fix

flash_vstream/model/language_model/vstream_llama.py

@@ -66,6 +66,7 @@ class VStreamLlamaForCausalLM(VStreamMetaForCausalLM, LlamaForCausalLM):
         images: Optional[torch.FloatTensor] = None,
         features: Optional[torch.FloatTensor] = None,
         return_dict: Optional[bool] = None,
+        cache_position=None,
     ) -> Union[Tuple, CausalLMOutputWithPast]:
         if inputs_embeds is None:
             if self.use_video_streaming_mode:

requirements.txt

@@ -5,7 +5,7 @@ numpy
 torch
 torchvision
 tokenizers
-transformers==4.37.2
+transformers==4.31.0
 accelerate
 bitsandbytes
 scikit-learn