ultravox/inference/oaieval_registry/solvers/vllm_server_config.yaml

generation/direct/vllm:
  class: evals.solvers.providers.fixie.fixie_solver:FixieSolver
  args:
    api_base: http://localhost:8000/v1
    completion_fn_options:
      model: fixie-ai/ultravox
      extra_options:
        temperature: 0
        # empirically, setting frequency_penalty to 1.0 seems to get us closer to the HF results
        # but it's not clear why this is necessary
        # TODO: investigate and try to match HF/VLLM results without this
        # UPDATE: freq penalty 1.0 hurts performance on big-bench-audio eval for ultravox8b. Switching back to 0.0.
        frequency_penalty: 0.0