pytorch-rocm (2.12.1+dfsg-0ubuntu1) stonking; urgency=medium

  [ Talha Can Havadar ]
  * Rebuild for ROCm support (non-free).
    The section of this package is changed from main to contrib
    due to non-free build-dependencies.
  * This changelog was automatically generated by src:pytorch via
    debian/rocmbuild.sh.
  * Please merge changes from the ROCm variant back into src:pytorch.
  * Upgraded as part of ROCm stack upgrade (LP #2164669)
  * d/control.rocm: relax libfmt-dev dependency >= 10.1.1
  * d/control.rocm: relax pkg-rocm-tools b-d to >= 0.9.8~
  * d/control.rocm: use mold for linking as lld causes timeouts on builds on LP
  * d/rules: ignore fmt_consteval for rocm builds as clang-22 and libfmt 10 is not compatible.
    This needs to be dropped once we have libfmt 11 available
  * d/rules: do not treat unused-command-line-arguments as error in clang
  * d/{p/2450,rules,control.rocm}: update b-d to use amd fork of LLVM.
    Add 2450-tensorexpr-llvm-jit-llvm23-linking-layer.patch to fix llvm-23
    build failures
  * d/control.rocm: only publish rocm packages for amd64
  * d/p/2460: guard nccl one-sided API behind device support
  * d/control.rocm: libamdhip64-dev required with fixed __no_inline__

  [ Mo Zhou ]
  * Get rid of tensorpipe from build dependencies.
  * Get rid of tensorpipe for cuda/rocm variants as well.
  * [ROCM] try to fix FTBFS. Gbp-Dch: Ignore
  * partial revert of the previous commit. Gbp-Dch: Ignore
  * Reword cuda/rocm build scripts.
  * [CUDA] Attempt to fix CUDA build but we still miss some dependencies.

 -- Talha Can Havadar <talha.can.havadar@canonical.com>  Fri, 25 Sep 2026 09:01:27 +0200
