diff --git a/mlx/backend/gpu/primitives.cpp b/mlx/backend/gpu/primitives.cpp
index 268d6290bf..0bc19c8171 100644
--- a/mlx/backend/gpu/primitives.cpp
+++ b/mlx/backend/gpu/primitives.cpp
@@ -246,8 +246,10 @@ void View::eval_gpu(const std::vector<array>& inputs, array& out) {
   // - type size is the same
   // - type size is smaller and the last axis is contiguous
   // - the entire array is row contiguous
-  if (ibytes == obytes || (obytes < ibytes && in.strides().back() == 1) ||
-      in.flags().row_contiguous) {
+  // and the offset is a multiple of the output type size
+  if (in.offset() % obytes == 0 &&
+      (ibytes == obytes || (obytes < ibytes && in.strides().back() == 1) ||
+       in.flags().row_contiguous)) {
     auto strides = in.strides();
     for (int i = 0; i < static_cast<int>(strides.size()) - 1; ++i) {
       strides[i] *= ibytes;
diff --git a/python/tests/test_ops.py b/python/tests/test_ops.py
index 7bda5c84b6..b3589580e5 100644
--- a/python/tests/test_ops.py
+++ b/python/tests/test_ops.py
@@ -4154,6 +4154,14 @@ def test_view(self):
             a_out = out.view(mx.int32)
             self.assertTrue(mx.array_equal(a_out, a, equal_nan=True))
 
+        # Byte offsets that are not a multiple of the new item size
+        a = mx.arange(64, dtype=mx.uint8)
+        a_np = np.arange(64, dtype=np.uint8)
+        for start in range(8):
+            out = a[start : start + 16].view(mx.int32)
+            expected = a_np[start : start + 16].view(np.int32)
+            self.assertTrue(np.array_equal(out + 0, expected))  # read in a kernel
+
     def _hadamard(self, N):
         # Matches scipy.linalg.hadamard
         H = np.array([[1]], dtype=np.int64)