[1/2] Introduce at::accelerator::Graph as a unified Graph interface#171269
Closed
guangyey wants to merge 20 commits intogh/guangyey/265/basefrom
Closed
[1/2] Introduce at::accelerator::Graph as a unified Graph interface#171269guangyey wants to merge 20 commits intogh/guangyey/265/basefrom
guangyey wants to merge 20 commits intogh/guangyey/265/basefrom
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/171269
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (2 Unrelated Failures)As of commit 62ab0c3 with merge base f72a552 ( FLAKY - The following jobs failed but were likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Dec 29, 2025
Collaborator
Author
|
Thanks @eellison |
Collaborator
Merge startedYour 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 |
Contributor
|
@pytorchbot revert -m 'This is currently breaking some internal build, I need to revert and reland this' -c ghfirst |
Collaborator
|
@pytorchbot successfully started a revert job. Check the current status here. |
pytorchmergebot
added a commit
that referenced
this pull request
Mar 11, 2026
…erface (#171269)" This reverts commit 2afd3c1. Reverted #171269 on behalf of https://github.com/huydhn due to This is currently breaking some internal build, I need to revert and reland this ([comment](#171269 (comment)))
Collaborator
|
@guangyey your PR has been successfully reverted. |
Contributor
|
@guangyey Please help do a rebase and reland this change |
Collaborator
Author
|
Try to reland. |
Collaborator
Merge startedYour 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 |
sandy-gags
pushed a commit
to sandy-gags/pytorch
that referenced
this pull request
Mar 12, 2026
ghstack-source-id: d3b7f3f Pull Request resolved: pytorch/pytorch#171269
EmanueleCoradin
pushed a commit
to EmanueleCoradin/pytorch
that referenced
this pull request
Mar 30, 2026
…ytorch#171269) # Motivation The original goal was to generalize `CUDAGraph` and share implementations and logic across different backends, as mentioned in pytorch#158827. However, after further offline discussions, we decided to take a more incremental approach: start by defining a unified interface, while allowing each backend to maintain its own implementation. This avoids premature coupling and addresses backend-specific concerns. This PR introduces `GraphImplInterface`, a lightweight, backend-agnostic interface that defines a unified API for graph capture and replay. Each backend (e.g., `CUDA`, `XPU`, `PrivateUse1`) provides its own implementation and registers it via `REGISTER_GRAPH_IMPL`. On top of this interface, we provide a unified graph API, `at::accelerator::Graph`, which transparently maps to: - `CUDAGraph` on CUDA - `XPUGraph` on XPU - and corresponding implementations for other backends (including `PrivateUse1`) This design establishes a common abstraction layer while preserving backend autonomy, and lays the groundwork for future sharing of logic once the interface and use cases have stabilized. An additional benefit is that, for `CUDA` and `XPU`, the backend-specific graph types (e.g., `cuda::CUDAGraph` and `xpu::XPUGraph`) can share the same underlying implementation as `accelerator::Graph` on each backend, avoiding code duplication and ensuring consistent behavior. For `PrivateUse1`, `accelerator::Graph` can be supported with minimal effort by reusing the existing `PU1Graph` implementation. Pull Request resolved: pytorch#171269 Approved by: https://github.com/EikanWang, https://github.com/eellison
EmanueleCoradin
pushed a commit
to EmanueleCoradin/pytorch
that referenced
this pull request
Mar 30, 2026
…erface (pytorch#171269)" This reverts commit 2afd3c1. Reverted pytorch#171269 on behalf of https://github.com/huydhn due to This is currently breaking some internal build, I need to revert and reland this ([comment](pytorch#171269 (comment)))
EmanueleCoradin
pushed a commit
to EmanueleCoradin/pytorch
that referenced
this pull request
Mar 30, 2026
…ytorch#171269) # Motivation The original goal was to generalize `CUDAGraph` and share implementations and logic across different backends, as mentioned in pytorch#158827. However, after further offline discussions, we decided to take a more incremental approach: start by defining a unified interface, while allowing each backend to maintain its own implementation. This avoids premature coupling and addresses backend-specific concerns. This PR introduces `GraphImplInterface`, a lightweight, backend-agnostic interface that defines a unified API for graph capture and replay. Each backend (e.g., `CUDA`, `XPU`, `PrivateUse1`) provides its own implementation and registers it via `REGISTER_GRAPH_IMPL`. On top of this interface, we provide a unified graph API, `at::accelerator::Graph`, which transparently maps to: - `CUDAGraph` on CUDA - `XPUGraph` on XPU - and corresponding implementations for other backends (including `PrivateUse1`) This design establishes a common abstraction layer while preserving backend autonomy, and lays the groundwork for future sharing of logic once the interface and use cases have stabilized. An additional benefit is that, for `CUDA` and `XPU`, the backend-specific graph types (e.g., `cuda::CUDAGraph` and `xpu::XPUGraph`) can share the same underlying implementation as `accelerator::Graph` on each backend, avoiding code duplication and ensuring consistent behavior. For `PrivateUse1`, `accelerator::Graph` can be supported with minimal effort by reusing the existing `PU1Graph` implementation. Pull Request resolved: pytorch#171269 Approved by: https://github.com/EikanWang, https://github.com/eellison
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
7 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):
Motivation
The original goal was to generalize
CUDAGraphand share implementations and logic across different backends, as mentioned in #158827. However, after further offline discussions, we decided to take a more incremental approach: start by defining a unified interface, while allowing each backend to maintain its own implementation. This avoids premature coupling and addresses backend-specific concerns.This PR introduces
GraphImplInterface, a lightweight, backend-agnostic interface that defines a unified API for graph capture and replay. Each backend (e.g.,CUDA,XPU,PrivateUse1) provides its own implementation and registers it viaREGISTER_GRAPH_IMPL.On top of this interface, we provide a unified graph API,
at::accelerator::Graph, which transparently maps to:CUDAGraphon CUDAXPUGraphon XPUPrivateUse1)This design establishes a common abstraction layer while preserving backend autonomy, and lays the groundwork for future sharing of logic once the interface and use cases have stabilized.
An additional benefit is that, for
CUDAandXPU, the backend-specific graph types (e.g.,cuda::CUDAGraphandxpu::XPUGraph) can share the same underlying implementation asaccelerator::Graphon each backend, avoiding code duplication and ensuring consistent behavior.For
PrivateUse1,accelerator::Graphcan be supported with minimal effort by reusing the existingPU1Graphimplementation.cc @albanD @eellison @EikanWang