[Feat] ernie4_5_vl_moe support CudaGraph (#3226)

* delete dynamic control flow for decode

* coda-style

* fix scatter/gather typos and use input stream instead default stream

* support 0-Size Tensor

* update runner and model

* using static mem address as input

* fix mem leak

* refine code

* update mm_buffer

* fix typo

* fix buffersize

* fix unk token

* refine code

* refine

* support other arch

* open cudagraph in vlci

* fix

* update

* update

* update

* fix cmd

* update

---------

Co-authored-by: aquagull <hongyuh@qq.com>
Co-authored-by: Yuanle Liu <yuanlehome@163.com>

custom_ops/gpu_ops/moe/moe_reduce.cu

@@ -59,6 +59,10 @@ paddle::Tensor MoeExpertReduceFunc(
 
   auto output = GetEmptyTensor({num_rows, hidden_size}, input_type, place);
 
+  if(num_rows == 0){
+    return output;
+  }
+
   switch (input_type) {
   case paddle::DataType::BFLOAT16:
     MoeReduceKernel<paddle::DataType::BFLOAT16>(