diff --git a/tests/kernels/quantization/test_block_fp8.py b/tests/kernels/quantization/test_block_fp8.py
index 52f87c403e..fd530e7ffb 100644
--- a/tests/kernels/quantization/test_block_fp8.py
+++ b/tests/kernels/quantization/test_block_fp8.py
@@ -406,5 +406,7 @@ def test_w8a8_block_fp8_b12x_matmul(M, N, K):
         ref_out.float().flatten(),
         dim=0,
     )
-    assert rel_diff < 0.002
-    assert cosine >= 0.9999
+    # Four-way split-K uses atomic BF16 reductions, so nondeterministic atomic
+    # ordering can flap an output between adjacent BF16 values one ULP apart.
+    assert rel_diff < 0.003
+    assert cosine >= 0.99999
diff --git a/tests/model_executor/kernels/test_b12x_linear.py b/tests/model_executor/kernels/test_b12x_linear.py
index 36a0896ea6..27368ead27 100644
--- a/tests/model_executor/kernels/test_b12x_linear.py
+++ b/tests/model_executor/kernels/test_b12x_linear.py
@@ -785,6 +785,11 @@ def test_b12x_backend_preserves_w4a16_fallback(monkeypatch) -> None:
 
     monkeypatch.setattr(linear_mod.current_platform, "_enum", PlatformEnum.CUDA)
     monkeypatch.setattr(linear_mod, "_get_linear_backend", lambda: "b12x")
+    monkeypatch.setitem(
+        linear_mod._POSSIBLE_NVFP4_KERNELS,
+        PlatformEnum.CUDA,
+        [B12xNvFp4LinearKernel, MarlinNvFp4LinearKernel],
+    )
     monkeypatch.setattr(
         MarlinNvFp4LinearKernel,
         "is_supported",