RuntimeError: CUDA error: device-side assert triggered

[A91A981E]

❓ Questions and Help

Happy Chinese New Year!
I tried to train this model with VG. I followed README to get start and met some problem with mix precision. So I use float32. When process went to 4812-th iteration with 12 batch size, this error occurred. Full content as follow:

Traceback (most recent call last):
  File "/root/.vscode-server/extensions/ms-python.python-2021.2.633441544/pythonFiles/lib/python/debugpy/_vendored/pydevd/pydevd.py", line 3215, in <module>
  File "/root/.vscode-server/extensions/ms-python.python-2021.2.633441544/pythonFiles/lib/python/debugpy/_vendored/pydevd/pydevd.py", line 3208, in main
    globals = debugger.run(setup['file'], None, None, is_module)
  File "/root/.vscode-server/extensions/ms-python.python-2021.2.633441544/pythonFiles/lib/python/debugpy/_vendored/pydevd/pydevd.py", line 2282, in run
    return self._exec(is_module, entry_point_fn, module_name, file, globals, locals)
  File "/root/.vscode-server/extensions/ms-python.python-2021.2.633441544/pythonFiles/lib/python/debugpy/_vendored/pydevd/pydevd.py", line 2289, in _exec
    pydev_imports.execfile(file, globals, locals)  # execute the script
  File "/root/.vscode-server/extensions/ms-python.python-2021.2.633441544/pythonFiles/lib/python/debugpy/_vendored/pydevd/_pydev_imps/_pydev_execfile.py", line 25, in execfile
    exec(compile(contents+"\n", file, 'exec'), glob, loc)
  File "tools/relation_train_net.py", line 383, in <module>
    main()
  File "tools/relation_train_net.py", line 376, in main
    model = train(cfg, args.local_rank, args.distributed, logger)
  File "tools/relation_train_net.py", line 164, in train
    scaled_losses.backward()
  File "/root/miniconda3/envs/scene_graph_benchmark/lib/python3.8/site-packages/torch/_tensor.py", line 307, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
  File "/root/miniconda3/envs/scene_graph_benchmark/lib/python3.8/site-packages/torch/autograd/__init__.py", line 154, in backward
    Variable._execution_engine.run_backward(
RuntimeError: CUDA error: device-side assert triggered
CUDA kernel errors might be asynchronously reported at some other API call,so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.

I noticed the NOTE in relation_train_net.py line 161, so I commented out:

with amp.scale_loss(losses, optimizer) as scaled_losses:
        scaled_losses.backward()

and use

losses.backward()

It's not working... And error came to:

Traceback (most recent call last):                                                                                                               
  File "tools/relation_train_net.py", line 384, in <module>                                                                                      
    main()                                                                                                                                       
  File "tools/relation_train_net.py", line 377, in main                                                                                          
    model = train(cfg, args.local_rank, args.distributed, logger)                                                                                
  File "tools/relation_train_net.py", line 165, in train                                                                                         
    losses.backward()                                                                                                                            
  File "/root/miniconda3/envs/scene_graph_benchmark/lib/python3.8/site-packages/torch/_tensor.py", line 307, in backward                         
    torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)                                                           
  File "/root/miniconda3/envs/scene_graph_benchmark/lib/python3.8/site-packages/torch/autograd/__init__.py", line 154, in backward               
    Variable._execution_engine.run_backward(                                                                                                     
RuntimeError: cuDNN error: CUDNN_STATUS_EXECUTION_FAILED

Have anyone met this issue before?

[dyang-TUM]

@A91A981E Did you solve this problem?