diff --git a/tests/models/multimodal/generation/test_vit_cudagraph.py b/tests/models/multimodal/generation/test_vit_cudagraph.py index 1db1134fb4..10dbd0c837 100644 --- a/tests/models/multimodal/generation/test_vit_cudagraph.py +++ b/tests/models/multimodal/generation/test_vit_cudagraph.py @@ -236,6 +236,13 @@ MODEL_CONFIGS: dict[str, VitCudagraphTestConfig] = { # that changes the output token count, so video uses the eager path. modalities=["image"], image_prompt=ernie45_vl_chat_template("What is in this image?"), + # Ernie4_5_VLMoeModel is deliberately not torch-compiled, since its + # split of text and vision experts breaks compilation, so piecewise + # cudagraphs have nothing to partition. Only the encoder graphs this + # test covers are captured. + compilation_config_overrides={ + "cudagraph_mode": 2, + }, # Shrink to 1 text + 1 vision layer with random weights so the test # runs on any CI GPU and skips the ~56 GiB weight download. The test # only validates encoder CG capture/replay, not output quality.