[MRV2] Support attention-free models (#52374)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Nick Hill
2026-08-14 23:54:50 +00:00
committed by GitHub
parent 9df9b0b0a1
commit d6f17f3d53
3 changed files with 7 additions and 2 deletions
@@ -873,6 +873,10 @@ def get_mamba_prefix_cache_step_configs(
def _run_mamba_prefix_cache_mrv1(
monkeypatch: pytest.MonkeyPatch, async_scheduling: bool
):
# This test patches the V1 model runner, so pin V1 explicitly: MoE/hybrid
# models like Qwen3-Next now default to the V2 runner.
monkeypatch.setenv("VLLM_USE_V2_MODEL_RUNNER", "0")
envs.disable_envs_cache()
global async_scheduling_mode
async_scheduling_mode = async_scheduling
run_ref_mamba_state_in_subprocess()