[Bug]: MM performance regression from upgrading to torch 2.9

[ywang96]

Your current environment

The output of python collect_env.py

Your output of `python collect_env.py` here

🐛 Describe the bug

There's a significant performance regression on nn.Conv3d which affects all multimodal models that use this module (Qwen3-VL, GLM4.1V, etc). This was caught by the unusually long multimodal profiling time in the latest nightly build, and a similar issue can be found at pytorch/pytorch#166122

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.