[BART] Fail successfully when fairseq is not installed

[klshuster]

Bug description
When using BART models, or any models that require BART (e.g. some of the hallucination or BB2, several folks have posted that ParlAI fails with Could not find pretrained model in parlai.zoo.bart.bart_large or parlai.zoo.bart.build; we should, instead, raise an error and note that fairseq should be installed.

Related Issues:

• Could not find pretrained model in parlai.zoo.bart.bart_large or parlai.zoo.bart.build #3975
• How can I train a BART FiD model on custom data with gold retrieved passages? #3872
• Unable to use BlenderBot2 model #3808
• No module named 'parlai.zoo.bart.bart_large' with BlenderBot 2 interactive #3970

Reproduction steps
See #3975 for repro.

Expected behavior
ParlAI should fail with an appropriate error, providing instructions for installing fairseq

Logs
See #3975 for logs.

[github-actions]

This issue has not had activity in 30 days. Please feel free to reopen if you have more issues. You may apply the "never-stale" tag to prevent this from happening.

[avinashsai]

Hi @klshuster I would like to work on this issue. Can you please more info where to add the error message.

[stephenroller]

ParlAI/parlai/core/build_data.py

Line 515 in 755b9dc

except (ImportError, AttributeError) as exc:

This except is catching too broadly an import error created by when we load the checkpoint here:

ParlAI/parlai/agents/bart/convert_fairseq_to_parlai.py

Line 254 in 755b9dc

raise ModuleNotFoundError(

The easiest solution is to change the latter to a RunTimeError.

[avinashsai]

Thank you! I will work on it.

[Guest400123064]

Hi, I am not sure if I should put my question here. Basically, I installed ParlAI from source (v1.7.1) and installed Fairseq (v0.12.2) but I am still not able to run BB2 with the same error ModuleNotFoundError: No module named 'parlai.zoo.bart.bart_large', which is followed by File "/home/wangy49/Projects/empathic-bot/packages/ParlAI/parlai/core/torch_generator_agent.py", line 1599, in advance self.partial_hyps[path_selection.hypothesis_ids.long()], RuntimeError: indices should be either on cpu or on the same device as the indexed tensor (cpu)

When running parlai interactive -m bart, I only get the second error.

I cannot find any solution. I wonder if there could be any possible cause? Thank you very much in advance!

===========================================================================
EDITED: seems that torch v1.13 will cause this problem while torch 1.12 does not. But torch 1.12 gives a warning Unable to load ngram blocking on GPU: /home/wangy49/.cache/torch_extensions/py39_cu116/ngram_repeat_block_cuda/ngram_repeat_block_cuda.so: undefined symbol: _ZN8pybind116detail11type_casterIN2at6TensorEvE4loadENS_6handleEb

[klshuster]

Thanks for flagging -- i would ignore the warning, if it's just that. If you could file a new issue regarding the first error, that would be great (so that we can track it separately)

[ye7iaserag]

I'm also having the indices should be either on cpu or on the same device error while converting bart.large after download
Using wsl2 with 3080 also have the same warning for the ngram blocks loading failure right before the failure of conversion

[klshuster]

@ye7iaserag could you try downgrading to pytorch 1.12?