In open-rs training setting, if I use a single machine with two gpus, the model would be automatically loaded into cpu memory, which leads to parameters in cpu, but data in gpu, this results in an error. Howerver, if I use a single machine with one gpu, here is an error arising from accelerate/accelerator.py in line 2321, self.deepspeed_engine_wrapped.backward(loss, **kwargs), with error info AttributeError: 'NoneType' object has no attribute 'backward'. I wonder if the code is complete, or the version of packages matters. If possible, can you release the env info in your training machine? Wish you a good day.
In open-rs training setting, if I use a single machine with two gpus, the model would be automatically loaded into cpu memory, which leads to parameters in cpu, but data in gpu, this results in an error. Howerver, if I use a single machine with one gpu, here is an error arising from accelerate/accelerator.py in line 2321, self.deepspeed_engine_wrapped.backward(loss, **kwargs), with error info AttributeError: 'NoneType' object has no attribute 'backward'. I wonder if the code is complete, or the version of packages matters. If possible, can you release the env info in your training machine? Wish you a good day.