#102 vLLM subclass implementation as inference backend and example updated.

[solankinitish]

Closes #102

Problem

The existing VLLM implementation (added in #106) was incomplete — it imported httpx mid-class to fetch max_model_len from the vLLM server at init time, making instantiation fail without a live server and introducing an undeclared dependency.

Changes

• Reimplemented VLLM as a clean subclass following the same pattern as Ollama — api_base and max_tokens passed explicitly, no network calls at init
• max_tokens defaults to 4096 and can be overridden by the user to match their model's context length
• Rate limiting and batch distribution handled by the existing LLM infrastructure
• Added vLLM to the LLM provider section in examples/Getting_started.ipynb

Usage

from datatune.llm.llm import VLLM

llm = VLLM(
    model_name="mistralai/Mistral-7B-Instruct-v0.1",
    api_base="http://localhost:8000/v1",
    max_tokens=4096
)