🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/111313

• 📄 Preview Python docs built from this PR
• 📄 Preview C++ docs built from this PR
• ❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

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

❌ 1 New Failure

As of commit 0689a86 with merge base 8da2f81 ():

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

The committers listed above are authorized under a signed CLA.

• ✅ login: stellaraccident / name: Stella Laurenzo (1259156, 684846c)
• ✅ login: nirvedhmeshram / name: Nirvedh Meshram (a957bb9, c335c9a, 3ef58ab, 9df7198, ed1ae3c, ad35ab8, 9484469, 9c5d56c, 7de8897, dbffaba, ee4dd33, 0689a86)

Oh this is a good point. I think it mostly reflects that Triton does not work on Windows. We should move the check to the Inductor backend and not put it at the top level for Dynamo.

Added the check to the inductor backend. I imagine that there are also some tests that can have their Windows excludes removed with this?

(took a quick look and it seems reasonable as-is)

@pytorchbot label "topic: not user facing"

Looks like lints are failing, you should be able to fix with:

make setup_lint
lintrunner -a

Thanks. Applied the lint patch.

@pytorchbot rebase

@pytorchbot started a rebase job onto refs/remotes/origin/viable/strict. Check the current status here

Successfully rebased dynamo_works_on_windows onto refs/remotes/origin/viable/strict, please pull locally before adding more changes (for example, via git checkout dynamo_works_on_windows && git pull --rebase)

@pytorchbot merge

Merge started

Your change will be merged once all checks pass (ETA 0-4 Hours).

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Merge failed

Reason: 1 jobs have failed, first few of them are: trunk / win-vs2019-cpu-py3 / test (default, 2, 3, windows.4xlarge.nonephemeral)

@stellaraccident the test failure above looks real

Thanks - can add it to my list to triage. Does anyone know how to trigger these specific tests on presubmit?

From the error log:

torch._dynamo.exc.Unsupported: 'skip function cond in file C:\actions-runner\_work\pytorch\pytorch\build\win_tmp\build\torch\_higher_order_ops\cond.py'', skipped according skipfiles.SKIP_DIRS'

It looks like the forced inlined file torch._higher_order_ops.cond doesn't get inlined correctly. Can anyone run this unit tests with TORCH_LOGS="+dynamo" and share output logs here?

Here is a gist of the output logs https://gist.github.com/nirvedhmeshram/69e1f8499e3c2a8341563df7e4ae6182

Can you run the the following code and share the output?

>>> import torch
>>> torch.cond.__code__
<code object cond at 0x7f44d43c5a50, file "/home/ybliang/local/pytorch/torch/_higher_order_ops/cond.py", line 36>
>>> from functorch.experimental.control_flow import cond
>>> cond.__code__
<code object cond at 0x7f44d43c5a50, file "/home/ybliang/local/pytorch/torch/_higher_order_ops/cond.py", line 36>

Can you run the the following code and share the output?

>>> import torch
>>> torch.cond.__code__
<code object cond at 0x7f44d43c5a50, file "/home/ybliang/local/pytorch/torch/_higher_order_ops/cond.py", line 36>
>>> from functorch.experimental.control_flow import cond
>>> cond.__code__
<code object cond at 0x7f44d43c5a50, file "/home/ybliang/local/pytorch/torch/_higher_order_ops/cond.py", line 36>

looks ok to me

>>> import torch
>>> torch.cond.__code__
<code object cond at 0x000001F241E78DB0, file "D:\shark_experiments\pytorch\torc
h\_higher_order_ops\cond.py", line 36>
>>> from functorch.experimental.control_flow import cond
>>> cond.__code__
<code object cond at 0x000001F241E78DB0, file "D:\shark_experiments\pytorch\torc
h\_higher_order_ops\cond.py", line 36>

@nirvedhmeshram This is weird, as we already put it in the force inline list:
https://github.com/pytorch/pytorch/blob/23b030a79caf88007eb7a17eaa14e2368eed8b06/torch/_dynamo/skipfiles.py#L173C12-L173C29

And it works well on other platform, not sure why it was wrapped as SkipFilesVariable on win, which is supposed to UserFunctionVariable.

@nirvedhmeshram This is weird, as we already put it in the force inline list: https://github.com/pytorch/pytorch/blob/23b030a79caf88007eb7a17eaa14e2368eed8b06/torch/_dynamo/skipfiles.py#L173C12-L173C29

And it works well on other platform, not sure why it was wrapped as SkipFilesVariable on win, which is supposed to UserFunctionVariable.

Its because of many places using "/" as seperator in file paths while windows need a different seperator so indeed the kind of fixes done here https://github.com/pytorch/pytorch/pull/109677/files are passing the tests for me. I need to apply the fixes properly and then will push up the fix, hopefully in a few hours.

Its because of many places using "/" as seperator in file paths while windows need a different seperator so indeed the kind of fixes done here https://github.com/pytorch/pytorch/pull/109677/files are passing the tests for me. I need to apply the fixes properly and then will push up the fix, hopefully in a few hours.

That makes sense, but I'm curious why our CI doesn't capture this, since the skip/inline rules are used everywhere.

I think this is dynamo specific (at least the changes I am making are all in /torch/_dynamo) and dynamo was disabled on windows so it never hit these issues.

I had to disable test_export_then_compile_tensor_ctor test for windows as it is using inductor which is not set up to find the cpp compiler as of now. Since this patch is trying to enable dynamo without trying to get inductor working, I hope that is okay. (Will ask for review after the CI passes.)

Looking at the CI failure here,

Public FX API(s) ['torch.fx.passes.backends.cudagraphs.CudaGraphsSupport', 'torch.fx.passes.backends.cudagraphs.partition_cudagraphs', 'torch.fx.passes.infra.partitioner.CapabilityBasedPartitioner', 'torch.fx.passes.infra.partitioner.Partition'] introduced but not given a backwards-compatibility classification! Please decorate these API(s) with `@torch.fx._compatibility.compatibility` to specify BC guarantees.

can these decorators simply be added as suggested, or is there more to this failure that isn't being addressed in the logs?

Closing because we think that #115969 addresses this.