Back to blog

Release

Stability under load (especially at 70B+)

Ink/charcoal doodle: a heavy stack is braced for stable large-model runs.

Fixes a GPU memory leak during reload fallback, hardens B200 scripts, and adds practical controls for acceptance ranges and overhead measurement.

1 min read
InvarLock Team

Release: InvarLock 0.3.1 — More knobs when reality gets messy

Highlights

  • GPU memory is freed before reload fallback (reducing OOM risk on big runs).
  • B200 scripts hardened (paths, size detection, error propagation).
  • New env vars for overhead skipping and primary metric acceptance ranges.

0.3.1 is a “make the big runs stop hurting” release. The memory leak fix is the kind of thing you only appreciate after you’ve watched a run fail repeatedly at scale—freeing GPU memory before reloading is a simple change with outsized impact.

The B200 script fixes are in that same category: fewer brittle edges, better failure behavior, less guessing. And the added operational knobs (skip overhead checks for large models, configure acceptance ranges without editing code) are there because real pipelines aren’t always neat—sometimes you need control, not ideology.

For more details, see CHANGELOG.md.

More from the blog

Continue through recent releases and implementation notes.