Merge branch 'main' into ncu-attach-mode

Signed-off-by: Alok Joshi <aljoshi@nvidia.com>

Author: Alok-Joshi

.github/workflows/ci.yaml

@@ -35,15 +35,27 @@ jobs:
           hatch run docs:build
 
   test:
-    runs-on: ${{ matrix.runner }}
+    name: test (${{ matrix.platform.runner }})
+    runs-on: ${{ matrix.platform.runner }}
     strategy:
       matrix:
-        runner: [linux-amd64-gpu-h100-latest-1, linux-arm64-gpu-l4-latest-1]
+        platform:
+          - runner: linux-amd64-gpu-h100-latest-1
+            ncu_installer_url: "https://developer.nvidia.com/downloads/assets/tools/secure/nsight-compute/2026_2_1/nsight_compute-linux-x86_64-2026.2.1.6.run"
+          - runner: linux-arm64-gpu-l4-latest-1
+            ncu_installer_url: "https://developer.nvidia.com/downloads/assets/tools/secure/nsight-compute/2026_2_1/nsight_compute-linux-sbsa-2026.2.1.6.run"
     container:
-      image: nvcr.io/nvidia/pytorch:26.05-py3
+      image: nvcr.io/nvidia/pytorch:26.06-py3
       options: "--cap-add=SYS_ADMIN --gpus all"
     steps:
       - uses: actions/checkout@v5
+      - name: "Install NCU 2026.2.1.0 (CUDA 13.3 Update 1)"
+        run: |
+          curl -L -o ncu_installer.run "${{ matrix.platform.ncu_installer_url }}"
+          chmod +x ncu_installer.run
+          ./ncu_installer.run -- -noprompt
+          rm ncu_installer.run
+          echo "/usr/local/NVIDIA-Nsight-Compute-2026.2" >> $GITHUB_PATH
       - name: "Running Tests"
         run: |
           pip install hatch

examples/test_examples.py

@@ -2,28 +2,10 @@
 # SPDX-License-Identifier: Apache-2.0
 
 import importlib
-import subprocess
-import sys
 
 import pytest
 
 
-def get_cuda_dev_cc_major(device_id: int) -> int:
-    # Run in a subprocess to avoid initializing CUDA in the pytest process during collection,
-    # which would prevent nsight's injection library from loading before CUDA init.
-    result = subprocess.run(
-        [
-            sys.executable,
-            "-c",
-            f"import torch; print(torch.cuda.get_device_properties({device_id}).major)",
-        ],
-        capture_output=True,
-        text=True,
-        check=True,
-    )
-    return int(result.stdout.strip())
-
-
 def test_00_minimal() -> None:
     minimal = importlib.import_module("examples.00_minimal")
     minimal.main()
@@ -85,11 +67,6 @@ def test_11_output_csv() -> None:
     output_csv.main()
 
 
-# skip cuTile test on CC 9.x as Cuda Toolkit 13.2 used for testing does not support cuda-tile on CC 9.x (Hopper)
-@pytest.mark.skipif(
-    get_cuda_dev_cc_major(0) == 9,
-    reason="cuda-tile not supported on CC 9.x in CUDA Toolkit 13.2",
-)  # type: ignore[untyped-decorator]
 def test_12_cutile() -> None:
     cutile = importlib.import_module("examples.12_cutile")
     cutile.main()