[ET-VK] Move rotary embedding custom op to be handled via graph pass instead of source transform #13450
Conversation
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
…instead of source transform ## Motivation Be able to test Vulkan lowering via optimum-executorch. ## Context Currently, ET-VK implements rotary embeddings via a custom op. This op is currently inserted into Transformer models by replacing Rotary Embedding modules with a custom module that executes the custom op via a source transform. The source transform approach makes it cumbersome to lower LLMs to Vulkan, since it requires the export logic to apply the source transform before calling `torch.export()`. This in turn makes it difficult to integrate Vulkan lowering into optimum-executorch, which tries to use a common export + lowering logic for all lowering paths. As an alternative, leverage `SubgraphMatcher` to detect fusable patterns and fuse the rotary embedding graph pattern into the custom op as part of the Vulkan delegate's graph passes. This removes the requirement to apply a custom source transform just for Vulkan. ## Changes * Introduce the `backends/vulkan/patterns` folder to store fusable graph patterns * Introduce a fusable graph pattern for rotary positional embeddings * Update partitioner logic to automatically include nodes that are part of a fusable graph pattern * Introduce a pass to fuse known patterns into custom ops / custom op sequence Differential Revision: [D80293301](https://our.internmc.facebook.com/intern/diff/D80293301/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/13450
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ❌ 2 New FailuresAs of commit 83b317e with merge base ecb639a ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
This PR needs a
|
SS-JIA
had a problem deploying
to
cherry-pick-bot
GitHub Actions
Failure
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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
Be able to test Vulkan lowering via optimum-executorch.
Context
Currently, ET-VK implements rotary embeddings via a custom op. This op is currently inserted into Transformer models by replacing Rotary Embedding modules with a custom module that executes the custom op via a source transform.
The source transform approach makes it cumbersome to lower LLMs to Vulkan, since it requires the export logic to apply the source transform before calling
torch.export(). This in turn makes it difficult to integrate Vulkan lowering into optimum-executorch, which tries to use a common export + lowering logic for all lowering paths.As an alternative, leverage
SubgraphMatcherto detect fusable patterns and fuse the rotary embedding graph pattern into the custom op as part of the Vulkan delegate's graph passes. This removes the requirement to apply a custom source transform just for Vulkan.Changes
backends/vulkan/patternsfolder to store fusable graph patternsDifferential Revision: D80293301