[test] test_topk_plain: parametrize sweep to fix collection-time OOM

[JohnQinAMD]

Problem

CI runs each op test as a script (python3 <file> in .github/scripts/aiter_test.sh), not via pytest. test_topk_plain.py sweeps 84 cases up to 3072 × 131072 fp32 and never frees tensors between cases, so on a memory-pressured gfx950 runner it OOMs mid-sweep and the whole file exits non-zero — producing intermittent "Standard Tests (MI35X, 8, 5)" failures unrelated to the code under test. On a clean GPU it passes; the failure is purely residual-memory dependent.

Fix (keeps the python3 <file> contract)

• Free each case's tensors (del + gc.collect() + torch.cuda.empty_cache()) so peak memory is one case, not the whole sweep — the actual OOM fix.
• Guard the sweep under if __name__ == "__main__": so import is side-effect-free.
• num_iters 1000 → 100 — this is a correctness check, not a perf gate; the high iteration count made it the slowest file in its shard for no benefit.
• Vectorize the per-row permutation (drops a batch_size-long Python loop).
• Assert on checkAllclose's returned error ratio — it does not raise, so the test previously would have passed even if topk_plain were incorrect.
• Summary table falls back to df.to_string when tabulate is absent, so the informational summary can never fail the run.

Validation

python3 op_tests/test_topk_plain.py on MI355X (gfx950): all 84 cases run, all passed~ (err 0), summary prints, exit 0.

Addresses both Copilot review comments (script-style no-op under python3; assert on the error ratio).

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

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

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 3934 --add-label <label>

[Copilot]

Pull request overview

This PR refactors op_tests/test_topk_plain.py to avoid collection/import-time GPU OOMs caused by a module-level sweep that allocates tensors for all cases up front. It converts the sweep into isolated, lazy-per-case execution using pytest parameterization, adds teardown to free GPU memory between cases, and preserves the end-of-run performance summary.

Changes:

• Replace the module-level sweep loop with @pytest.mark.parametrize test cases to prevent collection-time allocations and speed up collection.
• Add an autouse fixture to GC + torch.cuda.empty_cache() between cases and a session-scoped fixture to emit a single markdown perf summary (with a safe fallback when tabulate is missing).
• Vectorize the input permutation generation and reduce iteration counts (1000 → 20) to keep this as a correctness-focused test.

---

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