Update runtime packages

[klwetstone]

Update the runtime environment ahead of competition launch.

Removing / adding packages

To change the starting set of packages, I:

• incorporated recommendations from Monday's triage
• went through each package that was included in the runtime, checked its use, and removed any that seemed irrelevant
• briefly skimmed blog posts / past solutions to NLP challenges for commonly used packages

Packages removed:

• aiofiles
• aiohttp
• diskcache
• gensim
• keras
• lightgbm
• ray-default
• tensorflow

Packages added:

• bitsandbytes
• langchain
• llama-cpp-python
• nltk

Versioning

Steps used to update package versions:

• Removed all pinned versions (set to "*")
• Added a pypi dependency for vllm using the wheel compatible with cuda 11.8 (unfortunately it's not yet available on conda forge):

  [feature.base.pypi-dependencies]
  vllm = { url = "https:/vllm-project/vllm/releases/download/v0.4.0/vllm-0.4.0+cu118-cp310-cp310-manylinux1_x86_64.whl" }
  

• Ran pixi to solve dependencies
• Updated the pinned package versions to those compatible with vLLM. We are generally using more recent versions of all packages. Changes of note:
  • xformers 0.0.23 to 0.0.23.post1
  • pytorch 2.1.1 to 2.1.2
• Remove the vLLM dependency. So it will not be in the initial runtime, but the initial runtime will be compatible with it.

Comparison to pypi

Where do our versions differ from the most recent versions on pypi?

• numpy is 1.26.4, most recent is 2.1.0. v2 was only released in June 2024, so I'm not too suspicious about this resolving to v1
• We are using CUDA 11.8, CUDA 12 is the most recent version (we decided not to update). All other packages that are behind are likely a result of this
  • We use python 3.10, because there is a vllm wheel for CUDA 11.8 + python 3.10
  • We use torch 2.1.2, most recent is 2.4

The rest of the packages are the most recent version.

[klwetstone]

This is being replaced by #10 because of rebasing issues