Normalize PACKAGE_LINKS_ALLOW_LIST to the spelling lookups actually use

[atalman]

whl/test/cu134/typing-extensions/ does not exist, so every CUDA 13.4 smoke test on v2.14.0-rc2 fails:

Looking in indexes: https://download.pytorch.org/whl/test/cu134
ERROR: Could not find a version that satisfies the requirement
       typing-extensions>=4.10.0 (from torch) (from versions: none)

All 8 Python versions of manywheel-*-cuda13_4-test fail this way (run). from versions: none rather than a version complaint, because the smoke test installs from that single index with no PyPI fallback and the package directory simply is not there.

Root cause

typing-extensions is in PACKAGE_LINKS_ALLOW_LIST, but the membership test can never succeed.

Every lookup compares an underscore-spelled name:

• package names come from wheel filenames via obj_to_package_name — typing_extensions-4.16.0-...whl -> typing_extensions
• get_packages_to_copy_from_parent maps parent directory names through the same conversion before comparing:

# manage_v2.py:801-803
pkg_name_with_underscores = parts[0].replace("-", "_")
parent_packages.add(pkg_name_with_underscores)
...
if pkg_name.lower() in PACKAGE_LINKS_ALLOW_LIST:

But the set stored whatever spelling was typed, which for most entries is hyphenated. So "typing_extensions" in {..., "typing-extensions", ...} is False, and the copy step skips it — in every subdirectory, in every channel, always.

This had been hit before and worked around by listing both spellings, with a comment saying as much:

"cuda-bindings",
"cuda_bindings",
"cuda-toolkit",
# Underscore spelling is the one that matches; see "spmd_types" above.

Applied to four entries, missed on the rest.

Scope

77 of the 102 entries could never be copied. Only single-word names (filelock, sympy, networkx, fsspec, ...) and the four with an explicit underscore twin worked — which is exactly what is published today: cu134 has filelock/sympy/networkx/fsspec and is missing typing-extensions.

cuda-toolkit is on the broken list too, which explains the other cu134 gap reported earlier.

Fix

Normalize once at definition:

PACKAGE_LINKS_ALLOW_LIST = {
    x.lower().replace("-", "_")
    for x in [...]
}

The four underscore twins and the comment documenting the workaround become redundant, so they are dropped. 102 entries -> 97.

Verified by replaying the real code paths (parent_dir.replace("-","_") and wheel_name.split("-",1)[0]) against the parsed list:

packages copyable BEFORE: 20
packages copyable AFTER:  97
newly working: 77          (cuda_toolkit, nvidia_*, rocm_sdk_*, intel_*, onemkl_*, typing_extensions, ...)
regressions:   none
entries that can never match (contain '-'): 0

Please dry-run before applying

This makes the next index run copy ~77 additional package indexes into every arch subdirectory. That is the intended behaviour and each one is already allow-listed, but it is a large one-time change to published indexes, so it is worth running the update in dry-run first and eyeballing the plan.

Does not retroactively repair existing subdirectories

get_packages_to_copy_from_parent only copies packages "in parent but not in subdir", so a subdir is never refreshed once a package exists there. Two consequences:

• whl/test/cu134/typing-extensions/ will be created by this, since it is absent — so this does fix the rc2 breakage on the next run.
• Already-populated copies stay frozen. whl/test/cu126/typing-extensions/ still lists 5 versions topping out at 4.15.0 while the parent has 64 up to 4.16.0. Harmless while the floor is >=4.10.0, but it is a separate latent issue and not addressed here.

ruff format and ruff check clean. No test added: s3_management/ has no test harness and no CI runs pytest over it — happy to add one if you would like that set up.

[vercel]

@atalman is attempting to deploy a commit to the Meta Open Source Team on Vercel.

A member of the Team first needs to authorize it.