[Bugfix] Surface exceptions from non-blocking execute_model in UniProcExecutor...

[Bugfix] Surface exceptions from non-blocking execute_model in UniProcExecutor to avoid DP deadlocks (#35194)
Signed-off-by: fangyuchu <fangyuchu@qq.com>

vllm/v1/engine/core.py

@@ -443,9 +443,10 @@ class EngineCore:
         deferred_scheduler_output = None
         if self.scheduler.has_requests():
             scheduler_output = self.scheduler.schedule()
-            exec_future = self.model_executor.execute_model(
-                scheduler_output, non_block=True
-            )
+            with self.log_error_detail(scheduler_output):
+                exec_future = self.model_executor.execute_model(
+                    scheduler_output, non_block=True
+                )
             if self.is_ec_consumer:
                 model_executed = scheduler_output.total_num_scheduled_tokens > 0
 

vllm/v1/executor/uniproc_executor.py

@@ -100,12 +100,17 @@ class UniProcExecutor(Executor):
     def execute_model(  # type: ignore[override]
         self, scheduler_output: SchedulerOutput, non_block: bool = False
     ) -> ModelRunnerOutput | None | Future[ModelRunnerOutput | None]:
-        return self.collective_rpc(
+        output = self.collective_rpc(
             "execute_model",
             args=(scheduler_output,),
             non_block=non_block,
             single_value=True,
         )
+        # In non-blocking mode, surface any exception as early as possible.
+        if non_block and output.done():
+            # Raise the exception in-line if the task failed.
+            output.result()
+        return output
 
     def sample_tokens(  # type: ignore[override]
         self, grammar_output: GrammarOutput | None, non_block: bool = False