[ET-VK] Fix use-after-free in PrepackNode when TensorRefs are shared#18906
Merged
meta-codesync[bot] merged 2 commits intogh/SS-JIA/520/basefrom Apr 15, 2026
Merged
[ET-VK] Fix use-after-free in PrepackNode when TensorRefs are shared#18906meta-codesync[bot] merged 2 commits intogh/SS-JIA/520/basefrom
meta-codesync[bot] merged 2 commits intogh/SS-JIA/520/basefrom
Conversation
When a model has shared/tied weights (e.g. tied embeddings in transformers), the serialization deduplicates them into a single TensorRef that multiple PrepackNodes reference. Previously, `PrepackNode::create_staging_buffer()` called `tref->free_buffer()` unconditionally after copying weight data to a GPU staging buffer. This meant the first PrepackNode to execute would free the underlying host memory, and subsequent PrepackNodes sharing the same TensorRef would read from a dangling pointer — producing garbage/NaN values in prepacked weight and bias tensors on the GPU. The fix adds a `prepack_use_count` field to `TensorRef` that tracks how many PrepackNodes still need to read from it. Each PrepackNode increments the count in its constructor and decrements it after copying data. The buffer is only freed when the count reaches zero. This preserves the original eager-free behavior for non-shared weights (freeing immediately after the single consumer copies) while correctly deferring the free for shared weights until the last consumer is done — avoiding both the use-after-free and unnecessary peak memory increase. Differential Revision: [D101009402](https://our.internmc.facebook.com/intern/diff/D101009402/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18906
Note: Links to docs will display an error until the docs builds have been completed. ⏳ 1 Pending, 2 Unrelated FailuresAs of commit 5fd2bae with merge base 89e6416 ( BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
SS-JIA
pushed a commit
that referenced
this pull request
Apr 15, 2026
When a model has shared/tied weights (e.g. tied embeddings in transformers), the serialization deduplicates them into a single TensorRef that multiple PrepackNodes reference. Previously, `PrepackNode::create_staging_buffer()` called `tref->free_buffer()` unconditionally after copying weight data to a GPU staging buffer. This meant the first PrepackNode to execute would free the underlying host memory, and subsequent PrepackNodes sharing the same TensorRef would read from a dangling pointer — producing garbage/NaN values in prepacked weight and bias tensors on the GPU. The fix adds a `prepack_use_count` field to `TensorRef` that tracks how many PrepackNodes still need to read from it. Each PrepackNode increments the count in its constructor and decrements it after copying data. The buffer is only freed when the count reaches zero. This preserves the original eager-free behavior for non-shared weights (freeing immediately after the single consumer copies) while correctly deferring the free for shared weights until the last consumer is done — avoiding both the use-after-free and unnecessary peak memory increase. Differential Revision: [D101009402](https://our.internmc.facebook.com/intern/diff/D101009402/) ghstack-source-id: 367723453 Pull Request resolved: #18906
meta-cla
bot
added
the
CLA Signed
This PR needs a
|
…are shared" When a model has shared/tied weights (e.g. tied embeddings in transformers), the serialization deduplicates them into a single TensorRef that multiple PrepackNodes reference. Previously, `PrepackNode::create_staging_buffer()` called `tref->free_buffer()` unconditionally after copying weight data to a GPU staging buffer. This meant the first PrepackNode to execute would free the underlying host memory, and subsequent PrepackNodes sharing the same TensorRef would read from a dangling pointer — producing garbage/NaN values in prepacked weight and bias tensors on the GPU. The fix adds a `prepack_use_count` field to `TensorRef` that tracks how many PrepackNodes still need to read from it. Each PrepackNode increments the count in its constructor and decrements it after copying data. The buffer is only freed when the count reaches zero. This preserves the original eager-free behavior for non-shared weights (freeing immediately after the single consumer copies) while correctly deferring the free for shared weights until the last consumer is done — avoiding both the use-after-free and unnecessary peak memory increase. Differential Revision: [D101009402](https://our.internmc.facebook.com/intern/diff/D101009402/) [ghstack-poisoned]
SS-JIA
pushed a commit
that referenced
this pull request
Apr 15, 2026
Pull Request resolved: #18906 When a model has shared/tied weights (e.g. tied embeddings in transformers), the serialization deduplicates them into a single TensorRef that multiple PrepackNodes reference. Previously, `PrepackNode::create_staging_buffer()` called `tref->free_buffer()` unconditionally after copying weight data to a GPU staging buffer. This meant the first PrepackNode to execute would free the underlying host memory, and subsequent PrepackNodes sharing the same TensorRef would read from a dangling pointer — producing garbage/NaN values in prepacked weight and bias tensors on the GPU. The fix adds a `prepack_use_count` field to `TensorRef` that tracks how many PrepackNodes still need to read from it. Each PrepackNode increments the count in its constructor and decrements it after copying data. The buffer is only freed when the count reaches zero. This preserves the original eager-free behavior for non-shared weights (freeing immediately after the single consumer copies) while correctly deferring the free for shared weights until the last consumer is done — avoiding both the use-after-free and unnecessary peak memory increase. ghstack-source-id: 367726483 @exported-using-ghexport Differential Revision: [D101009402](https://our.internmc.facebook.com/intern/diff/D101009402/)
manuelcandales
approved these changes
Apr 15, 2026
meta-codesync
bot
merged commit cd7dc61
into
gh/SS-JIA/520/base
161 of 164 checks passed
SS-JIA
pushed a commit
that referenced
this pull request
Apr 15, 2026
Pull Request resolved: #18906 When a model has shared/tied weights (e.g. tied embeddings in transformers), the serialization deduplicates them into a single TensorRef that multiple PrepackNodes reference. Previously, `PrepackNode::create_staging_buffer()` called `tref->free_buffer()` unconditionally after copying weight data to a GPU staging buffer. This meant the first PrepackNode to execute would free the underlying host memory, and subsequent PrepackNodes sharing the same TensorRef would read from a dangling pointer — producing garbage/NaN values in prepacked weight and bias tensors on the GPU. The fix adds a `prepack_use_count` field to `TensorRef` that tracks how many PrepackNodes still need to read from it. Each PrepackNode increments the count in its constructor and decrements it after copying data. The buffer is only freed when the count reaches zero. This preserves the original eager-free behavior for non-shared weights (freeing immediately after the single consumer copies) while correctly deferring the free for shared weights until the last consumer is done — avoiding both the use-after-free and unnecessary peak memory increase. ghstack-source-id: 367726483 @exported-using-ghexport Differential Revision: [D101009402](https://our.internmc.facebook.com/intern/diff/D101009402/)
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
2 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.
Stack from ghstack (oldest at bottom):
When a model has shared/tied weights (e.g. tied embeddings in transformers), the serialization deduplicates them into a single TensorRef that multiple PrepackNodes reference. Previously,
PrepackNode::create_staging_buffer()calledtref->free_buffer()unconditionally after copying weight data to a GPU staging buffer. This meant the first PrepackNode to execute would free the underlying host memory, and subsequent PrepackNodes sharing the same TensorRef would read from a dangling pointer — producing garbage/NaN values in prepacked weight and bias tensors on the GPU.The fix adds a
prepack_use_countfield toTensorRefthat tracks how many PrepackNodes still need to read from it. Each PrepackNode increments the count in its constructor and decrements it after copying data. The buffer is only freed when the count reaches zero. This preserves the original eager-free behavior for non-shared weights (freeing immediately after the single consumer copies) while correctly deferring the free for shared weights until the last consumer is done — avoiding both the use-after-free and unnecessary peak memory increase.Differential Revision: D101009402