More changes for dropping 3.10 / adding 3.15

.github/actions/check-doc-links/action.yml

@@ -16,7 +16,7 @@ inputs:
     description: "check allows up to 10 rate-limited URLs; cache-warm warns on all link failures"
     default: check
 
-# Callers provide Python 3.10+ on PATH; the helpers use only its standard library.
+# Callers provide Python 3.11+ on PATH; the helpers use only its standard library.
 runs:
   using: composite
   steps:

.github/workflows/build-wheel.yml

@@ -442,14 +442,7 @@ jobs:
         uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # v7.0.0
         with:
           python-version: ${{ matrix.python-version }}
-          # TODO: remove allow-prereleases once 3.15 is officially supported
-          allow-prereleases: ${{ startsWith(matrix.python-version, '3.15') }}
-
-      - name: Enable Scientific Python Nightly Wheels for Python 3.15
-        if: ${{ (env.TEST_BINDINGS == 'true' || env.TEST_CORE == 'true') && startsWith(matrix.python-version, '3.15') }}
-        run: |
-          echo "PIP_EXTRA_INDEX_URL=https://pypi.anaconda.org/scientific-python-nightly-wheels/simple" >> "$GITHUB_ENV"
-          echo "PIP_ONLY_BINARY=numpy" >> "$GITHUB_ENV"
+          allow-prereleases: ${{ startsWith(matrix.python-version, '3.16') }}
 
       - name: verify free-threaded build
         if: ${{ (env.TEST_BINDINGS == 'true' || env.TEST_CORE == 'true') && endsWith(matrix.python-version, 't') }}

.github/workflows/test-wheel-linux.yml

@@ -110,8 +110,6 @@ jobs:
       fail-fast: false
       matrix: ${{ fromJSON(needs.compute-matrix.outputs.MATRIX) }}
     runs-on: "${{ matrix.FLAVOR || 'linux' }}-${{ matrix.ARCH }}-gpu-${{ matrix.GPU }}-${{ matrix.RUNNER_DRIVER }}-${{ matrix.GPU_COUNT }}"
-    # TODO: remove continue-on-error once 3.15 is officially supported
-    continue-on-error: ${{ startsWith(matrix.PY_VER, '3.15') }}
     # The build stage could fail but we want the CI to keep moving.
     if: ${{ github.repository_owner == 'nvidia' && !cancelled() }}
     # Our self-hosted runners require a container
@@ -329,19 +327,11 @@ jobs:
         uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # v7.0.0
         with:
           python-version: ${{ matrix.PY_VER }}
-          # TODO: remove allow-prereleases once 3.15 is officially supported
-          allow-prereleases: ${{ startsWith(matrix.PY_VER, '3.15') }}
+          allow-prereleases: ${{ startsWith(matrix.PY_VER, '3.16') }}
         env:
           # we use self-hosted runners on which setup-python behaves weirdly (Python include can't be found)...
           AGENT_TOOLSDIRECTORY: "/opt/hostedtoolcache"
 
-      - name: Enable Scientific Python Nightly Wheels for Python 3.15
-        if: ${{ (env.TEST_BINDINGS == 'true' || env.TEST_CORE == 'true' || env.TEST_PYTHON == 'true') &&
-                startsWith(matrix.PY_VER, '3.15') }}
-        run: |
-          echo "PIP_EXTRA_INDEX_URL=https://pypi.anaconda.org/scientific-python-nightly-wheels/simple" >> "$GITHUB_ENV"
-          echo "PIP_ONLY_BINARY=numpy" >> "$GITHUB_ENV"
-
       - name: Set up mini CTK
         if: ${{ matrix.LOCAL_CTK == '1' }}
         uses: ./.github/actions/fetch_ctk

.github/workflows/test-wheel-windows.yml

@@ -101,8 +101,6 @@ jobs:
       fail-fast: false
       matrix: ${{ fromJSON(needs.compute-matrix.outputs.MATRIX) }}
     if: ${{ github.repository_owner == 'nvidia' && !cancelled() }}
-    # TODO: remove continue-on-error once 3.15 is officially supported
-    continue-on-error: ${{ startsWith(matrix.PY_VER, '3.15') }}
     runs-on: "windows-${{ matrix.ARCH }}-gpu-${{ matrix.GPU }}-${{ matrix.RUNNER_DRIVER }}-${{ matrix.GPU_COUNT }}"
     steps:
       - name: Enable Git symlinks (Windows, must precede checkout)
@@ -311,16 +309,7 @@ jobs:
         uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # v7.0.0
         with:
           python-version: ${{ matrix.PY_VER }}
-          # TODO: remove allow-prereleases once 3.15 is officially supported
-          allow-prereleases: ${{ startsWith(matrix.PY_VER, '3.15') }}
-
-      - name: Enable Scientific Python Nightly Wheels for Python 3.15
-        if: ${{ (env.TEST_BINDINGS == 'true' || env.TEST_CORE == 'true' || env.TEST_PYTHON == 'true') &&
-                startsWith(matrix.PY_VER, '3.15') }}
-        shell: bash --noprofile --norc -xeuo pipefail {0}
-        run: |
-          echo "PIP_EXTRA_INDEX_URL=https://pypi.anaconda.org/scientific-python-nightly-wheels/simple" >> "$GITHUB_ENV"
-          echo "PIP_ONLY_BINARY=numpy" >> "$GITHUB_ENV"
+          allow-prereleases: ${{ startsWith(matrix.PY_VER, '3.16') }}
 
       - name: Verify LongPathsEnabled
         run: |

ci/ci-pipeline.mmd

@@ -34,7 +34,7 @@ flowchart TB
             direction LR
             B1["linux-64<br/>(Self-hosted)"] ~~~ B2["linux-aarch64<br/>(Self-hosted)"] ~~~ B3["win-64<br/>(GitHub-hosted)"]
         end
-        BUILD_DETAILS["Python versions: 3.10, 3.11, 3.12, 3.13, 3.14<br/>CUDA version: 13.0.0 (build-time)<br/>Components: cuda-core, cuda-bindings,<br/>cuda-pathfinder, cuda-python"]
+        BUILD_DETAILS["Python versions: 3.11, 3.12, 3.13, 3.14, 3.15<br/>CUDA version: 13.0.0 (build-time)<br/>Components: cuda-core, cuda-bindings,<br/>cuda-pathfinder, cuda-python"]
     end
 
     subgraph ARTIFACTS["ARTIFACT STORAGE"]

ci/tools/download-wheels

@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
 
-# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 #
 # SPDX-License-Identifier: Apache-2.0
 
@@ -56,13 +56,6 @@ do
         continue
     fi
 
-    # TODO: remove the Python 3.15 exclusion once 3.15 is officially supported
-    # exclude pre-release Python (3.15) wheels from releasing
-    if [[ "${p}" == *python315* ]]; then
-        echo "Skipping pre-release Python artifact: $p"
-        continue
-    fi
-
     # If we're not downloading "all", only process matching component
     if [[ "$COMPONENT" != "all" && "$p" != ${COMPONENT}* ]]; then
         continue

cuda_bindings/pixi.toml

@@ -21,7 +21,6 @@ pytest-repeat = "*"
 pyglet = ">=2.1.9"
 numpy = "*"
 packaging = "*"
-tomli = "*"
 matplotlib = "*"
 
 [feature.test.pypi-dependencies]

cuda_bindings/pyproject.toml

@@ -27,6 +27,7 @@ classifiers = [
     "Programming Language :: Python :: 3.12",
     "Programming Language :: Python :: 3.13",
     "Programming Language :: Python :: 3.14",
+    "Programming Language :: Python :: 3.15",
     "Environment :: GPU :: NVIDIA CUDA",
 ]
 dynamic = ["version", "readme"]
@@ -54,7 +55,6 @@ test = [
     "pytest-github-actions-annotate-failures==0.4.2",
     "pyglet==2.1.14",
     "packaging>=24",
-    "tomli>=2; python_version < '3.11'",
 ]
 test-ft = [
     "pytest-run-parallel==0.10.0",

cuda_core/AGENTS.md

@@ -224,8 +224,8 @@ below are for contributors. Reviewers and agents should flag violations.
   lock, and when a lock must stay held call `p_`, keep the status, and report
   after the lock is released (`deviceptr_import_ipc` is the model).
 - **Rollback failure**: the original exception propagates; the failed rollback
-  is attached to it with `attach_rollback_failure()` (a PEP 678 note on
-  Python 3.11+, reported out-of-band on 3.10), or chained with
+  is attached to it with `attach_rollback_failure()` (a PEP 678 note),
+  or chained with
   `raise ... from` when a second exception must be raised. Catching everything
   (bare `except:` or `except BaseException:`) is acceptable only for
   rollback-then-`raise` blocks, where the rollback must also run for

cuda_core/cuda/core/_cpp/rt/DESIGN.md

@@ -388,8 +388,7 @@ observes the caller's context. When restoration fails after the operation
 succeeded, the creation is undone and the restoration status is returned. When both fail, the
 operation status is returned. Either way the helper records a thread-local
 detail keyed to the returned status (`take_last_error_detail(status)`) that
-`_check_driver_error` attaches to the raised `CUDAError` as a PEP 678 note
-(appended to the message on Python 3.10), so the user learns that the caller's
+`_check_driver_error` attaches to the raised `CUDAError` as a PEP 678 note, so the user learns that the caller's
 context was not restored, which context is current and, for a double failure,
 why restoration failed. Keying the detail to its status narrows, but does not
 remove, misattribution: a caller that drops the status (an empty handle raised
@@ -428,7 +427,7 @@ exactly one of these; none is ever dropped.
 | Cython, on a path that can raise | `HANDLE_RETURN(status)` | Raises `CUDAError`. A restoration detail recorded by the C++ helper becomes a note on the exception. |
 | Cython, after a handle constructor returned an empty handle | `HANDLE_RETURN(get_last_error())`, immediately | Same. Transitional: #2760 makes constructors return the status instead. |
 | C++, a helper that runs an operation in another context | Return the `CUresult`; `exit_context` records the restoration detail | Cython raises it. Transitional: #2760 returns the restoration status as a second out-parameter. |
-| Cython, inside an `except` block whose rollback failed | `attach_rollback_failure(op, status, detail)` | Adds a note to the exception being handled. Reports instead if nothing is being handled or notes do not exist (Python 3.10). |
+| Cython, inside an `except` block whose rollback failed | `attach_rollback_failure(op, status, detail)` | Adds a note to the exception being handled. Reports instead if nothing is being handled. |
 | C++, a deleter or deferred cleanup | A `pw_*` wrapper, or `report_cuda_error()` / `report_message()` | Emits `CUDAWarning`. Never raises. |
 | Cython or Python, a `__dealloc__` or destructor-path callback | `warnings.warn(msg, CUDAWarning, stacklevel=2)` | Same. |
 | A CUDA callback thread | Nothing that needs the GIL. Hand the work to the deferred-cleanup queue with `Py_AddPendingCall` | CUDA forbids driver calls there, and acquiring the GIL there can deadlock with a GIL holder blocked in a driver call. GIL-free C API that only schedules work is fine. |

cuda_core/cuda/core/_memory/_copy_enums.py

@@ -6,11 +6,11 @@
 
 import dataclasses
 from collections.abc import Sequence
+from enum import StrEnum
 
 from cuda.core._device import Device
 from cuda.core._host import Host
 from cuda.core._utils.cuda_utils import driver
-from cuda.core._utils.pycompat import StrEnum
 
 __all__ = ["CopyOptions", "MemcpyOverlapMode", "MemcpySrcAccessOrder"]
 

cuda_core/cuda/core/_utils/cuda_utils.pyx

@@ -167,13 +167,8 @@ cdef object _NVRTC_SUCCESS = nvrtc.nvrtcResult.NVRTC_SUCCESS
 
 
 cdef inline void _attach_detail(exc, str detail):
-    # PEP 678 notes (Python 3.11+) keep the detail separable from the message;
-    # older interpreters get it appended to the message instead.
-    add_note = getattr(exc, "add_note", None)
-    if add_note is not None:
-        add_note(detail)
-    else:
-        exc.args = (f"{exc.args[0]} ({detail})", *exc.args[1:])
+    # PEP 678 notes keep the detail separable from the message.
+    exc.add_note(detail)
 
 
 cpdef inline int _check_driver_error(cydriver.CUresult error) except?-1 nogil:

cuda_core/cuda/core/_utils/pycompat.py

@@ -7,9 +7,8 @@
 """
 
 import sys
-from enum import StrEnum
 
-__all__ = ["BufferProtocol", "StrEnum"]
+__all__ = ["BufferProtocol"]
 
 
 if sys.version_info >= (3, 12):

cuda_core/cuda/core/system/typing.py

@@ -2,9 +2,10 @@
 #
 # SPDX-License-Identifier: Apache-2.0
 
+from enum import StrEnum
+
 from cuda.bindings import nvml as _nvml
 from cuda.bindings._internal._fast_enum import FastEnum as _FastEnum
-from cuda.core._utils.pycompat import StrEnum
 
 __all__ = [
     "AddressingMode",

cuda_core/pixi.toml

@@ -292,11 +292,6 @@ cuda-version = "*"
 numpy = "*"
 cuda-bindings = "*"
 cuda-pathfinder = "*"
-# Conda-forge ships per-Python variants of backports.strenum (real impl on
-# <3.11, no-op stub on >=3.12); pixi-build doesn't accept PEP 508-style
-# `python_version` selectors in [package.target.<X>], so list it
-# unconditionally and let the solver pick the right variant.
-"backports.strenum" = "*"
 
 [target.linux.tasks.build-cython-tests]
 cmd = ["$PIXI_PROJECT_ROOT/tests/cython/build_tests.sh"]

cuda_core/pyproject.toml

@@ -41,6 +41,7 @@ classifiers = [
     "Programming Language :: Python :: 3.12",
     "Programming Language :: Python :: 3.13",
     "Programming Language :: Python :: 3.14",
+    "Programming Language :: Python :: 3.15",
     "Programming Language :: Python :: Implementation :: CPython",
     "Environment :: GPU :: NVIDIA CUDA",
     "Environment :: GPU :: NVIDIA CUDA :: 12",
@@ -78,8 +79,7 @@ test = [
     "docutils==0.23",
     # Required by tests/test_graphics.py; pinned to match cuda_bindings.
     "pyglet==2.1.14",
-    # TODO: remove the Python 3.15 guard once 3.15 is officially supported
-    "cffi==2.0.0; python_version < '3.15'",
+    "cffi==2.0.0",
 ]
 # TODO: drop the Windows 3.15 guard once ml-dtypes publishes cp315 Windows wheels
 ml-dtypes = ["ml-dtypes>=0.5.4,<0.6.0; sys_platform != 'win32' or python_version < '3.15'"]

cuda_core/tests/graph/test_graph_builder.py

@@ -4,7 +4,6 @@
 """GraphBuilder stream capture tests."""
 
 import gc
-import sys
 import time
 import weakref
 
@@ -273,26 +272,19 @@ def test_graph_join_failure_survives_failing_close(init_cuda):
     Joining a builder from another capture is refused by the driver, which
     also invalidates that capture, so closing it afterwards fails. The cleanup
     must still close the remaining builders and raise the join error; the
-    failed close is attached to it as a note (Python 3.11+) or reported as a
-    CUDAWarning (Python 3.10).
+    failed close is attached to it as a note.
     """
-    has_notes = sys.version_info >= (3, 11)
     root = Device().create_graph_builder().begin_building()
     root, fork = root.split(2)
     other1 = Device().create_graph_builder().begin_building()
     other2 = Device().create_graph_builder().begin_building()
 
     with pytest.raises(CUDAError, match="^CUDA_ERROR_STREAM_CAPTURE_MERGE") as excinfo:
-        if has_notes:
-            GraphBuilder.join(root, other1, other2, fork)
-        else:
-            with pytest.warns(CUDAWarning, match="CUDA_ERROR_STREAM_CAPTURE_INVALIDATED"):
-                GraphBuilder.join(root, other1, other2, fork)
+        GraphBuilder.join(root, other1, other2, fork)
     assert other1.is_closed and other2.is_closed and fork.is_closed
-    if has_notes:
-        (note,) = excinfo.value.__notes__
-        assert "cuStreamEndCapture failed while closing a graph builder" in note
-        assert "CUDA_ERROR_STREAM_CAPTURE_INVALIDATED" in note
+    (note,) = excinfo.value.__notes__
+    assert "cuStreamEndCapture failed while closing a graph builder" in note
+    assert "CUDA_ERROR_STREAM_CAPTURE_INVALIDATED" in note
 
     # The refused wait invalidated the root's capture as well; ending it
     # raises the driver error and the builder still closes cleanly.

cuda_core/tests/test_enum_coverage.py

@@ -8,7 +8,6 @@
 # compatible cuda.bindings version.
 
 import inspect
-import sys
 from enum import StrEnum
 from typing import Any
 
@@ -387,7 +386,6 @@ def test_wrapper_covers_all_binding_members(binding, str_enum, mapping, binding_
             )
 
 
-@pytest.mark.skipif(sys.version_info < (3, 11), reason="Requires Python 3.11+ for StrEnum")
 @pytest.mark.parametrize(
     "module",
     _MODULES,

cuda_core/tests/test_error_handling.py

@@ -14,7 +14,6 @@
 """
 
 import ctypes
-import sys
 import warnings
 from contextlib import contextmanager
 
@@ -43,9 +42,6 @@
 INVALID_VALUE = int(driver.CUresult.CUDA_ERROR_INVALID_VALUE)
 DEINITIALIZED = int(driver.CUresult.CUDA_ERROR_DEINITIALIZED)
 
-# PEP 678 exception notes; on 3.10 the same information lands in the message.
-HAS_NOTES = sys.version_info >= (3, 11)
-
 thread_unsafe_context_fault = pytest.mark.thread_unsafe(
     reason="injects a thread-local restoration fault and mutates the CUDA context stack"
 )
@@ -101,10 +97,9 @@ def test_create_stream_raises_when_context_cannot_be_restored(init_cuda):
         assert "could not be restored" in text
         assert "CUDA_ERROR_INVALID_CONTEXT" in text
         assert "Device.set_current()" in text
-        if HAS_NOTES:
-            # The explanation is a note, separable from the driver error message.
-            assert "could not be restored" not in str(excinfo.value)
-            assert any("could not be restored" in note for note in excinfo.value.__notes__)
+        # The explanation is a note, separable from the driver error message.
+        assert "could not be restored" not in str(excinfo.value)
+        assert any("could not be restored" in note for note in excinfo.value.__notes__)
         # As documented, a failed restoration leaves the device's context current.
         assert current_context_handle() == int(dev.context.handle)
     assert current_context_handle() == int(dev.context.handle)
@@ -303,28 +298,20 @@ def as_bytes(buffer):
 @pytest.mark.agent_authored(model="claude-fable-5-1")
 def test_rollback_failure_is_attached_to_the_propagating_exception():
     """A failed rollback inside an except block becomes a note on the exception being
-    handled (Python 3.11+); on 3.10, or with no exception being handled, it is reported
-    as a CUDAWarning. CUDA_ERROR_DEINITIALIZED is neither attached nor reported."""
+    handled; with no exception being handled, it is reported as a CUDAWarning. CUDA_ERROR_DEINITIALIZED is neither attached nor reported."""
     with pytest.raises(RuntimeError) as excinfo:
         try:
             raise RuntimeError("primary failure")
         except RuntimeError:
-            if HAS_NOTES:
-                with assert_no_cuda_warning():
-                    _attach_rollback_failure_for_testing(INVALID_VALUE)
-            else:
-                with pytest.warns(CUDAWarning, match="cuTestOperation failed while testing"):
-                    _attach_rollback_failure_for_testing(INVALID_VALUE)
+            with assert_no_cuda_warning():
+                _attach_rollback_failure_for_testing(INVALID_VALUE)
             with assert_no_cuda_warning():
                 _attach_rollback_failure_for_testing(DEINITIALIZED)
             raise
     exc = excinfo.value
     assert str(exc) == "primary failure"
-    if HAS_NOTES:
-        assert len(exc.__notes__) == 1
-        assert "cuTestOperation failed while testing: CUDA_ERROR_INVALID_VALUE" in exc.__notes__[0]
-    else:
-        assert not hasattr(exc, "__notes__")
+    assert len(exc.__notes__) == 1
+    assert "cuTestOperation failed while testing: CUDA_ERROR_INVALID_VALUE" in exc.__notes__[0]
     # With no exception being handled there is nothing to attach to.
     with pytest.warns(CUDAWarning, match="cuTestOperation failed while testing"):
         _attach_rollback_failure_for_testing(INVALID_VALUE)

cuda_pathfinder/pyproject.toml

@@ -54,8 +54,7 @@ cu13 = [
     "nvidia-nvshmem-cu13<3.7; sys_platform != 'win32'",
 ]
 host = [
-    # TODO: remove the Python 3.15 guard once 3.15 is officially supported
-    "nvidia-cutlass; python_version < '3.15'",
+    "nvidia-cutlass",
     "nvpl-fft; platform_system == 'Linux' and platform_machine == 'aarch64'",
 ]