[Executorch][LLM] Use caching allocator for runner #15730
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
We observed that on iOS it improves perf by 6% because SDPA op does temp allocations. No significant difference on android though. Differential Revision: [D86120038](https://our.internmc.facebook.com/intern/diff/D86120038/) [ghstack-poisoned]
kimishpatel
requested review from
jackzhxng,
kirklandsign,
larryliu0820 and
mergennachin
as code owners
This was referenced Nov 10, 2025
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/15730
Note: Links to docs will display an error until the docs builds have been completed. ❌ 122 New FailuresAs of commit 5cecbfc with merge base 7600df8 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Nov 10, 2025
kimishpatel
added a commit
that referenced
this pull request
Nov 11, 2025
We observed that on iOS it improves perf by 6% because SDPA op does temp allocations. No significant difference on android though. Differential Revision: [D86120038](https://our.internmc.facebook.com/intern/diff/D86120038/) ghstack-source-id: 322321964 Pull Request resolved: #15730
meta-cla
bot
added
the
CLA Signed
This PR needs a
|
We observed that on iOS it improves perf by 6% because SDPA op does temp allocations. No significant difference on android though. Differential Revision: [D86120038](https://our.internmc.facebook.com/intern/diff/D86120038/) [ghstack-poisoned]
Contributor
There was a problem hiding this comment.
Pull Request Overview
This pull request integrates a CPU caching allocator into the LLM runner to improve performance by reducing temporary memory allocation overhead during inference. According to the PR description, this change provides a 6% performance improvement on iOS for operations like SDPA that perform temporary allocations, though no significant difference was observed on Android.
Key changes:
- Added CPUCachingAllocator as a temporary memory allocator for Module instances
- Updated build system dependencies in targets.bzl, CMakeLists.txt files
- Configured allocator with a 10MB cache size for temporary allocations
Reviewed Changes
Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.
| File | Description |
|---|---|
| extension/llm/runner/targets.bzl | Added dependency on cpu_caching_allocator for Buck build system |
| extension/llm/runner/llm_runner_helper.cpp | Instantiated CPUCachingAllocator with 10MB cache for Module temp allocations |
| extension/llm/runner/CMakeLists.txt | Added extension_memory_allocator to runner dependencies for CMake |
| CMakeLists.txt | Added memory_allocator subdirectory to build when LLM runner is enabled |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| endif() | ||
|
|
||
| if(EXECUTORCH_BUILD_EXTENSION_LLM_RUNNER) | ||
| add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/memory_allocator/runner) |
There was a problem hiding this comment.
The path extension/memory_allocator/runner does not exist in the repository. The memory allocator CMakeLists.txt is located at extension/memory_allocator/CMakeLists.txt. This line should be:
add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/memory_allocator)
Suggested change
| add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/memory_allocator/runner) | |
| add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/extension/memory_allocator) |
|
|
||
| // Create the Module | ||
| std::unique_ptr<Module> module; | ||
| uint32_t max_cached_memory_size_bytes_ = 1024 * 1024 * 10; // 10MB |
There was a problem hiding this comment.
The hardcoded value of 10MB for the caching allocator size should be documented or made configurable. According to the PR description, this improves performance by 6% on iOS for SDPA op temp allocations, but different models or use cases may benefit from different cache sizes. Consider:
- Adding a comment explaining why 10MB was chosen
- Making this value configurable through a parameter or constant
- Documenting the performance implications in code comments
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.
Stack from ghstack (oldest at bottom):
We observed that on iOS it improves perf by 6% because SDPA op does temp allocations.
No significant difference on android though.
Differential Revision: D86120038