Preserve values with dtype for NumpyOps/CupyOps.asarray

thinc/backends/cupy_ops.py

@@ -93,6 +93,8 @@ def asarray(self, data, dtype=None):
             array = tensorflow2xp(data)
         elif is_mxnet_gpu_array(data):
             array = mxnet2xp(data)
+        elif dtype is not None:
+            array = self.xp.array(data, dtype=dtype)
         else:
             array = self.xp.array(data)
 

thinc/backends/numpy_ops.pyx

@@ -71,6 +71,8 @@ class NumpyOps(Ops):
             array = data.numpy()
         elif hasattr(data, "get"):
             array = data.get()
+        elif dtype is not None:
+            array = self.xp.array(data, dtype=dtype)
         else:
             array = self.xp.array(data)
 

thinc/layers/strings2arrays.py

@@ -17,8 +17,10 @@ def strings2arrays() -> Model[InT, OutT]:
 
 
 def forward(model: Model[InT, OutT], Xs: InT, is_train: bool) -> Tuple[OutT, Callable]:
-    hashes = [[hash_unicode(word) for word in X] for X in Xs]
-    hash_arrays = [model.ops.asarray2i(h, dtype="uint64") for h in hashes]
+    hashes = model.ops.asarray2i(
+        [[hash_unicode(word) for word in X] for X in Xs], dtype="int32"
+    )
+    hash_arrays = [model.ops.asarray1i(h, dtype="uint64") for h in hashes]
     arrays = [model.ops.reshape2i(array, -1, 1) for array in hash_arrays]
 
     def backprop(dX: OutT) -> InT:

thinc/tests/backends/test_ops.py

@@ -1597,3 +1597,10 @@ def test_custom_kernel_compilation():
         assert compiled_kernel is not None
 
     assert compile_mmh() is not None
+
+
+@pytest.mark.parametrize("ops", ALL_OPS)
+def test_asarray_from_list_uint64(ops):
+    # list contains int values both above and below int64.max
+    uint64_list = [16, 11648197037703959513]
+    assert uint64_list == list(ops.asarray(uint64_list, dtype="uint64"))