[Build][JIT] Support core-only ROCm SDK environments

[apinge]

Motivation

Since ROCm 7.14, ROCm can be installed via python packages and is no longer necessarily located under /opt/rocm. While rocm-sdk-core already provides the hip headers, runtime libraries, lld, and device bitcode, it does not ship hip-config.cmake, and its resource layout differs from the traditional /opt/rocm structure.

This creates two issues:

1. CMake cannot locate hip via find_package(hip). Although [Build] Find ROCm through rocm-sdk before falling back to /opt/rocm* #1142 added support for discovering a pip-installed ROCm sdk via rocm-sdk path --root, that mechanism depends on rocm[devel], which is not present in some images such as rocm/pytorch:rocm7.14.1_ubuntu24.04_py3.13_pytorch_release_2.12.0.

2. GPU JIT still searches for lld and device bitcode under /opt/rocm, which can lead to lld invocation failed in environments that only provide the core wheel.

This PR only covers source builds and GPU JIT in core-only ROCm environments. The default external linking flow for AOT artifacts remains affected by the missing development linker name for libamdhip64.so (see ROCm/TheRock#1205) and is left for future work.

Technical Details

1. This PR preserves the existing lookup priority for the standard hip cmake package discovery mechanisms: hip_DIR, ROCM_PATH, rocm-sdk path --root, and /opt/rocm. Only if hip-config.cmake cannot be found in any of those locations will it fall back to reading the hip headers and runtime library from the _rocm_sdk_core package in the Python environment used for the build, while also providing the hip::host and hip::amdhip64 targets required by the FlyDSL runtime.
2. This PR introduces FlyRocmModuleToBinaryPass as a replacement for upstream llvm's GpuModuleToBinaryPass, and FlyRocmSerializer as a derived replacement for upstream llvm's AMDGPUSerializer. By overriding compileToBinary in the derived serializer, it avoids the hardcoded linker path used upstream. It solves the same problem as Auto-discover MLIR ROCm toolkit from rocm-sdk Python wheels #568 , but without requiring extra symbolic links, which makes the user environment easier to maintain. This makes it possible to resolve both the traditional installation path

toolkit = /opt/rocm
linker  = /opt/rocm/llvm/bin/ld.lld
bitcode = /opt/rocm/amdgcn/bitcode

and the new rocm-sdk-core path

toolkit = <core>/lib/llvm
linker  = <core>/lib/llvm/bin/ld.lld
bitcode = <core>/lib/llvm/amdgcn/bitcode

Test Plan

• Verified in a ROCm 7.14 core-only python environment that bash scripts/build.sh -j64 .
• Passed all tests in scripts/run_benchmark.sh.
• Passed all tests in scripts/run_tests.sh except tests/unit/test_aot_export.py.

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/TheRock/blob/main/GOVERNANCE.md#pull-requests.

[Copilot]

Copilot review overview

🟢 Approval recommended

The fallback preserves existing discovery priority and has focused automated coverage.

Review effort: Balanced
Findings: None

What changed in this PR

Adds HIP discovery fallback for minimal rocm-sdk-core Python installations.

Changes:

• Creates compatible hip::host and hip::amdhip64 CMake targets from core-wheel assets.
• Adds a focused configuration test for fallback discovery.

File	Description
lib/​Runtime/​ROCm/​CMakeLists.txt	Implements core-wheel HIP fallback.
tests/​unit/​test_backend_cmake_defaults.py	Verifies fallback target paths.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.