pytorch

name	pytorch
description	PyTorch deep learning framework — tensor compute, autograd, nn.Module training loops, DataLoader, mixed precision (torch.amp), torch.compile, GPU/MPS device handling, distributed (DDP/FSDP2), TorchScript/ONNX export. Use when: pytorch, torch, deep learning, nn.Module, autograd, requires_grad, DataLoader, Dataset, optimizer, training loop, torch.cuda, MPS, Apple Silicon, mixed precision, autocast, GradScaler, bfloat16, torch.compile, fullgraph, DDP, DistributedDataParallel, torchrun, FSDP, fully_shard, TorchScript, ONNX, GPU training, tensor, .to(device), .cuda(), CUDA OOM, NaN loss. SKIP: classical ML (→scikit-learn), GPU compute without DL (→cuda-python), tabular data prep (→pandas/polars), serving APIs (→fastapi).
stacks	["pytorch","python"]
tags	["pytorch","deep-learning","gpu","ml","training","inference"]
packages	["torch","torchvision","torchaudio"]
manifests	["pyproject.toml","requirements.txt"]
source	vechkasov-global-skills
risk	medium-stakes

🎯 Version Requirements (May 2026)

Primary pins:

• PyTorch: 2.12.x
• Python: 3.14.x

Source of truth: STACK_VERSIONS.md — verified 2026-05-16

Use this skill when

• Training neural networks (CNNs, transformers, MLPs) from scratch or fine-tuning pretrained models
• Writing custom nn.Module layers, losses, or model heads
• Building training loops with DataLoader, optimizer step, lr schedulers, checkpoints
• Optimizing throughput with mixed precision (torch.amp.autocast + GradScaler) or torch.compile
• Distributed training with DistributedDataParallel or FSDP2 (fully_shard)
• Running inference on GPU/CPU/MPS, exporting to ONNX or torch.export
• Debugging CUDA OOM, NaN losses, slow DataLoader, autograd errors

Do not use this skill when

• Classical (non-DL) ML — use scikit-learn instead
• Raw GPU compute without neural-net structure — use cuda-python instead
• Tabular feature engineering and EDA — use pandas / polars
• Serving a trained model behind an HTTP API — use fastapi (PyTorch handles the model; serving is separate)
• Authoring an LLM pipeline with HuggingFace abstractions — use the transformers skill cascade (when present)
• Native CUDA C++ kernels — out of scope; use cuda-python or write a custom extension

Purpose

PyTorch is the dominant research-to-production deep learning framework. It exposes a NumPy-like tensor library with a tape-based autograd engine, an extensible nn.Module system, eager execution by default, and JIT compilation through torch.compile. Production codebases use it for training (single-GPU, multi-GPU DDP, multi-node FSDP), fine-tuning, and inference across CUDA, ROCm, CPU, and Apple Metal (MPS).

This skill codifies the patterns that actually matter in real training jobs: device-agnostic code, mixed precision without NaNs, gradient accumulation, checkpointing, torch.compile with sensible fallbacks, and the operational gotchas ( in hot loops, forgotten , leaked autograd graph, OOM on the third epoch). It is opinionated toward modern PyTorch 2.x — over the deprecated , (FSDP2) over the original FSDP wrapper, over for pure inference, and over TorchScript.