MarkZakelj
commited on
Commit
·
dec981f
1
Parent(s):
341de7e
style lora fusion with character lora
Browse files- sequential_timer.py +25 -0
- serve_loras.py +123 -26
sequential_timer.py
ADDED
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from time import perf_counter
|
2 |
+
|
3 |
+
class SequentialTimer:
|
4 |
+
def __init__(self, make_print=False):
|
5 |
+
self.timings = []
|
6 |
+
self.make_print = make_print
|
7 |
+
|
8 |
+
def time(self, message: str):
|
9 |
+
if self.make_print:
|
10 |
+
print(message)
|
11 |
+
self.timings.append((perf_counter(), message))
|
12 |
+
|
13 |
+
def to_str(self) -> str:
|
14 |
+
s = ""
|
15 |
+
if len(self.timings) <= 1:
|
16 |
+
s = "No timings"
|
17 |
+
return s
|
18 |
+
t0 = self.timings[0][0]
|
19 |
+
for ((t1, m1), (t2, _)) in zip(self.timings, self.timings[1:]):
|
20 |
+
s += f"TIME: step: {t2 - t1:06.3f} | cum {t2 - t0:06.3f} - {m1}\n"
|
21 |
+
s += f"ALL TIME: {self.timings[-1][0] - self.timings[0][0]:07.3f}\n"
|
22 |
+
return s
|
23 |
+
|
24 |
+
def printall(self):
|
25 |
+
print(self.to_str())
|
serve_loras.py
CHANGED
@@ -5,7 +5,7 @@ import uuid
|
|
5 |
|
6 |
import diffusers
|
7 |
import torch
|
8 |
-
from diffusers import StableDiffusionXLPipeline
|
9 |
|
10 |
import numpy as np
|
11 |
import threading
|
@@ -14,13 +14,15 @@ import base64
|
|
14 |
from io import BytesIO
|
15 |
from PIL import Image
|
16 |
import numpy as np
|
17 |
-
import uuid
|
18 |
from tempfile import TemporaryFile
|
19 |
from google.cloud import storage
|
20 |
import sys
|
21 |
import sentry_sdk
|
22 |
from flask import Flask, request, jsonify
|
23 |
import os
|
|
|
|
|
|
|
24 |
|
25 |
logger = logging.getLogger(__name__)
|
26 |
logger.info("Diffusers version %s", diffusers.__version__)
|
@@ -34,6 +36,24 @@ sentry_sdk.init(
|
|
34 |
|
35 |
LORAS_DIR = './safetensors'
|
36 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
class DiffusersHandler(ABC):
|
38 |
"""
|
39 |
Diffusers handler class for text to image generation.
|
@@ -65,8 +85,31 @@ class DiffusersHandler(ABC):
|
|
65 |
torch_dtype=torch.float16,
|
66 |
use_safetensors=True,
|
67 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
68 |
|
69 |
-
logger.info("moving model to device: %s", device_str)
|
70 |
self.pipe.to(self.device)
|
71 |
|
72 |
logger.info(self.device)
|
@@ -86,20 +129,30 @@ class DiffusersHandler(ABC):
|
|
86 |
logger.info("Received requests: '%s'", raw_requests)
|
87 |
self.working = True
|
88 |
|
89 |
-
|
90 |
"prompt": raw_requests[0]["prompt"],
|
91 |
"negative_prompt": raw_requests[0].get("negative_prompt"),
|
92 |
"width": raw_requests[0].get("width"),
|
93 |
"height": raw_requests[0].get("height"),
|
94 |
-
"num_inference_steps": raw_requests[0].get("num_inference_steps",
|
95 |
-
"guidance_scale": raw_requests[0].get("guidance_scale",
|
96 |
-
"lora_weights": raw_requests[0].get("lora_name", None)
|
97 |
-
"cross_attention_kwargs": {"scale": raw_requests[0].get("lora_scale", 0.
|
98 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
99 |
|
100 |
-
logger.info("Processed request: '%s'",
|
101 |
-
axiom_logger.info("Processed request:" + str(
|
102 |
-
return
|
103 |
|
104 |
|
105 |
def inference(self, request):
|
@@ -111,29 +164,70 @@ class DiffusersHandler(ABC):
|
|
111 |
"""
|
112 |
|
113 |
# Handling inference for sequence_classification.
|
114 |
-
compel = Compel(tokenizer=[self.pipe.tokenizer, self.pipe.tokenizer_2] , text_encoder=[self.pipe.text_encoder, self.pipe.text_encoder_2], returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED, requires_pooled=[False, True])
|
|
|
|
|
|
|
|
|
|
|
|
|
115 |
|
116 |
-
|
117 |
-
|
|
|
|
|
118 |
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
124 |
|
125 |
# Handling inference for sequence_classification.
|
|
|
126 |
inferences = self.pipe(
|
127 |
prompt_embeds=conditioning,
|
128 |
pooled_prompt_embeds=pooled,
|
129 |
-
|
|
|
|
|
130 |
).images
|
131 |
|
132 |
-
if lora_weights is not None:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
133 |
self.pipe.unload_lora_weights()
|
134 |
|
135 |
-
|
136 |
-
|
|
|
|
|
137 |
return inferences
|
138 |
|
139 |
def postprocess(self, inference_outputs):
|
@@ -178,16 +272,19 @@ handlers = [DiffusersHandler() for i in range(gpu_count)]
|
|
178 |
for i in range(gpu_count):
|
179 |
handlers[i].initialize({"gpu_id": i})
|
180 |
|
181 |
-
|
182 |
-
|
|
|
183 |
|
184 |
@app.route('/generate', methods=['POST'])
|
185 |
def generate_image():
|
186 |
req_id = str(uuid.uuid4())
|
187 |
global handler_index
|
|
|
188 |
try:
|
189 |
# Extract raw requests from HTTP POST body
|
190 |
raw_requests = request.json
|
|
|
191 |
|
192 |
with handler_lock:
|
193 |
selected_handler = handlers[handler_index]
|
@@ -202,7 +299,7 @@ def generate_image():
|
|
202 |
return jsonify({"image_urls": outputs})
|
203 |
except Exception as e:
|
204 |
logger.error("Error during image generation: %s", str(e))
|
205 |
-
axiom_logger.critical("Error during image generation: " + str(e), request_id=req_id)
|
206 |
return jsonify({"error": "Failed to generate image", "details": str(e)}), 500
|
207 |
|
208 |
if __name__ == '__main__':
|
|
|
5 |
|
6 |
import diffusers
|
7 |
import torch
|
8 |
+
from diffusers import StableDiffusionXLPipeline, DiffusionPipeline
|
9 |
|
10 |
import numpy as np
|
11 |
import threading
|
|
|
14 |
from io import BytesIO
|
15 |
from PIL import Image
|
16 |
import numpy as np
|
|
|
17 |
from tempfile import TemporaryFile
|
18 |
from google.cloud import storage
|
19 |
import sys
|
20 |
import sentry_sdk
|
21 |
from flask import Flask, request, jsonify
|
22 |
import os
|
23 |
+
from sequential_timer import SequentialTimer
|
24 |
+
from safetensors.torch import load_file
|
25 |
+
import copy
|
26 |
|
27 |
logger = logging.getLogger(__name__)
|
28 |
logger.info("Diffusers version %s", diffusers.__version__)
|
|
|
36 |
|
37 |
LORAS_DIR = './safetensors'
|
38 |
|
39 |
+
handler_lock = threading.Lock()
|
40 |
+
handler_index = 0
|
41 |
+
|
42 |
+
class LoraCache():
|
43 |
+
def __init__(self, loras_dir: str = LORAS_DIR):
|
44 |
+
self.loras_dir = loras_dir
|
45 |
+
self.cache = {}
|
46 |
+
|
47 |
+
def load_lora(self, lora_name: str):
|
48 |
+
if lora_name.endswith('.safetensors'):
|
49 |
+
lora_name = lora_name.rstrip('.safetensors')
|
50 |
+
if lora_name not in self.cache:
|
51 |
+
lora = load_file(os.path.join(self.loras_dir, lora_name+'.safetensors'))
|
52 |
+
self.cache[lora_name] = lora
|
53 |
+
return copy.deepcopy(self.cache[lora_name])
|
54 |
+
|
55 |
+
lora_cache = LoraCache()
|
56 |
+
|
57 |
class DiffusersHandler(ABC):
|
58 |
"""
|
59 |
Diffusers handler class for text to image generation.
|
|
|
85 |
torch_dtype=torch.float16,
|
86 |
use_safetensors=True,
|
87 |
)
|
88 |
+
# self.refiner = DiffusionPipeline.from_pretrained(
|
89 |
+
# "stabilityai/stable-diffusion-xl-refiner-1.0",
|
90 |
+
# text_encoder_2=self.pipe.text_encoder_2,
|
91 |
+
# vae=self.pipe.vae,
|
92 |
+
# torch_dtype=torch.float16,
|
93 |
+
# use_safetensors=True,
|
94 |
+
# variant="fp16",
|
95 |
+
# )
|
96 |
+
# self.refiner.enable_model_cpu_offload(properties.get("gpu_id"))
|
97 |
+
# logger.info("Refiner initialized and o")
|
98 |
+
|
99 |
+
self.compel_base = Compel(
|
100 |
+
tokenizer=[self.pipe.tokenizer, self.pipe.tokenizer_2],
|
101 |
+
text_encoder=[self.pipe.text_encoder, self.pipe.text_encoder_2],
|
102 |
+
returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
|
103 |
+
requires_pooled=[False, True])
|
104 |
+
logger.info("Compel initialized")
|
105 |
+
|
106 |
+
# self.compel_refiner = Compel(
|
107 |
+
# tokenizer=[self.refiner.tokenizer_2],
|
108 |
+
# text_encoder=[self.refiner.text_encoder_2],
|
109 |
+
# returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
|
110 |
+
# requires_pooled=[True])
|
111 |
|
112 |
+
logger.info("moving base model to device: %s", device_str)
|
113 |
self.pipe.to(self.device)
|
114 |
|
115 |
logger.info(self.device)
|
|
|
129 |
logger.info("Received requests: '%s'", raw_requests)
|
130 |
self.working = True
|
131 |
|
132 |
+
model_args = {
|
133 |
"prompt": raw_requests[0]["prompt"],
|
134 |
"negative_prompt": raw_requests[0].get("negative_prompt"),
|
135 |
"width": raw_requests[0].get("width"),
|
136 |
"height": raw_requests[0].get("height"),
|
137 |
+
"num_inference_steps": raw_requests[0].get("num_inference_steps", 25),
|
138 |
+
"guidance_scale": raw_requests[0].get("guidance_scale", 8.5)
|
139 |
+
# "lora_weights": raw_requests[0].get("lora_name", None)
|
140 |
+
# "cross_attention_kwargs": {"scale": raw_requests[0].get("lora_scale", 0.0)}
|
141 |
}
|
142 |
+
|
143 |
+
extra_args = {
|
144 |
+
"seed": raw_requests[0].get("seed", None),
|
145 |
+
"style_lora": raw_requests[0].get("style_lora", None),
|
146 |
+
"style_scale": raw_requests[0].get("style_scale", 1.0),
|
147 |
+
"char_lora": raw_requests[0].get("char_lora", None),
|
148 |
+
"char_scale": raw_requests[0].get("char_scale", 1.0)
|
149 |
+
}
|
150 |
+
|
151 |
+
|
152 |
|
153 |
+
logger.info("Processed request: '%s'", model_args)
|
154 |
+
axiom_logger.info("Processed request:" + str(model_args), request_id=self.req_id, device=self.device_str)
|
155 |
+
return model_args, extra_args
|
156 |
|
157 |
|
158 |
def inference(self, request):
|
|
|
164 |
"""
|
165 |
|
166 |
# Handling inference for sequence_classification.
|
167 |
+
# compel = Compel(tokenizer=[self.pipe.tokenizer, self.pipe.tokenizer_2] , text_encoder=[self.pipe.text_encoder, self.pipe.text_encoder_2], returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED, requires_pooled=[False, True])
|
168 |
+
st = SequentialTimer()
|
169 |
+
model_args, extra_args = request
|
170 |
+
|
171 |
+
use_char_lora = extra_args['char_lora'] is not None
|
172 |
+
use_style_lora = extra_args['style_lora'] is not None
|
173 |
+
|
174 |
|
175 |
+
style_lora = extra_args['style_lora']
|
176 |
+
char_lora = extra_args['char_lora']
|
177 |
+
|
178 |
+
cross_attention_kwargs = {"scale": extra_args['char_scale'] if use_char_lora else extra_args['style_scale']}
|
179 |
|
180 |
+
generator = torch.Generator(device="cuda").manual_seed(extra_args['seed']) if extra_args['seed'] else None
|
181 |
+
|
182 |
+
|
183 |
+
self.prompt = model_args.pop("prompt")
|
184 |
+
|
185 |
+
st.time("Base compel embedding")
|
186 |
+
conditioning, pooled = self.compel_base(self.prompt)
|
187 |
+
|
188 |
+
if use_style_lora:
|
189 |
+
style_lora = os.path.join(LORAS_DIR, style_lora + '.safetensors')
|
190 |
+
st.time("Load style lora")
|
191 |
+
self.pipe.load_lora_weights(style_lora)
|
192 |
+
if use_char_lora:
|
193 |
+
st.time("Fuse style lora into model")
|
194 |
+
self.pipe.fuse_lora(lora_scale=extra_args['style_scale'], fuse_text_encoder=False)
|
195 |
+
|
196 |
+
if use_char_lora:
|
197 |
+
char_lora = os.path.join(LORAS_DIR, char_lora + '.safetensors')
|
198 |
+
st.time('load character lora')
|
199 |
+
self.pipe.load_lora_weights(char_lora)
|
200 |
+
|
201 |
+
# lora_weights = model_args.pop("lora_weights")
|
202 |
+
# if lora_weights is not None:
|
203 |
+
# lora_path = os.path.join(LORAS_DIR, lora_weights + '.safetensors')
|
204 |
+
# logger.info('LOADING LORA FROM: ' + lora_path)
|
205 |
+
# self.pipe.load_lora_weights(lora_path)
|
206 |
|
207 |
# Handling inference for sequence_classification.
|
208 |
+
st.time("base model inference")
|
209 |
inferences = self.pipe(
|
210 |
prompt_embeds=conditioning,
|
211 |
pooled_prompt_embeds=pooled,
|
212 |
+
generator=generator,
|
213 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
214 |
+
**model_args
|
215 |
).images
|
216 |
|
217 |
+
# if lora_weights is not None:
|
218 |
+
# self.pipe.unload_lora_weights()
|
219 |
+
if use_style_lora and use_char_lora:
|
220 |
+
st.time("unfuse lora weights")
|
221 |
+
self.pipe.unfuse_lora(unfuse_text_encoder=False)
|
222 |
+
|
223 |
+
if use_style_lora or use_char_lora:
|
224 |
+
st.time("unload lora weights")
|
225 |
self.pipe.unload_lora_weights()
|
226 |
|
227 |
+
st.time('end')
|
228 |
+
|
229 |
+
# logger.info("Generated image: '%s'", inferences)
|
230 |
+
axiom_logger.info("Generated images", request_id=self.req_id, device=self.device_str, timings=st.to_str())
|
231 |
return inferences
|
232 |
|
233 |
def postprocess(self, inference_outputs):
|
|
|
272 |
for i in range(gpu_count):
|
273 |
handlers[i].initialize({"gpu_id": i})
|
274 |
|
275 |
+
|
276 |
+
|
277 |
+
|
278 |
|
279 |
@app.route('/generate', methods=['POST'])
|
280 |
def generate_image():
|
281 |
req_id = str(uuid.uuid4())
|
282 |
global handler_index
|
283 |
+
selected_handler = None
|
284 |
try:
|
285 |
# Extract raw requests from HTTP POST body
|
286 |
raw_requests = request.json
|
287 |
+
axiom_logger.info(message="Received request", request_id=req_id, **raw_requests)
|
288 |
|
289 |
with handler_lock:
|
290 |
selected_handler = handlers[handler_index]
|
|
|
299 |
return jsonify({"image_urls": outputs})
|
300 |
except Exception as e:
|
301 |
logger.error("Error during image generation: %s", str(e))
|
302 |
+
axiom_logger.critical("Error during image generation: " + str(e), request_id=req_id, device=selected_handler.device_str)
|
303 |
return jsonify({"error": "Failed to generate image", "details": str(e)}), 500
|
304 |
|
305 |
if __name__ == '__main__':
|