Suppose I have a Python virtual environment named pytorch-benchmarks (created using python3 -m venv pytorch-benchmarks). This environment is set up with the PyTorch nightly build and various other packages for running PyTorch benchmarks. The benchmarks run fine without any runtime errors.

I also have another virtual environment named pytorch-test (created using python3 -m venv pytorch-test), which I use for building PyTorch from source. There is no compatibility issue between the two versions because I check out the git commit corresponding to the nightly build in the pytorch-benchmarks environment. In this pytorch-test environment, I have multiple PyTorch builds. After building PyTorch for commit ID x, I copy the torch directory from ~/pytorch-test/lib/python3.10/site-packages to torch-x.

To switch between different PyTorch versions, I do the following:

1. Back up the original PyTorch nightly in pytorch-benchmarks under ~/pytorch-benchmarks/lib/python3.10/site-packages/torch to ~/pytorch-benchmarks/lib/python3.10/site-packages/torch-nightly.
2. Remove the torch directory in ~/pytorch-benchmarks/lib/python3.10/site-packages.
3. Create a symlink from the specific version in pytorch-test to pytorch-benchmarks:

   ln -s /home/abhishek/pytorch-test/lib/python3.10/site-packages/torch-x /home/abhishek/pytorch-benchmarks/lib/python3.10/site-packages/torch
   

My concern is that this approach of switching between various versions of PyTorch (torch-x) could lead to silent erroneous execution due to the execution of stale bytecode from .pyc files. For instance, if I switch from torch-y to torch-x, could .pyc files from the old version torch-y cause problems?

Can anyone explain in detail what might go wrong and how to prevent the execution of stale code while using the symlink approach, without creating separate virtual environments for each version?