File size: 6,956 Bytes
ac2243f | 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 | import asyncio
import gc
import logging
import os
import random
import threading
from contextlib import asynccontextmanager
from dataclasses import dataclass
from typing import Any, Dict, Optional, Type
import torch
from fastapi import FastAPI, HTTPException, Request
from fastapi.concurrency import run_in_threadpool
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import FileResponse
from Pipelines import ModelPipelineInitializer
from pydantic import BaseModel
from utils import RequestScopedPipeline, Utils
@dataclass
class ServerConfigModels:
model: str = "stabilityai/stable-diffusion-3.5-medium"
type_models: str = "t2im"
constructor_pipeline: Optional[Type] = None
custom_pipeline: Optional[Type] = None
components: Optional[Dict[str, Any]] = None
torch_dtype: Optional[torch.dtype] = None
host: str = "0.0.0.0"
port: int = 8500
server_config = ServerConfigModels()
@asynccontextmanager
async def lifespan(app: FastAPI):
logging.basicConfig(level=logging.INFO)
app.state.logger = logging.getLogger("diffusers-server")
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:128,expandable_segments:True"
os.environ["CUDA_LAUNCH_BLOCKING"] = "0"
app.state.total_requests = 0
app.state.active_inferences = 0
app.state.metrics_lock = asyncio.Lock()
app.state.metrics_task = None
app.state.utils_app = Utils(
host=server_config.host,
port=server_config.port,
)
async def metrics_loop():
try:
while True:
async with app.state.metrics_lock:
total = app.state.total_requests
active = app.state.active_inferences
app.state.logger.info(f"[METRICS] total_requests={total} active_inferences={active}")
await asyncio.sleep(5)
except asyncio.CancelledError:
app.state.logger.info("Metrics loop cancelled")
raise
app.state.metrics_task = asyncio.create_task(metrics_loop())
try:
yield
finally:
task = app.state.metrics_task
if task:
task.cancel()
try:
await task
except asyncio.CancelledError:
pass
try:
stop_fn = getattr(model_pipeline, "stop", None) or getattr(model_pipeline, "close", None)
if callable(stop_fn):
await run_in_threadpool(stop_fn)
except Exception as e:
app.state.logger.warning(f"Error during pipeline shutdown: {e}")
app.state.logger.info("Lifespan shutdown complete")
app = FastAPI(lifespan=lifespan)
logger = logging.getLogger("DiffusersServer.Pipelines")
initializer = ModelPipelineInitializer(
model=server_config.model,
type_models=server_config.type_models,
)
model_pipeline = initializer.initialize_pipeline()
model_pipeline.start()
request_pipe = RequestScopedPipeline(model_pipeline.pipeline)
pipeline_lock = threading.Lock()
logger.info(f"Pipeline initialized and ready to receive requests (model ={server_config.model})")
app.state.MODEL_INITIALIZER = initializer
app.state.MODEL_PIPELINE = model_pipeline
app.state.REQUEST_PIPE = request_pipe
app.state.PIPELINE_LOCK = pipeline_lock
class JSONBodyQueryAPI(BaseModel):
model: str | None = None
prompt: str
negative_prompt: str | None = None
num_inference_steps: int = 28
num_images_per_prompt: int = 1
@app.middleware("http")
async def count_requests_middleware(request: Request, call_next):
async with app.state.metrics_lock:
app.state.total_requests += 1
response = await call_next(request)
return response
@app.get("/")
async def root():
return {"message": "Welcome to the Diffusers Server"}
@app.post("/api/diffusers/inference")
async def api(json: JSONBodyQueryAPI):
prompt = json.prompt
negative_prompt = json.negative_prompt or ""
num_steps = json.num_inference_steps
num_images_per_prompt = json.num_images_per_prompt
wrapper = app.state.MODEL_PIPELINE
initializer = app.state.MODEL_INITIALIZER
utils_app = app.state.utils_app
if not wrapper or not wrapper.pipeline:
raise HTTPException(500, "Model not initialized correctly")
if not prompt.strip():
raise HTTPException(400, "No prompt provided")
def make_generator():
g = torch.Generator(device=initializer.device)
return g.manual_seed(random.randint(0, 10_000_000))
req_pipe = app.state.REQUEST_PIPE
def infer():
gen = make_generator()
return req_pipe.generate(
prompt=prompt,
negative_prompt=negative_prompt,
generator=gen,
num_inference_steps=num_steps,
num_images_per_prompt=num_images_per_prompt,
device=initializer.device,
output_type="pil",
)
try:
async with app.state.metrics_lock:
app.state.active_inferences += 1
output = await run_in_threadpool(infer)
async with app.state.metrics_lock:
app.state.active_inferences = max(0, app.state.active_inferences - 1)
urls = [utils_app.save_image(img) for img in output.images]
return {"response": urls}
except Exception as e:
async with app.state.metrics_lock:
app.state.active_inferences = max(0, app.state.active_inferences - 1)
logger.error(f"Error during inference: {e}")
raise HTTPException(500, f"Error in processing: {e}")
finally:
if torch.cuda.is_available():
torch.cuda.synchronize()
torch.cuda.empty_cache()
torch.cuda.reset_peak_memory_stats()
torch.cuda.ipc_collect()
gc.collect()
@app.get("/images/{filename}")
async def serve_image(filename: str):
utils_app = app.state.utils_app
file_path = os.path.join(utils_app.image_dir, filename)
if not os.path.isfile(file_path):
raise HTTPException(status_code=404, detail="Image not found")
return FileResponse(file_path, media_type="image/png")
@app.get("/api/status")
async def get_status():
memory_info = {}
if torch.cuda.is_available():
memory_allocated = torch.cuda.memory_allocated() / 1024**3 # GB
memory_reserved = torch.cuda.memory_reserved() / 1024**3 # GB
memory_info = {
"memory_allocated_gb": round(memory_allocated, 2),
"memory_reserved_gb": round(memory_reserved, 2),
"device": torch.cuda.get_device_name(0),
}
return {"current_model": server_config.model, "type_models": server_config.type_models, "memory": memory_info}
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
if __name__ == "__main__":
import uvicorn
uvicorn.run(app, host=server_config.host, port=server_config.port)
|