Enable conversion of multimodal models to pooling tasks (#24451)

Signed-off-by: Max de Bayser <mbayser@br.ibm.com>

vllm/model_executor/models/gemma3_mm.py

@@ -512,7 +512,11 @@ class Gemma3ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP,
             architectures=["Gemma3ForCausalLM"],
         )
         logit_scale = getattr(config, "logit_scale", 1.0)
-        self.language_model.logits_processor.scale *= logit_scale
+
+        if hasattr(self.language_model, "logits_processor"):
+            # The logits processor can be unset if we're using
+            # automatic conversion to pooling model.
+            self.language_model.logits_processor.scale *= logit_scale
 
         self.make_empty_intermediate_tensors = (
             self.language_model.make_empty_intermediate_tensors)