Massive increase in memory consumption between 1.0 and 1.1 / 1.2

[IgorSwat]

🐛 Describe the bug

What changes were introduced between versions 1.0 and 1.1/1.2 that could explain why some models exported with XNNPACK delegates in ExecuTorch 1.1/1.2 consume nearly 3× more runtime memory compared to the same models exported with version 1.0?

I encountered this while exporting Kokoro modules using these scripts: first and second.

When profiling memory usage on Android (arm64-v8a), the combined runtime usage for both models is ~800 MB when exported with ExecuTorch 1.0, but suddenly increases to ~2.2 GB when exported with versions 1.1 or 1.2, without any changes in the models.

For reference, the export configurations used are:

python -m export.export_duration_predictor --bundled=true --input-size=inp-16 --max-tokens=32,64,128 --pad-lstm=true --dynamic=true --dtype=fp32

python -m export.export_synthesizer --input-size=inp-16 --dynamic=true --max-tokens=128 --max-duration=296 --dtype=fp32

Both models are exported with dynamic shapes enabled and include the LSTM padding fix (which pads LSTMs to a static sequence length).

Versions

Collecting environment information...
PyTorch version: 2.11.0
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: macOS 26.1 (arm64)
GCC version: Could not collect
Clang version: 17.0.0 (clang-1700.3.19.1)
CMake version: version 3.31.10
Libc version: N/A

Python version: 3.12.11 (main, Jun 3 2025, 15:41:47) [Clang 17.0.0 (clang-1700.0.13.3)] (64-bit runtime)
Python platform: macOS-26.1-arm64-arm-64bit
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
Is XPU available: False
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
Caching allocator config: N/A

CPU:
Apple M4 Pro

Versions of relevant libraries:
[pip3] executorch==1.2.0a0+0b0e2c5
[pip3] numpy==2.0.0
[pip3] pytorch_tokenizers==1.2.0
[pip3] torch==2.11.0
[pip3] torchao==0.17.0+git02105d46c
[pip3] torchaudio==2.11.0
[pip3] torchdata==0.11.0
[pip3] torchsr==1.0.4
[pip3] torchtune==0.0.0
[pip3] torchvision==0.26.0
[conda] Could not collect

cc @GregoryComer @digantdesai @cbilgin

[GregoryComer]

Thanks for the detailed report. I'm not aware of any expected reason for memory regression. I'll run through the instructions and see if I can reproduce the memory differential.

[GregoryComer]

@IgorSwat I ran through the export on release/1.0 and release/1.1 and can't repro the memory increase on Android or Mac. I'm assuming this just means I'm not reproducing the issue you saw correctly. Here's how I did it. Can you clarify if there's anything different I should try? Thanks.

Export (from cloned kokoro-export repo):

  python -m export.export_synthesizer \
    --input-size=inp-16 --dynamic=true --max-tokens=128 \
    --max-duration=296 --dtype=fp32 --output-dir=/tmp/kokoro-repro/v1.0

  python -m export.export_duration_predictor \
    --bundled=true --input-size=inp-16 --max-tokens=32,64,128 \
    --pad-lstm=true --dynamic=true --dtype=fp32 --output-dir=/tmp/kokoro-repro/v1.0

Benchmark:

• Cross compiled executor_runner for Android arm64-v8a.
• Ran via adb on Pixel 7 Pro.

Measured data on a Pixel 7 Pro:

	v1.0 runner + v1.0 pte	v1.1 runner + v1.1 pte
Synthesizer (after load)	1412 MB	1308 MB
Synthesizer (after exec)	1591 MB	1521 MB
Synthesizer planned buffer	277 MB	173 MB
Duration Pred (after load)	1343 MB	1343 MB
Duration Pred (after exec)	1346 MB	1347 MB
Duration Pred planned buffer	4.0 MB	4.6 MB