[ET-VK] Split up prepack command buffer #12533
Merged
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
Pull Request resolved: #12442 ## Changes * Introduce `run_prepack()` API which combines the functionality of `encode_prepack()` and `prepack()`, but submits prepacking shaders incrementally rather than all at once. * Introduce graph config options to control command buffer submission behaviour during prepacking. Note that the current default values for the prepack submission thresholds were determined through experimentation. I will leave determining optimal values for specific devices as a later exercise. The goal of this diff is simply to introduce this mechanism to fix the Llama model loading crash on Samsung S24 (described below). ## Context Currently, ET-VK will encode all prepacking shaders, and then perform prepacking by submitting only one command buffer. However, this approach has some drawbacks: * CPU/GPU parallelism is decreased, since the command buffer is submitted only after all commands have been encoded. * There can be performance issues at the Vulkan API level when processing a single "large" command buffer. By splitting up prepacking to occur over multiple command buffers, performance can be improved by avoiding both the aforementioned issues. ## Llama 3.2 1B crash on Samsung S24 I have also noticed that running large models (i.e. Llama 3.2 1B) on the Samsung S24 with ET-VK, the device's display will crash (causing the screen to go black and become unresponsive), and sometimes the device will shut down entirely. Fortunately, this change also fixes this behaviour, in addition to providing a significant performance boost to model load time for Llama models (from 9s to 3s). ## Performance Impact * Improves model load time, especially on larger models. ## Future Work * Deprecate the `encode_prepack()` + `prepack()` pattern in favor of the `run_prepack()` pattern ghstack-source-id: 296437695 @exported-using-ghexport Differential Revision: [D78275586](https://our.internmc.facebook.com/intern/diff/D78275586/)
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/12533
Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
SS-JIA
approved these changes
Jul 16, 2025
This PR needs a
|
lucylq
pushed a commit
that referenced
this pull request
Jul 17, 2025
## Changes * Introduce `run_prepack()` API which combines the functionality of `encode_prepack()` and `prepack()`, but submits prepacking shaders incrementally rather than all at once. * Introduce graph config options to control command buffer submission behaviour during prepacking. Note that the current default values for the prepack submission thresholds were determined through experimentation. I will leave determining optimal values for specific devices as a later exercise. The goal of this diff is simply to introduce this mechanism to fix the Llama model loading crash on Samsung S24 (described below). ## Context Currently, ET-VK will encode all prepacking shaders, and then perform prepacking by submitting only one command buffer. However, this approach has some drawbacks: * CPU/GPU parallelism is decreased, since the command buffer is submitted only after all commands have been encoded. * There can be performance issues at the Vulkan API level when processing a single "large" command buffer. By splitting up prepacking to occur over multiple command buffers, performance can be improved by avoiding both the aforementioned issues. ## Llama 3.2 1B crash on Samsung S24 I have also noticed that running large models (i.e. Llama 3.2 1B) on the Samsung S24 with ET-VK, the device's display will crash (causing the screen to go black and become unresponsive), and sometimes the device will shut down entirely. Fortunately, this change also fixes this behaviour, in addition to providing a significant performance boost to model load time for Llama models (from 9s to 3s). ## Performance Impact * Improves model load time, especially on larger models. ## Future Work * Deprecate the `encode_prepack()` + `prepack()` pattern in favor of the `run_prepack()` pattern Differential Revision: [D78275586](https://our.internmc.facebook.com/intern/diff/D78275586/)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
This PR was created by the merge bot to help merge the original PR into the main branch.
ghstack PR number: #12442 by @SS-JIA
^ Please use this as the source of truth for the PR details, comments, and reviews
ghstack PR base: https://github.com/pytorch/executorch/tree/gh/SS-JIA/257/base
ghstack PR head: https://github.com/pytorch/executorch/tree/gh/SS-JIA/257/head
Merge bot PR base: https://github.com/pytorch/executorch/tree/main
Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/SS-JIA/257/orig
@diff-train-skip-merge