-So wanted to create an fresh conda environment with CUDA11.0 Linux. python=3.9.19
-The only way to make it able to use CUDA with conda was with either pytorch=1.7.1 or 1.7.0 due to driver and cuda restrictions (tried mix and match cudatoolkit 11.0/10.1/10.2 with various pytorch versions, all failed, only the one indicated suitable for CUDA11.0 on pytorch official website[previous-versions] could work)
-Issue is, while CUDA10.2 allowed more recent pytorch(1.12) which could run the transformers import(but cause cuda to not work), CUDA11.0 which only allowed the 1.7.X are unable to import anything from transformers while throwing the error: Module error not found: “torch.utils._pytree”. This was attempted with various transformers versions (4.28.0,4.38.2,4.35,4.38,4.25 etc. , and latest version too as of this post). Funnily, transformers 4.28.1 yielded different error on the tokenizer not working…
Not sure how to fix this as I am currently unable to import anything in transformers due to this bug.