Merge pull request #25 from NCIC-AlphaSparse/main

gpu ci runner

Author: NCIC-AlphaSparse

.github/workflows/gpu-benchmark.yml

@@ -35,41 +35,59 @@ on:
 jobs:
   benchmark:
     name: Run GPU benchmark
-    runs-on:
-      - self-hosted
-      - linux
-      - gpu
-    timeout-minutes: 90
+    runs-on: test-flagsparse
+    timeout-minutes: 360
 
     steps:
       - name: Check out repository
         uses: actions/checkout@v4
 
-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.12"
-          cache: pip
-          cache-dependency-path: tools/ci/requirements-triton-smoke.lock.txt
+      - name: Initialize artifact paths
+        run: mkdir -p benchmark_results
+
+      - name: Verify runner Python
+        run: python3.12 --version
+
+      - name: Verify benchmark dependencies
+        run: |
+          python3.12 -m pip show build pytest torch
+          python3.12 - <<'PY'
+          import build
+          import pytest
+          import torch
+
+          print("build", getattr(build, "__version__", "unknown"))
+          print("pytest", pytest.__version__)
+          print("torch", torch.__version__)
+          PY
+
+      - name: Install FlagTree Triton
+        shell: bash
+        run: |
+          python3.12 -m pip uninstall -y triton
+          python3.12 -m pip install --upgrade flagtree===0.6.0rc2 \
+            --index-url=https://resource.flagos.net/repository/flagos-pypi-hosted/simple
+          python3.12 - <<'PY'
+          import triton
 
-      - name: Install benchmark dependencies
-        run: python -m pip install --upgrade -r tools/ci/requirements-triton-smoke.lock.txt
+          print("triton", getattr(triton, "__version__", "unknown"))
+          PY
 
       - name: Validate CUDA environment
         run: |
           nvidia-smi
-          python tools/ci/check_gpu_environment.py \
+          python3.12 tools/ci/check_gpu_environment.py \
             --require-cuda \
             --metadata benchmark_results/gpu_metadata.json
 
       - name: Build source and wheel distributions
-        run: python -m build
+        run: python3.12 -m build
 
       - name: Install package
-        run: python -m pip install --force-reinstall --no-deps dist/*.whl
+        run: python3.12 -m pip install --force-reinstall --no-deps dist/*.whl
 
       - name: Validate installed wheel import
-        run: python tools/ci/check_installed_wheel.py --expected-version 1.0.0
+        run: python3.12 tools/ci/check_installed_wheel.py --expected-version 1.0.0
 
       - name: Run GPU performance benchmarks
         shell: bash
@@ -88,7 +106,7 @@ jobs:
           if [ -n "${{ inputs.benchmark_args }}" ]; then
             args+=(--benchmark-args "${{ inputs.benchmark_args }}")
           fi
-          python "${args[@]}"
+          python3.12 "${args[@]}"
 
       - name: Upload benchmark artifacts
         if: always()

.github/workflows/gpu-ci.yml

@@ -22,41 +22,61 @@ on:
 jobs:
   accuracy-smoke:
     name: GPU accuracy smoke
-    runs-on:
-      - self-hosted
-      - linux
-      - gpu
+    runs-on: test-flagsparse
     timeout-minutes: 90
 
     steps:
       - name: Check out repository
         uses: actions/checkout@v4
 
-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.12"
-          cache: pip
-          cache-dependency-path: tools/ci/requirements-triton-smoke.lock.txt
+      - name: Initialize artifact paths
+        run: |
+          mkdir -p pytest_results
+          printf '{}\n' > gpu_ci_metadata.json
+
+      - name: Verify runner Python
+        run: python3.12 --version
+
+      - name: Verify GPU test dependencies
+        run: |
+          python3.12 -m pip show build pytest torch
+          python3.12 - <<'PY'
+          import build
+          import pytest
+          import torch
+
+          print("build", getattr(build, "__version__", "unknown"))
+          print("pytest", pytest.__version__)
+          print("torch", torch.__version__)
+          PY
+
+      - name: Install FlagTree Triton
+        shell: bash
+        run: |
+          python3.12 -m pip uninstall -y triton
+          python3.12 -m pip install --upgrade flagtree===0.6.0rc2 \
+            --index-url=https://resource.flagos.net/repository/flagos-pypi-hosted/simple
+          python3.12 - <<'PY'
+          import triton
 
-      - name: Install GPU test dependencies
-        run: python -m pip install --upgrade -r tools/ci/requirements-triton-smoke.lock.txt
+          print("triton", getattr(triton, "__version__", "unknown"))
+          PY
 
       - name: Validate CUDA environment
         run: |
           nvidia-smi
-          python tools/ci/check_gpu_environment.py \
+          python3.12 tools/ci/check_gpu_environment.py \
             --require-cuda \
             --metadata gpu_ci_metadata.json
 
       - name: Build source and wheel distributions
-        run: python -m build
+        run: python3.12 -m build
 
       - name: Install package
-        run: python -m pip install --force-reinstall --no-deps dist/*.whl
+        run: python3.12 -m pip install --force-reinstall --no-deps dist/*.whl
 
       - name: Validate installed wheel import
-        run: python tools/ci/check_installed_wheel.py --expected-version 1.0.0
+        run: python3.12 tools/ci/check_installed_wheel.py --expected-version 1.0.0
 
       - name: Run GPU accuracy tests
         shell: bash
@@ -70,7 +90,7 @@ jobs:
           if [ -n "${{ inputs.ops }}" ]; then
             args+=(--ops "${{ inputs.ops }}")
           fi
-          python "${args[@]}"
+          python3.12 "${args[@]}"
 
       - name: Upload GPU CI artifacts
         if: always()

README.md

@@ -148,8 +148,8 @@ outputs compare against CPU int32 references.
 - `.github/workflows/nightly-cpu.yml` is a `main`-branch-only nightly CPU check that repeats the package, lint, and shared-runtime smoke tests.
 - `.github/workflows/release.yml` builds source and wheel artifacts, then attaches them to GitHub Releases on `v*` tags.
 - `.github/workflows/triton-smoke.yml` is a manual opt-in job for triton-dependent smoke checks.
-- `.github/workflows/gpu-ci.yml` is a manual GPU accuracy smoke workflow for a self-hosted runner labeled `self-hosted`, `linux`, and `gpu`.
-- `.github/workflows/gpu-benchmark.yml` adds an Actions button for synthetic GPU benchmark runs on a self-hosted runner labeled `self-hosted`, `linux`, and `gpu`.
+- `.github/workflows/gpu-ci.yml` is a manual GPU accuracy smoke workflow that runs on the `test-flagsparse` Actions Runner Controller scale set.
+- `.github/workflows/gpu-benchmark.yml` adds an Actions button for synthetic GPU benchmark runs on the `test-flagsparse` Actions Runner Controller scale set.
 - `.github/workflows/release-drafter.yml` keeps draft release notes current from merged PRs.
 - `make help` lists the local entry points.
 - `make ci` / `make check` run the same CPU-only pipeline used by CI.

tests/ci/test_gpu_ci_policy.py

@@ -18,32 +18,53 @@ def _read(path: Path) -> str:
 def test_gpu_accuracy_workflow_is_self_hosted_and_manual():
     text = _read(WORKFLOWS_DIR / "gpu-ci.yml")
     assert "workflow_dispatch:" in text
-    for label in ["self-hosted", "linux", "gpu"]:
-        assert f"- {label}" in text
-    assert 'python-version: "3.12"' in text
-    assert "tools/ci/requirements-triton-smoke.lock.txt" in text
+    # Runner scale sets (ARC) must be referenced by name alone in `runs-on`;
+    # combining it with labels like self-hosted/linux breaks job matching.
+    assert "runs-on: test-flagsparse" in text
+    assert "actions/setup-python" not in text
+    assert "python3.12 --version" in text
+    assert "$GITHUB_PATH" not in text
+    assert (
+        "pip install --upgrade -r tools/ci/requirements-triton-smoke.lock.txt"
+        not in text
+    )
+    assert "python3.12 -m pip show build pytest torch" in text
+    assert "python3.12 -m pip uninstall -y triton" in text
+    assert "flagtree===0.6.0rc2" in text
+    assert "mkdir -p pytest_results" in text
+    assert "gpu_ci_metadata.json" in text
 
 
 def test_gpu_accuracy_workflow_checks_cuda_before_tests():
     text = _read(WORKFLOWS_DIR / "gpu-ci.yml")
     assert "nvidia-smi" in text
     assert "tools/ci/check_gpu_environment.py" in text
     assert "--require-cuda" in text
-    assert "pytest" in text
-    assert "tests/pytest" in text
+    assert "run_flagsparse_accuracy.py" in text
 
 
 def test_gpu_benchmark_workflow_uploads_artifacts():
     text = _read(WORKFLOWS_DIR / "gpu-benchmark.yml")
-    assert "self-hosted" in text
-    assert "gpu" in text
-    assert "tools/ci/run_gpu_benchmark.py" in text
+    assert "runs-on: test-flagsparse" in text
+    assert (
+        "pip install --upgrade -r tools/ci/requirements-triton-smoke.lock.txt"
+        not in text
+    )
+    assert "python3.12 -m pip show build pytest torch" in text
+    assert "python3.12 -m pip uninstall -y triton" in text
+    assert "flagtree===0.6.0rc2" in text
+    assert "mkdir -p benchmark_results" in text
+    assert "run_flagsparse_performance.py" in text
     assert "matrix_dir:" in text
-    for suite_name in ["alpha-spmm-alg1", "spmm-opt-alg2", "spgemm", "sddmm"]:
-        assert suite_name in text
     assert "actions/upload-artifact@v4" in text
 
 
+def test_gpu_dependency_bundle_leaves_triton_to_flagtree():
+    text = _read(TOOLS_DIR / "requirements-triton-smoke.lock.txt")
+    assert "torch==2.9.0" in text
+    assert "triton==" not in text
+
+
 def test_gpu_environment_check_can_run_without_gpu():
     text = _read(TOOLS_DIR / "check_gpu_environment.py")
     assert "--require-cuda" in text

tools/ci/requirements-triton-smoke.lock.txt

@@ -1,3 +1,2 @@
 -r requirements-ci.lock.txt
 torch==2.9.0
-triton==3.6.0

tools/ci/requirements-triton-smoke.txt

@@ -1,3 +1,2 @@
 -r requirements-ci.txt
 torch==2.9.0
-triton==3.6.0