It looks like the requirements.txt and requirements_py26.yaml define slightly different version ranges. Are both of these files supposed to be installed into the same environment, or do they provide two alternative options for installing the dependencies?

I would recommend to define

• one requirements.txt file pinning the major dependencies (e.g., tirex package version, pytorch) - the current requirements.txt file is already perfect
• one file with an exact snapshot of the dependencies used to run the benchmark (e.g., conda environment with exact == package versions used during environment, or a uv lock file https://docs.astral.sh/uv/pip/compile/#locking-requirements)

The naming requirements_py26.yaml is also a bit strange as it seems to imply that these a requirements for Python 2.6.

So in general it should work just installing the requirements.txt, however, we recommend installing TiRex inside the provided conda environment as we observe some people have problems with CUDA stuff otherwise. (TiRex uses custom CUDA kernels of the sLSTM)

I relaxed the major dependencies lock of the requirements.txt, and added a conda export of the environment on which i ran the test (freezed_test_environment.yaml)

In my opinion, it's okay to only keep requirements.txt and the frozen_test_environment.yaml. I would recommend pointing to the official installation instructions in the TiRex repo (the corresponding environment.yaml). This way neither you nor we will need to worry if the current requirements.yaml file stored in the fev repo is up-to-date.