Updated llama-cpp-python (and its dependencies with env variables)

[SnehalSwadhin]

Updated llama-cpp-python to 0.3.1 using Pypi.

Conda had 0.2.24 as the latest version. I was not able to run any ChatModel in langchain in this version (was getting an error that Llama.create_chat_completion() got an unexpected keyword argument 'logprobs')

I had to add a few environment variables and dependencies for updating llama-cpp-python

[klwetstone]

@SnehalSwadhin This represents a major change to the existing dependencies. Are you certain that llama-cpp-python 0.2.24 cannot be used for your solution? Is it possible to debug or work around the error in 0.2.24?

If not, it will take us some time to determine if all of these changes are compatible with solutions that worked in the previous runtime. This is a really impressive deep dive into advanced package installation with pixi 👏

[SnehalSwadhin]

I was able to get results with structured_output in langchain pretty easily with ollama (about a week). Just using llama.cpp has been a pain for more than a month now. I actually gave up at some point😅, but came back today to give it a last try.

Because it works on my device, its very difficult to debug. And I was not able to find any other fix for: Llama.create_chat_completion() got an unexpected keyword argument 'logprobs'

Anyway, it was fun figuring out something new🙌

[SnehalSwadhin]

At this point I just want to see one of my submission run successfully. Though the sad part is that it probably does not score as high as some submissions on the leaderboard.

[klwetstone]

If the change would be important to your solution we can certainly consider the PR, just let me know!

I also enjoyed reading your PR -- this is the first competition for which we've used Pixi to manage dependencies, and your code uses a lot Pixi functionality that I hadn't seen before.

FYI, there is a second DrivenData competition that uses the same data, but is an unsupervised challenge that does not rely on code execution for submissions. If you're interested, you could experiment with some of these methods for that challenge.

[SnehalSwadhin]

I've got a teammate working on the second track.
Let me try one more day with llama-cpp-python 0.2.24
If it still doesn't work, I'll tag you here.

[SnehalSwadhin]

@klwetstone Although I managed to bypass the logprobs error by overriding a method in ChatLlamaCpp, but my code ultimately does not seem to work in 0.2.24. Though it works without any issue in the newer version.

Invoking any llm with structured_output only returns None. I tried many combinations of different prompts, old & new models (mistral, meta-llama, granite, dolphin), and even tried the basic LamaCpp (non chat_models, which is deprecated to structured_output).

I'd request you to try and update the library.
If it seem to go beyond the rules of the competition, please do let me know. I appreciate your cooperation.