From a43065272f73a7468b1a35dd44fb5b0ed80f88c7 Mon Sep 17 00:00:00 2001 From: Konrad Zawora Date: Thu, 5 Dec 2024 17:47:46 +0100 Subject: [PATCH] [Misc][Gaudi] Avoid torch.compile and enable lazy collectives (#10897) Signed-off-by: Konrad Zawora --- vllm/plugins/__init__.py | 14 ++++++++++++++ 1 file changed, 14 insertions(+) diff --git a/vllm/plugins/__init__.py b/vllm/plugins/__init__.py index 81ee9975..ae6e5c0a 100644 --- a/vllm/plugins/__init__.py +++ b/vllm/plugins/__init__.py @@ -29,6 +29,20 @@ def load_general_plugins(): if current_platform.is_xpu(): # see https://github.com/pytorch/pytorch/blob/8cada5cbe5450e17c26fb8b358116785324537b2/torch/_dynamo/config.py#L158 # noqa os.environ['TORCH_COMPILE_DISABLE'] = 'True' + if current_platform.is_hpu(): + # NOTE(kzawora): PT HPU lazy backend (PT_HPU_LAZY_MODE = 1) + # does not support torch.compile + # Eager backend (PT_HPU_LAZY_MODE = 0) must be selected for + # torch.compile support + is_lazy = os.environ.get('PT_HPU_LAZY_MODE', '1') == '1' + if is_lazy: + # see https://github.com/pytorch/pytorch/blob/43c5f59/torch/_dynamo/config.py#L158 + torch._dynamo.config.disable = True + # NOTE(kzawora) multi-HPU inference with HPUGraphs (lazy-only) + # requires enabling lazy collectives + # see https://docs.habana.ai/en/latest/PyTorch/Inference_on_PyTorch/Inference_Using_HPU_Graphs.html # noqa: E501 + os.environ['PT_HPU_ENABLE_LAZY_COLLECTIVES'] = 'true' + global plugins_loaded if plugins_loaded: return