diff --git a/test/backend/test_multitensor.py b/test/backend/test_multitensor.py
index 9e0aa718d6..c8a580fff0 100644
--- a/test/backend/test_multitensor.py
+++ b/test/backend/test_multitensor.py
@@ -746,6 +746,11 @@ class TestMultiTensor(unittest.TestCase):
     t2.realize()
   def test_rand_like_on_shard_axis(self): self.test_rand_like_on_shard(0)
 
+  def test_rand_like_on_shard_axis_requires_grad(self):
+    t = Tensor.empty((16, 16)).shard(devices_2, axis=0)
+    self.assertIs(t.rand_like(requires_grad=True).requires_grad, True)
+    self.assertIs(t.rand_like(requires_grad=False).requires_grad, False)
+
   def test_rand_like_from_alu(self):
     a = Tensor.ones(4, 4).shard(devices_4, axis=0)
     aa = a + a
diff --git a/test/backend/test_tensor.py b/test/backend/test_tensor.py
index b98f88540f..95e8061f2a 100644
--- a/test/backend/test_tensor.py
+++ b/test/backend/test_tensor.py
@@ -260,6 +260,10 @@ class TestTinygrad(unittest.TestCase):
     b = Tensor.randperm(1000).realize()
     np.testing.assert_equal(set(b.numpy()), set(range(1000)))
 
+  def test_randperm_requires_grad(self):
+    self.assertIs(Tensor.randperm(5, requires_grad=True).requires_grad, True)
+    self.assertIs(Tensor.randperm(5, requires_grad=False).requires_grad, False)
+
   def test_randn_isnt_inf_on_zero(self):
     # simulate failure case of rand handing a zero to randn
     original_rand, Tensor.rand = Tensor.rand, Tensor.zeros
diff --git a/tinygrad/tensor.py b/tinygrad/tensor.py
index 694e571317..f06fe78515 100644
--- a/tinygrad/tensor.py
+++ b/tinygrad/tensor.py
@@ -618,7 +618,7 @@ class Tensor(OpMixin):
     if kwargs.get("device") is not None: raise RuntimeError("cannot specify `device` on `*_like` of a multi device tensor")
     if self.uop.axis is None: return fxn(self.shape, *args, dtype=dtype, **kwargs).shard(self.device)
     stacked = UOp.mstack(*[fxn(self.uop.shard_shape, *args, device=d, dtype=dtype, **kwargs).uop for d in self.device])
-    return Tensor(stacked.multi(self.uop.axis))
+    return Tensor(stacked.multi(self.uop.axis), requires_grad=kwargs.get("requires_grad"))
 
   def full_like(self, fill_value:ConstType, dtype=None, device=None, requires_grad=None) -> Tensor:
     """
@@ -816,7 +816,7 @@ class Tensor(OpMixin):
     print(Tensor.randperm(6).numpy())
     ```
     """
-    return Tensor.rand(n, device=device, **kwargs).argsort().cast(dtype)
+    return Tensor.rand(n, device=device, **kwargs).argsort().cast(dtype).requires_grad_(kwargs.get("requires_grad"))
 
   def multinomial(self:Tensor, num_samples:int = 1, replacement:bool = False) -> Tensor:
     """