Deprecate old TORCH_VERSION variables

[andrewor14]

Stack from ghstack (oldest at bottom):

• Replace export_for_training with torch.export.export #2724
• Remove old change_linear_weights_to_* APIs #2721
• Drop support for PyTorch 2.5 and before #2720
• -> Deprecate old TORCH_VERSION variables #2719

Summary: This commit deprecates the following variables in favor of simply calling torch_version_at_least.

TORCH_VERSION_AT_LEAST_2_8
TORCH_VERSION_AT_LEAST_2_7
TORCH_VERSION_AT_LEAST_2_6
TORCH_VERSION_AT_LEAST_2_5
TORCH_VERSION_AT_LEAST_2_4
TORCH_VERSION_AT_LEAST_2_3
TORCH_VERSION_AT_LEAST_2_2
TORCH_VERSION_AFTER_2_5
TORCH_VERSION_AFTER_2_4
TORCH_VERSION_AFTER_2_3
TORCH_VERSION_AFTER_2_2

Test Plan:

python test/test_utils.py -k torch_version_deprecation

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/2719

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 2b74165 with merge base 6cfa477 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[jerryzh168]

another thing we could fix, is that TORCH_VERSION_AT_LEAST does not match our common understanding of version naming, since:

e.g. for TORCH_VERSION_AT_LEAST("2.7.0"), commonly we might expect this to be 2.7.0 stable and 2.8.0.dev and above, and 2.7.0.dev is not greater than 2.7.0 (because after 2.6.0 release, the version will be 2.7.0.dev until 2.7.0 is released)
but currently, 2.7.0 and 2.7.0.dev will return True

[jerryzh168]

why is current torch version related to this? this is talking about deprecation of these variables right? if so, then these variables should be deprecated now, regardless of what the system torch version is?

[andrewor14]

yeah I can deprecate them all. I think these variables exist so we don't keep calling this function, but yeah we definitely don't need to expose them

[andrewor14]

Ok just deprecated them all

[jerryzh168]

OK thanks, also the deprecation doesn't need to depend on the current pytorch version I think

we could also add a cache for torch_version_at_least as well I think

[jerryzh168]

we need to check if these variables are used outside of torchao as well, TBH I'm not exactly sure if we want to deprecate these variables, also we should probably stop defining these variables, since these can be achieved by passing args to torch_version_at_least when we want to use them

[andrewor14]

e.g. for TORCH_VERSION_AT_LEAST("2.7.0"), commonly we might expect this to be 2.7.0 stable and 2.8.0.dev and above, and 2.7.0.dev is not greater than 2.7.0 (because after 2.6.0 release, the version will be 2.7.0.dev until 2.7.0 is released)
but currently, 2.7.0 and 2.7.0.dev will return True

You mean 2.7.0.dev should not be True because of the following ordering?

2.6.0 (stable) < 2.7.0.dev < 2.7.0 (stable)

But currently the code thinks:

# current state, but this is wrong
2.6.0 (stable) < 2.7.0 (stable) <= 2.7.0.dev

Is that right?

[andrewor14]

Filed an issue here: #2722. I think we should fix that separately because this PR doesn't change any semantics in its current state.

[jerryzh168]

e.g. for TORCH_VERSION_AT_LEAST("2.7.0"), commonly we might expect this to be 2.7.0 stable and 2.8.0.dev and above, and 2.7.0.dev is not greater than 2.7.0 (because after 2.6.0 release, the version will be 2.7.0.dev until 2.7.0 is released)
but currently, 2.7.0 and 2.7.0.dev will return True

You mean 2.7.0.dev should not be True because of the following ordering?

2.6.0 (stable) < 2.7.0.dev < 2.7.0 (stable)

But currently the code thinks:

# current state, but this is wrong
2.6.0 (stable) < 2.7.0 (stable) <= 2.7.0.dev

Is that right?

yeah exactly