Package triton-llvm is failing to build on aarch64 (and being retried) due to error “Output limit exceeded”. This hit the nixos-unstable channel.
Sadly one of those packages marked “big-parallel”
track record; Making sure you're not a bot!
Package triton-llvm is failing to build on aarch64 (and being retried) due to error “Output limit exceeded”. This hit the nixos-unstable channel.
Sadly one of those packages marked “big-parallel”
track record; Making sure you're not a bot!
@bertp better to raise Build failure issue on nixpkgs - GitHub · Where software is built then maintainers or whoever is managing staging-next can take a look.
I debated upstreaming this fix:
triton-llvm: build with BUILD_SHARED_LIBS to reduce output size
For x86_64-linux the directory tree size went from 1.8G down to ~200M.
diff --git a/pkgs/by-name/tr/triton-llvm/package.nix b/pkgs/by-name/tr/triton-llvm/package.nix
index b8072db53644..38cc399601c4 100644
--- a/pkgs/by-name/tr/triton-llvm/package.nix
+++ b/pkgs/by-name/tr/triton-llvm/package.nix
@@ -128,6 +128,7 @@ stdenv.mkDerivation (finalAttrs: {
(lib.cmakeBool "LLVM_INCLUDE_DOCS" (buildDocs || buildMan))
(lib.cmakeBool "MLIR_INCLUDE_DOCS" (buildDocs || buildMan))
(lib.cmakeBool "LLVM_BUILD_DOCS" (buildDocs || buildMan))
+ (lib.cmakeBool "BUILD_SHARED_LIBS" (!stdenv.hostPlatform.isStatic))
# Way too slow, only uses one core
# (lib.cmakeBool "LLVM_ENABLE_DOXYGEN" (buildDocs || buildMan))
(lib.cmakeBool "LLVM_ENABLE_SPHINX" (buildDocs || buildMan))
But @Lun mentioned that this might break downstream torch that might link to multiple LLVM versions.
Sorry for the miscommunication there, didn’t mean to block it, just wanted to get a chance to test it once ready.
Have turned it into this draft, will triton-llvm: build with BUILD_SHARED_LIBS to reduce output size by LunNova · Pull Request #500596 · NixOS/nixpkgs · GitHub mark ready once some long rebuilds up to leaf packages are done and I can smoke test.
Thanks, fwiw I didn’t think you were blocking it. It’s just that I got a bit tired of waiting for cuda rebuilds on my puny machine and didn’t get the energy to run those again. Thanks <3
Thanks for taking a look and proposing a fix!
I’ll create a “build failure” issue on the nixpkgs repo next time I encounter something similar.
Landed and then reverted because it ended up breaking keras: triton-llvm: build with BUILD_SHARED_LIBS to reduce output size by LunNova · Pull Request #500596 · NixOS/nixpkgs · GitHub
Landed this stopgap measure instead: triton-llvm: Set LLVM_TOOL_LLVM_DRIVER_BUILD=ON to use a single symlinked binary by LunNova · Pull Request #501365 · NixOS/nixpkgs · GitHub which should by now have reached unstable and resolved the immediate issue.
Could still do with a followup to try to reduce the size further, this only got us 500MiB of headroom, unsure if I’ll get around to looking at that any time soon.