Update app.py

app.py

@@ -16,8 +16,8 @@ from concurrent.futures import ThreadPoolExecutor
 app = Flask(__name__)
 logging.basicConfig(level=logging.INFO)
 
-MAX_CONTEXT_TOKENS = 1024 * 8
-MAX_GENERATION_TOKENS = 1024 * 8
+MAX_CONTEXT_TOKENS = 1024 * 6
+MAX_GENERATION_TOKENS = 1024 * 4
 
 with open('engines.json', 'r') as f:
     MODELS = json.load(f)
@@ -57,7 +57,7 @@ class LLMManager:
                     n_threads_batch=2,
                     use_mlock=True,
                     mmap=True,
-                    low_vram=False,
+                    low_vram=True,
                     vocab_only=False,
                     verbose=False,
                     logits_all=False,