From cf027b86af71251a8e937a56751636686b4429e4 Mon Sep 17 00:00:00 2001 From: Nick Hill Date: Sat, 13 Jun 2026 18:15:36 -0700 Subject: [PATCH] [Core] Simplify MRV2 async output handling (#45442) --- vllm/v1/executor/multiproc_executor.py | 11 ++++------- vllm/v1/executor/uniproc_executor.py | 2 ++ vllm/v1/worker/gpu/model_runner.py | 9 ++------- 3 files changed, 8 insertions(+), 14 deletions(-) diff --git a/vllm/v1/executor/multiproc_executor.py b/vllm/v1/executor/multiproc_executor.py index b0100c3d66..7bc81118e6 100644 --- a/vllm/v1/executor/multiproc_executor.py +++ b/vllm/v1/executor/multiproc_executor.py @@ -396,9 +396,7 @@ class MultiprocExecutor(Executor): return responses[0] if output_rank is not None else responses future = FutureWrapper( - self.futures_queue, - get_response=get_response, - aggregate=aggregate, + self.futures_queue, get_response=get_response, aggregate=aggregate ) return future if non_block else future.result() @@ -982,6 +980,9 @@ class WorkerProc: func = partial(cloudpickle.loads(method), self.worker) output = func(*args, **kwargs) + + if output_rank is None or self.rank == output_rank: + self.handle_output(output) except Exception as e: # Notes have been introduced in python 3.11 if hasattr(e, "add_note"): @@ -991,10 +992,6 @@ class WorkerProc: # string, only for logging purpose. if output_rank is None or self.rank == output_rank: self.handle_output(e) - continue - - if output_rank is None or self.rank == output_rank: - self.handle_output(output) @staticmethod def setup_proc_title_and_log_prefix(enable_ep: bool) -> None: diff --git a/vllm/v1/executor/uniproc_executor.py b/vllm/v1/executor/uniproc_executor.py index dd04b718d6..3bac65bf4f 100644 --- a/vllm/v1/executor/uniproc_executor.py +++ b/vllm/v1/executor/uniproc_executor.py @@ -90,6 +90,8 @@ class UniProcExecutor(Executor): if not non_block: result = run_method(self.driver_worker, method, args, kwargs) + if isinstance(result, AsyncModelRunnerOutput): + result = result.get_output() return result if single_value else [result] try: diff --git a/vllm/v1/worker/gpu/model_runner.py b/vllm/v1/worker/gpu/model_runner.py index 328b521bfc..31d31e971e 100644 --- a/vllm/v1/worker/gpu/model_runner.py +++ b/vllm/v1/worker/gpu/model_runner.py @@ -145,7 +145,6 @@ class GPUModelRunner(LoRAModelRunnerMixin): self.max_num_reqs = self.scheduler_config.max_num_seqs self.is_encoder_decoder = self.model_config.is_encoder_decoder - self.use_async_scheduling = self.scheduler_config.async_scheduling self.output_copy_stream = torch.cuda.Stream(self.device) # Pipeline parallelism. @@ -1457,9 +1456,7 @@ class GPUModelRunner(LoRAModelRunnerMixin): kv_connector_output = self.kv_connector.post_forward(finished_req_ids) model_runner_output.kv_connector_output = kv_connector_output - if self.use_async_scheduling: - return async_output - return async_output.get_output() + return async_output def take_draft_token_ids(self) -> DraftTokenIds | None: return self.draft_tokens_handler.get_draft_tokens() @@ -1503,9 +1500,7 @@ class GPUModelRunner(LoRAModelRunnerMixin): ) self.postprocess_num_computed_tokens(input_batch) - if self.use_async_scheduling: - return async_output - return async_output.get_output() + return async_output def postprocess_num_computed_tokens(self, input_batch: InputBatch) -> None: # Update the number of computed tokens.