| 1 | # MTP (Multi-Token Prediction) |
| 2 | |
| 3 | MTP is a speculative decoding method where the target model includes native |
| 4 | multi-token prediction capability. Unlike draft-model-based methods, you do not |
| 5 | need to provide a separate draft model. |
| 6 | |
| 7 | MTP is useful when: |
| 8 | |
| 9 | - Your model natively supports MTP. |
| 10 | - You want model-based speculative decoding with minimal extra configuration. |
| 11 | |
| 12 | ## Gemma 4 Assistant Models |
| 13 | |
| 14 | Gemma 4 assistant checkpoints use vLLM's Gemma 4 MTP path. They are not generic |
| 15 | draft models, even though they are passed through the `model` field in |
| 16 | `--speculative-config`. |
| 17 | |
| 18 | Use `"method": "mtp"` when serving Gemma 4 with an assistant checkpoint: |
| 19 | |
| 20 | ```bash |
| 21 | vllm serve google/gemma-4-E2B-it \ |
| 22 | --tensor-parallel-size 1 \ |
| 23 | --max-model-len 8192 \ |
| 24 | --speculative-config '{"method":"mtp","model":"gg-hf-am/gemma-4-E2B-it-assistant","num_speculative_tokens":1}' |
| 25 | ``` |
| 26 | |
| 27 | The E2B, E4B, 12B, 26B-A4B, and 31B Gemma 4 IT assistant checkpoints are supported. |
| 28 | Tower-based variants use `model_type: gemma4_assistant` and the encoder-free |
| 29 | Gemma 4 Unified variant (12B) uses `model_type: gemma4_unified_assistant`. |
| 30 | vLLM maps both to `Gemma4MTPModel` internally and wires the assistant layers |
| 31 | to share KV cache with the target model. |
| 32 | |
| 33 | If an older vLLM release logs `SpeculativeConfig(method='draft_model', ...)` |
| 34 | for a Gemma 4 assistant checkpoint, that release is treating the assistant as a |
| 35 | generic draft model and may fail during initialization for multimodal Gemma 4 |
| 36 | targets. Upgrade to a version with Gemma 4 MTP support instead. |
| 37 | |
| 38 | ## Offline Example |
| 39 | |
| 40 | ```python |
| 41 | from vllm import LLM, SamplingParams |
| 42 | |
| 43 | prompts = ["The future of AI is"] |
| 44 | sampling_params = SamplingParams(temperature=0.8, top_p=0.95) |
| 45 | |
| 46 | llm = LLM( |
| 47 | model="XiaomiMiMo/MiMo-7B-Base", |
| 48 | tensor_parallel_size=1, |
| 49 | speculative_config={ |
| 50 | "method": "mtp", |
| 51 | "num_speculative_tokens": 1, |
| 52 | }, |
| 53 | ) |
| 54 | outputs = llm.generate(prompts, sampling_params) |
| 55 | |
| 56 | for output in outputs: |
| 57 | prompt = output.prompt |
| 58 | generated_text = output.outputs[0].text |
| 59 | print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") |
| 60 | ``` |
| 61 | |
| 62 | ## Online Example |
| 63 | |
| 64 | ```bash |
| 65 | vllm serve XiaomiMiMo/MiMo-7B-Base \ |
| 66 | --tensor-parallel-size 1 \ |
| 67 | --speculative-config '{"method":"mtp","num_speculative_tokens":1}' |
| 68 | ``` |
| 69 | |
| 70 | ## Notes |
| 71 | |
| 72 | - MTP only works for model families that support MTP in vLLM. |
| 73 | - `num_speculative_tokens` controls speculative depth. A small value like `1` |
| 74 | is a good default to start with. |
| 75 | - If your model does not support MTP, use another method such as EAGLE or draft |
| 76 | model speculation. |