Seed the GDN CuteDSL correctness test inputs to fix flakiness (#32126)

2026-07-22 21:30:00 -07:00
parent 2d1a7be8c4
commit eb242b6c03
1 changed files with 4 additions and 0 deletions
@@ -39,6 +39,10 @@ from sglang.kernels.ops.attention.linear.gdn_blackwell import (  # noqa: E402
@pytest.mark.parametrize("num_seqs", [1, 5, 257])
@pytest.mark.parametrize("state_dtype", [torch.bfloat16, torch.float32])
 def test_gdn_chunk_cutedsl_correctness(num_seqs: int, state_dtype: torch.dtype):
+    # Fixed per-case seed: the mean-error assertions sit close to the observed
+    # error distribution (state_error.mean() threshold 6e-4 vs ~6.3e-4 seen on
+    # unlucky draws in CI), so unseeded inputs make the test flaky.
+    torch.manual_seed(num_seqs)
    seq_lens = torch.randint(1, 130, (num_seqs,), dtype=torch.int32)
    cu_seqlens = torch.zeros(num_seqs + 1, device="cuda", dtype=torch.int32)
    cu_seqlens[1:] = seq_lens.to(device="cuda").cumsum(0)