v0.2.0 1.10.0
dmtrKovalenko/odiffSep 2, 2026 → Sep 4, 20262 days between releases
What's New
Newer release introduces GPU batched linear-assignment solver and optimizer parameter merging, whereas the older version focused on AI API gateway features.
Breaking Changes
- Validation evaluates one model per epoch instead of two
- grad_accum_steps defaults to 1 instead of 4
- Optimizer parameter groups collapse from one-per-parameter to one-per-hyperparameter
- Dataset builders validate config instead of silently falling back to wrong defaults
- log_per_class_metrics defaults to False
New Features
- GPU batched linear-assignment solver for CUDA devices
- Optimizer parameter group merge for speedup
- Packed targets and draft JPEG decoding for DataLoader throughput
Improvements
- Training steps are ~23% faster at default batch size
- predict() is up to 46% faster on large input frames via zero-copy uint8 transfer
- 1.25–2.26× DataLoader throughput on input-bound runs
Bug Fixes
- Fixed WebSocket handling when receiving terminal event before closing
- Fixed model-level cooling incorrectly converting model_not_found to 429
- Fixed Anthropic fallback being passed through even when server-side-fallback beta is disabled
Upgrade Notes
Read the Migration guide before you retrain. The validation metrics now describe the EMA model only, and effective batch size has changed due to default grad_accum_steps shifting from 4 to 1.
The Two Releases
Sub2API 0.2.0
Sep 2, 2026· 2 months ago
Adds OpenAI Fast grouping strategy, model-specific reasoning effort configuration, and native Kimi Responses API forwarding support to the AI API Gateway platform.
Full release notesv1.10.0: Faster Training & Inference
Sep 4, 2026· 2 months ago
A training-focused release that reduces wall-clock time by 44% through optimizer parameter merging and GPU batched linear-assignment solvers, while introducing several behavioral changes to validation and defaults.
Full release notes