ResourcesChangelog

Changelog

A summary of MemScale releases. The authoritative, always-current changelog lives at memscale.id/changelog.

v1.2.0 — 2026-05-21 — “Intelligence Foundation”

Released. The ML policy ships off by default (auto_policy=False) — default behavior is bit-for-bit identical to v1.1.

The v1.2 cycle adds an ML-assisted optimization policy framework on top of the deterministic rule engine, plus the verification work behind it:

  • Two-stage optimization flow — Stage 1 strategy selection (opt-in ML policy via Config.auto_policy=True), Stage 2 the existing deterministic rule engine. A trained policy artifact ships with this release.
  • Expanded opt-in v2 telemetry — bucketed, non-identifying model and decision metadata as a training corpus. Off unless you enable it.
  • Empirical efficacy sweep — a 25-configuration LHS sweep on an RTX 3090 (apply_all_optimizations stack) measured a median ~59% peak-VRAM reduction (range 51–70%), reconciling the headline best-case figures with typical diverse workloads.
  • Cython memory-safety audit — 48 modules audited and hardened for the binary-distributed build.
  • Multi-platform wheels — 9 wheels across Linux, Windows, and macOS (CPython 3.10/3.11/3.12).

v1.1.0 — 2026-05-17 — “The Performance Release”

The first release with empirically-backed, reproducible benchmark numbers committed to the codebase.

  • Experimental async CPU offload — wrap(model, async_offload=True) enables a tier-aware async offload engine. Default off; existing users unaffected. Also settable via MEMSCALE_ASYNC_OFFLOAD=1.
  • Benchmark CLI — python -m memscale.benchmarks, a reproducible VRAM benchmark suite over a registry of standard models.
  • The Hugging Face Trainer integration now reports real samples/sec throughput (previously hard-coded to 0.0).
  • estimate_savings() now defaults to an honest same-GPU cost comparison.

v1.0.x — May 2026

  • v1.0.4 (2026-05-15) and v1.0.3 (2026-05-15) — stability and correctness fixes, including auto-disabling HuggingFace use_cache when activation checkpointing is enabled, to avoid a recompute shape mismatch.
  • v1.0.2 (2026-05-12) — no API key required; the library works fully offline. Optional anonymous opt-in telemetry introduced.
  • v1.0.1 (2026-05-09) — post-launch fixes.
  • v1.0.0 (2026-05-09) — Library Launch.

Earlier (0.x)

  • v0.3.0 (2026-04-26) — “Production Foundations.”
  • v0.2.0 (2026-04-26).
  • v0.1.0 (2026-04-26) — Initial public release.

Versioning

MemScale follows Semantic Versioning. The public API surface — wrap, optimize, detach, Config, OptimizationMode — is what semantic-versioning guarantees apply to; internal modules may change between releases.

GPU CompatibilityFAQ