old convert and lazy-way to convert to safetensors script
Browse files- bin2safetensors/convert_old.py +324 -0
bin2safetensors/convert_old.py
ADDED
@@ -0,0 +1,324 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import json
|
3 |
+
import os
|
4 |
+
import shutil
|
5 |
+
from collections import defaultdict
|
6 |
+
from inspect import signature
|
7 |
+
from tempfile import TemporaryDirectory
|
8 |
+
from typing import Dict, List, Optional, Set, Tuple
|
9 |
+
|
10 |
+
import torch
|
11 |
+
|
12 |
+
from huggingface_hub import CommitInfo, CommitOperationAdd, Discussion, HfApi, hf_hub_download
|
13 |
+
from huggingface_hub.file_download import repo_folder_name
|
14 |
+
from safetensors.torch import load_file, save_file
|
15 |
+
from transformers import AutoConfig
|
16 |
+
|
17 |
+
|
18 |
+
COMMIT_DESCRIPTION = """
|
19 |
+
This is an automated PR created with https://huggingface.co/spaces/safetensors/convert
|
20 |
+
|
21 |
+
This new file is equivalent to `pytorch_model.bin` but safe in the sense that
|
22 |
+
no arbitrary code can be put into it.
|
23 |
+
|
24 |
+
These files also happen to load much faster than their pytorch counterpart:
|
25 |
+
https://colab.research.google.com/github/huggingface/notebooks/blob/main/safetensors_doc/en/speed.ipynb
|
26 |
+
|
27 |
+
The widgets on your model page will run using this model even if this is not merged
|
28 |
+
making sure the file actually works.
|
29 |
+
|
30 |
+
If you find any issues: please report here: https://huggingface.co/spaces/safetensors/convert/discussions
|
31 |
+
|
32 |
+
Feel free to ignore this PR.
|
33 |
+
"""
|
34 |
+
|
35 |
+
ConversionResult = Tuple[List["CommitOperationAdd"], List[Tuple[str, "Exception"]]]
|
36 |
+
|
37 |
+
|
38 |
+
class AlreadyExists(Exception):
|
39 |
+
pass
|
40 |
+
|
41 |
+
|
42 |
+
def shared_pointers(tensors):
|
43 |
+
ptrs = defaultdict(list)
|
44 |
+
for k, v in tensors.items():
|
45 |
+
ptrs[v.data_ptr()].append(k)
|
46 |
+
failing = []
|
47 |
+
for ptr, names in ptrs.items():
|
48 |
+
if len(names) > 1:
|
49 |
+
failing.append(names)
|
50 |
+
return failing
|
51 |
+
|
52 |
+
|
53 |
+
def check_file_size(sf_filename: str, pt_filename: str):
|
54 |
+
sf_size = os.stat(sf_filename).st_size
|
55 |
+
pt_size = os.stat(pt_filename).st_size
|
56 |
+
|
57 |
+
if (sf_size - pt_size) / pt_size > 0.01:
|
58 |
+
raise RuntimeError(
|
59 |
+
f"""The file size different is more than 1%:
|
60 |
+
- {sf_filename}: {sf_size}
|
61 |
+
- {pt_filename}: {pt_size}
|
62 |
+
"""
|
63 |
+
)
|
64 |
+
|
65 |
+
|
66 |
+
def rename(pt_filename: str) -> str:
|
67 |
+
filename, ext = os.path.splitext(pt_filename)
|
68 |
+
local = f"{filename}.safetensors"
|
69 |
+
local = local.replace("pytorch_model", "model")
|
70 |
+
return local
|
71 |
+
|
72 |
+
|
73 |
+
def convert_multi(model_id: str, folder: str, token: Optional[str]) -> ConversionResult:
|
74 |
+
filename = hf_hub_download(repo_id=model_id, filename="pytorch_model.bin.index.json", token=token, cache_dir=folder)
|
75 |
+
with open(filename, "r") as f:
|
76 |
+
data = json.load(f)
|
77 |
+
|
78 |
+
filenames = set(data["weight_map"].values())
|
79 |
+
local_filenames = []
|
80 |
+
for filename in filenames:
|
81 |
+
pt_filename = hf_hub_download(repo_id=model_id, filename=filename, token=token, cache_dir=folder)
|
82 |
+
|
83 |
+
sf_filename = rename(pt_filename)
|
84 |
+
sf_filename = os.path.join(folder, sf_filename)
|
85 |
+
convert_file(pt_filename, sf_filename)
|
86 |
+
local_filenames.append(sf_filename)
|
87 |
+
|
88 |
+
index = os.path.join(folder, "model.safetensors.index.json")
|
89 |
+
with open(index, "w") as f:
|
90 |
+
newdata = {k: v for k, v in data.items()}
|
91 |
+
newmap = {k: rename(v) for k, v in data["weight_map"].items()}
|
92 |
+
newdata["weight_map"] = newmap
|
93 |
+
json.dump(newdata, f, indent=4)
|
94 |
+
local_filenames.append(index)
|
95 |
+
|
96 |
+
operations = [
|
97 |
+
CommitOperationAdd(path_in_repo=local.split("/")[-1], path_or_fileobj=local) for local in local_filenames
|
98 |
+
]
|
99 |
+
errors: List[Tuple[str, "Exception"]] = []
|
100 |
+
|
101 |
+
return operations, errors
|
102 |
+
|
103 |
+
|
104 |
+
def convert_single(model_id: str, folder: str, token: Optional[str]) -> ConversionResult:
|
105 |
+
pt_filename = hf_hub_download(repo_id=model_id, filename="pytorch_model.bin", token=token, cache_dir=folder)
|
106 |
+
|
107 |
+
sf_name = "model.safetensors"
|
108 |
+
sf_filename = os.path.join(folder, sf_name)
|
109 |
+
convert_file(pt_filename, sf_filename)
|
110 |
+
operations = [CommitOperationAdd(path_in_repo=sf_name, path_or_fileobj=sf_filename)]
|
111 |
+
errors: List[Tuple[str, "Exception"]] = []
|
112 |
+
return operations, errors
|
113 |
+
|
114 |
+
|
115 |
+
def convert_file(
|
116 |
+
pt_filename: str,
|
117 |
+
sf_filename: str,
|
118 |
+
):
|
119 |
+
loaded = torch.load(pt_filename, map_location="cpu")
|
120 |
+
if "state_dict" in loaded:
|
121 |
+
loaded = loaded["state_dict"]
|
122 |
+
shared = shared_pointers(loaded)
|
123 |
+
for shared_weights in shared:
|
124 |
+
for name in shared_weights[1:]:
|
125 |
+
loaded.pop(name)
|
126 |
+
|
127 |
+
# For tensors to be contiguous
|
128 |
+
loaded = {k: v.contiguous() for k, v in loaded.items()}
|
129 |
+
|
130 |
+
dirname = os.path.dirname(sf_filename)
|
131 |
+
#os.makedirs(dirname, exist_ok=True)
|
132 |
+
save_file(loaded, sf_filename, metadata={"format": "pt"})
|
133 |
+
check_file_size(sf_filename, pt_filename)
|
134 |
+
reloaded = load_file(sf_filename)
|
135 |
+
for k in loaded:
|
136 |
+
pt_tensor = loaded[k]
|
137 |
+
sf_tensor = reloaded[k]
|
138 |
+
if not torch.equal(pt_tensor, sf_tensor):
|
139 |
+
raise RuntimeError(f"The output tensors do not match for key {k}")
|
140 |
+
|
141 |
+
|
142 |
+
def create_diff(pt_infos: Dict[str, List[str]], sf_infos: Dict[str, List[str]]) -> str:
|
143 |
+
errors = []
|
144 |
+
for key in ["missing_keys", "mismatched_keys", "unexpected_keys"]:
|
145 |
+
pt_set = set(pt_infos[key])
|
146 |
+
sf_set = set(sf_infos[key])
|
147 |
+
|
148 |
+
pt_only = pt_set - sf_set
|
149 |
+
sf_only = sf_set - pt_set
|
150 |
+
|
151 |
+
if pt_only:
|
152 |
+
errors.append(f"{key} : PT warnings contain {pt_only} which are not present in SF warnings")
|
153 |
+
if sf_only:
|
154 |
+
errors.append(f"{key} : SF warnings contain {sf_only} which are not present in PT warnings")
|
155 |
+
return "\n".join(errors)
|
156 |
+
|
157 |
+
|
158 |
+
def check_final_model(model_id: str, folder: str, token: Optional[str]):
|
159 |
+
config = hf_hub_download(repo_id=model_id, filename="config.json", token=token, cache_dir=folder)
|
160 |
+
shutil.copy(config, os.path.join(folder, "config.json"))
|
161 |
+
config = AutoConfig.from_pretrained(folder)
|
162 |
+
|
163 |
+
import transformers
|
164 |
+
|
165 |
+
class_ = getattr(transformers, config.architectures[0])
|
166 |
+
(pt_model, pt_infos) = class_.from_pretrained(folder, output_loading_info=True)
|
167 |
+
(sf_model, sf_infos) = class_.from_pretrained(folder, output_loading_info=True)
|
168 |
+
|
169 |
+
if pt_infos != sf_infos:
|
170 |
+
error_string = create_diff(pt_infos, sf_infos)
|
171 |
+
raise ValueError(f"Different infos when reloading the model: {error_string}")
|
172 |
+
|
173 |
+
pt_params = pt_model.state_dict()
|
174 |
+
sf_params = sf_model.state_dict()
|
175 |
+
|
176 |
+
pt_shared = shared_pointers(pt_params)
|
177 |
+
sf_shared = shared_pointers(sf_params)
|
178 |
+
if pt_shared != sf_shared:
|
179 |
+
raise RuntimeError("The reconstructed model is wrong, shared tensors are different {shared_pt} != {shared_tf}")
|
180 |
+
|
181 |
+
sig = signature(pt_model.forward)
|
182 |
+
input_ids = torch.arange(10).unsqueeze(0)
|
183 |
+
pixel_values = torch.randn(1, 3, 224, 224)
|
184 |
+
input_values = torch.arange(1000).float().unsqueeze(0)
|
185 |
+
# Hardcoded for whisper basically
|
186 |
+
input_features = torch.zeros((1, 80, 3000))
|
187 |
+
kwargs = {}
|
188 |
+
if "input_ids" in sig.parameters:
|
189 |
+
kwargs["input_ids"] = input_ids
|
190 |
+
if "input_features" in sig.parameters:
|
191 |
+
kwargs["input_features"] = input_features
|
192 |
+
if "decoder_input_ids" in sig.parameters:
|
193 |
+
kwargs["decoder_input_ids"] = input_ids
|
194 |
+
if "pixel_values" in sig.parameters:
|
195 |
+
kwargs["pixel_values"] = pixel_values
|
196 |
+
if "input_values" in sig.parameters:
|
197 |
+
kwargs["input_values"] = input_values
|
198 |
+
if "bbox" in sig.parameters:
|
199 |
+
kwargs["bbox"] = torch.zeros((1, 10, 4)).long()
|
200 |
+
if "image" in sig.parameters:
|
201 |
+
kwargs["image"] = pixel_values
|
202 |
+
|
203 |
+
if torch.cuda.is_available():
|
204 |
+
pt_model = pt_model.cuda()
|
205 |
+
sf_model = sf_model.cuda()
|
206 |
+
kwargs = {k: v.cuda() for k, v in kwargs.items()}
|
207 |
+
|
208 |
+
try:
|
209 |
+
pt_logits = pt_model(**kwargs)[0]
|
210 |
+
except Exception as e:
|
211 |
+
try:
|
212 |
+
# Musicgen special exception.
|
213 |
+
decoder_input_ids = torch.ones((input_ids.shape[0] * pt_model.decoder.num_codebooks, 1), dtype=torch.long)
|
214 |
+
if torch.cuda.is_available():
|
215 |
+
decoder_input_ids = decoder_input_ids.cuda()
|
216 |
+
|
217 |
+
kwargs["decoder_input_ids"] = decoder_input_ids
|
218 |
+
pt_logits = pt_model(**kwargs)[0]
|
219 |
+
except Exception:
|
220 |
+
raise e
|
221 |
+
sf_logits = sf_model(**kwargs)[0]
|
222 |
+
|
223 |
+
torch.testing.assert_close(sf_logits, pt_logits)
|
224 |
+
print(f"Model {model_id} is ok !")
|
225 |
+
|
226 |
+
|
227 |
+
def previous_pr(api: "HfApi", model_id: str, pr_title: str) -> Optional["Discussion"]:
|
228 |
+
try:
|
229 |
+
main_commit = api.list_repo_commits(model_id)[0].commit_id
|
230 |
+
discussions = api.get_repo_discussions(repo_id=model_id)
|
231 |
+
except Exception:
|
232 |
+
return None
|
233 |
+
for discussion in discussions:
|
234 |
+
if discussion.status == "open" and discussion.is_pull_request and discussion.title == pr_title:
|
235 |
+
commits = api.list_repo_commits(model_id, revision=discussion.git_reference)
|
236 |
+
|
237 |
+
if main_commit == commits[1].commit_id:
|
238 |
+
return discussion
|
239 |
+
return None
|
240 |
+
|
241 |
+
|
242 |
+
def convert_generic(model_id: str, folder: str, filenames: Set[str], token: Optional[str]) -> ConversionResult:
|
243 |
+
operations = []
|
244 |
+
errors = []
|
245 |
+
|
246 |
+
extensions = set([".bin", ".ckpt"])
|
247 |
+
for filename in filenames:
|
248 |
+
prefix, ext = os.path.splitext(filename)
|
249 |
+
if ext in extensions:
|
250 |
+
pt_filename = hf_hub_download(model_id, filename=filename, token=token, cache_dir=folder)
|
251 |
+
dirname, raw_filename = os.path.split(filename)
|
252 |
+
if raw_filename == "pytorch_model.bin":
|
253 |
+
# XXX: This is a special case to handle `transformers` and the
|
254 |
+
# `transformers` part of the model which is actually loaded by `transformers`.
|
255 |
+
sf_in_repo = os.path.join(dirname, "model.safetensors")
|
256 |
+
else:
|
257 |
+
sf_in_repo = f"{prefix}.safetensors"
|
258 |
+
sf_filename = os.path.join(folder, sf_in_repo)
|
259 |
+
try:
|
260 |
+
convert_file(pt_filename, sf_filename)
|
261 |
+
operations.append(CommitOperationAdd(path_in_repo=sf_in_repo, path_or_fileobj=sf_filename))
|
262 |
+
except Exception as e:
|
263 |
+
errors.append((pt_filename, e))
|
264 |
+
return operations, errors
|
265 |
+
|
266 |
+
|
267 |
+
def convert(api: "HfApi", model_id: str, force: bool = False) -> Tuple["CommitInfo", List[Tuple[str, "Exception"]]]:
|
268 |
+
pr_title = "Adding `safetensors` variant of this model"
|
269 |
+
info = api.model_info(model_id)
|
270 |
+
filenames = set(s.rfilename for s in info.siblings)
|
271 |
+
|
272 |
+
with TemporaryDirectory() as d:
|
273 |
+
folder = os.path.join(d, repo_folder_name(repo_id=model_id, repo_type="models"))
|
274 |
+
os.makedirs(folder)
|
275 |
+
new_pr = None
|
276 |
+
try:
|
277 |
+
operations = None
|
278 |
+
pr = previous_pr(api, model_id, pr_title)
|
279 |
+
|
280 |
+
library_name = getattr(info, "library_name", None)
|
281 |
+
if any(filename.endswith(".safetensors") for filename in filenames) and not force:
|
282 |
+
raise AlreadyExists(f"Model {model_id} is already converted, skipping..")
|
283 |
+
elif pr is not None and not force:
|
284 |
+
url = f"https://huggingface.co/{model_id}/discussions/{pr.num}"
|
285 |
+
new_pr = pr
|
286 |
+
raise AlreadyExists(f"Model {model_id} already has an open PR check out {url}")
|
287 |
+
elif library_name == "transformers":
|
288 |
+
if "pytorch_model.bin" in filenames:
|
289 |
+
operations, errors = convert_single(model_id, folder, token=api.token)
|
290 |
+
elif "pytorch_model.bin.index.json" in filenames:
|
291 |
+
operations, errors = convert_multi(model_id, folder, token=api.token)
|
292 |
+
else:
|
293 |
+
raise RuntimeError(f"Model {model_id} doesn't seem to be a valid pytorch model. Cannot convert")
|
294 |
+
# check_final_model(model_id, folder, token=api.token)
|
295 |
+
else:
|
296 |
+
operations, errors = convert_generic(model_id, folder, filenames, token=api.token)
|
297 |
+
|
298 |
+
if operations:
|
299 |
+
new_pr = api.create_commit(
|
300 |
+
repo_id=model_id,
|
301 |
+
operations=operations,
|
302 |
+
commit_message=pr_title,
|
303 |
+
commit_description=COMMIT_DESCRIPTION,
|
304 |
+
create_pr=True,
|
305 |
+
)
|
306 |
+
print(f"Pr created at {new_pr.pr_url}")
|
307 |
+
else:
|
308 |
+
print("No files to convert")
|
309 |
+
finally:
|
310 |
+
shutil.rmtree(folder)
|
311 |
+
return new_pr, errors
|
312 |
+
|
313 |
+
|
314 |
+
if __name__ == "__main__":
|
315 |
+
DESCRIPTION = """
|
316 |
+
Simple utility tool to convert automatically some weights on the hub to `safetensors` format.
|
317 |
+
It is PyTorch exclusive for now.
|
318 |
+
It works by downloading the weights (PT), converting them locally, and uploading them back
|
319 |
+
as a PR on the hub.
|
320 |
+
"""
|
321 |
+
for i in range(1, 16): # Range starts at 1 and ends at 15
|
322 |
+
input_filename = f"jondurbin_airoboros-l2-70b-gpt4-1.4.1/pytorch_model-{i:05d}-of-00015.bin"
|
323 |
+
output_filename = f"pytorch_model-{i:05d}-of-00015.safetensors"
|
324 |
+
convert_file(input_filename, output_filename)
|