File size: 10,863 Bytes
dd5ad09
 
 
 
 
 
 
 
 
 
 
 
 
a3c5547
dd5ad09
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1788c41
 
dd5ad09
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0dee7c2
dd5ad09
 
 
 
 
 
 
a3c5547
dd5ad09
 
 
 
 
 
 
a3c5547
 
 
 
1788c41
a3c5547
 
 
 
 
 
 
 
 
 
89cc842
a3c5547
795abe7
a3c5547
 
 
 
 
8111116
dd5ad09
 
0dee7c2
dd5ad09
 
 
 
 
a3c5547
 
 
 
1788c41
a3c5547
 
 
8111116
dd5ad09
 
 
 
 
 
 
 
 
4a04f8f
 
 
 
 
 
 
 
dd5ad09
 
 
 
 
4a04f8f
dd5ad09
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0dee7c2
dd5ad09
 
 
 
a3c5547
 
dd5ad09
 
 
 
 
 
 
 
 
 
 
 
 
 
a3c5547
 
 
 
 
1788c41
a3c5547
 
 
 
 
89cc842
a3c5547
795abe7
a3c5547
 
 
 
 
dd5ad09
 
8111116
dd5ad09
 
 
 
 
 
 
 
 
 
 
 
1788c41
dd5ad09
 
 
 
 
 
 
 
 
 
8111116
dd5ad09
8111116
dd5ad09
 
 
8111116
a3c5547
 
dd5ad09
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
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, Tuple

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, _remove_duplicate_names


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.
"""
PR_TITLE = "Adding `safetensors` variant of this model"


ConversionResult = Tuple[List["CommitOperationAdd"], List[Tuple[str, "Exception"]]]


class AlreadyExists(Exception):
    pass


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, api: "HfApi") -> ConversionResult:
    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())


    
    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)

    
    new_pr = api.create_commit(
        repo_id=model_id,
        operations=[CommitOperationAdd(path_in_repo=index.split("/")[-1], path_or_fileobj=index)],
        commit_message=PR_TITLE,
        commit_description=COMMIT_DESCRIPTION,
        create_pr=True,
    )

    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)
        api.create_commit(
                repo_id=model_id,
                commit_message=f"Adds {sf_filename}",
                revision=new_pr.pr_revision,
                operations=[CommitOperationAdd(path_in_repo=sf_filename.split("/")[-1], path_or_fileobj=sf_filename)],
                create_pr=False,
        )
        os.remove(pt_filename)
        os.remove(sf_filename)
    return new_pr, []


def convert_single(model_id: str, folder: str, api: "HfApi") -> ConversionResult:
    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)

    new_pr = api.create_commit(
        repo_id=model_id,
        operations=[CommitOperationAdd(path_in_repo=sf_name, path_or_fileobj=sf_filename)],
        commit_message=PR_TITLE,
        commit_description=COMMIT_DESCRIPTION,
        create_pr=True,
    )
    return new_pr, []


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"]
    to_removes = _remove_duplicate_names(loaded)

    metadata = {"format": "pt"}
    for kept_name, to_remove_group in to_removes.items():
        for to_remove in to_remove_group:
            if to_remove not in metadata:
                metadata[to_remove] = kept_name
            del loaded[to_remove]
    # 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=metadata)
    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 previous_pr(api: "HfApi", model_id: str, pr_title: str) -> Optional["Discussion"]:
    try:
        main_commit = api.list_repo_commits(model_id)[0].commit_id
        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:
            commits = api.list_repo_commits(model_id, revision=discussion.git_reference)

            if main_commit == commits[1].commit_id:
                return discussion
    return None


def convert_generic(model_id: str, folder: str, filenames: Set[str], api: "HfApi") -> ConversionResult:
    operations = []
    errors = []

    extensions = set([".bin", ".ckpt"])

    new_pr = None
    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)
            try:
                convert_file(pt_filename, sf_filename)

                if new_pr is None:
                    new_pr = api.create_commit(
                        repo_id=model_id,
                        operations=[CommitOperationAdd(path_in_repo=sf_in_repo, path_or_fileobj=sf_filename)],
                        commit_message=PR_TITLE,
                        commit_description=COMMIT_DESCRIPTION,
                        create_pr=True,
                    )
                else:
                    api.create_commit(
                        repo_id=model_id,
                        commit_message=f"Adds {sf_filename}",
                        revision=new_pr.pr_revision,
                        operations=[CommitOperationAdd(path_in_repo=sf_in_repo, path_or_fileobj=sf_filename)],
                        create_pr=False,
                    )
                os.remove(pt_filename)
                os.remove(sf_filename)
            except Exception as e:
                errors.append((pt_filename, e))
    return new_pr, errors


def convert(api: "HfApi", model_id: str, force: bool = False) -> Tuple["CommitInfo", List["Exception"]]:
    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:
                    new_pr, errors = convert_single(model_id, folder, api)
                elif "pytorch_model.bin.index.json" in filenames:
                    new_pr, errors = convert_multi(model_id, folder, api)
                else:
                    raise RuntimeError(f"Model {model_id} doesn't seem to be a valid pytorch model. Cannot convert")
            else:
                new_pr, errors = convert_generic(model_id, folder, filenames, api)

            print(f"Pr created at {new_pr.pr_url}")
        finally:
            shutil.rmtree(folder)
        return new_pr, errors


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.",
    )
    parser.add_argument(
        "-y",
        action="store_true",
        help="Ignore safety prompt",
    )
    args = parser.parse_args()
    model_id = args.model_id
    api = HfApi()
    if args.y:
        txt = "y"
    else:
        txt = input(
            "This conversion script will unpickle a pickled file, which is inherently unsafe. If you do not trust this file, we invite you to use"
            " https://huggingface.co/spaces/safetensors/convert or google colab or other hosted solution to avoid potential issues with this file."
            " Continue [Y/n] ?"
        )
    if txt.lower() in {"", "y"}:
        _commit_info, _errors = convert(api, model_id, force=args.force)
    else:
        print(f"Answer was `{txt}` aborting.")