Crash when adding any PyTorch layer during script execution (works in terminal)

[ilya112358]

I am trying to run neural network inference by calling a Python script, but the process crashes. After narrowing the issue down, it appears that the crash happens whenever a layer is added to a PyTorch model.

import qtm
import torch.nn as nn

class SingleEventModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc = nn.Linear(18,128) # if I comment this line, it does not crash

def test_model():
    m = SingleEventModel()
    print(m)

def add_menu():
    """Function for adding custom commands and setting up the menu in QTM"""
    qtm.gui.add_command("test_model")
    qtm.gui.set_command_execute_function("test_model", test_model)
    # Add menu and button
    menu_id = qtm.gui.insert_menu_submenu(None, "My menu")
    qtm.gui.insert_menu_button(menu_id, "test_model", "test_model")
    
if __name__ == "__main__":
    clear() # qtm terminal clear
    add_menu()

Pressing test_model button crashes the application.

The same code runs correctly when executed directly in a terminal.

OS: Windows 11
Pip list:

I tried several solutions from ChatGPT like:

os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
os.environ["OMP_NUM_THREADS"] = "1"
os.environ["MKL_NUM_THREADS"] = "1"

Didn't help.
Any guidance on what might cause a crash and how to avoid it?

[Henrik-Norgren]

Hi,

I tried running your script, and for me it seems to deadlock rather than crash. Looking through debugger didn't reveal much other than it being stuck somewhere within Python/PyTorch itself:

I'd have to build Python with debug symbols in order to get more information.

Apparently it works fine if I import torch in the terminal before clicking the button. The only difference is that the terminal runs in a background thread. Likewise delegating the command to a background thread also works, even if we wait for it:

global m
def test_model_async():
    global m
    m = SingleEventModel()

def test_model():
    t = threading.Thread(target=test_model_async)
    t.start()
    t.join()
    print(m)

So for whatever reason PyTorch can't run on the Main/GUI thread. I don't know what it does internally so can't really tell why that might be.

[ilya112358]

I see. I guess, I will use a dedicated thread, or a subprocess, or run it externally. Thanks.