A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 3, 10:30 UTC
[MRv2] Default to MRv1 when a connector is present (#42955)
Signed-off-by: NickLucche <nlucches@redhat.com>
Nicolò Lucchesi committed
May 18, 2026 at 14:34 UTC
69c91d010a596bb74b553fe157497a1fd6edb47c
2 files changed
+11
tests/test_config.py
+7
@@ -148,6 +148,13 @@ def test_is_default_v2_model_runner_model(model_config, expected):
148
assert VllmConfig._is_default_v2_model_runner_model(config) is expected
149
150
151
+def test_use_v2_model_runner_defaults_to_v1_when_kv_connector_present():
152
+ config = SimpleNamespace(kv_transfer_config=object())
153
+ with patch.object(envs, "VLLM_USE_V2_MODEL_RUNNER", None):
154
+ result = VllmConfig.use_v2_model_runner.fget(config)
155
+ assert result is False
156
+
157
+
158
@pytest.mark.skip_global_cleanup
159
def test_with_hf_config_populates_missing_architectures_from_causal_lm_mapping(
160
monkeypatch,
vllm/config/vllm.py
+4
@@ -498,6 +498,10 @@ class VllmConfig:
498
if use_v2_model_runner is not None:
499
return use_v2_model_runner
500
501
+ # KVCache layout changes are breaking, let's stick with v1 for now (see #42846)
502
+ if self.kv_transfer_config is not None:
503
+ return False
504
+
505
if not self._is_default_v2_model_runner_model():
506
return False
507