Multiple models support

[CerineBnsd]

Description [ Updated ]

This pull request adds multiple models support.

What has been done so far:

• Replace the model attribute in Exchanger struct with a models: HashMap of models (The key is the uuid)
• Add an optional model_name argument to upload_model, (Not used for anything for instance, only saved in the InferenceModel object in the server side)
• Generate a uuid of the model, it is the key in the hashmap
• Return the uuid by upload_model
• Specifiy the uuid in run_model
• Add delete_model method to remove an uploaded model (by specifying its model_id)
• Add model_id to DataProof
• If no model_name / model_id are specified, default is used (to cover the existing cases).
• If the value of model_id passed to run_model is "default", (the default value), the last uploaded model is used.
• Edited test_distlbert to cover this features (except delete)

  What needs to be done

  OOM error handling in the server side.

Related Issue

Solves #39

Type of change

• [X] This change requires a documentation update
• [X] This change affects the client
• [X] This change affects the server
• [X] This change affects the API

How Has This Been Tested?

With a modified test_distelbert, and tested the other cases (like delete_model) locally

Checklist:

• [X] My code follows the style guidelines of this project
• [X] I have performed a self-review of my code
• [ ] I have commented my code, particularly in hard-to-understand areas
• [] My changes generate no new warnings
• [ ] I have updated the documentation according to my changes

[CerineBnsd]

@JoFrost When you have some time, would you please take a look at this so that it would be faster to integrate it tomorrow with Charles work? Thank youu!

[CerineBnsd]

A quick snippet on how it looks now (from the test)

        client.connect_server(
            addr="localhost",
            simulation=self.simulation,
            policy=policy_file,
            certificate=certificate_file,
        )

        response = client.upload_model(
            model=model_path,
            shape=inputs.shape,
            dtype=ModelDatumType.I64,
            model_name=self.model_name,
        )
        model_id = response.model_id #retreive the model_id

        response = client.run_model(run_inputs, model_id=model_id)

        client.delete_model(model_id)

[cchudant]

I merged this with master and i had some concerns. I talked a bit with Mehdi:

The models will take a lot of memory when running "upload_model" multiple times (during testing/getting started with thel ib

The solution we came up with was to deduplicate models based on their hash. If two models have the same hash, they will only be stored once in memory.

OOM (out of memory) handling

I will try to catch the panics and return an error, we'll see if it is practical/possible. If not, memory management / resource management will probably be left for future work (permissions model?)

Backward compatibility

We decided to break backward compability and update Daniel's blogpost. (add the model_id argument) The reason is that we would be forced to store the last UUID in the server side forever, which could be a security issue / technical dept long term.

Delete-all models / List-models commands

Left for future work.