[Bugfix] update neuron for version > 0.5.0 (#7175)

Signed-off-by: omrishiv <[email protected]>
Co-authored-by: Cyrus Leung <[email protected]>

vllm/engine/arg_utils.py

@@ -316,7 +316,7 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
         parser.add_argument('--block-size',
                             type=int,
                             default=EngineArgs.block_size,
-                            choices=[8, 16, 32],
+                            choices=[8, 16, 32, 128, 256, 512, 1024, 2048],
                             help='Token block size for contiguous chunks of '
                             'tokens.')
 

vllm/executor/neuron_executor.py

@@ -100,9 +100,8 @@ async def execute_model_async(
         self,
         execute_model_req: ExecuteModelRequest,
     ) -> List[SamplerOutput]:
-        output = await make_async(
-            self.driver_worker.execute_model
-        )(seq_group_metadata_list=execute_model_req.seq_group_metadata_list, )
+        output = await make_async(self.driver_worker.execute_model
+                                  )(execute_model_req=execute_model_req, )
         return output
 
     async def check_health_async(self) -> None:

vllm/worker/neuron_model_runner.py

@@ -197,6 +197,7 @@ def prepare_model_input(
         virtual_engine: int = 0,
         finished_requests_ids: Optional[List[str]] = None
     ) -> ModelInputForNeuron:
+        multi_modal_kwargs = None
         # NOTE: We assume that all sequences in the group are all prompts or
         # all decodes.
         is_prompt = seq_group_metadata_list[0].is_prompt

vllm/worker/neuron_worker.py

@@ -89,6 +89,9 @@ def prepare_worker_input(
         return WorkerInput(num_seq_groups=len(
             execute_model_req.seq_group_metadata_list), )
 
+    def execute_worker(self, worker_input: WorkerInput) -> None:
+        pass
+
     def get_cache_block_size_bytes(self) -> int:
         """Determine the size in bytes of a cache block.