Merge remote-tracking branch 'upstream/main' into cuda.core.system-affinity

Author: mdboom

.github/workflows/build-docs.yml

@@ -47,6 +47,12 @@ jobs:
       run:
         shell: bash -el {0}
     steps:
+      - name: validate build-ctk
+        run: |
+          if [ ! "${{ inputs.build-ctk-ver }}" =~ ^[0-9]+\.[0-9]+\.[0-9]+$ ]; then
+            echo "error: `build-ctk-ver` ${{ inputs.build-ctk-ver }} version does not match MAJOR.MINOR.MICRO" >&2
+            exit 1
+          fi
       - name: Checkout ${{ github.event.repository.name }}
         uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8  # v6.0.1
         with:

.github/workflows/ci.yml

@@ -62,7 +62,12 @@ jobs:
           echo "skip=${skip}" >> "$GITHUB_OUTPUT"
           echo "doc_only=${doc_only}" >> "$GITHUB_OUTPUT"
 
-  # WARNING: make sure all of the build jobs are in sync
+  # NOTE: Build jobs are intentionally split by platform rather than using a single
+  # matrix. This allows each test job to depend only on its corresponding build,
+  # so faster platforms can proceed through build & test without waiting for slower
+  # ones. Keep these job definitions textually identical except for:
+  #   - host-platform value
+  #   - if: condition (build-linux-64 omits doc-only check since it's needed for docs)
   build-linux-64:
     needs:
       - ci-vars
@@ -81,7 +86,7 @@ jobs:
       cuda-version: ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
       prev-cuda-version: ${{ needs.ci-vars.outputs.CUDA_PREV_BUILD_VER }}
 
-  # WARNING: make sure all of the build jobs are in sync
+  # See build-linux-64 for why build jobs are split by platform.
   build-linux-aarch64:
     needs:
       - ci-vars
@@ -100,7 +105,7 @@ jobs:
       cuda-version: ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
       prev-cuda-version: ${{ needs.ci-vars.outputs.CUDA_PREV_BUILD_VER }}
 
-  # WARNING: make sure all of the build jobs are in sync
+  # See build-linux-64 for why build jobs are split by platform.
   build-windows:
     needs:
       - ci-vars
@@ -119,7 +124,11 @@ jobs:
       cuda-version: ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
       prev-cuda-version: ${{ needs.ci-vars.outputs.CUDA_PREV_BUILD_VER }}
 
-  # WARNING: make sure both Linux test jobs are in sync
+  # NOTE: Test jobs are split by platform for the same reason as build jobs (see
+  # build-linux-64). Keep these job definitions textually identical except for:
+  #   - host-platform value
+  #   - build job under needs:
+  #   - uses: (test-wheel-linux.yml vs test-wheel-windows.yml)
   test-linux-64:
     strategy:
       fail-fast: false
@@ -141,21 +150,20 @@ jobs:
       host-platform: ${{ matrix.host-platform }}
       build-ctk-ver: ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
 
-  # WARNING: make sure both Linux test jobs are in sync
+  # See test-linux-64 for why test jobs are split by platform.
   test-linux-aarch64:
     strategy:
       fail-fast: false
       matrix:
         host-platform:
           - linux-aarch64
     name: Test ${{ matrix.host-platform }}
-    # Note: No doc-only check needed here - if build-linux-aarch64 is skipped,
-    # this job is automatically skipped due to the dependency.
-    if: ${{ github.repository_owner == 'nvidia' }}
+    if: ${{ github.repository_owner == 'nvidia' && !fromJSON(needs.should-skip.outputs.doc-only) }}
     permissions:
       contents: read  # This is required for actions/checkout
     needs:
       - ci-vars
+      - should-skip
       - build-linux-aarch64
     secrets: inherit
     uses: ./.github/workflows/test-wheel-linux.yml
@@ -164,20 +172,20 @@ jobs:
       host-platform: ${{ matrix.host-platform }}
       build-ctk-ver: ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
 
+  # See test-linux-64 for why test jobs are split by platform.
   test-windows:
     strategy:
       fail-fast: false
       matrix:
         host-platform:
           - win-64
     name: Test ${{ matrix.host-platform }}
-    # Note: No doc-only check needed here - if build-windows is skipped,
-    # this job is automatically skipped due to the dependency.
-    if: ${{ github.repository_owner == 'nvidia' }}
+    if: ${{ github.repository_owner == 'nvidia' && !fromJSON(needs.should-skip.outputs.doc-only) }}
     permissions:
       contents: read  # This is required for actions/checkout
     needs:
       - ci-vars
+      - should-skip
       - build-windows
     secrets: inherit
     uses: ./.github/workflows/test-wheel-windows.yml

.github/workflows/release.yml

@@ -135,6 +135,7 @@ jobs:
     needs:
       - check-tag
       - determine-run-id
+      - doc
     secrets: inherit
     uses: ./.github/workflows/release-upload.yml
     with:
@@ -148,6 +149,7 @@ jobs:
     needs:
       - check-tag
       - determine-run-id
+      - doc
     environment:
       name: ${{ inputs.wheel-dst }}
       url: https://${{ (inputs.wheel-dst == 'testpypi' && 'test.') || '' }}pypi.org/p/${{ inputs.component }}/

cuda_bindings/cuda/bindings/_nvml.pyx

@@ -1198,16 +1198,16 @@ class PowerMizerMode(_IntEnum):
 
 
 class DeviceArch(_IntEnum):
-    DEVICE_ARCH_KEPLER = 2
-    DEVICE_ARCH_MAXWELL = 3
-    DEVICE_ARCH_PASCAL = 4
-    DEVICE_ARCH_VOLTA = 5
-    DEVICE_ARCH_TURING = 6
-    DEVICE_ARCH_AMPERE = 7
-    DEVICE_ARCH_ADA = 8
-    DEVICE_ARCH_HOPPER = 9
-    DEVICE_ARCH_BLACKWELL = 10
-    DEVICE_ARCH_UNKNOWN = 0xFFFFFFFF
+    KEPLER = 2
+    MAXWELL = 3
+    PASCAL = 4
+    VOLTA = 5
+    TURING = 6
+    AMPERE = 7
+    ADA = 8
+    HOPPER = 9
+    BLACKWELL = 10
+    UNKNOWN = 0xFFFFFFFF
 
 
 class BusType(_IntEnum):

cuda_core/cuda/core/_event.pyx

@@ -172,7 +172,7 @@ cdef class Event:
             raise RuntimeError(explanation)
 
     def __hash__(self) -> int:
-        return hash((type(self), as_intptr(self._h_context), as_intptr(self._h_event)))
+        return hash((type(self), as_intptr(self._h_event)))
 
     def __eq__(self, other) -> bool:
         # Note: using isinstance because `Event` can be subclassed.

cuda_core/cuda/core/_stream.pyx

@@ -217,22 +217,12 @@ cdef class Stream:
         return (0, as_intptr(self._h_stream))
 
     def __hash__(self) -> int:
-        # Ensure context is initialized for hash consistency
-        Stream_ensure_ctx(self)
-        return hash((as_intptr(self._h_context), as_intptr(self._h_stream)))
+        return hash(as_intptr(self._h_stream))
 
     def __eq__(self, other) -> bool:
         if not isinstance(other, Stream):
             return NotImplemented
-        cdef Stream _other = <Stream>other
-        # Fast path: compare handles first
-        if as_intptr(self._h_stream) != as_intptr(_other._h_stream):
-            return False
-        # Ensure contexts are initialized for both streams
-        Stream_ensure_ctx(self)
-        Stream_ensure_ctx(_other)
-        # Compare contexts as well
-        return as_intptr(self._h_context) == as_intptr(_other._h_context)
+        return as_intptr(self._h_stream) == as_intptr((<Stream>other)._h_stream)
 
     @property
     def handle(self) -> cuda.bindings.driver.CUstream:

cuda_core/cuda/core/system/__init__.py

@@ -25,9 +25,12 @@
 if CUDA_BINDINGS_NVML_IS_COMPATIBLE:
     from ._device import *
     from ._device import __all__ as _device_all
+    from ._system_events import *
+    from ._system_events import __all__ as _system_events_all
     from .exceptions import *
     from .exceptions import __all__ as _exceptions_all
 
     __all__.append("get_nvml_version")
     __all__.extend(_device_all)
+    __all__.extend(_system_events_all)
     __all__.extend(_exceptions_all)