File size: 11,501 Bytes
66662af
 
 
dede4f0
b273b9d
 
a5c2203
 
66662af
 
 
a5c2203
dede4f0
a5c2203
dede4f0
 
66662af
 
7a97ac2
2a91a35
7a97ac2
 
5193ca2
7a97ac2
 
 
 
 
 
 
2a91a35
 
7a97ac2
 
 
 
9bf11f8
 
 
 
b273b9d
 
 
 
 
 
 
 
 
 
a5c2203
b273b9d
dede4f0
 
 
 
a5c2203
 
dede4f0
 
a5c2203
 
dede4f0
 
b273b9d
0c2d8eb
 
66662af
 
 
 
650d083
dede4f0
 
 
66662af
dede4f0
a5c2203
dede4f0
a5c2203
66662af
a5c2203
 
 
 
66662af
dede4f0
 
 
 
 
5af6059
dede4f0
66662af
a5c2203
 
 
66662af
dede4f0
 
 
b273b9d
a5c2203
 
 
 
 
 
 
dede4f0
a5c2203
 
 
 
 
c968ad4
b2f4808
 
b273b9d
 
 
 
 
 
 
 
0c2d8eb
a5c2203
 
 
 
 
 
 
 
 
dede4f0
 
a5c2203
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dede4f0
b273b9d
dede4f0
 
 
b273b9d
a5c2203
 
b273b9d
a5c2203
 
 
b273b9d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dede4f0
 
 
 
a5c2203
9bf11f8
 
 
 
 
 
7fb5a3e
 
 
 
dede4f0
b273b9d
a5c2203
 
 
 
 
 
 
 
66ed500
bac319c
 
 
66ed500
bac319c
 
a5c2203
 
 
 
 
 
 
b273b9d
66662af
 
dede4f0
b273b9d
 
 
 
 
 
9bf11f8
5af6059
 
bac319c
9bf11f8
b273b9d
 
 
9bf11f8
5af6059
a5c2203
 
 
 
 
 
b2f4808
b273b9d
a5c2203
b273b9d
 
849eff4
 
 
 
5193ca2
849eff4
 
a5c2203
 
 
b273b9d
 
 
66662af
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b273b9d
 
 
 
 
66662af
 
 
b273b9d
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
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
import argparse
import json
import os
import shutil
from collections import defaultdict
from inspect import signature
from tempfile import TemporaryDirectory
from typing import Dict, List, Optional, Set

import torch

from huggingface_hub import CommitInfo, CommitOperationAdd, Discussion, HfApi, hf_hub_download
from huggingface_hub.file_download import repo_folder_name
from safetensors.torch import load_file, save_file
from transformers import AutoConfig
from transformers.pipelines.base import infer_framework_load_model


COMMIT_DESCRIPTION = """
This is an automated PR created with https://huggingface.co/spaces/safetensors/convert

This new file is equivalent to `pytorch_model.bin` but safe in the sense that
no arbitrary code can be put into it.

These files also happen to load much faster than their pytorch counterpart:
https://colab.research.google.com/github/huggingface/notebooks/blob/main/safetensors_doc/en/speed.ipynb

The widgets on your model page will run using this model even if this is not merged
making sure the file actually works.

If you find any issues: please report here: https://huggingface.co/spaces/safetensors/convert/discussions

Feel free to ignore this PR.
"""


class AlreadyExists(Exception):
    pass


def shared_pointers(tensors):
    ptrs = defaultdict(list)
    for k, v in tensors.items():
        ptrs[v.data_ptr()].append(k)
    failing = []
    for ptr, names in ptrs.items():
        if len(names) > 1:
            failing.append(names)
    return failing


def check_file_size(sf_filename: str, pt_filename: str):
    sf_size = os.stat(sf_filename).st_size
    pt_size = os.stat(pt_filename).st_size

    if (sf_size - pt_size) / pt_size > 0.01:
        raise RuntimeError(
            f"""The file size different is more than 1%:
         - {sf_filename}: {sf_size}
         - {pt_filename}: {pt_size}
         """
        )


def rename(pt_filename: str) -> str:
    filename, ext = os.path.splitext(pt_filename)
    local = f"{filename}.safetensors"
    local = local.replace("pytorch_model", "model")
    return local


def convert_multi(model_id: str, folder: str) -> List["CommitOperationAdd"]:
    filename = hf_hub_download(repo_id=model_id, filename="pytorch_model.bin.index.json")
    with open(filename, "r") as f:
        data = json.load(f)

    filenames = set(data["weight_map"].values())
    local_filenames = []
    for filename in filenames:
        pt_filename = hf_hub_download(repo_id=model_id, filename=filename)

        sf_filename = rename(pt_filename)
        sf_filename = os.path.join(folder, sf_filename)
        convert_file(pt_filename, sf_filename)
        local_filenames.append(sf_filename)

    index = os.path.join(folder, "model.safetensors.index.json")
    with open(index, "w") as f:
        newdata = {k: v for k, v in data.items()}
        newmap = {k: rename(v) for k, v in data["weight_map"].items()}
        newdata["weight_map"] = newmap
        json.dump(newdata, f, indent=4)
    local_filenames.append(index)

    operations = [
        CommitOperationAdd(path_in_repo=local.split("/")[-1], path_or_fileobj=local) for local in local_filenames
    ]

    return operations


def convert_single(model_id: str, folder: str) -> List["CommitOperationAdd"]:
    pt_filename = hf_hub_download(repo_id=model_id, filename="pytorch_model.bin")

    sf_name = "model.safetensors"
    sf_filename = os.path.join(folder, sf_name)
    convert_file(pt_filename, sf_filename)
    operations = [CommitOperationAdd(path_in_repo=sf_name, path_or_fileobj=sf_filename)]
    return operations


def convert_file(
    pt_filename: str,
    sf_filename: str,
):
    loaded = torch.load(pt_filename, map_location="cpu")
    if "state_dict" in loaded:
        loaded = loaded["state_dict"]
    shared = shared_pointers(loaded)
    for shared_weights in shared:
        for name in shared_weights[1:]:
            loaded.pop(name)

    # For tensors to be contiguous
    loaded = {k: v.contiguous() for k, v in loaded.items()}

    dirname = os.path.dirname(sf_filename)
    os.makedirs(dirname, exist_ok=True)
    save_file(loaded, sf_filename, metadata={"format": "pt"})
    check_file_size(sf_filename, pt_filename)
    reloaded = load_file(sf_filename)
    for k in loaded:
        pt_tensor = loaded[k]
        sf_tensor = reloaded[k]
        if not torch.equal(pt_tensor, sf_tensor):
            raise RuntimeError(f"The output tensors do not match for key {k}")


def create_diff(pt_infos: Dict[str, List[str]], sf_infos: Dict[str, List[str]]) -> str:
    errors = []
    for key in ["missing_keys", "mismatched_keys", "unexpected_keys"]:
        pt_set = set(pt_infos[key])
        sf_set = set(sf_infos[key])

        pt_only = pt_set - sf_set
        sf_only = sf_set - pt_set

        if pt_only:
            errors.append(f"{key} : PT warnings contain {pt_only} which are not present in SF warnings")
        if sf_only:
            errors.append(f"{key} : SF warnings contain {sf_only} which are not present in PT warnings")
    return "\n".join(errors)


def check_final_model(model_id: str, folder: str):
    config = hf_hub_download(repo_id=model_id, filename="config.json")
    shutil.copy(config, os.path.join(folder, "config.json"))
    config = AutoConfig.from_pretrained(folder)

    _, (pt_model, pt_infos) = infer_framework_load_model(model_id, config, output_loading_info=True)
    _, (sf_model, sf_infos) = infer_framework_load_model(folder, config, output_loading_info=True)

    if pt_infos != sf_infos:
        error_string = create_diff(pt_infos, sf_infos)
        raise ValueError(f"Different infos when reloading the model: {error_string}")

    pt_params = pt_model.state_dict()
    sf_params = sf_model.state_dict()

    pt_shared = shared_pointers(pt_params)
    sf_shared = shared_pointers(sf_params)
    if pt_shared != sf_shared:
        raise RuntimeError("The reconstructed model is wrong, shared tensors are different {shared_pt} != {shared_tf}")

    sig = signature(pt_model.forward)
    input_ids = torch.arange(10).unsqueeze(0)
    pixel_values = torch.randn(1, 3, 224, 224)
    input_values = torch.arange(1000).float().unsqueeze(0)
    kwargs = {}
    if "input_ids" in sig.parameters:
        kwargs["input_ids"] = input_ids
    if "decoder_input_ids" in sig.parameters:
        kwargs["decoder_input_ids"] = input_ids
    if "pixel_values" in sig.parameters:
        kwargs["pixel_values"] = pixel_values
    if "input_values" in sig.parameters:
        kwargs["input_values"] = input_values
    if "bbox" in sig.parameters:
        kwargs["bbox"] = torch.zeros((1, 10, 4)).long()
    if "image" in sig.parameters:
        kwargs["image"] = pixel_values

    if torch.cuda.is_available():
        pt_model = pt_model.cuda()
        sf_model = sf_model.cuda()
        kwargs = {k: v.cuda() for k, v in kwargs.items()}

    pt_logits = pt_model(**kwargs)[0]
    sf_logits = sf_model(**kwargs)[0]

    torch.testing.assert_close(sf_logits, pt_logits)
    print(f"Model {model_id} is ok !")


def previous_pr(api: "HfApi", model_id: str, pr_title: str) -> Optional["Discussion"]:
    try:
        discussions = api.get_repo_discussions(repo_id=model_id)
    except Exception:
        return None
    for discussion in discussions:
        if discussion.status == "open" and discussion.is_pull_request and discussion.title == pr_title:
            details = api.get_discussion_details(repo_id=model_id, discussion_num=discussion.num)
            if details.target_branch == "refs/heads/main":
                return discussion


def convert_generic(model_id: str, folder: str, filenames: Set[str]) -> List["CommitOperationAdd"]:
    operations = []

    extensions = set([".bin", ".ckpt"])
    for filename in filenames:
        prefix, ext = os.path.splitext(filename)
        if ext in extensions:
            pt_filename = hf_hub_download(model_id, filename=filename)
            dirname, raw_filename = os.path.split(filename)
            if raw_filename == "pytorch_model.bin":
                # XXX: This is a special case to handle `transformers` and the
                # `transformers` part of the model which is actually loaded by `transformers`.
                sf_in_repo = os.path.join(dirname, "model.safetensors")
            else:
                sf_in_repo = f"{prefix}.safetensors"
            sf_filename = os.path.join(folder, sf_in_repo)
            convert_file(pt_filename, sf_filename)
            operations.append(CommitOperationAdd(path_in_repo=sf_in_repo, path_or_fileobj=sf_filename))
    return operations


def convert(api: "HfApi", model_id: str, force: bool = False) -> Optional["CommitInfo"]:
    pr_title = "Adding `safetensors` variant of this model"
    info = api.model_info(model_id)
    filenames = set(s.rfilename for s in info.siblings)

    with TemporaryDirectory() as d:
        folder = os.path.join(d, repo_folder_name(repo_id=model_id, repo_type="models"))
        os.makedirs(folder)
        new_pr = None
        try:
            operations = None
            pr = previous_pr(api, model_id, pr_title)

            library_name = getattr(info, "library_name", None)
            if any(filename.endswith(".safetensors") for filename in filenames) and not force:
                raise AlreadyExists(f"Model {model_id} is already converted, skipping..")
            elif pr is not None and not force:
                url = f"https://huggingface.co/{model_id}/discussions/{pr.num}"
                new_pr = pr
                raise AlreadyExists(f"Model {model_id} already has an open PR check out {url}")
            elif library_name == "transformers":
                if "pytorch_model.bin" in filenames:
                    operations = convert_single(model_id, folder)
                elif "pytorch_model.bin.index.json" in filenames:
                    operations = convert_multi(model_id, folder)
                else:
                    raise RuntimeError(f"Model {model_id} doesn't seem to be a valid pytorch model. Cannot convert")
                check_final_model(model_id, folder)
            else:
                operations = convert_generic(model_id, folder, filenames)

            if operations:
                new_pr = api.create_commit(
                    repo_id=model_id,
                    operations=operations,
                    commit_message=pr_title,
                    commit_description=COMMIT_DESCRIPTION,
                    create_pr=True,
                )
                print(f"Pr created at {new_pr.pr_url}")
            else:
                print("No files to convert")
        finally:
            shutil.rmtree(folder)
        return new_pr


if __name__ == "__main__":
    DESCRIPTION = """
    Simple utility tool to convert automatically some weights on the hub to `safetensors` format.
    It is PyTorch exclusive for now.
    It works by downloading the weights (PT), converting them locally, and uploading them back
    as a PR on the hub.
    """
    parser = argparse.ArgumentParser(description=DESCRIPTION)
    parser.add_argument(
        "model_id",
        type=str,
        help="The name of the model on the hub to convert. E.g. `gpt2` or `facebook/wav2vec2-base-960h`",
    )
    parser.add_argument(
        "--force",
        action="store_true",
        help="Create the PR even if it already exists of if the model was already converted.",
    )
    args = parser.parse_args()
    model_id = args.model_id
    api = HfApi()
    convert(api, model_id, force=args.force)