update params

app.py

@@ -15,8 +15,8 @@ DESCRIPTION = "# 真空ジェネレータ\n<p>Imitate 真空 (@vericava)'s posts
 if not torch.cuda.is_available():
     DESCRIPTION += "\n<p>Running on CPU 🥶 This demo does not work on CPU.</p>"
 
-MAX_MAX_NEW_TOKENS = 2048
-DEFAULT_MAX_NEW_TOKENS = 1024
+MAX_MAX_NEW_TOKENS = 768
+DEFAULT_MAX_NEW_TOKENS = 512
 MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "32768"))
 
 

pyproject.toml

@@ -8,6 +8,7 @@ dependencies = [
   "torch==2.4.0",
   "transformers>=4.48.3",
   "peft==0.14.0",
+  "sentencepiece==0.2.0",
 ]
 description = ""
 name = "chat-1"