Repository navigation
Expand file tree
/
Copy pathStableDiffusionControlnetONNXPipelineTrigger.py
More file actions
204 lines (183 loc) · 7.59 KB
/
Copy pathStableDiffusionControlnetONNXPipelineTrigger.py
File metadata and controls
204 lines (183 loc) · 7.59 KB
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
# Copyright Advanced Micro Devices, Inc.
#
# SPDX-License-Identifier: MIT
import time
import json
import os
import torch
import importlib
import logging as Logger
import numpy as np
from diffusers.utils import load_image
from PIL import Image
from transformers import CLIPTokenizer, CLIPTextModel, CLIPImageProcessor
from src.pipeline import StableDiffusionControlNetONNXPipelineAMD
from src.utils import common
from src.utils.canny import canny_edges
from src.utils.model_loader import ModelLoadConfig, ModelType, load_model
from src.trigger.base_trigger import BaseOnnxPipelineTrigger
class StableDiffusionControlnetONNXPipelineTrigger(BaseOnnxPipelineTrigger):
# Accept args namespace directly — eliminates runner getter boilerplate
def __init__(self, args):
super().__init__(args)
root_model_path = args.model_path
custom_op_path = args.custom_op_path
enable_compile = args.enable_compile
self.tokenizer = load_model(
ModelLoadConfig(
model_path=os.path.join(root_model_path, "tokenizer"),
model_type=ModelType.CLIP_TOKENIZER,
),
self.load_metrics_collector,
)
# Prefer the fp16 ONNX text encoder (model.onnx); fall back to the
# fp32 PyTorch safetensors version when it is missing / fails to load.
try:
self.text_encoder = load_model(
ModelLoadConfig(
model_path=os.path.join(root_model_path, "text_encoder"),
model_type=ModelType.ONNX_TEXT_ENCODER,
execution_providers=["DmlExecutionProvider", "CPUExecutionProvider"],
),
self.load_metrics_collector,
)
except Exception as e:
Logger.warning("ONNX text_encoder unavailable (%s); using PyTorch.", e)
self.text_encoder = load_model(
ModelLoadConfig(
model_path=os.path.join(root_model_path, "text_encoder"),
model_type=ModelType.CLIP_TEXT_MODEL,
),
self.load_metrics_collector,
)
self.scheduler = load_model(
ModelLoadConfig(
model_path=os.path.join(root_model_path, "scheduler"),
model_type=ModelType.SCHEDULER,
),
self.load_metrics_collector,
)
self.feature_extractor = load_model(
ModelLoadConfig(
model_path=os.path.join(root_model_path, "feature_extractor"),
model_type=ModelType.CLIP_IMAGE_PROCESSOR,
),
self.load_metrics_collector,
)
self.unet = load_model(
ModelLoadConfig(
model_path=os.path.join(root_model_path, "unet"),
model_type=ModelType.ONNX_MODEL,
execution_providers=self.execution_providers,
custom_op_path=custom_op_path,
enable_compile=enable_compile,
disable_dd_env="DISABLE_UNET_DD",
),
self.load_metrics_collector,
)
self.controlnet = load_model(
ModelLoadConfig(
model_path=os.path.join(root_model_path, "controlnet"),
model_type=ModelType.ONNX_MODEL,
execution_providers=["DmlExecutionProvider", "RyzenAILightExecutionProvider", "CPUExecutionProvider"],
custom_op_path=custom_op_path,
enable_compile=enable_compile,
disable_dd_env="DISABLE_CONTROLNET_DD",
),
self.load_metrics_collector,
)
self.vae_decoder = load_model(
ModelLoadConfig(
model_path=os.path.join(root_model_path, "vae_decoder"),
model_type=ModelType.ONNX_MODEL,
execution_providers=self.execution_providers,
custom_op_path=custom_op_path,
enable_compile=enable_compile,
disable_dd_env="DISABLE_VAE_DD",
),
self.load_metrics_collector,
)
def __enter__(self):
Logger.info(
"Initializing resources for StableDiffusionControlnetONNXPipelineTrigger."
)
return self
def __exit__(self, exc_type, exc_val, exc_tb):
del self.unet.model
del self.controlnet.model
del self.vae_decoder.model
import gc
gc.collect()
# bugcase 48820: must run after every ORT session using the
# dynamically-registered RyzenAI EP has been released and collected
# above -- otherwise process-exit EP/driver unload ordering races
# the AMD graphics UMD and crashes with 0xC0000409/0xC0000005 even
# though all sessions were already released cleanly.
common.unregister_ep()
Logger.debug(
"Models in StableDiffusionControlnetONNXPipelineTrigger are released"
)
@staticmethod
def _load_canny_image(image_path, low_threshold=100, high_threshold=200):
"""Moved from runner — canny preprocessing belongs with the controlnet trigger."""
try:
image = np.array(load_image(image_path))
image = canny_edges(image, low_threshold, high_threshold)
image = np.concatenate([image[:, :, None]] * 3, axis=2)
return Image.fromarray(image)
except Exception as exc:
Logger.warning("Failed to load control image: %s", exc)
return None
# Accept args + prompt + shape — eliminates runner get_run_kwargs boilerplate
def run(self, args, prompt, shape):
height, width = shape.height, shape.width
n_prompt = args.n_prompt
num_inference_steps = args.num_inference_steps
num_images_per_prompt = args.num_images_per_prompt
guidance_scale = args.guidance_scale
seed = args.seed
control_image = self._load_canny_image(args.control_image_path)
controlnet_conditioning_scale = args.controlnet_conditioning_scale
pipe = StableDiffusionControlNetONNXPipelineAMD(
unet_onnx=self.unet,
controlnet_onnx=self.controlnet,
vae_decoder_onnx=self.vae_decoder,
tokenizer=self.tokenizer,
text_encoder=self.text_encoder,
scheduler=self.scheduler,
feature_extractor=self.feature_extractor,
)
config_dict = {
"height": height,
"width": width,
"prompt": prompt,
"n_prompt": n_prompt,
"num_inference_steps": num_inference_steps,
"num_images_per_prompt": num_images_per_prompt,
"guidance_scale": guidance_scale,
"control_image": control_image,
"controlnet_conditioning_scale": controlnet_conditioning_scale,
"seed": seed,
}
common.print_config(config_dict)
def run_fn():
g = (
torch.Generator().manual_seed(seed)
if seed is not None
else torch.Generator()
)
return pipe.run(
prompt=prompt,
negative_prompt=n_prompt,
guidance_scale=guidance_scale,
controlnet_conditioning_scale=controlnet_conditioning_scale,
num_inference_steps=num_inference_steps,
generator=g,
image=control_image,
height=height,
width=width,
guess_mode=False,
return_dict=False,
device=torch.device("cpu"),
)
return super().run(pipe, run_fn, height, width, enable_profile=self.enable_profile)