Bug#1121021: rocblas: FTBFS on ppc64el
Control: block 1121021 by 1122634 [Cordell Bloor] > I'll double-check that and then file the request. The removal request is now pending in https://bugs.debian.org/1122634 >. -- Happy hacking Petter Reinholdtsen
Bug#1121021: rocblas: FTBFS on ppc64el
On Thu, 11 Dec 2025 09:37:29 +0100 Petter Reinholdtsen wrote: > One file a BTS report to the ftp masters asking for the arch binaries to > be removed. That is indeed the next step. I believe that with my upload of spfft 1.1.1-5 today, all reverse dependencies that had optional-at-build-time rocblas dependencies have now been updated to drop ROCm support on ppc64el. The remaining reverse dependencies require rocblas, and so should be removed along with it. I believe we need to remove the ppc64el binaries for rocblas, hipblas, rocsolver, hipsolver, and miopen. Additionally, we'll need to remove libggml0-backend-hip for ggml. That one is unique among the rocblas reverse dependencies in that rocblas is optional-at-build-time, but it produces a separate binary package. I'll double-check that and then file the request. Sincerely, Cory Bloor
Bug#1121021: rocblas: FTBFS on ppc64el
[Christian Kastner] > So if rocblas on ppc64el should no longer be supported, then this state > needs to be cleared (I don't immediately recall the process, I think the > RT is responsible). One file a BTS report to the ftp masters asking for the arch binaries to be removed. See https://bugs.debian.org/1122151 > for a randomly picked example. -- Happy hacking Petter Reinholdtsen
Bug#1121021: rocblas: FTBFS on ppc64el
Hi Cory, On 2025-11-20 07:04, Cordell Bloor wrote: > On 2025-11-19 09:18, Florent 'Skia' Jacquet wrote: > Thanks for looking into this. Instability in the compiler has been an > ongoing theme with ROCm on ppc64el. > > Christian, your ggml package is the primary reverse dependency affected. > While amd64 and arm64 can be binNMU'd, I think we might have trouble > with ppc64el for a while. As I understand it, our options are basically > to fix the bug in LLVM, find a workaround in rocblas, or RM rocblas > rdepends on ppc64el. Do you have any preferences on how to proceed? Seeing as all other ROCm packages have migrated to testing now, rocblas remains the last show-stopper for ggml. Regardless of ggml (which no longer depends rocblas on ppc64el), rocblas won't migrate to testing without a successful ppc64el build, as it has previously successfully built. So if rocblas on ppc64el should no longer be supported, then this state needs to be cleared (I don't immediately recall the process, I think the RT is responsible). Best, Christian
Bug#1121021: rocblas: FTBFS on ppc64el
Hi, On 2025-11-20 07:04, Cordell Bloor wrote: > Christian, your ggml package is the primary reverse dependency affected. > While amd64 and arm64 can be binNMU'd, I think we might have trouble > with ppc64el for a while. As I understand it, our options are basically > to fix the bug in LLVM, find a workaround in rocblas, or RM rocblas > rdepends on ppc64el. Do you have any preferences on how to proceed? I will be dropping the HIP backend from ggml over the weekend, as there are too many other issues with the ROCm stack open [1], and continuing to depend on it would impede ggml's migration to testing. The HIP backend will be re-added once librocblas4 has migrated to testing. Best, Christian [1]: I'll start a separate thread on this over the weekend
Bug#1121021: rocblas: FTBFS on ppc64el
Control: forwarded -1 https://github.com/llvm/llvm-project/issues/168727 Hi Skia, On 2025-11-19 09:18, Florent 'Skia' Jacquet wrote: I've had a quick look at the FTBFS on ppc64el, and this is actually a clang crash. The bug has been reported with more details upstream here: https://github.com/llvm/llvm-project/issues/168727 Thanks for looking into this. Instability in the compiler has been an ongoing theme with ROCm on ppc64el. Christian, your ggml package is the primary reverse dependency affected. While amd64 and arm64 can be binNMU'd, I think we might have trouble with ppc64el for a while. As I understand it, our options are basically to fix the bug in LLVM, find a workaround in rocblas, or RM rocblas rdepends on ppc64el. Do you have any preferences on how to proceed? Sincerely, Cory Bloor
Bug#1121021: rocblas: FTBFS on ppc64el
Source: rocblas Severity: serious Tags: ftbfs Justification: fails to build from source (but built successfully in the past) X-Debbugs-Cc: [email protected], [email protected], [email protected] User: [email protected] Usertags: ppc64el Hello there, I've had a quick look at the FTBFS on ppc64el, and this is actually a clang crash. The bug has been reported with more details upstream here: https://github.com/llvm/llvm-project/issues/168727 Cheers Skia
