FAQ

How is MemScale different from DeepSpeed or Accelerate?

DeepSpeed and Accelerate are powerful, but they ask you to restructure your training code and choose strategies — ZeRO stages, sharding, accelerator config — before you see a benefit.

MemScale optimizes for the shortest path from “my model OOMs” to “my model trains.” You call wrap() and keep your training loop unchanged. It is focused on single-GPU memory optimization (see the Multi-GPU guide), where it picks per-layer techniques for you automatically. If you need cluster-scale sharding, a framework like DeepSpeed remains the right tool.

Is MemScale production ready?

v1.1.0 is a stable release with empirically-backed, reproducible benchmark numbers committed to the codebase (see Benchmarking). The core techniques — checkpointing, offloading, tiling — are production paths. Features explicitly labelled experimental — async CPU offload, and the v1.2 ML policy — are off by default and should be treated as such.

Is the source code available?

MemScale is distributed as pre-built, compiled binary wheels on PyPI. The source is closed during 2026 — performance-sensitive modules are Cython-compiled to native extensions. You install and use the library normally; you just do not get the source tree.

Can I use MemScale commercially?

Yes — the MemScale library is free to use under its license. See the MemScale license for the exact terms.

What telemetry does MemScale collect?

By default, none. MemScale runs fully offline and needs no API key.

Telemetry is opt-in. When you enable it, MemScale records anonymous, bucketed, non-identifying metadata — architecture class, parameter-count buckets, which techniques were selected — used to improve optimization decisions and to build the training corpus for the v1.2 ML policy. It never collects your data, weights, or code. See the privacy policy for full terms.

How do I contribute?

Because the source is closed during 2026, MemScale does not currently take outside code contributions. The most useful things you can do:

• File issues and feature requests through the project’s GitHub.
• Enable opt-in telemetry, which directly informs optimization decisions.
• Share benchmark results from your own hardware.

MemScale optimized nothing — why?

For a small model on a large GPU there is nothing to gain, so the decision engine skips optimization by design. Force it with force_optimize=True — see Troubleshooting.

Where do I get help?

Start with Troubleshooting. For anything else, use the contact page.