[sgl-kernel] Support moe_sum_reduce cuda kernel (#10321)

Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>
Co-authored-by: Xiaoyu Zhang <35585791+BBuf@users.noreply.github.com>

sgl-kernel/python/sgl_kernel/__init__.py

@@ -112,6 +112,7 @@ from sgl_kernel.moe import (
     fp8_blockwise_scaled_grouped_mm,
     moe_align_block_size,
     moe_fused_gate,
+    moe_sum_reduce,
     prepare_moe_input,
     topk_softmax,
 )

sgl-kernel/python/sgl_kernel/moe.py

@@ -36,6 +36,18 @@ def topk_softmax(
     )
 
 
+def moe_sum_reduce(
+    input_tensor,
+    output_tensor,
+    routed_scaling_factor=0,
+):
+    torch.ops.sgl_kernel.moe_sum_reduce.default(
+        input_tensor,
+        output_tensor,
+        routed_scaling_factor,
+    )
+
+
 def moe_fused_gate(
     input_tensor,
     bias,