Upload sayak_lcm_benchmark.py
Browse files- sayak_lcm_benchmark.py +54 -0
sayak_lcm_benchmark.py
ADDED
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.utils.benchmark as benchmark
|
3 |
+
import argparse
|
4 |
+
from diffusers import DiffusionPipeline, LCMScheduler
|
5 |
+
|
6 |
+
PROMPT = "close-up photography of old man standing in the rain at night, in a street lit by lamps, leica 35mm summilux"
|
7 |
+
MODEL_ID = "stabilityai/stable-diffusion-xl-base-1.0"
|
8 |
+
LORA_ID = "latent-consistency/lcm-lora-sdxl"
|
9 |
+
|
10 |
+
|
11 |
+
def benchmark_fn(f, *args, **kwargs):
|
12 |
+
t0 = benchmark.Timer(
|
13 |
+
stmt="f(*args, **kwargs)", globals={"args": args, "kwargs": kwargs, "f": f}
|
14 |
+
)
|
15 |
+
return t0.blocked_autorange().mean * 1e6
|
16 |
+
|
17 |
+
|
18 |
+
def load_pipeline(standard_sdxl=False):
|
19 |
+
pipe = DiffusionPipeline.from_pretrained(MODEL_ID, variant="fp16")
|
20 |
+
if not standard_sdxl:
|
21 |
+
pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
|
22 |
+
pipe.load_lora_weights(LORA_ID)
|
23 |
+
|
24 |
+
pipe.to(device="cuda", dtype=torch.float16)
|
25 |
+
pipe.enable_model_cpu_offload()
|
26 |
+
return pipe
|
27 |
+
|
28 |
+
|
29 |
+
def call_pipeline(pipe, batch_size, num_inference_steps, guidance_scale):
|
30 |
+
images = pipe(
|
31 |
+
prompt=PROMPT,
|
32 |
+
num_inference_steps=num_inference_steps,
|
33 |
+
num_images_per_prompt=batch_size,
|
34 |
+
guidance_scale=guidance_scale,
|
35 |
+
).images[0]
|
36 |
+
|
37 |
+
|
38 |
+
if __name__ == "__main__":
|
39 |
+
parser = argparse.ArgumentParser()
|
40 |
+
parser.add_argument("--batch_size", type=int, default=1)
|
41 |
+
parser.add_argument("--standard_sdxl", action="store_true")
|
42 |
+
args = parser.parse_args()
|
43 |
+
|
44 |
+
pipeline = load_pipeline(args.standard_sdxl)
|
45 |
+
if args.standard_sdxl:
|
46 |
+
num_inference_steps = 25
|
47 |
+
guidance_scale = 5
|
48 |
+
else:
|
49 |
+
num_inference_steps = 4
|
50 |
+
guidance_scale = 1
|
51 |
+
|
52 |
+
time = benchmark_fn(call_pipeline, pipeline, args.batch_size, num_inference_steps, guidance_scale)
|
53 |
+
|
54 |
+
print(f"Batch size: {args.batch_size} in {time/1e6:.3f} seconds")
|