Every system I’ve seen for managing this kind of system has flaws, including the ones that I have written.
For instance scikit-learn implements excellent algorithms for model selection that would apply, in principle, to a model based on huggingface transformers that might take 2 hours to train. skl is a fast machine if memory fits in RAM on a single computer, but it is not up to task for multiple computers or anything mortal to a single process such as the computer bring turned off.
HF has model selection algorithms too, but not as nice. They don’t take the same kind of datasets as all so it would be a hassle to import my ski models into HF.
I have to be able to compare models generated with any kind of tools so I think I will build a universal model selection framework (builds and test models) but then you run into the problems langchain did where there is a lot of structure imposed and all sorts of quirks and performance losses because of that structure.
For instance my current skl selector wastes a lot of resources computing stuff from scratch over and over again and if the code were properly organized it could get the job done 3 times faster but the same trick wouldn’t work for every other experiment I might want to do.
So we are all running into hurdles and finding ways to jump over them, making a lot of mistakes because we are in a rush and don’t know better yet.
For instance scikit-learn implements excellent algorithms for model selection that would apply, in principle, to a model based on huggingface transformers that might take 2 hours to train. skl is a fast machine if memory fits in RAM on a single computer, but it is not up to task for multiple computers or anything mortal to a single process such as the computer bring turned off.
HF has model selection algorithms too, but not as nice. They don’t take the same kind of datasets as all so it would be a hassle to import my ski models into HF.
I have to be able to compare models generated with any kind of tools so I think I will build a universal model selection framework (builds and test models) but then you run into the problems langchain did where there is a lot of structure imposed and all sorts of quirks and performance losses because of that structure.
For instance my current skl selector wastes a lot of resources computing stuff from scratch over and over again and if the code were properly organized it could get the job done 3 times faster but the same trick wouldn’t work for every other experiment I might want to do.
So we are all running into hurdles and finding ways to jump over them, making a lot of mistakes because we are in a rush and don’t know better yet.