Fix num_hidden_layers in initialization of new model in Mamba (#30403)

Fix num_hidden_layers in initialization

Originally, the initialization was using config.num_layers instead of config.num_hidden_layers. This fixes that.

src/transformers/models/mamba/modeling_mamba.py

@@ -399,7 +399,7 @@ class MambaPreTrainedModel(PreTrainedModel):
                     # Having just p *= scale would repeatedly scale it down
                     nn.init.kaiming_uniform_(p, a=math.sqrt(5))
                     with torch.no_grad():
-                        p /= math.sqrt(self.config.num_layers)
+                        p /= math.sqrt(self.config.num_hidden_layers)
 
 
 @dataclass