Spaces:
Running
on
Zero
Running
on
Zero
adamelliotfields
commited on
Commit
•
af35186
1
Parent(s):
80551a9
Add timer context manager
Browse files- lib/__init__.py +2 -0
- lib/inference.py +4 -2
- lib/loader.py +38 -35
- lib/utils.py +13 -0
lib/__init__.py
CHANGED
@@ -10,6 +10,7 @@ from .utils import (
|
|
10 |
enable_progress_bars,
|
11 |
load_json,
|
12 |
read_file,
|
|
|
13 |
)
|
14 |
|
15 |
__all__ = [
|
@@ -24,4 +25,5 @@ __all__ = [
|
|
24 |
"generate",
|
25 |
"load_json",
|
26 |
"read_file",
|
|
|
27 |
]
|
|
|
10 |
enable_progress_bars,
|
11 |
load_json,
|
12 |
read_file,
|
13 |
+
timer,
|
14 |
)
|
15 |
|
16 |
__all__ = [
|
|
|
25 |
"generate",
|
26 |
"load_json",
|
27 |
"read_file",
|
28 |
+
"timer",
|
29 |
]
|
lib/inference.py
CHANGED
@@ -251,7 +251,9 @@ def generate(
|
|
251 |
loader.collect()
|
252 |
gc.collect()
|
253 |
|
254 |
-
|
|
|
|
|
255 |
if Info:
|
256 |
-
Info(
|
257 |
return images
|
|
|
251 |
loader.collect()
|
252 |
gc.collect()
|
253 |
|
254 |
+
end = time.perf_counter()
|
255 |
+
msg = f"Generated {len(images)} image{'s' if len(images) > 1 else ''} in {end - start:.2f}s"
|
256 |
+
print(msg)
|
257 |
if Info:
|
258 |
+
Info(msg)
|
259 |
return images
|
lib/loader.py
CHANGED
@@ -7,6 +7,7 @@ from diffusers.models import AutoencoderKL
|
|
7 |
|
8 |
from .config import Config
|
9 |
from .upscaler import RealESRGAN
|
|
|
10 |
|
11 |
|
12 |
class Loader:
|
@@ -23,13 +24,15 @@ class Loader:
|
|
23 |
cls._instance.upscaler = None
|
24 |
return cls._instance
|
25 |
|
26 |
-
|
|
|
27 |
if self.refiner is None:
|
28 |
return False
|
29 |
if self.model and self.model.lower() != model.lower():
|
30 |
return True
|
31 |
return False
|
32 |
|
|
|
33 |
def _should_unload_refiner(self, refiner=False):
|
34 |
if self.refiner is None:
|
35 |
return False
|
@@ -57,44 +60,45 @@ class Loader:
|
|
57 |
return True
|
58 |
return False
|
59 |
|
60 |
-
def
|
61 |
if self.refiner is not None:
|
62 |
-
self.refiner.to("cpu", silence_dtype_warnings=True)
|
63 |
self.refiner.vae = None
|
64 |
self.refiner.scheduler = None
|
65 |
self.refiner.tokenizer_2 = None
|
66 |
self.refiner.text_encoder_2 = None
|
67 |
|
68 |
def _unload_refiner(self):
|
69 |
-
|
70 |
-
|
|
|
71 |
|
72 |
def _unload_upscaler(self):
|
73 |
-
|
74 |
-
|
|
|
75 |
|
76 |
def _unload_deepcache(self):
|
77 |
if self.pipe.deepcache is not None:
|
78 |
-
print("
|
79 |
self.pipe.deepcache.disable()
|
80 |
delattr(self.pipe, "deepcache")
|
81 |
if self.refiner is not None:
|
82 |
if hasattr(self.refiner, "deepcache"):
|
83 |
-
print("Unloading DeepCache for refiner")
|
84 |
self.refiner.deepcache.disable()
|
85 |
delattr(self.refiner, "deepcache")
|
86 |
|
87 |
def _unload_pipeline(self):
|
88 |
-
|
89 |
-
|
|
|
90 |
|
91 |
def _unload(self, model, refiner, deepcache, scale):
|
92 |
to_unload = []
|
93 |
if self._should_unload_deepcache(deepcache): # remove deepcache first
|
94 |
self._unload_deepcache()
|
95 |
|
96 |
-
if self.
|
97 |
-
self.
|
98 |
|
99 |
if self._should_unload_refiner(refiner):
|
100 |
self._unload_refiner()
|
@@ -119,8 +123,8 @@ class Loader:
|
|
119 |
model = Config.REFINER_MODEL
|
120 |
pipeline = Config.PIPELINES["img2img"]
|
121 |
try:
|
122 |
-
|
123 |
-
|
124 |
except Exception as e:
|
125 |
print(f"Error loading {model}: {e}")
|
126 |
self.refiner = None
|
@@ -131,9 +135,9 @@ class Loader:
|
|
131 |
def _load_upscaler(self, scale=1):
|
132 |
if self.upscaler is None and scale > 1:
|
133 |
try:
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
except Exception as e:
|
138 |
print(f"Error loading {scale}x upscaler: {e}")
|
139 |
self.upscaler = None
|
@@ -144,7 +148,7 @@ class Loader:
|
|
144 |
return
|
145 |
if pipe_has_deepcache and self.pipe.deepcache.params["cache_interval"] == interval:
|
146 |
return
|
147 |
-
print("
|
148 |
self.pipe.deepcache = DeepCacheSDHelper(pipe=self.pipe)
|
149 |
self.pipe.deepcache.set_params(cache_interval=interval)
|
150 |
self.pipe.deepcache.enable()
|
@@ -155,7 +159,6 @@ class Loader:
|
|
155 |
return
|
156 |
if refiner_has_deepcache and self.refiner.deepcache.params["cache_interval"] == interval:
|
157 |
return
|
158 |
-
print("Loading DeepCache for refiner")
|
159 |
self.refiner.deepcache = DeepCacheSDHelper(pipe=self.refiner)
|
160 |
self.refiner.deepcache.set_params(cache_interval=interval)
|
161 |
self.refiner.deepcache.enable()
|
@@ -164,21 +167,21 @@ class Loader:
|
|
164 |
pipeline = Config.PIPELINES[kind]
|
165 |
if self.pipe is None:
|
166 |
try:
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
except Exception as e:
|
183 |
print(f"Error loading {model}: {e}")
|
184 |
self.model = None
|
|
|
7 |
|
8 |
from .config import Config
|
9 |
from .upscaler import RealESRGAN
|
10 |
+
from .utils import timer
|
11 |
|
12 |
|
13 |
class Loader:
|
|
|
24 |
cls._instance.upscaler = None
|
25 |
return cls._instance
|
26 |
|
27 |
+
# switching models
|
28 |
+
def _should_reset_refiner(self, model=""):
|
29 |
if self.refiner is None:
|
30 |
return False
|
31 |
if self.model and self.model.lower() != model.lower():
|
32 |
return True
|
33 |
return False
|
34 |
|
35 |
+
# switching refiner
|
36 |
def _should_unload_refiner(self, refiner=False):
|
37 |
if self.refiner is None:
|
38 |
return False
|
|
|
60 |
return True
|
61 |
return False
|
62 |
|
63 |
+
def _reset_refiner(self):
|
64 |
if self.refiner is not None:
|
|
|
65 |
self.refiner.vae = None
|
66 |
self.refiner.scheduler = None
|
67 |
self.refiner.tokenizer_2 = None
|
68 |
self.refiner.text_encoder_2 = None
|
69 |
|
70 |
def _unload_refiner(self):
|
71 |
+
if self.refiner is not None:
|
72 |
+
with timer("Unloading refiner"):
|
73 |
+
self.refiner.to("cpu", silence_dtype_warnings=True)
|
74 |
|
75 |
def _unload_upscaler(self):
|
76 |
+
if self.upscaler is not None:
|
77 |
+
with timer(f"Unloading {self.upscaler.scale}x upscaler"):
|
78 |
+
self.upscaler.to("cpu")
|
79 |
|
80 |
def _unload_deepcache(self):
|
81 |
if self.pipe.deepcache is not None:
|
82 |
+
print("Disabling DeepCache")
|
83 |
self.pipe.deepcache.disable()
|
84 |
delattr(self.pipe, "deepcache")
|
85 |
if self.refiner is not None:
|
86 |
if hasattr(self.refiner, "deepcache"):
|
|
|
87 |
self.refiner.deepcache.disable()
|
88 |
delattr(self.refiner, "deepcache")
|
89 |
|
90 |
def _unload_pipeline(self):
|
91 |
+
if self.pipe is not None:
|
92 |
+
with timer(f"Unloading {self.model}"):
|
93 |
+
self.pipe.to("cpu", silence_dtype_warnings=True)
|
94 |
|
95 |
def _unload(self, model, refiner, deepcache, scale):
|
96 |
to_unload = []
|
97 |
if self._should_unload_deepcache(deepcache): # remove deepcache first
|
98 |
self._unload_deepcache()
|
99 |
|
100 |
+
if self._should_reset_refiner(model):
|
101 |
+
self._reset_refiner()
|
102 |
|
103 |
if self._should_unload_refiner(refiner):
|
104 |
self._unload_refiner()
|
|
|
123 |
model = Config.REFINER_MODEL
|
124 |
pipeline = Config.PIPELINES["img2img"]
|
125 |
try:
|
126 |
+
with timer(f"Loading {model}"):
|
127 |
+
self.refiner = pipeline.from_pretrained(model, **kwargs).to("cuda")
|
128 |
except Exception as e:
|
129 |
print(f"Error loading {model}: {e}")
|
130 |
self.refiner = None
|
|
|
135 |
def _load_upscaler(self, scale=1):
|
136 |
if self.upscaler is None and scale > 1:
|
137 |
try:
|
138 |
+
with timer(f"Loading {scale}x upscaler"):
|
139 |
+
self.upscaler = RealESRGAN(scale, device=self.pipe.device)
|
140 |
+
self.upscaler.load_weights()
|
141 |
except Exception as e:
|
142 |
print(f"Error loading {scale}x upscaler: {e}")
|
143 |
self.upscaler = None
|
|
|
148 |
return
|
149 |
if pipe_has_deepcache and self.pipe.deepcache.params["cache_interval"] == interval:
|
150 |
return
|
151 |
+
print("Enabling DeepCache")
|
152 |
self.pipe.deepcache = DeepCacheSDHelper(pipe=self.pipe)
|
153 |
self.pipe.deepcache.set_params(cache_interval=interval)
|
154 |
self.pipe.deepcache.enable()
|
|
|
159 |
return
|
160 |
if refiner_has_deepcache and self.refiner.deepcache.params["cache_interval"] == interval:
|
161 |
return
|
|
|
162 |
self.refiner.deepcache = DeepCacheSDHelper(pipe=self.refiner)
|
163 |
self.refiner.deepcache.set_params(cache_interval=interval)
|
164 |
self.refiner.deepcache.enable()
|
|
|
167 |
pipeline = Config.PIPELINES[kind]
|
168 |
if self.pipe is None:
|
169 |
try:
|
170 |
+
with timer(f"Loading {model}"):
|
171 |
+
self.model = model
|
172 |
+
if model.lower() in Config.MODEL_CHECKPOINTS.keys():
|
173 |
+
self.pipe = pipeline.from_single_file(
|
174 |
+
f"https://huggingface.co/{model}/{Config.MODEL_CHECKPOINTS[model.lower()]}",
|
175 |
+
**kwargs,
|
176 |
+
).to("cuda")
|
177 |
+
else:
|
178 |
+
self.pipe = pipeline.from_pretrained(model, **kwargs).to("cuda")
|
179 |
+
if self.refiner is not None:
|
180 |
+
self.refiner.vae = self.pipe.vae
|
181 |
+
self.refiner.scheduler = self.pipe.scheduler
|
182 |
+
self.refiner.tokenizer_2 = self.pipe.tokenizer_2
|
183 |
+
self.refiner.text_encoder_2 = self.pipe.text_encoder_2
|
184 |
+
self.refiner.to(self.pipe.device)
|
185 |
except Exception as e:
|
186 |
print(f"Error loading {model}: {e}")
|
187 |
self.model = None
|
lib/utils.py
CHANGED
@@ -2,6 +2,8 @@ import functools
|
|
2 |
import inspect
|
3 |
import json
|
4 |
import os
|
|
|
|
|
5 |
from typing import Callable, TypeVar
|
6 |
|
7 |
import anyio
|
@@ -20,6 +22,17 @@ MAX_CONCURRENT_THREADS = 1
|
|
20 |
MAX_THREADS_GUARD = Semaphore(MAX_CONCURRENT_THREADS)
|
21 |
|
22 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
@functools.lru_cache()
|
24 |
def load_json(path: str) -> dict:
|
25 |
with open(path, "r", encoding="utf-8") as file:
|
|
|
2 |
import inspect
|
3 |
import json
|
4 |
import os
|
5 |
+
import time
|
6 |
+
from contextlib import contextmanager
|
7 |
from typing import Callable, TypeVar
|
8 |
|
9 |
import anyio
|
|
|
22 |
MAX_THREADS_GUARD = Semaphore(MAX_CONCURRENT_THREADS)
|
23 |
|
24 |
|
25 |
+
@contextmanager
|
26 |
+
def timer(message="Operation", logger=print):
|
27 |
+
start = time.perf_counter()
|
28 |
+
logger(message)
|
29 |
+
try:
|
30 |
+
yield
|
31 |
+
finally:
|
32 |
+
end = time.perf_counter()
|
33 |
+
logger(f"{message} took {end - start:.2f}s")
|
34 |
+
|
35 |
+
|
36 |
@functools.lru_cache()
|
37 |
def load_json(path: str) -> dict:
|
38 |
with open(path, "r", encoding="utf-8") as file:
|