diff --git a/dpctl/tensor/_indexing_functions.py b/dpctl/tensor/_indexing_functions.py
index 188f3cd291..374e1bb58f 100644
--- a/dpctl/tensor/_indexing_functions.py
+++ b/dpctl/tensor/_indexing_functions.py
@@ -120,7 +120,7 @@ def take(x, indices, /, *, axis=None, mode="wrap"):
     return res
 
 
-def put(x, indices, vals, /, *, axis=None, mode="wrap"):
+def put(x, indices, vals, /, *, axis=None, mode="wrap", unique_only=False):
     """put(x, indices, vals, axis=None, mode="wrap")
 
     Puts values of an array into another array
@@ -144,6 +144,10 @@ def put(x, indices, vals, /, *, axis=None, mode="wrap"):
           negative indices.
           "clip" - clips indices to (0 <= i < n)
           Default: `"wrap"`.
+       unique_only:
+          If True, only unique indices will be used.
+          This will prevent a undefined behavior when indices repeat,
+          but may negatively impact performance.
     """
     if not isinstance(x, dpt.usm_ndarray):
         raise TypeError(
@@ -175,6 +179,22 @@ def put(x, indices, vals, /, *, axis=None, mode="wrap"):
                 indices.dtype
             )
         )
+    if not isinstance(unique_only, bool):
+        raise TypeError("`unique_only` expected `bool`, got `{}`".format(
+            type(unique_only)
+        ))
+
+    not_unique = False
+    if unique_only:
+        indices = dpt.flip(indices)
+        _, local_indices = np.unique(indices, return_index=True)
+        not_unique = len(indices) != len(local_indices)
+
+    if not_unique:
+        local_indices = dpt.asarray(local_indices)
+        indices = dpt.take(indices, local_indices)
+        indices = dpt.flip(indices)
+
     queues_.append(indices.sycl_queue)
     usm_types_.append(indices.usm_type)
     exec_q = dpctl.utils.get_execution_queue(queues_)
@@ -207,7 +227,10 @@ def put(x, indices, vals, /, *, axis=None, mode="wrap"):
         vals = dpt.asarray(
             vals, dtype=x.dtype, usm_type=vals_usm_type, sycl_queue=exec_q
         )
-
+    if not_unique:
+        vals = dpt.flip(vals)
+        vals = dpt.take(vals, local_indices)
+        vals = dpt.flip(vals)
     vals = dpt.broadcast_to(vals, val_shape)
 
     hev, _ = ti._put(x, (indices,), vals, axis, mode, sycl_queue=exec_q)