[Bugfix][Diffusion] Drop aborted request's orphaned async output (fix #6413) (#6439)

Signed-off-by: summer <128961079+zhang-keliang@users.noreply.github.com>
Co-authored-by: multica-agent <github@multica.ai>

vllm_omni/diffusion/diffusion_engine.py

@@ -1609,6 +1609,11 @@ class DiffusionEngine:
             raise RuntimeError(f"Diffusion scheduler lost state for request {request_id}.")
 
         if state.status == DiffusionRequestStatus.FINISHED_ABORTED:
+            # An aborted request is never waited on, so a pending async output
+            # would be cached forever by the executor result pump (issue #6413).
+            # Tell the executor to drop it before returning the aborted result.
+            if runner_output is not None and runner_output.async_output_id is not None:
+                self.executor.drop_output(runner_output.async_output_id)
             # Preserve runner-provided abort details when available.
             if runner_output is not None and runner_output.result is not None and runner_output.result.aborted:
                 return runner_output.result