Update app.py

app.py

@@ -11,6 +11,8 @@ import gradio as gr
 from huggingface_hub import hf_hub_download
 from ui import css, PLACEHOLDER
 
+llm = None
+llm_model = None
 # hf_hub_download(repo_id="bartowski/dolphin-2.9.1-yi-1.5-34b-GGUF", filename="dolphin-2.9.1-yi-1.5-34b-Q6_K.gguf",  local_dir = "./models")
 hf_hub_download(repo_id="crusoeai/dolphin-2.9.1-llama-3-70b-GGUF", filename="dolphin-2.9.1-llama-3-70b.Q3_K_M.gguf",  local_dir = "./models")
 # hf_hub_download(repo_id="kroonen/dolphin-2.9.2-Phi-3-Medium-GGUF", filename="dolphin-2.9.2-Phi-3-Medium-Q6_K.gguf",  local_dir = "./models")
@@ -26,14 +28,19 @@ def respond(
     top_p,
     top_k,
     repeat_penalty,
-):  
-    llm = Llama(
-        model_path=f"models/{model}",
-        flash_attn=True,
-        n_gpu_layers=81,
-        n_batch=1024,
-        n_ctx=8192,
-    )
+):
+    global llm
+    global llm_model
+
+    if llm is None or llm_model != model:
+        llm = Llama(
+            model_path=f"models/{model}",
+            flash_attn=True,
+            n_gpu_layers=81,
+            n_batch=1024,
+            n_ctx=8192,
+        )
+        llm_model=model
     provider = LlamaCppPythonProvider(llm)
 
     agent = LlamaCppAgent(