| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import torch |
| 4 | |
| 5 | from vllm.config import VllmConfig |
| 6 | |
| 7 | |
| 8 | def init_speculator(vllm_config: VllmConfig, device: torch.device): |
| 9 | speculative_config = vllm_config.speculative_config |
| 10 | assert speculative_config is not None |
| 11 | if speculative_config.method == "extract_hidden_states": |
| 12 | from vllm.v1.worker.gpu.spec_decode.extract_hidden_states import ( |
| 13 | ExtractHiddenStatesSpeculator, |
| 14 | ) |
| 15 | |
| 16 | return ExtractHiddenStatesSpeculator(vllm_config, device) |
| 17 | elif speculative_config.method == "dflash": |
| 18 | if "LiLiCorrDraftModel" in speculative_config.draft_model_config.architectures: |
| 19 | from vllm.v1.worker.gpu.spec_decode.lilicorr.speculator import ( |
| 20 | LiLiCorrSpeculator, |
| 21 | ) |
| 22 | |
| 23 | return LiLiCorrSpeculator(vllm_config, device) |
| 24 | if "DFlash2DraftModel" in speculative_config.draft_model_config.architectures: |
| 25 | from vllm.v1.worker.gpu.spec_decode.dflash2.speculator import ( |
| 26 | DFlash2Speculator, |
| 27 | ) |
| 28 | |
| 29 | return DFlash2Speculator(vllm_config, device) |
| 30 | from vllm.v1.worker.gpu.spec_decode.dflash.speculator import ( |
| 31 | DFlashSpeculator, |
| 32 | ) |
| 33 | |
| 34 | return DFlashSpeculator(vllm_config, device) |
| 35 | elif speculative_config.method == "dspark": |
| 36 | from vllm.v1.worker.gpu.spec_decode.dspark.speculator import ( |
| 37 | DSparkSpeculator, |
| 38 | ) |
| 39 | |
| 40 | return DSparkSpeculator(vllm_config, device) |
| 41 | elif speculative_config.use_gemma4_mtp(): |
| 42 | from vllm.v1.worker.gpu.spec_decode.gemma4.speculator import ( |
| 43 | Gemma4Speculator, |
| 44 | ) |
| 45 | |
| 46 | return Gemma4Speculator(vllm_config, device) |
| 47 | elif speculative_config.use_multi_module_mtp(): |
| 48 | from vllm.v1.worker.gpu.spec_decode.multi_module_mtp.speculator import ( |
| 49 | MultiModuleMTPSpeculator, |
| 50 | ) |
| 51 | |
| 52 | return MultiModuleMTPSpeculator(vllm_config, device) |
| 53 | elif speculative_config.method == "mtp": |
| 54 | from vllm.v1.worker.gpu.spec_decode.mtp.speculator import MTPSpeculator |
| 55 | |
| 56 | return MTPSpeculator(vllm_config, device) |
| 57 | elif speculative_config.use_eagle(): |
| 58 | from vllm.v1.worker.gpu.spec_decode.eagle.speculator import ( |
| 59 | EagleSpeculator, |
| 60 | ) |
| 61 | |
| 62 | return EagleSpeculator(vllm_config, device) |
| 63 | elif speculative_config.uses_draft_model(): |
| 64 | from vllm.v1.worker.gpu.spec_decode.draft_model.speculator import ( |
| 65 | PlainDraftModelSpeculator, |
| 66 | ) |
| 67 | |
| 68 | return PlainDraftModelSpeculator(vllm_config, device) |
| 69 | else: |
| 70 | raise NotImplementedError(f"{speculative_config.method} is not supported yet.") |