[Fix] fix expert_parallel bug in decoder stage (#2848)

Author: freeliuzc

fastdeploy/model_executor/layers/moe/fused_moe_backend_base.py

@@ -49,7 +49,7 @@ def init_ep(self, layer: nn.Layer) -> None:
                 from .ep import EPDecoderRunner
                 self.ep_decoder_runner = EPDecoderRunner(
                     layer.top_k, layer.hidden_size, layer.num_experts,
-                    layer.model_config.num_max_dispatch_tokens_per_rank,
+                    layer.fd_config.model_config.num_max_dispatch_tokens_per_rank,
                     layer.ep_size, layer.ep_rank)
             else:
                 from .ep import EPPrefillRunner

fastdeploy/model_executor/layers/moe/fused_moe_deepgemm_backend.py

@@ -241,7 +241,7 @@ def apply_ep_decode(
             [
                 layer.num_local_experts,
                 layer.ep_size *
-                layer.model_config.num_max_dispatch_tokens_per_rank,
+                layer.fd_config.model_config.num_max_dispatch_tokens_per_rank,
                 layer.moe_intermediate_size * 2,
             ],
             dtype=paddle.bfloat16,
@@ -251,7 +251,7 @@ def apply_ep_decode(
             [
                 layer.num_local_experts,
                 layer.ep_size *
-                layer.model_config.num_max_dispatch_tokens_per_rank,
+                layer.fd_config.model_config.num_max_dispatch_tokens_per_rank,
                 layer.hidden_size,
             ],
             dtype=paddle.bfloat16,