CI: reduce packaging, profiling, and unit-test overhead

.github/workflows/_packaging.yml

@@ -41,11 +41,22 @@ jobs:
         if: inputs.setup_variant == 'github'
         run: |
           if [[ "${{ runner.os }}" == "Linux" ]]; then
-            sudo apt-get update
-            sudo apt-get install -y cmake ninja-build g++
+            missing_packages=()
+            command -v ninja >/dev/null 2>&1 || missing_packages+=(ninja-build)
+            command -v ccache >/dev/null 2>&1 || missing_packages+=(ccache)
+            if ! command -v gcc >/dev/null 2>&1 || ! command -v g++ >/dev/null 2>&1; then
+              missing_packages+=(g++)
+            fi
+            if (( ${#missing_packages[@]} )); then
+              sudo apt-get update
+              sudo apt-get install -y "${missing_packages[@]}"
+            fi
           else
-            brew install ninja cmake
-            brew install gcc@15 || brew install gcc
+            command -v ninja >/dev/null 2>&1 || brew install ninja
+            command -v ccache >/dev/null 2>&1 || brew install ccache
+            if ! command -v gcc-15 >/dev/null 2>&1 || ! command -v g++-15 >/dev/null 2>&1; then
+              brew install gcc@15 || brew install gcc
+            fi
           fi
 
       - name: Set up Python ${{ inputs.python_version }}
@@ -61,9 +72,28 @@ jobs:
       - name: Install build + runtime deps in venv
         uses: ./.github/actions/setup-venv
         with:
-          packages: "scikit-build-core nanobind cmake>=3.15 pytest"
+          packages: "scikit-build-core>=0.10.0 nanobind>=2.0.0 cmake>=3.15 pytest>=6.0"
 
       - name: Run packaging matrix
         run: |
           . .venv/bin/activate
+          if [[ "${{ runner.os }}" == "macOS" ]]; then
+            export MACOSX_DEPLOYMENT_TARGET
+            MACOSX_DEPLOYMENT_TARGET="$(sw_vers -productVersion | cut -d. -f1,2)"
+          fi
+          if command -v ccache >/dev/null 2>&1; then
+            export CCACHE_DIR="$RUNNER_TEMP/simpler-packaging-ccache"
+            # Keep one cold build per job while reusing identical objects
+            # across the remaining clean-install modes.
+            if ccache --clear && ccache --max-size=1G && ccache --zero-stats; then
+              export CMAKE_C_COMPILER_LAUNCHER=ccache
+              export CMAKE_CXX_COMPILER_LAUNCHER=ccache
+              trap 'ccache --show-stats || true' EXIT
+            else
+              unset CCACHE_DIR
+              echo "::warning::ccache initialization failed; running without compiler cache"
+            fi
+          else
+            echo "::notice::ccache not found; running packaging matrix without compiler cache"
+          fi
           bash tools/verify_packaging.sh

.github/workflows/_profiling-flags-smoke.yml

@@ -60,24 +60,26 @@ jobs:
       - name: Run all profiling flag combos x 2 arches
         run: |
           COMBO_NAMES=(dfx-off orch orch-tensormap sched orch-sched all-on)
-          COMBO_DEFS=(
-            "-DSIMPLER_DFX=0"
-            "-DSIMPLER_ORCH_PROFILING=1"
-            "-DSIMPLER_ORCH_PROFILING=1 -DSIMPLER_TENSORMAP_PROFILING=1"
-            "-DSIMPLER_SCHED_PROFILING=1"
-            "-DSIMPLER_ORCH_PROFILING=1 -DSIMPLER_SCHED_PROFILING=1"
-            "-DSIMPLER_ORCH_PROFILING=1 -DSIMPLER_SCHED_PROFILING=1 -DSIMPLER_TENSORMAP_PROFILING=1"
-          )
+          COMBO_DFX=(0 1 1 1 1 1)
+          COMBO_ORCH=(0 1 1 0 1 1)
+          COMBO_SCHED=(0 0 0 1 1 1)
+          COMBO_TENSORMAP=(0 0 1 0 0 1)
           ARCHES=(a2a3sim a5sim)
           FAIL=()
 
-          for ARCH in "${ARCHES[@]}"; do
-            ARCH_DIR=${ARCH%sim}
-            for i in "${!COMBO_NAMES[@]}"; do
-              NAME=${COMBO_NAMES[$i]}
-              DEFS=${COMBO_DEFS[$i]}
-              echo "::group::$ARCH / $NAME - CXX defines: $DEFS"
-              if ! CXX="g++ $DEFS" .venv/bin/python simpler_setup/build_runtimes.py --platforms "$ARCH"; then
+          # Each architecture keeps one persistent CMake tree. Updating the
+          # generated config header rebuilds only objects that include it.
+          for i in "${!COMBO_NAMES[@]}"; do
+            NAME=${COMBO_NAMES[$i]}
+            for ARCH in "${ARCHES[@]}"; do
+              ARCH_DIR=${ARCH%sim}
+              echo "::group::$ARCH / $NAME"
+              if ! .venv/bin/python simpler_setup/build_runtimes.py \
+                   --platforms "$ARCH" \
+                   --profiling-dfx "${COMBO_DFX[$i]}" \
+                   --profiling-orch "${COMBO_ORCH[$i]}" \
+                   --profiling-sched "${COMBO_SCHED[$i]}" \
+                   --profiling-tensormap "${COMBO_TENSORMAP[$i]}"; then
                 echo "::error::Build failed: $ARCH / $NAME"
                 FAIL+=("$ARCH/$NAME (build)")
                 echo "::endgroup::"

.github/workflows/_st-pod.yml

@@ -195,13 +195,10 @@ jobs:
         if: always()
         uses: ./.github/actions/pod-teardown
 
-      # Both sides' device logs and each daemon's output land under the run
-      # directory and stay on the runner otherwise. A device-side failure — a
-      # scheduler timeout on the peer, say — names its sub-class only there, so
-      # without this the run page shows the host traceback and nothing that
-      # explains it.
+      # Artifact service availability must not override the POD test result.
       - name: Upload pod logs
         if: always()
+        continue-on-error: true
         uses: actions/upload-artifact@v4
         with:
           name: pod-ci-${{ github.run_id }}-${{ github.run_attempt }}

.github/workflows/_ut-no-hardware.yml

@@ -49,11 +49,7 @@ jobs:
         run: |
           if [[ "${{ runner.os }}" == "Linux" ]]; then
             sudo apt-get update
-            sudo apt-get install -y cmake libgtest-dev
-            cd /usr/src/googletest
-            sudo cmake -B build
-            sudo cmake --build build
-            sudo cmake --install build
+            sudo apt-get install -y libgtest-dev
           else
             brew install googletest
           fi
@@ -65,7 +61,10 @@ jobs:
         if: inputs.setup_variant == 'github'
         run: |
           pip install torch --index-url https://download.pytorch.org/whl/cpu
-          pip install '.[test]'
+          pip install 'scikit-build-core>=0.10.0' 'nanobind>=2.0.0' 'cmake>=3.15'
+          # Worker integration tests exercise both a2a3sim and a5sim.
+          pip install --no-build-isolation \
+            --config-settings=build.targets=build_package_sim -e '.[test]'
 
       - name: venv + deps
         if: inputs.setup_variant == 'self-cpu'

.github/workflows/_ut-npu-a2a3.yml

@@ -46,24 +46,19 @@ jobs:
           system-site-packages: "true"
           source-cann: "true"
 
-      - name: Run Python hardware unit tests (a2a3)
+      - name: Build C++ hardware unit tests (a2a3)
         run: |
           source /usr/local/Ascend/cann/set_env.sh
           source .venv/bin/activate
-          if [ "$(uname -m)" = "x86_64" ]; then
-            python -m pytest tests/ut -m requires_hardware --platform a2a3 --device ${DEVICE_RANGE} -v
-          else
-            task-submit --timeout 1800 --max-time 1800 --device auto --device-num "$DEVICE_NUM" \
-              --run "python -m pytest tests/ut -m requires_hardware --platform a2a3 --device \$TASK_DEVICE -v"
-          fi
+          cmake -B tests/ut/cpp/build -S tests/ut/cpp -DSIMPLER_ENABLE_HARDWARE_TESTS=ON
+          cmake --build tests/ut/cpp/build --target test_comm_lifecycle
 
-      - name: Build and run C++ hardware unit tests (a2a3)
+      - name: Run Python and C++ hardware unit tests (a2a3)
         run: |
           source /usr/local/Ascend/cann/set_env.sh
           source .venv/bin/activate
-          cmake -B tests/ut/cpp/build -S tests/ut/cpp -DSIMPLER_ENABLE_HARDWARE_TESTS=ON
-          cmake --build tests/ut/cpp/build --target test_comm_lifecycle
           if [ "$(uname -m)" = "x86_64" ]; then
+            python -m pytest tests/ut -m requires_hardware --platform a2a3 --device ${DEVICE_RANGE} -v
             python3 -c "
           import json, os
           p = os.environ['DEVICE_RANGE'].split('-'); s, e = p[0], p[-1]
@@ -74,7 +69,7 @@ jobs:
             ctest --test-dir tests/ut/cpp/build -L '^requires_hardware(_a2a3)?$' --resource-spec-file $PWD/tests/ut/cpp/build/resources.json -j$(nproc) --output-on-failure
           else
             task-submit --timeout 1800 --max-time 1800 --device auto --device-num "$DEVICE_NUM" \
-              --run "python3 -c \"import json,os; ids=os.environ['TASK_DEVICE'].split(','); json.dump({'version':{'major':1,'minor':0},'local':[{'npus':[{'id':i,'slots':1} for i in ids]}]}, open('tests/ut/cpp/build/resources.json','w'))\" && ctest --test-dir tests/ut/cpp/build -L '^requires_hardware(_a2a3)?\$' --resource-spec-file $PWD/tests/ut/cpp/build/resources.json -j$(nproc) --output-on-failure"
+              --run "python -m pytest tests/ut -m requires_hardware --platform a2a3 --device \$TASK_DEVICE -v && python3 -c \"import json,os; ids=os.environ['TASK_DEVICE'].split(','); json.dump({'version':{'major':1,'minor':0},'local':[{'npus':[{'id':i,'slots':1} for i in ids]}]}, open('tests/ut/cpp/build/resources.json','w'))\" && ctest --test-dir tests/ut/cpp/build -L '^requires_hardware(_a2a3)?\$' --resource-spec-file $PWD/tests/ut/cpp/build/resources.json -j$(nproc) --output-on-failure"
           fi
 
       - name: Build cann-examples/query

cmake/profiling_build_config.h.in

@@ -0,0 +1,18 @@
+/* Generated by CMake. Do not edit. */
+#ifndef SIMPLER_PROFILING_BUILD_CONFIG_H_
+#define SIMPLER_PROFILING_BUILD_CONFIG_H_
+
+#ifndef SIMPLER_DFX
+#define SIMPLER_DFX @SIMPLER_DFX@
+#endif
+#ifndef SIMPLER_ORCH_PROFILING
+#define SIMPLER_ORCH_PROFILING @SIMPLER_ORCH_PROFILING@
+#endif
+#ifndef SIMPLER_SCHED_PROFILING
+#define SIMPLER_SCHED_PROFILING @SIMPLER_SCHED_PROFILING@
+#endif
+#ifndef SIMPLER_TENSORMAP_PROFILING
+#define SIMPLER_TENSORMAP_PROFILING @SIMPLER_TENSORMAP_PROFILING@
+#endif
+
+#endif  // SIMPLER_PROFILING_BUILD_CONFIG_H_

cmake/profiling_config.cmake

@@ -0,0 +1,54 @@
+# Copyright (c) PyPTO Contributors.
+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
+# CANN Open Software License Agreement Version 2.0 (the "License").
+# Please refer to the License for details. You may not use this file except in compliance with the License.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+# See LICENSE in the root of the software repository for the full text of the License.
+# -----------------------------------------------------------------------------------------------------------
+
+set(_SIMPLER_PROFILING_CONFIG_DIR "${CMAKE_CURRENT_LIST_DIR}")
+
+function(simpler_configure_profiling target)
+    foreach(name IN ITEMS
+            SIMPLER_DFX
+            SIMPLER_ORCH_PROFILING
+            SIMPLER_SCHED_PROFILING
+            SIMPLER_TENSORMAP_PROFILING)
+        if(DEFINED ${name} AND NOT "${${name}}" MATCHES "^[01]$")
+            message(FATAL_ERROR "${name} must be 0 or 1")
+        endif()
+    endforeach()
+
+    if(NOT DEFINED SIMPLER_DFX)
+        set(SIMPLER_DFX 1)
+    endif()
+    if(NOT DEFINED SIMPLER_ORCH_PROFILING)
+        set(SIMPLER_ORCH_PROFILING 0)
+    endif()
+    if(NOT DEFINED SIMPLER_SCHED_PROFILING)
+        set(SIMPLER_SCHED_PROFILING 0)
+    endif()
+    if(NOT DEFINED SIMPLER_TENSORMAP_PROFILING)
+        set(SIMPLER_TENSORMAP_PROFILING 0)
+    endif()
+
+    if(SIMPLER_ORCH_PROFILING AND NOT SIMPLER_DFX)
+        message(FATAL_ERROR "SIMPLER_ORCH_PROFILING requires SIMPLER_DFX=1")
+    endif()
+    if(SIMPLER_SCHED_PROFILING AND NOT SIMPLER_DFX)
+        message(FATAL_ERROR "SIMPLER_SCHED_PROFILING requires SIMPLER_DFX=1")
+    endif()
+    if(SIMPLER_TENSORMAP_PROFILING AND NOT SIMPLER_ORCH_PROFILING)
+        message(FATAL_ERROR "SIMPLER_TENSORMAP_PROFILING requires SIMPLER_ORCH_PROFILING=1")
+    endif()
+
+    set(generated_dir "${CMAKE_CURRENT_BINARY_DIR}/generated")
+    file(MAKE_DIRECTORY "${generated_dir}")
+    configure_file(
+        "${_SIMPLER_PROFILING_CONFIG_DIR}/profiling_build_config.h.in"
+        "${generated_dir}/simpler_profiling_build_config.h"
+        @ONLY
+    )
+    target_include_directories(${target} PRIVATE "${generated_dir}")
+endfunction()

docs/ci.md

@@ -88,12 +88,12 @@ means adding a `test_*.py` wrapper with `@scene_level(SceneTestLevel.POD)`, not
 editing `_st-pod.yml`. `pod_remote_device_count` stays as the peer-resource
 declaration; `pod` is the runner topology, not a pytest selection marker.
 
-Pod logs go to `output/pod-ci-<run>-<attempt>/pytest/` and the whole directory
-is uploaded as one artifact. Parent-side `ASCEND_PROCESS_LOG_PATH` is split per
-pytest nodeid by `st_pod_logs`; peer-side daemon/device logs are grouped for the
-pytest session. Reach for the artifact first on a device-side failure: the host
-traceback only says the peer's scheduler gave up, and the sub-class saying why
-is printed on the device.
+Pod logs go to `output/pod-ci-<run>-<attempt>/pytest/` while the job is running.
+Parent-side `ASCEND_PROCESS_LOG_PATH` is split per pytest nodeid by
+`st_pod_logs`; peer-side daemon/device logs are grouped for the pytest session.
+The directory is uploaded on a best-effort basis: artifact-service failures are
+ignored so they cannot override the POD test result. Preserve relevant
+diagnostics in the inline job log as well when investigating a failure.
 
 Writing an example — the files, the entry module, the `run(...)` entry point,
 the `test_*.py` pod wrapper, and the manual `run_parent.sh` — is covered in

docs/getting-started.md

@@ -107,15 +107,16 @@ All workflows assume an activated project-local venv (see [`.claude/rules/venv-i
 ```bash
 python3 -m venv --system-site-packages .venv
 source .venv/bin/activate
-pip install --no-build-isolation scikit-build-core nanobind 'cmake>=3.15' pytest torch
+pip install --no-build-isolation \
+  'scikit-build-core>=0.10.0' 'nanobind>=2.0.0' 'cmake>=3.15' 'pytest>=6.0' 'torch>=2.3'
 pip install --no-build-isolation -e .
 ```
 
 Editing Python is instant (editable install). Editing C++ requires re-running `pip install --no-build-isolation -e .` (scikit-build-core's auto-rebuild is disabled because it interacts badly with pip's ephemeral build env — see `docs/python-packaging.md`).
 
 **Other supported paths:** `pip install .` (non-editable), `pip install --no-build-isolation .`, `pip install -e .`, and `cmake + PYTHONPATH` (no pip). Full comparison of all 5 paths — what lands where, which entry points work under each, trade-offs — lives in [`docs/python-packaging.md`](python-packaging.md).
 
-**Verifying an install:** the single source of truth is `tools/verify_packaging.sh`, which exercises all 5 install paths × 4 entry points from a fully clean state. CI runs the same script on macOS + Ubuntu (see the `packaging-matrix` job in `.github/workflows/ci.yml`).
+**Verifying an install:** the single source of truth is `tools/verify_packaging.sh`, which exercises all 5 install paths × 2 entry points from a fully clean state. CI runs the same script on macOS + Ubuntu (see the `packaging-matrix` job in `.github/workflows/ci.yml`).
 
 ## Build Process
 

docs/python-packaging.md

@@ -123,7 +123,7 @@ For developers who don't want to use pip at all:
 
 ```bash
 . .venv/bin/activate
-pip install scikit-build-core nanobind 'cmake>=3.15' pytest torch  # one-time
+pip install 'scikit-build-core>=0.10.0' 'nanobind>=2.0.0' 'cmake>=3.15' 'pytest>=6.0' 'torch>=2.3'  # one-time
 cmake -S . -B build/cmake_only -Dnanobind_DIR=$(python -c 'import nanobind; print(nanobind.cmake_dir())')
 cmake --build build/cmake_only
 export PYTHONPATH=$(pwd):$(pwd)/python
@@ -158,11 +158,11 @@ Any change that touches:
 ```bash
 # Locally — same script CI runs.
 source .venv/bin/activate
-pip install scikit-build-core nanobind 'cmake>=3.15' pytest torch  # one-time
+pip install 'scikit-build-core>=0.10.0' 'nanobind>=2.0.0' 'cmake>=3.15' 'pytest>=6.0' 'torch>=2.3'  # one-time
 bash tools/verify_packaging.sh
 ```
 
-The script wipes `build/`, uninstalls `simpler`, and re-runs the install + smoke check from scratch for every mode, so a previous mode's cached binaries cannot mask a regression in the next. Slow (~25–45 min for all 5 modes) but reliable. Both CI and the local invocation use the same script — there's no second source of truth that can drift.
+The script wipes `build/`, uninstalls `simpler`, and re-runs the install + smoke check from scratch for every mode, so a previous mode's cached binaries cannot mask a regression in the next. Smoke checks run from a temporary directory outside the repository, preventing source files from shadowing an incomplete wheel install. Slow (~25–45 min for all 5 modes) but reliable. Both CI and the local invocation use the same script — there's no second source of truth that can drift.
 
 The smoke check itself only verifies imports and each entry point's `--help`. Functional tests (real runtime compilation, scene tests) live in `ut-py`, `ut-cpp`, `st-sim-*`, and the hardware self-hosted jobs.