Update app.py

app.py

@@ -1,20 +1,17 @@
 from flask import Flask, request, jsonify
 import subprocess
 
-# Absolute paths
-MODEL_PATH = "/app/model/model.gguf"
-BINARY_PATH = "/llama.cpp/build/bin/llama-cli"  # Correct path
-MAX_TOKENS = "256"
-THREADS = "4"
+MODEL_PATH = "/app/model/qwen2.5-coder-1.5b-instruct-q4_k_m.gguf"
+BINARY_PATH = "/llama.cpp/build/bin/llama-cli"
+
+MAX_TOKENS = "128"   # ↓ reduce for speed (important)
+THREADS = "2"        # ↓ safer for HF CPU
 
 app = Flask(__name__)
 
 def query_model(system: str, user: str) -> str:
-    """
-    Run the local GGUF model using llama.cpp binary.
-    Combines system + user prompt if system is provided.
-    """
     prompt = f"System: {system}\nUser: {user}" if system else user
+
     try:
         cmd = [
             BINARY_PATH,
@@ -23,20 +20,31 @@ def query_model(system: str, user: str) -> str:
             "--prompt", prompt,
             "--n_predict", MAX_TOKENS
         ]
-        result = subprocess.run(cmd, capture_output=True, text=True, check=True)
+
+        result = subprocess.run(
+            cmd,
+            capture_output=True,
+            text=True,
+            timeout=60  # 🔥 prevents hanging forever
+        )
+
         return result.stdout.strip() or "No output from model."
+
+    except subprocess.TimeoutExpired:
+        return "Error: Model took too long (timeout)."
     except subprocess.CalledProcessError as e:
         return f"Error running model:\n{e.stderr or str(e)}"
     except Exception as e:
         return f"Unexpected error:\n{str(e)}"
 
+
 @app.route("/api/query", methods=["GET", "POST"])
 def api_query():
     if request.method == "POST":
         data = request.json or {}
         user_prompt = data.get("prompt", "")
         system_prompt = data.get("system", "")
-    else:  # GET request
+    else:
         user_prompt = request.args.get("prompt", "")
         system_prompt = request.args.get("system", "")
 
@@ -46,5 +54,10 @@ def api_query():
     output = query_model(system_prompt, user_prompt)
     return jsonify({"output": output})
 
+
+@app.route("/")
+def home():
+    return "API is running"
+
 if __name__ == "__main__":
     app.run(host="0.0.0.0", port=7860)