Add incubate port.

Author: A-nnonymous

paddle/phi/kernels/moe_fuse_bwd_op.h

@@ -17,7 +17,7 @@
 #include "paddle/common/exception.h"
 #include "paddle/phi/kernels/moe_kernel_impl.h"
 
-
+p/
 template<typename T, int64_t vec_size>
 __global__ void gather_with_mask_permute_kernel(const T* dy, // [s*k, d]
                                 const int* scatter_index, // [s, k]

python/paddle/incubate/nn/functional/__init__.py

@@ -50,6 +50,9 @@
 from .build_src_rank_and_local_expert_id import build_src_rank_and_local_expert_id
 from .int_bincount import int_bincount
 from .fused_rms_norm_ext import fused_rms_norm_ext
+from .moe_gate_dispatch import moe_gate_dispatch
+from .moe_gate_dispatch_permute import moe_gate_dispatch_permute
+from .moe_ops_partial_nosoftmaxtopk import moe_gate_dispatch_partial_nosoftmaxtopk
 
 __all__ = [
     'fused_multi_head_attention',
@@ -75,4 +78,8 @@
     "build_src_rank_and_local_expert_id"
     "int_bincount",
     "fused_rms_norm_ext",
+    "moe_gate_dispatch",
+    "moe_gate_dispatch_permute",
+    "moe_gate_dispatch_partial_nosoftmaxtopk",
+
 ]