From b581b2250474a4277459426b05aeeacee8b1f0d8 Mon Sep 17 00:00:00 2001 From: Cheng Wan <54331508+ch-wan@users.noreply.github.com> Date: Fri, 30 May 2025 01:42:08 -0700 Subject: [PATCH] Fix one bug in the grouped-gemm triton kernel (#6772) --- python/sglang/srt/layers/moe/ep_moe/kernels.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/python/sglang/srt/layers/moe/ep_moe/kernels.py b/python/sglang/srt/layers/moe/ep_moe/kernels.py index 56c6c7db7..e00562f58 100644 --- a/python/sglang/srt/layers/moe/ep_moe/kernels.py +++ b/python/sglang/srt/layers/moe/ep_moe/kernels.py @@ -621,7 +621,7 @@ def grouped_gemm_triton_kernel( b_ptr += BLOCK_SIZE_K if use_fp8_w8a8 and not (group_k > 0 and group_n > 0): - scale_a_value = tl.load(scale_a + expert_id) + scale_a_value = tl.load(scale_a + m_range_start + offs_am[:, None]) scale_b_value = tl.load(scale_b + expert_id) accumulator *= scale_a_value * scale_b_value