[Optimization] Refine row parallel bias and nranks and moe all_reduce (#5247)

* rename nranks to tp_size and fix bias in v1 loader

* fix

* update

fastdeploy/model_executor/layers/moe/moe.py

@@ -21,6 +21,7 @@ from paddle import nn
 from paddleformers.utils.log import logger
 
 from fastdeploy import envs
+from fastdeploy.distributed.communication import tensor_model_parallel_all_reduce
 from fastdeploy.model_executor.layers.utils import get_tensor
 from fastdeploy.model_executor.utils import h2d_copy, slice_fn
 from fastdeploy.platforms import current_platform
@@ -634,4 +635,7 @@ class FusedMoE(nn.Layer):
             out = self.forward_split_allgather(x, gate)
         else:
             out = self.quant_method.apply(self, x, gate)
+
+        if self.reduce_results and self.tp_size > 1:
+            out = tensor_model_parallel_all_reduce(out, self.tp_group)
         return out