Architecture

A high-level view of how MemScale is structured, for users who want to understand what runs where.

The optimization pipeline

Every optimized run flows through the same stages, described in Core Concepts: Overview:

wrap() / optimize()
      │
      ▼
  Profiler  ──►  Decision Engine  ──►  Execution Plan  ──►  Executor
 (hardware +     (rule-based,                              (attaches
  graph)          deterministic)                            hooks)

In v1.2 an opt-in ML policy adds a strategy selection stage in front of the decision engine; the deterministic per-layer expansion is unchanged.

Two-tier model

MemScale is built as a freemium product with a clean split:

• MemScale OSS — the pip install memscale library. Core memory optimization (checkpointing, offloading, tiling), the auto-tuner, the OOM predictor, and the Hugging Face / Lightning integrations. Runs entirely on your hardware with no required network calls.
• MemScale Pro / hosted services — the dashboard at app.memscale.id and the API at api.memscale.id, which add accounts, run history, and team features on top of the library.

The documentation you are reading covers the OSS library.

Runs locally, offline

The library performs no mandatory network I/O. There is no API key. The only network feature is opt-in anonymous telemetry, which is off unless you enable it. Profiling, decision-making, and execution all happen in-process on your machine.

Distribution

MemScale is published to PyPI as pre-built, compiled wheels. Performance- sensitive modules are Cython-compiled to native extensions rather than shipped as plain Python; the Python layer orchestrates them. This is why the PyPI package is binary and the source is not public during 2026 — see the FAQ.

Ecosystem and deployment

The MemScale product spans four sites — the memscale.id landing site, this docs site (docs.memscale.id), the app.memscale.id dashboard, and the api.memscale.id backend. The public-facing sites sit behind a Cloudflare-fronted, layered network defense; the library itself, running on your hardware, is not part of that surface.

Where to go deeper

See Internals for the module layout, or Decision Engine for how plans are produced.