[STORY] Enable Caikit Python API library

[heyselbi]

Caikit python library - so it can be accessed from a notebook. It would be a wrapper around grpcio/requests for the API. It can be pip installed in the notebook.

Task includes:

• Develop an abstraction layer a python client API to access the Caikit core APIs from a notebook. Could we do a REST interface with swagger enablement? KServe APIs already supports swagger.
• Look into generating pb2 files during runtime. Is this an option? Is there an impact on inferencing performance?
• Create a Caikit Python API documentation that is data scientist persona focused.
• Long term: can we make these APIs be compatible with OpenAI APIs and HG TGI APIs?

Related issues:

• handle self signed certificates for http client caikit-nlp-client#7
• handle self signed certificates for grpc client caikit-nlp-client#8
• Get tests working for the streaming endpoints caikit-nlp-client#6
• ci: setup PyPi releases caikit-nlp-client#9
• docs/examples caikit-nlp-client#10

[heyselbi]

Useful reference: https://github.com/rh-aiservices-bu/llm-on-openshift/blob/main/examples/caikit/caikit_grpc_query_example.ipynb
cc @guimou

[danielezonca]

Note: the code of the example is still WIP so check with @guimou before doing code changes

[guimou]

Yeah, I have a few changes that I should make by the end of the day. Principally for the channel timeout, and change some parameters to remove anything hard coded.

…

On Mon., Oct. 2, 2023, 12:56 Daniele Zonca, ***@***.***> wrote: Note: the code of the example is still WIP so check with @guimou <https://github.com/guimou> before doing code changes — Reply to this email directly, view it on GitHub <#15 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA6C4YSUX7F3IEFOCUXZR63X5LW3TAVCNFSM6AAAAAA5PUMB66VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTONBTGM4TSOJQGI> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[guimou]

Quick note on this "Look into generating pb2 files during runtime. Is this an option? Is there an impact on inferencing performance?".
The bad news is that the Python grpc-reflection package does not implement on-the-fly stub generation. It's builtin in Java, available in Go through a third-party package, but nothing in Python. You can only retrieve the protofiles. Of course you could make a shell call to protoc, but that's really dirty, and you'd have to bundle protoc as well, for all architectures.
At the moment it's easier to keep the pb2 files...

[guimou]

This is also an interesting avenue. A wrapper around different serving providers to allow direct use of OpenAI API: https://github.com/BerriAI/litellm

[krrishdholakia]

Hey @guimou - i'm the comaintainer of litellm. Happy to help out via PR. What's the problem you're hoping to solve here with litellm?

[z103cb]

@Xaenalt and @heyselbi,
@vaibhavjainwiz and I have met discussed how we think we should be tackling this issue (I am also summarising what we discussed on slack).

The library requirements:

1. Creating a connection, specify HTTP or gRPC, optionally provide a cert to trust (required for gRPC with self-signed certs, there's more we can discuss about potentially fixing that upstream too)
2. Calling TextGenerationTaskPredict and the streaming version on that http/grpc connection, with inputs like fn_name(text="sometext")
3. Ensuring that we can pass other keyword args to it, like min_new_tokens, max_new_tokens, etc

The implementation:

1. The library will be housed in a new repo under the Caikit organisation. Tentatively named Caikit-nlp-client.
2. In the library we will provide a mechanism to create from .proto files static _pb2.py which will be used to provide the serialisation mechanisms and GRPC client (stub).
3. The .proto will be generated from executing:

RUNTIME_LIBRARY=caikit_nlp python -m caikit.runtime.dump_services $grpc_interface_dir

4. The _pb2.py files will be generated from executing the python generation from .proto (the following example command line might not be 100% accurate):

python -m grpc_tools.protoc -I./grpc/ --python_out=. --pyi_out=. --grpc_python_out=. grpc/*.proto

5. On top of the generated python we will write a client class to provide a simple and straightforward way to make the GRPC calls to the NLP service.

6. For now we plan on using the generated _pb2.py DTO (request and responses) as the model for the HTTP client. I am not 100% sure that using those objects for the http client would work (a cursory google search, stack overflow would indicate that is possible). I will need to prototype that to make sure it would work.

7. Provide for some automated tests:

• Spin up the Caikit NLP server on the test machine (no certificates) with at least two simple models running
• Connect to the running server
• Exercise the endpoints, focus of error conditions (un-happy path).

[vaibhavjainwiz]

Implement insecure HTTP client
https://github.com/vaibhavjainwiz/caikit-nlp-client/pull/31

[dtrifiro]

Initial implementation (wip): opendatahub-io/caikit-nlp-client#1

[z103cb]

@heyselbi I think we should still keep this open (but I will defer) to your better judgement.

[dtrifiro]

First version (0.0.2) was released on PyPi https://pypi.org/project/caikit-nlp-client/. See https://github.com/opendatahub-io/caikit-nlp-client/releases for releases.