Fix transformers compatibility: pin versions and rename past_key_value to past_key_values

requirements.txt

@@ -1,6 +1,6 @@
-torch>=2.0.0
-transformers>=4.37.0
-gradio>=4.0.0
+torch>=2.0.0,<3.0.0
+transformers>=4.51.0,<4.55.0
+gradio>=4.0.0,<5.0.0
 bitsandbytes>=0.41.0
-accelerate>=0.25.0
+accelerate>=0.25.0,<1.0.0
 huggingface_hub>=0.19.0

src/inference.py

@@ -717,7 +717,7 @@ class DSSDecoder:
                     hidden_states=hidden_states,
                     position_ids=position_ids,
                     attention_mask=None,
-                    past_key_value=None,
+                    past_key_values=None,
                     position_embeddings=position_embeddings,
                     use_cache=False,
                     cache_position=cache_position,

src/model_adapters.py

@@ -33,7 +33,7 @@ class ModelAdapter(ABC):
         hidden_states: Tensor,
         position_ids: Tensor,
         attention_mask: Optional[Tensor],
-        past_key_value: Optional[Tuple],
+        past_key_values: Optional[Tuple],
         position_embeddings: Optional[Tuple],
         use_cache: bool = True,
         cache_position: Optional[Tensor] = None,
@@ -97,7 +97,7 @@ class LlamaStyleAdapter(ModelAdapter):
         hidden_states: Tensor,
         position_ids: Tensor,
         attention_mask: Optional[Tensor],
-        past_key_value: Optional[Tuple],
+        past_key_values: Optional[Tuple],
         position_embeddings: Optional[Tuple],
         use_cache: bool = True,
         cache_position: Optional[Tensor] = None,
@@ -107,7 +107,7 @@ class LlamaStyleAdapter(ModelAdapter):
             hidden_states,
             attention_mask=attention_mask,
             position_ids=position_ids,
-            past_key_value=past_key_value,
+            past_key_values=past_key_values,
             use_cache=use_cache,
             position_embeddings=position_embeddings,
             cache_position=cache_position,