fix zero scheduler on v0.8.4

vllm/zero_overhead/v0/llm_engine.py

@@ -16,7 +16,7 @@ from vllm.logger import init_logger
 from vllm.executor.executor_base import ExecutorBase
 from vllm.inputs import INPUT_REGISTRY
 from vllm.inputs.data import ProcessorInputs
-from vllm.inputs.parse import is_encoder_decoder_inputs
+from vllm.inputs.parse import split_enc_dec_inputs
 from vllm.inputs.preprocess import InputPreprocessor
 from vllm.inputs.registry import InputRegistry
 from vllm.lora.request import LoRARequest
@@ -573,12 +573,7 @@ class ZeroOverheadEngine(LLMEngine):
         seq_id = next(self.seq_counter)
         eos_token_id = self.input_preprocessor.get_eos_token_id(lora_request)
 
-        if is_encoder_decoder_inputs(processed_inputs):
-            decoder_inputs = processed_inputs["decoder"]
-            encoder_inputs = processed_inputs["encoder"]
-        else:
-            decoder_inputs = processed_inputs
-            encoder_inputs = None
+        encoder_inputs, decoder_inputs = split_enc_dec_inputs(processed_inputs)
 
         seq = ZeroOverheadSequence(seq_id, decoder_inputs, block_size, eos_token_id,
                        lora_request, prompt_adapter_request)