Skip to content

Commit d610c2d

Browse files
committed
balck code formatting
1 parent 014f5bc commit d610c2d

6 files changed

Lines changed: 1117 additions & 601 deletions

File tree

sae_lens/toolkit/pretrained_sae_loaders.py

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -22,7 +22,7 @@ def __call__(
2222
) -> tuple[dict[str, Any], dict[str, torch.Tensor], Optional[torch.Tensor]]: ...
2323

2424

25-
'''
25+
"""
2626
def sae_lens_loader_oldself(
2727
repo_id: str,
2828
folder_name: str,
@@ -75,7 +75,7 @@ def sae_lens_loader_oldself(
7575
log_sparsity = None
7676
7777
return cfg_dict, state_dict, log_sparsity
78-
'''
78+
"""
7979

8080

8181
def sae_lens_loader(

tutorials/convert_llamaSae_to_Saelens.py

Lines changed: 32 additions & 27 deletions
Original file line numberDiff line numberDiff line change
@@ -4,28 +4,30 @@
44
from tqdm import tqdm
55
from sae_lens.config import LOCAL_SAE_MODEL_PATH
66

7-
#base_dir = '/data02/tuwenming/SAE/Pretrained_SAEs/sae-llama-3-8b-32x-v2/layers.{}/'
8-
base_dir = LOCAL_SAE_MODEL_PATH + '/EleutherAI/sae-llama-3-8b-32x/layers.{}/'
7+
# base_dir = '/data02/tuwenming/SAE/Pretrained_SAEs/sae-llama-3-8b-32x-v2/layers.{}/'
8+
base_dir = LOCAL_SAE_MODEL_PATH + "/EleutherAI/sae-llama-3-8b-32x/layers.{}/"
9+
910

1011
def delete_original_sae(layer_index):
1112
# 构建文件路径
1213
layer_dir = base_dir.format(layer_index)
13-
sae_file_path = os.path.join(layer_dir, 'sae.safetensors')
14-
14+
sae_file_path = os.path.join(layer_dir, "sae.safetensors")
15+
1516
# 检查文件是否存在
1617
if os.path.exists(sae_file_path):
1718
os.remove(sae_file_path)
1819
print(f"Deleted {sae_file_path}")
1920
else:
2021
print(f"File {sae_file_path} does not exist, skipping.")
2122

23+
2224
def process_layer(layer_index):
2325
# 构建文件路径
2426
layer_dir = base_dir.format(layer_index)
25-
sae_file_path = os.path.join(layer_dir, 'sae.safetensors')
26-
cfg_file_path = os.path.join(layer_dir, 'cfg.json')
27-
new_sae_file_path = os.path.join(layer_dir, 'sae_weights.safetensors')
28-
27+
sae_file_path = os.path.join(layer_dir, "sae.safetensors")
28+
cfg_file_path = os.path.join(layer_dir, "cfg.json")
29+
new_sae_file_path = os.path.join(layer_dir, "sae_weights.safetensors")
30+
2931
# 处理 sae.safetensors 文件
3032
state_dict = load_file(sae_file_path)
3133
new_state_dict = {}
@@ -38,33 +40,36 @@ def process_layer(layer_index):
3840
# 对权重进行转置
3941
new_state_dict[new_key] = state_dict[key].transpose(0, 1).contiguous()
4042
continue # 跳过下面的赋值语句,避免重复添加
41-
43+
4244
new_state_dict[new_key] = state_dict[key]
43-
45+
4446
# 保存修改后的state_dict到新的safetensors文件
4547
save_file(new_state_dict, new_sae_file_path)
46-
48+
4749
# 处理 cfg.json 文件
48-
with open(cfg_file_path, 'r') as cfg_file:
50+
with open(cfg_file_path, "r") as cfg_file:
4951
cfg_data = json.load(cfg_file)
50-
52+
5153
# 添加新的键值对
52-
cfg_data.update({
53-
"d_sae": 131072,
54-
"dtype": "float32",
55-
"dataset_path": "togethercomputer/RedPajama-Data-1T-Sample",
56-
"context_size": 256,
57-
"model_name": "meta-llama/Meta-Llama-3-8B",
58-
"hook_name": f"blocks.{layer_index}.hook_resid_post",
59-
"hook_layer": str(layer_index),
60-
"hook_head_index": None
61-
})
62-
54+
cfg_data.update(
55+
{
56+
"d_sae": 131072,
57+
"dtype": "float32",
58+
"dataset_path": "togethercomputer/RedPajama-Data-1T-Sample",
59+
"context_size": 256,
60+
"model_name": "meta-llama/Meta-Llama-3-8B",
61+
"hook_name": f"blocks.{layer_index}.hook_resid_post",
62+
"hook_layer": str(layer_index),
63+
"hook_head_index": None,
64+
}
65+
)
66+
6367
# 将修改后的内容写回 cfg.json
64-
with open(cfg_file_path, 'w') as cfg_file:
68+
with open(cfg_file_path, "w") as cfg_file:
6569
json.dump(cfg_data, cfg_file, indent=4)
6670

71+
6772
# 遍历所有层次 (i 从 0 到 30)
68-
for i in tqdm(range(3,30,1)):
73+
for i in tqdm(range(3, 30, 1)):
6974
process_layer(i)
70-
delete_original_sae(i)
75+
delete_original_sae(i)

0 commit comments

Comments
 (0)