[Feature] support compute shared experts before combine for better overlap (#6697)

* [Feature] support compute shared experts before combine for better overlap

* fix test

* fix xpu

* fix

fastdeploy/model_executor/layers/quantization/mxfp4.py

@@ -450,7 +450,12 @@ class MXFP4MoeMethod(MoEMethodBase):
             )
 
     def apply(
-        self, layer: nn.Layer, x: paddle.Tensor, router: nn.Layer, topk_ids_hookfunc: Callable = None
+        self,
+        layer: nn.Layer,
+        x: paddle.Tensor,
+        router: nn.Layer,
+        topk_ids_hookfunc: Callable = None,
+        shared_experts: nn.Layer = None,
     ) -> paddle.Tensor:
         router_out = router(x.cast("float32"))
 

fastdeploy/model_executor/layers/quantization/nvfp4.py

@@ -14,9 +14,10 @@
 # limitations under the License.
 """
 
-from typing import Optional
+from typing import Callable, Optional
 
 import paddle
+from paddle import nn
 from paddleformers.utils.log import logger
 
 import fastdeploy
@@ -553,7 +554,8 @@ class ModelOptNvFp4FusedMoE(QuantMethodBase):
         layer,
         x,
         gate,
-        topk_ids_hookfunc=None,
+        topk_ids_hookfunc: Callable = None,
+        shared_experts: nn.Layer = None,
     ):
         """
         flashinfer nvfp4 fusedmoe for Model Optimizer