fix: [pytorch] Resolve ROCm deps via RUNPATHs

[venkat1361]

Motivation

Commit 6a55012 added rocprofiler-sdk to PyTorch’s Linux startup preload list to fix nightly smoke tests after PyTorch/Kineto gained a runtime dependency on librocprofiler-sdk.so.1.

When PyTorch is launched under an external rocprofv3, the profiler may already have loaded and initialized one rocprofiler-sdk provider. PyTorch startup can then explicitly load the wheel-bundled rocprofiler-sdk, creating a duplicate-provider path in the same process. In that scenario, the rocprofv3 tool configuration path can be invoked after rocprofiler-sdk has already closed its configuration window, leading to:

ROCPROFILER_STATUS_ERROR_CONFIGURATION_LOCKED
Configuration request occurred outside of valid rocprofiler configuration period

This PR treats the original issue as a packaging/runtime dependency issue instead of an import-time preload issue. PyTorch’s ROCm-linked shared libraries should resolve their DT_NEEDED dependencies through relocatable wheel RUNPATHs.

JIRA ID: AIPROFSDK-969

Technical Details

This PR removes rocprofiler-sdk from PyTorch’s generated _rocm_init.py preload list.

After pytorch build step unpacks the generated torch wheel, finds ROCm-linked torch shared objects, removes non-relocatable absolute RUNPATH entries, appends $ORIGIN-relative RUNPATH entries pointing to the installed ROCm Python package library directories, and repacks the wheel before copying.

This lets dependencies such as librocprofiler-sdk.so.1 resolve through the dynamic loader instead of through explicit ctypes/rocm_sdk.initialize_process() preloading. If an external rocprofv3 has already loaded a compatible SDK provider with the same SONAME, the loader can use the already-loaded provider instead of PyTorch forcing another absolute SDK load.

This PR also updates package promotion logic so torch ELF RUNPATHs stay valid when ROCm package versions are promoted. Text files such as _rocm_init.py, version.py, and metadata were already updated during promotion, but RUNPATH strings embedded in ELF files also need their ROCm version fragments updated when promoting from nightly/RC versions to release versions.

Test Plan

• Added test in build_tools/packaging/tests/promote_keep_list_test.py

Test Result

• https://github.com/ROCm/TheRock/actions/runs/29043018157

Submission Checklist

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

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
🌿 Branch Name	✅ Pass	—
📝 PR Title/Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🔎 pre-commit	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—
🤖 therock-pr-bot	✅ Pass	—

🎉 All checks passed! This PR is ready for review.

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

[therock-pr-bot]

🎉 All checks passed! This PR is ready for review.

[Copilot]

Pull request overview

This PR changes PyTorch wheel packaging to resolve ROCm shared-library dependencies via relocatable ELF RUNPATHs (embedded in torch shared objects) instead of preloading rocprofiler-sdk at import/startup time, avoiding duplicate-provider initialization when users run PyTorch under external rocprofv3.

Changes:

• Removes rocprofiler-sdk from PyTorch’s generated Linux preload list.
• Post-processes the built torch wheel on Linux to strip absolute RUNPATH entries and append $ORIGIN-relative RUNPATHs pointing at ROCm Python package library directories.
• Extends package promotion to also rewrite ROCm version fragments embedded in torch ELF RUNPATHs (with unit tests for the string rewrite helper).

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
external-builds/pytorch/build_prod_wheels.py	Drops rocprofiler-sdk preload and adds wheel repack step that patches torch .so RUNPATHs via patchelf.
build_tools/packaging/promote_packages.py	Updates torch ELF RUNPATHs during version promotion so promoted wheels keep valid ROCm package paths.
build_tools/packaging/tests/promote_keep_list_test.py	Adds unit tests covering ROCm version substitution within RUNPATH strings.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[geomin12]

lgtm as it is passing