rocm-runtime is still using gcc, is about to become broken by gcc13

[ghost]

Heads up: even though it tries to overrideCC with clang, our rocm-runtime expression still uses gcc for a ton of stuff.

Try changing default-gcc-version in pkgs/top-level/all-packages.nix to 13 and you'll see the breakage. I'm about to merge

• default-gcc-version: 12 -> 13 if !isDarwin && !isAarch64 #268097

which does this, and it's going to break rocmPackages. I tried to fix it, but the problem is very deep: you guys think you're using only clang, but in fact you're not.

Wanted to give you a heads-up.

cc: @lovesegfault @Madouura @Flakebi

[ghost]

I recommend you guys merge a PR which pins rocmPackages to gcc12Stdenv wherever you aren't really using llvm/clang. I tried to do this, but it's far from clear where this needs to be done.

[ghost]

Also thank you all for the awesome hard work you did to make this happen. I was totally shocked when I was able to get 8*MI25s to run Falcon-180B in llama.cpp with only a few days of effort. Obviously that's because you folks had already put in months of work to make it so smooth...

[Madouura]

you guys think you're using only clang, but in fact you're not.

Unfortunately, I'm aware. I was working on making rocmPackages use only libc/libcxx/etc, but it's rather WIP.

I tried to do this, but it's far from clear where this needs to be done.

It's likely all the way to the root - ROCm LLVM. It has a GCC backend which is used.

Any chance you could give me the error log for rocm-runtime so I don't have to rebuild ROCm LLVM and a bunch of other stuff?

[ghost]

Unfortunately, I'm aware.

Ah, that's actually somewhat of a relief. At least the problem is known.

Log is attached. I'm happy to test fixes; I have plenty of compute. After the new year I will be working on making the buildfarm available to other projects; you'll be on the shortlist if you're interested.
8q1vy1pyk4079vjq3pfqkjmffdvksif2-rocm-llvm-runtimes-5.7.1.drv.log.gz

[ghost]

Also if you need to get my attention I hang out in #tvl on irc.hackint.org. It's bridged to Matrix.

[Madouura]

8q1vy1pyk4079vjq3pfqkjmffdvksif2-rocm-llvm-runtimes-5.7.1.drv.log.gz

Okay, this is rocm-llvm-runtimes, not rocm-runtime. This might be a slightly easier fix then.

I'm happy to test fixes; I have plenty of compute. After the new year I will be working on making the buildfarm available to other projects; you'll be on the shortlist if you're interested.

I appreciate that, when I put the PR through I might request that you do a nixpkgs-review, possibly a few times if the dependency hell is really bad.

[Madouura]

Actually, I think I may take your suggestion and just pin a GCC version until a new ROCm release.
This will likely be fixed upstream.

[Madouura]

Fixed by #271954
(Didn't auto close for some reason?)