File size: 9,821 Bytes
f1069cc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import argparse
import os
import shutil
from pathlib import Path

import onnx
import torch
from packaging import version
from torch.onnx import export

from diffusers import OnnxRuntimeModel, OnnxStableDiffusionPipeline, StableDiffusionPipeline


is_torch_less_than_1_11 = version.parse(version.parse(torch.__version__).base_version) < version.parse("1.11")


def onnx_export(
    model,
    model_args: tuple,
    output_path: Path,
    ordered_input_names,
    output_names,
    dynamic_axes,
    opset,
    use_external_data_format=False,
):
    output_path.parent.mkdir(parents=True, exist_ok=True)
    # PyTorch deprecated the `enable_onnx_checker` and `use_external_data_format` arguments in v1.11,
    # so we check the torch version for backwards compatibility
    if is_torch_less_than_1_11:
        export(
            model,
            model_args,
            f=output_path.as_posix(),
            input_names=ordered_input_names,
            output_names=output_names,
            dynamic_axes=dynamic_axes,
            do_constant_folding=True,
            use_external_data_format=use_external_data_format,
            enable_onnx_checker=True,
            opset_version=opset,
        )
    else:
        export(
            model,
            model_args,
            f=output_path.as_posix(),
            input_names=ordered_input_names,
            output_names=output_names,
            dynamic_axes=dynamic_axes,
            do_constant_folding=True,
            opset_version=opset,
        )


@torch.no_grad()
def convert_models(model_path: str, output_path: str, opset: int, fp16: bool = False):
    dtype = torch.float16 if fp16 else torch.float32
    if fp16 and torch.cuda.is_available():
        device = "cuda"
    elif fp16 and not torch.cuda.is_available():
        raise ValueError("`float16` model export is only supported on GPUs with CUDA")
    else:
        device = "cpu"
    pipeline = StableDiffusionPipeline.from_pretrained(model_path, torch_dtype=dtype).to(device)
    output_path = Path(output_path)

    # TEXT ENCODER
    num_tokens = pipeline.text_encoder.config.max_position_embeddings
    text_hidden_size = pipeline.text_encoder.config.hidden_size
    text_input = pipeline.tokenizer(
        "A sample prompt",
        padding="max_length",
        max_length=pipeline.tokenizer.model_max_length,
        truncation=True,
        return_tensors="pt",
    )
    onnx_export(
        pipeline.text_encoder,
        # casting to torch.int32 until the CLIP fix is released: https://github.com/huggingface/transformers/pull/18515/files
        model_args=(text_input.input_ids.to(device=device, dtype=torch.int32)),
        output_path=output_path / "text_encoder" / "model.onnx",
        ordered_input_names=["input_ids"],
        output_names=["last_hidden_state", "pooler_output"],
        dynamic_axes={
            "input_ids": {0: "batch", 1: "sequence"},
        },
        opset=opset,
    )
    del pipeline.text_encoder

    # UNET
    unet_in_channels = pipeline.unet.config.in_channels
    unet_sample_size = pipeline.unet.config.sample_size
    unet_path = output_path / "unet" / "model.onnx"
    onnx_export(
        pipeline.unet,
        model_args=(
            torch.randn(2, unet_in_channels, unet_sample_size, unet_sample_size).to(device=device, dtype=dtype),
            torch.randn(2).to(device=device, dtype=dtype),
            torch.randn(2, num_tokens, text_hidden_size).to(device=device, dtype=dtype),
            False,
        ),
        output_path=unet_path,
        ordered_input_names=["sample", "timestep", "encoder_hidden_states", "return_dict"],
        output_names=["out_sample"],  # has to be different from "sample" for correct tracing
        dynamic_axes={
            "sample": {0: "batch", 1: "channels", 2: "height", 3: "width"},
            "timestep": {0: "batch"},
            "encoder_hidden_states": {0: "batch", 1: "sequence"},
        },
        opset=opset,
        use_external_data_format=True,  # UNet is > 2GB, so the weights need to be split
    )
    unet_model_path = str(unet_path.absolute().as_posix())
    unet_dir = os.path.dirname(unet_model_path)
    unet = onnx.load(unet_model_path)
    # clean up existing tensor files
    shutil.rmtree(unet_dir)
    os.mkdir(unet_dir)
    # collate external tensor files into one
    onnx.save_model(
        unet,
        unet_model_path,
        save_as_external_data=True,
        all_tensors_to_one_file=True,
        location="weights.pb",
        convert_attribute=False,
    )
    del pipeline.unet

    # VAE ENCODER
    vae_encoder = pipeline.vae
    vae_in_channels = vae_encoder.config.in_channels
    vae_sample_size = vae_encoder.config.sample_size
    # need to get the raw tensor output (sample) from the encoder
    vae_encoder.forward = lambda sample, return_dict: vae_encoder.encode(sample, return_dict)[0].sample()
    onnx_export(
        vae_encoder,
        model_args=(
            torch.randn(1, vae_in_channels, vae_sample_size, vae_sample_size).to(device=device, dtype=dtype),
            False,
        ),
        output_path=output_path / "vae_encoder" / "model.onnx",
        ordered_input_names=["sample", "return_dict"],
        output_names=["latent_sample"],
        dynamic_axes={
            "sample": {0: "batch", 1: "channels", 2: "height", 3: "width"},
        },
        opset=opset,
    )

    # VAE DECODER
    vae_decoder = pipeline.vae
    vae_latent_channels = vae_decoder.config.latent_channels
    vae_out_channels = vae_decoder.config.out_channels
    # forward only through the decoder part
    vae_decoder.forward = vae_encoder.decode
    onnx_export(
        vae_decoder,
        model_args=(
            torch.randn(1, vae_latent_channels, unet_sample_size, unet_sample_size).to(device=device, dtype=dtype),
            False,
        ),
        output_path=output_path / "vae_decoder" / "model.onnx",
        ordered_input_names=["latent_sample", "return_dict"],
        output_names=["sample"],
        dynamic_axes={
            "latent_sample": {0: "batch", 1: "channels", 2: "height", 3: "width"},
        },
        opset=opset,
    )
    del pipeline.vae

    # SAFETY CHECKER
    if pipeline.safety_checker is not None:
        safety_checker = pipeline.safety_checker
        clip_num_channels = safety_checker.config.vision_config.num_channels
        clip_image_size = safety_checker.config.vision_config.image_size
        safety_checker.forward = safety_checker.forward_onnx
        onnx_export(
            pipeline.safety_checker,
            model_args=(
                torch.randn(
                    1,
                    clip_num_channels,
                    clip_image_size,
                    clip_image_size,
                ).to(device=device, dtype=dtype),
                torch.randn(1, vae_sample_size, vae_sample_size, vae_out_channels).to(device=device, dtype=dtype),
            ),
            output_path=output_path / "safety_checker" / "model.onnx",
            ordered_input_names=["clip_input", "images"],
            output_names=["out_images", "has_nsfw_concepts"],
            dynamic_axes={
                "clip_input": {0: "batch", 1: "channels", 2: "height", 3: "width"},
                "images": {0: "batch", 1: "height", 2: "width", 3: "channels"},
            },
            opset=opset,
        )
        del pipeline.safety_checker
        safety_checker = OnnxRuntimeModel.from_pretrained(output_path / "safety_checker")
        feature_extractor = pipeline.feature_extractor
    else:
        safety_checker = None
        feature_extractor = None

    onnx_pipeline = OnnxStableDiffusionPipeline(
        vae_encoder=OnnxRuntimeModel.from_pretrained(output_path / "vae_encoder"),
        vae_decoder=OnnxRuntimeModel.from_pretrained(output_path / "vae_decoder"),
        text_encoder=OnnxRuntimeModel.from_pretrained(output_path / "text_encoder"),
        tokenizer=pipeline.tokenizer,
        unet=OnnxRuntimeModel.from_pretrained(output_path / "unet"),
        scheduler=pipeline.scheduler,
        safety_checker=safety_checker,
        feature_extractor=feature_extractor,
        requires_safety_checker=safety_checker is not None,
    )

    onnx_pipeline.save_pretrained(output_path)
    print("ONNX pipeline saved to", output_path)

    del pipeline
    del onnx_pipeline
    _ = OnnxStableDiffusionPipeline.from_pretrained(output_path, provider="CPUExecutionProvider")
    print("ONNX pipeline is loadable")


if __name__ == "__main__":
    parser = argparse.ArgumentParser()

    parser.add_argument(
        "--model_path",
        type=str,
        required=True,
        help="Path to the `diffusers` checkpoint to convert (either a local directory or on the Hub).",
    )

    parser.add_argument("--output_path", type=str, required=True, help="Path to the output model.")

    parser.add_argument(
        "--opset",
        default=14,
        type=int,
        help="The version of the ONNX operator set to use.",
    )
    parser.add_argument("--fp16", action="store_true", default=False, help="Export the models in `float16` mode")

    args = parser.parse_args()

    convert_models(args.model_path, args.output_path, args.opset, args.fp16)