TECHNOLOGY PROFILE · TECH-0142
Lightning Fabric
Lightweight training infrastructure around PyTorch loops.
Reference link only — not reviewed this edition. This profile preserves the prepared catalog’s original descriptions. Advantages depend on the workload; comparisons are editorial, not benchmark results.
Where it has an edge
Supplies device, precision and strategy handling without requiring a full trainer abstraction.
Best-fit problems
Custom research loops needing portable training infrastructure.
Tradeoffs & cautions
Debugging still requires knowledge of the selected backend strategy.
What to test before choosing
Scale the same workload across devices; measure useful throughput, peak memory, communication and recovery.
Runtime & hardware
Multi-device or multi-node
Access & licensing
Open-source; backend terms vary
Execution
Local / self-hosted
Lifecycle
Not independently audited
Comparison basis
Editorial problem-fit judgement; no head-to-head benchmark performed
Connected concepts
Alternatives to investigate
Hugging Face Accelerate · PyTorch DistributedDataParallel