From 4883121477f8a6d79079eab59aeb1c6f9c7e84ce Mon Sep 17 00:00:00 2001 From: Lucas Liebenwein <11156568+lucaslie@users.noreply.github.com> Date: Thu, 15 May 2025 19:19:30 -0400 Subject: [PATCH] [AutoDeploy] fix: disable overlap scheduler until supported (#4365) Signed-off-by: Lucas Liebenwein <11156568+lucaslie@users.noreply.github.com> --- tensorrt_llm/_torch/auto_deploy/shim/interface.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/tensorrt_llm/_torch/auto_deploy/shim/interface.py b/tensorrt_llm/_torch/auto_deploy/shim/interface.py index ecb77ed83b..b49c3ea500 100644 --- a/tensorrt_llm/_torch/auto_deploy/shim/interface.py +++ b/tensorrt_llm/_torch/auto_deploy/shim/interface.py @@ -120,3 +120,6 @@ class AutoDeployConfig(PyTorchConfig): # gets replaced by the user provided one. We don't want that though. f_default = self.__dataclass_fields__["model_kwargs"].default_factory() setattr(self, "model_kwargs", {**f_default, **getattr(self, "model_kwargs")}) + + # TODO (https://github.com/NVIDIA/TensorRT-LLM/issues/4364) support overlap scheduler + self.disable_overlap_scheduler = True