File size: 11,796 Bytes
687d3ce 9522bfe 687d3ce 9522bfe 687d3ce 9522bfe 687d3ce 9522bfe 687d3ce 9522bfe 687d3ce 9522bfe 687d3ce 9522bfe 687d3ce 9522bfe 687d3ce 9522bfe 687d3ce 9522bfe 687d3ce 9522bfe |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 |
from diffusers.models.attention_processor import FluxAttnProcessor2_0
from safetensors import safe_open
import re
import torch
from .layers_cache import MultiDoubleStreamBlockLoraProcessor, MultiSingleStreamBlockLoraProcessor
# 移除全局 device = "cuda"，改为通过参数传递
def load_safetensors(path):
tensors = {}
with safe_open(path, framework="pt", device="cpu") as f:
for key in f.keys():
tensors[key] = f.get_tensor(key)
return tensors
def get_lora_rank(checkpoint):
for k in checkpoint.keys():
if k.endswith(".down.weight"):
return checkpoint[k].shape[0]
def load_checkpoint(local_path):
if local_path is not None:
if '.safetensors' in local_path:
print(f"Loading .safetensors checkpoint from {local_path}")
checkpoint = load_safetensors(local_path)
else:
print(f"Loading checkpoint from {local_path}")
checkpoint = torch.load(local_path, map_location='cpu')
return checkpoint
def update_model_with_lora(checkpoint, lora_weights, transformer, cond_size, device="cpu"):
number = len(lora_weights)
ranks = [get_lora_rank(checkpoint) for _ in range(number)]
lora_attn_procs = {}
double_blocks_idx = list(range(19))
single_blocks_idx = list(range(38))
for name, attn_processor in transformer.attn_processors.items():
match = re.search(r'\.(\d+)\.', name)
if match:
layer_index = int(match.group(1))
if name.startswith("transformer_blocks") and layer_index in double_blocks_idx:
lora_state_dicts = {}
for key, value in checkpoint.items():
if re.search(r'\.(\d+)\.', key):
checkpoint_layer_index = int(re.search(r'\.(\d+)\.', key).group(1))
if checkpoint_layer_index == layer_index and key.startswith("transformer_blocks"):
lora_state_dicts[key] = value
lora_attn_procs[name] = MultiDoubleStreamBlockLoraProcessor(
dim=3072,
ranks=ranks,
network_alphas=ranks,
lora_weights=lora_weights,
device=device, # 使用传入的 device 参数
dtype=torch.bfloat16,
cond_width=cond_size,
cond_height=cond_size,
n_loras=number
)
# Load weights and move to specified device
for n in range(number):
lora_attn_procs[name].q_loras[n].down.weight.data = lora_state_dicts.get(f'{name}.q_loras.{n}.down.weight', None)
lora_attn_procs[name].q_loras[n].up.weight.data = lora_state_dicts.get(f'{name}.q_loras.{n}.up.weight', None)
lora_attn_procs[name].k_loras[n].down.weight.data = lora_state_dicts.get(f'{name}.k_loras.{n}.down.weight', None)
lora_attn_procs[name].k_loras[n].up.weight.data = lora_state_dicts.get(f'{name}.k_loras.{n}.up.weight', None)
lora_attn_procs[name].v_loras[n].down.weight.data = lora_state_dicts.get(f'{name}.v_loras.{n}.down.weight', None)
lora_attn_procs[name].v_loras[n].up.weight.data = lora_state_dicts.get(f'{name}.v_loras.{n}.up.weight', None)
lora_attn_procs[name].proj_loras[n].down.weight.data = lora_state_dicts.get(f'{name}.proj_loras.{n}.down.weight', None)
lora_attn_procs[name].proj_loras[n].up.weight.data = lora_state_dicts.get(f'{name}.proj_loras.{n}.up.weight', None)
lora_attn_procs[name].to(device) # 使用传入的 device
elif name.startswith("single_transformer_blocks") and layer_index in single_blocks_idx:
lora_state_dicts = {}
for key, value in checkpoint.items():
if re.search(r'\.(\d+)\.', key):
checkpoint_layer_index = int(re.search(r'\.(\d+)\.', key).group(1))
if checkpoint_layer_index == layer_index and key.startswith("single_transformer_blocks"):
lora_state_dicts[key] = value
lora_attn_procs[name] = MultiSingleStreamBlockLoraProcessor(
dim=3072,
ranks=ranks,
network_alphas=ranks,
lora_weights=lora_weights,
device=device, # 使用传入的 device 参数
dtype=torch.bfloat16,
cond_width=cond_size,
cond_height=cond_size,
n_loras=number
)
# Load weights and move to specified device
for n in range(number):
lora_attn_procs[name].q_loras[n].down.weight.data = lora_state_dicts.get(f'{name}.q_loras.{n}.down.weight', None)
lora_attn_procs[name].q_loras[n].up.weight.data = lora_state_dicts.get(f'{name}.q_loras.{n}.up.weight', None)
lora_attn_procs[name].k_loras[n].down.weight.data = lora_state_dicts.get(f'{name}.k_loras.{n}.down.weight', None)
lora_attn_procs[name].k_loras[n].up.weight.data = lora_state_dicts.get(f'{name}.k_loras.{n}.up.weight', None)
lora_attn_procs[name].v_loras[n].down.weight.data = lora_state_dicts.get(f'{name}.v_loras.{n}.down.weight', None)
lora_attn_procs[name].v_loras[n].up.weight.data = lora_state_dicts.get(f'{name}.v_loras.{n}.up.weight', None)
lora_attn_procs[name].to(device) # 使用传入的 device
else:
lora_attn_procs[name] = FluxAttnProcessor2_0()
transformer.set_attn_processor(lora_attn_procs)
def update_model_with_multi_lora(checkpoints, lora_weights, transformer, cond_size, device="cpu"): # 顺便更新此函数
ck_number = len(checkpoints)
cond_lora_number = [len(ls) for ls in lora_weights]
cond_number = sum(cond_lora_number)
ranks = [get_lora_rank(checkpoint) for checkpoint in checkpoints]
multi_lora_weight = []
for ls in lora_weights:
for n in ls:
multi_lora_weight.append(n)
lora_attn_procs = {}
double_blocks_idx = list(range(19))
single_blocks_idx = list(range(38))
for name, attn_processor in transformer.attn_processors.items():
match = re.search(r'\.(\d+)\.', name)
if match:
layer_index = int(match.group(1))
if name.startswith("transformer_blocks") and layer_index in double_blocks_idx:
lora_state_dicts = [{} for _ in range(ck_number)]
for idx, checkpoint in enumerate(checkpoints):
for key, value in checkpoint.items():
if re.search(r'\.(\d+)\.', key):
checkpoint_layer_index = int(re.search(r'\.(\d+)\.', key).group(1))
if checkpoint_layer_index == layer_index and key.startswith("transformer_blocks"):
lora_state_dicts[idx][key] = value
lora_attn_procs[name] = MultiDoubleStreamBlockLoraProcessor(
dim=3072,
ranks=ranks,
network_alphas=ranks,
lora_weights=multi_lora_weight,
device=device, # 使用传入的 device 参数
dtype=torch.bfloat16,
cond_width=cond_size,
cond_height=cond_size,
n_loras=cond_number
)
num = 0
for idx in range(ck_number):
for n in range(cond_lora_number[idx]):
lora_attn_procs[name].q_loras[num].down.weight.data = lora_state_dicts[idx].get(f'{name}.q_loras.{n}.down.weight', None)
lora_attn_procs[name].q_loras[num].up.weight.data = lora_state_dicts[idx].get(f'{name}.q_loras.{n}.up.weight', None)
lora_attn_procs[name].k_loras[num].down.weight.data = lora_state_dicts[idx].get(f'{name}.k_loras.{n}.down.weight', None)
lora_attn_procs[name].k_loras[num].up.weight.data = lora_state_dicts[idx].get(f'{name}.k_loras.{n}.up.weight', None)
lora_attn_procs[name].v_loras[num].down.weight.data = lora_state_dicts[idx].get(f'{name}.v_loras.{n}.down.weight', None)
lora_attn_procs[name].v_loras[num].up.weight.data = lora_state_dicts[idx].get(f'{name}.v_loras.{n}.up.weight', None)
lora_attn_procs[name].proj_loras[num].down.weight.data = lora_state_dicts[idx].get(f'{name}.proj_loras.{n}.down.weight', None)
lora_attn_procs[name].proj_loras[num].up.weight.data = lora_state_dicts[idx].get(f'{name}.proj_loras.{n}.up.weight', None)
lora_attn_procs[name].to(device) # 使用传入的 device
num += 1
elif name.startswith("single_transformer_blocks") and layer_index in single_blocks_idx:
lora_state_dicts = [{} for _ in range(ck_number)]
for idx, checkpoint in enumerate(checkpoints):
for key, value in checkpoint.items():
if re.search(r'\.(\d+)\.', key):
checkpoint_layer_index = int(re.search(r'\.(\d+)\.', key).group(1))
if checkpoint_layer_index == layer_index and key.startswith("single_transformer_blocks"):
lora_state_dicts[idx][key] = value
lora_attn_procs[name] = MultiSingleStreamBlockLoraProcessor(
dim=3072,
ranks=ranks,
network_alphas=ranks,
lora_weights=multi_lora_weight,
device=device, # 使用传入的 device 参数
dtype=torch.bfloat16,
cond_width=cond_size,
cond_height=cond_size,
n_loras=cond_number
)
num = 0
for idx in range(ck_number):
for n in range(cond_lora_number[idx]):
lora_attn_procs[name].q_loras[num].down.weight.data = lora_state_dicts[idx].get(f'{name}.q_loras.{n}.down.weight', None)
lora_attn_procs[name].q_loras[num].up.weight.data = lora_state_dicts[idx].get(f'{name}.q_loras.{n}.up.weight', None)
lora_attn_procs[name].k_loras[num].down.weight.data = lora_state_dicts[idx].get(f'{name}.k_loras.{n}.down.weight', None)
lora_attn_procs[name].k_loras[num].up.weight.data = lora_state_dicts[idx].get(f'{name}.k_loras.{n}.up.weight', None)
lora_attn_procs[name].v_loras[num].down.weight.data = lora_state_dicts[idx].get(f'{name}.v_loras.{n}.down.weight', None)
lora_attn_procs[name].v_loras[num].up.weight.data = lora_state_dicts[idx].get(f'{name}.v_loras.{n}.up.weight', None)
lora_attn_procs[name].to(device) # 使用传入的 device
num += 1
else:
lora_attn_procs[name] = FluxAttnProcessor2_0()
transformer.set_attn_processor(lora_attn_procs)
def set_single_lora(transformer, local_path, lora_weights=[], cond_size=512, device="cpu"):
checkpoint = load_checkpoint(local_path)
update_model_with_lora(checkpoint, lora_weights, transformer, cond_size, device=device)
def set_multi_lora(transformer, local_paths, lora_weights=[[]], cond_size=512, device="cpu"): # 顺便更新此函数
checkpoints = [load_checkpoint(local_path) for local_path in local_paths]
update_model_with_multi_lora(checkpoints, lora_weights, transformer, cond_size, device=device)
def unset_lora(transformer):
lora_attn_procs = {}
for name, attn_processor in transformer.attn_processors.items():
lora_attn_procs[name] = FluxAttnProcessor2_0()
transformer.set_attn_processor(lora_attn_procs) |