vLLM Engine Initialization

The initialize_llm_engine step creates a vllm.LLM instance and stores it in the experiment context.

Installation

pip install --upgrade --pre crystallize-extras[vllm]

Usage

from crystallize_extras.vllm_step import initialize_llm_engine

init_engine = initialize_llm_engine(engine_options={"model": "mistral"})
pipeline = Pipeline([init_engine, generate_predictions()])

• The step runs during pipeline setup, storing a resource_factory under the context key (default llm_engine).
• Access the cached engine inside later steps: engine = ctx.get("llm_engine")(ctx).
• Pass a different context_key if you need multiple engines.

If vllm is not installed, the step raises an informative error suggesting pip install crystallize-extras[vllm].