[Distributed] Bump torch version

[kwen2501]

Distributed inference (pipeline parallel part, to be specific) requires two features landed in pytorch nightly:

• [Pipelining] Make PipelineStage support meta initialization

• [Pipelining] Allow non-0 stages to accept kwargs

Thus bumping the torch version to 2.6.0.dev20240925.

Tested locally.

Cc: @Jack-Khuu @lessw2020

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchchat/1225

• 📄 Preview Python docs built from this PR

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

❌ 1 New Failure

As of commit 5e3c3ce with merge base 24d00ea ():

NEW FAILURE - The following job has failed:

• pull / test-torchao-experimental (macos-14-xlarge) (gh)
  torch._dynamo.exc.Unsupported: 'skip function dequantize_per_channel_group in file /Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/site-packages/torch/ao/quantization/fx/_decomposed.py'

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

[lessw2020]

we need to bump the nightly vision version to avoid an error.
i.e. this:
VISION_NIGHTLY_VERSION=dev20240901

Bumping to same as PYTORCH_NIGHTLY_VERSION fixes this error for me.
See inline comment.

[Jack-Khuu]

Vision and PT being on the same nightly pin usually solves this

[Jack-Khuu]

I'll also be bumping the version over on ExecuTorch pytorch/executorch#5549

I can match you at 925

[kwen2501]

@metascroy Wondering if you could have a look at the CI issue? Thanks!

[metascroy]

@metascroy Wondering if you could have a look at the CI issue? Thanks!

At an initial glance, I'm not sure why it is failing. I'll investigate more after lunch, but I'm wondering if PyTorch changed how custom ops work in the pin bump.

[kwen2501]

Cc @zou3519 any recent changes in custom op support that may be relevant to the CI failure here?

[zou3519]

Maybe @jerryzh168 ?

[metascroy]

@metascroy Wondering if you could have a look at the CI issue? Thanks!

At an initial glance, I'm not sure why it is failing. I'll investigate more after lunch, but I'm wondering if PyTorch changed how custom ops work in the pin bump.

@kwen2501 I figured out the issue.

ExecuTorch uses PyTorch 9/1, so when ExecuTorch installs, it installs 9/1 and the two different PyTorch versions cause an issue. In #1235, I bump the PT pin, but remove the ET parts of the test and it passes.

So before PT pin can be updated, it needs to be updated in ET first and the PT/ET pin then need to be updated together in torchchat cc @Jack-Khuu

[kwen2501]

Thanks @metascroy !

I'll also be bumping the version over on ExecuTorch pytorch/executorch#5549

I can match you at 925

@Jack-Khuu Looks like this bump here needs to wait for your bump in ET to land first :)