[Bugfix][Dynamo] Fix Sparse tensors by graph break in Dynamo by Lucaskabela · Pull Request #164873 · pytorch/pytorch

Lucaskabela · 2025-10-07T21:04:15Z

Fixes #164823 by making lack of support for sparse tensors very explicit (in fake tensor, inductor, and lowering code)

cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben

pytorch-bot · 2025-10-07T21:04:19Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/164873

📄 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

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

✅ No Failures

As of commit d9b2166 with merge base 14af1dc ():
💚 Looks good so far! There are no failures yet. 💚

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

Lucaskabela · 2025-10-07T21:07:32Z

@pytorchbot label "topic: not user facing"

Lucaskabela · 2025-10-09T15:43:36Z

There are perhaps two ways of going about this;

We can graph break on any sparse tensor in graph capture (approach taken here)
We can error inside inductor itself when we see sparse tensors

I am partial to 1, since this is more defensive and clear for users; however we then preclude folks using dynamo/aot_eager for sparse tensor capture. However, I think this is fine since we already have a handful of graph breaks for sparse tensors and operations today (see

pytorch/torch/_dynamo/variables/builtin.py

Line 2336 in 724463d

if isinstance(obj, TensorVariable):

)

williamwen42 · 2025-10-09T20:13:33Z

torch/_dynamo/variables/builder.py

+            not tx.export or not config.capture_sparse_compute
+        ):
+            unimplemented_v2(
+                gb_type="Attempted to wrap sparse Tensor",


Suggested change

gb_type="Attempted to wrap sparse Tensor",

gb_type="Attempted to wrap sparse Tensor with VariableTracker",

williamwen42 · 2025-10-09T20:13:49Z

torch/_dynamo/variables/builder.py

+        ):
+            unimplemented_v2(
+                gb_type="Attempted to wrap sparse Tensor",
+                context="",


Suggested change

context="",

context=str(example_value),

StrongerXi · 2025-10-09T20:20:12Z

If it works with aot_eager, I'd expect there's a way for inductor to skip the sparse ops during codegen and run them via aten calls... Did you check with inductor folks?

Lucaskabela · 2025-10-09T22:42:21Z

Sure let me check - cc @laithsakka @eellison

eellison

Should this be in FakeTensor instead ? E.g if an operator decomposed to use sparse tensors, or had sparsity in the backward. See

pytorch/torch/_subclasses/fake_impls.py

Line 1081 in 70925bd

"torch.compile does not support strided NestedTensor"

eellison · 2025-10-10T12:59:31Z

@StrongerXi yes, there is this check to fallback for unsupported tensor

pytorch/torch/_inductor/lowering.py

Lines 2136 to 2139 in 70925bd

    
           if t.is_complex(): 
        
               # Complex views are supported with IR ComplexView 
        
               _warn_complex_not_supported() 
        
               return True

Lucaskabela · 2025-10-10T17:32:58Z

@StrongerXi yes, there is this check to fallback for unsupported tensor

pytorch/torch/_inductor/lowering.py

Lines 2136 to 2139 in 70925bd

if t.is_complex():

# Complex views are supported with IR ComplexView

_warn_complex_not_supported()

return True

Okay - let me see if we can move here as opposed to another graph break :)

Lucaskabela · 2025-10-13T15:31:54Z

@pytorchmergebot merge -i "Test failure from cuda error not related to changes"

pytorch-bot · 2025-10-13T15:31:57Z

❌ 🤖 pytorchbot command failed:

@pytorchbot: error: unrecognized arguments: Test failure from cuda error not related to changes

usage: @pytorchbot [-h] {merge,revert,rebase,label,drci,cherry-pick} ...

Try @pytorchbot --help for more info.

Lucaskabela · 2025-10-13T15:33:21Z

@pytorchmergebot merge -i

pytorchmergebot · 2025-10-13T15:35:15Z

Merge started

Your change will be merged while ignoring the following 1 checks: inductor / inductor-test / test (inductor_torchbench, 1, 2, linux.g5.4xlarge.nvidia.gpu)

Learn more about merging in the wiki.

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

Advanced Debugging

Check the merge workflow status
here

pytorchmergebot · 2025-10-13T16:44:27Z

Merge failed

Reason: 1 jobs have failed, first few of them are: trunk / linux-jammy-cuda12.8-py3.10-gcc11 / test (default, 3, 5, lf.linux.g6.4xlarge.experimental.nvidia.gpu)

Details for Dev Infra team

Raised by workflow job

Lucaskabela · 2025-10-13T22:31:54Z

test/inductor/test_cudagraph_trees.py


-            for _ in range(3):
-                self.assertEqual(foo_opt(view_6, buf31), foo(view_6, buf31))
+            with self.assertRaises(torch._dynamo.exc.BackendCompilerFailed):


This was previously not supported but just silently didn't cudagraph (see

pytorch/torch/_inductor/ir.py

Line 5829 in f368345

if isinstance(t, torch.Tensor) and t.is_sparse:

)

Letting users know loudly seems preferable

pytorchmergebot · 2025-10-15T23:37:55Z

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

pytorchmergebot · 2025-10-15T23:38:00Z

Successfully rebased lucaskabela/fix_164823 onto refs/remotes/origin/viable/strict, please pull locally before adding more changes (for example, via git checkout lucaskabela/fix_164823 && git pull --rebase)

Lucaskabela · 2025-10-16T14:58:39Z

@pytorchmergebot merge

pytorchmergebot · 2025-10-16T15:00:37Z

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

Advanced Debugging

Check the merge workflow status
here

…#164873) Fixes pytorch#164823 by making lack of support for sparse tensors very explicit (in fake tensor, inductor, and lowering code) Pull Request resolved: pytorch#164873 Approved by: https://github.com/williamwen42, https://github.com/eellison, https://github.com/mlazos

pytorch-bot bot added ciflow/inductor module: dynamo labels Oct 7, 2025

pytorch-bot bot added topic: not user facing topic category module: inductor labels Oct 7, 2025

Lucaskabela force-pushed the lucaskabela/fix_164823 branch from 7963d0e to d778061 Compare October 8, 2025 22:57

Lucaskabela marked this pull request as ready for review October 9, 2025 17:28

Lucaskabela requested review from StrongerXi, anijain2305 and williamwen42 October 9, 2025 17:29

Lucaskabela changed the title ~~[Draft][Bugfix] Fix by graph break on sparse output ops~~ [Bugfix][Dynamo] Fix Sparse tensors by graph break in Dynamo Oct 9, 2025

williamwen42 approved these changes Oct 9, 2025

View reviewed changes

Lucaskabela requested review from eellison and laithsakka October 9, 2025 22:42

eellison approved these changes Oct 10, 2025

View reviewed changes

pytorch-bot bot added the ciflow/trunk Trigger trunk jobs on your pull request label Oct 13, 2025

pytorchmergebot added the merging label Oct 13, 2025

pytorchmergebot removed the merging label Oct 13, 2025

Lucaskabela commented Oct 13, 2025

View reviewed changes

Lucaskabela added 17 commits October 15, 2025 23:37

Fix by graph break on sparse output ops

3a474ce

format and fix tests

d10f3c1

fix i32 sparse too

2309b7b

Fix graph break for consistency

ed32d0b

Fix suggestions on graphbreak message

b876cce

Minimize to inductor/fake tensor level

eea0c14

Ensure fix localized to inductor backend

7bafd31

Formalize support

8275a85

Update test to remove comments'

b2de91c

Redisable tests we expect to graphbreak on

ed1d854

formatting

072bff7

Fix failing sparsity test (make explicit)

0acbf17

Remove fake_impl error (needed for autograd check)

be89f27

Fix test (still trace but fail on inductor)

82f9ffd

Readd graphbreak (necessary for clear semantics)

d3789ea

Should pass tests now

291b0dc

Fix breakage of dlrm model

d9b2166

pytorchmergebot force-pushed the lucaskabela/fix_164823 branch from e702058 to d9b2166 Compare October 15, 2025 23:38

pytorchmergebot added the merging label Oct 16, 2025

pytorchmergebot added the Merged label Oct 16, 2025

pytorchmergebot closed this in e6d9d68 Oct 16, 2025

pytorchmergebot removed the merging label Oct 16, 2025

github-actions bot deleted the lucaskabela/fix_164823 branch November 16, 2025 02:21

	gb_type="Attempted to wrap sparse Tensor",
	gb_type="Attempted to wrap sparse Tensor with VariableTracker",

Conversation

Lucaskabela commented Oct 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-bot bot commented Oct 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/164873

✅ No Failures

Uh oh!

Lucaskabela commented Oct 7, 2025

Uh oh!

Lucaskabela commented Oct 9, 2025

Uh oh!

williamwen42 Oct 9, 2025

Choose a reason for hiding this comment

Uh oh!

williamwen42 Oct 9, 2025

Choose a reason for hiding this comment

Uh oh!

StrongerXi commented Oct 9, 2025

Uh oh!

Lucaskabela commented Oct 9, 2025

Uh oh!

eellison left a comment

Choose a reason for hiding this comment

Uh oh!

eellison commented Oct 10, 2025

Uh oh!

Lucaskabela commented Oct 10, 2025

Uh oh!

Lucaskabela commented Oct 13, 2025

Uh oh!

pytorch-bot bot commented Oct 13, 2025

Uh oh!

Lucaskabela commented Oct 13, 2025

Uh oh!

pytorchmergebot commented Oct 13, 2025

Merge started

Uh oh!

pytorchmergebot commented Oct 13, 2025

Merge failed

Uh oh!

Lucaskabela Oct 13, 2025

Choose a reason for hiding this comment

Uh oh!

pytorchmergebot commented Oct 15, 2025

Uh oh!

pytorchmergebot commented Oct 15, 2025

Uh oh!

Lucaskabela commented Oct 16, 2025

Uh oh!

pytorchmergebot commented Oct 16, 2025

Merge started

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

Lucaskabela commented Oct 7, 2025 •

edited

Loading

pytorch-bot bot commented Oct 7, 2025 •

edited

Loading