Update handler.py
Browse files- handler.py +69 -16
handler.py
CHANGED
|
@@ -8,7 +8,54 @@ from para_attn.first_block_cache.diffusers_adapters import apply_cache_on_pipe
|
|
| 8 |
import time
|
| 9 |
import uuid
|
| 10 |
from huggingface_hub import HfApi
|
| 11 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
class EndpointHandler:
|
| 13 |
def __init__(self, path=""):
|
| 14 |
self.pipe = FluxPipeline.from_pretrained(
|
|
@@ -22,9 +69,14 @@ class EndpointHandler:
|
|
| 22 |
self.pipe.vae = torch.compile(
|
| 23 |
self.pipe.vae, mode="max-autotune-no-cudagraphs",
|
| 24 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
|
| 26 |
def __call__(self, data: Dict[str, Any]) -> str:
|
| 27 |
-
global record
|
| 28 |
logger.info(f"Received incoming request with {data=}")
|
| 29 |
|
| 30 |
if "inputs" in data and isinstance(data["inputs"], str):
|
|
@@ -36,8 +88,7 @@ class EndpointHandler:
|
|
| 36 |
"Provided input body must contain either the key `inputs` or `prompt` with the"
|
| 37 |
" prompt to use for the image generation, and it needs to be a non-empty string."
|
| 38 |
)
|
| 39 |
-
|
| 40 |
-
return record
|
| 41 |
parameters = data.pop("parameters", {})
|
| 42 |
|
| 43 |
num_inference_steps = parameters.get("num_inference_steps", 28)
|
|
@@ -48,20 +99,22 @@ class EndpointHandler:
|
|
| 48 |
# seed generator (seed cannot be provided as is but via a generator)
|
| 49 |
seed = parameters.get("seed", 0)
|
| 50 |
generator = torch.manual_seed(seed)
|
| 51 |
-
record+=1
|
| 52 |
start_time = time.time()
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
# ).images[0]
|
| 62 |
end_time = time.time()
|
| 63 |
time_taken = end_time - start_time
|
| 64 |
print(f"Time taken: {time_taken:.2f} seconds")
|
| 65 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 66 |
|
| 67 |
-
return
|
|
|
|
| 8 |
import time
|
| 9 |
import uuid
|
| 10 |
from huggingface_hub import HfApi
|
| 11 |
+
|
| 12 |
+
from pyngrok import ngrok
|
| 13 |
+
import subprocess
|
| 14 |
+
from fastapi import FastAPI
|
| 15 |
+
from fastapi.responses import FileResponse
|
| 16 |
+
import uvicorn
|
| 17 |
+
import threading
|
| 18 |
+
|
| 19 |
+
image_directory='./images'
|
| 20 |
+
if not os.path.exists(image_directory):
|
| 21 |
+
os.makedirs(image_directory)
|
| 22 |
+
|
| 23 |
+
app = FastAPI()
|
| 24 |
+
|
| 25 |
+
@app.get("/images/{image_name}")
|
| 26 |
+
async def get_image(image_name: str):
|
| 27 |
+
image_path = os.path.join(image_directory, image_name)
|
| 28 |
+
|
| 29 |
+
if os.path.exists(image_path):
|
| 30 |
+
return FileResponse(image_path)
|
| 31 |
+
else:
|
| 32 |
+
return {"error": "Image not found"}
|
| 33 |
+
|
| 34 |
+
def run_uvicorn():
|
| 35 |
+
uvicorn.run(app, host="127.0.0.1", port=6000)
|
| 36 |
+
|
| 37 |
+
uvicorn_thread = threading.Thread(target=run_uvicorn)
|
| 38 |
+
uvicorn_thread.daemon = True
|
| 39 |
+
uvicorn_thread.start()
|
| 40 |
+
|
| 41 |
+
authtoken = "2cvqFKWc1eb9b0aN7pRLDUBfEtC_2FUehxFL8CAKXRkW3Hfjo"
|
| 42 |
+
commands = [
|
| 43 |
+
# "snap install ngrok",
|
| 44 |
+
f"ngrok config add-authtoken {authtoken}"
|
| 45 |
+
]
|
| 46 |
+
for command in commands:
|
| 47 |
+
try:
|
| 48 |
+
subprocess.run(command, shell=True, check=True)
|
| 49 |
+
logger.info(f"SUCCESS CMD: {command}")
|
| 50 |
+
except subprocess.CalledProcessError as e:
|
| 51 |
+
logger.info(f"Failed CMD: {e}")
|
| 52 |
+
def run_ngrok():
|
| 53 |
+
subprocess.run(["ngrok", "http", "5000"])
|
| 54 |
+
ngrok_thread = threading.Thread(target=run_ngrok)
|
| 55 |
+
ngrok_thread.daemon = True
|
| 56 |
+
ngrok_thread.start()
|
| 57 |
+
logger.info("ngrok is running in the background")
|
| 58 |
+
|
| 59 |
class EndpointHandler:
|
| 60 |
def __init__(self, path=""):
|
| 61 |
self.pipe = FluxPipeline.from_pretrained(
|
|
|
|
| 69 |
self.pipe.vae = torch.compile(
|
| 70 |
self.pipe.vae, mode="max-autotune-no-cudagraphs",
|
| 71 |
)
|
| 72 |
+
# ngrok
|
| 73 |
+
self.public_url = ngrok.connect(6000).public_url
|
| 74 |
+
# subprocess.Popen(['ngrok', 'http', '6000'], stdout=subprocess.PIPE, stderr=subprocess.PIPE)
|
| 75 |
+
|
| 76 |
+
logger.info("Ngrok is running in the background.")
|
| 77 |
+
|
| 78 |
|
| 79 |
def __call__(self, data: Dict[str, Any]) -> str:
|
|
|
|
| 80 |
logger.info(f"Received incoming request with {data=}")
|
| 81 |
|
| 82 |
if "inputs" in data and isinstance(data["inputs"], str):
|
|
|
|
| 88 |
"Provided input body must contain either the key `inputs` or `prompt` with the"
|
| 89 |
" prompt to use for the image generation, and it needs to be a non-empty string."
|
| 90 |
)
|
| 91 |
+
|
|
|
|
| 92 |
parameters = data.pop("parameters", {})
|
| 93 |
|
| 94 |
num_inference_steps = parameters.get("num_inference_steps", 28)
|
|
|
|
| 99 |
# seed generator (seed cannot be provided as is but via a generator)
|
| 100 |
seed = parameters.get("seed", 0)
|
| 101 |
generator = torch.manual_seed(seed)
|
|
|
|
| 102 |
start_time = time.time()
|
| 103 |
+
result = self.pipe( # type: ignore
|
| 104 |
+
prompt,
|
| 105 |
+
height=height,
|
| 106 |
+
width=width,
|
| 107 |
+
guidance_scale=guidance_scale,
|
| 108 |
+
num_inference_steps=num_inference_steps,
|
| 109 |
+
generator=generator,
|
| 110 |
+
).images[0]
|
|
|
|
| 111 |
end_time = time.time()
|
| 112 |
time_taken = end_time - start_time
|
| 113 |
print(f"Time taken: {time_taken:.2f} seconds")
|
| 114 |
+
filename = f"{uuid.uuid4()}.png"
|
| 115 |
+
image_path = f"/images/{filename}"
|
| 116 |
+
|
| 117 |
+
result.save(image_path)
|
| 118 |
+
image_url = f"{self.public_url+image_path}"
|
| 119 |
|
| 120 |
+
return image_url
|