fix

Author: KuilongCui

tests/e2e_test/test_e2e.py

@@ -63,7 +63,7 @@ def generate_launch_command(result_filename: str = "", launch_ray_cluster: bool
         f"--trust-remote-code "
         f"--request-migration-policy LCFS "
         f"--migration-backend {migration_backend} "
-        f"--migration-cache-blocks 32 "
+        f"--migration-buffer-blocks 32 "
         f"--tensor-parallel-size 1 "
         f"--request-output-queue-port {1234+port} "
         f"{'--enable-pd-disagg ' if enable_pd_disagg else ''} "

tests/unit_test/llumlet/test_engine_step_exception.py

@@ -37,7 +37,7 @@ def __init__(self, *args, **kwargs) -> None:
     def set_error_step(self, broken: bool):
         self.backend_engine._stop_event.set()
 
-        time.sleep(3)
+        time.sleep(30)
 
         assert self.backend_engine.state == EngineState.STOPPED
 
@@ -55,7 +55,7 @@ async def raise_error_step():
 @pytest.mark.skipif(torch.cuda.device_count() < 1, reason="Need at least 1 GPU to run the test.")
 def test_engine_step_exception(setup_ray_env):
     engine_args = EngineArgs(model="facebook/opt-125m", max_model_len=8, worker_use_ray=True)
-    migration_config = MigrationConfig("LCFS", "rpc", 16, 1, 4, 5, 20, 1)
+    migration_config = MigrationConfig("LCFS", "rpc", 16, 1, 4, 5, 20, 2)
     node_id = ray.get_runtime_context().get_node_id()
     scheduling_strategy = NodeAffinitySchedulingStrategy(node_id=node_id, soft=False)