From 6f892047ecad16e265bfc5461c5051714fc5c60b Mon Sep 17 00:00:00 2001 From: Baizhou Zhang Date: Mon, 18 May 2026 14:51:32 -0700 Subject: [PATCH] [misc] Throw error when single batch overlap is enabled on Hopper (#25509) --- python/sglang/srt/layers/moe/utils.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/python/sglang/srt/layers/moe/utils.py b/python/sglang/srt/layers/moe/utils.py index 78bd3abd5..c6ac46ee8 100644 --- a/python/sglang/srt/layers/moe/utils.py +++ b/python/sglang/srt/layers/moe/utils.py @@ -279,6 +279,11 @@ def initialize_moe_config(server_args: ServerArgs): DEEPEP_CONFIG = server_args.deepep_config or "" IS_TBO_ENABLED = server_args.enable_two_batch_overlap IS_SBO_ENABLED = server_args.enable_single_batch_overlap + if IS_SBO_ENABLED and torch.cuda.is_available(): + if torch.cuda.get_device_capability()[0] == 9: + raise ValueError( + "SBO (single batch overlap) is not supported on SM90 GPUs with latest sgl-deep-gemm wheel. Please try removing --enable-single-batch-overlap argument." + ) TBO_TOKEN_DISTRIBUTION_THRESHOLD = server_args.tbo_token_distribution_threshold DISABLE_FLASHINFER_CUTLASS_MOE_FP4_ALLGATHER = ( server_args.disable_flashinfer_cutlass_moe_fp4_allgather