Why not just conda?

If you are thinking our [cuda] to pypi dependencies pipeline is "just reinventing conda", you are absolutely right.

Combo detection, torch-family pinning, resolving wheel URLs by scraping a PEP 503 page, +cuXtorchY local-version tags, #sha256= fragment plumbing... all of it exists because the CUDA/torch variant axis does not exist in PyPI's model, and conda's model has it natively.

A conda solver does every one of those jobs for free, and comfy-env would gladly delegate them, but the PyTorch team does not publish conda packages.

I feel like this is quite egregious, given torch is the poster child for the exact problems conda exists to solve (bundled libomp copies, import-numpy-before-torch-or-was-it-the-other-way-around native loading order).

A personal note about the PyTorch situation because I'm really not happy about it

PyTorch saying "we will shut down conda support because only 5% of our downloads come through there" is like a hospital saying: "Only 5% of our patients arrive here by ambulance, so ambulances are clearly low ROI and we shouldn't support them anymore"

Conda-forge's community builds, healthy as they are, are ~30% of what is published on pypi.

Compiling torch for ALL python/os/cuda versions that it is compiled for in pypi is one of the items in the roadmap.