Add h2oGPT Client

[this]

Fixes #80

Usage

from h2ogpt_client import Client

client = Client("http://0.0.0.0:7860")

# text completion
response = client.text_completion.create("Hello world")
response = await client.text_completion.create_async("Hello world")

# chat completion
chat_context = client.chat_completion.create()
chat = chat_context.chat("Hey!")
print(chat["user"])  # prints user prompt, i.e. "Hey!"
print(chat["gpt"])   # prints reply of the h2oGPT
chat_history = chat_context.chat_history()

[pseudotensor]

@this Looking great! Make sure you merge main into your PR as the langchain had breaking changes with 1 extra API item.

Also, gradio team fixed the Chat part:

gradio-app/gradio#4081 (comment)

I'll finish this PR now: #117

[pseudotensor]

@this How's this PR going?

[this]

How's this PR going?

@pseudotensor I'm currently working on the chat_completion API. I was hoping to include that in this PR, but we can merge this and I can create a new PR.

[pseudotensor]

We should avoid dup with rest of code as this is constantly extended. I moved to enum the original version. Is there some reason to stick to FOO_BAR = "foo_bar" format?

[this]

Is there some reason to stick to FOO_BAR = "foo_bar" format?

@pseudotensor I simply capitalized the value to derive the name of the enum. Id these names are not appropriate, please suggest good names.

All capitalized names were used by following the Python convention https://docs.python.org/3.8/library/enum.html#creating-an-enum. I noticed that PromptType enum names are all simple. Shall we change that?

[pseudotensor]

Yes, let's change to whatever is normal convention. Requires going through and changing conditionals where PromptType used.

[this]

I'll create a separate PR for this.

[pseudotensor]

For the list of parameters, can you (when building the client) verify that the parameters match those passed to evaluate, so it is up to date? i.e. using eval_func_param_names . It won't break things when they are not up to date if you are using string version of API, but would be good when building client that everything is synced so everything possible can be done with this client. Thanks!

[pseudotensor]

E.g. you can at build time check in python that set of inspection arguments for "create" is larger than set from eval_func_param_names . I do stuff like that in h2oGPT.

[this]

ya, we can do that. Shall we create a new PR for this? In order to make eval_func_param_names isolated (so that it can be used in the client) I had to make changes in places where eval_func_param_names is used. I fell like if I add those changes into this PR, it will clutter the PR.

[pseudotensor]

Let's merge! Thanks! Let's do more testing and refactors in another PR.

[pseudotensor]

Non-manual test failures on this branch:

============================================================================================================ short test summary info ============================================================================================================
FAILED tests/test_pipeline.py::test_export_copy - ImportError: cannot import name 'inject_newline' from 'test_output.h2oai_pipeline' (/home/jon/h2ogpt/test_output/h2oai_pipeline.py)
FAILED tests/test_requirements.py::test_requirements - ValueError: ['accelerate==0.19.0', 'transformers==4.28.1', 'git+https://github.com/huggingface/peft.git@3714aa2fff158fdfa637b2b65952580801d890b2']
FAILED tests/test_ui.py::test_newline_replace - AssertionError: assert 'You can use ... sorted list.' == 'You can use ... sorted list.'
=========================================================================== 15 failed, 99 passed, 36 skipped, 1 xfailed, 1 xpassed, 12 warnings in 3745.08s (1:02:25) ===========================================================================
(h2ollm) jon@pseudotensor:~/h2ogpt$ 

Unrelated to this PR.