[paddle-adapt] model_optimizations: 690 PASS with torch.sort compat patch

[BingooYang]

📌 Description

Adapt tests/model_optimizations for Paddle compat mode.

The key fix is a monkey-patch for torch.sort in tests/model_optimizations/conftest.py, which addresses two Paddle compat issues:

1. Paddle compat torch.sort returns values-only Tensor (not a (values, indices) namedtuple like PyTorch), causing ValueError: not enough values to unpack (expected 2, got 1)
2. Paddle compat torch.sort does not accept dim= keyword argument (uses axis= internally), causing TypeError: paddle.sort() received unexpected keyword argument dim

The patch introduces a _SortResult class wrapping (values, indices) where indices are computed via argsort with stable=True to preserve correct bfloat16 tie-breaking semantics.

Result: 690 passed, 4164 skipped (SM architecture env-level), 0 failed

🔍 Related Issues

Part of the FlashInfer Paddle compat adaptation effort.

Changes

• tests/model_optimizations/conftest.py (new): Paddle compat monkey-patches §44-§48 (assert_close bfloat16, torch.equal, tensor.multiply, clamp_min/max) + §51 (torch.sort)
• scripts/paddle_all_test_cases.sh: Added tests/model_optimizations/ entry

🚀 Pull Request Checklist

✅ Pre-commit Checks

• Installed pre-commit hooks
• pre-commit run --all-files (all hooks passed)

🧪 Tests

• tests/model_optimizations: 690 passed, 4164 skipped, 0 failed
• Regression: tests/utils/test_topk.py 1335 PASS, tests/grouped_mm 4 PASS

Reviewer Notes

Key insight: torch.sort must use stable=True in argsort for bfloat16 cases. Without stable=True, many tied values in bfloat16 are broken differently than PyTorch, causing ~210 new failures. The fix must live in conftest.py (not inline in test files) to apply the stable sort globally.

Refs: MISMATCH_EXPERIMENT §51