HPU only release binary (#302)

Signed-off-by: yiliu30 <[email protected]>
Signed-off-by: Sun, Xuehao <[email protected]>
Co-authored-by: Sun, Xuehao <[email protected]>

Author: yiliu30

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

@@ -18,7 +18,7 @@ LOG_DIR=/auto-round/log_dir
 mkdir -p ${LOG_DIR}
 ut_log_name=${LOG_DIR}/ut.log
 
-find . -name "test*.py" | sed "s,\.\/,python -m pytest --cov=\"${auto_round_path}\" --cov-report term --html=report.html --self-contained-html  --cov-report xml:coverage.xml --cov-append -vs --disable-warnings ,g" > run.sh
+find . -name "test*.py" ! -name "*hpu_only*.py" | sed "s,\.\/,python -m pytest --cov=\"${auto_round_path}\" --cov-report term --html=report.html --self-contained-html  --cov-report xml:coverage.xml --cov-append -vs --disable-warnings ,g" > run.sh
 cat run.sh
 bash run.sh 2>&1 | tee ${ut_log_name}
 
@@ -33,4 +33,4 @@ fi
 # if ut pass, collect the coverage file into artifacts
 cp .coverage ${LOG_DIR}/.coverage
 
-echo "UT finished successfully! "
+echo "UT finished successfully! "

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

@@ -0,0 +1,36 @@
+#!/bin/bash
+set -xe
+
+# install requirements
+echo "set up UT env..."
+pip install pytest-cov pytest-html
+pip list
+
+cd /auto-round/test || exit 1
+find . -type f -exec sed -i '/sys\.path\.insert(0, "\.\.")/d' {} +
+
+export LD_LIBRARY_PATH=/usr/local/lib/:$LD_LIBRARY_PATH
+export FORCE_BF16=1
+export COVERAGE_RCFILE=/auto-round/.azure-pipelines/scripts/ut/.coverage
+auto_round_path=$(python -c 'import auto_round; print(auto_round.__path__[0])')
+
+LOG_DIR=/auto-round/log_dir
+mkdir -p ${LOG_DIR}
+ut_log_name=${LOG_DIR}/ut.log
+
+find . -name "test*hpu_only.py" | sed "s,\.\/,python -m pytest --cov=\"${auto_round_path}\" --cov-report term --html=report.html --self-contained-html  --cov-report xml:coverage.xml --cov-append -vs --disable-warnings ,g" > run.sh
+cat run.sh
+bash run.sh 2>&1 | tee ${ut_log_name}
+
+cp report.html ${LOG_DIR}/
+cp coverage.xml ${LOG_DIR}/
+
+if [ $(grep -c '== FAILURES ==' ${ut_log_name}) != 0 ] || [ $(grep -c '== ERRORS ==' ${ut_log_name}) != 0 ] || [ $(grep -c ' passed' ${ut_log_name}) == 0 ]; then
+    echo "Find errors in pytest case, please check the output..."
+    exit 1
+fi
+
+# if ut pass, collect the coverage file into artifacts
+cp .coverage ${LOG_DIR}/.coverage
+
+echo "UT finished successfully! "

.azure-pipelines/template/docker-template.yml

@@ -16,6 +16,9 @@ parameters:
   - name: repo
     type: string
     default: "https://github.com/intel/auto-round"
+  - name: imageSource
+    type: string
+    default: "build"
 
 steps:
   - task: Bash@3
@@ -24,7 +27,7 @@ steps:
       script: |
         docker ps -a
         if [[ $(docker ps -a | grep -i '${{ parameters.containerName }}'$) ]]; then
-            docker start $(docker ps -aq)
+            docker start ${{ parameters.containerName }}
             echo "remove left files through container ..."
             docker exec ${{ parameters.containerName }} bash -c "ls -a /auto-round && rm -fr /auto-round/* && rm -fr /auto-round/.* && ls -a /auto-round || true"
         fi
@@ -33,7 +36,6 @@ steps:
   - ${{ if eq(parameters.dockerConfigName, 'commonDockerConfig') }}:
       - script: |
           rm -fr ${BUILD_SOURCESDIRECTORY} || sudo rm -fr ${BUILD_SOURCESDIRECTORY} || true
-          echo y | docker system prune
         displayName: "Clean workspace"
 
       - checkout: self
@@ -45,7 +47,6 @@ steps:
           rm -fr ${BUILD_SOURCESDIRECTORY} || sudo rm -fr ${BUILD_SOURCESDIRECTORY} || true
           mkdir ${BUILD_SOURCESDIRECTORY}
           chmod 777 ${BUILD_SOURCESDIRECTORY}
-          echo y | docker system prune
         displayName: "Clean workspace"
 
       - checkout: none
@@ -57,20 +58,26 @@ steps:
           git checkout main
         displayName: "Checkout out main"
 
-  - script: |
-      if [[ ! $(docker images | grep -i ${{ parameters.repoName }}:${{ parameters.repoTag }}) ]]; then
-        cd ${BUILD_SOURCESDIRECTORY}
-        docker build -f .azure-pipelines/docker/${{parameters.dockerFileName}}.devel -t ${{ parameters.repoName }}:${{ parameters.repoTag }} .
-      fi
-      docker images | grep -i ${{ parameters.repoName }}
-      if [[ $? -ne 0 ]]; then
-        echo "NO Such Repo"
-        exit 1
-      fi
-    displayName: "Build develop docker image"
+  - ${{ if eq(parameters.imageSource, 'build') }}:
+      - script: |
+          docker image prune -a -f
+          if [[ ! $(docker images | grep -i ${{ parameters.repoName }}:${{ parameters.repoTag }}) ]]; then
+            docker build -f ${BUILD_SOURCESDIRECTORY}/.azure-pipelines/docker/${{parameters.dockerFileName}}.devel -t ${{ parameters.repoName }}:${{ parameters.repoTag }} .
+          fi
+          docker images | grep -i ${{ parameters.repoName }}
+          if [[ $? -ne 0 ]]; then
+            echo "NO Such Repo"
+            exit 1
+          fi
+        displayName: "Build develop docker image"
+
+  - ${{ if eq(parameters.imageSource, 'pull') }}:
+      - script: |
+            docker pull vault.habana.ai/gaudi-docker/1.18.0/ubuntu22.04/habanalabs/pytorch-installer-2.4.0:latest
+        displayName: "Pull habana docker image"
 
   - script: |
-      docker stop $(docker ps -aq)
+      docker stop ${{ parameters.containerName }}
       docker rm -vf ${{ parameters.containerName }} || true
       env | sort
     displayName: "Clean docker container"
@@ -80,8 +87,15 @@ steps:
         inputs:
           targetType: "inline"
           script: |
-            docker run -dit --disable-content-trust --privileged --name=${{ parameters.containerName }} --shm-size="2g" \
-            -v ${BUILD_SOURCESDIRECTORY}:/auto-round -v /tf_dataset:/tf_dataset -v /tf_dataset2:/tf_dataset2 ${{ parameters.repoName }}:${{ parameters.repoTag }}
+            if [[ "${{ parameters.imageSource }}" == "build" ]]; then
+                docker run -dit --disable-content-trust --privileged --name=${{ parameters.containerName }} --shm-size="2g" \
+                -v ${BUILD_SOURCESDIRECTORY}:/auto-round -v /tf_dataset:/tf_dataset -v /tf_dataset2:/tf_dataset2 \
+                ${{ parameters.repoName }}:${{ parameters.repoTag }}
+            else
+                docker run -dit --disable-content-trust --privileged --name=${{ parameters.containerName }} --shm-size="2g" \
+                --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --net=host --ipc=host \
+                -v ${BUILD_SOURCESDIRECTORY}:/auto-round vault.habana.ai/gaudi-docker/1.18.0/ubuntu22.04/habanalabs/pytorch-installer-2.4.0:latest
+            fi
             echo "Show the container list after docker run ... "
             docker ps -a
         displayName: "Docker run - ${{ parameters.containerName }} Container"

.azure-pipelines/template/ut-template.yml

@@ -17,6 +17,9 @@ parameters:
   - name: utContainerName
     type: string
     default: "AutoRoundUnitTest"
+  - name: imageSource
+    type: string
+    default: "build"
 
 steps:
   - template: docker-template.yml
@@ -27,14 +30,23 @@ steps:
       dockerFileName: "Dockerfile"
       containerName: ${{ parameters.utContainerName }}
       repo: ${{ parameters.repo }}
+      imageSource: ${{ parameters.imageSource }}
 
-  - script: |
-      docker exec ${{ parameters.utContainerName }} bash -c "cd /auto-round \
-        && pip install -r requirements.txt \
-        && python setup.py bdist_wheel && rm -rf build \
-        && pip install dist/auto_round*.whl \
-        && pip list"
-    displayName: "Env Setup"
+  - ${{ if eq(parameters.imageSource, 'build') }}:
+    - script: |
+        docker exec ${{ parameters.utContainerName }} bash -c "cd /auto-round \
+          && pip install -r requirements.txt \
+          && python setup.py bdist_wheel && rm -rf build \
+          && pip install dist/auto_round*.whl \
+          && pip list"
+      displayName: "Env Setup"
+
+  - ${{ if eq(parameters.imageSource, 'pull') }}:
+    - script: |
+        docker exec ${{ parameters.utContainerName }} bash -c "cd /auto-round \
+          && python setup.py install hpu \
+          && pip list"
+      displayName: "HPU Env Setup"
 
   - script: |
       docker exec ${{ parameters.utContainerName }} bash -c "cd /auto-round/.azure-pipelines/scripts \
@@ -54,4 +66,6 @@ steps:
       targetType: "inline"
       script: |
         docker exec ${{ parameters.utContainerName }} bash -c "rm -rf /auto-round/* && rm -rf /auto-round/.* || true"
+        docker stop ${{ parameters.utContainerName }}
+        docker rm -vf ${{ parameters.utContainerName }} || true
     displayName: "Docker clean up"

.azure-pipelines/unit-test-hpu.yml

@@ -0,0 +1,113 @@
+trigger: none
+
+pr:
+  autoCancel: true
+  drafts: false
+  branches:
+    include:
+      - main
+  paths:
+    include:
+      - auto_round
+      - test/test*hpu*'
+      - setup.py
+      - requirements-hpu.txt
+      - .azure-pipelines/scripts/ut
+    exclude:
+      - auto_round/export/export_to_autogptq
+      - auto_round/export/export_to_awq
+
+pool: GAUDI
+
+variables:
+  IMAGE_NAME: "auto-round"
+  IMAGE_TAG: "py310"
+  UPLOAD_PATH: $(Build.SourcesDirectory)/log_dir
+  DOWNLOAD_PATH: $(Build.SourcesDirectory)/log_dir
+  ARTIFACT_NAME: "UT_coverage_report"
+  REPO: $(Build.Repository.Uri)
+
+stages:
+  - stage: Unit_test
+    displayName: Unit Test
+    dependsOn: []
+    jobs:
+      - job:
+        displayName: Unit Test
+        steps:
+          - template: template/ut-template.yml
+            parameters:
+              imageSource: "pull"
+              dockerConfigName: "commonDockerConfig"
+              utScriptFileName: "run_ut_hpu"
+              uploadPath: $(UPLOAD_PATH)
+              utArtifact: "ut"
+
+  - stage: Unit_test_baseline
+    displayName: Unit Test Baseline
+    dependsOn: []
+    jobs:
+      - job:
+        displayName: Unit Test
+        steps:
+          - template: template/ut-template.yml
+            parameters:
+              imageSource: "pull"
+              dockerConfigName: "gitCloneDockerConfig"
+              utScriptFileName: "run_ut_hpu"
+              uploadPath: $(UPLOAD_PATH)
+              utArtifact: "ut_baseline"
+              repo: $(REPO)
+
+  - stage: Coverage
+    displayName: "Coverage Compare"
+    pool:
+      vmImage: "ubuntu-latest"
+    dependsOn: [Unit_test, Unit_test_baseline]
+    jobs:
+      - job: CollectDatafiles
+        steps:
+          - script: |
+              if [[ ! $(docker images | grep -i ${IMAGE_NAME}:${IMAGE_TAG}) ]]; then
+                docker build -f ${BUILD_SOURCESDIRECTORY}/.azure-pipelines/docker/Dockerfile.devel -t ${IMAGE_NAME}:${IMAGE_TAG} .
+              fi
+              docker images | grep -i ${IMAGE_NAME}
+              if [[ $? -ne 0 ]]; then
+                echo "NO Such Repo"
+                exit 1
+              fi
+            displayName: "Build develop docker image"
+
+          - task: DownloadPipelineArtifact@2
+            inputs:
+              artifact:
+              patterns: "*_coverage/.coverage"
+              path: $(DOWNLOAD_PATH)
+
+          - script: |
+              echo "--- create container ---"
+              docker run -d -it --name="collectLogs"  -v ${BUILD_SOURCESDIRECTORY}:/auto-round  ${IMAGE_NAME}:${IMAGE_TAG} /bin/bash
+              echo "--- docker ps ---"
+              docker ps
+              echo "--- collect logs ---"
+              docker exec collectLogs bash -c "cd /auto-round \
+                && python setup.py install hpu \
+                && pip list"
+              docker exec collectLogs /bin/bash +x -c "cd /auto-round/.azure-pipelines/scripts \
+                && bash ut/collect_log.sh"
+            displayName: "Collect UT Coverage"
+
+          - task: PublishPipelineArtifact@1
+            condition: succeededOrFailed()
+            inputs:
+              targetPath: $(UPLOAD_PATH)
+              artifact: $(ARTIFACT_NAME)
+              publishLocation: "pipeline"
+
+          - task: Bash@3
+            condition: always()
+            inputs:
+              targetType: "inline"
+              script: |
+                docker exec collectLogs bash -c "rm -fr /auto-round/* && rm -fr /auto-round/.* || true"
+            displayName: "Docker clean up"

.azure-pipelines/unit-test.yml

@@ -12,8 +12,10 @@ pr:
       - auto_round_extension
       - test
       - setup.py
-      - requirements_pt.txt
+      - requirements.txt
       - .azure-pipelines/scripts/ut
+    exclude:
+      - test/test*hpu*
 
 pool: ICX-16C
 

auto_round/auto_quantizer.py

@@ -46,10 +46,6 @@
 
 from auto_round.backend import get_layer_backend, dynamic_import_inference_linear
 
-import auto_round_extension.qbits.qlinear_qbits as qlinear_qbits
-import auto_round_extension.qbits.qlinear_qbits_gptq as qlinear_qbits_gptq
-import auto_round_extension.ipex.qlinear_ipex_gptq as qlinear_ipex_gptq
-import auto_round_extension.ipex.qlinear_ipex_awq as qlinear_ipex_awq
 from auto_round.backend import BackendInfos
 from transformers.utils.versions import require_version
 from enum import Enum
@@ -110,9 +106,6 @@ def is_auto_round_available():
             )
 
 
-if is_auto_round_available():
-    from auto_round_extension.cuda.post_init import autoround_post_init
-
 
 #
 def get_device(obj: Union[torch.Tensor, nn.Module]):
@@ -588,12 +581,14 @@ def cpu_post_init(self, model):
 
         for n, layer in tqdm(layers, desc=message, total=len(layers),
                              leave=True):
-            if isinstance(layer, (qlinear_qbits.QuantLinear, qlinear_qbits_gptq.QuantLinear)):
+            from auto_round_extension.qbits import qbits_qlinear_classes
+            from auto_round_extension.ipex import ipex_qlinear_classes
+            if isinstance(layer, qbits_qlinear_classes):
                 if dep_check:
                     layer.req_check()
                 layer.post_init()
                 dep_check = False
-            if isinstance(layer, (qlinear_ipex_gptq.QuantLinear, qlinear_ipex_awq.QuantLinear)):
+            if isinstance(layer, ipex_qlinear_classes):
                 layer.post_init()
 
         return model
@@ -702,7 +697,7 @@ class StoreAttr(object):
                             "marlin format requires gptqmodel to be installed, "
                             "`pip install -v gptqmodel --no-build-isolation `")
             self.repack_marlin(model)
-
+        from auto_round_extension.cuda.post_init import autoround_post_init
         model = autoround_post_init(model)
         # there are no side-effects after call qbits_post_init when model quant-type not equal to qbits.
         if self.target_device == "cpu":

auto_round/autoround.py

@@ -1216,7 +1216,7 @@ def save_quantized(self, output_dir=None, format="auto_round", inplace=True, **k
         format = format.split(":")[0]
         if format not in EXPORT_FORMAT:
             logger.error(f"export format only supports {EXPORT_FORMAT.keys()}")
-            exit(-1)
+            raise ValueError(f"export format only supports {EXPORT_FORMAT.keys()}, but got {format}")
         save_quantized_as_format = EXPORT_FORMAT.get(format)
         if "gptq" in format and not self.sym:
             logger.warning(