Export fails: sdpa_with_kv_cache not registered despite clean build from source (_portable_lib missing)

Issue

Despite following both:

• the official ExecuTorch “Building from source” guide:
  https://pytorch.org/executorch/stable/using-executorch-building-from-source.html
• and Arm’s LLama 3 Chat App tutorial for ExecuTorch/XNNPACK:
  https://learn.arm.com/learning-paths/mobile-graphics-and-gaming/build-llama3-chat-android-app-using-executorch-and-xnnpack/4-prepare-llama-models/

I was unable to export a Llama model using the SDPA path. The custom op torch.ops.llama.sdpa_with_kv_cache is never registered, and _portable_lib.so is missing entirely from the build artifacts. The exporter crashes as soon as it tries to import executorch.extension.llm.custom_ops.

---

Environment

Component	Version
OS	Ubuntu 22.04 LTS
Python	3.10.18 (venv)
CMake	3.22.1
GCC	11.4
Ninja	1.11.1
Git + LFS	git 2.34.1 / git-lfs 3.4.0
Buck2	Auto-downloaded (2024-02-15-x86_64)
ExecuTorch version	0.6.0 (cloned from main branch)

---

Steps followed

git clone --recursive https://github.com/pytorch/executorch.git
cd executorch
git submodule sync --recursive
git submodule update --init --recursive --jobs 8
git lfs pull

python3 -m venv .venv
source .venv/bin/activate
pip install -U pip
./install_requirements.sh         # also tried: pip install -e ".[examples]"

./install_executorch.sh --clean
./install_executorch.sh           

Everything completes without errors. A wheel is produced and installed into the venv.

---

What goes wrong

Running this sanity check:

from executorch.extension.llm.custom_ops import custom_ops
import torch
print("SDPA op present:", hasattr(custom_ops, "sdpa_with_kv_cache"))
print("TorchAO op present:", hasattr(torch.ops.torchao, "dequantize_affine"))

Prints:

SDPA op present: False
TorchAO op present: True

Further:

>>> from executorch.extension.llm.custom_ops import custom_ops
Traceback (most recent call last):
  File ".../custom_ops.py", line 20, in <module>
    op = torch.ops.llama.sdpa_with_kv_cache.default
AttributeError: '_OpNamespace' 'llama' object has no attribute 'sdpa_with_kv_cache'

...

ModuleNotFoundError: No module named 'executorch.extension.pybindings._portable_lib'

Unzipping the wheel confirms: _portable_lib.so is missing entirely.

---

ARM's Tutorial Failure

Exporting Llama-3.2-1B fails immediately at the attention replacement phase:

python -m executorch.examples.models.llama.export_llama \
  --checkpoint ~/.llama/checkpoints/Llama3.2-1B-Instruct/consolidated.00.pth \
  --params     ~/.llama/checkpoints/Llama3.2-1B-Instruct/params.json \
  -kv --use_sdpa_with_kv_cache -X --xnnpack-extended-ops \
  --max_seq_length 1024 --max_context_length 32000 \
  --output_name llama3.pte ...

Fails with:

ModuleNotFoundError: No module named 'executorch.extension.pybindings._portable_lib'
AttributeError: '_OpNamespace' 'llama' object has no attribute 'sdpa_with_kv_cache'

---

Expected

from executorch.extension.llm.custom_ops import custom_ops
assert hasattr(custom_ops, "sdpa_with_kv_cache")  # → True

and the exporter should succeed with --use_sdpa_with_kv_cache.

Versions

PyTorch version: 2.7.0+cpu
Is debug build: False
CUDA used to build PyTorch: Could not collect
ROCM used to build PyTorch: N/A

OS: Ubuntu 24.04.2 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-9ubuntu1) 11.4.0
Clang version: Could not collect
CMake version: version 3.31.6
Libc version: glibc-2.39

Python version: 3.10.18 (main, Jun 4 2025, 08:56:00) [GCC 13.3.0] (64-bit runtime)
Python platform: Linux-6.8.0-51-generic-x86_64-with-glibc2.39
Is CUDA available: False
CUDA runtime version: 12.6.85
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 2070
Nvidia driver version: 560.35.05
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.10.2
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.10.2
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.10.2
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.10.2
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.10.2
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.10.2
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.10.2
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.10.2
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

Versions of relevant libraries:
[pip3] numpy==2.2.6
[pip3] torch==2.7.0+cpu
[pip3] torchao==0.10.0+git8b264ce1
[pip3] torchaudio==2.7.0+cpu
[pip3] torchdata==0.11.0
[pip3] torchsr==1.0.4
[pip3] torchtune==0.6.1
[pip3] torchvision==0.22.0+cpu