Support `transformers>=4.56`

modeling_plamo.py

@@ -1426,8 +1426,10 @@ class Plamo2Model(Plamo2PreTrainedModel):
                 past_key_values_prev = past_key_values
                 past_key_values = Plamo2Cache(self.config)
 
-                # If `past_key_values` is a `DynamicCache` object, it must be empty
-                assert len(past_key_values_prev) == 0
+                # If `past_key_values` is a `DynamicCache` object, it must be empty or all layer caches have zero sequence length.
+                assert len(past_key_values_prev) == 0 or not any(
+                    layer_cache.get_seq_length() for layer_cache in past_key_values_prev.layers
+                )
             assert isinstance(past_key_values, Plamo2Cache)
             past_key_values_length = past_key_values.get_seq_length()
             seq_length_with_past = seq_length_with_past + past_key_values_length
@@ -1633,7 +1635,11 @@ class Plamo2ForCausalLM(Plamo2PreTrainedModel):
         image_features: Optional[torch.Tensor] = None,
         **kwargs: Any,
     ) -> Dict[str, Any]:
-        if past_key_values:
+        # Starting from transformers v4.54, `DynamicCache` is passed to `past_key_values` during the prefill stage,
+        # and its length becomes non-zero from v4.56 onward.
+        # `Plamo2Model.forward` converts it into a `Plamo2Cache` on the first call, 
+        # se we use the type of `past_key_values` to distinguish between the prefill and decode stages.
+        if isinstance(past_key_values, Plamo2Cache):
             input_ids = input_ids[:, -1:]
             if image_features is not None:
                 image_features = image_features[:, -1:, :]
@@ -1643,7 +1649,7 @@ class Plamo2ForCausalLM(Plamo2PreTrainedModel):
             # create position_ids on the fly for batch generation
             position_ids = attention_mask.long().cumsum(-1) - 1
             position_ids.masked_fill_(attention_mask == 0, 1)
-            if past_key_values:
+            if isinstance(past_key_values, Plamo2Cache):
                 position_ids = position_ids[:, -1].unsqueeze(-1)
 
         # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
@@ -1714,4 +1720,4 @@ class Bias(nn.Module):
         self,
         x: torch.Tensor,
     ) -> torch.Tensor:
-        return x + self._bias
+        return x + self._bias