Release: NDSL 2026.03.00

.github/PULL_REQUEST_TEMPLATE/release-patch.md

@@ -29,4 +29,4 @@ What to do to actually release:
 What to do after a release:
 
 - [ ] update the pace PR from the pre-commit checklist to include the released version of NDSL and merge it.
-- [ ] in NDSL, merge `main` back into `develop` (potentially adding a commit to fix the issue "properly")
+- [ ] in NDSL, merge `main` back into `develop` (potentially adding a commit to fix the issue "properly") to have all changes in develop and ensure `setuptools_scm` finds the latest release tag

.github/PULL_REQUEST_TEMPLATE/release.md

@@ -22,5 +22,6 @@ What to do to actually release:
 
 What to do after a release:
 
+- [ ] merge `main` down into `develop` to ensure `setuptools_scm` finds the latest release tag
 - [ ] update the pace PR from the pre-commit checklist to include the released version of NDSL and merge it.
 - [ ] merge breaking changes in NDSL (e.g. search for deprecation warnings)

.github/workflows/fv3_translate_tests.yaml

@@ -10,10 +10,7 @@ on:
 
 jobs:
   fv3_translate_tests:
-    # TODO
-    # restore once NDSL 2026.02.00 is released and pyFV3 is updated.
-    # uses: NOAA-GFDL/pyFV3/.github/workflows/translate.yaml@develop
-    uses: romanc/pyFV3/.github/workflows/translate.yaml@noop
+    uses: NOAA-GFDL/pyFV3/.github/workflows/translate.yaml@develop
     with:
       component_trigger: true
       component_name: NDSL

.github/workflows/pace_tests.yaml

@@ -10,10 +10,7 @@ on:
 
 jobs:
   pace_main_tests:
-    # TODO
-    # restore once NDSL 2026.02.00 is released and pace is updated.
-    # uses: NOAA-GFDL/pace/.github/workflows/main_unit_tests.yaml@develop
-    uses: romanc/pace/.github/workflows/main_unit_tests.yaml@noop
+    uses: NOAA-GFDL/pace/.github/workflows/main_unit_tests.yaml@develop
     with:
       component_trigger: true
       component_name: NDSL

.github/workflows/shield_tests.yaml

@@ -10,10 +10,7 @@ on:
 
 jobs:
   shield_translate_tests:
-    # TODO
-    # restore once NDSL 2026.02.00 is released and pySHiELD is updated.
-    # uses: NOAA-GFDL/pySHiELD/.github/workflows/translate.yaml@develop
-    uses: romanc/pySHiELD/.github/workflows/translate.yaml@noop
+    uses: NOAA-GFDL/pySHiELD/.github/workflows/translate.yaml@develop
     with:
       component_trigger: true
       component_name: NDSL

.github/workflows/unit_tests.yaml

@@ -18,7 +18,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: ['3.11', '3.12']
+        python-version: ['3.11', '3.12', '3.13']
     name: Python ${{ matrix.python-version }}
     steps:
       - name: Checkout repository

examples/NDSL/03_orchestration_basics.ipynb

@@ -35,7 +35,7 @@
     "    orchestrate,\n",
     "    QuantityFactory,\n",
     ")\n",
-    "from ndsl.constants import X_DIM, Y_DIM, Z_DIM\n",
+    "from ndsl.constants import I_DIM, J_DIM, K_DIM\n",
     "from ndsl.dsl.typing import FloatField, Float\n",
     "from ndsl.boilerplate import get_factories_single_tile_orchestrated"
    ]
@@ -93,7 +93,7 @@
     "            domain=grid_indexing.domain_compute(),\n",
     "        )\n",
     "        self._tmp_field = quantity_factory.zeros(\n",
-    "            [X_DIM, Y_DIM, Z_DIM], \"n/a\", dtype=dtype\n",
+    "            [I_DIM, J_DIM, K_DIM], \"n/a\", dtype=dtype\n",
     "        )\n",
     "        self._n_halo = quantity_factory.sizer.n_halo\n",
     "\n",
@@ -134,9 +134,9 @@
     "    )\n",
     "    local_sum = LocalSum(stencil_factory, qty_factory)\n",
     "\n",
-    "    in_field = qty_factory.zeros([X_DIM, Y_DIM, Z_DIM], \"n/a\", dtype=dtype)\n",
+    "    in_field = qty_factory.zeros([I_DIM, J_DIM, K_DIM], \"n/a\", dtype=dtype)\n",
     "    in_field.view[:] = 2.0\n",
-    "    out_field = qty_factory.zeros([X_DIM, Y_DIM, Z_DIM], \"n/a\", dtype=dtype)\n",
+    "    out_field = qty_factory.zeros([I_DIM, J_DIM, K_DIM], \"n/a\", dtype=dtype)\n",
     "\n",
     "    # Run\n",
     "    local_sum(in_field, out_field)"

ndsl/buffer.py

@@ -2,9 +2,10 @@
 
 import contextlib
 from collections.abc import Callable, Generator, Iterable
+from typing import Any
 
 import numpy as np
-from numpy.lib.index_tricks import IndexExpression
+import numpy.typing as npt
 
 from ndsl.performance.timer import NullTimer, Timer
 from ndsl.types import Allocator
@@ -16,7 +17,7 @@
 )
 
 
-BufferKey = tuple[Callable, Iterable[int], type]
+BufferKey = tuple[Callable, Iterable[int], npt.DTypeLike]
 BUFFER_CACHE: dict[BufferKey, list["Buffer"]] = {}
 
 
@@ -41,7 +42,7 @@ def __init__(self, key: BufferKey, array: np.ndarray):
 
     @classmethod
     def pop_from_cache(
-        cls, allocator: Allocator, shape: Iterable[int], dtype: type
+        cls, allocator: Allocator, shape: Iterable[int], dtype: npt.DTypeLike
     ) -> Buffer:
         """Retrieve or insert then retrieve of buffer from cache.
 
@@ -78,8 +79,8 @@ def finalize_memory_transfer(self) -> None:
     def assign_to(
         self,
         destination_array: np.ndarray,
-        buffer_slice: IndexExpression = np.index_exp[:],
-        buffer_reshape: IndexExpression = None,
+        buffer_slice: Any = np.index_exp[:],
+        buffer_reshape: Any | None = None,
     ) -> None:
         """Assign internal array to destination_array.
 
@@ -95,7 +96,7 @@ def assign_to(
             )
 
     def assign_from(
-        self, source_array: np.ndarray, buffer_slice: IndexExpression = np.index_exp[:]
+        self, source_array: np.ndarray, buffer_slice: Any = np.index_exp[:]
     ) -> None:
         """Assign source_array to internal array.
 
@@ -107,7 +108,7 @@ def assign_from(
 
 @contextlib.contextmanager
 def array_buffer(
-    allocator: Allocator, shape: Iterable[int], dtype: type
+    allocator: Allocator, shape: Iterable[int], dtype: npt.DTypeLike
 ) -> Generator[Buffer, Buffer, None]:
     """
     A context manager providing a contiguous array, which may be re-used between calls.
@@ -132,7 +133,7 @@ def send_buffer(
     allocator: Callable,
     array: np.ndarray,
     timer: Timer | None = None,
-) -> np.ndarray:
+) -> Generator[np.ndarray]:
     """A context manager ensuring that `array` is contiguous in a context where it is
     being sent as data, copying into a recycled buffer array if necessary.
 
@@ -166,7 +167,7 @@ def recv_buffer(
     allocator: Callable,
     array: np.ndarray,
     timer: Timer | None = None,
-) -> np.ndarray:
+) -> Generator[np.ndarray]:
     """A context manager ensuring that array is contiguous in a context where it is
     being used to receive data, using a recycled buffer array and then copying the
     result into array if necessary.

ndsl/comm/communicator.py

@@ -2,6 +2,7 @@
 
 import abc
 from collections.abc import Mapping, Sequence
+from types import ModuleType
 from typing import Any, Self, cast
 
 import numpy as np
@@ -16,7 +17,6 @@
 from ndsl.optional_imports import cupy
 from ndsl.performance.timer import NullTimer, Timer
 from ndsl.quantity import Quantity, QuantityHaloSpec, QuantityMetadata
-from ndsl.types import NumpyModule
 
 
 def to_numpy(array, dtype=None) -> np.ndarray:  # type: ignore[no-untyped-def]
@@ -83,7 +83,7 @@ def size(self) -> int:
         """Total number of ranks in this communicator"""
         return self.comm.Get_size()
 
-    def _maybe_force_cpu(self, module: NumpyModule) -> NumpyModule:
+    def _maybe_force_cpu(self, module: ModuleType) -> ModuleType:
         """
         Get a numpy-like module depending on configuration and
         Quantity original allocator.
@@ -223,7 +223,7 @@ def _get_gather_recv_quantity(
     ) -> Quantity:
         """Initialize a Quantity for use when receiving global data during gather"""
         recv_quantity = Quantity(
-            send_metadata.np.zeros(global_extent, dtype=send_metadata.dtype),  # type: ignore
+            send_metadata.np.zeros(global_extent, dtype=send_metadata.dtype),
             dims=send_metadata.dims,
             units=send_metadata.units,
             origin=tuple([0 for dim in send_metadata.dims]),
@@ -238,7 +238,7 @@ def _get_scatter_recv_quantity(
     ) -> Quantity:
         """Initialize a Quantity for use when receiving subtile data during scatter"""
         recv_quantity = Quantity(
-            send_metadata.np.zeros(shape, dtype=send_metadata.dtype),  # type: ignore
+            send_metadata.np.zeros(shape, dtype=send_metadata.dtype),
             dims=send_metadata.dims,
             units=send_metadata.units,
             backend=send_metadata.backend,
@@ -837,7 +837,7 @@ def _get_gather_recv_quantity(
         # needs to change the quantity dimensions since we add a "tile" dimension,
         # unlike for tile scatter/gather which retains the same dimensions
         recv_quantity = Quantity(
-            metadata.np.zeros(global_extent, dtype=metadata.dtype),  # type: ignore
+            metadata.np.zeros(global_extent, dtype=metadata.dtype),
             dims=(constants.TILE_DIM,) + metadata.dims,
             units=metadata.units,
             origin=(0,) + tuple([0 for dim in metadata.dims]),
@@ -859,7 +859,7 @@ def _get_scatter_recv_quantity(
         # needs to change the quantity dimensions since we remove a "tile" dimension,
         # unlike for tile scatter/gather which retains the same dimensions
         recv_quantity = Quantity(
-            metadata.np.zeros(shape, dtype=metadata.dtype),  # type: ignore
+            metadata.np.zeros(shape, dtype=metadata.dtype),
             dims=metadata.dims[1:],
             units=metadata.units,
             backend=metadata.backend,

ndsl/constants.py

@@ -38,13 +38,13 @@ def _get_constant_version(
 # Common constants
 #####################
 
-I_DIM = X_DIM = "i"
-I_INTERFACE_DIM = X_INTERFACE_DIM = "i_interface"
-J_DIM = Y_DIM = "j"
-J_INTERFACE_DIM = Y_INTERFACE_DIM = "j_interface"
-K_DIM = Z_DIM = "k"
-K_INTERFACE_DIM = Z_INTERFACE_DIM = "k_interface"
-K_SOIL_DIM = Z_SOIL_DIM = "k_soil"
+I_DIM = "i"
+I_INTERFACE_DIM = "i_interface"
+J_DIM = "j"
+J_INTERFACE_DIM = "j_interface"
+K_DIM = "k"
+K_INTERFACE_DIM = "k_interface"
+K_SOIL_DIM = "k_soil"
 
 I_DIMS = (I_DIM, I_INTERFACE_DIM)
 J_DIMS = (J_DIM, J_INTERFACE_DIM)

ndsl/dsl/__init__.py

@@ -1,6 +1,5 @@
 # Literal precision for both GT4Py & NDSL
 import os
-import platform
 import sys
 from typing import Literal
 
@@ -36,15 +35,6 @@ def _get_literal_precision(default: Literal["32", "64"] = "64") -> Literal["32",
 os.environ["GT4PY_LITERAL_INT_PRECISION"] = str(NDSL_GLOBAL_PRECISION)
 os.environ["GT4PY_LITERAL_FLOAT_PRECISION"] = str(NDSL_GLOBAL_PRECISION)
 
-# OpenMP handling
-
-detected_macos = platform.system() == "Darwin"
-if detected_macos:
-    ndsl_log.warning(
-        "Multithreading is deactivated under MacOS due to apple-clang not handling OpenMP by default."
-    )
-os.environ["GT4PY_CARTESIAN_ENABLE_OPENMP"] = "False" if detected_macos else "True"
-
 
 # Set cache names for default gt backends workflow
 import gt4py.cartesian.config  # noqa: E402

ndsl/dsl/dace/dace_config.py

@@ -7,6 +7,7 @@
 
 import dace.config
 from gt4py.cartesian.config import GT4PY_COMPILE_OPT_LEVEL
+from gt4py.cartesian.utils.compiler import cxx_compiler_defaults, gpu_configuration
 
 from ndsl import LocalComm
 from ndsl.comm.communicator import Communicator
@@ -226,23 +227,18 @@ def __init__(
             else:
                 dace.config.Config.set("compiler", "build_type", value="Release")
 
-            # Required to True for gt4py storage/memory
-            dace.config.Config.set(
-                "compiler",
-                "allow_view_arguments",
-                value=True,
-            )
             # Resolve "march/mtune" option for GPU
             # - turn on numeric-centric SSE by default
             # - Neoverse-V2 Grace CPU is too new for GCC 14 and -march=native will fail
             # - use alternative march=armv8-a instead
             march_cpu = "armv8-a" if is_arm_neoverse else "native"
             # Removed --fmath
+            cxx_defaults = cxx_compiler_defaults(GT4PY_COMPILE_OPT_LEVEL)
             dace.config.Config.set(
                 "compiler",
                 "cpu",
                 "args",
-                value=f"-march={march_cpu} -std=c++17 -fPIC -Wall -Wextra -O{optimization_level}",
+                value=f"-march={march_cpu} -std=c++17 -fPIC -Wall -Wextra -O{optimization_level} {cxx_defaults.cxx_compile_flags}",
             )
             # Potentially buggy - deactivate
             dace.config.Config.set(
@@ -257,11 +253,12 @@ def __init__(
             # - use alternative mcpu=native instead
             march_option = "-mcpu=native" if is_arm_neoverse else "-march=native"
             # Removed --fast-math
+            gpu_config = gpu_configuration(GT4PY_COMPILE_OPT_LEVEL)
             dace.config.Config.set(
                 "compiler",
                 "cuda",
                 "args",
-                value=f"-std=c++14 -Xcompiler -fPIC -O3 -Xcompiler {march_option}",
+                value=f"-std=c++14 -Xcompiler -fPIC -O{optimization_level} -Xcompiler {march_option} {gpu_config.gpu_compile_flags}",
             )
 
             cuda_sm = cp.cuda.Device(0).compute_capability if cp else 60
@@ -280,6 +277,14 @@ def __init__(
                 "max_concurrent_streams",
                 value=-1,  # no concurrent streams, every kernel on defaultStream
             )
+
+            # Required to True for gt4py storage/memory
+            dace.config.Config.set(
+                "compiler",
+                "allow_view_arguments",
+                value=True,
+            )
+
             # Speed up built time
             dace.config.Config.set(
                 "compiler",

ndsl/dsl/dace/orchestration.py

@@ -122,7 +122,7 @@ def _to_gpu(sdfg: SDFG) -> None:
 def _simplify(
     sdfg: SDFG,
     *,
-    validate: bool = True,
+    validate: bool = False,
     validate_all: bool = False,
     verbose: bool = False,
 ) -> dict | None:
@@ -146,9 +146,6 @@ def _build_sdfg(
     backend_name = config.get_backend()
 
     if is_compiling:
-        with DaCeProgress(config, "Validate original SDFG"):
-            sdfg.validate()
-
         # Fully specialize all known symbols and then propagate these changes in the simplify
         # pass that follows. This is not only a smart idea in general, but also simplifies (haha)
         # the schedule tree (optimization) roundtrip.
@@ -271,9 +268,6 @@ def _build_sdfg(
                 negative_delp_checker(sdfg)
                 negative_qtracers_checker(sdfg)
 
-        with DaCeProgress(config, "Validate before compile"):
-            sdfg.validate()
-
         # Compile
         with DaCeProgress(config, "Codegen & compile"):
             sdfg.compile()
@@ -646,9 +640,7 @@ def __call__(self, *arg, **kwarg):  # type: ignore[no-untyped-def]
                 return wrapped(*arg, **kwarg)
 
             def __sdfg__(self, *args, **kwargs):  # type: ignore[no-untyped-def]
-                sdfg = wrapped.__sdfg__(*args, **kwargs)
-                sdfg.validate()
-                return sdfg
+                return wrapped.__sdfg__(*args, **kwargs)
 
             def __sdfg_closure__(self, reevaluate=None):  # type: ignore[no-untyped-def]
                 return wrapped.__sdfg_closure__(reevaluate)