[Bugfix] Fix Ernie4_5_MoeForCausalLM shared experts (#21717)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>

vllm/model_executor/models/ernie45_moe.py

@@ -109,8 +109,8 @@ class Ernie4_5_MoeMoE(nn.Module):
         layer_idx = extract_layer_index(prefix)
         self.layer_idx = layer_idx
         self.tp_size = get_tensor_model_parallel_world_size()
-        self.moe_num_shared_experts = getattr(config, "moe_num_shared_experts",
-                                              None)
+        self.has_shared_experts = (getattr(config, "moe_num_shared_experts", 0)
+                                   > 0)
 
         if self.tp_size > config.moe_num_experts:
             raise ValueError(
@@ -137,7 +137,7 @@ class Ernie4_5_MoeMoE(nn.Module):
             prefix=f"{prefix}.experts",
             e_score_correction_bias=self.gate.e_score_correction_bias)
 
-        if self.moe_num_shared_experts is not None:
+        if self.has_shared_experts:
             intermediate_size = (config.moe_intermediate_size *
                                  config.moe_num_shared_experts)
             self.shared_experts = Ernie4_5_MoeMLP(
@@ -153,7 +153,8 @@ class Ernie4_5_MoeMoE(nn.Module):
         orig_shape = hidden_states.shape
         hidden_dim = hidden_states.shape[-1]
         hidden_states = hidden_states.view(-1, hidden_dim)
-        if self.moe_num_shared_experts is not None:
+        shared_output = None
+        if self.has_shared_experts:
             shared_output = self.shared_experts(hidden_states)
 
         router_logits, _ = self.gate(hidden_states)
@@ -161,7 +162,7 @@ class Ernie4_5_MoeMoE(nn.Module):
         final_hidden_states = self.experts(hidden_states=hidden_states,
                                            router_logits=router_logits)
 
-        if self.moe_num_shared_experts is not None and \
+        if self.has_shared_experts and \
               shared_output is not None:
             final_hidden_states = final_hidden_states + shared_output