Overview
This release adds mixed precision support to PPO and Distillation along with a number of bug fixes.
Full Changelog: v5.4.2...v5.5.0
Added
- Adds bfloat16 mixed precision to PPO and Distillation by @lzyang2000 in #219
- Adds a warning for wasted rollout steps in distillation
Fixed
- Fixes missing class names from the logged configuration by @ClemensSchwarke in #230
- Synchronizes normalization statistics across GPUs by @NeoZng in #226
- Updates normalizers after optimization by @NeoZng in #225
- Fixes the LSTM hidden state type returned by storage
- Fixes the term order in console logs
Breaking Changes
logger_typegot removed from theLoggerclass. Please check the type of thewriterinstead.
New Contributors
- @NeoZng made their first contribution in #226
- @lzyang2000 made their first contribution in #219