【Inference Optimize】Support ERNIE-4_5-300B-A47B-2BITS-Paddle model TP2/TP4 Inference (#2666)

* Support TP2&TP4 Wint

* Support TP2&TP4 Wint2 Inference

fastdeploy/model_executor/layers/moe/fused_moe_wint2_backend.py

@@ -18,6 +18,8 @@ import paddle
 from paddle import nn
 
 import fastdeploy
+from fastdeploy.distributed.communication_op import \
+    tensor_model_parallel_all_reduce
 
 from ..quantization.quant_base import QuantMethodBase
 from ..utils import create_and_set_parameter, get_tensor
@@ -222,7 +224,6 @@ class TritonWint2FusedMoeMethod(Wint2MoeMethod):
         )
 
         from fastdeploy.model_executor.ops.gpu import moe_expert_reduce
-
         fused_moe_out = moe_expert_reduce(
             ffn_out,
             topk_weights,
@@ -233,4 +234,7 @@ class TritonWint2FusedMoeMethod(Wint2MoeMethod):
             routed_scaling_factor=1.0,
         )
 
+        if layer.tp_size > 1:
+            tensor_model_parallel_all_reduce(fused_moe_out)
+
         return fused_moe_out