Invalid resource handle when deleting ctranslate2.Translator

[opennmt]

Invalid resource handle when deleting ctranslate2.Translator

[guillaumekln]

Thanks for reporting. When testing, I found another behavior that should be improved: #127.

Let me know if there are other issues like these. I did not test the device placement thoroughly.

[francoishernandez]

Thanks for the quick fix.
This seems to fix the python case, but not the translate cli though.
(Pulled repo and rebuilt docker image from scratch.)

[guillaumekln]

Right. The API usage is slightly different.

[francoishernandez]

The issue seem to be appearing again on master:

>>> import ctranslate2
>>> translator = ctranslate2.Translator("ende_ctranslate2", device="cuda", device_index=0)
>>> translator.translate_batch([["▁H", "ello", "▁world", "!"]])
[[{'score': -1.736800193786621, 'tokens': ['▁Hallo', '▁Welt', '!']}]]
>>> del translator
>>> translator = ctranslate2.Translator("ende_ctranslate2", device="cuda", device_index=1)
>>> translator.translate_batch([["▁H", "ello", "▁world", "!"]])
[[{'score': -1.736800193786621, 'tokens': ['▁Hallo', '▁Welt', '!']}]]
>>> del translator
terminate called after throwing an instance of 'std::runtime_error'
  what():  /root/ctranslate2-dev/src/primitives/cuda.cu:72: CUDA failed with error invalid resource handle
Aborted (core dumped)

The error does not happen when deleting the object right after building it, without performing any translation.

[guillaumekln]

Thanks. I can reproduce this.

I will look to provide a more robust fix to this issue.