fix destructors flow and remove finish_measurements

vllm/engine/llm_engine.py

@@ -1325,9 +1325,6 @@ def _advance_to_next_step(
                 else:
                     seq.append_token_id(sample.output_token, sample.logprobs)
 
-    def finish_measurements(self):
-        self.model_executor.finish_measurements()
-
     def step(self) -> List[Union[RequestOutput, EmbeddingRequestOutput]]:
         """Performs one decoding iteration and returns newly generated results.
 

vllm/entrypoints/llm.py

@@ -236,10 +236,6 @@ def set_tokenizer(self, tokenizer: AnyTokenizer) -> None:
         else:
             tokenizer_group.tokenizer = get_cached_tokenizer(tokenizer)
 
-    def finish_measurements(self):
-        assert not envs.VLLM_USE_V1, "INC does not support vLLM V1"
-        self.llm_engine.finish_measurements()  # type: ignore[attr-defined]
-
     @overload  # LEGACY: single (prompt + optional token ids)
     def generate(
         self,

vllm/executor/hpu_executor.py

@@ -26,6 +26,7 @@ class HPUExecutor(ExecutorBase):
     def _init_executor(self) -> None:
         """Initialize the worker and load the model."""
         self._init_worker()
+        self.shutdown_inc = True
 
     def _get_worker_kwargs(
             self,
@@ -94,9 +95,6 @@ def initialize_cache(self, num_gpu_blocks: int, num_cpu_blocks) -> None:
         msg = f"init_cache_engine took {cache_init_m.get_summary_string()}"
         logger.info(msg)
 
-    def finish_measurements(self):
-        self.driver_worker.finish_measurements()
-
     def execute_model(
             self,
             execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]:
@@ -201,8 +199,10 @@ def stop_profile(self) -> None:
         self.driver_worker.stop_profile()
 
     def shutdown(self) -> None:
-        if hasattr(self.driver_worker, 'shutdown_inc'):
-            self.driver_worker.shutdown_inc()
+        if getattr(self, 'shutdown_inc', False):
+            if hasattr(self.driver_worker, 'shutdown_inc'):
+                self.driver_worker.shutdown_inc()
+            self.shutdown_inc = False
 
 
 class HPUExecutorAsync(HPUExecutor, ExecutorAsyncBase):

vllm/executor/ray_hpu_executor.py

@@ -70,7 +70,12 @@ def _init_executor(self) -> None:
         self.output_decoder = msgspec.msgpack.Decoder(
             Optional[List[SamplerOutput]])
 
+        self.shutdown_inc = True
+
     def shutdown(self) -> None:
+        if getattr(self, 'shutdown_inc', False):
+            self._run_workers("shutdown_inc")
+            self.shutdown_inc = False
         for worker in self.workers:
             worker.__ray_terminate__.remote()
         if hasattr(self, "forward_dag") and self.forward_dag is not None:
@@ -80,9 +85,6 @@ def shutdown(self) -> None:
                 ray.kill(worker)
             self.forward_dag = None
 
-    def finish_measurements(self):
-        self._run_workers("finish_measurements")
-
     def _get_worker_module_and_class(
         self
     ) -> Tuple[str, str, Optional[Callable[[],

vllm/worker/hpu_model_runner.py

@@ -1777,10 +1777,6 @@ def prepare_model_input(
                                    is_prompt=is_prompt,
                                    virtual_engine=virtual_engine)
 
-    def finish_measurements(self):
-        from neural_compressor.torch.quantization import finalize_calibration
-        finalize_calibration(self.model.model)
-
     def _check_config(self, batch_size, seq_len, is_prompt, warmup_mode):
         cfg = (batch_size, seq_len, is_prompt)
         seen = cfg in self.seen_configs
@@ -2161,5 +2157,3 @@ def shutdown_inc(self):
             finalize_calibration(self.model.model)
             self._is_inc_finalized = True
 
-    def __del__(self):
-        self.shutdown_inc()

vllm/worker/hpu_worker.py

@@ -251,9 +251,6 @@ def _warm_up_model(self) -> None:
         # the model initialization and profiling.
         set_random_seed(self.model_config.seed)
 
-    def finish_measurements(self):
-        self.model_runner.finish_measurements()
-
     @property
     def do_metadata_broadcast(self) -> bool:
         return self.parallel_config.tensor_parallel_size > 1