[XPU] refactor moe ffn (#5501)

- remove BKCL_DISPATCH_ALL_GATHER
- support sparse mode
- support moe quant_method

custom_ops/xpu_ops/src/ops/pybind/pybind.cc

@@ -143,6 +143,8 @@ std::vector<paddle::Tensor> WeightOnlyLinear(
     const int arch,
     const int group_size);
 
+std::vector<paddle::Tensor> Quant2dPerToken(const paddle::Tensor& x);
+
 std::vector<paddle::Tensor> MoeEPCombine(const paddle::Tensor& ffn_out,
                                          const paddle::Tensor& moe_index,
                                          const paddle::Tensor& weights,
@@ -1275,6 +1277,9 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
         py::arg("arch"),
         py::arg("group_size") = -1);
 
+  m.def(
+      "quant2d_per_token", &Quant2dPerToken, py::arg("x"), "quant x per token");
+
   m.def("xpu_moe_layer",
         &MoeLayer,
         py::arg("x"),