Why

• by default, the CUDA version of torch is installled which takes up a bunch of space, yet we don't support GPUs on replit
  • we also have similar overrides for tensorflow for this reason:

    upm/internal/backends/python/pypi_map.override.go

    Line 3 in a3cbf8f

    var tf_variants = [...]string{"tensorflow-cpu", "tensorflow-cpu-aws", "tensorflow-gpu", "tensorflow-intel", "tf-nightly", "tf-nightly-cpu", "tf-nightly-cpu-aws", "tf-nightly-gpu", "tf-nightly-intel"}

• this leads to deployments failing because their image size is too large, let's try using the cpu version instead

What changed

• unfortunately we can't do the override in the same way as tensorflow since the cpu version of torch isn't available on pypi
• instead, add extra indexes that will be checked before pypi, described here: https://docs.astral.sh/uv/configuration/indexes/

Test plan

• make upm on a linux machine
• make a python project that uses uv
• do upm add torch and see the following output, notably torch==2.5.1+cpu:

--> uv add torch --index https://download.pytorch.org/whl/cpu --index-strategy unsafe-best-match
Resolved 11 packages in 513ms
Installed 10 packages in 135ms
 + filelock==3.16.1
 + fsspec==2024.10.0
 + jinja2==3.1.4
 + markupsafe==3.0.2
 + mpmath==1.3.0
 + networkx==3.4.2
 + setuptools==75.6.0
 + sympy==1.13.1
 + torch==2.5.1+cpu
 + typing-extensions==4.12.2