Update main.py

main.py

@@ -3,9 +3,9 @@ from fastapi import FastAPI, Form
 from pydantic import BaseModel
 
 #Model loading
-llm = AutoModelForCausalLM.from_pretrained("Meta-Llama-3-8B-Instruct-Q4_K_M.gguf",
+llm = AutoModelForCausalLM.from_pretrained("TheBloke/Llama-2-7B-Chat-GGUF",
 model_type='llama',
-# max_new_tokens = 1096,
+# max_new_tokens = 4096,
 threads = 3,
 )