gunicorn: fix deadlock when preload + multiple workers was enabled.

Fix Gunicorn hanging on startup when preloading and multiple workers
were enabled. There was some bad interaction between multithreading in
Pytorch and preloading in Gunicorn (possibly because Pytorch can't
survive a fork?). Disabling threading in Pytorch fixes it.

app.py

@@ -7,11 +7,17 @@ from base64 import b64encode
 from fastai.vision.core import PILImage
 from flask import Flask, request, render_template, jsonify, abort
 from werkzeug.exceptions import HTTPException
+import torch
 
 load_dotenv()
 model_path = getenv("MODEL_PATH", "models/model.pth")
 autotagger = Autotagger(model_path)
 
+# This is necessary for Gunicorn to work with multiple workers and preloading enabled.
+torch.set_num_threads(1)
+autotagger.learn.model.eval()
+autotagger.learn.model.share_memory()
+
 app = Flask(__name__)
 app.config["JSON_SORT_KEYS"] = False
 app.config["JSON_PRETTYPRINT_REGULAR"] = True

gunicorn.conf.py

@@ -10,3 +10,5 @@ errorlog = getenv("GUNICORN_ERRORLOG", "-")
 loglevel = getenv("GUNICORN_LOGLEVEL", "info")
 access_log_format = getenv("GUNICORN_ACCESS_LOG_FORMAT", '{"time":"%(t)s","id":"%({X-Request-Id}i)s","ip":"%(h)s","method":"%(m)s","url":"%(U)s","status":"%(s)s","contentType":"%(Content-Type)s","userAgent":"%(a)s","referer":"%(f)s","sent":"%(B)s","duration":"%(D)s"}')
 preload_app = bool(strtobool(getenv("GUNICORN_PRELOAD", "True")))
+max_requests = int(getenv("GUNICORN_MAX_REQUESTS", 0))
+max_requests_jitter = int(getenv("GUNICORN_MAX_REQUESTS_JITTER", 0))