Fix return_dict_in_generate bug in InstructBlip generate function (#25246)

Fix bug in InstructBlip generate function

Previously, the postprocessing conducted on generated sequences in InstructBlip's generate function assumed these sequences were tensors (i.e. that `return_dict_in_generate == False`).

This commit checks whether the result of the call to the wrapped language model `generate()` is a tensor, and if not attempts to postprocess the sequence attribute of the returned results object.

src/transformers/models/instructblip/modeling_instructblip.py

@@ -1559,6 +1559,9 @@ class InstructBlipForConditionalGeneration(InstructBlipPreTrainedModel):
         # with the tokenizer's bos token being set to </s> which has ID=2,
         # whereas the model's text config has bos token id = 0
         if self.config.text_config.architectures[0] == "LLaMAForCausalLM":
+            if isinstance(outputs, torch.Tensor):
                 outputs[outputs == 0] = 2
+            else:
+                outputs.sequences[outputs.sequences == 0] = 2
 
         return outputs