Spaces:
Sleeping
Sleeping
File size: 4,849 Bytes
ca1ecab |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 |
from typing import List, Optional
import torch
from vllm.utils import in_wsl
class LoRALayerWeights:
"""LoRA weights for a layer composed of two low rank matrixes."""
def __init__(
self,
module_name: str,
rank: int,
lora_alpha: int,
lora_a: torch.Tensor,
lora_b: torch.Tensor,
embeddings_tensor: Optional[torch.Tensor] = None,
scaling: Optional[float] = None,
) -> None:
self.module_name = module_name
self.rank = rank
self.lora_alpha = lora_alpha
self.lora_a = lora_a
self.lora_b = lora_b
self.embeddings_tensor = embeddings_tensor
if scaling is None:
self.scaling = self.lora_alpha / self.rank
else:
self.scaling = scaling
def optimize(self) -> "LoRALayerWeights":
"""Optimize the LoRA by merging the scaling into lora_b."""
if self.scaling == 1:
return
self.lora_b *= self.scaling
self.scaling = 1
return self
@property
def input_dim(self) -> int:
return self.lora_a.shape[0]
@property
def output_dim(self) -> int:
return self.lora_b.shape[1]
@property
def is_packed(self) -> bool:
return False
@property
def extra_vocab_size(self) -> int:
return self.embeddings_tensor.shape[
0] if self.embeddings_tensor is not None else 0
@classmethod
def create_dummy_lora_weights(
cls,
module_name: str,
input_dim: int,
output_dim: int,
rank: int,
dtype: torch.dtype,
device: torch.device,
embeddings_tensor_dim: Optional[int] = None) -> "LoRALayerWeights":
pin_memory = str(device) == "cpu" and not in_wsl()
lora_a = torch.zeros([input_dim, rank],
dtype=dtype,
device=device,
pin_memory=pin_memory)
lora_b = torch.zeros([rank, output_dim],
dtype=dtype,
device=device,
pin_memory=pin_memory)
embeddings_tensor = torch.rand(
10,
embeddings_tensor_dim,
dtype=dtype,
device=device,
pin_memory=pin_memory) if embeddings_tensor_dim else None
return cls(
module_name,
rank=rank,
lora_alpha=1,
lora_a=lora_a,
lora_b=lora_b,
embeddings_tensor=embeddings_tensor,
)
class PackedLoRALayerWeights(LoRALayerWeights):
"""LoRA used for packed layers (eg. qkv_proj)."""
def __init__(
self,
module_name: str,
rank: int,
lora_alphas: List[int],
lora_a: List[torch.Tensor],
lora_b: List[torch.Tensor],
scaling: Optional[List[float]] = None,
) -> None:
super().__init__(
module_name=module_name,
rank=rank,
lora_alpha=0,
lora_a=lora_a,
lora_b=lora_b,
scaling=scaling,
embeddings_tensor=None,
)
self.lora_alphas = lora_alphas
if scaling is None:
self.scaling = [
lora_alpha / self.rank for lora_alpha in self.lora_alphas
]
@classmethod
def pack(cls, loras: List["LoRALayerWeights"]) -> "PackedLoRALayerWeights":
"""Pack a list of LoRAs into a single LoRA.
If LoRA is None, it signifies that the submodule does not have a LoRA.
"""
first_lora = next(lora for lora in loras if lora is not None)
for lora in loras:
if lora is None:
continue
lora.optimize()
rank = first_lora.rank
module_name = first_lora.module_name
obj = cls(
module_name,
rank,
[lora.lora_alpha if lora is not None else None for lora in loras],
[lora.lora_a if lora is not None else None for lora in loras],
[lora.lora_b if lora is not None else None for lora in loras],
scaling=[1 if lora is not None else None for lora in loras])
return obj
def optimize(self) -> "PackedLoRALayerWeights":
"""Optimize the LoRA by merging the scaling into lora_b."""
for i in range(len(self.lora_b)):
if self.scaling[i] == 1 or self.lora_b[i] is None:
continue
self.lora_b[i] *= self.scaling[i]
self.scaling[i] = 1
return self
@property
def input_dim(self) -> int:
raise NotImplementedError()
@property
def output_dim(self) -> int:
raise NotImplementedError()
@property
def is_packed(self) -> bool:
return True
|