With v0.9.0 I cannot instantiate DeepSpeed. I get the error "AttributeError: 'NoneType' object has no attribute 'optimizer_name'"

[ilanshib]

Describe the bug With v0.9.0 I cannot instantiate DeepSpeed. I get the error "AttributeError: 'NoneType' object has no attribute 'optimizer_name'"

To Reproduce Here is the code that used to work when istantiating DeepSpeed:

class Engine(DeepSpeedEngine): def init(self, *args, **kwargs): ... ... super().init(None, *args, **kwargs) ... ...

Expected behavior A usable DeepSpeed instance will be created.

Actual result* I get the following error traceback: .... .... File "/usr/local/lib/python3.10/site-packages/deepspeed/runtime/engine.py", line 264, in init self._do_sanity_check() File "/usr/local/lib/python3.10/site-packages/deepspeed/runtime/engine.py", line 988, in _do_sanity_check if self.optimizer_name() is not None: File "/usr/local/lib/python3.10/site-packages/deepspeed/runtime/engine.py", line 648, in optimizer_name return (self.client_optimizer.class.name if self.client_optimizer else self._config.optimizer_name) AttributeError: 'NoneType' object has no attribute 'optimizer_name'

It seems that cfg is now mandatory during insanitation. It wasn't mandatory before.

[mrwyattii]

With the release of 0.9.0, we slightly changed the params for DeepSpeedEngine initialization: https://github.com/microsoft/DeepSpeed/blob/4de4d2acc6077707c806a8abbc26fba1e2779745/deepspeed/init.py#L142

Specifically, we moved the creation of the DeepSpeedConfig outside of engine and expect that to be passed in as config_class. Could you please share more about your use case? Typically users will interact with DeepSpeed via deepspeed.initialize. Knowing more about how you are initializing DeepSpeed will help us avoid breaking changes in the future.

[rituk]

I'm facing the same "optimzer_name" error.

[JonathanColetti]

me aswell. Has anyone solved this