Update app.py

app.py

@@ -1,23 +1,4 @@
-import os
 import gradio as gr
-from huggingface_hub import InferenceClient
-
-SYSTEM_PROMPT = os.environ["SYSTEM_PROMPT"]
-client = InferenceClient("HuggingFaceTB/SmolLM2-1.7B-Instruct")
-
-def chat(message, history):
-    messages = [{"role": "system", "content": SYSTEM_PROMPT}]
-    messages.extend(history)
-    messages.append({"role": "user", "content": message})
-    
-    result = ""
-    for chunk in client.chat.completions.create(
-        messages,
-        max_tokens=1024,
-        stream=True
-    ):
-        result += chunk.choices[0].delta.content
-        yield result
 
 with gr.Blocks(
     title="💬 SmolLM Instruct",
@@ -29,12 +10,7 @@ with gr.Blocks(
     )
 ) as app:
     gr.Markdown("# 💬 SmolLM Instruct")
-    
-    gr.ChatInterface(
-        fn=chat,
-        type="messages"
-    )
-
+    gr.load("models/HuggingFaceTB/SmolLM2-1.7B-Instruct")
     with gr.Accordion("ℹ️ About", open=False):
         gr.Markdown(f"""
         * Created by [🍒 cherry-ghosts community](https://hf.co/cherry-ghosts)