Add echo/useless mode to improve testing efficiency without loading models

app.py

@@ -20,6 +20,7 @@ HEALTH_PORT = int(os.getenv("HEALTH_PORT", "8080"))
 GRADIO_HOST = os.getenv("GRADIO_HOST", "0.0.0.0")
 GRADIO_PORT = int(os.getenv("GRADIO_PORT", "7860"))
 DEFAULT_MAX_NEW_TOKENS = int(os.getenv("DEFAULT_MAX_NEW_TOKENS", "128"))
+ECHO_MODE = os.getenv("ECHO_MODE", "off").lower()  # 'off'|'echo'|'useless'
 
 
 # ---------------- Logging ----------------
@@ -131,6 +132,17 @@ def predict(instruction: str,
     if not instruction or not instruction.strip():
         return "⚠️ مهرباني وکړئ یوه لارښوونه ولیکئ."  # please provide an instruction
 
+    # Fast path: echo/useless mode avoids loading large models during testing.
+    if ECHO_MODE in ("echo", "useless"):
+        # If echo mode, return exactly the combined prompt
+        prompt = instruction.strip()
+        if input_text and input_text.strip():
+            prompt += "\n" + input_text.strip()
+        if ECHO_MODE == "echo":
+            return prompt
+        # If useless mode, return a simple placeholder text
+        return "This is a useless placeholder response."
+
     # Build a simple prompt: instruction (+ input if provided)
     prompt = instruction.strip()
     if input_text and input_text.strip():