diff --git a/mjx/mujoco/mjx/_src/io.py b/mjx/mujoco/mjx/_src/io.py
index 4199936a..670c93e1 100644
--- a/mjx/mujoco/mjx/_src/io.py
+++ b/mjx/mujoco/mjx/_src/io.py
@@ -903,17 +903,6 @@ def _make_data_warp(
 
   data = jax.device_put(data, device=device)
 
-  with wp.ScopedDevice('cuda:0'):  # pylint: disable=undefined-variable
-    # Warm-up the warp kernel cache.
-    # TODO(robotics-simulation): remove this warmup compilation once warp
-    # stops unloading modules during XLA graph capture for tile kernels.
-    # pylint: disable=undefined-variable
-    dw = mjwp.make_data(m, nworld=1, naconmax=naconmax, njmax=njmax)
-    mw = mjwp.put_model(m)
-    _ = mjwp.step(mw, dw)
-    # pylint: enable=undefined-variable
-  del dw, mw
-
   return data
 
 
@@ -1363,6 +1352,44 @@ def _put_data_cpp(
   return _strip_weak_type(data)
 
 
+def _put_data_warp(
+    m: mujoco.MjModel,
+    d: mujoco.MjData,
+    device: Optional[jax.Device] = None,
+    naconmax: Optional[int] = None,
+    njmax: Optional[int] = None,
+) -> types.Data:
+  """Puts mujoco.MjData onto a device, resulting in mjx.Data."""
+
+  with wp.ScopedDevice('cpu'):  # pylint: disable=undefined-variable
+    dw = mjwp.put_data(m, d, nworld=1, naconmax=naconmax, njmax=njmax)  # pylint: disable=undefined-variable
+
+  fields = _put_data_public_fields(d)
+  for k in fields:
+    if not hasattr(dw, k):
+      continue
+    field = _wp_to_np_type(getattr(dw, k))
+    if mjxw.types._BATCH_DIM['Data'][k]:  # pylint: disable=protected-access
+      field = field.reshape(field.shape[1:])
+    fields[k] = field
+
+  impl_fields = {}
+  for k in mjxw.types.DataWarp.__annotations__.keys():
+    field = _get_nested_attr(dw, k, split='__')
+    field = _wp_to_np_type(field)
+    if mjxw.types._BATCH_DIM['Data'][k]:  # pylint: disable=protected-access
+      field = field.reshape(field.shape[1:])
+    impl_fields[k] = field
+
+  data = types.Data(
+      **fields,
+      _impl=mjxw.types.DataWarp(**impl_fields),
+  )
+
+  data = jax.device_put(data, device=device)
+  return data
+
+
 def put_data(
     m: mujoco.MjModel,
     d: mujoco.MjData,
@@ -1393,7 +1420,6 @@ def put_data(
     an mjx.Data placed on device
     DeprecationWarning: if nconmax is used
   """
-  del njmax
   if nconmax is not None:
     warnings.warn(
         'nconmax will be deprecated in mujoco-mjx>=3.5. Use naconmax instead.',
@@ -1410,8 +1436,10 @@ def put_data(
     return _put_data_cpp(
         m, d, device, dummy_arg_for_batching=dummy_arg_for_batching
     )
-
-  # TODO(robotics-team): implement put_data_warp
+  elif impl == types.Impl.WARP:
+    _check_warp_installed()
+    naconmax = nconmax if naconmax is None else naconmax
+    return _put_data_warp(m, d, device, naconmax, njmax)
 
   raise NotImplementedError(
       f'put_data for implementation "{impl}" not implemented yet.'
diff --git a/mjx/mujoco/mjx/_src/io_test.py b/mjx/mujoco/mjx/_src/io_test.py
index 8382a6cf..1cd0afe9 100644
--- a/mjx/mujoco/mjx/_src/io_test.py
+++ b/mjx/mujoco/mjx/_src/io_test.py
@@ -494,9 +494,15 @@ class DataIOTest(parameterized.TestCase):
     self.assertEqual(d._impl.contact__dist.shape[0], 9)
     self.assertEqual(d._impl.efc__pos.shape[0], 23)
 
-  @parameterized.parameters('jax', 'c', 'cpp')
+  @parameterized.parameters('jax', 'c', 'cpp', 'warp')
   def test_put_data(self, impl: str):
     """Test that put_data puts the correct data for dense and sparse."""
+    if impl == 'warp':
+      if not mjxw.WARP_INSTALLED:
+        self.skipTest('Warp is not installed.')
+      if not mjx_io.has_cuda_gpu_device():
+        self.skipTest('No CUDA GPU device.')
+
     m = mujoco.MjModel.from_xml_string(_MULTIPLE_CONSTRAINTS)
     d = mujoco.MjData(m)
     mujoco.mj_step(m, d, 2)
@@ -517,6 +523,15 @@ class DataIOTest(parameterized.TestCase):
         )
     )
 
+    # xmat, ximat, geom_xmat are all shape transformed
+    np.testing.assert_allclose(dx.xmat.reshape((-1, 9)), d.xmat)
+    np.testing.assert_allclose(dx.ximat.reshape((-1, 9)), d.ximat)
+    np.testing.assert_allclose(dx.geom_xmat.reshape((-1, 9)), d.geom_xmat)
+    np.testing.assert_allclose(dx.site_xmat.reshape((-1, 9)), d.site_xmat)
+
+    # tendon length is correct
+    np.testing.assert_allclose(dx.ten_length, d.ten_length)
+
     if impl == 'jax':
       # check that qM is transformed properly
       qm = np.zeros((m.nv, m.nv), dtype=np.float64)
@@ -530,6 +545,21 @@ class DataIOTest(parameterized.TestCase):
       self.assertTrue(hasattr(dx._impl, 'pointer_lo'))
       self.assertTrue(hasattr(dx._impl, 'pointer_hi'))
       return  # cpp does not populate other fields in _impl
+    elif impl == 'warp':
+      qm = np.zeros((m.nv, m.nv), dtype=np.float64)
+      mujoco.mj_fullM(m, qm, d.qM)
+      np.testing.assert_allclose(dx._impl.qM, qm)
+      # TODO(taylorhowell): test efc__J
+      np.testing.assert_allclose(dx._impl.efc__aref[:3], d.efc_aref[:3])
+
+    # tendon impl data is correct
+    np.testing.assert_equal(dx._impl.ten_wrapadr, np.zeros((1,)))
+    np.testing.assert_equal(dx._impl.ten_wrapnum, np.zeros((1,)))
+    np.testing.assert_equal(dx._impl.wrap_obj, np.zeros((2, 2)))
+    np.testing.assert_equal(dx._impl.wrap_xpos, np.zeros((2, 6)))
+
+    if impl == 'warp':
+      return
 
     # 4 contacts, 2 for each capsule against the plane
     self.assertEqual(dx._impl.contact.dist.shape, (4,))
@@ -542,23 +572,6 @@ class DataIOTest(parameterized.TestCase):
     )
     np.testing.assert_allclose(dx._impl.contact.frame[1:], 0)
 
-    # xmat, ximat, geom_xmat are all shape transformed
-    self.assertEqual(dx.xmat.shape, (3, 3, 3))
-    self.assertEqual(dx.ximat.shape, (3, 3, 3))
-    self.assertEqual(dx.geom_xmat.shape, (3, 3, 3))
-    self.assertEqual(dx.site_xmat.shape, (1, 3, 3))
-    np.testing.assert_allclose(dx.xmat.reshape((3, 9)), d.xmat)
-    np.testing.assert_allclose(dx.ximat.reshape((3, 9)), d.ximat)
-    np.testing.assert_allclose(dx.geom_xmat.reshape((3, 9)), d.geom_xmat)
-    np.testing.assert_allclose(dx.site_xmat.reshape((1, 9)), d.site_xmat)
-
-    # tendon data is correct
-    np.testing.assert_allclose(dx.ten_length, d.ten_length)
-    np.testing.assert_equal(dx._impl.ten_wrapadr, np.zeros((1,)))
-    np.testing.assert_equal(dx._impl.ten_wrapnum, np.zeros((1,)))
-    np.testing.assert_equal(dx._impl.wrap_obj, np.zeros((2, 2)))
-    np.testing.assert_equal(dx._impl.wrap_xpos, np.zeros((2, 6)))
-
     # efc_ are also shape transformed and padded
     self.assertEqual(dx._impl.efc_J.shape, (45, 8))  # nefc, nv
     d_efc_j = d.efc_J.reshape((-1, 8))
@@ -583,7 +596,9 @@ class DataIOTest(parameterized.TestCase):
     d = mujoco.MjData(m)
     mujoco.mj_step(m, d, 2)
     dx_sparse = mjx.put_data(m, d, impl=impl)
-    np.testing.assert_allclose(dx_sparse._impl.efc_J, dx._impl.efc_J, atol=1e-8)
+    np.testing.assert_allclose(
+        dx_sparse._impl.efc_J, dx._impl.efc_J, atol=1e-8
+    )
 
     # check sparse mass matrices are correct
     np.testing.assert_allclose(dx_sparse._impl.qM, d.qM, atol=1e-8)
@@ -604,6 +619,31 @@ class DataIOTest(parameterized.TestCase):
     elif impl == 'c':
       np.testing.assert_allclose(dx_from_dense._impl.qM, d.qM, atol=1e-8)
 
+  def test_put_data_warp_ndim(self):
+    """Tests that put_data produces expected dimensions for Warp fields."""
+    if not mjxw.WARP_INSTALLED:
+      self.skipTest('Warp is not installed.')
+    if not mjx_io.has_cuda_gpu_device():
+      self.skipTest('No CUDA GPU device.')
+
+    m = mujoco.MjModel.from_xml_string(_MULTIPLE_CONSTRAINTS)
+    d = mujoco.MjData(m)
+    mujoco.mj_step(m, d, 2)
+    dx = mjx.put_data(m, d, impl='warp')
+
+    def check_ndim(path, x):
+      k = _get_name_from_path(path)
+      if k not in mjxw_types._NDIM['Data']:
+        return
+      is_batched = mjxw_types._BATCH_DIM['Data'][k]
+      expected_ndim = mjxw_types._NDIM['Data'][k] - is_batched
+      if not hasattr(x, 'ndim'):
+        return
+      msg = f'Field {k} has ndim {x.ndim} but expected {expected_ndim}'
+      self.assertEqual(x.ndim, expected_ndim, msg)
+
+    _ = jax.tree.map_with_path(check_ndim, dx)
+
   @parameterized.parameters(
       ('jax', False), ('jax', True), ('c', False), ('c', True)
   )
diff --git a/mjx/mujoco/mjx/viewer.py b/mjx/mujoco/mjx/viewer.py
index 075e243e..ddd88696 100644
--- a/mjx/mujoco/mjx/viewer.py
+++ b/mjx/mujoco/mjx/viewer.py
@@ -88,15 +88,9 @@ def _main(argv: Sequence[str]) -> None:
     m = mujoco.MjModel.from_xml_path(_MODEL_PATH.value)
   d = mujoco.MjData(m)
   mx = mjx.put_model(m, impl=_IMPL.value)
-  if _IMPL.value == 'warp':
-    # TODO(btaba): use put_data.
-    dx = mjx.make_data(
-        m, impl=_IMPL.value, naconmax=_NACONMAX.value, njmax=_NJMAX.value
-    )
-  else:
-    dx = mjx.put_data(
-        m, d, impl=_IMPL.value, naconmax=_NACONMAX.value, njmax=_NJMAX.value
-    )
+  dx = mjx.put_data(
+      m, d, impl=_IMPL.value, naconmax=_NACONMAX.value, njmax=_NJMAX.value
+  )
 
   print(f'Default backend: {jax.default_backend()}')
   step_fn = mjx.step