Improve QNN backend build-from-source user experience

[abhinaykukkadapu]

Stack from ghstack (oldest at bottom):

• -> Improve QNN backend build-from-source user experience #17990
• Improve QNN backend AOT user experience #17989

• Remove EXECUTORCH_BUILD_WHEEL_DO_NOT_USE gate in CMakeLists.txt so
  QNN SDK auto-downloads for all build paths (editable installs,
  build.sh), not just wheel builds
• Use shared cache dir (~/.cache/executorch/qnn/) instead of cmake
  build dir, so SDK survives build dir cleans and is shared across
  all build flows
• Remove hard QNN_SDK_ROOT requirement from build.sh — cmake handles
  auto-download during configure when SDK is not set
• Auto-download Android NDK in build.sh via install_qnn_sdk.sh when
  ANDROID_NDK_ROOT is not set
• Fix PYTHON_EXECUTABLE default in build.sh (was checked before being
  used, with a buggy test that never triggered)
• Default EXECUTORCH_BUILD_QNN to ON on Linux x86 in pybind preset,
  matching wheel behavior so editable installs build QNN out of the box
• Redirect download_qnn_sdk.py progress output to stderr when
  --print-sdk-path is used, so cmake can capture the path cleanly

Test Plan/ User journeys

Verified end-to-end on Linux x86_64 with QNN SDK v2.37.0.250724.

AOT user: pip install executorch and import

First-time import (no SDK on disk):
Unset QNN_SDK_ROOT, clear ~/.cache/executorch/qnn/, then:

>>> import executorch.backends.qualcomm
[QNN] Downloading Qualcomm AI Runtime SDK v2.37.0.250724...
[QNN] Downloading: 645/1291 MB (50%)
[QNN] Downloading: 1291/1291 MB (100%)
[QNN] Download complete.
[QNN] Extracting SDK...
[QNN] Downloading libc++ (LLVM 14.0.0)...
[QNN] QNN SDK v2.37.0.250724 ready at ~/.cache/executorch/qnn/sdk-2.37.0.250724

User sees download progress instead of a silent multi-minute hang. SDK is stored in ~/.cache/, not inside the pip package directory.

Second import (SDK already cached):

>>> import executorch.backends.qualcomm
[QNN] Using cached QNN SDK v2.37.0.250724 at ~/.cache/executorch/qnn/sdk-2.37.0.250724

Instant — no re-download.

User has their own SDK installation:

$ export QNN_SDK_ROOT=/opt/qnn/2.37.0
>>> import executorch.backends.qualcomm
[QNN] Using QNN SDK at /opt/qnn/2.37.0 (from QNN_SDK_ROOT)

Respects the user's path, no download attempted.

Scripted usage (--print-sdk-path):

$ SDK_PATH=$(python3 backends/qualcomm/scripts/download_qnn_sdk.py --print-sdk-path)
$ echo $SDK_PATH
/home/user/.cache/executorch/qnn/sdk-2.37.0.250724

Stdout contains only the path — progress output goes to stderr, so shell capture works cleanly.

Build-from-source user: pip install -e . or build.sh

Zero-config build (no env vars set):

$ pip install -e . --no-build-isolation
  ...
  -- QNN_SDK_ROOT not set. Auto-downloading QNN SDK...
  -- EXECUTORCH_BUILD_QNN: ON
  ...
  Building target qnn_executorch_backend

QNN backend is built automatically. The SDK is auto-downloaded to ~/.cache/ during CMake configure.

Opting out of QNN:

$ CMAKE_ARGS="-DEXECUTORCH_BUILD_QNN=OFF" pip install -e . --no-build-isolation
  ...
  -- EXECUTORCH_BUILD_QNN: OFF

No SDK download, no QNN targets built.

Build with pre-existing SDK:

$ export QNN_SDK_ROOT=~/.cache/executorch/qnn/sdk-2.37.0.250724
$ pip install -e . --no-build-isolation
  ...
  -- EXECUTORCH_BUILD_QNN: ON
  Building target qnn_executorch_backend

Uses the provided SDK. No "Auto-downloading" message.

Shell build (build.sh) follows the same pattern:

$ bash backends/qualcomm/scripts/build.sh --skip_linux_android
[QNN] QNN_SDK_ROOT not set. SDK will be auto-downloaded during cmake configure.
  -- Auto-downloading QNN SDK...
  -- EXECUTORCH_BUILD_QNN: ON

With QNN_SDK_ROOT set, it skips the download and uses the provided path.

SDK survives build directory clean:

$ rm -rf build-x86/
$ bash backends/qualcomm/scripts/build.sh --skip_linux_android

No re-download — SDK lives in ~/.cache/, independent of the build directory.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17990

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 1 New Failure, 3 Unrelated Failures

As of commit 716c04b with merge base 043a3a1 ():

NEW FAILURE - The following job has failed:

• Build Presets / apple (profiling) / build (gh)
  The process '/opt/homebrew/bin/git' failed with exit code 128

BROKEN TRUNK - The following jobs failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• pull / test-moshi-linux / linux-job (gh) (trunk failure)
  E: Failed to fetch http://security.ubuntu.com/ubuntu/pool/main/libs/libssh/libssh-gcrypt-4_0.9.6-2ubuntu0.22.04.6_amd64.deb 404 Not Found [IP: 185.125.190.82 80]
• pull / unittest / windows / windows-job (gh) (trunk failure)
  ##[error]The operation was canceled.
• pull / unittest-editable / windows / windows-job (gh) (trunk failure)
  ##[error]The operation was canceled.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

[abhinaykukkadapu]

Hi @cccclai, @haowhsu-quic, @shewu-quic, any thoughts or improvements you have in mind to improve user experience of ET- Qualcomm. Please see the userflows in both the PRs. Thanks!

[winskuo-quic]

Hi @abhinaykukkadapu,
Thanks for improving QNN backend build experience.
For the automated download QNN SDK flow, I would like to know if it exports QNN_SDK_ROOT as environment variable after downloading QNN SDK. If not, do you think export QNN_SDK_ROOT for user will improve the user experience? The reason is that in Qualcomm python example scripts, we will try to get QNN_SDK_ROOT from env, like following: (

executorch/examples/qualcomm/scripts/inception_v3.py

Line 77 in 2cd88e7

qnn_sdk=os.getenv("QNN_SDK_ROOT"),

)

[abhinaykukkadapu]

Hi @abhinaykukkadapu, Thanks for improving QNN backend build experience. For the automated download QNN SDK flow, I would like to know if it exports QNN_SDK_ROOT as environment variable after downloading QNN SDK. If not, do you think export QNN_SDK_ROOT for user will improve the user experience? The reason is that in Qualcomm python example scripts, we will try to get QNN_SDK_ROOT from env, like following: (

executorch/examples/qualcomm/scripts/inception_v3.py

Line 77 in 2cd88e7

qnn_sdk=os.getenv("QNN_SDK_ROOT"),

)

Thanks, so, right now we set the QNN_SDK_ROOT, LD_LIBRARY_PATH for the python process, if someone runs this script, there are imports such as from executorch.backends.qualcomm so that specific python process would get it.

[abhinaykukkadapu]

@haowhsu-quic bumping this PR up for review thanks, i will merge both at once including #17989

[haowhsu-quic]

Could you keep this for Yocto linux toolchain guide? thanks.