Remove the dummy dependency hack and use extra-ling-args instead.

[LaurentMazare]

This is only available in cargo 1.50 and in nightly mode at the moment. This can be tried out via:

cargo +nightly run --example basics -Zextra-link-arg --verbose

[LaurentMazare]

This should be stable in the upcoming 1.56 Rust release. rust-lang/cargo#9557

[riesha]

any update on this?

[oovm]

This feature should be able to merge, any update?

[LaurentMazare]

Sadly things have changed a bit since this PR was crafted and now the flags are only used for the crate specifying them and not by dependent crates (or at least that's my understanding). This means that the change here doesn't work, I've put in #642 a similar change that specifies the flags at the tch level rather than torch-sys, so this should work for binaries, examples, and tests from the tch crate but not for crates that use tch. Such crates will also need to specify the flag whereas with the current hack there is no need for any custom configuration on dependent crates.
It's a bit unclear what the best thing to do is here so we should probably think about it a bit more.

[LaurentMazare]

Closing this one as it doesn't seem possible to pass flags to the final linking step from an intermediary crate library (see this comment). The current best practice is for final crates to have a build.rs that set the proper flags, e.g. in diffusers-rs.

Related issues: cargo#5077. wip for supporting as-needed rust#99424.

[bharrisau]

FYI - I have the following in my project that seems to work on nightly for my needs with pytorch/TensorRT.

#![feature(native_link_modifiers_as_needed)]
#[link(name = "torchtrt_runtime", kind = "dylib", modifiers = "-as-needed")]
extern {}

Once the feature is stabilised you might be able to println!("cargo:rustc-link-lib=dylib:-as-needed=torch_cuda"). I couldn't work out how to use unstable features from cargo build.rs. I guess you could just gate the #[link()] behind a #[cfg(tch-cuda)] and use println!("cargo:rustc-cfg=tch-cuda") from the build script.

[LaurentMazare]

Thanks for pointing that out @bharrisau that could indeed be super helpful in removing the dummy cuda dependency and providing a less hacky linking setup. My previous post on this thread has a link to the stabilization tracking issue for this but sadly it seems a bit stale a the moment.