v1.0.147 1.10.0

high risk

0xNyk/council-of-high-intelligenceSep 3, 2026 → Sep 4, 20261 day between releases

What's New

Major performance release with default parameter changes and behavioral shifts

Breaking Changes

  • `grad_accum_steps` defaults to `1` instead of `4`
  • Optimizer parameter groups collapse from per-parameter to per-hyperparameter
  • Dataset builders now validate their config instead of falling back
  • `log_per_class_metrics` defaults to `False`
  • `compute_val_loss` defaults to `"auto"`

New Features

  • GPU batched linear-assignment solver
  • Optimizer parameter group merge
  • Validation evaluates one model per epoch

Improvements

  • Up to 44% faster training and 46% faster prediction
  • Optimizer parameter group merge reduces per-step cost by 23%
  • DataLoader throughput increased via packed targets and draft JPEG decoding
  • GPU batched linear-assignment solver for CUDA

Upgrade Notes

Review the Migration guide before retraining to understand behavioral changes. Adjust `grad_accum_steps` if expecting the previous default of 4. Expect potential config validation errors if previous configs relied on fallback behavior.

The Two Releases

v1.0.147

v1.0.147

Sep 3, 2026· 2 months ago

Minimal release notes: "**Full Changelog**: https://github.com/iOfficeAI/OfficeCLI/compare/v1.0.146...v1.0.147"

Full release notes
1.10.0

v1.10.0: Faster Training & Inference

Sep 4, 2026· 2 months ago

A major performance release significantly faster training and inference, accompanied by several behavioral changes.

Full release notes