(bugfix) mori bug on RoCM714

[QizhouZhang97]

Fix MORI build + multi-process topology crash on ROCm 7.14

Summary

Two fixes that let MORI build and run on ROCm 7.14 (no regression on 7.2).

1. CMake: missing numa::numa target

ROCm 7.14's hsakmt-config.cmake declares that hsakmt::hsakmt links numa::numa, but ships no find_package/finder to define it, so configuration aborts:

The link interface of target "hsakmt::hsakmt" contains: numa::numa but the target was not found.

Fix: define a numa::numa IMPORTED target from the system libnuma before find_package(hsakmt).

2. Runtime: rsmi_init() crash under multi-process (rsmi symbol split-brain)

On the ROCm 7.14 docker, the rsmi_* symbols resolve to two different shared objects in the same process: some calls (e.g. rsmi_init, rsmi_is_P2P_accessible) bind to libamd_smi.so (amdsmi), while the rest (rsmi_num_monitor_devices, rsmi_dev_pci_id_get, …) bind to librocm_smi64. amdsmi sits in the global symbol scope and interposes the subset of the rsmi ABI it re-exports.

The two libraries keep separate global singletons, so rsmi_init() initializes one while rsmi_num_monitor_devices() queries the other (uninitialized) → RSMI_STATUS_INIT_ERROR, and the check's exit(-1) kills every rank. This does not happen on ROCm 7.2, where all symbols resolve to a single library.

Fix: pin the library ourselves. TopoSystemGpu::Load() now dlopens librocm_smi64 with RTLD_NOW | RTLD_LOCAL and resolves every rsmi_* symbol from that one handle, so all calls hit the same instance and amdsmi can no longer interpose. The link-time dependency on librocm_smi64 is removed from CMake, which also means the library is loaded lazily, after torch has initialized HIP — its NEEDED libamdhip64 then resolves to the already-loaded copy by SONAME instead of pulling in a second HIP runtime.

An optional MORI_ROCM_SMI_PATH env var overrides the library path.

Test plan

• pip install . --no-build-isolation builds on ROCm 7.14.
• 8-rank test_dispatch_combine_internode.py --cmd stress completes 5000/5000 dispatch + combine with no RSMI_STATUS_INIT_ERROR (no LD_PRELOAD needed).
• No regression on ROCm 7.2.

[jhchouuu]

Overall: Right direction, should land — but please don't remove the P2P topology capability in this PR.

✅ CMake numa::numa fix — LGTM. Defining the imported target from system libnuma to work around ROCm 7.14's missing target is correct.

✅ rsmi → HIP for bus id — correct. This avoids rocm-smi's cross-process mutex contention under /dev/shm/rocm_smi_* (concurrent rsmi_init() races return RSMI_STATUS_INIT_ERROR, and ROCM_SMI_CHECK's exit(-1) kills every rank). HIP's query APIs are read-only and multi-process safe.

⚠️ Don't drop the GPU↔GPU P2P topology. The crash root cause is rsmi's process safety, not P2P collection itself. The rack solution will need this P2P graph (link type/hops/weight) for scale-up routing and cost modeling. Suggest either:

1. (Preferred) keep it, swap the source — rebuild via hipDeviceCanAccessPeer + hipDeviceGetP2PAttribute; or
2. (Minimum) keep the TopoNodeGpuP2pLink struct and p2ps fields with a TODO, drop only the rsmi collection.

Removing the capability is a rack-roadmap decision and shouldn't ride along in a bugfix.

[jhchouuu]

LGTM need minor modification