[ryan-li-ryanllli-aws]

Summary

• add model.local_compile_backend with an inductor default
• use that backend for every enabled local compile region
• late-bind FlexAttention and block-mask helpers so custom device backends do not require a process-wide torch.compile patch
• preserve inductor-specific options only for inductor and keep the deterministic ROCm eager fallback

Why

TorchTitan-owned compile sites currently select the default backend before model configuration is available. A custom device backend therefore has to patch torch.compile process-wide, which also affects user and third-party compile calls. This change gives those sites one scoped backend setting while preserving current default behavior.

Test plan

• Python 3.11 plus nightly CPU PyTorch: 53 focused CPU tests passed
• formatting, flake8, pydoclint, codespell, AST, whitespace, merge-conflict, license, and link checks passed
• git diff --check passed

[ryan-li-ryanllli-aws]

Superseded by #13, which retains the lazy helper reset/rebinding design and includes HF config initialization, typed torch.compile calls, compatibility defaults, and engine propagation coverage.