default art to vllm default generation config, instead of model

Author: Saumya Gandhi

src/art/dev/model.py

@@ -43,6 +43,7 @@ def get_model_config(
         # which is the fallback for devices with compute capability < 8.0
         num_scheduler_steps=16 if torch.cuda.get_device_capability()[0] >= 8 else 1,
         enable_sleep_mode=enable_sleep_mode,
+        generation_config="vllm",
     )
     engine_args.update(config.get("engine_args", {}))
     init_args.update(config.get("init_args", {}))

src/art/dev/openai_server.py

@@ -27,6 +27,7 @@ def get_openai_server_config(
         num_scheduler_steps=16,
         served_model_name=base_model,
         disable_log_requests=True,
+        generation_config="vllm",
     )
     engine_args.update(config.get("engine_args", {}))
     return OpenAIServerConfig(