fix issue 23977

[zihaomu]

Overview

This solution is taken from #23977. Credits to @mapatel-amd for identifying the root cause and proposing the fix.

Test Hard ware env:

Strix Halo 128G

Test Software env:

  Local environment:
  - OS: Ubuntu 24.04.4 LTS, Linux 6.17.0-23-generic, x86_64
  - Hardware: AMD Ryzen AI MAX+ 395 with Radeon 8060S Graphics
  - ROCm device: gfx1151 / Radeon 8060S Graphics
  - Reported ROCm memory: 122880 MiB
  - HIP: 7.12.60610-2bd1678d3d
  - AMD clang: 22.0.0git
  - CMake: 3.28.3
  - Ninja: 1.11.1
  - Base commit: 59917d392
  - Test branch commit: 1bf74b70e

Build configuration:

  PATH=/opt/rocm/core-7.12/bin:$PATH \
  HIPCXX=/opt/rocm/core-7.12/lib/llvm/bin/clang \
  HIP_PATH=/opt/rocm/core-7.12 \
  cmake -S . -B build-hip-issue23977 -G Ninja \
    -DGGML_HIP=ON \
    -DGPU_TARGETS=gfx1151 \
    -DCMAKE_BUILD_TYPE=Release \
    -DLLAMA_BUILD_TESTS=ON
  cmake --build build-hip-issue23977 --target llama-bench -j 16
  cmake --build build-hip-issue23977 --target test-backend-ops -j 16

A/B validation method:

I used a small local probe, not committed to the repo, to query:

• ggml_backend_cuda_host_buffer_type()
• ggml_backend_dev_supports_buft(dev, ROCm_Host)

The same machine and same HIP build configuration were used before and after the patch.

A/B result:

Before:

  device=0 name=ROCm0 type=2 host_buft=ROCm_Host supports_host_buft=0

After:

  device=0 name=ROCm0 type=2 host_buft=ROCm_Host supports_host_buft=1

[ggml-gh-bot]

Hi @zihaomu, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• Multiple open PRs from a new contributor: We limit new contributors (those without a previously merged PR) to 1 open PR at a time. You currently have 2 open PRs.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.