Migrate cpu sdpa from NS flash attn

.azure-pipelines/compatibility-test.yml

@@ -10,8 +10,10 @@ pr:
     include:
       - auto_round
       - auto_round_extension
+      - pyproject.toml
       - setup.py
       - setup.cfg
+      - MANIFEST.in
       - requirements.txt
       - requirements-cpu.txt
       - .azure-pipelines/compatibility-test.yml
@@ -43,6 +45,9 @@ stages:
           Python314_Linux:
             python_version: '3.14'
             vmImage: 'ubuntu-latest'
+          Python314t_Linux:
+            python_version: '3.14t'
+            vmImage: 'ubuntu-latest'
 
           Python310_Windows:
             python_version: '3.10'
@@ -59,18 +64,29 @@ stages:
           Python314_Windows:
             python_version: '3.14'
             vmImage: 'windows-latest'
+          Python314t_Windows:
+            python_version: '3.14t'
+            vmImage: 'windows-latest'
 
       pool:
         vmImage: $(vmImage)
 
       steps:
-      - task: UsePythonVersion@0
-        inputs:
-          versionSpec: '$(python_version)'
-        displayName: 'Use Python $(python_version)'
+      - bash: |
+          curl -LsSf https://astral.sh/uv/install.sh | sh
+          echo "##vso[task.prependpath]$HOME/.local/bin"
+        condition: ne(variables['Agent.OS'], 'Windows_NT')
+        displayName: 'Install uv (Linux)'
+
+      - powershell: |
+          irm https://astral.sh/uv/install.ps1 | iex
+          Write-Host "##vso[task.prependpath]$env:USERPROFILE\.local\bin"
+        condition: eq(variables['Agent.OS'], 'Windows_NT')
+        displayName: 'Install uv (Windows)'
 
       - bash: |
-          python -m pip install --upgrade pip uv
+          uv python install $(python_version)
+          uv venv --python $(python_version)
           uv pip install -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cpu
           uv build
           uv pip install dist/*.tar.gz && uv pip uninstall auto-round
@@ -79,9 +95,8 @@ stages:
         env:
           PYTHONUNBUFFERED: '1'
           UV_NO_PROGRESS: '1'
-          UV_SYSTEM_PYTHON: '1'
         displayName: 'Install dependencies'
 
       - bash: |
-          python -c "import auto_round; print(auto_round.__version__)"
+          uv run python .azure-pipelines/scripts/compat_smoke_test.py
         displayName: 'Run compatibility test'

.azure-pipelines/nightly-test-xpu.yml

@@ -0,0 +1,64 @@
+# Nightly XPU pipeline: runs the XPU unit tests plus the heavier XPU LLMC
+# integration test that is excluded from the fast PR pipeline (unit-test-xpu.yml).
+# Scheduled daily; can also be triggered manually from the Azure DevOps UI.
+
+trigger: none
+pr: none
+
+schedules:
+  - cron: "0 19 * * *"   # 03:00 Asia/Shanghai every day
+    displayName: Daily XPU nightly run
+    branches:
+      include:
+        - main
+    always: true
+
+# use XPU BMG B60 agent pool to run tests
+pool: B60
+
+variables:
+  IMAGE_NAME: "auto-round"
+  IMAGE_TAG: "py312-xpu"
+  DOCKERFILE_NAME: "Dockerfile_xpu"
+  UPLOAD_PATH: $(Build.SourcesDirectory)/log_dir
+  DOWNLOAD_PATH: $(Build.SourcesDirectory)/log_dir
+  ARTIFACT_NAME: "Nightly_XPU_coverage_report"
+  REPO: $(Build.Repository.Uri)
+
+stages:
+  - template: template/lib-build-template.yml
+    parameters:
+      enableChangeDetection: false
+      publishToTestPyPI: false
+
+  - stage: Nightly_xpu_test
+    displayName: Nightly XPU Test
+    dependsOn: [BuildArkWheel]
+    condition: in(dependencies.BuildArkWheel.result, 'Succeeded', 'Skipped')
+    jobs:
+      - job:
+        displayName: Nightly XPU Test
+        timeoutInMinutes: 180
+        steps:
+          - template: template/ut-template.yml
+            parameters:
+              dockerConfigName: "commonDockerConfig"
+              utScriptFileName: "run_ut_xpu"
+              # "integration" enables the XPU LLMC integration suite in run_ut_xpu.sh
+              utTestMode: "integration"
+              uploadPath: $(UPLOAD_PATH)
+              utArtifact: "nightly-xpu"
+              imageTag: $(IMAGE_TAG)
+              dockerFileName: $(DOCKERFILE_NAME)
+              utContainerName: "AutoRoundNightlyXPU$(NODE_LABEL)"
+              buildARKWheel: "false"
+
+          - task: UseDotNet@2
+            displayName: 'Use .NET Core sdk 7.0.x'
+            inputs:
+              version: 7.0.x
+
+          - task: PublishCodeCoverageResults@2
+            condition: succeededOrFailed()
+            inputs:
+              summaryFileLocation: $(UPLOAD_PATH)/coverage.xml

.azure-pipelines/nightly-test.yml

@@ -0,0 +1,51 @@
+# Nightly pipeline: runs the slower CPU integration + e2e suites that are
+# excluded from the fast PR unit-test pipeline. Scheduled daily; can also be
+# triggered manually from the Azure DevOps UI.
+
+trigger: none
+pr: none
+
+schedules:
+  - cron: "0 18 * * *"   # 02:00 Asia/Shanghai every day
+    displayName: Daily nightly run
+    branches:
+      include:
+        - main
+    always: true
+
+pool: ICX-16C
+
+variables:
+  IMAGE_NAME: "auto-round"
+  IMAGE_TAG: "py312"
+  UPLOAD_PATH: $(Build.SourcesDirectory)/log_dir
+  DOWNLOAD_PATH: $(Build.SourcesDirectory)/log_dir
+  ARTIFACT_NAME: "Nightly_coverage_report"
+  REPO: $(Build.Repository.Uri)
+
+stages:
+  - template: template/lib-build-template.yml
+    parameters:
+      enableChangeDetection: false
+
+  - stage: Nightly_test
+    displayName: Nightly Integration + E2E Test
+    dependsOn: [BuildArkWheel]
+    condition: in(dependencies.BuildArkWheel.result, 'Succeeded', 'Skipped')
+    jobs:
+      - job:
+        timeoutInMinutes: 180
+        steps:
+          - template: template/ut-template.yml
+            parameters:
+              dockerConfigName: "commonDockerConfig"
+              utScriptFileName: "run_nightly"
+              uploadPath: $(UPLOAD_PATH)
+              utArtifact: "nightly"
+              utContainerName: "AutoRoundNightly$(NODE_LABEL)"
+              buildARKWheel: "false"
+
+          - task: PublishCodeCoverageResults@2
+            condition: succeededOrFailed()
+            inputs:
+              summaryFileLocation: $(UPLOAD_PATH)/coverage.xml

.azure-pipelines/scripts/compat_smoke_test.py

@@ -0,0 +1,30 @@
+"""Post-install smoke test for the compatibility pipeline.
+
+Run as a standalone script (``python .azure-pipelines/scripts/compat_smoke_test.py``)
+so that ``import auto_round`` resolves to the *installed* package instead of the
+source tree at the repository root (``sys.path[0]`` becomes this script's directory,
+and the current working directory is not added for script execution).
+
+It validates that:
+  * the package imports and exposes ``__version__``;
+  * the public ``AutoRound`` entry class is importable;
+  * a registered console script (``auto-round``) is installed and runnable.
+"""
+
+import subprocess
+import sys
+
+import auto_round
+from auto_round import AutoRound
+
+print(f"auto_round imported from: {auto_round.__file__}")
+print(f"auto_round {auto_round.__version__} imported successfully (AutoRound={AutoRound.__name__})")
+
+# Verify the console_scripts entry point was installed and is runnable.
+result = subprocess.run(["auto-round", "--help"], capture_output=True, text=True)
+if result.returncode != 0:
+    sys.stderr.write(result.stdout)
+    sys.stderr.write(result.stderr)
+    raise SystemExit(f"`auto-round --help` failed with exit code {result.returncode}")
+
+print("console script `auto-round` is installed and runnable")

.azure-pipelines/scripts/cuda_unit_test/run_cuda_ut.sh

@@ -65,8 +65,8 @@ function run_unit_test() {
     uv pip install torch==2.13.0 torchvision torchao --index-url https://download.pytorch.org/whl/cu130
     uv pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu130
     uv pip install 'git+https://github.com/ggml-org/llama.cpp.git#subdirectory=gguf-py'
-    uv pip install -r test/test_cuda/requirements.txt
-    uv pip install -r test/test_cuda/requirements_diffusion.txt
+    uv pip install -r test/unit/test_cuda/requirements.txt
+    uv pip install -r test/unit/test_cuda/requirements_diffusion.txt
     uv pip install -U transformers chardet
     uv pip install -U pytest-cov
     uv pip install kernels==0.15.2 # For sm120: https://github.com/huggingface/transformers/blob/v5.13.1/setup.py#L93
@@ -80,17 +80,17 @@ function run_unit_test() {
 
     cd "${BUILD_SOURCESDIRECTORY}/test" || exit 1
 
-    find ./test_cuda -type f -name "test_*.py" | grep -Ev "vlms|llmc|sglang|vllm|multiple_card" | sort > all_tests.txt
+    find ./unit/test_cuda -type f -name "test_*.py" | grep -Ev "vlms|llmc|sglang|vllm|multiple_card" | sort > all_tests.txt
     total_lines=$(wc -l < all_tests.txt)
     NUM_CHUNKS=2
     q=$(( total_lines / NUM_CHUNKS ))
     r=$(( total_lines % NUM_CHUNKS ))
-    if [ "$test_part" -le "$r" ]; then
+    if [ "$test_part" -lt "$r" ]; then
         chunk_size=$(( q + 1 ))
-        start_line=$(( (test_part - 1) * chunk_size + 1 ))
+        start_line=$(( test_part * chunk_size + 1 ))
     else
         chunk_size=$q
-        start_line=$(( r * (q + 1) + (test_part - r - 1) * q + 1 ))
+        start_line=$(( r * (q + 1) + (test_part - r) * q + 1 ))
     fi
     end_line=$(( start_line + chunk_size - 1 ))
     selected_files=$(sed -n "${start_line},${end_line}p" all_tests.txt)
@@ -115,7 +115,7 @@ function run_unit_test_llmc() {
     uv venv --python=3.12 /root/.venv
     uv pip install -U pytest-cov
     BUILD_TYPE="nightly" uv pip install \
-        -r test/test_cuda/requirements_llmc.txt \
+        -r test/integration/test_cuda/requirements_llmc.txt \
         --extra-index-url https://download.pytorch.org/whl/cu130 \
         --index-strategy unsafe-best-match
     uv pip install -U chardet
@@ -127,7 +127,7 @@ function run_unit_test_llmc() {
 
     export COVERAGE_RCFILE="${BUILD_SOURCESDIRECTORY}/.azure-pipelines/scripts/ut/.coverage"
 
-    for test_file in $(find ./test_cuda -name "test_llmc*.py" | sort); do
+    for test_file in $(find ./integration/test_cuda -name "test_llmc*.py" | sort); do
         echo "##[group]Running ${test_file}..."
         local test_basename=$(basename ${test_file} .py)
         local ut_log_name=${LOG_DIR}/unittest_cuda_llmc_${test_basename}.log
@@ -145,7 +145,7 @@ function run_unit_test_sglang() {
     rm -rf /root/.venv
     uv venv --python=3.12 /root/.venv
     uv pip install -U pytest-cov
-    uv pip install -r test/test_cuda/requirements_sglang.txt \
+    uv pip install -r test/integration/test_cuda/requirements_sglang.txt \
         --prerelease=allow \
         --extra-index-url https://download.pytorch.org/whl/cu130 \
         --index-strategy unsafe-best-match
@@ -158,7 +158,7 @@ function run_unit_test_sglang() {
     cd "${BUILD_SOURCESDIRECTORY}/test" || exit 1
     export COVERAGE_RCFILE="${BUILD_SOURCESDIRECTORY}/.azure-pipelines/scripts/ut/.coverage"
 
-    for test_file in $(find ./test_cuda -name "test_sglang*.py" | sort); do
+    for test_file in $(find ./integration/test_cuda ./e2e/test_cuda -name "test_sglang*.py" | sort); do
         echo "##[group]Running ${test_file}..."
         local test_basename=$(basename ${test_file} .py)
         local ut_log_name=${LOG_DIR}/unittest_cuda_sglang_${test_basename}.log
@@ -176,7 +176,7 @@ function run_unit_test_vllm() {
     rm -rf /root/.venv
     uv venv --python=3.12 /root/.venv
     uv pip install -U pytest-cov
-    uv pip install -r test/test_cuda/requirements_vllm.txt \
+    uv pip install -r test/integration/test_cuda/requirements_vllm.txt \
         --extra-index-url https://download.pytorch.org/whl/cu130 \
         --index-strategy unsafe-best-match
     local flashinfer_version=$(uv pip show flashinfer-python 2>/dev/null | grep -i "^Version" | awk '{print $2}')
@@ -189,7 +189,7 @@ function run_unit_test_vllm() {
     cd "${BUILD_SOURCESDIRECTORY}/test" || exit 1
     export COVERAGE_RCFILE="${BUILD_SOURCESDIRECTORY}/.azure-pipelines/scripts/ut/.coverage"
 
-    for test_file in $(find ./test_cuda -name "test_vllm*.py" | sort); do
+    for test_file in $(find ./integration/test_cuda ./e2e/test_cuda -name "test_vllm*.py" | sort); do
         echo "##[group]Running ${test_file}..."
         local test_basename=$(basename ${test_file} .py)
         local ut_log_name=${LOG_DIR}/unittest_cuda_vllm_${test_basename}.log
@@ -203,16 +203,14 @@ function run_unit_test_vllm() {
 
 function main() {
     setup_environment
-    if [ "${test_case}" == "vlm" ]; then
-        run_unit_test_vlm
-    elif [ "${test_case}" == "specific" ]; then
+    if [ "${test_case}" == "nightly" ]; then
         run_unit_test_sglang
         run_unit_test_llmc
         run_unit_test_vllm
-    elif [ "${test_case}" == "all" ]; then
+    elif [ "${test_case}" == "ci" ]; then
         run_unit_test
     else
-        echo "##[error]Invalid test case specified: ${test_case}. Please use 'vlm', 'specific', or 'all'."
+        echo "##[error]Invalid test case specified: ${test_case}. Please use 'nightly' or 'ci'."
         exit 1
     fi
     check_storage_usage

.azure-pipelines/scripts/ut/collect_result.py

@@ -58,6 +58,8 @@ class XmlAnalyzer:
         "unittest_cuda_llmc_",
         "unittest_cuda_",
         "unittest_",
+        "integration_",
+        "e2e_",
     )
 
     def __init__(self, log_dir: Path, log_pattern: str = "*.log"):

.azure-pipelines/scripts/ut/run_ut.sh

@@ -17,14 +17,16 @@ function setup_environment() {
     export TQDM_MININTERVAL=120
     export HF_HUB_DISABLE_PROGRESS_BARS=1
 
-    uv pip install pytest-cov
-    uv pip install -U chardet
-    uv pip list
-
     # install latest gguf for ut test
     cd ~ || exit 1
     git clone -b master --quiet --single-branch https://github.com/ggml-org/llama.cpp.git && cd llama.cpp/gguf-py && uv pip install .
 
+    # install unit report dependencies
+    uv pip install pytest-cov
+    uv pip install -U chardet
+    uv pip list
+
+    # install auto-round for unit tests
     cd /auto-round && uv pip install .
 
     export LD_LIBRARY_PATH=${HOME}/.venv/lib/:$LD_LIBRARY_PATH
@@ -65,8 +67,10 @@ function check_storage_usage() {
 function run_unit_test() {
     cd /auto-round/test || exit 1
 
-    # Split test files into 5 parts
-    find ./test_cpu -name "test*.py" | grep -Ev "test_llmc|test_inc" | sort > all_tests.txt
+    # Split test files into 5 parts.
+    # Only fast unit tests run in PR CI; integration (inc/llmc) and e2e suites
+    # run in the nightly/weekly pipelines (see nightly-test.yml / weekly-test.yml).
+    find ./unit/test_cpu -name "test*.py" | sort > all_tests.txt
     total_lines=$(wc -l < all_tests.txt)
     NUM_CHUNKS=5
     q=$(( total_lines / NUM_CHUNKS ))
@@ -95,7 +99,7 @@ function run_unit_test() {
 
 function run_inc_unit_test() {
     echo "##[group]set up INC UT env..."
-    INC_PT_ONLY=1 uv pip install -r /auto-round/test/test_cpu/requirements_inc.txt --extra-index-url https://download.pytorch.org/whl/cpu
+    INC_PT_ONLY=1 uv pip install -r /auto-round/test/integration/test_cpu/requirements_inc.txt --extra-index-url https://download.pytorch.org/whl/cpu
     echo "##[endgroup]"
 
     cd /auto-round/test || exit 1
@@ -114,7 +118,7 @@ function run_inc_unit_test() {
 
 function run_llmc_unit_test() {
     echo "##[group]set up LLMC UT env..."
-    BUILD_TYPE="nightly" uv pip install -r /auto-round/test/test_cpu/requirements_llmc.txt --extra-index-url https://download.pytorch.org/whl/cpu
+    BUILD_TYPE="nightly" uv pip install -r /auto-round/test/integration/test_cpu/requirements_llmc.txt --extra-index-url https://download.pytorch.org/whl/cpu
     uv pip uninstall auto-round
     cd /auto-round && uv pip install .
     echo "##[endgroup]"
@@ -143,10 +147,6 @@ function collect_log() {
 function main() {
     setup_environment
     run_unit_test
-    if [ "$test_part" -eq 5 ]; then
-        run_inc_unit_test
-        run_llmc_unit_test
-    fi
     collect_log
     check_storage_usage
     print_summary