transformer-deploy
transformer-deploy copied to clipboard
ValueError: Unrecognized configuration class <class 'transformers.models.t5.configuration_t5.T5Config'> for this kind of AutoModel: AutoModelForSequenceClassification.
Hi - I am trying to accelerate some T5 models and I get this error. How do I fix this?
Command to reproduce:
convert_model -m "valhalla/t5-small-qa-qg-hl" --backend tensorrt onnx --seq-len 16 16 16
I understand that this requires a AutoModelForSeq2SeqLM, or MT5ForConditionalGeneration class. How do I add these and how to add a new --task seq2seq
apart from the existing classes? I get this error when I add a new task seq2seq
-task: invalid choice: 'seq2seq' (choose from 'classification', 'embedding', 'text-generation')
Thanks
Convert command doesn't manage t5 for now, you may follow the notebook to get the trt model
Any plans to add T5 support for the convert command in the near future?
Yes the Triton server with onnx 1.12 (required for that) has been released few days ago and we were waiting for it. However I have no ETA to share for now.