Merge branch 'main' into ci-nightly-optdeps

Author: leofang

.gitattributes

@@ -5,6 +5,7 @@
 *.h binary
 *.hpp binary
 # Exception: headers we own
+benchmarks/cuda_bindings/benchmarks/cpp/*.hpp -binary text diff
 cuda_bindings/cuda/bindings/_bindings/*.h -binary text diff
 cuda_bindings/cuda/bindings/_lib/*.h -binary text diff
 cuda_core/cuda/core/_cpp/*.h -binary text diff

.github/workflows/ci.yml

@@ -343,7 +343,7 @@ jobs:
       build-type: pull-request
       host-platform: ${{ matrix.host-platform }}
       build-ctk-ver: ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
-      nruns: ${{ (github.event_name == 'schedule' && 100) || 1}}
+      nruns: ${{ (github.event_name == 'schedule' && 5) || 1}}
       skip-bindings-test: ${{ !fromJSON(needs.detect-changes.outputs.test_bindings) }}
 
   # See test-linux-64 for why test jobs are split by platform.
@@ -368,7 +368,7 @@ jobs:
       build-type: pull-request
       host-platform: ${{ matrix.host-platform }}
       build-ctk-ver: ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
-      nruns: ${{ (github.event_name == 'schedule' && 100) || 1}}
+      nruns: ${{ (github.event_name == 'schedule' && 5) || 1}}
       skip-bindings-test: ${{ !fromJSON(needs.detect-changes.outputs.test_bindings) }}
 
   # See test-linux-64 for why test jobs are split by platform.
@@ -393,7 +393,7 @@ jobs:
       build-type: pull-request
       host-platform: ${{ matrix.host-platform }}
       build-ctk-ver: ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
-      nruns: ${{ (github.event_name == 'schedule' && 100) || 1}}
+      nruns: ${{ (github.event_name == 'schedule' && 5) || 1}}
       skip-bindings-test: ${{ !fromJSON(needs.detect-changes.outputs.test_bindings) }}
 
   doc:

.github/workflows/release.yml

@@ -20,7 +20,6 @@ on:
           - cuda-bindings
           - cuda-pathfinder
           - cuda-python
-          - all
       git-tag:
         description: "The release git tag"
         required: true
@@ -89,6 +88,30 @@ jobs:
             gh release create "${{ inputs.git-tag }}" --draft --repo "${{ github.repository }}" --title "Release ${{ inputs.git-tag }}" --notes "Release ${{ inputs.git-tag }}"
           fi
 
+  check-release-notes:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout Source
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          ref: ${{ inputs.git-tag }}
+
+      - name: Set up Python
+        uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065  # v5.6.0
+        with:
+          python-version: "3.12"
+
+      - name: Self-test release-notes checker
+        run: |
+          pip install pytest
+          pytest ci/tools/tests
+
+      - name: Check versioned release notes exist
+        run: |
+          python ci/tools/check_release_notes.py \
+            --git-tag "${{ inputs.git-tag }}" \
+            --component "${{ inputs.component }}"
+
   doc:
     name: Build release docs
     if: ${{ github.repository_owner == 'nvidia' }}
@@ -99,6 +122,7 @@ jobs:
       pull-requests: write
     needs:
       - check-tag
+      - check-release-notes
       - determine-run-id
     secrets: inherit
     uses: ./.github/workflows/build-docs.yml
@@ -114,6 +138,7 @@ jobs:
       contents: write
     needs:
       - check-tag
+      - check-release-notes
       - determine-run-id
       - doc
     secrets: inherit
@@ -128,11 +153,12 @@ jobs:
     runs-on: ubuntu-latest
     needs:
       - check-tag
+      - check-release-notes
       - determine-run-id
       - doc
     environment:
       name: testpypi
-      url: https://test.pypi.org/${{ inputs.component != 'all' && format('p/{0}/', inputs.component) || '' }}
+      url: https://test.pypi.org/p/${{ inputs.component }}/
     permissions:
       id-token: write
     steps:
@@ -162,7 +188,7 @@ jobs:
       - publish-testpypi
     environment:
       name: pypi
-      url: https://pypi.org/${{ inputs.component != 'all' && format('p/{0}/', inputs.component) || '' }}
+      url: https://pypi.org/p/${{ inputs.component }}/
     permissions:
       id-token: write
     steps:

.spdx-ignore

@@ -10,5 +10,7 @@ cuda_bindings/examples/*
 
 # Vendored
 cuda_core/cuda/core/_include/dlpack.h
+cuda_core/cuda/core/_include/aoti_shim.h
+cuda_core/cuda/core/_include/aoti_shim.def
 
 qa/ctk-next.drawio.svg

benchmarks/cuda_bindings/README.md

@@ -47,12 +47,14 @@ To run the benchmarks combine the environment and task:
 ```bash
 # Run the Python benchmarks in the wheel environment
 pixi run -e wheel bench
+pixi run -e wheel bench --min-time 0.1
 
 # Run the Python benchmarks in the source environment
 pixi run -e source bench
 
 # Run the C++ benchmarks
 pixi run -e wheel bench-cpp
+pixi run -e wheel bench-cpp --min-time 0.1
 ```
 
 Both runners automatically save results to JSON files in the benchmarks

benchmarks/cuda_bindings/benchmarks/cpp/bench_event.cpp

@@ -45,6 +45,11 @@ int main(int argc, char** argv) {
     check_cu(cuStreamSynchronize(stream), "cuStreamSynchronize failed");
 
     bench::BenchmarkSuite suite(options);
+    // Drain the persistent stream after calibration so event_record (which
+    // enqueues onto the stream) and event_synchronize start from a known state.
+    suite.set_post_calibrate([&]() {
+        check_cu(cuStreamSynchronize(stream), "post-calibrate sync failed");
+    });
 
     // --- event_create_destroy ---
     {

benchmarks/cuda_bindings/benchmarks/cpp/bench_launch.cpp

@@ -238,6 +238,12 @@ int main(int argc, char** argv) {
     void* struct_params[] = {&struct_2048B};
 
     bench::BenchmarkSuite suite(options);
+    // After calibration, drain the persistent stream so the first measured
+    // sample does not start on a backlogged stream. Calibration for enqueue-
+    // style ops (kernel launches) may queue many thousands of operations.
+    suite.set_post_calibrate([&]() {
+        check_cu(cuStreamSynchronize(stream), "post-calibrate sync failed");
+    });
 
     suite.run("launch.launch_empty_kernel", [&]() {
         check_cu(cuLaunchKernel(empty_kernel, 1, 1, 1, 1, 1, 1, 0, stream, nullptr, nullptr),

benchmarks/cuda_bindings/benchmarks/cpp/bench_memory.cpp

@@ -52,6 +52,11 @@ int main(int argc, char** argv) {
     uint8_t host_dst[COPY_SIZE] = {};
 
     bench::BenchmarkSuite suite(options);
+    // Drain the persistent stream after calibration so async benchmarks
+    // (mem_alloc_async_free_async) don't start measurement on a backlogged stream.
+    suite.set_post_calibrate([&]() {
+        check_cu(cuStreamSynchronize(stream), "post-calibrate sync failed");
+    });
 
     // --- mem_alloc_free ---
     {

benchmarks/cuda_bindings/benchmarks/cpp/bench_stream.cpp

@@ -38,6 +38,12 @@ int main(int argc, char** argv) {
     check_cu(cuStreamCreate(&stream, CU_STREAM_NON_BLOCKING), "cuStreamCreate failed");
 
     bench::BenchmarkSuite suite(options);
+    // Drain the persistent stream after calibration for completeness.
+    // stream_create_destroy uses a local stream, but stream_query/synchronize
+    // observe the persistent one.
+    suite.set_post_calibrate([&]() {
+        check_cu(cuStreamSynchronize(stream), "post-calibrate sync failed");
+    });
 
     // --- stream_create_destroy ---
     {