Fix nondeterministic FlashInfer GDN alignment test

[nvpohanh]

[by Codex]

Summary

• initialize the misaligned a and b test views with deterministic finite values
• keep their values distinct so the test still verifies that both FlashInfer arguments are copied independently and without corruption
• avoid comparing allocator-provided torch.empty contents, which can contain NaNs and make assert_close fail nondeterministically

What failure does this fix?

This fixes the base-a-test-cpu failure observed while testing #33922:

• CI job: https://github.com/sgl-project/sglang/actions/runs/33459311788/job/99706092418
• test: test/registered/unit/layers/attention/test_gdn_flashinfer_alignment.py
• test case: TestFlashInferGDNAlignment::test_decode_repairs_read_only_arguments_before_flashinfer
• CI signature: torch.testing.assert_close(captured["a"], a) reported 2/8 mismatched elements with a greatest absolute difference of nan

The same test failed standalone on latest main in lmsysorg/sglang:dev-cu13 on a computelab node. In that run, captured["b"] reported 8/8 mismatched elements with a greatest absolute difference of nan. Which tensor/elements fail depends on the contents returned by the allocator, but both failures have the same root cause.

Additional confirmed CI occurrences

I scanned 600 failed PR Test Base workflow runs after #35758 introduced this test and inspected the logs of their failed base-a-test-cpu root jobs. The exact failure was confirmed in three pipeline runs across two unrelated PRs:

• Run 33353343812 / job 99370860899 — PR [CPU] Fix shm allreduce collision and sglang-router import #37179, 2026-08-31 03:16 UTC
• Run 33459311788 / job 99706092418 — PR Fix Qwen3.5 GDN multi-item scoring #33922, 2026-09-01 01:35 UTC
• Run 33460931970 / job 99710903533 — PR [CPU] Fix shm allreduce collision and sglang-router import #37179, 2026-09-01 02:00 UTC

All three failed in TestFlashInferGDNAlignment::test_decode_repairs_read_only_arguments_before_flashinfer with the same signature: 2/8 mismatched elements and a nan difference at index (0, 0, 2). The recurrence across unrelated PRs and separate workflow runs supports that this is an intermittent main-branch test defect rather than a regression caused by any one target PR.

Root cause

_view_with_pointer_mod() intentionally builds misaligned views from torch.empty backing storage. In this test, a and b were never initialized before kernel.decode():

a = _view_with_pointer_mod((1, 1, 8), torch.bfloat16, 16)
b = _view_with_pointer_mod((1, 1, 8), torch.bfloat16, 16)

The alignment-repair path correctly copies these views into aligned scratch buffers. The test then compares the copied tensors against the original uninitialized tensors with torch.testing.assert_close. If the allocator-provided bit patterns contain NaNs, the comparison fails because assert_close does not consider corresponding NaNs equal by default.

Therefore:

• the failure is intermittent in manifestation, because it depends on allocator reuse and the uninitialized bit patterns;
• it is not an infrastructure-only flake—there is a concrete test-code defect;
• it is not caused by Fix Qwen3.5 GDN multi-item scoring #33922; the exact test also reproduces on latest main without that PR;
• the defective test was introduced by #35758 (qwen 3.8 rebase), merge commit 5f216fc33f315ade378bc7cd295ee88f5322d4d3, which added this test file and the uninitialized setup.

Fix

Initialize a and b with distinct finite sequences before calling decode(). This removes allocator-dependent NaNs while preserving the test's intended coverage:

• both source views remain deliberately 16-byte-offset/misaligned;
• the test still verifies 32-byte alignment at the FlashInfer call boundary;
• distinct values ensure the two live argument buffers cannot be accidentally aliased or swapped;
• the final assert_close checks still verify lossless copy behavior.

Validation

After rebasing onto 60f881b40c:

• exact failing pytest node: passed
• full test_gdn_flashinfer_alignment.py: 10 passed
• 500 allocator-poison stress iterations: passed
• BLACK_NUM_WORKERS=1 SKIP=no-commit-to-branch pre-commit run --all-files --show-diff-on-failure: passed

---

CI States

Latest PR Test (Base): ❌ Run #33467025020
Latest PR Test (Extra): ❌ Run #33467024850
Latest PR Test (AMD ROCm 7.2): ❌ Run #33467024983

[nvpohanh]

/rerun-tests test/registered/unit/layers/attention/test_gdn_flashinfer_alignment.py

[github-actions]

Results for /rerun-tests test/registered/unit/layers/attention/test_gdn_flashinfer_alignment.py:

🚀 ubuntu-latest (1 test): ✅ View workflow run

cd test/ && python3 registered/unit/layers/attention/test_gdn_flashinfer_alignment.py