Update app.py

app.py

@@ -9,25 +9,26 @@ model_path = hf_hub_download(
     filename="Atlas-FirstAid.Q4_K_M.gguf"
 )
 
-# Load once at startup, not inside the GPU-decorated function.
-# Since this build is CPU-only, loading inside the GPU context
-# gave no benefit — it just added load time to the first request.
-llm = Llama(
-    model_path=model_path,
-    n_ctx=2048,
-    n_threads=4,        # bumped from 2 — ZeroGPU containers typically
-                         # have more CPU cores available than that
-    n_batch=512,         # faster prompt (input) processing
-    n_gpu_layers=0,      # explicit: this build has no CUDA, so don't
-                         # even attempt GPU offload — avoids wasted checks
-    verbose=False,
-)
-
+# Initialize globally as None
+llm = None
 
-@spaces.GPU(duration=20)  # trimmed from 30 — shorter hold on the quota
+@spaces.GPU(duration=30)
 def chat_response(message, history):
+    global llm
+    
+    # Lazy load the model inside the GPU context so CUDA is detected
+    if llm is None:
+        llm = Llama(
+            model_path=model_path,
+            n_ctx=2048,
+            n_threads=2,
+            n_gpu_layers=-1, 
+            verbose=False,
+        )
+
     messages = []
 
+    # Preserve conversation history (Gradio 5 dict format)
     for msg in history:
         messages.append({
             "role": msg["role"],
@@ -39,14 +40,17 @@ def chat_response(message, history):
         "content": message,
     })
 
+    # Enable streaming for instant response times
     response_stream = llm.create_chat_completion(
         messages=messages,
         temperature=0.7,
         max_tokens=512,
-        stream=True
+        stream=True 
     )
 
     partial_message = ""
+    
+    # Iterate through the stream and yield tokens instantly
     for chunk in response_stream:
         if "choices" in chunk and chunk["choices"]:
             delta = chunk["choices"][0].get("delta", {})