Skip to content

Latest commit

 

History

History
55 lines (48 loc) · 2.51 KB

CHANGELOG.md

File metadata and controls

55 lines (48 loc) · 2.51 KB

TODO

  • [general] NSR loss mode?
  • [falor] Add tests
  • [lockd] Idea - add some options to customize decomposition process (e.g, type of deconvolution decomposition)
  • [trainer] Add requirements.txt
  • [trainer] Check device switching handling (cpu decomposition, gpu training etc.)
  • [trainer] Specify device via config
  • [trainer_llm] Remove temporarily added llm_pipelines from build of trainer_llm

DONE

  • [dwain] Add reduction_factor and max_accepted_ppl_diff parameters in dwain method
  • [dwain] Clean-up metrics logging

ptdeco llm trainer 0.1.0

  • [trainer_llm] Implements decomposing llms with dwain method and finetuning of

ptdeco 0.5.0

  • [dwain] Add new decomposition method - dwain

ptdeco 0.4.0, ptdeco vision trainer 0.8.0

  • [general] Implement and test building a wheel package
  • [general] Move to src project layout
  • [utils] Add logging num of decomposed modules of each type in apply_decompose_config_in_place
  • [falor] Fix nsr loss bug
  • [lockd] Improve typing in losses
  • [trainer] Add timing to every run
  • [trainer] Add saving final decompose config and state dict of finetuned model
  • [trainer] Fix typing
  • [trainer] Move notebooks to a separate dir

ptdeco vision trainer 0.7.0

  • [tranier] Add logging state dict stats
  • [trainer] Fix state dict bug
  • [trainer] Unify logged stats of decomposed models in lockd/falor/finetune

ptdeco 0.3.0, ptdeco vision trainer 0.6.0

  • [falor] Refactor fal -> falor and move module to separate directory
  • [general] Refactor wrapping class names
  • [lockd] Get rid of needless module copies in lockd wrapping
  • [lockd] Refactor - split ptdeco.py into multiple modules
  • [lockd] Remove uwrap_in_place
  • [trainer] Refactor compile config to Optional in trainer

ptdeco 0.2.0, ptdeco vision trainer 0.5.0

  • [falor] Add decomposition params handling through config
  • [general] Add checking if all blacklisted modules are present in the model
  • [general] Clean-up configurator getters to make use of pydantic models
  • [general] Switch from dict to pydantic models
  • [lockd] Fix lockd decomposition script
  • [general] Add pydantic validators to trainer
  • [general] Refactor replace submodule in place
  • [lockd] Add per module metadata in decompose config
  • [falor] Add per module metadata in decompose config
  • [direct] Add creation of decompose config
  • [direct] Add reporting flops/params before and after decomposition