diff --git a/test/test_nn.py b/test/test_nn.py
index 72fc0d77df..f85aa43f59 100755
--- a/test/test_nn.py
+++ b/test/test_nn.py
@@ -13,29 +13,6 @@ from test.helpers import not_support_multi_device
 
 @unittest.skipIf(CI and Device.DEFAULT in {"CUDA", "NV"}, "slow")
 class TestNN(unittest.TestCase):
-  def test_sparse_categorical_crossentropy(self):
-    # create in tinygrad
-    input_tensor = Tensor.randn(6, 5) # not square to test that mean scaling uses the correct dimension
-    target = Tensor([0, 0, 0, 1, 2, 3])  # torch doesn't support target=-1
-    torch_input = torch.tensor(input_tensor.numpy())
-    torch_target = torch.tensor(target.numpy(), dtype=torch.long)
-
-    for smoothing in [0.0, 0.1, 0.5, 1.0]:
-      for ignore_index in [-1, 0, 2]:
-        for reduction in ["none", "sum", "mean"]:
-          loss = input_tensor.sparse_categorical_crossentropy(target, label_smoothing=smoothing, ignore_index=ignore_index, reduction=reduction)
-          torch_loss = torch.nn.CrossEntropyLoss(reduction=reduction, label_smoothing=smoothing, ignore_index=ignore_index)(torch_input, torch_target)
-          np.testing.assert_allclose(loss.numpy(), torch_loss.detach().numpy(), atol=1e-5, rtol=1e-6)
-
-          # also test with a batch dimension (of size 1)
-          loss = input_tensor.unsqueeze(0).sparse_categorical_crossentropy(
-            target.unsqueeze(0), label_smoothing=smoothing, ignore_index=ignore_index, reduction=reduction
-          )
-          torch_loss = torch.nn.CrossEntropyLoss(reduction=reduction, label_smoothing=smoothing, ignore_index=ignore_index)(
-            torch_input.unsqueeze(0).permute(0,2,1), torch_target.unsqueeze(0)
-          )
-          np.testing.assert_allclose(loss.numpy(), torch_loss.detach().numpy(), atol=1e-5, rtol=1e-6)
-
   def test_batchnorm2d(self, training=False, threed=False, track_running_stats=True):
     with Tensor.train(training):
       szs = [4, 8, 16, 32]
diff --git a/test/test_ops.py b/test/test_ops.py
index cb94a4294f..6b6ef9563e 100644
--- a/test/test_ops.py
+++ b/test/test_ops.py
@@ -2950,6 +2950,39 @@ class TestOps(unittest.TestCase):
       helper_test_op([(32,10)], lambda x: torch.nn.functional.cross_entropy(x, torch.tensor(classes), label_smoothing=ls),
                                 lambda x: x.cross_entropy(Tensor(classes), label_smoothing=ls))
 
+  def test_sparse_categorical_crossentropy(self):
+    classes = np.random.randint(0, 10, (12,), dtype=np.int32).tolist()
+    helper_test_op([(12,10)], lambda x: torch.nn.CrossEntropyLoss()(x, torch.tensor(classes)),
+                              lambda x: x.sparse_categorical_crossentropy(Tensor(classes)))
+
+    # combine args
+    helper_test_op([(12,10)],
+                   lambda x: torch.nn.CrossEntropyLoss(reduction="mean", ignore_index=classes[0], label_smoothing=0.3)(x, torch.tensor(classes)),
+                   lambda x: x.sparse_categorical_crossentropy(Tensor(classes), reduction="mean", ignore_index=classes[0], label_smoothing=0.3))
+
+    # with batch. somehow this does not match torch
+    classes = np.random.randint(0, 10, (3,12), dtype=np.int32).tolist()
+    helper_test_op([(3,12,10)], lambda x: torch.nn.CrossEntropyLoss()(x.permute(0,2,1), torch.tensor(classes)),
+                                lambda x: x.sparse_categorical_crossentropy(Tensor(classes)))
+
+  def test_sparse_categorical_crossentropy_reductions(self):
+    for r in ("mean", "sum", "none"):
+      classes = np.random.randint(0, 10, (12,), dtype=np.int32).tolist()
+      helper_test_op([(12,10)], lambda x: torch.nn.CrossEntropyLoss(reduction=r)(x, torch.tensor(classes)),
+                                lambda x: x.sparse_categorical_crossentropy(Tensor(classes), reduction=r))
+
+  def test_sparse_categorical_crossentropy_ignore_index(self):
+    classes = [0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3]
+    for i in (-1, 0, 3):
+      helper_test_op([(12,10)], lambda x: torch.nn.CrossEntropyLoss(ignore_index=i)(x, torch.tensor(classes)),
+                                lambda x: x.sparse_categorical_crossentropy(Tensor(classes), ignore_index=i))
+
+  def test_sparse_categorical_crossentropy_label_smoothing(self):
+    for s in (0.3, 0.9):
+      classes = np.random.randint(0, 10, (12,), dtype=np.int32).tolist()
+      helper_test_op([(12,10)], lambda x: torch.nn.CrossEntropyLoss(label_smoothing=s)(x, torch.tensor(classes)),
+                                lambda x: x.sparse_categorical_crossentropy(Tensor(classes), label_smoothing=s))
+
   def test_nll_loss(self):
     target = np.random.randint(0, 10, (32,), dtype=np.int32).tolist()
     helper_test_op([(32,10)],