MarkZakelj
commited on
Commit
·
e85fe1a
1
Parent(s):
c339505
loras serve
Browse files- gunicorn_config.py +12 -0
- serve_loras.py +25 -17
- serve_loras_prod.py +318 -0
gunicorn_config.py
ADDED
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# gunicorn_config.py
|
2 |
+
import os
|
3 |
+
|
4 |
+
worker_id_counter = 0
|
5 |
+
|
6 |
+
def pre_fork(server, worker):
|
7 |
+
global worker_id_counter
|
8 |
+
worker_id_counter += 1
|
9 |
+
|
10 |
+
def post_fork(server, worker):
|
11 |
+
worker_id = worker_id_counter - 1
|
12 |
+
os.environ['WORKER_ID'] = str(worker_id % 4)
|
serve_loras.py
CHANGED
@@ -9,6 +9,7 @@ from diffusers import StableDiffusionXLPipeline, DiffusionPipeline
|
|
9 |
|
10 |
import numpy as np
|
11 |
import threading
|
|
|
12 |
|
13 |
import base64
|
14 |
from io import BytesIO
|
@@ -23,6 +24,7 @@ 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__)
|
@@ -36,23 +38,25 @@ sentry_sdk.init(
|
|
36 |
|
37 |
LORAS_DIR = './safetensors'
|
38 |
|
|
|
|
|
39 |
handler_lock = threading.Lock()
|
40 |
handler_index = 0
|
41 |
|
42 |
-
class LoraCache():
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
|
55 |
-
lora_cache = LoraCache()
|
56 |
|
57 |
class DiffusersHandler(ABC):
|
58 |
"""
|
@@ -134,7 +138,7 @@ class DiffusersHandler(ABC):
|
|
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",
|
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)}
|
@@ -167,6 +171,7 @@ class DiffusersHandler(ABC):
|
|
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
|
@@ -188,7 +193,8 @@ class DiffusersHandler(ABC):
|
|
188 |
if use_style_lora:
|
189 |
style_lora = os.path.join(LORAS_DIR, style_lora + '.safetensors')
|
190 |
st.time("Load style lora")
|
191 |
-
|
|
|
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)
|
@@ -196,7 +202,8 @@ class DiffusersHandler(ABC):
|
|
196 |
if use_char_lora:
|
197 |
char_lora = os.path.join(LORAS_DIR, char_lora + '.safetensors')
|
198 |
st.time('load character lora')
|
199 |
-
|
|
|
200 |
|
201 |
# lora_weights = model_args.pop("lora_weights")
|
202 |
# if lora_weights is not None:
|
@@ -287,6 +294,8 @@ def generate_image():
|
|
287 |
axiom_logger.info(message="Received request", request_id=req_id, **raw_requests)
|
288 |
|
289 |
with handler_lock:
|
|
|
|
|
290 |
selected_handler = handlers[handler_index]
|
291 |
handler_index = (handler_index + 1) % gpu_count # Rotate to the next handler
|
292 |
selected_handler.req_id = req_id
|
@@ -295,7 +304,6 @@ def generate_image():
|
|
295 |
inferences = selected_handler.inference(processed_request)
|
296 |
outputs = selected_handler.postprocess(inferences)
|
297 |
selected_handler.req_id = None
|
298 |
-
|
299 |
return jsonify({"image_urls": outputs})
|
300 |
except Exception as e:
|
301 |
logger.error("Error during image generation: %s", str(e))
|
|
|
9 |
|
10 |
import numpy as np
|
11 |
import threading
|
12 |
+
import mmap
|
13 |
|
14 |
import base64
|
15 |
from io import BytesIO
|
|
|
24 |
from sequential_timer import SequentialTimer
|
25 |
from safetensors.torch import load_file
|
26 |
import copy
|
27 |
+
import gc
|
28 |
|
29 |
logger = logging.getLogger(__name__)
|
30 |
logger.info("Diffusers version %s", diffusers.__version__)
|
|
|
38 |
|
39 |
LORAS_DIR = './safetensors'
|
40 |
|
41 |
+
lora_lock = threading.Lock()
|
42 |
+
|
43 |
handler_lock = threading.Lock()
|
44 |
handler_index = 0
|
45 |
|
46 |
+
# class LoraCache():
|
47 |
+
# def __init__(self, loras_dir: str = LORAS_DIR):
|
48 |
+
# self.loras_dir = loras_dir
|
49 |
+
# self.cache = {}
|
50 |
+
|
51 |
+
# def load_lora(self, lora_name: str):
|
52 |
+
# if lora_name.endswith('.safetensors'):
|
53 |
+
# lora_name = lora_name.rstrip('.safetensors')
|
54 |
+
# if lora_name not in self.cache:
|
55 |
+
# lora = load_file(os.path.join(self.loras_dir, lora_name+'.safetensors'))
|
56 |
+
# self.cache[lora_name] = lora
|
57 |
+
# return copy.deepcopy(self.cache[lora_name])
|
58 |
|
59 |
+
# lora_cache = LoraCache()
|
60 |
|
61 |
class DiffusersHandler(ABC):
|
62 |
"""
|
|
|
138 |
"negative_prompt": raw_requests[0].get("negative_prompt"),
|
139 |
"width": raw_requests[0].get("width"),
|
140 |
"height": raw_requests[0].get("height"),
|
141 |
+
"num_inference_steps": raw_requests[0].get("num_inference_steps", 30),
|
142 |
"guidance_scale": raw_requests[0].get("guidance_scale", 8.5)
|
143 |
# "lora_weights": raw_requests[0].get("lora_name", None)
|
144 |
# "cross_attention_kwargs": {"scale": raw_requests[0].get("lora_scale", 0.0)}
|
|
|
171 |
# 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])
|
172 |
st = SequentialTimer()
|
173 |
model_args, extra_args = request
|
174 |
+
global lora_cache
|
175 |
|
176 |
use_char_lora = extra_args['char_lora'] is not None
|
177 |
use_style_lora = extra_args['style_lora'] is not None
|
|
|
193 |
if use_style_lora:
|
194 |
style_lora = os.path.join(LORAS_DIR, style_lora + '.safetensors')
|
195 |
st.time("Load style lora")
|
196 |
+
with lora_lock:
|
197 |
+
self.pipe.load_lora_weights(style_lora)
|
198 |
if use_char_lora:
|
199 |
st.time("Fuse style lora into model")
|
200 |
self.pipe.fuse_lora(lora_scale=extra_args['style_scale'], fuse_text_encoder=False)
|
|
|
202 |
if use_char_lora:
|
203 |
char_lora = os.path.join(LORAS_DIR, char_lora + '.safetensors')
|
204 |
st.time('load character lora')
|
205 |
+
with lora_lock:
|
206 |
+
self.pipe.load_lora_weights(char_lora)
|
207 |
|
208 |
# lora_weights = model_args.pop("lora_weights")
|
209 |
# if lora_weights is not None:
|
|
|
294 |
axiom_logger.info(message="Received request", request_id=req_id, **raw_requests)
|
295 |
|
296 |
with handler_lock:
|
297 |
+
if handler_index == 0:
|
298 |
+
gc.collect()
|
299 |
selected_handler = handlers[handler_index]
|
300 |
handler_index = (handler_index + 1) % gpu_count # Rotate to the next handler
|
301 |
selected_handler.req_id = req_id
|
|
|
304 |
inferences = selected_handler.inference(processed_request)
|
305 |
outputs = selected_handler.postprocess(inferences)
|
306 |
selected_handler.req_id = None
|
|
|
307 |
return jsonify({"image_urls": outputs})
|
308 |
except Exception as e:
|
309 |
logger.error("Error during image generation: %s", str(e))
|
serve_loras_prod.py
ADDED
@@ -0,0 +1,318 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from compel import Compel, ReturnedEmbeddingsType
|
2 |
+
import logging
|
3 |
+
from abc import ABC
|
4 |
+
import uuid
|
5 |
+
|
6 |
+
import diffusers
|
7 |
+
import torch
|
8 |
+
from diffusers import StableDiffusionXLPipeline, DiffusionPipeline
|
9 |
+
|
10 |
+
import numpy as np
|
11 |
+
import threading
|
12 |
+
|
13 |
+
import base64
|
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, current_app
|
22 |
+
import os
|
23 |
+
from sequential_timer import SequentialTimer
|
24 |
+
from safetensors.torch import load_file
|
25 |
+
from dotenv import load_dotenv
|
26 |
+
import copy
|
27 |
+
import gc
|
28 |
+
|
29 |
+
|
30 |
+
logger = logging.getLogger(__name__)
|
31 |
+
logger.info("Diffusers version %s", diffusers.__version__)
|
32 |
+
|
33 |
+
from axiom_logger import AxiomLogger
|
34 |
+
axiom_logger = AxiomLogger()
|
35 |
+
|
36 |
+
sentry_sdk.init(
|
37 |
+
dsn="https://f750d1b039d66541f344ee6151d38166@o4505891057696768.ingest.sentry.io/4506071735205888",
|
38 |
+
)
|
39 |
+
|
40 |
+
LORAS_DIR = './safetensors'
|
41 |
+
|
42 |
+
load_dotenv()
|
43 |
+
|
44 |
+
lora_lock = threading.Lock()
|
45 |
+
|
46 |
+
# handler_lock = threading.Lock()
|
47 |
+
# handler_index = 0
|
48 |
+
|
49 |
+
# class LoraCache():
|
50 |
+
# def __init__(self, loras_dir: str = LORAS_DIR):
|
51 |
+
# self.loras_dir = loras_dir
|
52 |
+
# self.cache = {}
|
53 |
+
|
54 |
+
# def load_lora(self, lora_name: str):
|
55 |
+
# if lora_name.endswith('.safetensors'):
|
56 |
+
# lora_name = lora_name.rstrip('.safetensors')
|
57 |
+
# if lora_name not in self.cache:
|
58 |
+
# lora = load_file(os.path.join(self.loras_dir, lora_name+'.safetensors'))
|
59 |
+
# self.cache[lora_name] = lora
|
60 |
+
# return copy.deepcopy(self.cache[lora_name])
|
61 |
+
|
62 |
+
# lora_cache = LoraCache()
|
63 |
+
|
64 |
+
class DiffusersHandler(ABC):
|
65 |
+
"""
|
66 |
+
Diffusers handler class for text to image generation.
|
67 |
+
"""
|
68 |
+
|
69 |
+
def __init__(self):
|
70 |
+
self.initialized = False
|
71 |
+
self.req_id = None
|
72 |
+
|
73 |
+
def initialize(self, properties):
|
74 |
+
"""In this initialize function, the Stable Diffusion model is loaded and
|
75 |
+
initialized here.
|
76 |
+
Args:
|
77 |
+
ctx (context): It is a JSON Object containing information
|
78 |
+
pertaining to the model artefacts parameters.
|
79 |
+
"""
|
80 |
+
|
81 |
+
logger.info("Loading diffusion model")
|
82 |
+
logger.info("I'm totally new and updated")
|
83 |
+
|
84 |
+
|
85 |
+
device_str = "cuda:" + str(properties.get("gpu_id")) if torch.cuda.is_available() and properties.get("gpu_id") is not None else "cpu"
|
86 |
+
self.device_str = device_str
|
87 |
+
|
88 |
+
print("my device is " + device_str)
|
89 |
+
self.device = torch.device(device_str)
|
90 |
+
self.pipe = StableDiffusionXLPipeline.from_pretrained(
|
91 |
+
"./",
|
92 |
+
torch_dtype=torch.float16,
|
93 |
+
use_safetensors=True,
|
94 |
+
)
|
95 |
+
# self.refiner = DiffusionPipeline.from_pretrained(
|
96 |
+
# "stabilityai/stable-diffusion-xl-refiner-1.0",
|
97 |
+
# text_encoder_2=self.pipe.text_encoder_2,
|
98 |
+
# vae=self.pipe.vae,
|
99 |
+
# torch_dtype=torch.float16,
|
100 |
+
# use_safetensors=True,
|
101 |
+
# variant="fp16",
|
102 |
+
# )
|
103 |
+
# self.refiner.enable_model_cpu_offload(properties.get("gpu_id"))
|
104 |
+
# logger.info("Refiner initialized and o")
|
105 |
+
|
106 |
+
self.compel_base = Compel(
|
107 |
+
tokenizer=[self.pipe.tokenizer, self.pipe.tokenizer_2],
|
108 |
+
text_encoder=[self.pipe.text_encoder, self.pipe.text_encoder_2],
|
109 |
+
returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
|
110 |
+
requires_pooled=[False, True])
|
111 |
+
logger.info("Compel initialized")
|
112 |
+
|
113 |
+
# self.compel_refiner = Compel(
|
114 |
+
# tokenizer=[self.refiner.tokenizer_2],
|
115 |
+
# text_encoder=[self.refiner.text_encoder_2],
|
116 |
+
# returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
|
117 |
+
# requires_pooled=[True])
|
118 |
+
|
119 |
+
logger.info("moving base model to device: %s", device_str)
|
120 |
+
self.pipe.to(self.device)
|
121 |
+
|
122 |
+
logger.info(self.device)
|
123 |
+
logger.info("Diffusion model from path %s loaded successfully")
|
124 |
+
axiom_logger.info("Diffusion model initialized", device=self.device_str)
|
125 |
+
|
126 |
+
self.initialized = True
|
127 |
+
|
128 |
+
def preprocess(self, raw_requests):
|
129 |
+
"""Basic text preprocessing, of the user's prompt.
|
130 |
+
Args:
|
131 |
+
requests (str): The Input data in the form of text is passed on to the preprocess
|
132 |
+
function.
|
133 |
+
Returns:
|
134 |
+
list : The preprocess function returns a list of prompts.
|
135 |
+
"""
|
136 |
+
logger.info("Received requests: '%s'", raw_requests)
|
137 |
+
self.working = True
|
138 |
+
|
139 |
+
model_args = {
|
140 |
+
"prompt": raw_requests[0]["prompt"],
|
141 |
+
"negative_prompt": raw_requests[0].get("negative_prompt"),
|
142 |
+
"width": raw_requests[0].get("width"),
|
143 |
+
"height": raw_requests[0].get("height"),
|
144 |
+
"num_inference_steps": raw_requests[0].get("num_inference_steps", 30),
|
145 |
+
"guidance_scale": raw_requests[0].get("guidance_scale", 8.5)
|
146 |
+
# "lora_weights": raw_requests[0].get("lora_name", None)
|
147 |
+
# "cross_attention_kwargs": {"scale": raw_requests[0].get("lora_scale", 0.0)}
|
148 |
+
}
|
149 |
+
|
150 |
+
extra_args = {
|
151 |
+
"seed": raw_requests[0].get("seed", None),
|
152 |
+
"style_lora": raw_requests[0].get("style_lora", None),
|
153 |
+
"style_scale": raw_requests[0].get("style_scale", 1.0),
|
154 |
+
"char_lora": raw_requests[0].get("char_lora", None),
|
155 |
+
"char_scale": raw_requests[0].get("char_scale", 1.0),
|
156 |
+
"scene_prompt": raw_requests[0].get("scene_prompt", None)
|
157 |
+
}
|
158 |
+
|
159 |
+
|
160 |
+
logger.info("Processed request: '%s'", model_args)
|
161 |
+
axiom_logger.info("Processed request:" + str(model_args), request_id=self.req_id, device=self.device_str)
|
162 |
+
return model_args, extra_args
|
163 |
+
|
164 |
+
|
165 |
+
def inference(self, request):
|
166 |
+
"""Generates the image relevant to the received text.
|
167 |
+
Args:
|
168 |
+
inputs (list): List of Text from the pre-process function is passed here
|
169 |
+
Returns:
|
170 |
+
list : It returns a list of the generate images for the input text
|
171 |
+
"""
|
172 |
+
|
173 |
+
# Handling inference for sequence_classification.
|
174 |
+
# 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])
|
175 |
+
st = SequentialTimer()
|
176 |
+
model_args, extra_args = request
|
177 |
+
|
178 |
+
use_char_lora = extra_args['char_lora'] is not None
|
179 |
+
use_style_lora = extra_args['style_lora'] is not None
|
180 |
+
|
181 |
+
|
182 |
+
style_lora = extra_args['style_lora']
|
183 |
+
char_lora = extra_args['char_lora']
|
184 |
+
|
185 |
+
cross_attention_kwargs = {"scale": extra_args['char_scale'] if use_char_lora else extra_args['style_scale']}
|
186 |
+
|
187 |
+
generator = torch.Generator(device="cuda").manual_seed(extra_args['seed']) if extra_args['seed'] else None
|
188 |
+
|
189 |
+
|
190 |
+
prompt = model_args.pop("prompt")
|
191 |
+
negative_prompt = model_args.pop('negative_prompt')
|
192 |
+
scene_prompt = extra_args['scene_prompt']
|
193 |
+
if scene_prompt:
|
194 |
+
prompt = f'("{prompt}", "{scene_prompt}").and()'
|
195 |
+
st.time("Base compel embedding")
|
196 |
+
conditioning, pooled = self.compel_base(prompt)
|
197 |
+
negative_conditioning, negative_pooled = self.compel_base(negative_prompt)
|
198 |
+
|
199 |
+
[conditioning, negative_conditioning] = self.compel_base.pad_conditioning_tensors_to_same_length([conditioning, negative_conditioning])
|
200 |
+
|
201 |
+
if use_style_lora:
|
202 |
+
style_lora = os.path.join(LORAS_DIR, style_lora + '.safetensors')
|
203 |
+
st.time("Load style lora")
|
204 |
+
self.pipe.load_lora_weights(style_lora)
|
205 |
+
if use_char_lora:
|
206 |
+
st.time("Fuse style lora into model")
|
207 |
+
self.pipe.fuse_lora(lora_scale=extra_args['style_scale'], fuse_text_encoder=False)
|
208 |
+
|
209 |
+
if use_char_lora:
|
210 |
+
char_lora = os.path.join(LORAS_DIR, char_lora + '.safetensors')
|
211 |
+
st.time('load character lora')
|
212 |
+
self.pipe.load_lora_weights(char_lora)
|
213 |
+
|
214 |
+
# lora_weights = model_args.pop("lora_weights")
|
215 |
+
# if lora_weights is not None:
|
216 |
+
# lora_path = os.path.join(LORAS_DIR, lora_weights + '.safetensors')
|
217 |
+
# logger.info('LOADING LORA FROM: ' + lora_path)
|
218 |
+
# self.pipe.load_lora_weights(lora_path)
|
219 |
+
|
220 |
+
# Handling inference for sequence_classification.
|
221 |
+
st.time("base model inference")
|
222 |
+
inferences = self.pipe(
|
223 |
+
prompt_embeds=conditioning,
|
224 |
+
pooled_prompt_embeds=pooled,
|
225 |
+
negative_prompt_embeds=negative_conditioning,
|
226 |
+
negative_pooled_prompt_embeds=negative_pooled,
|
227 |
+
generator=generator,
|
228 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
229 |
+
**model_args
|
230 |
+
).images
|
231 |
+
|
232 |
+
if use_style_lora and use_char_lora:
|
233 |
+
st.time("unfuse lora weights")
|
234 |
+
self.pipe.unfuse_lora(unfuse_text_encoder=False)
|
235 |
+
|
236 |
+
if use_style_lora or use_char_lora:
|
237 |
+
st.time("unload lora weights")
|
238 |
+
self.pipe.unload_lora_weights()
|
239 |
+
|
240 |
+
st.time('end')
|
241 |
+
|
242 |
+
# logger.info("Generated image: '%s'", inferences)
|
243 |
+
axiom_logger.info("Generated images", request_id=self.req_id, device=self.device_str, timings=st.to_str())
|
244 |
+
return inferences
|
245 |
+
|
246 |
+
def postprocess(self, inference_outputs):
|
247 |
+
"""Post Process Function converts the generated image into Torchserve readable format.
|
248 |
+
Args:
|
249 |
+
inference_outputs (list): It contains the generated image of the input text.
|
250 |
+
Returns:
|
251 |
+
(list): Returns a list of the images.
|
252 |
+
"""
|
253 |
+
bucket_name = "outputs-storage-prod"
|
254 |
+
client = storage.Client()
|
255 |
+
self.working = False
|
256 |
+
bucket = client.get_bucket(bucket_name)
|
257 |
+
outputs = []
|
258 |
+
for image in inference_outputs:
|
259 |
+
image_name = str(uuid.uuid4())
|
260 |
+
|
261 |
+
blob = bucket.blob(image_name + '.png')
|
262 |
+
|
263 |
+
with TemporaryFile() as tmp:
|
264 |
+
image.save(tmp, format="png")
|
265 |
+
tmp.seek(0)
|
266 |
+
blob.upload_from_file(tmp, content_type='image/png')
|
267 |
+
|
268 |
+
# generate txt file with the image name and the prompt inside
|
269 |
+
# blob = bucket.blob(image_name + '.txt')
|
270 |
+
# blob.upload_from_string(self.prompt)
|
271 |
+
url_name = 'https://storage.googleapis.com/' + bucket_name + '/' + image_name + '.png'
|
272 |
+
outputs.append(url_name)
|
273 |
+
axiom_logger.info("Pushed image to google cloud: "+ url_name, request_id=self.req_id, device=self.device_str)
|
274 |
+
return outputs
|
275 |
+
|
276 |
+
|
277 |
+
app = Flask(__name__)
|
278 |
+
|
279 |
+
# Initialize the handler on startup
|
280 |
+
gpu_count = torch.cuda.device_count()
|
281 |
+
if gpu_count == 0:
|
282 |
+
raise ValueError("No GPUs available!")
|
283 |
+
|
284 |
+
worker_id = os.environ.get('WORKER_ID', 'Unknown')
|
285 |
+
if worker_id == 'Unknown':
|
286 |
+
raise ValueError("No worker id")
|
287 |
+
logger.critical("cant get worker ID")
|
288 |
+
logger.info(f"WORKER ID: {worker_id}")
|
289 |
+
handler = DiffusersHandler()
|
290 |
+
handler.initialize({"gpu_id": worker_id})
|
291 |
+
|
292 |
+
|
293 |
+
@app.route('/generate', methods=['POST'])
|
294 |
+
def generate_image():
|
295 |
+
req_id = str(uuid.uuid4())
|
296 |
+
selected_handler = None
|
297 |
+
try:
|
298 |
+
# Extract raw requests from HTTP POST body
|
299 |
+
raw_requests = request.json
|
300 |
+
axiom_logger.info(message="Received request", request_id=req_id, **raw_requests)
|
301 |
+
|
302 |
+
gc.collect()
|
303 |
+
torch.cuda.empty_cache()
|
304 |
+
selected_handler = handler
|
305 |
+
selected_handler.req_id = req_id
|
306 |
+
|
307 |
+
processed_request = selected_handler.preprocess([raw_requests])
|
308 |
+
inferences = selected_handler.inference(processed_request)
|
309 |
+
outputs = selected_handler.postprocess(inferences)
|
310 |
+
selected_handler.req_id = None
|
311 |
+
return jsonify({"image_urls": outputs})
|
312 |
+
except Exception as e:
|
313 |
+
logger.error("Error during image generation: %s", str(e))
|
314 |
+
axiom_logger.critical("Error during image generation: " + str(e), request_id=req_id, device=selected_handler.device_str)
|
315 |
+
return jsonify({"error": "Failed to generate image", "details": str(e)}), 500
|
316 |
+
|
317 |
+
if __name__ == '__main__':
|
318 |
+
app.run(host='0.0.0.0', port=3000, threaded=False)
|