File size: 8,033 Bytes
ca822d3
 
 
 
 
 
 
 
 
 
 
b6e0a71
ae27e5e
ca822d3
 
 
 
 
 
 
 
dd9c27c
 
b6e0a71
ca822d3
 
249f661
 
ca39fe5
 
ca822d3
dd9c27c
 
 
 
7cdc8db
 
 
dd9c27c
7e2dc74
 
 
 
 
dd9c27c
 
 
 
 
ca822d3
 
 
 
 
 
 
 
 
 
 
 
 
ca39fe5
 
 
 
 
b6e0a71
dd9c27c
3cb2c68
dd9c27c
 
 
 
 
 
 
 
 
 
 
 
 
ca822d3
 
 
249f661
 
 
 
 
 
 
 
 
ca39fe5
249f661
ca822d3
f419fcc
ca822d3
ae27e5e
96014db
ca822d3
249f661
ca822d3
249f661
 
 
3e47535
ca822d3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b6e0a71
ca822d3
 
 
 
 
 
 
 
 
 
 
dd9c27c
1383dae
 
 
ca822d3
 
1383dae
ca822d3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1383dae
 
 
 
ca39fe5
 
1383dae
f419fcc
 
 
1383dae
 
ca39fe5
 
 
 
 
ca822d3
dd9c27c
 
249f661
ca39fe5
dd9c27c
1383dae
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ca822d3
 
 
 
 
 
 
 
 
 
 
 
f419fcc
 
ca822d3
 
 
 
 
 
 
dd9c27c
ca822d3
1383dae
ca822d3
 
 
 
 
 
 
 
 
 
 
 
 
 
dd9c27c
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
import asyncio
import json
import logging
import traceback
from pydantic import BaseModel

from fastapi import FastAPI, WebSocket, HTTPException, WebSocketDisconnect
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import StreamingResponse, JSONResponse
from fastapi.staticfiles import StaticFiles

from diffusers import DiffusionPipeline, AutoencoderTiny
from compel import Compel
import torch
from PIL import Image
import numpy as np
import gradio as gr
import io
import uuid
import os
import time
import psutil

MAX_QUEUE_SIZE = int(os.environ.get("MAX_QUEUE_SIZE", 0))
TIMEOUT = float(os.environ.get("TIMEOUT", 0))
SAFETY_CHECKER = os.environ.get("SAFETY_CHECKER", None)
WIDTH = 512
HEIGHT = 512
# disable tiny autoencoder for better quality speed tradeoff
USE_TINY_AUTOENCODER=True

# check if MPS is available OSX only M1/M2/M3 chips
mps_available = hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
torch_device = device

# change to torch.float16 to save GPU memory
torch_dtype = torch.float32

print(f"TIMEOUT: {TIMEOUT}")
print(f"SAFETY_CHECKER: {SAFETY_CHECKER}")
print(f"MAX_QUEUE_SIZE: {MAX_QUEUE_SIZE}")
print(f"device: {device}")

if mps_available:
    device = torch.device("mps")
    torch_device = "cpu"
    torch_dtype = torch.float32

if SAFETY_CHECKER == "True":
    pipe = DiffusionPipeline.from_pretrained(
        "SimianLuo/LCM_Dreamshaper_v7",
        custom_pipeline="latent_consistency_img2img.py",
        custom_revision="main",
    )
else:
    pipe = DiffusionPipeline.from_pretrained(
        "SimianLuo/LCM_Dreamshaper_v7",
        safety_checker=None,
        custom_pipeline="latent_consistency_img2img.py",
        custom_revision="main",
    )

if USE_TINY_AUTOENCODER:
    pipe.vae = AutoencoderTiny.from_pretrained(
        "madebyollin/taesd", torch_dtype=torch_dtype, use_safetensors=True
    )
pipe.set_progress_bar_config(disable=True)
pipe.to(torch_device=torch_device, torch_dtype=torch_dtype).to(device)
pipe.unet.to(memory_format=torch.channels_last)

if psutil.virtual_memory().total < 64 * 1024**3:
    pipe.enable_attention_slicing()

if not mps_available:
    pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
    pipe(prompt="warmup", image=[Image.new("RGB", (512, 512))])

compel_proc = Compel(
    tokenizer=pipe.tokenizer,
    text_encoder=pipe.text_encoder,
    truncate_long_prompts=False,
)
user_queue_map = {}


class InputParams(BaseModel):
    prompt: str
    seed: int = 2159232
    guidance_scale: float = 8.0
    strength: float = 0.5
    width: int = WIDTH
    height: int = HEIGHT


def predict(input_image: Image.Image, params: InputParams, prompt_embeds: torch.Tensor = None):
    generator = torch.manual_seed(params.seed)
    # Can be set to 1~50 steps. LCM support fast inference even <= 4 steps. Recommend: 1~8 steps.
    num_inference_steps = 3
    results = pipe(
        prompt_embeds=prompt_embeds,
        generator=generator,
        image=input_image,
        strength=params.strength,
        num_inference_steps=num_inference_steps,
        guidance_scale=params.guidance_scale,
        width=params.width,
        height=params.height,
        lcm_origin_steps=50,
        output_type="pil",
    )
    nsfw_content_detected = (
        results.nsfw_content_detected[0]
        if "nsfw_content_detected" in results
        else False
    )
    if nsfw_content_detected:
        return None
    return results.images[0]


app = FastAPI()
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)


@app.websocket("/ws")
async def websocket_endpoint(websocket: WebSocket):
    await websocket.accept()
    if MAX_QUEUE_SIZE > 0 and len(user_queue_map) >= MAX_QUEUE_SIZE:
        print("Server is full")
        await websocket.send_json({"status": "error", "message": "Server is full"})
        await websocket.close()
        return

    try:
        uid = str(uuid.uuid4())
        print(f"New user connected: {uid}")
        await websocket.send_json(
            {"status": "success", "message": "Connected", "userId": uid}
        )
        user_queue_map[uid] = {"queue": asyncio.Queue()}
        await websocket.send_json(
            {"status": "start", "message": "Start Streaming", "userId": uid}
        )
        await handle_websocket_data(websocket, uid)
    except WebSocketDisconnect as e:
        logging.error(f"WebSocket Error: {e}, {uid}")
        traceback.print_exc()
    finally:
        print(f"User disconnected: {uid}")
        queue_value = user_queue_map.pop(uid, None)
        queue = queue_value.get("queue", None)
        if queue:
            while not queue.empty():
                try:
                    queue.get_nowait()
                except asyncio.QueueEmpty:
                    continue


@app.get("/queue_size")
async def get_queue_size():
    queue_size = len(user_queue_map)
    return JSONResponse({"queue_size": queue_size})


@app.get("/stream/{user_id}")
async def stream(user_id: uuid.UUID):
    uid = str(user_id)
    try:
        user_queue = user_queue_map[uid]
        queue = user_queue["queue"]
        async def generate():
            last_prompt: str = None
            prompt_embeds: torch.Tensor = None
            while True:
                data = await queue.get()
                input_image = data["image"]
                params = data["params"]
                if input_image is None:
                    continue
                # avoid recalculate prompt embeds
                if last_prompt != params.prompt:
                    print("new prompt")
                    prompt_embeds = compel_proc(params.prompt)
                    last_prompt = params.prompt

                image = predict(
                    input_image,
                    params,
                    prompt_embeds,
                )
                if image is None:
                    continue
                frame_data = io.BytesIO()
                image.save(frame_data, format="JPEG")
                frame_data = frame_data.getvalue()
                if frame_data is not None and len(frame_data) > 0:
                    yield b"--frame\r\nContent-Type: image/jpeg\r\n\r\n" + frame_data + b"\r\n"

                await asyncio.sleep(1.0 / 120.0)

        return StreamingResponse(
            generate(), media_type="multipart/x-mixed-replace;boundary=frame"
        )
    except Exception as e:
        logging.error(f"Streaming Error: {e}, {user_queue_map}")
        traceback.print_exc()
        return HTTPException(status_code=404, detail="User not found")


async def handle_websocket_data(websocket: WebSocket, user_id: uuid.UUID):
    uid = str(user_id)
    user_queue = user_queue_map[uid]
    queue = user_queue["queue"]
    if not queue:
        return HTTPException(status_code=404, detail="User not found")
    last_time = time.time()
    try:
        while True:
            data = await websocket.receive_bytes()
            params = await websocket.receive_json()
            params = InputParams(**params)
            pil_image = Image.open(io.BytesIO(data))

            while not queue.empty():
                try:
                    queue.get_nowait()
                except asyncio.QueueEmpty:
                    continue
            await queue.put({"image": pil_image, "params": params})
            if TIMEOUT > 0 and time.time() - last_time > TIMEOUT:
                await websocket.send_json(
                    {
                        "status": "timeout",
                        "message": "Your session has ended",
                        "userId": uid,
                    }
                )
                await websocket.close()
                return

    except Exception as e:
        logging.error(f"Error: {e}")
        traceback.print_exc()


app.mount("/", StaticFiles(directory="img2img", html=True), name="public")