Build ROCm against TheRock 10.0.0, the ROCm XLA uses

[wsmoses]

Enzyme-JAX's JAX bump (EnzymeAD/Enzyme-JAX#3368) fails the ROCm Reactant_jll build. The new XLA builds its MORI collective kernels (xla/backends/gpu/collectives/mori_kernels.cu.cc, under if_rocm_is_configured), and MORI calls __builtin_amdgcn_global_load_b128:

external/roc_mori/include/mori/collective/XLA/collectives_common.hpp:128:14: error: use of undeclared identifier '__builtin_amdgcn_global_load_b128'

Why a ROCm bump:

• Where the builtin lives: it is only in AMD's downstream clang; upstream LLVM has no such builtin. Our "7.1" toolchain is the TheRock nightly 7.10.0a20251103, whose clang doesn't have it.
• What XLA uses: XLA's hermetic ROCm is now the TheRock multiarch 10.0.0 release (third_party/gpus/rocm/rocm_redist.bzl, same sha256 as here).
• Checked locally: its clang (AMD clang 23.0.0git) compiles the builtin to global_load_dwordx4 / global_load_b128 for every target in TF_ROCM_AMDGPU_TARGETS: gfx908, gfx90a, gfx942, gfx950, gfx1030, gfx1100, gfx1101, gfx1200 and gfx1201.

Changes:

• Source: therock-dist-linux-multiarch-10.0.0.tar.gz, replacing the gfx120X nightly.
• Platform tag: gpu_version becomes 10.0 instead of 7.1, in the build matrix and in platform_augmentation.jl, which picks the ROCm build when an AMD driver is present. Reactant.jl doesn't refer to the tag.
• Sonames: the nightly-specific renames (*.so.7.1.25442-19ae9ff849) become a pattern that strips TheRock's build suffix. 10.0 names them *.so.7.15.26333-0000000. I checked the loop on the 10.0 files: the .so.7 links resolve.
• Clang resource directory: 10.0 ships clang/23; 7.1 shipped clang/22.
  • The cuda_wrappers overrides now read the version from the tarball.
  • Enzyme-JAX's workspace.bzl patches XLA's ROCm crosstool with a hardcoded lib/llvm/lib/clang/22/include: rocm_configure.bzl takes its realpath, and the crosstool declares it to the inclusion check. The first CI run failed there (/workspace/srcdir/lib/llvm/lib/clang/22 (No such file or directory)). The recipe now symlinks clang/22 → clang/23 and spells hipcc's -isystem paths through clang/22, so both declared spellings match what the compiler reports. Checked on the 10.0 tree: clang/22/include resolves to clang/23/include, including stddef.h and cuda_wrappers.
• amdgcn link: 10.0 already ships amdgcn -> lib/llvm/amdgcn. The recipe's own ln -s followed it and created lib/llvm/amdgcn/amdgcn, a link back to its own directory, so the second CI run failed globbing rocm_dist/amdgcn/** (Symbolic link loop). That ln is removed.
• Removed libraries: 10.0 has no libhipsolver_fortran and no libamd_comgr_loader (libamd_comgr.so.3 is still there). Their installs and the libhipsolver_fortran product are dropped; Reactant.jl uses neither. Every other installed library and product exists in 10.0.

Not verified here: the full BinaryBuilder build; that's what this PR's CI runs. Changing the ROCm toolchain also invalidates the cached host tools on that job, so it will hit the __throw_bad_array_new_length failure until #236 is merged.

🤖 Generated with Claude Code

https://claude.ai/code/session_016zErYp7upmqr4NHfhod9UD