From 1b98ae1060b765f2efe22540f52691b8c00a83f1 Mon Sep 17 00:00:00 2001 From: Sayak Paul Date: Sat, 22 Aug 2026 13:13:24 +0530 Subject: [PATCH] [tests] skip hunyuan framepack group offloading tests (#14551) skip hunyuan framepack group offloading tests --- .../hunyuan_video/test_hunyuan_video_framepack.py | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/tests/pipelines/hunyuan_video/test_hunyuan_video_framepack.py b/tests/pipelines/hunyuan_video/test_hunyuan_video_framepack.py index 3187aa88dd..ace06371be 100644 --- a/tests/pipelines/hunyuan_video/test_hunyuan_video_framepack.py +++ b/tests/pipelines/hunyuan_video/test_hunyuan_video_framepack.py @@ -389,6 +389,14 @@ class HunyuanVideoFramepackPipelineFastTests( # this test because of MHA (example: HunyuanDiT because of AttentionPooling layer). pass + @unittest.skip("The image_encoder uses SiglipVisionModel, which does not support group offloading.") + def test_pipeline_level_group_offloading_inference(self): + # Same root cause as the sequential CPU offloading skips above: the attention layer is a + # torch.nn.MultiheadAttention, which passes `self.out_proj.weight` to + # `torch.nn.functional.multi_head_attention_forward` instead of calling `self.out_proj`. The leaf-level + # onload hook on `out_proj` is therefore never triggered and its weights stay on the CPU. + pass + # TODO(aryan): Create a dummy gemma model with smol vocab size @unittest.skip( "A very small vocab size is used for fast tests. So, Any kind of prompt other than the empty default used in other tests will lead to a embedding lookup error. This test uses a long prompt that causes the error."