File size: 12,929 Bytes
610d9c2 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 | # -------------------------------------------------------------------------
# Copyright (c) Microsoft Corporation. All rights reserved.
# Licensed under the MIT License.
# --------------------------------------------------------------------------
import gc
import importlib.util
import time
from statistics import mean
import torch
from demo_utils import PipelineInfo
from diffusers import (
AutoencoderKL,
ControlNetModel,
DiffusionPipeline,
EulerAncestralDiscreteScheduler,
StableDiffusionXLControlNetPipeline,
)
from engine_builder import EngineType, get_engine_paths
from pipeline_stable_diffusion import StableDiffusionPipeline
"""
Benchmark script for SDXL-Turbo with control net for engines like PyTorch or Stable Fast.
Setup for Stable Fast (see https://github.com/chengzeyi/stable-fast/blob/main/README.md for more info):
git clone https://github.com/chengzeyi/stable-fast.git
cd stable-fast
git submodule update --init
pip3 install torch torchvision torchaudio ninja
pip3 install -e '.[dev,xformers,triton,transformers,diffusers]' -v
sudo apt install libgoogle-perftools-dev
export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libtcmalloc.so
"""
def get_canny_image():
import cv2 # noqa: PLC0415
import numpy as np # noqa: PLC0415
from PIL import Image # noqa: PLC0415
# Test Image can be downloaded from https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png
image = Image.open("input_image_vermeer.png").convert("RGB")
image = np.array(image)
image = cv2.Canny(image, 100, 200)
image = image[:, :, None]
image = np.concatenate([image, image, image], axis=2)
return Image.fromarray(image)
def compile_stable_fast(pipeline, enable_cuda_graph=True):
from sfast.compilers.stable_diffusion_pipeline_compiler import CompilationConfig, compile # noqa: PLC0415
config = CompilationConfig.Default()
if importlib.util.find_spec("xformers") is not None:
config.enable_xformers = True
if importlib.util.find_spec("triton") is not None:
config.enable_triton = True
config.enable_cuda_graph = enable_cuda_graph
pipeline = compile(pipeline, config)
return pipeline
def compile_torch(pipeline, use_nhwc=False):
if use_nhwc:
pipeline.unet.to(memory_format=torch.channels_last)
pipeline.unet = torch.compile(pipeline.unet, mode="reduce-overhead", fullgraph=True)
if hasattr(pipeline, "controlnet"):
if use_nhwc:
pipeline.controlnet.to(memory_format=torch.channels_last)
pipeline.controlnet = torch.compile(pipeline.controlnet, mode="reduce-overhead", fullgraph=True)
return pipeline
def load_pipeline(name, engine, use_control_net=False, use_nhwc=False, enable_cuda_graph=True):
gc.collect()
torch.cuda.empty_cache()
before_memory = torch.cuda.memory_allocated()
scheduler = EulerAncestralDiscreteScheduler.from_pretrained(name, subfolder="scheduler")
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16).to("cuda")
if use_control_net:
assert "xl" in name
controlnet = ControlNetModel.from_pretrained("diffusers/controlnet-canny-sdxl-1.0", torch_dtype=torch.float16)
pipeline = StableDiffusionXLControlNetPipeline.from_pretrained(
name,
controlnet=controlnet,
vae=vae,
scheduler=scheduler,
variant="fp16",
use_safetensors=True,
torch_dtype=torch.float16,
).to("cuda")
else:
pipeline = DiffusionPipeline.from_pretrained(
name,
vae=vae,
scheduler=scheduler,
variant="fp16",
use_safetensors=True,
torch_dtype=torch.float16,
).to("cuda")
pipeline.safety_checker = None
gc.collect()
after_memory = torch.cuda.memory_allocated()
print(f"Loaded model with {after_memory - before_memory} bytes allocated")
if engine == "stable_fast":
pipeline = compile_stable_fast(pipeline, enable_cuda_graph=enable_cuda_graph)
elif engine == "torch":
pipeline = compile_torch(pipeline, use_nhwc=use_nhwc)
pipeline.set_progress_bar_config(disable=True)
return pipeline
def get_prompt():
return "little cute gremlin wearing a jacket, cinematic, vivid colors, intricate masterpiece, golden ratio, highly detailed"
def load_ort_cuda_pipeline(name, engine, use_control_net=False, enable_cuda_graph=True, work_dir="."):
version = PipelineInfo.supported_models()[name]
guidance_scale = 0.0
pipeline_info = PipelineInfo(
version,
use_vae=True,
use_fp16_vae=True,
do_classifier_free_guidance=(guidance_scale > 1.0),
controlnet=["canny"] if use_control_net else [],
)
engine_type = EngineType.ORT_CUDA if engine == "ort_cuda" else EngineType.ORT_TRT
onnx_dir, engine_dir, output_dir, framework_model_dir, _ = get_engine_paths(
work_dir=work_dir, pipeline_info=pipeline_info, engine_type=engine_type
)
pipeline = StableDiffusionPipeline(
pipeline_info,
scheduler="EulerA",
max_batch_size=32,
use_cuda_graph=enable_cuda_graph,
framework_model_dir=framework_model_dir,
output_dir=output_dir,
engine_type=engine_type,
)
pipeline.backend.build_engines(
engine_dir=engine_dir,
framework_model_dir=framework_model_dir,
onnx_dir=onnx_dir,
device_id=torch.cuda.current_device(),
)
return pipeline
def test_ort_cuda(
pipeline,
batch_size=1,
steps=4,
control_image=None,
warmup_runs=3,
test_runs=10,
seed=123,
verbose=False,
image_height=512,
image_width=512,
):
if batch_size > 4 and pipeline.pipeline_info.version == "xl-1.0":
pipeline.backend.enable_vae_slicing()
pipeline.load_resources(image_height, image_width, batch_size)
warmup_prompt = "warm up"
for _ in range(warmup_runs):
images, _ = pipeline.run(
[warmup_prompt] * batch_size,
[""] * batch_size,
image_height=image_height,
image_width=image_width,
denoising_steps=steps,
guidance=0.0,
seed=seed,
controlnet_images=[control_image],
controlnet_scales=torch.FloatTensor([0.5]),
output_type="image",
)
assert len(images) == batch_size
generator = torch.Generator(device="cuda")
generator.manual_seed(seed)
prompt = get_prompt()
latency_list = []
images = None
for _ in range(test_runs):
torch.cuda.synchronize()
start_time = time.perf_counter()
images, _ = pipeline.run(
[prompt] * batch_size,
[""] * batch_size,
image_height=image_height,
image_width=image_width,
denoising_steps=steps,
guidance=0.0,
seed=seed,
controlnet_images=[control_image],
controlnet_scales=torch.FloatTensor([0.5]),
output_type="pil",
)
torch.cuda.synchronize()
seconds = time.perf_counter() - start_time
latency_list.append(seconds)
if verbose:
print(latency_list)
return images, latency_list
def test(pipeline, batch_size=1, steps=4, control_image=None, warmup_runs=3, test_runs=10, seed=123, verbose=False):
control_net_args = {}
if hasattr(pipeline, "controlnet"):
control_net_args = {
"image": control_image,
"controlnet_conditioning_scale": 0.5,
}
warmup_prompt = "warm up"
for _ in range(warmup_runs):
images = pipeline(
prompt=warmup_prompt,
num_inference_steps=steps,
num_images_per_prompt=batch_size,
guidance_scale=0.0,
**control_net_args,
).images
assert len(images) == batch_size
generator = torch.Generator(device="cuda")
generator.manual_seed(seed)
prompt = get_prompt()
latency_list = []
images = None
for _ in range(test_runs):
torch.cuda.synchronize()
start_time = time.perf_counter()
images = pipeline(
prompt=prompt,
num_inference_steps=steps,
num_images_per_prompt=batch_size,
guidance_scale=0.0,
generator=generator,
**control_net_args,
).images
torch.cuda.synchronize()
seconds = time.perf_counter() - start_time
latency_list.append(seconds)
if verbose:
print(latency_list)
return images, latency_list
def arguments():
import argparse # noqa: PLC0415
parser = argparse.ArgumentParser(description="Benchmark Stable Diffusion pipeline (optional control net for SDXL)")
parser.add_argument(
"--engine",
type=str,
default="torch",
choices=["torch", "stable_fast", "ort_cuda", "ort_trt"],
help="Backend engine: torch, stable_fast or ort_cuda",
)
parser.add_argument(
"--name",
type=str,
choices=list(PipelineInfo.supported_models().keys()),
default="stabilityai/sdxl-turbo",
help="Stable diffusion model name. Default is stabilityai/sdxl-turbo",
)
parser.add_argument(
"--work-dir",
type=str,
default=".",
help="working directory for ort_cuda or ort_trt",
)
parser.add_argument(
"--use_control_net",
action="store_true",
help="Use control net diffusers/controlnet-canny-sdxl-1.0",
)
parser.add_argument(
"--batch_size",
type=int,
default=1,
help="Batch size",
)
parser.add_argument(
"--steps",
type=int,
default=1,
help="Denoising steps",
)
parser.add_argument(
"--warmup_runs",
type=int,
default=3,
help="Number of warmup runs before measurement",
)
parser.add_argument(
"--use_nhwc",
action="store_true",
help="use channel last format for torch compile",
)
parser.add_argument(
"--enable_cuda_graph",
action="store_true",
help="enable cuda graph for stable fast",
)
parser.add_argument(
"--verbose",
action="store_true",
help="print more information",
)
args = parser.parse_args()
return args
def main():
args = arguments()
with torch.no_grad():
if args.engine == "ort_cuda":
pipeline = load_ort_cuda_pipeline(
args.name,
args.engine,
use_control_net=args.use_control_net,
enable_cuda_graph=args.enable_cuda_graph,
work_dir=args.work_dir,
)
else:
pipeline = load_pipeline(
args.name,
args.engine,
use_control_net=args.use_control_net,
use_nhwc=args.use_nhwc,
enable_cuda_graph=args.enable_cuda_graph,
)
canny_image = get_canny_image()
if args.engine == "ort_cuda":
images, latency_list = test_ort_cuda(
pipeline,
args.batch_size,
args.steps,
control_image=canny_image,
warmup_runs=args.warmup_runs,
verbose=args.verbose,
)
elif args.engine == "stable_fast":
from sfast.utils.compute_precision import low_compute_precision # noqa: PLC0415
with low_compute_precision():
images, latency_list = test(
pipeline,
args.batch_size,
args.steps,
control_image=canny_image,
warmup_runs=args.warmup_runs,
verbose=args.verbose,
)
else:
images, latency_list = test(
pipeline,
args.batch_size,
args.steps,
control_image=canny_image,
warmup_runs=args.warmup_runs,
verbose=args.verbose,
)
# Save the first output image to inspect the result.
if images:
images[0].save(
f"{args.engine}_{args.name.replace('/', '_')}_{args.batch_size}_{args.steps}_c{int(args.use_control_net)}.png"
)
result = {
"engine": args.engine,
"batch_size": args.batch_size,
"steps": args.steps,
"control_net": args.use_control_net,
"nhwc": args.use_nhwc,
"enable_cuda_graph": args.enable_cuda_graph,
"average_latency_in_ms": mean(latency_list) * 1000,
}
print(result)
main()
|