Discrepancy between stated minimal runtime library size (50KB) and actual build output

[goka-wu]

Hi team,

I am trying to build Executorch from source to get the minimal runtime library. The documentation states that the minimal runtime library size is around 50kB (The original statement in the official documentation (https://docs.pytorch.org/executorch/stable/runtime-overview.html) reads: "The core runtime library is less than 50kB when built without kernels or backends." ).

However, when I built the project using the following command, I observed a significant size difference in the output libraries:

cmake .. -DEXECUTORCH_LOG_LEVEL=Fatal -DCMAKE_BUILD_TYPE=Release
make

Here are the file sizes I found:

• libexecutorch.a: 50kB
• libexecutorch_core.a: 174kB

Since libexecutorch.a depends on libexecutorch_core.a, this seems to contradict the documentation.

The source code is from main branch with the following commit info：

commit b22a2be9cef46730d1ec4d82a24c043910bc6a08 (HEAD -> main, origin/main, origin/HEAD)
Author: Oscar Andersson <[email protected]>
Date:   Tue Jun 17 10:00:19 2025 +0200

    Arm backend: Add support for per-channel quantization (#11752)
    
    - Adds support for per-channel quantization in TosaQuantizer and
    TosaBackend
    - Enables per-channel quantization for MobilneNetV2 test cases
    
    
    cc @digantdesai @freddan80 @per @zingo
    
    ---------
    
    Signed-off-by: Oscar Andersson <[email protected]>

Thank you for your time.

[GregoryComer]

Regarding libexecutorch_core.a, libexecutorch.a should depend on libexecutorch_core.a. If you're seeing it the other way around, that would be unexpected. Here's the relevant line from CMake:

executorch/CMakeLists.txt

Line 450 in 9053089

target_link_libraries(executorch PRIVATE executorch_core)

Here's how the runtime size is build for the size check in CI: https://github.com/pytorch/executorch/blob/main/test/build_size_test.sh. Setting -DEXECUTORCH_OPTIMIZE_SIZE=ON, disabling RTTI and exceptions, and disabling fatal level logging can potentially trim a little more size.

@digantdesai @lucylq Do you have any other suggestions on reducing runtime size?

[goka-wu]

Thank you for the reminder. You're right, libexecutorch.a does depend on libexecutorch_core.a. I've corrected my understanding and updated the issue description.

I checked the CI results on GitHub and found that the script you mentioned validates the size of the executorch program size_test which must be less than 51,752KB. This is not a constraint on the size of libexecutorch.a or libexecutorch_core.a.

The 50K mentioned on the official website refers to programs likesize_test, not to libexecutorch_core.a or libexecutorch.a, is that correct?

[lucylq]

Hi @goka-wu, the 50K refers to the core executorch components (without operators). Size test builds libexecutorch.a and then links it to a minimal executor runner. The executor runner can't really run any models because no operators are linked, but gives an approximation of the core runtime size.

See the size test sources:
https://github.com/pytorch/executorch/blob/main/test/CMakeLists.txt#L46-L55

Seconding @GregoryComer's comment on build flags to trim size.

[goka-wu]

Okay, I’ve got it. Thanks for your assistance.