Because TextGenerationTGIS is not a subclass but a backend type for TextGeneration, any additional tasks declared on the former were not actually getting added. By adding the additional tasks on TextGeneration and PeftPromptTuning respectively, this allows the tokenization tasks to be available on the TGIS backend implementations. Unimplemented functions have to be added or there will be errors.
Tokenization run functions could eventually be implemented since each LLM could be reasonably expected to have a tokenizer.
Std getting imported with LaunchConfig started to error in latest torch 2.3.0. It was observed the Std object is at https://github.com/pytorch/pytorch/blame/main/torch/distributed/elastic/multiprocessing/api.py and not in the LaunchConfig and was potentially getting imported through another import to the launcher api. tee is no longer an arg on LaunchConfig. Since this was a breaking change, torch is now pinned below 2.3.0.
Closes https://github.com/caikit/caikit-nlp/issues/350
Because
TextGenerationTGIS
is not a subclass but a backend type forTextGeneration
, any additional tasks declared on the former were not actually getting added. By adding the additional tasks onTextGeneration
andPeftPromptTuning
respectively, this allows the tokenization tasks to be available on the TGIS backend implementations. Unimplemented functions have to be added or there will be errors.Tokenization run functions could eventually be implemented since each LLM could be reasonably expected to have a tokenizer.
Std
getting imported withLaunchConfig
started to error in latest torch2.3.0
. It was observed theStd
object is at https://github.com/pytorch/pytorch/blame/main/torch/distributed/elastic/multiprocessing/api.py and not in theLaunchConfig
and was potentially getting imported through another import to the launcher api.tee
is no longer an arg onLaunchConfig
. Since this was a breaking change, torch is now pinned below2.3.0
.