Incompatibility with PyTorch

[ElliottKasoar]

Most MLIP packages that I'm aware of are built on PyTorch. However, this seems to lead to conflicts with tensorflow[and-cuda] due to differing nvidia-cublas-cu12 requirements, e.g.:

      we can conclude that torch>2.2,<2.5.0 depends on nvidia-cublas-cu12{platform_machine == 'x86_64' and sys_platform == 'linux'}==12.1.3.1.
      And because torch>=2.5.0,<=2.6.0 depends on nvidia-cublas-cu12{platform_machine == 'x86_64' and sys_platform == 'linux'}==12.4.5.8 and
      nvidia-cublas-cu12{platform_machine == 'x86_64' and sys_platform == 'linux'}==12.6.4.1, we can conclude that nvidia-cublas-cu12!=12.1.3.1,
      torch>2.2,<2.8.0, all of:
          nvidia-cublas-cu12{platform_machine == 'x86_64' and sys_platform == 'linux'}<12.4.5.8
          nvidia-cublas-cu12{platform_machine == 'x86_64' and sys_platform == 'linux'}>12.4.5.8,<12.6.4.1
          nvidia-cublas-cu12{platform_machine == 'x86_64' and sys_platform == 'linux'}>12.6.4.1
       are incompatible.
      And because torch>=2.8.0 depends on nvidia-cublas-cu12{platform_machine == 'x86_64' and sys_platform == 'linux'}==12.8.4.1 and
      tensorflow[and-cuda]==2.19.1 depends on nvidia-cublas-cu12==12.5.3.2, we can conclude that tensorflow[and-cuda]==2.19.1 and torch>2.2 are
      incompatible.
      And because we know from (8) that torch>2.2 and all of:
          tensorflow[and-cuda]<2.5.0
          tensorflow[and-cuda]>2.15.1,<2.19.1
       are incompatible, we can conclude that all of:
          tensorflow<2.5.0
          tensorflow>2.15.1
      , torch>2.2, tensorflow[and-cuda]<2.20 are incompatible.
      And because tensorpotential[cuda] depends on tensorflow[and-cuda]<2.20 and your project depends on tensorflow>2.16,<2.20, we can conclude that
      your project, tensorpotential[cuda], torch>2.2 are incompatible.
      And because your project depends on torch>2.2 and your project requires tensorpotential[cuda], we can conclude that your project's requirements
      are unsatisfiable.

This inhibits running multiple models and understanding their performance.

Would it be possible to move tensorflow[and-cuda] to an optional extra, along these lines: master...ElliottKasoar:grace-tensorpotential:master?

This still wouldn't allow running on GPU conflict-free, but would at least allow simple comparisons.