Disable inductor config setter by default

[HDCharles]

This was requested in the huggingface x pytorch slack by @BenjaminBossan

This setting for Torchao should be disabled by default to avoid side effects (https://github.com/pytorch/ao/blob/24c966cb8931507ef389715bc5f19a11c28b0484/torchao/quantization/utils.py#L574)

torch._inductor.config.coordinate_descent_tuning = True
torch._inductor.config.coordinate_descent_check_all_directions = True
torch._inductor.config.force_fuse_int_mm_with_mul = True
torch._inductor.config.fx_graph_cache = True
torch._inductor.config.triton.unique_kernel_names = True
torch.set_float32_matmul_precision("high")

though the default performance will be worse, these settings can be set manually.

@SunMarc

[github-actions]

Hi 👋, thank you for opening this pull request! The pull request is converted to draft by default. When it is ready for review, please click the Ready for review button (at the bottom of the PR page).

[BenjaminBossan]

If the recommended settings generally work better, would it make sense to document that somewhere and tell users how to enable them?

[HDCharles]

If the recommended settings generally work better, would it make sense to document that somewhere and tell users how to enable them?

I've added a note to the .md file as well.

[HDCharles]

note: the failing test seems to be on trunk or at least unrelated, other PRs are also failing with teh same unrelated error: #36610

[HDCharles]

looks like 2 HTTPS errors?

[SunMarc]

Thanks for the update !

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[HDCharles]

This is ready to land @SunMarc @ArthurZucker

[ArthurZucker]

LGTM thanks for pushing through