| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from typing import Literal |
| 5 | |
| 6 | from torch import nn |
| 7 | |
| 8 | from vllm.config import ModelConfig, VllmConfig |
| 9 | from vllm.config.load import LoadConfig |
| 10 | from vllm.logger import init_logger |
| 11 | from vllm.model_executor.model_loader.base_loader import BaseModelLoader |
| 12 | from vllm.model_executor.model_loader.default_loader import DefaultModelLoader |
| 13 | from vllm.model_executor.model_loader.dummy_loader import DummyModelLoader |
| 14 | from vllm.model_executor.model_loader.modelexpress_loader import ( |
| 15 | ModelExpressModelLoader, |
| 16 | ) |
| 17 | from vllm.model_executor.model_loader.runai_streamer_loader import ( |
| 18 | RunaiModelStreamerLoader, |
| 19 | ) |
| 20 | from vllm.model_executor.model_loader.sharded_state_loader import ShardedStateLoader |
| 21 | from vllm.model_executor.model_loader.tensorizer_loader import TensorizerLoader |
| 22 | from vllm.model_executor.model_loader.utils import ( |
| 23 | get_architecture_class_name, |
| 24 | get_model_architecture, |
| 25 | get_model_cls, |
| 26 | ) |
| 27 | from vllm.model_executor.model_loader.weight_cache.ipc_loader import IpcModelLoader |
| 28 | |
| 29 | logger = init_logger(__name__) |
| 30 | |
| 31 | # Reminder: Please update docstring in `LoadConfig` |
| 32 | # if a new load format is added here |
| 33 | LoadFormats = Literal[ |
| 34 | "auto", |
| 35 | "hf", |
| 36 | "dummy", |
| 37 | "fastsafetensors", |
| 38 | "instanttensor", |
| 39 | "ipc_cache", |
| 40 | "mistral", |
| 41 | "modelexpress", |
| 42 | "npcache", |
| 43 | "pt", |
| 44 | "runai_streamer", |
| 45 | "runai_streamer_sharded", |
| 46 | "safetensors", |
| 47 | "sharded_state", |
| 48 | "tensorizer", |
| 49 | ] |
| 50 | _LOAD_FORMAT_TO_MODEL_LOADER: dict[str, type[BaseModelLoader]] = { |
| 51 | "auto": DefaultModelLoader, |
| 52 | "hf": DefaultModelLoader, |
| 53 | "dummy": DummyModelLoader, |
| 54 | "fastsafetensors": DefaultModelLoader, |
| 55 | "instanttensor": DefaultModelLoader, |
| 56 | "ipc_cache": IpcModelLoader, |
| 57 | "mistral": DefaultModelLoader, |
| 58 | "modelexpress": ModelExpressModelLoader, |
| 59 | "npcache": DefaultModelLoader, |
| 60 | "pt": DefaultModelLoader, |
| 61 | "runai_streamer": RunaiModelStreamerLoader, |
| 62 | "runai_streamer_sharded": ShardedStateLoader, |
| 63 | "safetensors": DefaultModelLoader, |
| 64 | "sharded_state": ShardedStateLoader, |
| 65 | "tensorizer": TensorizerLoader, |
| 66 | } |
| 67 | |
| 68 | |
| 69 | def register_model_loader(load_format: str): |
| 70 | """Register a customized vllm model loader. |
| 71 | |
| 72 | When a load format is not supported by vllm, you can register a customized |
| 73 | model loader to support it. |
| 74 | |
| 75 | Args: |
| 76 | load_format (str): The model loader format name. |
| 77 | |
| 78 | Examples: |
| 79 | >>> from vllm.config.load import LoadConfig |
| 80 | >>> from vllm.model_executor.model_loader import ( |
| 81 | ... get_model_loader, |
| 82 | ... register_model_loader, |
| 83 | ... ) |
| 84 | >>> from vllm.model_executor.model_loader.base_loader import BaseModelLoader |
| 85 | >>> |
| 86 | >>> @register_model_loader("my_loader") |
| 87 | ... class MyModelLoader(BaseModelLoader): |
| 88 | ... def download_model(self): |
| 89 | ... pass |
| 90 | ... |
| 91 | ... def load_weights(self): |
| 92 | ... pass |
| 93 | >>> |
| 94 | >>> load_config = LoadConfig(load_format="my_loader") |
| 95 | >>> type(get_model_loader(load_config)) |
| 96 | <class 'MyModelLoader'> |
| 97 | |
| 98 | """ # noqa: E501 |
| 99 | |
| 100 | def _wrapper(model_loader_cls): |
| 101 | if load_format in _LOAD_FORMAT_TO_MODEL_LOADER: |
| 102 | logger.warning( |
| 103 | "Load format `%s` is already registered, and will be " |
| 104 | "overwritten by the new loader class `%s`.", |
| 105 | load_format, |
| 106 | model_loader_cls, |
| 107 | ) |
| 108 | if not issubclass(model_loader_cls, BaseModelLoader): |
| 109 | raise ValueError( |
| 110 | "The model loader must be a subclass of `BaseModelLoader`." |
| 111 | ) |
| 112 | _LOAD_FORMAT_TO_MODEL_LOADER[load_format] = model_loader_cls |
| 113 | logger.info( |
| 114 | "Registered model loader `%s` with load format `%s`", |
| 115 | model_loader_cls, |
| 116 | load_format, |
| 117 | ) |
| 118 | return model_loader_cls |
| 119 | |
| 120 | return _wrapper |
| 121 | |
| 122 | |
| 123 | def get_model_loader(load_config: LoadConfig) -> BaseModelLoader: |
| 124 | """Get a model loader based on the load format.""" |
| 125 | load_format = load_config.load_format |
| 126 | if load_format not in _LOAD_FORMAT_TO_MODEL_LOADER: |
| 127 | raise ValueError(f"Load format `{load_format}` is not supported") |
| 128 | return _LOAD_FORMAT_TO_MODEL_LOADER[load_format](load_config) |
| 129 | |
| 130 | |
| 131 | def get_model( |
| 132 | *, |
| 133 | vllm_config: VllmConfig, |
| 134 | model_config: ModelConfig | None = None, |
| 135 | prefix: str = "", |
| 136 | load_config: LoadConfig | None = None, |
| 137 | ) -> nn.Module: |
| 138 | loader = get_model_loader(load_config or vllm_config.load_config) |
| 139 | if model_config is None: |
| 140 | model_config = vllm_config.model_config |
| 141 | return loader.load_model( |
| 142 | vllm_config=vllm_config, model_config=model_config, prefix=prefix |
| 143 | ) |
| 144 | |
| 145 | |
| 146 | __all__ = [ |
| 147 | "get_model", |
| 148 | "get_model_loader", |
| 149 | "get_architecture_class_name", |
| 150 | "get_model_architecture", |
| 151 | "get_model_cls", |
| 152 | "register_model_loader", |
| 153 | "BaseModelLoader", |
| 154 | "ModelExpressModelLoader", |
| 155 | "DefaultModelLoader", |
| 156 | "DummyModelLoader", |
| 157 | "IpcModelLoader", |
| 158 | "RunaiModelStreamerLoader", |
| 159 | "ShardedStateLoader", |
| 160 | "TensorizerLoader", |
| 161 | ] |