Add environment variable definition to documentation and disable timeout by default

app/faster_whisper/core.py

@@ -20,12 +20,13 @@ model_lock = Lock()
 #   https://opennmt.net/CTranslate2/quantization.html
 
 last_activity_time = time.time()
-idle_timeout = int(os.getenv("IDLE_TIMEOUT", 300))  # default to 5 minutes
+idle_timeout = int(os.getenv("IDLE_TIMEOUT", 0))  # default to being disabled
 
 def monitor_idleness():
     global model
+    if(idle_timeout <= 0): return
     while True:
-        time.sleep(60)  # check every minute
+        time.sleep(15)
         if time.time() - last_activity_time > idle_timeout:
             with model_lock:
                 release_model()

app/openai_whisper/core.py

@@ -15,12 +15,13 @@ model = None
 model_lock = Lock()
 
 last_activity_time = time.time()
-idle_timeout = int(os.getenv("IDLE_TIMEOUT", 300))  # default to 5 minutes
+idle_timeout = int(os.getenv("IDLE_TIMEOUT", 0))  # default to being disabled
 
 def monitor_idleness():
     global model
+    if(idle_timeout <= 0): return
     while True:
-        time.sleep(60)  # check every minute
+        time.sleep(15)  # check every minute
         if time.time() - last_activity_time > idle_timeout:
             with model_lock:
                 release_model()

docs/environmental-variables.md

@@ -24,3 +24,11 @@ For English-only applications, the `.en` models tend to perform better, especial
 ```sh
 export ASR_MODEL_PATH=/data/whisper
 ```
+
+### Configuring the `Model Unloading Timeout`
+
+```sh
+export IDLE_TIMEOUT=300
+```
+
+After no activity for this period (in seconds), unload the model until it is requested again. When set to 0, there will be no timeout.