fix(rocm): add gfx1151 support and expose AMDGPU_TARGETS build-arg

[keithmattix]

Summary

Adds gfx1151 (AMD Strix Halo / Ryzen AI MAX) to the default AMDGPU_TARGETS list and fixes a bug where custom GPU architecture targets passed via CMAKE_ARGS were silently overridden.

Fixes #9374

Problem

When building the llama-cpp backend for a non-default GPU architecture (e.g. gfx1151), the documented approach of passing -DAMDGPU_TARGETS=gfx1151 through CMAKE_ARGS does not work:

docker buildx build -f backend/Dockerfile.llama-cpp \
    --build-arg BUILD_TYPE=hipblas \
    --build-arg CMAKE_ARGS="-DAMDGPU_TARGETS=gfx1151" \
    ...

The Makefile's hipblas block (line 37) appends its own -DAMDGPU_TARGETS=$(AMDGPU_TARGETS) using the default arch list after the user's CMAKE_ARGS value. Since CMake uses the last -D value for a given variable, the user's target is silently overridden.

At runtime, this causes the gRPC backend to crash with a GPF in libamdhip64.so:

hip_fatbin.cpp:687: No compatible code objects found for: gfx1151
traps: grpcpp_sync_ser[...] general protection fault ... in libamdhip64.so.7.2.70201

This was initially misdiagnosed as a gRPC/HIP threading conflict (#9374), but AMD_LOG_LEVEL=4 tracing revealed the binary simply didn't contain code objects for the target GPU.

Fix

1. Add gfx1151 to the default AMDGPU_TARGETS in backend/cpp/llama-cpp/Makefile. ROCm 7.2.1 ships Tensile libraries for gfx1151, so it should be included in default builds.

2. Expose AMDGPU_TARGETS as an ARG/ENV in Dockerfile.llama-cpp so users can override the target list correctly:

docker buildx build -f backend/Dockerfile.llama-cpp \
    --build-arg BUILD_TYPE=hipblas \
    --build-arg AMDGPU_TARGETS=gfx1151 \
    ...

This sets the Make variable AMDGPU_TARGETS before the Makefile's ?= conditional assignment, so it takes precedence.

Testing

Tested on AMD Ryzen AI MAX+ 395 (gfx1151), ROCm 7.2.1, Ubuntu 24.04, kernel 6.17.0-1017-oem:

Scenario	Result
Before fix: CMAKE_ARGS=\"-DAMDGPU_TARGETS=gfx1151\"	GPF — binary missing gfx1151 code objects
After fix: --build-arg AMDGPU_TARGETS=gfx1151	PASS — model loads, inference works
Stock llama.cpp (same machine)	PASS — baseline control

Signed commits

Yes, I signed my commits.