Update app.py

app.py

@@ -132,7 +132,7 @@ def load_models_and_documents():
             chat_format="llama-3",
             n_ctx=16384,
             n_gpu_layers=32,
-            flash_attn=True,
+            #flash_attn=True,
             type_k=GGML_TYPE_I8,
             type_v=GGML_TYPE_I8
         )
@@ -149,10 +149,10 @@ def load_models_and_documents():
         )
 
         data_dir = os.path.join(os.getenv('HF_HOME'), 'nltk_data')
-        os.makedirs(os.path.join(data_dir, 'taggers'), exist_ok=True)
-        os.makedirs(os.path.join(data_dir, 'tokenizers'), exist_ok=True)
-
         nltk.data.path.append(data_dir)
+
+        nltk.download('punkt', data_dir, halt_on_error=False)
+        nltk.download('averaged_perceptron_tagger', data_dir, halt_on_error=False)
         
         tagger_path = nltk.data.find('taggers/averaged_perceptron_tagger/averaged_perceptron_tagger.pickle')
         with open(os.path.join(data_dir, 'taggers', 'averaged_perceptron_tagger.pickle'), 'wb') as f: