Update app.py

app.py

@@ -20,13 +20,29 @@ print("Model downloaded:", MODEL_PATH)
 server = subprocess.Popen([
     "./llama/llama-server",
     "-m", MODEL_PATH,
-    "-c", "2048",
-    "-t", "2",
+
+    # smaller context
+    "-c", "1024",
+
+    # use more CPU threads
+    "-t", "8",
+
+    # batch processing
+    "-b", "512",
+
+    # no GPU on free CPU spaces
     "-ngl", "0",
-    "--host", "127.0.0.1",
+
+    # mmap + mlock optimizations
+    "--mlock",
+    "--no-mmap",   # sometimes faster on HF ephemeral disk
+
+    "--host", "0.0.0.0",
     "--port", "8080"
 ])
-time.sleep(6)
+
+time.sleep(10)
+
 
 # =========================
 # 3. GRADIO CHAT FUNCTION