| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from collections.abc import Iterable |
| 5 | from typing import TYPE_CHECKING, overload |
| 6 | |
| 7 | import torch |
| 8 | import torch.nn as nn |
| 9 | from transformers import PreTrainedConfig |
| 10 | |
| 11 | from vllm.config.lora import LoRAConfig |
| 12 | |
| 13 | if TYPE_CHECKING: |
| 14 | from vllm.lora.punica_wrapper import PunicaWrapperBase |
| 15 | |
| 16 | |
| 17 | class BaseLayerWithLoRA(nn.Module): |
| 18 | def __getattr__(self, name): |
| 19 | d = self.__dict__ |
| 20 | if name in d.get("_parameters", ()): |
| 21 | return d["_parameters"][name] |
| 22 | if name in d.get("_buffers", ()): |
| 23 | return d["_buffers"][name] |
| 24 | if name in d.get("_modules", ()): |
| 25 | return d["_modules"][name] |
| 26 | # Forward public misses to ``base_layer``; private names are framework |
| 27 | # bookkeeping and must stay local. |
| 28 | if not name.startswith("_"): |
| 29 | base_layer = d.get("_modules", {}).get("base_layer") |
| 30 | if base_layer is not None: |
| 31 | return getattr(base_layer, name) |
| 32 | raise AttributeError( |
| 33 | f"{type(self).__name__!r} object has no attribute {name!r}" |
| 34 | ) |
| 35 | |
| 36 | def load_weights( |
| 37 | self, weights: Iterable[tuple[str, torch.Tensor]] |
| 38 | ) -> Iterable[str]: |
| 39 | """Load checkpoint weights into the wrapped base layer.""" |
| 40 | base_load_weights = getattr(self.base_layer, "load_weights", None) |
| 41 | if callable(base_load_weights): |
| 42 | return base_load_weights(weights) |
| 43 | |
| 44 | from vllm.model_executor.models.utils import AutoWeightsLoader |
| 45 | |
| 46 | return AutoWeightsLoader(self.base_layer).load_weights(weights) |
| 47 | |
| 48 | @overload |
| 49 | def slice_lora_a( |
| 50 | self, lora_a: list[torch.Tensor | None] |
| 51 | ) -> list[torch.Tensor | None]: ... |
| 52 | @overload |
| 53 | def slice_lora_a(self, lora_a: torch.Tensor) -> torch.Tensor: ... |
| 54 | def slice_lora_a( |
| 55 | self, lora_a: torch.Tensor | list[torch.Tensor | None] |
| 56 | ) -> torch.Tensor | list[torch.Tensor | None]: |
| 57 | """Slice lora a if splitting for tensor parallelism.""" |
| 58 | ... |
| 59 | |
| 60 | @overload |
| 61 | def slice_lora_b( |
| 62 | self, lora_b: list[torch.Tensor | None] |
| 63 | ) -> list[torch.Tensor | None]: ... |
| 64 | @overload |
| 65 | def slice_lora_b(self, lora_b: torch.Tensor) -> torch.Tensor: ... |
| 66 | def slice_lora_b( |
| 67 | self, lora_b: torch.Tensor | list[torch.Tensor | None] |
| 68 | ) -> torch.Tensor | list[torch.Tensor | None]: |
| 69 | """Slice lora b if splitting with tensor parallelism.""" |
| 70 | ... |
| 71 | |
| 72 | def create_lora_weights( |
| 73 | self, |
| 74 | max_loras: int, |
| 75 | lora_config: LoRAConfig, |
| 76 | model_config: PreTrainedConfig | None = None, |
| 77 | ) -> None: |
| 78 | """Initializes lora matrices.""" |
| 79 | ... |
| 80 | |
| 81 | def reset_lora(self, index: int): |
| 82 | """Resets the lora weights at index back to 0.""" |
| 83 | ... |
| 84 | |
| 85 | def set_lora( |
| 86 | self, |
| 87 | index: int, |
| 88 | lora_a: torch.Tensor | list[torch.Tensor], |
| 89 | lora_b: torch.Tensor | list[torch.Tensor], |
| 90 | ): |
| 91 | """Overwrites lora tensors at index.""" |
| 92 | ... |
| 93 | |
| 94 | def set_mapping( |
| 95 | self, |
| 96 | punica_wrapper, |
| 97 | ): |
| 98 | self.punica_wrapper: PunicaWrapperBase = punica_wrapper |
| 99 | |
| 100 | @classmethod |
| 101 | def can_replace_layer( |
| 102 | cls, |
| 103 | source_layer: nn.Module, |
| 104 | lora_config: LoRAConfig, |
| 105 | packed_modules_list: list, |
| 106 | model_config: PreTrainedConfig | None = None, |
| 107 | ) -> bool: |
| 108 | """Returns True if the layer can be replaced by this LoRA layer.""" |
| 109 | raise NotImplementedError |