diff --git a/test/backend/test_multitensor.py b/test/backend/test_multitensor.py
index 39bdec3d13..bdda8656d7 100644
--- a/test/backend/test_multitensor.py
+++ b/test/backend/test_multitensor.py
@@ -305,7 +305,8 @@ class TestMultiTensor(unittest.TestCase):
     Xs = X.shard(device, shard_x)
     Ws = W.shard(device, shard_w)
     O = (Xs@Ws)
-    np.testing.assert_allclose(X.numpy() @ W.numpy(), O.to(Device.DEFAULT).numpy(), atol=1e-5)
+    with np.errstate(all='ignore'):
+      np.testing.assert_allclose(X.numpy() @ W.numpy(), O.to(Device.DEFAULT).numpy(), atol=1e-5)
 
   def _test_double_matmul_shard_axis(self, shard_x, shard_w, device):
     X = Tensor.kaiming_uniform(N, N).realize()
@@ -315,7 +316,8 @@ class TestMultiTensor(unittest.TestCase):
     W1s = W1.shard(device, shard_w)
     W2s = W2.shard(device, shard_w)
     O = (Xs@W1s)@W2s
-    np.testing.assert_allclose((X.numpy() @ W1.numpy()) @ W2.numpy(), O.to(Device.DEFAULT).numpy(), atol=1e-5)
+    with np.errstate(all='ignore'):
+      np.testing.assert_allclose((X.numpy() @ W1.numpy()) @ W2.numpy(), O.to(Device.DEFAULT).numpy(), atol=1e-5)
 
   def test_matmul_shard_none(self): return self._test_matmul_shard_axis(None, None, devices_2)
   def test_matmul_shard_X_0(self): return self._test_matmul_shard_axis(0, None, devices_2)
diff --git a/test/backend/test_opt_gemm.py b/test/backend/test_opt_gemm.py
index 12d1bda436..83d8933563 100644
--- a/test/backend/test_opt_gemm.py
+++ b/test/backend/test_opt_gemm.py
@@ -12,7 +12,8 @@ class TestOptGemm(unittest.TestCase):
     N = 64
     cls.a = Tensor.randn(N, N).contiguous().realize()
     cls.b = Tensor.randn(N, N).contiguous().realize()
-    cls.res = cls.a.T.numpy() @ cls.b.T.numpy()
+    with np.errstate(all='ignore'):
+      cls.res = cls.a.T.numpy() @ cls.b.T.numpy()
 
   def _test_gemm_unrolled_permute_l(self, opts=[]):
     t = self.a.T @ self.b.T
diff --git a/test/backend/test_schedule.py b/test/backend/test_schedule.py
index 9723273e31..494f0079f0 100644
--- a/test/backend/test_schedule.py
+++ b/test/backend/test_schedule.py
@@ -312,8 +312,9 @@ class TestSchedule(unittest.TestCase):
     np.testing.assert_allclose(out0.numpy(), np_out0:=np.exp2(a.numpy().sum()), atol=1e-4, rtol=1e-4)
     np.testing.assert_allclose(out1.numpy(), np_out1:=a.numpy().sum()+np_out0, atol=1e-4, rtol=1e-4)
     np_b = (a.numpy() + np_out0 + np_out1)
-    np.testing.assert_allclose(out2.numpy(), np_out2:=np.exp2(np_b.sum()), atol=1e-4, rtol=1e-4)
-    np.testing.assert_allclose(out3.numpy(), np_b.sum()+np_out2, atol=1e-4, rtol=1e-4)
+    with np.errstate(over='ignore'):
+      np.testing.assert_allclose(out2.numpy(), np_out2:=np.exp2(np_b.sum()), atol=1e-4, rtol=1e-4)
+      np.testing.assert_allclose(out3.numpy(), np_b.sum()+np_out2, atol=1e-4, rtol=1e-4)
 
   def test_reduce_ext_reduce_child(self):
     Tensor.manual_seed(0)
diff --git a/test/unit/test_dtype_spec.py b/test/unit/test_dtype_spec.py
index e0f8f514f7..477faf22bb 100644
--- a/test/unit/test_dtype_spec.py
+++ b/test/unit/test_dtype_spec.py
@@ -64,7 +64,8 @@ class TestTypeSpec(unittest.TestCase):
     tested = 0
     for dtype_str, dtype in [
       ("bool", dtypes.bool), ("int8", dtypes.int8), ("int", dtypes.int), ("uint32", dtypes.uint32), ("float32", dtypes.float32)]:
-      np.testing.assert_equal(Tensor(n, dtype=dtype_str).numpy(), Tensor(n, dtype=dtype).numpy())
+      with np.errstate(invalid='ignore'):
+        np.testing.assert_equal(Tensor(n, dtype=dtype_str).numpy(), Tensor(n, dtype=dtype).numpy())
       np.testing.assert_equal(Tensor(n).cast(dtype_str).numpy(), Tensor(n).cast(dtype).numpy())
       if dtype.itemsize == 4:
         np.testing.assert_equal(Tensor(n).bitcast(dtype_str).numpy(), Tensor(n).bitcast(dtype).numpy())
diff --git a/test/unit/test_gguf.py b/test/unit/test_gguf.py
index dfd587f29c..3b611a2488 100644
--- a/test/unit/test_gguf.py
+++ b/test/unit/test_gguf.py
@@ -162,7 +162,8 @@ class TestGGUFGEMV(unittest.TestCase):
     _, tensors = gguf_load(Tensor(np.frombuffer(buf, dtype=np.uint8)).to(None))
 
     x = rng.standard_normal(cols).astype(np.float32)
-    np.testing.assert_allclose((tensors["weight"] @ Tensor(x)).numpy(), ref @ x, atol=1e-2, rtol=1e-2)
+    with np.errstate(all='ignore'):
+      np.testing.assert_allclose((tensors["weight"] @ Tensor(x)).numpy(), ref @ x, atol=1e-2, rtol=1e-2)
     if qtype == GGMLQuantizationType.BF16 or is_dtype_supported(dtypes.half): np.testing.assert_equal(tensors["weight"].numpy(), ref)
     assert np.isfinite(ref).all() and np.isfinite(tensors["weight"].numpy()).all(), f"{qtype.name} has NaN/Inf"