[Feature]: Inductor partitioning should decide what ops to partition on dynamically

[ProExpertProg]

🚀 The feature, motivation and pitch

Currently, if use_inductor_graph_partition is set, we always partition on attention by making it cudagraph unsafe - instead we should respect splitting_ops (or likely make a new field). Use register_should_partition_rule from pytorch/pytorch#163310 and pytorch/pytorch#163395.

Alternatives

No response

Additional context

No response

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.

[baonudesifeizhai]

the latest torch i can get is torch: 2.9.0.dev20250909+cu128 ..... if you guys really in hurry i will try build torch from source

[ProExpertProg]

@baonudesifeizhai if you're working on this feel free to open a pr!