Update app.py

app.py

@@ -1,22 +1,31 @@
 from llama_cpp import Llama
 import gradio as gr
 
-# 🦙 Llama model (bartowski GGUF)
-llm = Llama.from_pretrained(
-    repo_id="bartowski/Llama-3.2-3B-Instruct-uncensored-GGUF",
-    filename="Llama-3.2-3B-Instruct-uncensored-Q4_K_M.gguf",
-    n_ctx=4096,
-    n_threads=2,
-    verbose=False
-)
+# 🔁 Lazy loading (important to avoid OOM)
+llm = None
+
+def load_model():
+    global llm
+    if llm is None:
+        llm = Llama.from_pretrained(
+            repo_id="bartowski/microsoft_Phi-4-mini-instruct-GGUF",
+            filename="microsoft_Phi-4-mini-instruct-Q4_K_M.gguf",
+            n_ctx=2048,      # lower = less RAM
+            n_threads=2,
+            verbose=False
+        )
+    return llm
+
 
 def chat(message, history):
+    model = load_model()
+
     prompt = f"""<|im_start|>user
 {message}<|im_end|>
 <|im_start|>assistant
 """
 
-    output = llm(
+    output = model(
         prompt,
         max_tokens=512,
         temperature=0.7,
@@ -29,4 +38,8 @@ def chat(message, history):
         response += chunk["choices"][0]["text"]
         yield response
 
-gr.ChatInterface(chat, title="🦙 Llama 3.2 3B Chat").launch()
+
+gr.ChatInterface(
+    chat,
+    title="🧠 Phi-4 Mini (CPU Stable Space)"
+).launch()