| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """MiniMax M3 model — hardware-isolated entry point. |
| 4 | |
| 5 | The implementation lives under ``nvidia/`` and ``amd/``; this module picks the |
| 6 | right one for the current platform and re-exports the public classes used by |
| 7 | the model registry. (Mirrors ``vllm.models.deepseek_v4``.) |
| 8 | """ |
| 9 | |
| 10 | from typing import TYPE_CHECKING |
| 11 | |
| 12 | from vllm.platforms import current_platform |
| 13 | |
| 14 | # The NVIDIA branch is the static default that type-checkers see; the ROCm |
| 15 | # branch overrides it at runtime (kept type-compatible via type: ignore). |
| 16 | if TYPE_CHECKING or not current_platform.is_rocm(): |
| 17 | from .nvidia.model import ( |
| 18 | MiniMaxM3SparseForCausalLM, |
| 19 | MiniMaxM3SparseForConditionalGeneration, |
| 20 | ) |
| 21 | from .nvidia.mtp import MiniMaxM3MTP |
| 22 | else: |
| 23 | from .amd.model import ( # type: ignore[assignment] |
| 24 | MiniMaxM3SparseForCausalLM, |
| 25 | MiniMaxM3SparseForConditionalGeneration, |
| 26 | ) |
| 27 | from .amd.mtp import MiniMaxM3MTP # type: ignore[assignment] |
| 28 | |
| 29 | __all__ = [ |
| 30 | "MiniMaxM3MTP", |
| 31 | "MiniMaxM3SparseForCausalLM", |
| 32 | "MiniMaxM3SparseForConditionalGeneration", |
| 33 | ] |