vLLM Deployment#

Configure vLLM as the deployment backend for serving models during evaluation.

Configuration Parameters#

Basic Settings#

deployment:
  type: vllm
  image: vllm/vllm-openai:latest
  checkpoint_path: /path/to/model  # Model path (local or HuggingFace ID)
  served_model_name: your-model-name
  port: 8000

Parallelism Configuration#

deployment:
  tensor_parallel_size: 8
  pipeline_parallel_size: 1
  data_parallel_size: 1

  • tensor_parallel_size: Number of GPUs to split the model across (default: 8)

  • pipeline_parallel_size: Number of pipeline stages (default: 1)

  • data_parallel_size: Number of model replicas (default: 1)

Extra Arguments and Endpoints#

deployment:
  extra_args: "--max-model-len 4096"
  
  endpoints:
    chat: /v1/chat/completions
    completions: /v1/completions
    health: /health

The extra_args field passes extra arguments to the vllm serve command.

Complete Example#

defaults:
  - execution: slurm/default
  - deployment: vllm
  - _self_

deployment:
  checkpoint_path: /path/to/checkpoint
  served_model_name: llama-3.1-8b-instruct
  tensor_parallel_size: 1
  data_parallel_size: 8
  extra_args: "--max-model-len 4096"

execution:
  account: your-account
  output_dir: /path/to/output
  walltime: 02:00:00

evaluation:
  tasks:
    - name: ifeval
    - name: gpqa_diamond

Reference#

The following example configuration files are available in the examples/ directory:

  • lepton_vllm_llama_3_1_8b_instruct.yaml - vLLM deployment on Lepton platform

  • slurm_llama_3_1_8b_instruct.yaml - vLLM deployment on SLURM cluster

  • slurm_llama_3_1_8b_instruct_hf.yaml - vLLM deployment using HuggingFace model ID

Use nemo-evaluator-launcher run --dry-run to check your configuration before running.