| 1 | # Draft Models |
| 2 | |
| 3 | The following code configures vLLM in an offline mode to use speculative decoding with a draft model, speculating 5 tokens at a time. |
| 4 | |
| 5 | ```python |
| 6 | from vllm import LLM, SamplingParams |
| 7 | |
| 8 | prompts = ["The future of AI is"] |
| 9 | sampling_params = SamplingParams(temperature=0.8, top_p=0.95) |
| 10 | |
| 11 | llm = LLM( |
| 12 | model="Qwen/Qwen3-8B", |
| 13 | tensor_parallel_size=1, |
| 14 | speculative_config={ |
| 15 | "model": "Qwen/Qwen3-0.6B", |
| 16 | "num_speculative_tokens": 5, |
| 17 | "method": "draft_model", |
| 18 | }, |
| 19 | ) |
| 20 | outputs = llm.generate(prompts, sampling_params) |
| 21 | |
| 22 | for output in outputs: |
| 23 | prompt = output.prompt |
| 24 | generated_text = output.outputs[0].text |
| 25 | print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") |
| 26 | ``` |
| 27 | |
| 28 | To perform the equivalent launch in online mode, use the following server-side code: |
| 29 | |
| 30 | ```bash |
| 31 | vllm serve Qwen/Qwen3-4B-Thinking-2507 \ |
| 32 | --host 0.0.0.0 \ |
| 33 | --port 8000 \ |
| 34 | --seed 42 \ |
| 35 | -tp 1 \ |
| 36 | --max-model-len 2048 \ |
| 37 | --gpu-memory-utilization 0.8 \ |
| 38 | --speculative-config '{"model": "Qwen/Qwen3-0.6B", "num_speculative_tokens": 5, "method": "draft_model"}' |
| 39 | ``` |
| 40 | |
| 41 | The code used to request completions as a client remains unchanged: |
| 42 | |
| 43 | ??? code |
| 44 | |
| 45 | ```python |
| 46 | from openai import OpenAI |
| 47 | |
| 48 | # Modify OpenAI's API key and API base to use vLLM's API server. |
| 49 | openai_api_key = "EMPTY" |
| 50 | openai_api_base = "http://localhost:8000/v1" |
| 51 | |
| 52 | client = OpenAI( |
| 53 | # defaults to os.environ.get("OPENAI_API_KEY") |
| 54 | api_key=openai_api_key, |
| 55 | base_url=openai_api_base, |
| 56 | ) |
| 57 | |
| 58 | models = client.models.list() |
| 59 | model = models.data[0].id |
| 60 | |
| 61 | # Completion API |
| 62 | stream = False |
| 63 | completion = client.completions.create( |
| 64 | model=model, |
| 65 | prompt="The future of AI is", |
| 66 | echo=False, |
| 67 | n=1, |
| 68 | stream=stream, |
| 69 | ) |
| 70 | |
| 71 | print("Completion results:") |
| 72 | if stream: |
| 73 | for c in completion: |
| 74 | print(c) |
| 75 | else: |
| 76 | print(completion) |
| 77 | ``` |
| 78 | |
| 79 | ## Draft Model Method with heterogeneous vocabs |
| 80 | |
| 81 | By default, vLLM requires the draft and target models to share the same vocabulary. Setting `use_heterogeneous_vocab: true` enables the **Token-Level Intersection (TLI)** algorithm, which allows draft models from a different model family with a different tokenizer. |
| 82 | |
| 83 | Currently, `use_heterogeneous_vocab` requires `draft_sample_method='greedy'` (the default). Probabilistic draft sampling is not yet supported and will be added in a |
| 84 | future release. |
| 85 | |
| 86 | ```python |
| 87 | from vllm import LLM, SamplingParams |
| 88 | |
| 89 | llm = LLM( |
| 90 | model="Qwen/Qwen3-8B", |
| 91 | speculative_config={ |
| 92 | "method": "draft_model", |
| 93 | "model": "HuggingFaceTB/SmolLM2-135M-Instruct", |
| 94 | "num_speculative_tokens": 3, |
| 95 | "use_heterogeneous_vocab": True, |
| 96 | }, |
| 97 | gpu_memory_utilization=0.5, |
| 98 | ) |
| 99 | outputs = llm.generate(prompts, sampling_params) |
| 100 | |
| 101 | for output in outputs: |
| 102 | prompt = output.prompt |
| 103 | generated_text = output.outputs[0].text |
| 104 | print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") |
| 105 | ``` |
| 106 | |
| 107 | !!! warning |
| 108 | Note: Please use `--speculative-config` to set all configurations related |
| 109 | to speculative decoding. The previous method of specifying the model |
| 110 | through `--speculative-model` and adding related parameters such as |
| 111 | `--num-speculative-tokens` separately has been deprecated. For supported |
| 112 | keys and examples, see the [`--speculative-config` schema](README.md#--speculative-config-schema). |