[Bugfix] Fix Positive Feature Layers in Llava Models (#13514)

Signed-off-by: Alex-Brooks <Alex.brooks@ibm.com>

vllm/model_executor/models/clip.py

@@ -251,7 +251,7 @@ class CLIPEncoder(nn.Module):
     def forward(
         self, inputs_embeds: torch.Tensor, return_all_hidden_states: bool
     ) -> Union[torch.Tensor, list[torch.Tensor]]:
-        hidden_states_pool = []
+        hidden_states_pool = [inputs_embeds]
         hidden_states = inputs_embeds
 
         for encoder_layer in self.layers: