Use ACT2FN to fetch ReLU activation (#16874)

- all activations should be fetched through ACT2FN
- it returns ReLU as `nn.Module`, which allows attaching hooks on the activation function and prints it to stdout when `print(model)`

src/transformers/models/t5/modeling_t5.py

@@ -282,10 +282,11 @@ class T5DenseReluDense(nn.Module):
         self.wi = nn.Linear(config.d_model, config.d_ff, bias=False)
         self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
         self.dropout = nn.Dropout(config.dropout_rate)
+        self.relu_act = ACT2FN["relu"]
 
     def forward(self, hidden_states):
         hidden_states = self.wi(hidden_states)
-        hidden_states = nn.functional.relu(hidden_states)
+        hidden_states = self.relu_act(hidden_states)
         hidden_states = self.dropout(hidden_states)
         hidden_states = self.wo(hidden_states)
         return hidden_states