| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """CPU DeepSeek-V4 MTP (speculative decoding) — deferred, not implemented. |
| 4 | |
| 5 | ``DeepSeekV4MTP`` is only instantiated when |
| 6 | ``vllm/model_executor/models/registry.py``'s speculative-decoding registry |
| 7 | resolves it by name for a speculative-decoding config; a plain (non-spec) |
| 8 | DeepSeek-V4 CPU model never constructs this class. This stub exists solely |
| 9 | so ``vllm.models.deepseek_v4.cpu`` is import-clean. |
| 10 | """ |
| 11 | |
| 12 | from vllm.config import VllmConfig |
| 13 | |
| 14 | |
| 15 | class DeepSeekV4MTP: |
| 16 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 17 | raise NotImplementedError( |
| 18 | "DeepSeek-V4 speculative decoding (MTP) is not implemented on CPU." |
| 19 | ) |