Using PyTorch Engineering

Overview

This meta-skill routes you to the right PyTorch specialist based on symptoms. PyTorch engineering problems fall into distinct categories that require specialized knowledge. Load this skill when you encounter PyTorch-specific issues but aren't sure which specialized skill to use.

Core Principle: Different PyTorch problems require different specialists. Match symptoms to the appropriate specialist skill. Don't guess at solutions—route to the expert.

API surface calibrated to PyTorch 2.9+ (verified against torch 2.9.1), reviewed 2026-08. Deprecated torch.cuda.amp aliases have been migrated to torch.amp; FairScale ZeRO references are replaced with native FSDP1/FSDP2. Modern features (torch.compile, FlexAttention, CUDA Graphs, NVTX/Nsight Systems, DTensor, expandable_segments, channels_last) are covered as first-class topics. FP8 appears only as a debugging topic (NaN/Inf triage) — torch.amp has no FP8 autocast path, so FP8 recipes and strategy live in yzmir-training-optimization.

About This Pack's API Currency

Reconciliation gate — what you can rely on inside this pack: