Update main.py

main.py

@@ -21,9 +21,9 @@ embeddingModel = AutoModel.from_pretrained('./multilingual-e5-base')
 
 # chatGpt replacement
 inferenceTokenizer = AutoTokenizer.from_pretrained(
-    "./ct2fast-flan-alpaca-xxl")
+    "./ct2fast-flan-alpaca-xl")
 inferenceTranslator = Translator(
-    "./ct2fast-flan-alpaca-xxl", compute_type="int8", device="cpu")
+    "./ct2fast-flan-alpaca-xl", compute_type="int8", device="cpu", inter_threads=1, intra_threads=2)
 
 
 class EmbeddingRequest(BaseModel):