Releases: rjojjr/torch-tuner
Releases · rjojjr/torch-tuner
v2.2.0
This release contains updated default arguments and argument descriptions as well as CPU only optimizations.
Changes
- update
--lora-dropout
CLI argument default value - update default optimizer
- update
--max-saved
CLI argument to accept 'None' to keep all checkpoint saves - optimize CPU only tuning
- add ability to serve models on CPU only
- update argument descriptions
v2.1.7
v2.1.6
This release brings multiple additional CLI arguments and optimizations.
Changes
- add
--show-token-metrics
CLI argument - fix evaluation mode
- add
--load-best-before-save
CLI argument - add
--max-parallel-requests
CLI argument for serve mode - load-dataset optimizations
v2.1.5
v2.1.4
This release include minor code optimizations and a hotfix for serve mode broken by the previous release.
Changes
- general code optimizations
- fix serve mode broken by missing dependency
- fix serve mode API request body parsing
- add serve endpoints return proper
finish_reason
v2.1.3
v2.1.2
v2.1.1
v2.1.0
This release adds Windows OS file path support, additional CLI configuration arguments and exception handling optimizations.
Changes
- add
--neftune-noise-alpha
CLI argument - add
--overwrite-output
CLI argument - refactor exceptions to be more informative
- add Windows OS file path support
- change default output directory(POSSIBLY BREAKING CHANGE)