Error: invalid input 'encoder_attention_mask' about transformers.js

[xenova]

Could you provide the command used to convert it? Which task did you use? It may be because you are not exporting using -with-past (but that is just a guess).

The command I would use is:

python ./scripts/convert.py --model_id facebook/bart-large-cnn --from_hub --quantize --task seq2seq-lm-with-past

[spencekim]

Yep that's the exact command I used. If you'd like, I could upload the files to your HuggingFace repo (https://huggingface.co/Xenova/transformers.js/tree/main) to make sure they look ok.

[xenova]

No worries - I'll convert it on my side. Then it's most likely something I'd need to fix. The strange thing is, other model types (with the same model type, i.e., "bart") need the encoder attention mask (vs. what you are showing, which doesn't).

[xenova]

Okay yep I was able to reproduce it on my side. As I said above, it's very strange that some bart models need it, while others dont... I'll make it work either way :)

EDIT: Okay I got it working (basically I just made it only add it if explicitly required by the decoder:

However, the output I got was a little strange:
[{"summary_text":"TheTower is 324 metres ( \"). It is the tallest in The Eiffel Tower is the tallest. The tower is the second tallest free-standing structure in France."}]

I assume it has a different type of tokenization that isn't yet supported. I'm off to bed now, but I'll look into it tomorrow. Once that's fixed, I'll push the changes :)

[xenova]

Got it working :)
[{"summary_text":"The Eiffel Tower is the tallest structure in Paris. It is 324 metres (1,063 ft) tall, about the same height as an 81-storey building."}]

Will push changes soon.

[xenova]

Looks like the commit automatically closed the issue. Whoops.

Can you confirm the changes work (I think you can install an npm package from GitHub), and if so, I will push this in version 1.2.6

[spencekim]

Sorry for the late response. It works great!

Side note, any plans to set up Typescript types, and/or support onnxruntime-node? Both would be super useful. I'd be open to helping implement it as well.

[xenova]

Sorry for the late response. It works great!

Great!

Side note, any plans to set up Typescript types, and/or support onnxruntime-node? Both would be super useful. I'd be open to helping implement it as well.

Absolutely! We currently have a couple of people working on the typescript side of things (#26 and #28)

For onnxruntime-node, I suppose it would be a question of how to support both web and node versions of onnxruntime without duplicating code. If you have any ideas, or know how to do that, feel free to open up a PR!