Merge branch 'main' into tkuczynski/enable_test_device_assert

Author: dev-tomek

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

@@ -42,7 +42,7 @@ on:
       use_pyenv_python:
         description: Use Python built with pyenv
         type: boolean
-        default: false
+        default: true
 
 permissions: read-all
 
@@ -64,4 +64,4 @@ jobs:
       skip_list: ${{ inputs.skip_list }}
       run_name: ${{ inputs.run_name || format('Build and test {0}', inputs.runner_label) }}
       enable_unskip: ${{ inputs.enable_unskip }}
-      use_pyenv_python: ${{ inputs.use_pyenv_python || false }}
+      use_pyenv_python: ${{ inputs.use_pyenv_python || true }}

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

@@ -118,10 +118,6 @@ jobs:
           ref: ${{ inputs.pytorch_ref }}
           mode: ${{ inputs.pytorch_mode }}
 
-      - name: Build Proton with XPU support
-        run: |
-          echo TRITON_BUILD_PROTON_XPU=1 | tee -a $GITHUB_ENV
-
       - name: Build Triton
         uses: ./.github/actions/setup-triton
         with:
@@ -287,6 +283,8 @@ jobs:
       - name: Build PTI && Run Proton tests
         if: matrix.suite == 'rest'
         run: |
+          # `intel-pti` can be installed in "Setup PyTorch" step with `pytorch_mode==wheels`
+          pip uninstall intel-pti -y
           PTI_COMMIT_ID="$(<.github/pins/pti.txt)"
           git clone https://github.com/intel/pti-gpu.git
           cd pti-gpu

.github/workflows/sglang-tests.yml

@@ -104,21 +104,18 @@ jobs:
           python transform_results.py $REPORTS/moe-gemm-performance.csv $REPORTS/moe-gemm-fp8-report.csv --tag $TAG --benchmark moe-fp8-benchmark
 
       - name: Run Liger-Kernel benchmarks
-        if: ${{ steps.install.outcome == 'success' && !cancelled() && (inputs.benchmarks == '' || contains(fromJson(inputs.benchmarks || '[]'), 'liger-kernel')) }}
+        if: ${{ steps.install.outcome == 'success' && !cancelled() && (inputs.benchmarks == '' || contains(fromJson(inputs.benchmarks || '[]'), 'liger')) }}
         run: |
           source ./scripts/capture-hw-details.sh
 
-          cd benchmarks/third_party/liger_kernels
-
-          git clone https://github.com/linkedin/Liger-Kernel
-          pip install -e Liger-Kernel
+          ./scripts/test-triton.sh --install-liger --skip-pip-install --skip-pytorch-install
 
           # To remember return code, but still copy results
           RET_CODE=0
-          bash ./run_benchmarks.sh || RET_CODE=$?
+          bash benchmarks/third_party/liger/run_benchmarks.sh || RET_CODE=$?
 
           cp Liger-Kernel/benchmark/data/all_benchmark_data.csv $REPORTS/liger-raw.csv
-          python transform.py $REPORTS/liger-raw.csv $REPORTS/liger-report.csv --tag $TAG
+          python benchmarks/third_party/liger/transform.py $REPORTS/liger-raw.csv $REPORTS/liger-report.csv --tag $TAG
 
           # Return the captured return code at the end
           exit "$RET_CODE"

.github/workflows/third-party-benchmarks.yml

@@ -1,4 +1,4 @@
-name: Third party tests [liger-kernels, vllm]
+name: Third party tests [liger-kernels, vllm, sglang]
 
 on:
   workflow_dispatch:
@@ -28,12 +28,12 @@ env:
   TAG: ${{ inputs.tag || (github.event_name == 'pull_request' && format('pr-{0}', github.event.number)) || (github.event_name == 'schedule' && 'ci') || 'test' }}
 
 jobs:
-  build:
-    name: Third party tests [liger-kernels, vllm]
+  small-tests:
+    name: Third party tests [vllm, sglang]
     runs-on:
       - linux
       - ${{ inputs.runner_label || 'max1550' }}
-    timeout-minutes: 720
+    timeout-minutes: 120
     defaults:
       run:
         shell: bash -noprofile --norc -eo pipefail -c "source /opt/intel/oneapi/setvars.sh > /dev/null; source {0}"
@@ -47,14 +47,7 @@ jobs:
       - name: Checkout repository
         uses: actions/checkout@v5
 
-      - name: Install Python
-        if: ${{ !(inputs.use_pyenv_python || false) }}
-        uses: actions/setup-python@v6
-        with:
-          python-version: ${{ env.PYTHON_VERSION }}
-
       - name: Install Python (from pyenv) ${{ inputs.python_version }}
-        if: ${{ inputs.use_pyenv_python }}
         uses: ./.github/actions/setup-pyenv-python
         with:
           python-version: ${{ env.PYTHON_VERSION }}
@@ -71,34 +64,84 @@ jobs:
       - name: Setup PyTorch
         uses: ./.github/actions/setup-pytorch
 
-      - name: Build Triton wheels
-        uses: ./.github/actions/setup-triton
-        with:
-          command: DEBUG=1 python -m build --wheel --no-isolation
-
-      - name: Install Triton
+      - name: Setup Triton
         id: install
-        run: |
-          pip install dist/*.whl
+        uses: ./.github/actions/setup-triton
 
       - name: Create reports dir
         run: |
           mkdir reports
           echo "REPORTS=$PWD/reports" >> $GITHUB_ENV
 
+      - name: Run SGLANG tests
+        if: ${{ steps.install.outcome == 'success' && !cancelled() }}
+        run: |
+          ./scripts/test-triton.sh --sglang --skip-pip-install --skip-pytorch-install
+
       - name: Run VLLM tests
         if: ${{ steps.install.outcome == 'success' && !cancelled() }}
         run: |
           ./scripts/test-triton.sh --vllm --skip-pip-install --skip-pytorch-install
 
-      - name: Run Liger-Kernel tests
+      - name: Upload test report
         if: ${{ steps.install.outcome == 'success' && !cancelled() }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: test-main-reports
+          path: reports
+  # We run all tests for Liger, so it's slow and we test it separately
+  liger:
+    name: Liger testing
+    runs-on:
+      - linux
+      - ${{ inputs.runner_label || 'max1550' }}
+    timeout-minutes: 120
+    defaults:
+      run:
+        shell: bash -noprofile --norc -eo pipefail -c "source /opt/intel/oneapi/setvars.sh > /dev/null; source {0}"
+    steps:
+      - name: Print inputs
+        run: |
+          cat <<EOF
+          ${{ toJSON(inputs) }}
+          EOF
+
+      - name: Checkout repository
+        uses: actions/checkout@v5
+
+      - name: Install Python (from pyenv) ${{ inputs.python_version }}
+        uses: ./.github/actions/setup-pyenv-python
+        with:
+          python-version: ${{ env.PYTHON_VERSION }}
+
+      - name: Identify Python version
+        run: |
+          PYTHON_VERSION="$(python -c 'import sys; print(f"{sys.version_info[0]}.{ sys.version_info[1]}")')"
+          echo "PYTHON_VERSION=$PYTHON_VERSION" | tee -a $GITHUB_ENV
+
+      - name: Install Python build dependencies
+        run: |
+          pip install cmake
+
+      - name: Setup PyTorch
+        uses: ./.github/actions/setup-pytorch
+
+      - name: Setup Triton
+        id: install
+        uses: ./.github/actions/setup-triton
+
+      - name: Create reports dir
+        run: |
+          mkdir reports
+          echo "REPORTS=$PWD/reports" >> $GITHUB_ENV
+
+      - name: Run Liger-Kernel tests
         run: |
           ./scripts/test-triton.sh --liger --skip-pip-install --skip-pytorch-install
 
       - name: Upload test report
         if: ${{ steps.install.outcome == 'success' && !cancelled() }}
         uses: actions/upload-artifact@v4
         with:
-          name: test-reports
+          name: test-liger-reports
           path: reports

.github/workflows/third-party-tests.yml

@@ -18,6 +18,7 @@ on:
         options:
           - ELAPSED_TIME
           - UPSTREAM_PYTORCH_PROFILER
+          - PROTON_PROFILER
         default: UPSTREAM_PYTORCH_PROFILER
       verify:
         description: Verify the benchmark results
@@ -141,6 +142,14 @@ jobs:
           python build_report.py $REPORTS/softmax-performance.csv $REPORTS/softmax-xetla-report.csv --benchmark softmax --compiler xetla --param_cols "N" --tflops_col XeTLA-TFlops --hbm_col "XeTLA-GB/s" --tag $TAG
           python build_report.py $REPORTS/softmax-performance.csv $REPORTS/softmax-onednn-report.csv --benchmark softmax --compiler onednn --param_cols "N" --tflops_col oneDNN-TFlops --hbm_col "oneDNN-GB/s" --tag $TAG
 
+      - name: Run Triton Softmax kernel benchmark with Proton
+        if: ${{ steps.install.outcome == 'success' && !cancelled() && (inputs.benchmarks == '' || contains(fromJson(inputs.benchmarks || '[]'), 'fused_softmax.py')) && !contains(fromJson(inputs.skip_benchmarks || '[]'), 'fused_softmax.py') }}
+        run: |
+          export LD_LIBRARY_PATH=$PTI_LIBS_DIR:$LD_LIBRARY_PATH
+          cd benchmarks/triton_kernels_benchmark
+          BENCHMARKING_METHOD=PROTON_PROFILER python fused_softmax.py
+          source ../../scripts/capture-hw-details.sh
+
       - name: Run Triton GEMM kernel benchmark
         if: ${{ steps.install.outcome == 'success' && !cancelled() && (inputs.benchmarks == '' || contains(fromJson(inputs.benchmarks || '[]'), 'gemm_benchmark.py')) && !contains(fromJson(inputs.skip_benchmarks || '[]'), 'gemm_benchmark.py') }}
         run: |