Drop pyproject.toml entries, start windows fix

.github/workflows/array-api.yml

@@ -13,6 +13,10 @@ concurrency:
 
 jobs:
   array-api-tests:
+    # Run if the commit message contains 'run array-api tests' or if the job is triggered on schedule
+    if: >-
+      contains(github.event.head_commit.message, 'run array-api tests') ||
+      github.event_name == 'schedule'
     name: Array API test
     timeout-minutes: 90
     runs-on: ubuntu-latest-8core

.github/workflows/ci.yml

@@ -32,13 +32,13 @@ jobs:
       fail-fast: false
       matrix:
         os:
-          - ubuntu-latest
-          - macos-latest
+          # - ubuntu-latest
+          # - macos-latest
           - windows-latest
         environment:
-          - py310
+          # - py310
           - py311
-          - py312
+          # - py312
     steps:
       - name: Checkout branch
         uses: actions/checkout@v4

ndonnx/_core/_impl.py

@@ -540,7 +540,10 @@ def unique_all(self, x):
 
         # FIXME: I think we can simply use arange/ones+cumsum or something for the indices
         # maybe: indices = opx.cumsum(ones_like(flattened, dtype=dtypes.i64), axis=ndx.asarray(0))
-        indices = opx.squeeze(opx.ndindex(opx.shape(flattened._core())), opx.const([1]))
+        indices = opx.squeeze(
+            opx.ndindex(opx.shape(flattened._core())),
+            opx.const([1], dtype=dtypes.int64),
+        )
 
         ret = namedtuple("ret", ["values", "indices", "inverse_indices", "counts"])
 

ndonnx/_funcs.py

@@ -618,7 +618,9 @@ def expand_dims(x, axis=0):
     if (out := x.dtype._ops.expand_dims(x, axis)) is not NotImplemented:
         return out
     return x._transmute(
-        lambda corearray: opx.unsqueeze(corearray, axes=opx.const([axis]))
+        lambda corearray: opx.unsqueeze(
+            corearray, axes=opx.const([axis], dtype=dtypes.int64)
+        )
     )
 
 
@@ -694,7 +696,12 @@ def roll(x, shift, axis=None):
         shift_single = opx.add(opx.const(-sh), len_single)
         # Find the needed element index and then gather from it
         range = opx.cast(
-            opx.range(opx.const(0), len_single, opx.const(1)), to=dtypes.int64
+            opx.range(
+                opx.const(0, dtype=len_single.dtype),
+                len_single,
+                opx.const(1, dtype=len_single.dtype),
+            ),
+            to=dtypes.int64,
         )
         new_indices = opx.mod(opx.add(range, shift_single), len_single)
         x = take(x, _from_corearray(new_indices), axis=ax)

ndonnx/_opset_extensions.py

@@ -184,6 +184,8 @@ def const(value, dtype: CoreType | None = None) -> _CoreArray:
 
 @eager_propagate
 def squeeze(data: _CoreArray, axes: _CoreArray | None = None) -> _CoreArray:
+    if axes is not None and axes.dtype != dtypes.int64:
+        raise ValueError(f"Expected axes to be of type int64, got {axes.dtype}")
     return _CoreArray(op.squeeze(data.var, axes=axes.var if axes is not None else None))
 
 
@@ -364,6 +366,8 @@ def concat(inputs: list[_CoreArray], axis: int) -> _CoreArray:
 
 @eager_propagate
 def unsqueeze(data: _CoreArray, axes: _CoreArray) -> _CoreArray:
+    if axes.dtype != dtypes.int64:
+        raise TypeError(f"axes must be int64, got {axes.dtype}")
     return _CoreArray(op.unsqueeze(data.var, axes.var))
 
 
@@ -503,7 +507,7 @@ def getitem_null(corearray: _CoreArray, index: _CoreArray) -> _CoreArray:
     if get_rank(index) == 0:
 
         def extend_shape(x: Var) -> Var:
-            return op.concat([op.const([1]), op.shape(x)], axis=0)
+            return op.concat([op.const([1], dtype=np.int64), op.shape(x)], axis=0)
 
         var = op.reshape(var, extend_shape(var), allowzero=True)
         index_var = op.reshape(index.var, extend_shape(index.var), allowzero=True)
@@ -514,7 +518,8 @@ def extend_shape(x: Var) -> Var:
             reshaped_var, reshaped_index = var, index.var
         else:
             ret_shape = op.concat(
-                [op.const([-1]), op.shape(var, start=get_rank(index))], axis=0
+                [op.const([-1], dtype=np.int64), op.shape(var, start=get_rank(index))],
+                axis=0,
             )
             reshaped_var = op.reshape(var, ret_shape, allowzero=True)
             reshaped_index = op.reshape(index.var, op.const([-1]), allowzero=True)
@@ -583,7 +588,7 @@ def getitem(
     index_filtered = [x for x in index if isinstance(x, (type(None), slice))]
     axis_new_axes = [ind for ind, x in enumerate(index_filtered) if x is None]
     if len(axis_new_axes) != 0:
-        var = op.unsqueeze(var, axes=op.const(axis_new_axes))
+        var = op.unsqueeze(var, axes=op.const(axis_new_axes, dtype=np.int64))
 
     return _CoreArray(var)
 
@@ -640,14 +645,19 @@ def ndindex(shape: _CoreArray, to_reverse=None, axes_permutation=None) -> _CoreA
         axes_indices = [axes_permutation.index(i) for i in builtins.range(rank)]
 
     shape_var = shape.var
+    dtype = shape_var.unwrap_tensor().dtype
     ranges = [
         (
-            op.range(op.const(0), op.gather(shape_var, op.const(i)), op.const(1))
+            op.range(
+                op.const(0, dtype=dtype),
+                op.gather(shape_var, op.const(i)),
+                op.const(1, dtype=dtype),
+            )
             if i not in to_reverse
             else op.range(
-                op.sub(op.gather(shape_var, op.const(i)), op.const(1)),
-                op.const(-1),
-                op.const(-1),
+                op.sub(op.gather(shape_var, op.const(i)), op.const(1, dtype=dtype)),
+                op.const(-1, dtype=dtype),
+                op.const(-1, dtype=dtype),
             )
         )
         for i in builtins.range(rank)
@@ -657,7 +667,8 @@ def ndindex(shape: _CoreArray, to_reverse=None, axes_permutation=None) -> _CoreA
         op.unsqueeze(
             r,
             op.const(
-                [j for j in builtins.range(rank) if axes_indices[i] != j], dtype=np.int_
+                [j for j in builtins.range(rank) if axes_indices[i] != j],
+                dtype=np.int64,
             ),
         )
         for i, r in enumerate(ranges)
@@ -669,7 +680,7 @@ def ndindex(shape: _CoreArray, to_reverse=None, axes_permutation=None) -> _CoreA
     expanded_ranges = [op.expand(r, shape_var) for r in fit_ranges]
 
     ret = op.concat(
-        [op.unsqueeze(r, op.const([-1])) for r in expanded_ranges],
+        [op.unsqueeze(r, op.const([-1], dtype=np.int64)) for r in expanded_ranges],
         axis=-1,
     )
 
@@ -692,6 +703,8 @@ def static_map(
     input: _CoreArray, mapping: Mapping[KeyType, ValueType], default: ValueType | None
 ) -> _CoreArray:
     keys = np.array(tuple(mapping.keys()))
+    if keys.dtype == np.int32:
+        keys = keys.astype(np.int64)
     values = np.array(tuple(mapping.values()))
     value_dtype = values.dtype
     if default is None:

pyproject.toml

@@ -82,12 +82,3 @@ exclude = ["docs/"]
 
 [tool.typos.default]
 extend-ignore-identifiers-re = ["scatter_nd", "arange"]
-
-[tool.pixi.project]
-channels = ["conda-forge"]
-platforms = ["osx-arm64"]
-
-[tool.pixi.pypi-dependencies]
-ndonnx = { path = ".", editable = true }
-
-[tool.pixi.tasks]

tests/ndonnx/test_constant_propagation.py

@@ -67,13 +67,13 @@ def dynamic_masking_model(mode: Literal["lazy", "constant"]):
 
 def constant_indexing_model(mode: Literal["lazy", "constant"]):
     if mode == "constant":
-        a = ndx.asarray([0, 1, 2, 3])
+        a = ndx.asarray([0, 1, 2, 3], dtype=ndx.int64)
     else:
         a = ndx.array(
             shape=("N",),
             dtype=ndx.int64,
         )
-    b = ndx.asarray([5, 7, 8, 8, 9, 9, 234])
+    b = ndx.asarray([5, 7, 8, 8, 9, 9, 234], dtype=ndx.int64)
     idx = ndx.asarray([1, 3, 5, 0])
     result = a * b[idx]
     return ndx.build({"a": a} if mode == "lazy" else {}, {"y": result})

tests/ndonnx/test_core.py

@@ -81,9 +81,11 @@ def test_null_promotion():
 
 
 def test_asarray():
-    a = ndx.asarray([1, 2, 3])
+    a = ndx.asarray([1, 2, 3], dtype=ndx.int64)
     assert a.dtype == ndx.int64
-    np.testing.assert_array_equal(np.array([1, 2, 3]), a.to_numpy(), strict=True)
+    np.testing.assert_array_equal(
+        np.array([1, 2, 3], np.int64), a.to_numpy(), strict=True
+    )
 
 
 def test_asarray_masked():
@@ -366,7 +368,7 @@ def test_matrix_transpose():
     model = ndx.build({"a": a}, {"b": b})
     np.testing.assert_equal(
         npx.matrix_transpose(npx.reshape(npx.arange(3 * 2 * 3), (3, 2, 3))),
-        run(model, {"a": np.arange(3 * 2 * 3).reshape(3, 2, 3)})["b"],
+        run(model, {"a": np.arange(3 * 2 * 3, dtype=np.int64).reshape(3, 2, 3)})["b"],
     )
 
 
@@ -377,7 +379,7 @@ def test_matrix_transpose_attribute():
     model = ndx.build({"a": a}, {"b": b})
     np.testing.assert_equal(
         npx.reshape(npx.arange(3 * 2 * 3), (3, 2, 3)).mT,
-        run(model, {"a": np.arange(3 * 2 * 3).reshape(3, 2, 3)})["b"],
+        run(model, {"a": np.arange(3 * 2 * 3, dtype=np.int64).reshape(3, 2, 3)})["b"],
     )
 
 
@@ -388,7 +390,7 @@ def test_transpose_attribute():
     model = ndx.build({"a": a}, {"b": b})
     np.testing.assert_equal(
         npx.reshape(npx.arange(3 * 2), (3, 2)).T,
-        run(model, {"a": np.arange(3 * 2).reshape(3, 2)})["b"],
+        run(model, {"a": np.arange(3 * 2, dtype=np.int64).reshape(3, 2)})["b"],
     )
 
 
@@ -399,7 +401,7 @@ def test_array_spox_interoperability():
     model = ndx.build({"a": a}, {"b": b})
     expected = npx.reshape(npx.arange(3 * 2), (3, 2)) + 5
     input = np.ma.masked_array(
-        np.arange(3 * 2).reshape(3, 2), mask=np.ones((3, 2), dtype=bool)
+        np.arange(3 * 2, dtype=np.int64).reshape(3, 2), mask=np.ones((3, 2), dtype=bool)
     )
     actual = run(model, {"a": input})["b"]
     np.testing.assert_equal(expected, actual)