Load any 13b/20b/30b model in 8-bit mode when no flags are supplied

server.py

@@ -108,7 +108,7 @@ def load_model(model_name):
 
     # Default settings
     if not (args.cpu or args.load_in_8bit or args.auto_devices or args.disk or args.gpu_memory is not None or args.cpu_memory is not None or args.deepspeed):
-        if model_name.lower().startswith(('gpt-neo', 'opt-', 'galactica')) and any(size in model_name.lower() for size in ('13b', '20b', '30b')):
+        if any(size in model_name.lower() for size in ('13b', '20b', '30b')):
             model = AutoModelForCausalLM.from_pretrained(Path(f"models/{model_name}"), device_map='auto', load_in_8bit=True)
         else:
             model = AutoModelForCausalLM.from_pretrained(Path(f"models/{model_name}"), low_cpu_mem_usage=True, torch_dtype=torch.bfloat16 if args.bf16 else torch.float16).cuda()