[Bugfix][Structured Output] Disallow MRV1 + PP>1 + async sched + structured output (#56250)

Signed-off-by: Artem Perevedentsev <aperevedents@nvidia.com>
Co-authored-by: CNE Pierre FICHEPOIL <pierre-1.fichepoil@gendarmerie.interieur.gouv.fr>

tests/test_config.py

@@ -1082,10 +1082,11 @@ def test_models_default_to_v2_model_runner(model_config, expected, monkeypatch):
 
 def test_v1_model_runner_rejects_v2_only_features():
     config = SimpleNamespace(
-        parallel_config=SimpleNamespace(
+        parallel_config=ParallelConfig(
             prefill_context_parallel_size=2,
-            enable_batch_sharded_sampling=False,
+            distributed_executor_backend="mp",
         ),
+        scheduler_config=SchedulerConfig.default_factory(async_scheduling=False),
         speculative_config=None,
         model_config=None,
     )
@@ -1245,6 +1246,47 @@ def test_async_scheduling_with_pipeline_parallelism_is_allowed():
     assert cfg.scheduler_config.async_scheduling is True
 
 
+def test_v1_model_runner_drops_async_scheduling_with_pipeline_parallelism(monkeypatch):
+    """PP>1 must stay buildable whenever the V1 model runner is selected, such
+    as the external_launcher fallback; async scheduling is dropped instead."""
+    monkeypatch.setenv("VLLM_USE_V2_MODEL_RUNNER", "0")
+
+    cfg = VllmConfig(
+        scheduler_config=SchedulerConfig(
+            max_model_len=8192,
+            is_encoder_decoder=False,
+        ),
+        parallel_config=ParallelConfig(
+            pipeline_parallel_size=2,
+            distributed_executor_backend="mp",
+            nnodes=2,
+        ),
+    )
+    assert cfg.scheduler_config.async_scheduling is False
+
+
+def test_v1_model_runner_rejects_pipeline_parallelism_with_async_scheduling():
+    """Only the async combination desyncs the grammar FSM (#45014), so plain
+    PP>1 must stay usable on the V1 model runner."""
+    config = SimpleNamespace(
+        parallel_config=ParallelConfig(
+            pipeline_parallel_size=2,
+            distributed_executor_backend="mp",
+        ),
+        scheduler_config=SchedulerConfig.default_factory(async_scheduling=False),
+        speculative_config=None,
+        model_config=None,
+    )
+    config._dflash_needs_multi_kv_group = lambda: False
+    config._is_dflash2_draft = lambda: False
+
+    assert VllmConfig._get_v1_model_runner_unsupported_features(config) == []
+
+    config.scheduler_config.async_scheduling = True
+    unsupported = VllmConfig._get_v1_model_runner_unsupported_features(config)
+    assert "pipeline parallelism with async scheduling" in unsupported
+
+
 def test_data_parallel_rpc_port_has_fixed_default():
     assert ParallelConfig().data_parallel_rpc_port == 29550
 

vllm/config/vllm.py

@@ -1522,6 +1522,15 @@ class VllmConfig:
                     "DP+EP generation accuracy."
                 )
                 self.scheduler_config.async_scheduling = False
+            elif (
+                self.parallel_config.pipeline_parallel_size > 1
+                and not self.use_v2_model_runner
+            ):
+                logger.warning_once(
+                    "Async scheduling is disabled because the V1 model runner "
+                    "does not support it with pipeline parallelism."
+                )
+                self.scheduler_config.async_scheduling = False
             else:
                 self.scheduler_config.async_scheduling = True
 
@@ -2962,6 +2971,20 @@ class VllmConfig:
         if self.parallel_config.prefill_context_parallel_size > 1:
             unsupported.append("prefill context parallel")
 
+        # Note(arpera):
+        # MRV1 + PP>1 + async sched + structured output
+        # does not work in vLLM. For more info see:
+        # https://github.com/vllm-project/vllm/issues/45014
+        # Since recently MRV1 has been deprecated then
+        # there was decided not to fix the issue but instead
+        # to disallow such configuration.
+        # At the same time MRV2 works fine in this case.
+        if (
+            self.parallel_config.pipeline_parallel_size > 1
+            and self.scheduler_config.async_scheduling
+        ):
+            unsupported.append("pipeline parallelism with async scheduling")
+
         # DSpark is implemented only by the V2 GPU model runner.
         if self.speculative_config:
             if self.speculative_config.method == "dspark":