metal-kernel
Implement native Metal shaders for PyTorch operators on Apple Silicon.
Install
mkdir -p .claude/skills/metal-kernel && curl -L -o skill.zip "https://agentskills.codes/api/skills/download/2590" && unzip -o skill.zip -d .claude/skills/metal-kernel && rm skill.zipInstalls to .claude/skills/metal-kernel
Activation
This is the description your AI agent reads to decide when to run this skill — the better it matches your request, the more reliably it fires.
Write Metal/MPS kernels for PyTorch operators. Use when adding MPS device support to operators, implementing Metal shaders, or porting CUDA kernels to Apple Silicon. Covers native_functions.yaml dispatch, host-side operators, and Metal kernel implementation.Key capabilities
- →Update native_functions.yaml for MPS dispatch
- →Implement Metal kernels using c10/metal/ infrastructure
- →Create host-side stubs for operator dispatch
- →Register unary and binary operations
- →Handle complex number types in Metal
How it works
It uses the c10/metal/ infrastructure to register custom Metal kernels and host-side stubs, replacing legacy MPSGraph implementations.
Inputs & outputs
When to use metal-kernel
- →Implement new MPS operators
- →Port CUDA kernels to Apple Silicon
- →Optimize PyTorch performance
About metal-kernel
Guides the implementation of native Metal kernels via the `c10/metal/` infrastructure. It covers dispatch updates, kernel writing, and host-side stub creation.
Write Metal/MPS kernels for PyTorch operators. Use when adding MPS device support to operators, implementing Metal shaders, or porting CUDA kernels to Apple Silicon. Covers native_functions.yaml dispatch, host-side operators, and Metal kernel implementation.
When not to use it
- →Using MPSGraph for new operator implementation
- →Performing host-side synchronization for error checking
Prerequisites
Limitations
- →Requires native Metal kernel implementation
- →No support for V100 or CPU inference
How it compares
This approach provides native Metal control and performance, avoiding the overhead and limitations of MPSGraph.
Compared to similar skills
metal-kernel side by side with the closest alternatives in the catalog.
| Skill | Installs | Updated | Safety | Difficulty |
|---|---|---|---|---|
| metal-kernel (this skill) | 3 | 6mo | Review | Advanced |
| gguf-quantization | 6 | 10mo | Review | Intermediate |
| pytorch-fsdp2 | 6 | 9mo | No flags | Advanced |
| huggingface-accelerate | 3 | 10mo | Review | Intermediate |
Try saying
Example prompts that trigger this skill in your AI assistant.
More by pytorch
View all by pytorch →You might also like
gguf-quantization
davila7
GGUF format and llama.cpp quantization for efficient CPU/GPU inference. Use when deploying models on consumer hardware, Apple Silicon, or when needing flexible quantization from 2-8 bit without GPU requirements.
pytorch-fsdp2
Orchestra-Research
Adds PyTorch FSDP2 (fully_shard) to training scripts with correct init, sharding, mixed precision/offload config, and distributed checkpointing. Use when models exceed single-GPU memory or when you need DTensor-based sharding with DeviceMesh.
huggingface-accelerate
davila7
Simplest distributed training API. 4 lines to add distributed support to any PyTorch script. Unified API for DeepSpeed/FSDP/Megatron/DDP. Automatic device placement, mixed precision (FP16/BF16/FP8). Interactive config, single launch command. HuggingFace ecosystem standard.
model-pruning
davila7
Reduce LLM size and accelerate inference using pruning techniques like Wanda and SparseGPT. Use when compressing models without retraining, achieving 50% sparsity with minimal accuracy loss, or enabling faster inference on hardware accelerators. Covers unstructured pruning, structured pruning, N:M sparsity, magnitude pruning, and one-shot methods.
optimizing-attention-flash
davila7
Optimizes transformer attention with Flash Attention for 2-4x speedup and 10-20x memory reduction. Use when training/running transformers with long sequences (>512 tokens), encountering GPU memory issues with attention, or need faster inference. Supports PyTorch native SDPA, flash-attn library, H100 FP8, and sliding window attention.
serving-llms-vllm
davila7
Serves LLMs with high throughput using vLLM's PagedAttention and continuous batching. Use when deploying production LLM APIs, optimizing inference latency/throughput, or serving models with limited GPU memory. Supports OpenAI-compatible endpoints, quantization (GPTQ/AWQ/FP8), and tensor parallelism.