Summary
Starting with the 7.14.0a20260609 nightly tarball, libhsa-runtime64 segfaults during hsa_init on gfx1151 (Radeon 8060S, Strix Halo). The 7.14.0a20260608 nightly works fine. The crash happens deep in rocr::AMD::GpuAgent::InitDma() (control jumps to a bogus address 0x100000001), so any HIP program that enumerates devices dies immediately.
Affected builds
Tarballs from https://therock-nightly-tarball.s3.amazonaws.com/therock-dist-linux-<target>-<version>.tar.gz. Discriminating by the libamdhip64.so build hash (the version string in release notes is unreliable):
| Nightly date |
libamdhip64 build hash |
gfx1151 result |
| 7.14.0a20260607 |
d34cbb6409 |
works |
| 7.14.0a20260608 |
d34cbb6409 |
works |
| 7.14.0a20260609 |
1b2a555677 |
segfault |
| 7.14.0a20260610 |
1b2a555677 |
segfault |
The regression boundary is precisely June 8 → June 9.
Environment
- GPU: AMD Radeon 8060S Graphics, gfx1151 (Strix Halo). Also reproduced on gfx1150.
- Kernel:
6.14.0-1019-oem (in-kernel amdgpu/amdkfd, no amdgpu-dkms), /dev/kfd present.
- System ROCm: 7.14.0.
- ROCm userspace under test: TheRock nightly S3 tarballs (userspace-only, no kernel components).
Reproduction
LD_LIBRARY_PATH=$PWD/lib ./llama-cli -m Qwen3-0.6B-Q4_0.gguf -ngl 99 -p "hi" -n 8 -no-cnv
(Any HIP program that calls hipGetDeviceCount / triggers hsa_init will crash the same way.)
gdb backtrace
Crash in libhsa-runtime64.so.1 — control transfers to 0x100000001
rocr::AMD::GpuAgent::InitDma()
hsa_init
hipGetDeviceCount
ggml_cuda_init
Isolation
- Swapping only
libhsa-runtime64.so* from the 20260608 (or system 7.14.0) build into the 20260609 tarball fixes the crash.
- Swapping only
libamdhip64.so* does not fix it.
So the regression is isolated to libhsa-runtime64 between the 20260608 (d34cbb6409) and 20260609 (1b2a555677) builds.
Expectation
hsa_init should succeed on gfx1151/gfx1150 with the in-kernel amdkfd on kernel 6.14, as it did through 20260608.
Summary
Starting with the
7.14.0a20260609nightly tarball,libhsa-runtime64segfaults duringhsa_initon gfx1151 (Radeon 8060S, Strix Halo). The7.14.0a20260608nightly works fine. The crash happens deep inrocr::AMD::GpuAgent::InitDma()(control jumps to a bogus address0x100000001), so any HIP program that enumerates devices dies immediately.Affected builds
Tarballs from
https://therock-nightly-tarball.s3.amazonaws.com/therock-dist-linux-<target>-<version>.tar.gz. Discriminating by thelibamdhip64.sobuild hash (the version string in release notes is unreliable):The regression boundary is precisely June 8 → June 9.
Environment
6.14.0-1019-oem(in-kernelamdgpu/amdkfd, no amdgpu-dkms),/dev/kfdpresent.Reproduction
(Any HIP program that calls
hipGetDeviceCount/ triggershsa_initwill crash the same way.)gdb backtrace
Isolation
libhsa-runtime64.so*from the 20260608 (or system 7.14.0) build into the 20260609 tarball fixes the crash.libamdhip64.so*does not fix it.So the regression is isolated to
libhsa-runtime64between the 20260608 (d34cbb6409) and 20260609 (1b2a555677) builds.Expectation
hsa_initshould succeed on gfx1151/gfx1150 with the in-kernel amdkfd on kernel 6.14, as it did through 20260608.