[zufayu]

The ROCm 10.0 Nightly Docker Release fails while testing the image because loader worker initialization resolves omp_set_num_threads from the process-global symbol table. PyTorch may load its OpenMP dependency with RTLD_LOCAL, so ctypes.CDLL(None) raises undefined symbol: omp_set_num_threads and breaks the loader thread pool.

Resolve the symbol through the already-loaded torch._C extension and its dependency graph. This selects the same OpenMP runtime used by PyTorch while preserving the per-loader-thread limit. Add a regression test that verifies the lookup uses the torch extension rather than the global process handle.

Failure evidence: https://github.com/ROCm/ATOM/actions/runs/41456304159

Validation:

• python -m compileall atom/model_loader/loading_core.py tests/test_loader_threads.py
• git diff --check
• PyTorch, Black, and Ruff are unavailable in the local Windows environment; repository CI will run the executable test and formatting checks.

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 2936 --add-label <label>

[valar-cop]

one thing to confirm before merge..

post-fix run in the rocm 10 nightly image showing the loader pool starts?

ci green.

[zufayu]

Status: review thread resolved, all required checks green, ready to squash-merge — except the for_main gate needs one code-owner approval (I pushed the branch, so I can't self-approve).

On the inline question — the fix resolves omp_set_num_threads strictly through torch._C's dependency chain, i.e. torch's own bundled OpenMP runtime. If the image preloads libomp separately, that is exactly the runtime torch._C already depends on, so the lookup still selects torch's OMP rather than a second implementation. The definitive confirmation is the ROCm 10.0 nightly (loader pool init on the built image); if it still fails there, we revert.

@gyohuangxin could you take a final look / approve when you get a chance?

[valar-cop]

ci triage for 7d37484:

ATOM SGLang Test passed. https://github.com/ROCm/ATOM/actions/runs/41587269821

ATOM vLLM Test passed. https://github.com/ROCm/ATOM/actions/runs/41587269787

ATOM Test passed. https://github.com/ROCm/ATOM/actions/runs/41587269855