Fix cache type in Idefics2 (#30729)

standardize cache in idefics2

src/transformers/models/idefics2/modeling_idefics2.py

@@ -1591,8 +1591,9 @@ class Idefics2Model(Idefics2PreTrainedModel):
             raise ValueError("You have to specify either input_ids or inputs_embeds")
 
         past_seen_tokens = 0
-        if use_cache:
-            if not isinstance(past_key_values, Cache):
+        return_legacy_cache = False
+        if use_cache and not isinstance(past_key_values, Cache):  # kept for BC (non `Cache` `past_key_values` inputs)
+            return_legacy_cache = True
             past_key_values = DynamicCache.from_legacy_cache(past_key_values)
             past_seen_tokens = past_key_values.get_usable_length(seq_length)
 
@@ -1667,6 +1668,9 @@ class Idefics2Model(Idefics2PreTrainedModel):
             return_dict=return_dict,
         )
 
+        if return_legacy_cache:
+            outputs.past_key_values = outputs.past_key_values.to_legacy_cache()
+
         if not return_dict:
             return tuple(v for v in [*outputs, image_hidden_states] if v is not None)