Skip to content

FlashAlign 1.2.0

Latest

Choose a tag to compare

@Kurcide Kurcide released this 06 Nov 04:56
061f6bd

FlashAlign 1.2.0 adds live profile management, better diagnostics and more reliable agent setup.

  • Manage profiles without restarting compatible engines. Stage, activate, roll back and retire profiles. Requests already running keep their original generation. Profile management.
  • See what your deployment is doing. New status, doctor and terminal dashboard commands report runtime metadata. Operations.
  • Keep Hermes ownership stable across duplicate loads and reloads. Request audits now include fixed identity-verification reason codes without recording message text or contact identifiers. Identity integration.
  • Recover and stop managed engines more reliably. Preserve profile state, track interrupted starts, and verify external resources have stopped before reporting shutdown complete.
  • Install from a complete Python package. Setup assets, engine kits and starter-profile metadata are included, with pinned engine downloads checked by size and SHA-256.
  • Collect reproducible evaluation results. FR443 collection, semantic judging, artifact-bound evaluation and result exports are available. Benchmarks.

The final code passed 1,083 core tests on Python 3.10 and 1,083 on Python 3.12, plus 16 offline packaging, installation and command checks. These are software and distribution checks; model and engine results retain their recorded scope. Validation · Testing policy.

The native DeepSeek Docker wizard's full-model activation remains unqualified end to end. DeepSeek's TensorFold path has separate validation. Qwen's SGLang graph path remains opt-in, with eager execution the default. See compatibility and the Qwen quality results.