remove to.cuda

app.py

@@ -10,7 +10,6 @@ import json
 import bs4
 import os
 os.environ["USE_FLASH_ATTENTION"] = "0"
-device = "cuda" if torch.cuda.is_available() else "cpu"
 
 product_strings = []
 recipe_strings = []
@@ -172,7 +171,6 @@ def respond(
             Frage: {query}
         <|im_end|>
         <|im_start|>assistant"""
-    system_message = system_message.to("cuda") if torch.cuda.is_available() else system_message
     refined_context = generator[1](system_message, do_sample=True, padding=True, truncation=True, top_p=0.95, max_new_tokens=150)
     # Retrieve relevant context from Qdrant
     if "rezept" in query.lower() or "gericht" in query.lower():