[CI] Isolate standalone CUTLASS workflow changes

.github/actions/workflow-build/build-workflow.py

@@ -62,6 +62,7 @@
 import re
 import struct
 import sys
+from pathlib import Path
 
 import yaml
 
@@ -1060,21 +1061,22 @@ def natural_sort_key(key):
     return workflow_dispatch_groups
 
 
-def find_workflow_line_number(workflow_name):
+def find_workflow_line_number(workflow_name, filename=None):
+    filename = filename or matrix_yaml["filename"]
     regex = re.compile(f"^( )*{workflow_name}:", re.IGNORECASE)
     line_number = 0
-    with open(matrix_yaml["filename"], "r") as f:
+    with open(filename, "r") as f:
         for line in f:
             line_number += 1
             if regex.match(line):
                 return line_number
     raise Exception(
-        f"Workflow '{workflow_name}' not found in {matrix_yaml['filename]']} (could not match regex: {regex})"
+        f"Workflow '{workflow_name}' not found in {filename} (could not match regex: {regex})"
     )
 
 
-def get_matrix_job_origin(matrix_job, workflow_name, workflow_location):
-    filename = matrix_yaml["filename"]
+def get_matrix_job_origin(matrix_job, workflow_name, workflow_location, filename=None):
+    filename = filename or matrix_yaml["filename"]
     original_matrix_job = json.dumps(matrix_job, indent=None, separators=(", ", ": "))
     original_matrix_job = original_matrix_job.replace('"', "")
     return {
@@ -1350,9 +1352,9 @@ def parse_workflow_matrix_jobs(workflow_name, filter_projects=None):
             workflow_location = (
                 f"{matrix_yaml['filename']}:{workflow_line_number} (job {idx + 1})"
             )
-            matrix_job["origin"] = get_matrix_job_origin(
-                matrix_job, workflow_name, workflow_location
-            )
+            matrix_job["origin"] = matrix_yaml.get("project_workflow_origins", {}).get(
+                (workflow_name, idx)
+            ) or get_matrix_job_origin(matrix_job, workflow_name, workflow_location)
 
     # Fill in default values, explode lists.
     matrix_jobs = preprocess_matrix_jobs(matrix_jobs, is_exclusion_matrix)
@@ -1548,6 +1550,53 @@ def print_devcontainer_info(args):
     print(json.dumps(devcontainer_json, indent=2))
 
 
+def load_project_workflows(matrix, matrix_file):
+    """Append optional project-owned workflow rows and retain their source locations."""
+    allowed_workflows = {
+        "pull_request",
+        "pull_request_lite",
+        "pull_request_fork",
+        "nightly",
+        "weekly",
+    }
+    origins = {}
+    for project, config in matrix["projects"].items():
+        workflows_file = config.get("workflows_file")
+        if workflows_file is None:
+            continue
+        path = Path(matrix_file).parent / workflows_file
+        if not path.exists():
+            continue
+        workflows = yaml.safe_load(path.read_text(encoding="utf-8"))
+        if not isinstance(workflows, dict):
+            raise ValueError(f"{path}: expected a mapping of workflows to job lists")
+        for workflow_name, jobs in workflows.items():
+            if (
+                workflow_name not in allowed_workflows
+                or workflow_name not in matrix["workflows"]
+            ):
+                raise ValueError(
+                    f"{path}: unsupported project workflow '{workflow_name}'"
+                )
+            if not isinstance(jobs, list):
+                raise ValueError(
+                    f"{path}: workflow '{workflow_name}' must contain a job list"
+                )
+            existing_jobs = matrix["workflows"][workflow_name] or []
+            line_number = find_workflow_line_number(workflow_name, str(path))
+            for idx, job in enumerate(jobs):
+                if not isinstance(job, dict) or job.get("project") != project:
+                    raise ValueError(
+                        f"{path}: every job must select only project '{project}'"
+                    )
+                location = f"{path}:{line_number} (job {idx + 1})"
+                origins[(workflow_name, len(existing_jobs) + idx)] = (
+                    get_matrix_job_origin(job, workflow_name, location, str(path))
+                )
+            matrix["workflows"][workflow_name] = existing_jobs + jobs
+    return origins
+
+
 def preprocess_matrix_yaml(matrix):
     # Numeric YAML keys lose the spelling of dotted versions.
     if not all(isinstance(version, str) for version in matrix["ctk_versions"]):
@@ -1616,6 +1665,9 @@ def main():
         matrix_yaml = yaml.safe_load(f)
         matrix_yaml = preprocess_matrix_yaml(matrix_yaml)
         matrix_yaml["filename"] = args.matrix_file
+        matrix_yaml["project_workflow_origins"] = load_project_workflows(
+            matrix_yaml, args.matrix_file
+        )
 
     if args.workflows:
         print_gha_workflow(args)

.github/workflows/ci-workflow-pull-request.yml

@@ -77,6 +77,11 @@ jobs:
           echo "pr_number=${{ fromJSON(steps.get-pr-info.outputs.pr-info).number }}" | tee -a "${GITHUB_OUTPUT}"
           git fetch origin "${BASE_SHA}" -q
           echo "merge_base_sha=$(git merge-base "${BASE_SHA}" "${GITHUB_SHA}")" | tee -a "${GITHUB_OUTPUT}"
+      - name: Inspect compile-time benchmark inputs
+        id: inspect-compile-time-changes
+        env:
+          BASE_SHA: ${{ steps.export-pr-info.outputs.base_sha }}
+        run: python3 ci/inspect_compile_time_changes.py --refs "${BASE_SHA}" "${GITHUB_SHA}"
       - name: Build benchmark dispatch matrix
         id: build-bench-matrix
         run: |
@@ -109,6 +114,15 @@ jobs:
             exit 0
           fi
 
+          if [[ "${{ steps.inspect-compile-time-changes.outputs.compile_time_changed }}" != "true" ]]; then
+            echo "Public headers and compile-time benchmark inputs are unchanged; skipping dispatch."
+            {
+              echo "compile_time_enabled=false"
+              echo 'compile_time_matrix={"include":[]}'
+            } >> "${GITHUB_OUTPUT}"
+            exit 0
+          fi
+
           compile_time_matrix_json="$(./ci/compile_time/parse_matrix.py "ci/matrix.yaml" --workflow pull_request)"
           compile_time_enabled="$(
             jq -r 'if (.include // []) == [] then "false" else "true" end' \

ci/inspect_compile_time_changes.py

@@ -0,0 +1,67 @@
+#!/usr/bin/env python3
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""Select PR compile-time benchmarks from their inputs before allocating runners."""
+
+from __future__ import annotations
+
+from typing import Sequence
+
+import inspect_changes
+
+# The benchmark builds public headers, not Python packages or C++ library tests.
+# Reuse their ownership from the matrix classifier.
+PUBLIC_PROJECTS = ("libcudacxx_public", "cub_public", "thrust_public")
+BENCHMARK_INPUTS = (
+    r"ci/compile_time/",
+    r"ci/build_compile_time_bench\.sh$",
+    r"\.github/workflows/compile-time-bench\.yml$",
+)
+
+
+def compile_time_changed(
+    config: inspect_changes.Config,
+    dirty_files: Sequence[str],
+    ignored_files: Sequence[str],
+) -> bool:
+    # Include benchmark inputs ignored by the main matrix, while continuing to
+    # ignore documentation-only edits.
+    all_files = [
+        path for path in (*dirty_files, *ignored_files) if not path.endswith(".md")
+    ]
+    inputs = inspect_changes.compile_patterns(BENCHMARK_INPUTS)
+    benchmark_changed = any(
+        inspect_changes.matches_any(inputs, path) for path in all_files
+    )
+    project_files = inspect_changes.build_project_dirty_map(config, dirty_files)
+
+    # Preserve the classifier's conservative fallback for unknown infrastructure.
+    infrastructure_changed = bool(project_files[inspect_changes.CORE_PROJECT_KEY])
+    public_changed = any(project_files[key] for key in PUBLIC_PROJECTS)
+    # Package config, install rules, and header-test generation live beside the
+    # public headers. Their changes also affect compile-time data.
+    cmake_inputs = inspect_changes.compile_patterns(
+        (
+            r"(libcudacxx|cub|thrust|cudax)/(cmake/|CMakeLists\.txt$)",
+            r"cccl-version\.json$",
+        )
+    )
+    cmake_changed = any(
+        inspect_changes.matches_any(cmake_inputs, path) for path in all_files
+    )
+    return (
+        infrastructure_changed or public_changed or cmake_changed or benchmark_changed
+    )
+
+
+def main(argv: Sequence[str] | None = None) -> int:
+    args = inspect_changes.parse_args(argv)
+    config = inspect_changes.load_config(inspect_changes.CONFIG_PATH)
+    dirty_files, ignored_files = inspect_changes.collect_dirty_files(args, config)
+    enabled = compile_time_changed(config, dirty_files, ignored_files)
+    inspect_changes.write_output("compile_time_changed", str(enabled).lower())
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

ci/matrix.yaml

@@ -864,6 +864,11 @@ jobs:
   test_py_coop_numba_mlir_compile: { name: "Compile cuda.coop Numba-CUDA-MLIR (GPU hidden)", gpu: false, needs: 'build', force_producer_ctk: "pybuild", invoke: { prefix: 'test', args: '-stage numba-mlir-compile'} }
   test_py_coop_numba_mlir_runtime: { name: "Test cuda.coop Numba-CUDA-MLIR (L4)", gpu: true, needs: 'build', force_producer_ctk: "pybuild", invoke: { prefix: 'test', args: '-stage numba-mlir-runtime'} }
 
+  # Standalone CUTLASS compiler tests build and qualify the runtime wheel.
+  test_py_coop_cutlass_host: { name: "Test cuda.coop.cutlass (host)", gpu: false, invoke: { prefix: 'test', args: '-stage host'} }
+  test_py_coop_cutlass_compile: { name: "Compile cuda.coop.cutlass (GPU hidden)", gpu: false, invoke: { prefix: 'test', args: '-stage compile'} }
+  test_py_coop_cutlass_runtime: { name: "Test cuda.coop.cutlass", gpu: true, needs: 'test_py_coop_cutlass_compile', invoke: { prefix: 'test', args: '-stage runtime'} }
+
   # Run jobs for 'target' project (ci/util/build_and_test_targets.sh):
   run_cpu: { gpu: false }
   run_gpu: { gpu: true  }
@@ -948,6 +953,9 @@ projects:
       test:  ['test_py_compute_minimal']
   cuda_coop_python:
     name: "Python (cuda.coop)"
+  cuda_coop_cutlass:
+    name: "Python (cuda.coop.cutlass)"
+    workflows_file: ../python/cuda_coop_cutlass/ci/matrix.yaml
   cccl_c_parallel:
     name: 'CCCL C Parallel'
     stds: [20]

ci/project_files_and_dependencies.yaml

@@ -199,6 +199,23 @@ projects:
     - "^pyproject\\.toml$"
     exclude_regexes: []
 
+  cuda_coop_cutlass:
+    name: "cuda.coop.cutlass"
+    matrix_project: "cuda_coop_cutlass"
+    lite_dependencies: [libcudacxx_public, cub_public, thrust_public]
+    full_dependencies: []
+    include_regexes:
+      - 'python/cuda_coop_cutlass/'
+      - 'ci/(build|test)_cuda_coop_cutlass\.sh$'
+      - 'ci/windows/build_cuda_coop_cutlass\.ps1$'
+      - 'ci/cuda_coop_cutlass_requirements\.txt$'
+      - 'ci/validate_cuda_coop_cutlass_artifacts\.py$'
+      - 'ci/test/test_cuda_coop_cutlass_ci\.py$'
+      - '^LICENSE$'
+      - '^cub/LICENSE\.TXT$'
+      - '^libcudacxx/LICENSE\.TXT$'
+      - '^thrust/LICENSE$'
+
   packaging:
     name: "CCCL Packaging"
     matrix_project: "packaging"
@@ -213,6 +230,7 @@ projects:
       - cudax_internal
       - cudax_public
     full_dependencies: []
+    exclude_project_files: [cuda_coop_cutlass]
     include_regexes:
       - "examples/"
       - "test/cmake/"

ci/test/inspect_changes/core_dirty.output

@@ -1,2 +1,2 @@
-FULL_BUILD=libcudacxx cub thrust cudax cccl_c_parallel cccl_c_parallel_v2 python_v2 python_tsan cccl_c_stf python cuda_coop_python packaging stdpar nvbench_helper nvrtcc tidy
+FULL_BUILD=libcudacxx cub thrust cudax cccl_c_parallel cccl_c_parallel_v2 python_v2 python_tsan cccl_c_stf python cuda_coop_python cuda_coop_cutlass packaging stdpar nvbench_helper nvrtcc tidy
 LITE_BUILD=

ci/test/inspect_changes/cuda_coop_cutlass.dirty_files

@@ -0,0 +1 @@
+python/cuda_coop_cutlass/cuda/coop/cutlass/_compiler/_runtime.py

ci/test/inspect_changes/cuda_coop_cutlass.output

@@ -0,0 +1,2 @@
+FULL_BUILD=cuda_coop_cutlass
+LITE_BUILD=

ci/test/inspect_changes/cuda_coop_python_license.output

@@ -1,2 +1,2 @@
-FULL_BUILD=cuda_coop_python
+FULL_BUILD=cuda_coop_python cuda_coop_cutlass
 LITE_BUILD=

ci/test/inspect_changes/libcudacxx_both.output

@@ -1,2 +1,2 @@
 FULL_BUILD=libcudacxx tidy
-LITE_BUILD=cub thrust cudax cccl_c_parallel cccl_c_parallel_v2 python_v2 python_tsan cccl_c_stf python cuda_coop_python packaging stdpar nvbench_helper
+LITE_BUILD=cub thrust cudax cccl_c_parallel cccl_c_parallel_v2 python_v2 python_tsan cccl_c_stf python cuda_coop_python cuda_coop_cutlass packaging stdpar nvbench_helper

ci/test/inspect_changes/libcudacxx_public_only.output

@@ -1,2 +1,2 @@
 FULL_BUILD=libcudacxx tidy
-LITE_BUILD=cub thrust cudax cccl_c_parallel cccl_c_parallel_v2 python_v2 python_tsan cccl_c_stf python cuda_coop_python packaging stdpar nvbench_helper
+LITE_BUILD=cub thrust cudax cccl_c_parallel cccl_c_parallel_v2 python_v2 python_tsan cccl_c_stf python cuda_coop_python cuda_coop_cutlass packaging stdpar nvbench_helper

ci/test/inspect_changes/libcudacxx_thrust.output

@@ -1,2 +1,2 @@
 FULL_BUILD=libcudacxx thrust tidy
-LITE_BUILD=cub cudax cccl_c_parallel cccl_c_parallel_v2 python_v2 python_tsan cccl_c_stf python cuda_coop_python packaging stdpar nvbench_helper
+LITE_BUILD=cub cudax cccl_c_parallel cccl_c_parallel_v2 python_v2 python_tsan cccl_c_stf python cuda_coop_python cuda_coop_cutlass packaging stdpar nvbench_helper