Enables support for vLLM. To use, you must specify the model field in the task parameter of the pgml.transform function and you must add "backend": "vllm" in the task parameters. For example,

SELECT * FROM pgml.transform(
    task => '{"model":"tiiuae/falcon-7b","backend":"vllm"}'::JSONB,
    inputs => Array['hello']
);

A list of supported models for vLLM can be found here.

Only one vLLM model can be loaded per client connection process due to a limitation in vLLM. The first call to pgml.transform with a given model will load the model ("cold start"), but subsequent calls will use the cached model. If you change the specified model in the same client connection, the cached model will be replaced with the new one.