Skip to content

Serve in vLLM

This tutorial shows you how to deploy your trained speculator model for production inference using vLLM.

After training a speculator, deploying it with vLLM is straightforward. vLLM automatically recognizes the speculators_config in your model and enables speculative decoding.

Basic Serving

vllm serve ./checkpoints/checkpoint_best

That's it! vLLM will:

  1. Load your speculator model
  2. Read the speculators_config from config.json
  3. Load the target/verifier model
  4. Enable speculative decoding automatically

Long form command

vllm serve Qwen/Qwen3-8B \
  -tp 1 \
  --speculative-config '{
    "model": "RedHatAI/Qwen3-8B-speculator.eagle3",
    "num_speculative_tokens": 3,
    "method": "eagle3"
  }'

For the long form command, pass in the target model first (e.g. Qwen/Qwen3-8B) and then specify the draft model in the speculative config (e.g. RedHatAI/Qwen3-8B-speculator.eagle3).

The long form command can be used to change the num_speculative_tokens or to use a different target model with the speculator. This can be used to combine a quantized target model, with a speculative decoding model for even better performance.