| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import os |
| 5 | from dataclasses import dataclass |
| 6 | |
| 7 | import safetensors |
| 8 | import torch |
| 9 | |
| 10 | from vllm.logger import init_logger |
| 11 | from vllm.lora.lora_weights import LoRAFullModuleWeights, LoRALayerWeights |
| 12 | from vllm.lora.peft_helper import PEFTHelper |
| 13 | from vllm.lora.utils import ( |
| 14 | get_lora_id, |
| 15 | is_base_embedding_weights, |
| 16 | parse_fine_tuned_lora_name, |
| 17 | ) |
| 18 | from vllm.model_executor.model_loader.tensorizer import TensorizerConfig |
| 19 | from vllm.model_executor.models.utils import WeightsMapper |
| 20 | from vllm.utils.torch_utils import PIN_MEMORY |
| 21 | |
| 22 | logger = init_logger(__name__) |
| 23 | |
| 24 | |
| 25 | @dataclass(frozen=True) |
| 26 | class MoEEPLoadSpec: |
| 27 | """Per-expert-parallel slicing metadata for one FusedMoEFactory LoRA module. |
| 28 | |
| 29 | Threaded into the LoRA loader so per-expert weights from EP ranks |
| 30 | other than this one can be skipped before they ever hit CPU memory. |
| 31 | """ |
| 32 | |
| 33 | ep_rank: int |
| 34 | local_num_experts: int |
| 35 | global_num_experts: int |
| 36 | |
| 37 | |
| 38 | _EXPERTS_SEPARATOR = ".experts." |
| 39 | |
| 40 | |
| 41 | def _is_remote_expert_key(raw_name: str, spec: "MoEEPLoadSpec") -> bool: |
| 42 | """Decide whether a checkpoint key belongs to a non-local expert.""" |
| 43 | pos = raw_name.find(_EXPERTS_SEPARATOR) |
| 44 | if pos < 0: |
| 45 | return False |
| 46 | idx_start = pos + len(_EXPERTS_SEPARATOR) |
| 47 | idx_end = raw_name.find(".", idx_start) |
| 48 | if idx_end < 0: |
| 49 | return False |
| 50 | idx_str = raw_name[idx_start:idx_end] |
| 51 | if not idx_str.isdigit(): |
| 52 | return False |
| 53 | expert_idx = int(idx_str) |
| 54 | local_start = spec.ep_rank * spec.local_num_experts |
| 55 | return not (local_start <= expert_idx < local_start + spec.local_num_experts) |
| 56 | |
| 57 | |
| 58 | class LoRAModel: |
| 59 | """A LoRA fine-tuned model.""" |
| 60 | |
| 61 | def __init__( |
| 62 | self, |
| 63 | lora_model_id: int, |
| 64 | rank: int, |
| 65 | loras: dict[str, LoRALayerWeights], |
| 66 | is_3d_lora_weight: bool = False, |
| 67 | *, |
| 68 | modules_to_save: dict[str, LoRAFullModuleWeights] | None = None, |
| 69 | ) -> None: |
| 70 | """Args: |
| 71 | lora_model_id: The integer id for the lora model. |
| 72 | rank: lora rank. |
| 73 | loras: module name -> weights for lora-replaced layers. |
| 74 | is_3d_lora_weight: Whether the on-disk MoE adapter is in the 3D |
| 75 | fused (gate_up_proj / down_proj) layout. Propagated from the |
| 76 | originating LoRARequest. Only consulted by the LoRA model |
| 77 | manager when enable_mixed_moe_lora_format is on. |
| 78 | |
| 79 | """ |
| 80 | self.id = lora_model_id |
| 81 | |
| 82 | assert lora_model_id > 0, ( |
| 83 | f"a valid lora id should be greater than 0, got {self.id}" |
| 84 | ) |
| 85 | self.rank = rank |
| 86 | self.loras: dict[str, LoRALayerWeights] = loras |
| 87 | self.modules_to_save = modules_to_save or {} |
| 88 | self.is_3d_lora_weight = is_3d_lora_weight |
| 89 | |
| 90 | def clone(self, lora_model_id: int) -> "LoRAModel": |
| 91 | """Return a copy of the object with different ids. |
| 92 | |
| 93 | Will share the underlying tensors.""" |
| 94 | return self.__class__( |
| 95 | lora_model_id, |
| 96 | rank=self.rank, |
| 97 | loras=self.loras.copy(), |
| 98 | is_3d_lora_weight=self.is_3d_lora_weight, |
| 99 | modules_to_save=self.modules_to_save.copy(), |
| 100 | ) |
| 101 | |
| 102 | def get_lora(self, module_name: str) -> LoRALayerWeights | None: |
| 103 | """Get LoRA for a given module by name.""" |
| 104 | return self.loras.get(module_name, None) |
| 105 | |
| 106 | def get_module_to_save(self, module_name: str) -> LoRAFullModuleWeights | None: |
| 107 | return self.modules_to_save.get(module_name) |
| 108 | |
| 109 | def check_lora_name(self, lora_name: str) -> bool: |
| 110 | return lora_name in self.loras |
| 111 | |
| 112 | @staticmethod |
| 113 | def _should_skip_module(module_name: str, skip_prefixes: list[str]) -> bool: |
| 114 | """Check if a module should be skipped based on skip prefixes.""" |
| 115 | for prefix in skip_prefixes: |
| 116 | if f".{prefix}" in module_name or module_name.startswith(prefix): |
| 117 | return True |
| 118 | return False |
| 119 | |
| 120 | @classmethod |
| 121 | def from_lora_tensors( |
| 122 | cls, |
| 123 | lora_model_id: int, |
| 124 | tensors: dict[str, torch.Tensor], |
| 125 | peft_helper: PEFTHelper, |
| 126 | device: str = "cuda", |
| 127 | dtype: torch.dtype | None = None, |
| 128 | model_vocab_size: int | None = None, |
| 129 | weights_mapper: WeightsMapper | None = None, |
| 130 | skip_prefixes: list[str] | None = None, |
| 131 | ) -> "LoRAModel": |
| 132 | """Create a LoRAModel from a dictionary of tensors.""" |
| 133 | pin_memory = str(device) == "cpu" and PIN_MEMORY |
| 134 | |
| 135 | modules_to_save_names = peft_helper.modules_to_save or [] |
| 136 | full_parameters: dict[str, dict[str, torch.Tensor]] = {} |
| 137 | loras: dict[str, LoRALayerWeights] = {} |
| 138 | for tensor_name, tensor in tensors.items(): |
| 139 | if is_base_embedding_weights(tensor_name): |
| 140 | continue |
| 141 | # Skip modules based on model-defined prefixes (e.g., MTP layers) |
| 142 | if skip_prefixes and cls._should_skip_module(tensor_name, skip_prefixes): |
| 143 | continue |
| 144 | module_name, is_lora_a = parse_fine_tuned_lora_name( |
| 145 | tensor_name, weights_mapper |
| 146 | ) |
| 147 | if module_name in modules_to_save_names: |
| 148 | full_parameters.setdefault(module_name, {})[ |
| 149 | tensor_name.split(".")[-1] |
| 150 | ] = tensor.to(device=device) |
| 151 | continue |
| 152 | |
| 153 | if module_name not in loras: |
| 154 | loras[module_name] = LoRALayerWeights.from_config( |
| 155 | module_name, peft_helper |
| 156 | ) |
| 157 | |
| 158 | if is_lora_a: |
| 159 | if ( |
| 160 | "lora_embedding_A" in tensor_name |
| 161 | and model_vocab_size is not None |
| 162 | and model_vocab_size != tensor.shape[1] |
| 163 | ): |
| 164 | raise RuntimeError( |
| 165 | f"The embedding LoRA size({tensor.shape[1]}) must be consistent" |
| 166 | f" with the base model's vocabulary size({model_vocab_size})." |
| 167 | ) |
| 168 | loras[module_name].lora_a = tensor.to(device=device, dtype=dtype) |
| 169 | if pin_memory: |
| 170 | loras[module_name].lora_a = loras[module_name].lora_a.pin_memory() |
| 171 | else: |
| 172 | loras[module_name].lora_b = tensor.to(device=device, dtype=dtype) |
| 173 | |
| 174 | if pin_memory: |
| 175 | loras[module_name].lora_b = loras[module_name].lora_b.pin_memory() |
| 176 | |
| 177 | modules_to_save = {} |
| 178 | for module_name, parameters in full_parameters.items(): |
| 179 | weight = parameters.get("weight") |
| 180 | if weight is None: |
| 181 | raise ValueError(f"Full module {module_name!r} is missing weight.") |
| 182 | modules_to_save[module_name] = LoRAFullModuleWeights( |
| 183 | module_name=module_name, |
| 184 | weight=weight, |
| 185 | bias=parameters.get("bias"), |
| 186 | ) |
| 187 | if len(modules_to_save) > 1: |
| 188 | raise ValueError( |
| 189 | "Only one full classification module is supported per " |
| 190 | f"adapter, received {sorted(modules_to_save)}." |
| 191 | ) |
| 192 | |
| 193 | return cls(lora_model_id, peft_helper.r, loras, modules_to_save=modules_to_save) |
| 194 | |
| 195 | @classmethod |
| 196 | def from_local_checkpoint( |
| 197 | cls, |
| 198 | lora_dir: str, |
| 199 | expected_lora_modules: set[str], |
| 200 | peft_helper: PEFTHelper, |
| 201 | *, |
| 202 | lora_model_id: int | None = None, |
| 203 | device: str = "cuda", |
| 204 | dtype: torch.dtype | None = None, |
| 205 | model_vocab_size: int | None = None, |
| 206 | weights_mapper: WeightsMapper | None = None, |
| 207 | tensorizer_config_dict: dict | None = None, |
| 208 | skip_prefixes: list[str] | None = None, |
| 209 | moe_ep_spec: MoEEPLoadSpec | None = None, |
| 210 | ) -> "LoRAModel": |
| 211 | """Create a LoRAModel from a local checkpoint. |
| 212 | |
| 213 | Args: |
| 214 | lora_dir: The local path that has lora data. |
| 215 | expected_lora_modules: Name of modules that are expected to be |
| 216 | replaced by lora. |
| 217 | peft_helper: Loaded lora configuration information. |
| 218 | lora_model_id: LoRA model id. If not given, automatically set by |
| 219 | a global counter. |
| 220 | device: Device where the lora model is loaded. |
| 221 | dtype: dtype of the lora model weights. |
| 222 | model_vocab_size: Vocab size of the base model, used to size the |
| 223 | embedding deltas. |
| 224 | weights_mapper: Optional mapper rewriting checkpoint weight names |
| 225 | to vLLM names. |
| 226 | tensorizer_config_dict: Optional tensorizer config used to load |
| 227 | the checkpoint via tensorizer instead of from disk. |
| 228 | skip_prefixes: List of module name prefixes to skip during loading. |
| 229 | Models can define this to skip modules not used in inference |
| 230 | (e.g., MTP layers). Format: ["mtp."] |
| 231 | moe_ep_spec: When 2D FusedMoEFactory LoRA modules are present with |
| 232 | expert parallelism enabled, the (ep_rank, local, global) |
| 233 | slicing metadata shared across all MoE layers. Non-local |
| 234 | expert weights are skipped at read time instead of being |
| 235 | loaded and discarded later. |
| 236 | |
| 237 | Returns: |
| 238 | Loaded LoRA Model. |
| 239 | |
| 240 | """ |
| 241 | lora_tensor_path = os.path.join(lora_dir, "adapter_model.safetensors") |
| 242 | lora_bin_file_path = os.path.join(lora_dir, "adapter_model.bin") |
| 243 | lora_pt_file_path = os.path.join(lora_dir, "adapter_model.pt") |
| 244 | |
| 245 | tensors: dict[str, torch.Tensor] = {} |
| 246 | unexpected_modules: list[list[str] | str] = [] |
| 247 | |
| 248 | def check_unexpected_modules(modules: dict): |
| 249 | for lora_module in modules.keys(): # noqa |
| 250 | if is_base_embedding_weights(lora_module): |
| 251 | continue |
| 252 | # Handle PEFT file format where experts.base_layer is the |
| 253 | # gate_up_proj and experts is the down_proj |
| 254 | if "base_layer" in lora_module: |
| 255 | continue |
| 256 | # Skip modules based on model-defined prefixes |
| 257 | if skip_prefixes and cls._should_skip_module( |
| 258 | lora_module, skip_prefixes |
| 259 | ): |
| 260 | continue |
| 261 | module_name, _ = parse_fine_tuned_lora_name(lora_module, weights_mapper) |
| 262 | base_name = module_name.rsplit(".", 1)[-1] |
| 263 | # Case for expert lora weights |
| 264 | if ".experts" in module_name: |
| 265 | expert_idx = module_name.find(".experts") |
| 266 | expert_suffix = module_name[expert_idx + 1 :] |
| 267 | if expert_suffix not in expected_lora_modules: |
| 268 | unexpected_modules.append(module_name) |
| 269 | |
| 270 | elif base_name not in expected_lora_modules and base_name not in ( |
| 271 | peft_helper.modules_to_save or () |
| 272 | ): |
| 273 | unexpected_modules.append(module_name) |
| 274 | |
| 275 | if unexpected_modules: |
| 276 | raise ValueError( |
| 277 | f"While loading {lora_dir}, expected" |
| 278 | f" target modules in {expected_lora_modules}" |
| 279 | f" but received {unexpected_modules}." |
| 280 | f" Please verify that the loaded LoRA module is correct" |
| 281 | ) |
| 282 | |
| 283 | if tensorizer_config_dict: |
| 284 | from tensorizer import TensorDeserializer |
| 285 | |
| 286 | tensorizer_config = TensorizerConfig(**tensorizer_config_dict) |
| 287 | tensorizer_dir = tensorizer_config.tensorizer_dir |
| 288 | if tensorizer_dir is None: |
| 289 | raise ValueError("tensorizer_dir must be set in tensorizer config.") |
| 290 | lora_tensor_path = os.path.join(tensorizer_dir, "adapter_model.tensors") |
| 291 | tensorizer_args = tensorizer_config._construct_tensorizer_args() |
| 292 | tensors = TensorDeserializer( |
| 293 | lora_tensor_path, |
| 294 | dtype=tensorizer_config.dtype, |
| 295 | device=device, |
| 296 | **tensorizer_args.deserialization_kwargs, |
| 297 | ) |
| 298 | check_unexpected_modules(tensors) |
| 299 | |
| 300 | elif os.path.isfile(lora_tensor_path): |
| 301 | # Find unexpected modules. |
| 302 | # Use safetensor key as a source of truth to find expected modules. |
| 303 | # in peft if you have target_modules A, B, C and C does not exist |
| 304 | # in the model it won’t error and model will be trained with A, B |
| 305 | # loraified. C won’t exist in the safetensor but it will exist in |
| 306 | # the target_modules of the adapter_config.json. |
| 307 | unexpected_modules = [] |
| 308 | with safetensors.safe_open(lora_tensor_path, framework="pt") as f: # type: ignore |
| 309 | # Load tensors if there are only expected modules. |
| 310 | check_unexpected_modules(f) |
| 311 | for module in f.keys(): # noqa |
| 312 | if moe_ep_spec is not None and _is_remote_expert_key( |
| 313 | module, moe_ep_spec |
| 314 | ): |
| 315 | continue |
| 316 | tensors[module] = f.get_tensor(module) |
| 317 | elif os.path.isfile(lora_bin_file_path) or os.path.isfile(lora_pt_file_path): |
| 318 | lora_file_path = ( |
| 319 | lora_bin_file_path |
| 320 | if os.path.isfile(lora_bin_file_path) |
| 321 | else lora_pt_file_path |
| 322 | ) |
| 323 | tensors = torch.load(lora_file_path, map_location=device, weights_only=True) |
| 324 | check_unexpected_modules(tensors) |
| 325 | if moe_ep_spec is not None: |
| 326 | # `.bin`/`.pt` adapters can't be lazy-loaded, but pruning |
| 327 | # the dict here still frees the non-local expert tensors |
| 328 | # before the dtype cast / pin_memory work that follows. |
| 329 | tensors = { |
| 330 | k: v |
| 331 | for k, v in tensors.items() |
| 332 | if not _is_remote_expert_key(k, moe_ep_spec) |
| 333 | } |
| 334 | else: |
| 335 | raise ValueError(f"{lora_dir} doesn't contain tensors") |
| 336 | |
| 337 | return cls.from_lora_tensors( |
| 338 | lora_model_id=get_lora_id() if lora_model_id is None else lora_model_id, |
| 339 | tensors=tensors, |
| 340 | peft_helper=peft_helper, |
| 341 | device=device, |
| 342 | dtype=dtype, |
| 343 | model_vocab_size=model_vocab_size, |
| 344 | weights_mapper=weights_mapper, |
| 345 | skip_prefixes=skip_prefixes, |
| 346 | ) |