[dynamo] handle setting .data on a tensor #113080

voznesenskym · 2023-11-06T22:16:40Z

Stack from ghstack (oldest at bottom):

-> [dynamo] handle setting .data on a tensor #113080

Dynamo

We don't want setattr in the graph. Setting data has interesting implications on both aliasing and on the autograd engine.

The safe recipe is:

Disable grad
Call set_()
Manually lower the version counter on the object to hide it from the autograd engine

This is effectively the same exact thing as setting .data, and it composes properly with aot_autograd and inductor.

aot_autograd

For aot_autograd, there's another snag.

Specifically, when we invoke aot_autograd, we call fake_mode.from_tensor(), relying on memo to get the right tensor out. For .data mutations, this doesn't work, because the memoized fake_tensor is in the state it will be in at the end of the trace, not at the beginning. This means that the .data call is already applied, and the tensor shape (as in the case of these tests) mismatches. aot_autograd produces an invalid graph, with illegal calls like torch.ops.aten.view.default(primals_2, [0]) where primals is actually sized ([6]) on input.

The new plan here is to:

Record tensor fakification policy in dynamo
provide a fresh fake mode to all backends
Invoke from_tensor with the stored policy to get fresh new fake tensors in aot_autograd

cc @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @aakhundov @kadeng

cleanup fix tests [ghstack-poisoned]

pytorch-bot · 2023-11-06T22:16:43Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/113080

📄 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.

✅ No Failures

As of commit 374d303 with merge base ec124b9 ():
💚 Looks good so far! There are no failures yet. 💚

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

cleanup fix tests cc penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 aakhundov kadeng [ghstack-poisoned]

torch/_dynamo/variables/builtin.py

test/dynamo/test_repros.py

voznesenskym · 2023-11-07T01:38:18Z

Need to make a few changes, holdup

We don't want setattr in the graph. Setting data has interesting implications on both aliasing and on the autograd engine. The safe recipe is: 1) Disable grad 2) Call set_() 3) Manually lower the version counter on the object to hide it from the autograd engine This is effectively the same exact thing as setting .data, and it composes properly with aot_autograd and inductor. cc penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 aakhundov kadeng [ghstack-poisoned]

**Dynamo** We don't want setattr in the graph. Setting data has interesting implications on both aliasing and on the autograd engine. The safe recipe is: 1) Disable grad 2) Call set_() 3) Manually lower the version counter on the object to hide it from the autograd engine This is effectively the same exact thing as setting .data, and it composes properly with aot_autograd and inductor. **aot_autograd** For aot_autograd, there's another snag. Specifically, when we invoke aot_autograd, we call `fake_mode.from_tensor()`, relying on memo to get the right tensor out. For .data mutations, this doesn't work, because the memoized fake_tensor is in the state it will be in at the end of the trace, not at the beginning. This means that the .data call is already applied, and the tensor shape (as in the case of these tests) mismatches. aot_autograd produces an invalid graph, with illegal calls like `torch.ops.aten.view.default(primals_2, [0])` where primals is actually sized `([6])` on input. The plan/RFC here is to force re-creating the fake_tensor using the passed in input, so that aot_autograd does not choke on the graph it makes at runtime. cc penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 aakhundov kadeng [ghstack-poisoned]

voznesenskym · 2023-11-07T07:09:27Z

torch/_subclasses/meta_utils.py

+        if force_fresh:
+            del self.tensor_memo[WeakIdRef(t)]


@ezyang this might have weird implications for dynamic shapes, we should discuss. Will ping you.

**Dynamo** We don't want setattr in the graph. Setting data has interesting implications on both aliasing and on the autograd engine. The safe recipe is: 1) Disable grad 2) Call set_() 3) Manually lower the version counter on the object to hide it from the autograd engine This is effectively the same exact thing as setting .data, and it composes properly with aot_autograd and inductor. **aot_autograd** For aot_autograd, there's another snag. Specifically, when we invoke aot_autograd, we call `fake_mode.from_tensor()`, relying on memo to get the right tensor out. For .data mutations, this doesn't work, because the memoized fake_tensor is in the state it will be in at the end of the trace, not at the beginning. This means that the .data call is already applied, and the tensor shape (as in the case of these tests) mismatches. aot_autograd produces an invalid graph, with illegal calls like `torch.ops.aten.view.default(primals_2, [0])` where primals is actually sized `([6])` on input. The plan/RFC here is to force re-creating the fake_tensor using the passed in input, so that aot_autograd does not choke on the graph it makes at runtime. cc penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 aakhundov kadeng [ghstack-poisoned]

cleanup fix tests ghstack-source-id: ff1a162 Pull Request resolved: #113080 fix tests fix tests fix tests [dynamo] Not working yet - .add_() w/ aot_autograd (missing synthetic_base_info) ghstack-source-id: ff1a162 Pull Request resolved: #113081 fix tests aot_autograd fixes fix tests

voznesenskym · 2023-11-07T07:27:05Z

torch/_functorch/aot_autograd.py

+                    try:
+                        original_inpt.set_(updated_inpt)
+                    except RuntimeError:
+                        original_inpt.data = updated_inpt


@bdhirsh this passes, but is garbage, let's figure out a better way to do this? Maybe what we need here is to track if the original_inpt set_() was due to a .data change, since that bypasses size safety.

May I ask what error is raised? Why does set_ check size safety?

Is the error due to checking size when it consists of dynamic shapes?

Hmm, what about this:

# We can technically run under no_grad and avoid the VC bump to avoid any issues at runtime # For full correctness, we should probably make sure that we **still** bump the VC, if the eager mode code also would have bumped the VC ctx = nullcontext() if detected_vc_did_not_change_at_compile_time: ctx = torch.autograd._unsafe_preserve_version_counter(original_inpt) with torch.no_grad(), ctx: original_inpt.set_(updated_inpt)

If we're saying "the only forms of .data we will support in torch.compile are those that can be emulated through set_() + VC setting", then it should be safe to assume that if we got here at runtime, we should be able to use those same API's?

torch/_subclasses/fake_tensor.py

jon-chuang · 2023-11-07T15:40:07Z

As a follow-up / corner case, could you also check if

t.real = y
t.complex = y

Require the same handling? I think t.real = may desugar to t.data = ?

Not sure what t.complex = does for non complex tensors, but you can actually set it... (maybe ignore this)

bdhirsh · 2023-11-30T19:54:57Z

torch/_dynamo/variables/builtin.py

+                    # new .data value size and shape differences will cause
+                    # tracked fakes to produce incorrect guards. This is sound because the TensorVariable
+                    # coming out of set_() below will be a new one, and get
+                    # installed in tracked fakes.


Hmm I'm a little unclear on what tracked_fakes is used for, and why we still need this extra handling now that we have the "immutable FakeTensorMode infra. Just leaving the comment but I'll read the tracked_fakes code some more

oh at first I was going to say set_() is an inplace op so it will just return the same (mutated) FakeTensor as the input".

But I see that tracked_fakes is a List not a Set. So I think this makes sense to me. One question I have is: does that mean that we'll end up with duplicate FakeTensors in the tracked_fakes list whenever we have an ordinary inplace op? And I guess that's fine because that list is only really used by ShapEnv, which is only using it to generate guards on the FakeTensor symbols.

Potentially? We should test for this.

torch/_dynamo/variables/builtin.py

**Dynamo** We don't want setattr in the graph. Setting data has interesting implications on both aliasing and on the autograd engine. The safe recipe is: 1) Disable grad 2) Call set_() 3) Manually lower the version counter on the object to hide it from the autograd engine This is effectively the same exact thing as setting .data, and it composes properly with aot_autograd and inductor. **aot_autograd** For aot_autograd, there's another snag. Specifically, when we invoke aot_autograd, we call `fake_mode.from_tensor()`, relying on memo to get the right tensor out. For .data mutations, this doesn't work, because the memoized fake_tensor is in the state it will be in at the end of the trace, not at the beginning. This means that the .data call is already applied, and the tensor shape (as in the case of these tests) mismatches. aot_autograd produces an invalid graph, with illegal calls like `torch.ops.aten.view.default(primals_2, [0])` where primals is actually sized `([6])` on input. The new plan here is to: 1) Record tensor fakification policy in dynamo 2) provide a fresh fake mode to all backends 3) Invoke from_tensor with the stored policy to get fresh new fake tensors in aot_autograd cc penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 aakhundov kadeng [ghstack-poisoned]

bdhirsh · 2023-11-30T21:32:25Z

torch/_dynamo/variables/builtin.py

+                        (out.as_proxy(),),
+                        {},
+                    )
+                    obj.as_proxy().node.meta['example_value'] = _lower_version_count_by_1(obj.as_proxy().node.meta['example_value'])


doens't really matter but no need to re-assign (we're just directly mutating the FakeTensor`'s VC)

**Dynamo** We don't want setattr in the graph. Setting data has interesting implications on both aliasing and on the autograd engine. The safe recipe is: 1) Disable grad 2) Call set_() 3) Manually lower the version counter on the object to hide it from the autograd engine This is effectively the same exact thing as setting .data, and it composes properly with aot_autograd and inductor. **aot_autograd** For aot_autograd, there's another snag. Specifically, when we invoke aot_autograd, we call `fake_mode.from_tensor()`, relying on memo to get the right tensor out. For .data mutations, this doesn't work, because the memoized fake_tensor is in the state it will be in at the end of the trace, not at the beginning. This means that the .data call is already applied, and the tensor shape (as in the case of these tests) mismatches. aot_autograd produces an invalid graph, with illegal calls like `torch.ops.aten.view.default(primals_2, [0])` where primals is actually sized `([6])` on input. The new plan here is to: 1) Record tensor fakification policy in dynamo 2) provide a fresh fake mode to all backends 3) Invoke from_tensor with the stored policy to get fresh new fake tensors in aot_autograd cc penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 aakhundov kadeng [ghstack-poisoned]

cleanup fix tests ghstack-source-id: a87bc01 Pull Request resolved: #113080 fix tests fix tests fix tests [dynamo] Not working yet - .add_() w/ aot_autograd (missing synthetic_base_info) ghstack-source-id: a87bc01 Pull Request resolved: #113081 fix tests aot_autograd fixes fix tests Refactor Revert "[dynamo] Graph break on `setattr(Tensor, "data", Tensor)` (#113043)" This reverts commit ddfe572. ghstack-source-id: a87bc01 Pull Request resolved: #113205 Refactor Refactor Refactor Refactor Refactor Refactor cache to shape env ghstack-source-id: a87bc01 Pull Request resolved: #113295 cache to shape env fake policy fake policy fake policy fake policy fake policy fake policy fake policy fake policy fake policy Builder fix Builder fix lint bad test bad test bad tests Alternate impl ghstack-source-id: a87bc01 Pull Request resolved: #113576 Alternate impl undo wip Fixes Fixes Fixes wip lint, test fix

**Dynamo** We don't want setattr in the graph. Setting data has interesting implications on both aliasing and on the autograd engine. The safe recipe is: 1) Disable grad 2) Call set_() 3) Manually lower the version counter on the object to hide it from the autograd engine This is effectively the same exact thing as setting .data, and it composes properly with aot_autograd and inductor. **aot_autograd** For aot_autograd, there's another snag. Specifically, when we invoke aot_autograd, we call `fake_mode.from_tensor()`, relying on memo to get the right tensor out. For .data mutations, this doesn't work, because the memoized fake_tensor is in the state it will be in at the end of the trace, not at the beginning. This means that the .data call is already applied, and the tensor shape (as in the case of these tests) mismatches. aot_autograd produces an invalid graph, with illegal calls like `torch.ops.aten.view.default(primals_2, [0])` where primals is actually sized `([6])` on input. The new plan here is to: 1) Record tensor fakification policy in dynamo 2) provide a fresh fake mode to all backends 3) Invoke from_tensor with the stored policy to get fresh new fake tensors in aot_autograd cc penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 aakhundov kadeng [ghstack-poisoned]

cleanup fix tests ghstack-source-id: 7a543f8 Pull Request resolved: #113080 fix tests fix tests fix tests [dynamo] Not working yet - .add_() w/ aot_autograd (missing synthetic_base_info) ghstack-source-id: 7a543f8 Pull Request resolved: #113081 fix tests aot_autograd fixes fix tests Refactor Revert "[dynamo] Graph break on `setattr(Tensor, "data", Tensor)` (#113043)" This reverts commit ddfe572. ghstack-source-id: 7a543f8 Pull Request resolved: #113205 Refactor Refactor Refactor Refactor Refactor Refactor cache to shape env ghstack-source-id: 7a543f8 Pull Request resolved: #113295 cache to shape env fake policy fake policy fake policy fake policy fake policy fake policy fake policy fake policy fake policy Builder fix Builder fix lint bad test bad test bad tests Alternate impl ghstack-source-id: 7a543f8 Pull Request resolved: #113576 Alternate impl undo wip Fixes Fixes Fixes wip lint, test fix woopsie fix test

**Dynamo** We don't want setattr in the graph. Setting data has interesting implications on both aliasing and on the autograd engine. The safe recipe is: 1) Disable grad 2) Call set_() 3) Manually lower the version counter on the object to hide it from the autograd engine This is effectively the same exact thing as setting .data, and it composes properly with aot_autograd and inductor. **aot_autograd** For aot_autograd, there's another snag. Specifically, when we invoke aot_autograd, we call `fake_mode.from_tensor()`, relying on memo to get the right tensor out. For .data mutations, this doesn't work, because the memoized fake_tensor is in the state it will be in at the end of the trace, not at the beginning. This means that the .data call is already applied, and the tensor shape (as in the case of these tests) mismatches. aot_autograd produces an invalid graph, with illegal calls like `torch.ops.aten.view.default(primals_2, [0])` where primals is actually sized `([6])` on input. The new plan here is to: 1) Record tensor fakification policy in dynamo 2) provide a fresh fake mode to all backends 3) Invoke from_tensor with the stored policy to get fresh new fake tensors in aot_autograd cc penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 aakhundov kadeng [ghstack-poisoned]

cleanup fix tests ghstack-source-id: 5343720 Pull Request resolved: #113080 fix tests fix tests fix tests [dynamo] Not working yet - .add_() w/ aot_autograd (missing synthetic_base_info) ghstack-source-id: 5343720 Pull Request resolved: #113081 fix tests aot_autograd fixes fix tests Refactor Revert "[dynamo] Graph break on `setattr(Tensor, "data", Tensor)` (#113043)" This reverts commit ddfe572. ghstack-source-id: 5343720 Pull Request resolved: #113205 Refactor Refactor Refactor Refactor Refactor Refactor cache to shape env ghstack-source-id: 5343720 Pull Request resolved: #113295 cache to shape env fake policy fake policy fake policy fake policy fake policy fake policy fake policy fake policy fake policy Builder fix Builder fix lint bad test bad test bad tests Alternate impl ghstack-source-id: 5343720 Pull Request resolved: #113576 Alternate impl undo wip Fixes Fixes Fixes wip lint, test fix woopsie fix test rm replace

bdhirsh

lgtm

**Dynamo** We don't want setattr in the graph. Setting data has interesting implications on both aliasing and on the autograd engine. The safe recipe is: 1) Disable grad 2) Call set_() 3) Manually lower the version counter on the object to hide it from the autograd engine This is effectively the same exact thing as setting .data, and it composes properly with aot_autograd and inductor. **aot_autograd** For aot_autograd, there's another snag. Specifically, when we invoke aot_autograd, we call `fake_mode.from_tensor()`, relying on memo to get the right tensor out. For .data mutations, this doesn't work, because the memoized fake_tensor is in the state it will be in at the end of the trace, not at the beginning. This means that the .data call is already applied, and the tensor shape (as in the case of these tests) mismatches. aot_autograd produces an invalid graph, with illegal calls like `torch.ops.aten.view.default(primals_2, [0])` where primals is actually sized `([6])` on input. The new plan here is to: 1) Record tensor fakification policy in dynamo 2) provide a fresh fake mode to all backends 3) Invoke from_tensor with the stored policy to get fresh new fake tensors in aot_autograd cc penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 aakhundov kadeng [ghstack-poisoned]

cleanup fix tests ghstack-source-id: c89b426 Pull Request resolved: #113080 fix tests fix tests fix tests [dynamo] Not working yet - .add_() w/ aot_autograd (missing synthetic_base_info) ghstack-source-id: c89b426 Pull Request resolved: #113081 fix tests aot_autograd fixes fix tests Refactor Revert "[dynamo] Graph break on `setattr(Tensor, "data", Tensor)` (#113043)" This reverts commit ddfe572. ghstack-source-id: c89b426 Pull Request resolved: #113205 Refactor Refactor Refactor Refactor Refactor Refactor cache to shape env ghstack-source-id: c89b426 Pull Request resolved: #113295 cache to shape env fake policy fake policy fake policy fake policy fake policy fake policy fake policy fake policy fake policy Builder fix Builder fix lint bad test bad test bad tests Alternate impl ghstack-source-id: c89b426 Pull Request resolved: #113576 Alternate impl undo wip Fixes Fixes Fixes wip lint, test fix woopsie fix test rm replace rm replace

voznesenskym · 2023-12-02T00:31:57Z

@pytorchbot merge

pytorchmergebot · 2023-12-02T00:35:22Z

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

**Dynamo** We don't want setattr in the graph. Setting data has interesting implications on both aliasing and on the autograd engine. The safe recipe is: 1) Disable grad 2) Call set_() 3) Manually lower the version counter on the object to hide it from the autograd engine This is effectively the same exact thing as setting .data, and it composes properly with aot_autograd and inductor. **aot_autograd** For aot_autograd, there's another snag. Specifically, when we invoke aot_autograd, we call `fake_mode.from_tensor()`, relying on memo to get the right tensor out. For .data mutations, this doesn't work, because the memoized fake_tensor is in the state it will be in at the end of the trace, not at the beginning. This means that the .data call is already applied, and the tensor shape (as in the case of these tests) mismatches. aot_autograd produces an invalid graph, with illegal calls like `torch.ops.aten.view.default(primals_2, [0])` where primals is actually sized `([6])` on input. The new plan here is to: 1) Record tensor fakification policy in dynamo 2) provide a fresh fake mode to all backends 3) Invoke from_tensor with the stored policy to get fresh new fake tensors in aot_autograd Pull Request resolved: pytorch#113080 Approved by: https://github.com/bdhirsh

[dynamo] handle setting .data on a tensor

3c7141b

cleanup fix tests [ghstack-poisoned]

pytorch-bot bot added the release notes: AO frontend label Nov 6, 2023

This was referenced Nov 6, 2023

AOTAutograd: handle set_(), detect metadata mutations that cancel out #111554

Closed

[dynamo] Not working yet - .add_() w/ aot_autograd (missing synthetic_base_info) #113081

Closed

github-actions bot requested review from SherlockNoMad, albanD, antoniojkim, bdhirsh, ezyang, miladm and wconstab November 6, 2023 22:16

github-actions bot added module: dynamo ciflow/inductor labels Nov 6, 2023

voznesenskym mentioned this pull request Nov 6, 2023

[dynamo] handle setting .data on a tensor #113069

Closed

voznesenskym added 2 commits November 6, 2023 14:19

Update on "[dynamo] handle setting .data on a tensor"

59c7ec0

cleanup fix tests cc penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 aakhundov kadeng [ghstack-poisoned]

Update on "[dynamo] handle setting .data on a tensor"

f1d9f25

cleanup fix tests cc penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 aakhundov kadeng [ghstack-poisoned]

ezyang reviewed Nov 7, 2023

View reviewed changes

torch/_dynamo/variables/builtin.py Show resolved Hide resolved

voznesenskym marked this pull request as draft November 7, 2023 01:11

jon-chuang reviewed Nov 7, 2023

View reviewed changes

test/dynamo/test_repros.py Outdated Show resolved Hide resolved

voznesenskym added 2 commits November 6, 2023 18:07

voznesenskym commented Nov 7, 2023

View reviewed changes

jon-chuang reviewed Nov 7, 2023

View reviewed changes

torch/_subclasses/fake_tensor.py Outdated Show resolved Hide resolved

bdhirsh reviewed Nov 30, 2023

View reviewed changes

torch/_dynamo/variables/builtin.py Show resolved Hide resolved

bdhirsh reviewed Nov 30, 2023

View reviewed changes

torch/_dynamo/variables/builtin.py Outdated Show resolved Hide resolved

bdhirsh reviewed Nov 30, 2023

View reviewed changes

bdhirsh approved these changes Dec 1, 2023

View reviewed changes

pytorchmergebot added the merging label Dec 2, 2023

pytorchmergebot added the Merged label Dec 2, 2023

pytorchmergebot removed the merging label Dec 2, 2023

pytorchmergebot closed this in 4cfe997 Dec 2, 2023

facebook-github-bot deleted the gh/voznesenskym/265/head branch December 5, 2023 15:28

bdhirsh mentioned this pull request Jul 10, 2024

[Traceable FSDP2][AOTAutograd] Move .set_ and .resize_ to end of joint graph #129852

Closed

[dynamo] handle setting .data on a tensor #113080

[dynamo] handle setting .data on a tensor #113080

Uh oh!

Conversation

voznesenskym commented Nov 6, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-bot bot commented Nov 6, 2023 • 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/113080

✅ No Failures

Uh oh!

Uh oh!

Uh oh!

voznesenskym commented Nov 7, 2023

Uh oh!

voznesenskym Nov 7, 2023

Choose a reason for hiding this comment

Uh oh!

voznesenskym Nov 7, 2023

Choose a reason for hiding this comment

Uh oh!

jon-chuang Nov 7, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

bdhirsh Nov 7, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

jon-chuang commented Nov 7, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

bdhirsh Nov 30, 2023

Choose a reason for hiding this comment

Uh oh!

bdhirsh Nov 30, 2023

Choose a reason for hiding this comment

Uh oh!

voznesenskym Dec 1, 2023

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

bdhirsh Nov 30, 2023

Choose a reason for hiding this comment

Uh oh!

bdhirsh left a comment

Choose a reason for hiding this comment

Uh oh!

voznesenskym commented Dec 2, 2023

Uh oh!

pytorchmergebot commented Dec 2, 2023

Merge started

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

voznesenskym commented Nov 6, 2023 •

edited

Loading

pytorch-bot bot commented Nov 6, 2023 •

edited

Loading

jon-chuang Nov 7, 2023 •

edited

Loading

bdhirsh Nov 7, 2023 •

edited

Loading

jon-chuang commented Nov 7, 2023 •

edited

Loading