Fix one bug in the grouped-gemm triton kernel (#6772)

b581b225 · Cheng Wan · GitHub · 69dd878b · b581b225
Unverified Commit b581b225 authored May 30, 2025 by Cheng Wan Committed by GitHub May 30, 2025
Hide whitespace changes
Inline Side-by-side

Showing with 1 addition and 1 deletion

python/sglang/srt/layers/moe/ep_moe/kernels.py python/sglang/srt/layers/moe/ep_moe/kernels.py +1 -1

No files found.
--- 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