From 01f558d905a0f0a0b5f6ba0da7476da83543520d Mon Sep 17 00:00:00 2001 From: JINO ROHIT Date: Tue, 21 Jul 2026 12:28:27 +0530 Subject: [PATCH] Sm120 scatter fallback (#31669) Co-authored-by: github-actions[bot] --- .../srt/layers/moe/fused_moe_triton/triton_kernels_moe.py | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/python/sglang/srt/layers/moe/fused_moe_triton/triton_kernels_moe.py b/python/sglang/srt/layers/moe/fused_moe_triton/triton_kernels_moe.py index bb3ba2dd2..f36d77b20 100644 --- a/python/sglang/srt/layers/moe/fused_moe_triton/triton_kernels_moe.py +++ b/python/sglang/srt/layers/moe/fused_moe_triton/triton_kernels_moe.py @@ -17,11 +17,17 @@ from triton_kernels.matmul_ogs import ( ScatterIndx, matmul_ogs, ) +from triton_kernels.matmul_ogs_details.opt_flags import update_opt_flags_constraints from triton_kernels.numerics import InFlexData from triton_kernels.swiglu import swiglu_fn from triton_kernels.tensor import FP4 from sglang.srt.utils import is_cuda +from sglang.srt.utils.common import is_sm120_supported + +if is_sm120_supported(): + # use the regular gather/scatter implementation for unsupported devices. + update_opt_flags_constraints({"is_persistent": False}) if is_cuda(): from sglang.jit_kernel.activation import gelu_and_mul, silu_and_mul