Standalone pytorch build on windows 10 fails silently on imports after execution

• Nuitka version, full Python version, flavor, OS, etc. as output by this command (it does more than you think, and we are adding more all the time):

  python -m nuitka --version

0.7.6
Commercial: None
Python: 3.7.9 (tags/v3.7.9:13c94747c7, Aug 17 2020, 18:58:18) [MSC v.1900 64 bit (AMD64)]
Flavor: Unknown
Executable: C:\Users\begemot\build\build_test\venv\Scripts\python.exe
OS: Windows
Arch: x86_64
WindowsRelease: 10

Latest 0.8rc7 also gives same results

• How did you install Nuitka and Python

Python was installed using official installer, downloaded from python.org. Nuitka was installed in virturalenv using pip install nuitka for 0.7.6 and pip install -U "https://github.com/Nuitka/Nuitka/archive/develop.zip" for 0.8rc7

• The specific PyPI names and versions

  python -m pip freeze

numpy==1.21.5
torch==1.11.0

• Many times when you get an error from Nuitka, your setup may be special

“Hello world” works perfectly. Small sample exmaples, using numpy also work as expected.

• Also supply a Short, Self Contained, Correct, Example

import torch

print("hello torch")

• Provide in your issue the Nuitka options used

python -m nuitka --remove-output --plugin-enable=torch --plugin-enable=numpy --plugin-enable=pylint-warnings --windows-force-stdout-spec=stdout.log --windows-force-stderr-spec=stderr.log --standalone main_torch_min.py

• Note if this is a regression

This is my first try to use Nuitka, but it seemed from docs and issues that building torch worked fine.

• Error I get

Nuitka builds “hello torch” example without any errors.

And there is no error message after running compiled file, actually. EXE just fails silently. Nothing is printed. But “Application Error 1000” error is reported in windows event log (which contains no useful information also). stderr.log and stdout.log are also empty. Also tried compiling without --windows-force-*-spec keys with same result.

After turning on PYTHONVERBOSE=3 and PYTHONDEBUG=1 in environment it seems that execution stops on importing _bootlocale.py.

Full log

(venv) PS C:\Users\begemot\build\build_test\main_torch_min.dist> .\main_torch_min.exe
import _frozen_importlib # frozen
import _imp # builtin
import '_thread' # <class '_frozen_importlib.BuiltinImporter'>
import '_warnings' # <class '_frozen_importlib.BuiltinImporter'>
import '_weakref' # <class '_frozen_importlib.BuiltinImporter'>
# installing zipimport hook
import 'zipimport' # <class '_frozen_importlib.BuiltinImporter'>
# installed zipimport hook
import '_frozen_importlib_external' # <class '_frozen_importlib.FrozenImporter'>
import '_io' # <class '_frozen_importlib.BuiltinImporter'>
import 'marshal' # <class '_frozen_importlib.BuiltinImporter'>
import 'nt' # <class '_frozen_importlib.BuiltinImporter'>
import _thread # previously loaded ('_thread')
import '_thread' # <class '_frozen_importlib.BuiltinImporter'>
import _weakref # previously loaded ('_weakref')
import '_weakref' # <class '_frozen_importlib.BuiltinImporter'>
import 'winreg' # <class '_frozen_importlib.BuiltinImporter'>
import '_codecs' # <class '_frozen_importlib.BuiltinImporter'>
import 'codecs' # <class '_frozen_importlib.FrozenImporter'>
import 'encodings.aliases' # <class '_frozen_importlib.FrozenImporter'>
import 'encodings' # <class '_frozen_importlib.FrozenImporter'>
import 'encodings.utf_8' # <class '_frozen_importlib.FrozenImporter'>
import '_signal' # <class '_frozen_importlib.BuiltinImporter'>
import 'encodings.latin_1' # <class '_frozen_importlib.FrozenImporter'>
import '_abc' # <class '_frozen_importlib.BuiltinImporter'>
import 'abc' # <class '_frozen_importlib.FrozenImporter'>
import 'io' # <class '_frozen_importlib.FrozenImporter'>
# trying C:\Users\begemot\build\build_test\main_torch_min.dist\_bootlocale.cp37-win_amd64.pyd
# trying C:\Users\begemot\build\build_test\main_torch_min.dist\_bootlocale.pyd
# trying C:\Users\begemot\build\build_test\main_torch_min.dist\_bootlocale.py
# trying C:\Users\begemot\build\build_test\main_torch_min.dist\_bootlocale.pyw
# trying C:\Users\begemot\build\build_test\main_torch_min.dist\_bootlocale.pyc
# trying C:\Users\begemot\build\BUILD_~1\MAIN_T~2.DIS\_bootlocale.cp37-win_amd64.pyd
# trying C:\Users\begemot\build\BUILD_~1\MAIN_T~2.DIS\_bootlocale.pyd
# trying C:\Users\begemot\build\BUILD_~1\MAIN_T~2.DIS\_bootlocale.py
# trying C:\Users\begemot\build\BUILD_~1\MAIN_T~2.DIS\_bootlocale.pyw
# trying C:\Users\begemot\build\BUILD_~1\MAIN_T~2.DIS\_bootlocale.pyc
import 'encodings.ascii' # <class '_frozen_importlib.FrozenImporter'>
# trying C:\Users\begemot\build\build_test\main_torch_min.dist\_bootlocale.cp37-win_amd64.pyd
# trying C:\Users\begemot\build\build_test\main_torch_min.dist\_bootlocale.pyd
# trying C:\Users\begemot\build\build_test\main_torch_min.dist\_bootlocale.py
# trying C:\Users\begemot\build\build_test\main_torch_min.dist\_bootlocale.pyw
# trying C:\Users\begemot\build\build_test\main_torch_min.dist\_bootlocale.pyc
# trying C:\Users\begemot\build\BUILD_~1\MAIN_T~2.DIS\_bootlocale.cp37-win_amd64.pyd
# trying C:\Users\begemot\build\BUILD_~1\MAIN_T~2.DIS\_bootlocale.pyd
# trying C:\Users\begemot\build\BUILD_~1\MAIN_T~2.DIS\_bootlocale.py
# trying C:\Users\begemot\build\BUILD_~1\MAIN_T~2.DIS\_bootlocale.pyw
# trying C:\Users\begemot\build\BUILD_~1\MAIN_T~2.DIS\_bootlocale.pyc
(venv) PS C:\Users\begemot\build\build_test\main_torch_min.dist>

[lup-]

As a workaround I tried to include torch as source-code module with this minimal options

python -m nuitka ^
       --mingw64 ^
       --nofollow-import-to=torch ^
       --include-package=numpy.core.multiarray ^
       --include-data-dir=C:\Users\begemot\build\build_test\venv\Lib\site-packages\torch=torch ^
       --standalone ^
       main_torch_min.py

Works smoothly, but I had to include multiprocessing module in my example (--import-package and --import-module skipped this module because it was “unused” without following torch).

import multiprocessing
import torch

print("hello torch")

[kayhayen]

Latest Torch is supposed to work with the plugin activated on 0.8.1