skills.rest
steleman
stelemanCOMMUNITY
Explore all skills by this author →
GitHub

metal-kernel

Implement Metal kernels and host stubs for PyTorch operators on Apple Silicon.

1|Updated Apr 5, 2026
Search Tags:#dispatch#metal#kernel#pytorch#apple-silicon#mps#host-stub
One-click install
npx skills add https://github.com/steleman/pytorch-cuda-2.11.0 --skill metal-kernel
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: metal-kernel
Source: https://github.com/steleman/pytorch-cuda-2.11.0/tree/main/.claude/skills/metal-kernel
Command: npx skills add https://github.com/steleman/pytorch-cuda-2.11.0 --skill metal-kernel

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Metal kernel development for PyTorch on Apple Silicon to enable native GPU acceleration of operators.

Core Features & Use Cases

  • Implement Metal kernels for PyTorch operators in aten/src/ATen/native/mps/kernels/.
  • Add Dispatch mappings to native_functions.yaml for MPS support and ensure proper host-side stubs are wired in aten/src/ATen/native/mps/operations/.
  • Provide battle-tested host-side integration via TensorIterator-based workflows and stepwise migration from MPSGraph to native Metal.

Quick Start

Run the task to implement a Metal kernel and a corresponding host stub for a new PyTorch operator.

Frequently Asked Questions about metal-kernel

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I implement Metal kernels for PyTorch operators on Apple Silicon?▼

Implement Metal kernels for PyTorch operators by writing kernel code in aten/src/ATen/native/mps/kernels/ and wiring host-side stubs in Operations to enable native Apple Silicon GPU acceleration.

How do I add MPS dispatch mappings to native_functions.yaml?▼

Add MPS dispatch mappings to native_functions.yaml to register new PyTorch operators, ensuring proper host-side stubs are integrated in aten/src/ATen/native/mps/operations/ for correct dispatch routing.

Can I port existing CUDA kernels to Metal for MPS support?▼

Porting CUDA kernels to Metal for MPS support involves migrating to native Metal implementations in kernels/ and replacing MPSGraph workflows with TensorIterator-based host-side integration.

What is the process for migrating from MPSGraph to native Metal in PyTorch?▼

Migrating from MPSGraph to native Metal requires writing Metal kernels in kernels/ and implementing TensorIterator-based host-side workflows to achieve native Apple Silicon acceleration for PyTorch operators.

Do I need host-side stubs to wire Metal kernels into PyTorch?▼

Host-side stubs are required to wire Metal kernels into PyTorch, ensuring the Operations layer correctly interfaces with the kernel implementations and updates native_functions.yaml dispatch mappings.

Why use native Metal kernels instead of MPSGraph for PyTorch operators?▼

Native Metal kernels provide battle-tested host-side integration and stepwise migration from MPSGraph, enabling optimized native Apple Silicon acceleration directly within the PyTorch operator stack.

Related Skills

Explore More Software Engineering Skills→
plurigridplurigrid
60

exo-distributed

Coordinate distributed LLM inference across Apple Silicon devices with auto peer discovery.

Official
Advanced
Just-itJust-it
33

ascend-benchmark-evaluator

Assess Ascend C operator correctness and runtime using NPUKernelBench datasets.

Community
Advanced
yoshisadayoshisada

trim-push

Push code components to Penpot and update component mappings.

Community
Advanced
tondevreltondevrel
19

pytorch

Develop and train deep learning models using PyTorch with tensor operations and automatic differentiation.

Community
Advanced