fixes

Author: manueldeprada

src/transformers/models/prophetnet/modeling_prophetnet.py

@@ -26,7 +26,7 @@
 from torch.nn import LayerNorm
 
 from ...activations import ACT2FN
-from ...cache_utils import Cache, EncoderDecoderCache
+from ...cache_utils import Cache, DynamicCache, EncoderDecoderCache
 from ...generation import GenerationMixin
 from ...modeling_layers import GradientCheckpointingLayer
 from ...modeling_outputs import BaseModelOutput
@@ -1235,7 +1235,7 @@ def forward(
                 use_cache = False
 
         if use_cache and past_key_values is None:
-            past_key_values = EncoderDecoderCache()
+            past_key_values = EncoderDecoderCache() if encoder_hidden_states is not None else DynamicCache()
         return_legacy_cache = False
         if use_cache and isinstance(past_key_values, tuple):
             logger.warning_once(

tests/generation/test_utils.py

@@ -2071,7 +2071,12 @@ def test_generate_with_quant_cache(self):
                 model.generate(past_key_valyes=DynamicCache(), **generation_kwargs, **inputs_dict)
 
             # setting incorrect cache_config args should raise an Error, i.e. nbits=60 does not make sense
-            generation_kwargs["cache_config"] = {"nbits": 60, "q_group_size": 8, "residual_length": 128}
+            generation_kwargs["cache_config"] = {
+                "backend": "quanto",
+                "nbits": 60,
+                "q_group_size": 8,
+                "residual_length": 128,
+            }
             with self.assertRaises(ValueError):
                 model.generate(**generation_kwargs, **inputs_dict)