Fix test_subprocess.py/test_print by installing and using dev PTI

.github/workflows/build-test-reusable.yml

@@ -267,7 +267,7 @@ jobs:
           pip install intel_pti-*.whl
           PTI_LIBS_DIR=$(python ./scripts/pti_lib.py)
           ls $PTI_LIBS_DIR
-          echo "PTI_LIBS_DIR=$PTI_LIBS_DIR" | tee -a $GITHUB_ENV
+          echo "LD_LIBRARY_PATH=$PTI_LIBS_DIR:$LD_LIBRARY_PATH" | tee -a $GITHUB_ENV
 
       - name: Report environment details
         run: |
@@ -305,7 +305,6 @@ jobs:
       - name: Run Proton tests
         if: matrix.suite == 'rest'
         run: |
-          export LD_LIBRARY_PATH=${{ env.PTI_LIBS_DIR }}:$LD_LIBRARY_PATH
           cd third_party/proton/test
           pytest test_api.py test_cmd.py test_lib.py test_profile.py test_viewer.py --device xpu -s -v
           cd ..

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

@@ -122,6 +122,15 @@ jobs:
           pip install -U pybind11 cmake
           pip install -v '.[build,tests,tutorials]'
 
+      - name: Build && Install PTI
+        shell: bash
+        run: |
+          cd ${{ env.NEW_WORKSPACE }}
+          ./scripts/install-pti.sh --build-level-zero
+          PTI_LIBS_DIR=$(python ./scripts/pti_lib.py)
+          ls $PTI_LIBS_DIR
+          echo "PATH=$PTI_LIBS_DIR:$PATH" | tee -a $GITHUB_ENV
+
       - name: Triton version
         run: |
           .venv\Scripts\activate.ps1

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

@@ -101,6 +101,15 @@ jobs:
           .venv\Scripts\activate.ps1
           python -c 'import triton; print(triton.__version__)'
 
+      - name: Build && Install PTI
+        shell: bash
+        run: |
+          cd ${{ env.NEW_WORKSPACE }}
+          ./scripts/install-pti.sh --build-level-zero
+          PTI_LIBS_DIR=$(python ./scripts/pti_lib.py)
+          ls $PTI_LIBS_DIR
+          echo "PATH=$PTI_LIBS_DIR:$PATH" | tee -a $GITHUB_ENV
+
       - name: Install test dependencies
         run: |
           .venv\Scripts\activate.ps1

.github/workflows/pip-test.yml

@@ -55,6 +55,13 @@ jobs:
       - name: Install Triton
         uses: ./.github/actions/setup-triton
 
+      - name: Build && Install PTI
+        run: |
+          ./scripts/install-pti.sh --build-level-zero
+          PTI_LIBS_DIR=$(python ./scripts/pti_lib.py)
+          ls $PTI_LIBS_DIR
+          echo "LD_LIBRARY_PATH=$PTI_LIBS_DIR:$LD_LIBRARY_PATH" | tee -a $GITHUB_ENV
+
       - name: Install runtime dependencies
         run: |
           curl -sSLO --retry 10 https://raw.githubusercontent.com/pytorch/pytorch/$(<.github/pins/pytorch.txt)/.github/scripts/generate_binary_build_matrix.py

python/test/unit/language/test_subprocess.py

@@ -53,9 +53,6 @@ def test_print(func_type: str, data_type: str, device: str):
         return
 
     outs = [line for line in proc.stdout.decode("UTF-8").splitlines() if line]
-    # FIXME: workaround for DLE 2025.2, namely PTI 0.13.1
-    # https://github.com/intel/intel-xpu-backend-for-triton/issues/4998
-    outs = filter(lambda elem: "Another subscriber already subscribed to Sycl runtime events" not in elem, outs)
     # The total number of elements in the 1-D tensor to print.
     N = 128