mcloud_oaieval.yaml

# Ultravox eval with vllm configuration
name: ultravox-oaieval
image: mosaicml/composer:latest
compute:
  gpus: 8
  cluster: r18z1p1
integrations:
  - integration_type: git_repo
    git_repo: fixie-ai/ultravox
    git_branch: main
scheduling:
  max_duration: 4
command: |
  cd ultravox &&
  source $VENV_PATH/bin/activate &&
  poetry install --only main &&
  pip install vllm==0.10.0 &&
  pip install vllm --index-url https://wheels.vllm.ai/${VLLM_COMMIT} --no-deps --force-reinstall &&
  if [ ! -z "${MODEL_SOUP}" ]; then python3 -m ultravox.tools.model_averaging $MODEL_SOUP; fi &&
  python3 -m ultravox.inference.run_vllm_inference $EVAL_ARGS

env_variables:
  EVAL_ARGS: --evalset audio-core --model fixie-ai/ultravox-v0_6-llama-3_3-70b --push_to_hub False
  VLLM_COMMIT: 62965de5fe8be8e3622952a9b5cda86973cf9c51
  HF_HUB_ENABLE_HF_TRANSFER: 1
  # You can use MODEL_SOUP to average multiple models together.
  # Just make sure to set the right model_paths and update EVAL_ARGS to use updated_model
  # MODEL_SOUP: --model_paths <model_1> <model_2> <model_3> ... --output_path updated_model