[HereThereBeDragons]

Cherry pick of 37ca1c5
JIRA ID: AIRDEL-111

Motivation

ROCm SMI now ships with deprecation notices (ROCm/rocm-systems#13131), so build it again instead of skipping it. Partially reverts #6852.

This is intended to add rocm-smi back in for 10.2, but only for external users who didn't get the deprecation notice. This is meant to be a final build with a warning in the output for users to move over to amd-smi.

Technical Details

File	Change
base/CMakeLists.txt	Drop -DROCM_SMI_DISABLE=True
rocm-sdk-core setup.py, _cli.py	Restore the rocm-smi console script
rocm_sdk/_dist_info.py	Restore the rocm_smi64 core library entry
rocm_sdk/tests/core_test.py	Restore the rocm-smi console-script test
amdrocm-core / amdrocm-developer-tools postinst + prerm	Add rocm-smi next to amd-smi in the update-alternatives lists (dropped by #7672 while rocm-smi-lib was disabled)

Not restored: rocm_smi64 in the PyTorch wheel preload list (external-builds/pytorch/build_prod_wheels.py).

Test Plan

• CI build + rocm-sdk core tests
• python3 -m pytest build_tools/tests/sdk_targets_test.py
• Render the 4 postinst/prerm templates (deb), bash -n, run against a fake prefix with a stub update-alternatives

Test Result

• sdk_targets_test.py: 5 passed
• _dist_info registers rocm_smi64 in core
• All 4 scripts pass bash -n; postinst installs and prerm removes /usr/bin/rocm-smi alongside amd-smi

Issue Tracking

JIRA ID: ROCM-21935

[peymanr]

PR Pundit review

(testing PR Pundit Code Review agent. Not all comments may be relevant. Please reach out to @peymanr for suggestions and feedback.)

#		Finding	Where
1	🟠	rocm_smi64 not restored to LINUX_LIBRARY_PRELOADS	external-builds/pytorch/build_prod_wheels.py:175
2	🔵	Shared rocm-smi alternative removed when either owning package uninstalled	build_tools/packaging/linux/template/scripts/amdrocm-developer-tools-prerm.j2:20
3	🔵	rocm_smi64 entry misplaced and required rather than optional	build_tools/packaging/python/templates/rocm/src/rocm_sdk/_dist_info.py:295

1. 🟠 Should fix: rocm_smi64 not restored to LINUX_LIBRARY_PRELOADS

external-builds/pytorch/build_prod_wheels.py:175

This PR undoes #6852 except for one change: 50dc7263a also removed "rocm_smi64" from LINUX_LIBRARY_PRELOADS in external-builds/pytorch/build_prod_wheels.py, and that entry is not restored. Now that librocm_smi64 is built and listed again (LibraryEntry("rocm_smi64", ...)), the generated torch/_rocm_init.py will not preload it with RTLD_GLOBAL. The preload was originally added in #4440 because PyTorch smoke tests failed with librocm_smi64.so.1: cannot open shared object file. Without it, wheel installs risk that failure again, or a fallback to a system /opt/rocm copy. Re-add "rocm_smi64" to LINUX_LIBRARY_PRELOADS to complete the revert.

Who is affected: Linux PyTorch wheels built with build_prod_wheels.py against ROCm Python packages from this PR, where torch or its ROCm libraries load librocm_smi64 at runtime.

Verified by running it.

Related past reviews: #898 on build_tools/packaging/python/templates/rocm/src/rocm_sdk/_dist_info.py, #2779 on build_tools/packaging/python/templates/rocm/src/rocm_sdk/tests/core_test.py

2. 🔵 Nit: Shared rocm-smi alternative removed when either owning package uninstalled

build_tools/packaging/linux/template/scripts/amdrocm-developer-tools-prerm.j2:20

rocm-smi is added to the update-alternatives lists of both amdrocm-core and amdrocm-developer-tools, but the binary ships in amdrocm-base. Both packages register the same path, and update-alternatives does not count registrations. So removing either metapackage runs --remove rocm-smi and deletes /usr/bin/rocm-smi, even though the binary and the other metapackage are still installed. The base scripts already do exactly this for amd-smi, rocminfo and rocm_agent_enumerator, so this PR adds one more instance of an existing pattern. A proper fix (register shared tools from one package, or from amdrocm-base) belongs in a separate cleanup.

Who is affected: Native deb/rpm users who install both amdrocm-core and amdrocm-developer-tools and then remove one of them lose the /usr/bin/rocm-smi link (as they already do for amd-smi/rocminfo).

Verified by running it.

Related past reviews: #7672 on build_tools/packaging/linux/template/scripts/amdrocm-developer-tools-postinst.j2

3. 🔵 Nit: rocm_smi64 entry misplaced and required rather than optional

build_tools/packaging/python/templates/rocm/src/rocm_sdk/_dist_info.py:295

LibraryEntry("rocm_smi64", ...) is inserted between rocm_kpack and rocdecode. Its sibling amd_smi is three lines below, so grouping the two SMI libraries would read better. Making the entry optional=True is not needed: on Linux rocm_smi_lib is now always built as part of base, which every core package includes. On Windows the empty DLL pattern already makes find_libraries skip it.

Who is affected: Readers and maintainers of _dist_info.py; this has no functional effect on any build.

Related past reviews: #5283 on build_tools/packaging/python/templates/rocm/src/rocm_sdk/_dist_info.py, #1191 on build_tools/packaging/python/templates/rocm/src/rocm_sdk/_dist_info.py, #5708 on build_tools/packaging/python/templates/rocm/src/rocm_sdk/_dist_info.py

_{Shared from a pilot of PR Pundit, an agentic code reviewer that builds and tests the change and reads the project's past reviews. Reviewed at 2f89a764; findings checked and posted by @peymanr; points already raised on this PR are left out.}