[Static Runtime] Fix MemoryPlanner dtor crash in debug mode#79942
Closed
mikeiovine wants to merge 1 commit intopytorch:masterfrom
Closed
[Static Runtime] Fix MemoryPlanner dtor crash in debug mode#79942mikeiovine wants to merge 1 commit intopytorch:masterfrom
mikeiovine wants to merge 1 commit intopytorch:masterfrom
Conversation
Contributor
🔗 Helpful links
✅ No Failures (0 Pending)As of commit 3f9c479 (more details on the Dr. CI page): Expand to see more💚 💚 Looks good so far! There are no failures yet. 💚 💚 This comment was automatically generated by Dr. CI (expand for details).Please report bugs/suggestions to the (internal) Dr. CI Users group. |
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D37303728 |
…79942) Summary: Pull Request resolved: pytorch#79942 Memory planner destruction was hitting [this assertion](https://www.internalfb.com/code/fbsource/[f8baf8a0bab462c860d2eb7491a4e3fb40d2907a]/fbcode/caffe2/c10/util/intrusive_ptr.h?lines=117) in debug mode for a few models. Here's what was going on: 1) The set of unmanaged `IValue`s acquires one or more owning refs of a managed `StorageImpl` 2) Then, one or more tensors in that storage group have their `StorageImpl` swapped out during execution 3) During `deallocateManagedTensors`, we swap the correct `StorageImpl` back in, [calling `unsafe_adapt_non_heap_allocated` again and resetting the refcount](https://www.internalfb.com/code/fbsource/[f8baf8a0bab462c860d2eb7491a4e3fb40d2907a]/fbcode/caffe2/torch/csrc/jit/runtime/static/memory_planner.cpp?lines=446-452) 4) The unmanaged `IValues` are deallocated, decrementing the refcount into the danger zone. So, we just have to make sure that unmanaged `IValue`s are destructed before we deallocate the managed tensors. Test Plan: CI Differential Revision: D37303728 fbshipit-source-id: dffc079c84dc92b5dad44758cb2bafbf60dc0d0e
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D37303728 |
298674f to
3f9c479
Compare
tenpercent
approved these changes
Jun 21, 2022
Contributor
|
@pytorchbot merge (Initiating merge automatically since Phabricator Diff has merged) |
Collaborator
|
@pytorchbot successfully started a merge job. Check the current status here |
Collaborator
|
@mikeiovine your PR has been successfully merged. |
Contributor
|
Hey @mikeiovine. |
facebook-github-bot
pushed a commit
that referenced
this pull request
Jun 21, 2022
Summary: Pull Request resolved: #79942 Memory planner destruction was hitting [this assertion](https://www.internalfb.com/code/fbsource/[f8baf8a0bab462c860d2eb7491a4e3fb40d2907a]/fbcode/caffe2/c10/util/intrusive_ptr.h?lines=117) in debug mode for a few models. Here's what was going on: 1) The set of unmanaged `IValue`s acquires one or more owning refs of a managed `StorageImpl` 2) Then, one or more tensors in that storage group have their `StorageImpl` swapped out during execution 3) During `deallocateManagedTensors`, we swap the correct `StorageImpl` back in, [calling `unsafe_adapt_non_heap_allocated` again and resetting the refcount](https://www.internalfb.com/code/fbsource/[f8baf8a0bab462c860d2eb7491a4e3fb40d2907a]/fbcode/caffe2/torch/csrc/jit/runtime/static/memory_planner.cpp?lines=446-452) 4) The unmanaged `IValues` are deallocated, decrementing the refcount into the danger zone. So, we just have to make sure that unmanaged `IValue`s are destructed before we deallocate the managed tensors. Test Plan: CI Reviewed By: ajyu, tenpercent Differential Revision: D37303728 fbshipit-source-id: 82b58221d45ab04a30cb7358c8bbeb124f71129d
miladm
pushed a commit
to miladm/pytorch
that referenced
this pull request
Jun 27, 2022
…79942) Summary: Memory planner destruction was hitting [this assertion](https://www.internalfb.com/code/fbsource/[f8baf8a0bab462c860d2eb7491a4e3fb40d2907a]/fbcode/caffe2/c10/util/intrusive_ptr.h?lines=117) in debug mode for a few models. Here's what was going on: 1) The set of unmanaged `IValue`s acquires one or more owning refs of a managed `StorageImpl` 2) Then, one or more tensors in that storage group have their `StorageImpl` swapped out during execution 3) During `deallocateManagedTensors`, we swap the correct `StorageImpl` back in, [calling `unsafe_adapt_non_heap_allocated` again and resetting the refcount](https://www.internalfb.com/code/fbsource/[f8baf8a0bab462c860d2eb7491a4e3fb40d2907a]/fbcode/caffe2/torch/csrc/jit/runtime/static/memory_planner.cpp?lines=446-452) 4) The unmanaged `IValues` are deallocated, decrementing the refcount into the danger zone. So, we just have to make sure that unmanaged `IValue`s are destructed before we deallocate the managed tensors. Test Plan: CI Differential Revision: D37303728 Pull Request resolved: pytorch#79942 Approved by: https://github.com/tenpercent
laurentdupin
pushed a commit
to laurentdupin/pytorch
that referenced
this pull request
Apr 25, 2026
…79942) Summary: Memory planner destruction was hitting [this assertion](https://www.internalfb.com/code/fbsource/[f8baf8a0bab462c860d2eb7491a4e3fb40d2907a]/fbcode/caffe2/c10/util/intrusive_ptr.h?lines=117) in debug mode for a few models. Here's what was going on: 1) The set of unmanaged `IValue`s acquires one or more owning refs of a managed `StorageImpl` 2) Then, one or more tensors in that storage group have their `StorageImpl` swapped out during execution 3) During `deallocateManagedTensors`, we swap the correct `StorageImpl` back in, [calling `unsafe_adapt_non_heap_allocated` again and resetting the refcount](https://www.internalfb.com/code/fbsource/[f8baf8a0bab462c860d2eb7491a4e3fb40d2907a]/fbcode/caffe2/torch/csrc/jit/runtime/static/memory_planner.cpp?lines=446-452) 4) The unmanaged `IValues` are deallocated, decrementing the refcount into the danger zone. So, we just have to make sure that unmanaged `IValue`s are destructed before we deallocate the managed tensors. Test Plan: CI Differential Revision: D37303728 Pull Request resolved: pytorch#79942 Approved by: https://github.com/tenpercent
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary:
Memory planner destruction was hitting this assertion in debug mode for a few models.
Here's what was going on:
IValues acquires one or more owning refs of a managedStorageImplStorageImplswapped out during executiondeallocateManagedTensors, we swap the correctStorageImplback in, callingunsafe_adapt_non_heap_allocatedagain and resetting the refcountIValuesare deallocated, decrementing the refcount into the danger zone.So, we just have to make sure that unmanaged
IValues are destructed before we deallocate the managed tensors.Test Plan: CI
Differential Revision: D37303728