[dmitrii-galantsev]

Motivation

TheRock's rocm-sdk-core wheel carries the amdsmi Python module only at _rocm_sdk_core/share/amd_smi (since #9340). Copied into site-packages, import amdsmi; amdsmi.amdsmi_init() fails with libamd_smi.so.27: cannot open shared object file unless torch was imported first, because only PyTorch's preload makes the library findable.

Technical Details

_load_library() in amdsmi_wrapper.py (mirrored in tools/generator.py):

• _rocm_sdk_core lookup. When no system library is found or the one found is refused, the loader takes lib/libamd_smi.so.27 from the _rocm_sdk_core package (sys.modules, else sys.path). It checks paths only and imports nothing, so import guards like rocprofiler-compute's cannot break it (test(amdsmi): cover loader resilience under a restrictive import guard #9879).
• _AMDSMI_ROCM_SDK_CORE flag. TheRock sets it in the copy it ships in rocm-sdk-core (Mark the amdsmi Python module shipped in rocm-sdk-core TheRock#9578). That copy then uses the relocatable path and _rocm_sdk_core only, never a system libamd_smi.so from another release. The rpm/deb copy keeps its order and only gains the fallback.
• Reuse before load. Each candidate is first opened with RTLD_NOLOAD. PyTorch 2.14's _amdsmi_cdll_hook (in the ROCm 10.1.0 wheels; narrowed upstream in [ROCm] Preserve amdsmi library selection during import pytorch/pytorch#197482) redirects any libamd_smi.so* load to $ROCM_PATH/lib/libamd_smi.so first. With ROCM_PATH set and a system ROCm installed, the module loaded that second copy and both torch.cuda.device_count() and amdsmi reported 0 GPUs. Asked with RTLD_NOLOAD, the hook can only return a copy already in the process.
• Library version check. SONAME 27 changed struct sizes between 27.1 and 27.2: amdsmi_fabric_info_t grew from 320 to 552 bytes when its union gained v2, and 10.2 nightlies built before revert(amdsmi): restore single-field driver version info #13926 have a 1792-byte amdsmi_driver_info_t that amdsmi_get_gpu_driver_info() zeroes in full, against the 768 bytes a 27.1 module allocates. A library newer than the module's _version.py is closed and skipped, so it cannot write past the module's buffers. An older one warns. AMDSMI_LIB_OVERRIDE still loads anything.
• Version stays in sync. get_version_from_file makes CMake reconfigure when amdsmi.h changes, so _version.py cannot lag the library in an incremental build.
• Docs. docs/packaging.md, docs/install/install.md and py-interface/README.md describe the new loader order, the flag, RTLD_NOLOAD reuse and the version check.

This does not make share/amd_smi pip-installable again. A module installed from there would not depend on rocm-sdk-core and would go stale on the next ROCm upgrade.

Issue Tracking

JIRA ID: ROCM-33336
Related: ROCM-31613, ROCm/TheRock#9226 (with #14409), ROCm/TheRock#3054

Test Plan

• Python loader tests (tests/python/test_abi_compat.py): flag on and off, the fallback, RTLD_NOLOAD reuse, newer and older libraries, closing a refused library, unusable sys.path entries. Each new check was disabled once to confirm its test fails.
• The 10.1.0 rocm-sdk-core module patched with this loader, on the shipped wheels (rocm-sdk-core 10.1.0, torch 2.14.0+rocm10.1.0), 2× MI210, host with a 10.2 ROCm in ld.so.cache and ROCM_PATH set:
  • copied into site-packages and via PYTHONPATH;
  • imported alone, before torch and after torch;
  • also with /opt and ld.so.cache hidden;
  • also with the 10.2 library loaded first.
• The amd-smi CLI from an install tree and from rocm-sdk-core's entry point.
• A version bump in amdsmi.h followed by ninja: reconfigures and updates _version.py.
• The DESTDIR install: the site-packages copy matches share/amd_smi, _version.py included.

Test Result

• Python tests: 55 passed.
• Module, every case: one libamd_smi mapped, torch and amdsmi each report 2 GPUs.
• With a newer 10.2 library already loaded before PyTorch 2.14's hook imports the module, the module refuses it with the version error rather than binding it.
• amd-smi version and amd-smi list: work from both locations.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md.

[therock-pr-bot]

✅ All Policy Checks Passed

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

🎉 All policy checks passed!

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[Copilot]

🟡 Changes recommended

Library-loading paths still let PyTorch hooks load a system AMD SMI copy instead of reusing or opening the intended SDK library, risking duplicate-library GPU-enumeration failures.

3 open issues

• Preserve SDK path when no library is preloaded · New
• Reuse loaded libraries in SONAME branch · New
• Do not skip empty sys.path entries · New

What changed in this PR

Updates AMD SMI’s Python loader so modules copied from TheRock’s rocm-sdk-core wheel can locate the SDK library outside share/amd_smi, while adding library-version checks.

Changes:

• Adds SDK package discovery and SDK-only loading behavior, mirrored in the wrapper generator.
• Adds reuse-before-load attempts, version compatibility checks, and loader regression tests.
• Triggers CMake reconfiguration when the version header changes to keep Python version metadata synchronized.

File	Description
projects/​amdsmi/​tools/​generator.py	Generates the updated SDK discovery and library-loading logic.
projects/​amdsmi/​tests/​python/​test_abi_compat.py	Tests SDK fallback, library reuse, version checks, and unusable search paths.
projects/​amdsmi/​py-interface/​amdsmi_wrapper.py	Implements SDK discovery, loading controls, and version validation.
projects/​amdsmi/​cmake_modules/​utils.cmake	Tracks the version header as a configure dependency.
projects/​amdsmi/​CHANGELOG.md	Documents copied-module support and compatibility safeguards.

_{Review effort: Balanced}