Installation Instruction

1. CPU installation (default)
   pip install its-hub

1. CUDA installation (with vLLM support)
   pip install 'its-hub[vllm]'{}

Requested Package Name and Version

dependencies = [
    "openai>=1.75.0",
    "tqdm>=4.65.0",
    "typing-extensions>=4.0.0",
    "reward-hub==0.1.5",
    "transformers==4.53.2",  # Pin to exact version that worked in CI to avoid aimv2 config conflict with vLLM 0.9.1
    "backoff>=2.2.0",
    "click>=8.1.0",
    "fastapi>=0.115.0",
    "uvicorn<0.30.0",
    "pydantic>=2.0.0",
    "numpy>=1.24.0",
    "requests>=2.28.0",
    "aiohttp>=3.8.0",
    "litellm>=1.0.0",
]

Optional extras (for CUDA builds):

pip install 'its-hub[vllm]'

Repository:
🔗 https://github.com/Red-Hat-AI-Innovation-Team/its_hub
📦 Release v0.3.1

Brief Explanation for Request

New updates to its-hub package have been released and we would like to update the package build [both CPU and CUDA build] to new version in the RH PyPi Mirror Index(ices). Please refer to parent issue for original build request.

Summary of Changes

• Resolved CPU build issues: vLLM has been made optional, removing the GPU dependency barrier for CPU environments.

• Improved installation flexibility:

• • CPU build: pip install its-hub

• • CUDA build: pip install 'its-hub[vllm]'

• Dependencies cleaned up: Removed accelerate and reduced non-essential packages.

• Enhanced functionality:

• • Added Inference-Time-Scaling (ITS) as a configurable API endpoint.

• • Updated notebooks to demonstrate ITS self-consistency algorithms.

These changes ensure broader deployability of ITS-Hub in enterprise and on-prem environments where GPU hardware may not be available.

QE User Acceptance Tests

Objective:
Validate that inference-time-scaling (ITS) works consistently across both CPU and CUDA builds, with correct accuracy-compute tradeoff behavior.

Testing Focus:

• ITS inference accuracy scaling via API endpoint

• Compatibility with FastAPI and litellm

• Correct operation in CPU-only mode (no vLLM dependency)

• Optional CUDA path validation with [vllm] extras

Expected Outcome:
ITS endpoints should perform deterministically and allow users to adjust compute scaling during inference for improved accuracy.

Package License

License: Apache License 2.0
Compliance: Approved per Fedora Allowed Licenses List