lm-eval run --model vllm --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct,tensor_parallel_size=1 --tasks hellaswag

Evaluates hellaswag using vLLM as the inference engine, which is dramatically faster than the transformers backend on big models. --model_args takes comma-separated key=value settings, here passing tensor_parallel_size=1. vLLM must be installed (`pip install vllm`) and the harness's vllm extra enabled.