docs: add conversion script
Browse files- convert_qwen2_to_llama.py +182 -0
convert_qwen2_to_llama.py
ADDED
@@ -0,0 +1,182 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Converts the 2nd version of the Qwen models in the same format as LLaMA2.
|
3 |
+
# Usage: python convert_qwen2_to_llama.py --input_dir magnum-72b-v1 --output_dir magnum-72b-v1-llamaify --save_safetensors --continue_conversion
|
4 |
+
# Original script: https://github.com/Minami-su/character_AI_open/blob/main/llamafy_qwen_v2.py
|
5 |
+
|
6 |
+
import json
|
7 |
+
import os
|
8 |
+
from collections import OrderedDict
|
9 |
+
from typing import Any, Dict, Optional
|
10 |
+
|
11 |
+
import fire
|
12 |
+
import torch
|
13 |
+
from safetensors import safe_open
|
14 |
+
from safetensors.torch import save_file
|
15 |
+
from tqdm import tqdm
|
16 |
+
from transformers.modeling_utils import (
|
17 |
+
SAFE_WEIGHTS_INDEX_NAME,
|
18 |
+
SAFE_WEIGHTS_NAME,
|
19 |
+
WEIGHTS_INDEX_NAME,
|
20 |
+
WEIGHTS_NAME,
|
21 |
+
shard_checkpoint,
|
22 |
+
)
|
23 |
+
from transformers.utils import check_min_version
|
24 |
+
|
25 |
+
try:
|
26 |
+
check_min_version("4.34.0")
|
27 |
+
except Exception:
|
28 |
+
raise ValueError("Please upgrade `transformers` to 4.34.0")
|
29 |
+
|
30 |
+
CONFIG_NAME = "config.json"
|
31 |
+
|
32 |
+
|
33 |
+
def load_existing_shards(
|
34 |
+
output_dir: str, save_safetensors: bool
|
35 |
+
) -> Dict[str, torch.Tensor]:
|
36 |
+
existing_state_dict = OrderedDict()
|
37 |
+
weights_name = SAFE_WEIGHTS_NAME if save_safetensors else WEIGHTS_NAME
|
38 |
+
index_name = SAFE_WEIGHTS_INDEX_NAME if save_safetensors else WEIGHTS_INDEX_NAME
|
39 |
+
|
40 |
+
if os.path.exists(os.path.join(output_dir, index_name)):
|
41 |
+
with open(os.path.join(output_dir, index_name), "r", encoding="utf-8") as f:
|
42 |
+
index = json.load(f)
|
43 |
+
|
44 |
+
for shard_file in tqdm(
|
45 |
+
index["weight_map"].values(), desc="Loading existing shards"
|
46 |
+
):
|
47 |
+
if os.path.exists(os.path.join(output_dir, shard_file)):
|
48 |
+
if save_safetensors:
|
49 |
+
with safe_open(
|
50 |
+
os.path.join(output_dir, shard_file),
|
51 |
+
framework="pt",
|
52 |
+
device="cpu",
|
53 |
+
) as f:
|
54 |
+
for key in f.keys():
|
55 |
+
existing_state_dict[key] = f.get_tensor(key)
|
56 |
+
else:
|
57 |
+
shard = torch.load(
|
58 |
+
os.path.join(output_dir, shard_file), map_location="cpu"
|
59 |
+
)
|
60 |
+
existing_state_dict.update(shard)
|
61 |
+
|
62 |
+
return existing_state_dict
|
63 |
+
|
64 |
+
|
65 |
+
def save_weight(
|
66 |
+
input_dir: str,
|
67 |
+
output_dir: str,
|
68 |
+
shard_size: str,
|
69 |
+
save_safetensors: bool,
|
70 |
+
continue_conversion: bool,
|
71 |
+
) -> str:
|
72 |
+
qwen_state_dict: Dict[str, torch.Tensor] = OrderedDict()
|
73 |
+
for filepath in tqdm(os.listdir(input_dir), desc="Load weights"):
|
74 |
+
if os.path.isfile(os.path.join(input_dir, filepath)) and filepath.endswith(
|
75 |
+
".safetensors"
|
76 |
+
):
|
77 |
+
with safe_open(
|
78 |
+
os.path.join(input_dir, filepath), framework="pt", device="cpu"
|
79 |
+
) as f:
|
80 |
+
for key in f.keys():
|
81 |
+
qwen_state_dict[key] = f.get_tensor(key)
|
82 |
+
|
83 |
+
llama2_state_dict: Dict[str, torch.Tensor] = OrderedDict()
|
84 |
+
if continue_conversion:
|
85 |
+
llama2_state_dict = load_existing_shards(output_dir, save_safetensors)
|
86 |
+
|
87 |
+
torch_dtype = None
|
88 |
+
for key, value in tqdm(qwen_state_dict.items(), desc="Convert format"):
|
89 |
+
if torch_dtype is None:
|
90 |
+
torch_dtype = value.dtype
|
91 |
+
if "self_attn.o_proj" in key:
|
92 |
+
llama2_state_dict[key] = value
|
93 |
+
bias_key = key.replace(".weight", ".bias")
|
94 |
+
if bias_key not in llama2_state_dict:
|
95 |
+
llama2_state_dict[bias_key] = torch.zeros_like(value[:, 0]).squeeze()
|
96 |
+
else:
|
97 |
+
llama2_state_dict[key] = value
|
98 |
+
|
99 |
+
weights_name = SAFE_WEIGHTS_NAME if save_safetensors else WEIGHTS_NAME
|
100 |
+
shards, index = shard_checkpoint(
|
101 |
+
llama2_state_dict, max_shard_size=shard_size, weights_name=weights_name
|
102 |
+
)
|
103 |
+
|
104 |
+
for shard_file, shard in tqdm(shards.items(), desc="Save weights"):
|
105 |
+
if save_safetensors:
|
106 |
+
save_file(
|
107 |
+
shard, os.path.join(output_dir, shard_file), metadata={"format": "pt"}
|
108 |
+
)
|
109 |
+
else:
|
110 |
+
torch.save(shard, os.path.join(output_dir, shard_file))
|
111 |
+
|
112 |
+
if index is None:
|
113 |
+
print(f"Model weights saved in {os.path.join(output_dir, weights_name)}")
|
114 |
+
else:
|
115 |
+
index_name = SAFE_WEIGHTS_INDEX_NAME if save_safetensors else WEIGHTS_INDEX_NAME
|
116 |
+
with open(os.path.join(output_dir, index_name), "w", encoding="utf-8") as f:
|
117 |
+
json.dump(index, f, indent=2, sort_keys=True)
|
118 |
+
print(f"Model weights saved in {output_dir}")
|
119 |
+
|
120 |
+
return str(torch_dtype).replace("torch.", "")
|
121 |
+
|
122 |
+
|
123 |
+
def save_config(input_dir: str, output_dir: str, torch_dtype: str):
|
124 |
+
with open(os.path.join(input_dir, CONFIG_NAME), "r", encoding="utf-8") as f:
|
125 |
+
qwen_config_dict: Dict[str, Any] = json.load(f)
|
126 |
+
|
127 |
+
llama2_config_dict: Dict[str, Any] = OrderedDict()
|
128 |
+
llama2_config_dict["architectures"] = ["LlamaForCausalLM"]
|
129 |
+
llama2_config_dict["attention_bias"] = True
|
130 |
+
llama2_config_dict["attention_dropout"] = qwen_config_dict["attention_dropout"]
|
131 |
+
llama2_config_dict["hidden_act"] = "silu"
|
132 |
+
llama2_config_dict["hidden_size"] = qwen_config_dict["hidden_size"]
|
133 |
+
llama2_config_dict["initializer_range"] = qwen_config_dict["initializer_range"]
|
134 |
+
llama2_config_dict["intermediate_size"] = qwen_config_dict["intermediate_size"]
|
135 |
+
llama2_config_dict["max_position_embeddings"] = 32767 # Qwen2-72B-Instruct
|
136 |
+
llama2_config_dict["max_window_layers"] = qwen_config_dict["max_window_layers"]
|
137 |
+
llama2_config_dict["model_type"] = "llama"
|
138 |
+
llama2_config_dict["num_attention_heads"] = qwen_config_dict["num_attention_heads"]
|
139 |
+
llama2_config_dict["num_hidden_layers"] = qwen_config_dict["num_hidden_layers"]
|
140 |
+
llama2_config_dict["num_key_value_heads"] = qwen_config_dict["num_key_value_heads"]
|
141 |
+
llama2_config_dict["pretraining_tp"] = 1
|
142 |
+
llama2_config_dict["rms_norm_eps"] = qwen_config_dict["rms_norm_eps"]
|
143 |
+
llama2_config_dict["rope_theta"] = qwen_config_dict["rope_theta"]
|
144 |
+
llama2_config_dict["rope_scaling"] = None
|
145 |
+
llama2_config_dict["sliding_window"] = qwen_config_dict["sliding_window"]
|
146 |
+
llama2_config_dict["tie_word_embeddings"] = qwen_config_dict["tie_word_embeddings"]
|
147 |
+
llama2_config_dict["torch_dtype"] = torch_dtype
|
148 |
+
llama2_config_dict["transformers_version"] = "4.37.0"
|
149 |
+
llama2_config_dict["use_cache"] = True
|
150 |
+
llama2_config_dict["use_sliding_window"] = qwen_config_dict["use_sliding_window"]
|
151 |
+
llama2_config_dict["vocab_size"] = qwen_config_dict["vocab_size"]
|
152 |
+
|
153 |
+
with open(os.path.join(output_dir, CONFIG_NAME), "w", encoding="utf-8") as f:
|
154 |
+
json.dump(llama2_config_dict, f, indent=2)
|
155 |
+
print(f"Model config saved in {os.path.join(output_dir, CONFIG_NAME)}")
|
156 |
+
|
157 |
+
|
158 |
+
def llamafy_qwen_v2(
|
159 |
+
input_dir: str,
|
160 |
+
output_dir: str,
|
161 |
+
shard_size: Optional[str] = "4GB",
|
162 |
+
save_safetensors: Optional[bool] = False,
|
163 |
+
continue_conversion: Optional[bool] = False,
|
164 |
+
):
|
165 |
+
if not continue_conversion:
|
166 |
+
try:
|
167 |
+
os.makedirs(output_dir, exist_ok=False)
|
168 |
+
except Exception as e:
|
169 |
+
raise ValueError(
|
170 |
+
"Output dir already exists. Use --continue_conversion to resume."
|
171 |
+
) from e
|
172 |
+
else:
|
173 |
+
os.makedirs(output_dir, exist_ok=True)
|
174 |
+
|
175 |
+
torch_dtype = save_weight(
|
176 |
+
input_dir, output_dir, shard_size, save_safetensors, continue_conversion
|
177 |
+
)
|
178 |
+
save_config(input_dir, output_dir, torch_dtype)
|
179 |
+
|
180 |
+
|
181 |
+
if __name__ == "__main__":
|
182 |
+
fire.Fire(llamafy_qwen_v2)
|