Use ONNX models from HuggingFace

[seitzquest]

What does this PR do?

Changes onnx inference model's source from local file to HuggingFace repository

• Added "onnx_use_quantized" field to model_config to specify if quantized model should be used or not
• Removed model_path and replaced decoder_path with is_encoder_decoder to use Optimum's naming convention

We updated the existing onnx tests to use models from HF and added tests for quantized qa

We provide model configurations for extractive, categorical, multiple-choice and abstractive QA

• See "model-inference/model_inference/api.dev.local.http" for query examples

Minor additional changes:

• Corrected the generation task's return message in model-inference/model_inference/app/api/routes/prediction.py
• Removed ONNX section in model-inference/README.md -> onnx workers are now created like any other model type

Who can review?

@HaritzPuerto
@Rachneet

[timbmg]

should these services really be part of the "minimal" version of square?

[seitzquest]

We copied the dependencies from the tasb_worker. When I try to remove rabbit or redis from any of our services, I get the following errors during requests:

• amqp.exceptions.AccessRefused: (0, 0): (403) ACCESS_REFUSED - Login was refused using authentication mechanism PLAIN. For details see the broker logfile.
• redis.exceptions.ResponseError: WRONGPASS invalid username-password pair or user is disabled.

I think that 'minimal" in this case refers to the UI not being included

• Maybe we should rename the file to "docker-compose.ytt.no-frontend.yaml"

[timbmg]

Right, services like Rabbit, Redis, Mongo etc they are dependencies of other core services, so no need to remove them.

However, I was more talking about the additional models that are added here. IMHO, this file should just deploy a single model and single datastore. This is basically just a demo of the each component. Therefore, I think we do not need additional models in this file. But please pitch in @Rachneet or @kwang2049 if you have a different opinion.

[timbmg]

@Rachneet do you want these model services to be added to the docker-compose.ytt.yaml or should they be deployed dynamically?

[Rachneet]

They should be deployed dynamically. For testing purposes only, these services can be added to docker-compose.models.yaml file.

[timbmg]

Thanks for your work and PR! Just a couple of small changes and clarifications from my side. However, I have just looked superficially. @Rachneet can give more feedback.