Stephen Roller
Stephen Roller
Good catch. I’d welcome a PR
https://github.com/facebookresearch/ParlAI/blob/755b9dcb778deb5a82029d69ae3260579c6450f1/parlai/core/build_data.py#L515 This except is catching too broadly an import error created by when we load the checkpoint here: https://github.com/facebookresearch/ParlAI/blob/755b9dcb778deb5a82029d69ae3260579c6450f1/parlai/agents/bart/convert_fairseq_to_parlai.py#L254 The easiest solution is to change the latter to a RunTimeError.
@moyapchen it must be one of the model files you released was trained with -m fb:bart or whatever and needs to be overridden. Can you look at this?
The following script produces the exact same output for me, regardless of batchsize: ```python #!/usr/bin/env python3 BS = 2 from parlai.core.agents import create_agent_from_model_file agent = create_agent_from_model_file( "zoo:blender/blender_3B/model", {'model_parallel': False} )...
We must be using different models. Mine never said it has a dog. I expect small floating point errors, but BS=50 is wildly off jeez. BS=2 is too. What GPU...
Hm, we haven't tried pytorch 1.8 in open source land yet, so I can't vouch for that. We've used in plenty in internal use cases and not had problems, but...
With the reversion and BS-50 I still observe a weird few observations