Sure, done: https://github.com/NixOS/nixpkgs/pull/478681

EDIT: merged by happysalada

Suppose you’ve got a project where different object files are compiled with different versions of NVCC/GCC. Under what conditions can they be linked into shared objects?

As an example, assume you’ve got a project which builds with an older version of CUDA, and you’re using stdenv (not cudaPackages.backendStdenv). The CUDA portion of the build would produce object files using an older, NVCC-compatible GCC (by virtue of how we wrap NVCC so it always sees a compatible GCC), but the rest would produce object files using the GCC provided by stdenv. Linking happens with some combination of stdenv’s linker, the linker corresponding to the bintools (I think) of the GCC available to NVCC, and NVLink.

Under what conditions would linking succeed?

I ask because I remember trying to use multiple versions of LLVM-produced object files and getting something along the lines of a bitstream version mismatch, but I don’t remember if that was because I was using LTO or something else.

In reply to @glepage:matrix.org
Yes, because this package is not available for cuda 12.8

After a shameful number of hours spend on building the latest ORT, I finally succeeded!
The PR has been opened since Oct. 10 2025!

Thanks to @yuyurureka who found the last missing patch to make the thing build!
Not thanks to Microslop for shipping broken software.
https://github.com/NixOS/nixpkgs/pull/450587

After a shameful number of hours spend on building the latest ORT, I finally succeeded!
The PR has been opened since Oct. 10 2025!

Thanks to @yuyurureka who found the last missing patch to make the thing build!
Not thanks to Microslop for shipping broken software.
https://github.com/NixOS/nixpkgs/pull/450587

EDIT: still fails with cudaSupport and rocmSupport :')

After a shameful number of hours spend on building the latest ORT, I finally succeeded!
The PR has been opened since Oct. 10 2025!

Thanks to @yuyurureka who found the last missing patch to make the thing build!
Not thanks to Microslop for shipping broken software.
https://github.com/NixOS/nixpkgs/pull/450587

EDIT: still fails with cudaSupport and rocmSupport :')
EDIT2: turns out they seem to have removed Rocm support (https://github.com/microsoft/onnxruntime/pull/25181)

NOTE ROCm Execution Provider has been removed since 1.23 release. Please Migrate your applications to use the MIGraphX Execution Provider

Thanks for the heads up regarding Madouura's situation.

Considering the amount of code removed, I wondered whether we could still do something.

Thanks for the review connor (burnt/out) (UTC-8). Anything left to consider before merging?

I didn't spot any major regressions.

hexa (UTC+1) suggested that it could be merged into master.

I'll try to build it on the other platforms.