From 850dafea92fde38e62385981ad35f52526f233b4 Mon Sep 17 00:00:00 2001 From: Tyler Michael Smith Date: Fri, 20 Jun 2025 19:57:07 +0000 Subject: [PATCH] update Signed-off-by: Tyler Michael Smith --- vllm/distributed/device_communicators/all2all.py | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/vllm/distributed/device_communicators/all2all.py b/vllm/distributed/device_communicators/all2all.py index 35f2fd0ba9e22..c8c373b8adc9b 100644 --- a/vllm/distributed/device_communicators/all2all.py +++ b/vllm/distributed/device_communicators/all2all.py @@ -138,9 +138,11 @@ class DeepEPAll2AllManagerBase(All2AllManagerBase): super().__init__(cpu_group) self.handle_cache = Cache() - # This is the DeepEP default. Stick to it till we can establish - # reasonable defaults based on profiling. - self.num_sms = 20 + # Use all SMs for all2all communication + # This will need to be adjusted for dual-batch overlap + device = self.dp_group.device + props = torch.cuda.get_device_properties(device) + self.num_sms = props.multi_processor_count def get_handle(self, kwargs): raise NotImplementedError