From 354ee46c26b4c0c63760234a10c57c9498c45199 Mon Sep 17 00:00:00 2001 From: cctry Date: Wed, 9 Sep 2026 11:56:30 -0700 Subject: [PATCH] Keep VMM capability votes on CPU (#38722) --- python/sglang/srt/utils/cuda_vmm_utils.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/python/sglang/srt/utils/cuda_vmm_utils.py b/python/sglang/srt/utils/cuda_vmm_utils.py index 6e1f56283..8f4cfe9ca 100644 --- a/python/sglang/srt/utils/cuda_vmm_utils.py +++ b/python/sglang/srt/utils/cuda_vmm_utils.py @@ -667,7 +667,7 @@ class VmmReservation: def all_ranks_ok(group: ProcessGroup, ok: bool) -> bool: """True iff ``ok`` holds on every rank in ``group`` (BAND all-reduce).""" - flag = torch.tensor([1 if ok else 0], dtype=torch.int32) + flag = torch.tensor([1 if ok else 0], dtype=torch.int32, device="cpu") dist.all_reduce(flag, op=dist.ReduceOp.BAND, group=group) return flag.item() == 1