Upload main.py

changed to gguf

main.py

@@ -19,8 +19,9 @@ app = FastAPI(docs_url="/")
 
 # pipe = pipeline(model="TheBloke/Llama-2-7B-GGML/llama-2-7b.ggmlv3.q4_0.bin")
 
-model_name_or_path = "TheBloke/Llama-2-7B-GGML"
-model_basename = "llama-2-7b.ggmlv3.q4_0.bin"
+model_name_or_path = "TheBloke/Llama-2-7B-GGUF"
+model_basename = "llama-2-7b.Q4_0.gguf"
+
 model_path = hf_hub_download(repo_id=model_name_or_path, filename=model_basename)
 
 llm = Llama(model_path=model_path)