Add a CUDA kernel for fusing mapping and weighted sum for MoE. (#6916)

Co-authored-by: Elfie Guo <elfiegxf@gmail.com>

sgl-kernel/python/sgl_kernel/__init__.py

@@ -48,6 +48,7 @@ from sgl_kernel.gemm import (
 )
 from sgl_kernel.grammar import apply_token_bitmask_inplace_cuda
 from sgl_kernel.moe import (
+    apply_shuffle_mul_sum,
     cutlass_fp4_group_mm,
     ep_moe_post_reorder,
     ep_moe_pre_reorder,

sgl-kernel/python/sgl_kernel/moe.py

@@ -178,6 +178,17 @@ def prepare_moe_input(
     )
 
 
+def apply_shuffle_mul_sum(
+    input,
+    output,
+    permutation,
+    factors,
+):
+    torch.ops.sgl_kernel.apply_shuffle_mul_sum.default(
+        input, output, permutation, factors
+    )
+
+
 def cutlass_fp4_group_mm(
     a_fp4,
     b_fp4,