rocm-specs

T

Sakura286 af542a7f67 python-triton: keep the nvidia C++ backend in the cmake build

Dropping nvidia from TRITON_CODEGEN_BACKENDS does not build the NVGPU/NVWS
dialect TableGen output under third_party/nvidia, but Triton core hard-
depends on it (TritonGPUTransforms via Passes.h, TritonInstrumentToLLVM),
so the wheel build failed with

  fatal error: 'nvidia/include/Dialect/NVGPU/IR/Dialect.h.inc' file not found
  fatal error: 'nvidia/include/Dialect/NVWS/IR/Dialect.h.inc' file not found

Keep the nvidia C++ libraries in the cmake build (they only need the
in-tree LLVM NVPTX target, no CUDA SDK, and stay offline) while still
packaging only the AMD Python backend, so neither ptxas nor the
proprietary libdevice.10.bc lands in the RPM.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

2026-06-11 10:00:25 +08:00

SPECS

python-triton: keep the nvidia C++ backend in the cmake build

2026-06-11 10:00:25 +08:00