[Bugfix][CI] Set cudagraph_mode=FULL for the Ernie4.5-VL ViT cudagraph test (#54957)
Signed-off-by: Stefan Koncarevic <stefan.koncarevic@amd.com>
stefankoncarevic committed
Sep 2, 2026 at 20:18 UTC
2691c6cc535692d173986e93384c84970a0fe009
1 file changed
+7
tests/models/multimodal/generation/test_vit_cudagraph.py
+7
@@ -236,6 +236,13 @@ MODEL_CONFIGS: dict[str, VitCudagraphTestConfig] = {
236
# that changes the output token count, so video uses the eager path.
237
modalities=["image"],
238
image_prompt=ernie45_vl_chat_template("What is in this image?"),
239
+ # Ernie4_5_VLMoeModel is deliberately not torch-compiled, since its
240
+ # split of text and vision experts breaks compilation, so piecewise
241
+ # cudagraphs have nothing to partition. Only the encoder graphs this
242
+ # test covers are captured.
243
+ compilation_config_overrides={
244
+ "cudagraph_mode": 2,
245
+ },
246
# Shrink to 1 text + 1 vision layer with random weights so the test
247
# runs on any CI GPU and skips the ~56 GiB weight download. The test
248
# only validates encoder CG capture/replay, not output quality.