs on Android or iOS hardware.