From 44bc478863052efdc455b0282b49830f68fdefb3 Mon Sep 17 00:00:00 2001 From: Shuwen Wang <47200617+alphabetc1@users.noreply.github.com> Date: Sat, 10 Oct 2026 21:49:55 +0800 Subject: [PATCH] [HiSparse] Avoid host synchronization in eager backup (#41446) --- python/sglang/srt/managers/hisparse_coordinator.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/python/sglang/srt/managers/hisparse_coordinator.py b/python/sglang/srt/managers/hisparse_coordinator.py index fcc7d9bec5..4f65dc999a 100644 --- a/python/sglang/srt/managers/hisparse_coordinator.py +++ b/python/sglang/srt/managers/hisparse_coordinator.py @@ -755,9 +755,11 @@ class HiSparseCoordinator: if not backup_indices: return + # Use a pinned, non-blocking H2D copy to avoid synchronizing the host + # with pending work on the schedule stream. backup_indices_gpu = torch.tensor( - backup_indices, dtype=torch.int64, device=self.device - ) + backup_indices, dtype=torch.int64, device="cpu", pin_memory=True + ).to(self.device, non_blocking=True) backup_req_indices = req_pool_indices[backup_indices_gpu] # The previous compressed token's position and its device buffer slot: