Migration to PyTorch core

[jansel]

We have merged github.com/pytorch/torchdynamo into github.com/pytorch/pytorch. Being in a separate repo has allowed us to move fast, but it creates some installation roadblocks for users and version pinning has been a growing headache as the project scales.

• import torchdynamo is now import torch._dynamo
• import torchinductor is now import torch._inductor
• The standalone C extensions are added to torch._C
• torchdynamo/benchmarks is now pytorch/benchmarks/dynamo
• torchdynamo/test/dynamo is now pytorch/test/dynamo
• torchdynamo/test/inductor is now pytorch/test/inductor

Instructions for migrating your PR from pytorch/torchdynamo to pytorch/pytorch

1. Merge/rebase your PR to the correct hash of TorchDynamo: 6380959

cd ../torchdynamo
git pull origin 6380959be21851bfda99424392cc08fda29d073d

2. Checkout a new branch with the correct version of PyTorch: pytorch/pytorch@054a2fd
   (the two versions must match so only your changes are copied over):

cd ../pytorch
git fetch origin master 
git checkout -b <your_new_branch_name> 054a2fd6c2fc361796663eed4772368c287d6c83

3. Run the rsync+auto-format script

bash ../torchdynamo/copy_to_core.sh

4. Commit your changes, rebase to the latest version of PyTorch, and create a pull request as normal for the PyTorch repo.

If you are copying files by hand, make sure to run make setup_lint && lintrunner -a to switch to the new auto-formatting.

The Plan

Phase 1 (complete)

In the first phase we will move the code over to core, which consists of these two PRs:
#1538
pytorch/pytorch#86461

In this phase (now complete), the dynamo source of truth remained pytorch/torchdynamo. Any changes will be copied over by the copy_to_core.sh script which we will rerun at the end of phase one.

After these initial PRs land, dynamo will be working in core -- but most of the testing will be disabled. The initial PR will have testing disable, but then we will enable tests in smaller followup PRs.

Phase 2 (going on now)

At this switchover point, the source of truth will become PyTorch core. All open PR's will be closed and asked to resubmit to pytorch/pytorch. We will replace torchdynamo in this repo with from torch._dynamo import *, and same for torchinductor.

The criteria for reaching this point is the PRs above land in core, sync to fbcode (to check for reverts), and basic testing infra is setup.

We will continue using this project for github issues.

Followup Tasks

• Land initial migration in core (this blocks most other items) @jansel
• Land dynamo size of that migration with copy_to_core.sh script @jansel
• Enable all unit tests in core @jansel
• Add smaller deps such as Sympy/Filelock and to PyTorch CI
• test_recompile_ux.py has a bug where it only works in pytest and not using PyTorch's test infra @wconstab
• Add Triton to PyTorch CI Add triton to CI pytorch#86988
• Update torchdynamo tests in core to use torch._dynamo and stop installing torchdynamo Migrate dynamo CI test shards to torch._dynamo pytorch#87039
• Figure out how to include triton/etc properly for end users installing nightlies. Perhaps vend our own version of it in the short term. @malfet
• Update performance dashboards to use the new system. @anijain2305
• Move docs over to pytorch core (these won't be covered in the initial move)
• Update internal fbcode buck target and internal users to use the new packages, remove the old one
• Port over our circleci integration tests, we may need to convert these to github actions in the process. Since these are expensive, we may want to gate them on a tag. @desertfire
• Fix followup issues uncovered by @albanD, @voznesenskym

I am sure there are other things I am forgetting. Please edit this or comment with suggestions.

[wconstab]

what's the best way to run/repro the test_recompile_ux bug?

[voznesenskym]

One more action item:

I plan to apply @albanD's suggestions and feedback to the dynamo eval_frame code.

Another up for grabs:

Windows test enablement

And another:

Py3.11 eventual support

[jansel]

what's the best way to run/repro the test_recompile_ux bug?

1. Checkout Changes related to moving to PyTorch core #1538
2. Uncomment these lines:
   

   torchdynamo/test/dynamo/test_recompile_ux.py

   Lines 201 to 204 in 151fc25

   	# TODO(jansel): these pass with pytest, but not with pytorch CI
   	# if __name__ == "__main__":
   	# from torchdynamo.testing import run_tests
   	# run_tests()

3. python test/dynamo/test_recompile_ux.py should fail (while pytest passes)

[jansel]

Windows test enablement

pytorch/pytorch#93736

Py3.11 eventual support

#106

[malfet]

Consider splitting pytorch/pytorch#86461 into two parts: native plugins and pure python code.

[jansel]

Consider splitting pytorch/pytorch#86461 into two parts: native plugins and pure python code.

This is done with pytorch/pytorch#86621

[wconstab]

what's the best way to run/repro the test_recompile_ux bug?

@jansel is it important for torchdynamo.git/test/dynamo/test_recompile_ux.py to pass? I did scratch my head debugging this and was not able to figure out why the logger config is not working there. However, pytorch.git/test/dynamo/test_recompile_ux.py does pass for me. Isn't that all that matters at this point?

[jansel]

I think I disabled the test, so it might be passing because of that. If it is too hard I am ok just disabling it.

[wconstab]

I double checked my tmux scrollback and i did run it with -v and see that the tests really ran. (Ran 6 tests in 2.375s) -- i'm pretty sure something is messed up with the version of the test on the torchdynamo repo i think things are ok on the pytorch side if you just uncomment that block

-# if __name__ == "__main__":
-#     from torch._dynamo.testing import run_tests
-#     run_tests()

I am also ok with it being disabled transiently, so if you find that enabling it doesn't work just disable and file an issue on me, i'll debug it once the transition has settled.

[malfet]

Do we need followup task for fixing internal build config to point to the "right" version for torchdynamo?

[jansel]

That is what I meant by "Update internal fbcode buck target and internal users to use the new packages, remove the old one" in the list above