Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

HPU only release binary (#302)

Yi Liu5cd255edbb5b
diff --git a/.azure-pipelines/scripts/ut/run_ut.sh b/.azure-pipelines/scripts/ut/run_ut.sh
index 5c4289f56..d9ccaebf7 100644
--- a/.azure-pipelines/scripts/ut/run_ut.sh
+++ b/.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 ${LOG_DIR}/.coverage
-echo "UT finished successfully! "
\ No newline at end of file
+echo "UT finished successfully! "
diff --git a/.azure-pipelines/scripts/ut/run_ut_hpu.sh b/.azure-pipelines/scripts/ut/run_ut_hpu.sh
new file mode 100644
index 000000000..77619df96
--- /dev/null
+++ b/.azure-pipelines/scripts/ut/run_ut_hpu.sh
@@ -0,0 +1,36 @@
+#!/bin/bash
+set -xe
+
+# install requirements
+echo "set up UT env..."
+pip install pytest-cov pytest-html
+pip list
+
+cd /auto-round/test || exit 1
+find . -type f -exec sed -i '/sys\.path\.insert(0, "\.\.")/d' {} +
+
+export LD_LIBRARY_PATH=/usr/local/lib/:$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! "
\ No newline at end of file
diff --git a/.azure-pipelines/template/docker-template.yml b/.azure-pipelines/template/docker-template.yml
index cbde3680c..f71123fa3 100644
--- a/.azure-pipelines/template/docker-template.yml
+++ b/.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"
diff --git a/.azure-pipelines/template/ut-template.yml b/.azure-pipelines/template/ut-template.yml
index 46addafa8..2464da86f 100644
--- a/.azure-pipelines/template/ut-template.yml
+++ b/.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"
diff --git a/.azure-pipelines/unit-test-hpu.yml b/.azure-pipelines/unit-test-hpu.yml
new file mode 100644
index 000000000..b0b6bf3d5
--- /dev/null
+++ b/.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"
diff --git a/.azure-pipelines/unit-test.yml b/.azure-pipelines/unit-test.yml
index eff997208..e0cca55cd 100644
--- a/.azure-pipelines/unit-test.yml
+++ b/.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
diff --git a/auto_round/auto_quantizer.py b/auto_round/auto_quantizer.py
index 0cfbf0729..a91cf589c 100644
--- a/auto_round/auto_quantizer.py
+++ b/auto_round/auto_quantizer.py
@@ -46,10 +46,6 @@ from auto_round.utils import get_module, set_module, is_hpu_supported
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 @@ class AutoRoundQuantizer(HfQuantizer):
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 AutoRoundQuantizer(HfQuantizer):
"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":
diff --git a/auto_round/autoround.py b/auto_round/autoround.py
index 86dd36d46..2401a732b 100644
--- a/auto_round/autoround.py
+++ b/auto_round/autoround.py
@@ -1216,7 +1216,7 @@ class AutoRound(object):
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(
diff --git a/auto_round/export/__init__.py b/auto_round/export/__init__.py
index ec3047930..149ede560 100644
--- a/auto_round/export/__init__.py
+++ b/auto_round/export/__init__.py
@@ -12,9 +12,39 @@
# See the License for the specific language governing permissions and
# limitations under the License.
-from .register import EXPORT_FORMAT
-from auto_round.export.export_to_autogptq.export import save_quantized_as_autogptq
-from .export_to_itrex import save_quantized_as_itrex, QuantConfig
-from .export_to_autoround.export import save_quantized_as_autoround
-from .export_to_awq.export import save_quantized_as_autoawq
+from auto_round.export.register import EXPORT_FORMAT, register_format
+
+@register_format("auto_gptq")
+def _save_quantized_as_autogptq(*args, **kwargs):
+ from auto_round.export.export_to_autogptq.export import save_quantized_as_autogptq
+
+ return save_quantized_as_autogptq(*args, **kwargs)
+
+
+@register_format("itrex")
+def _save_quantized_as_itrex(*args, **kwargs):
+ from auto_round.export.export_to_itrex.export import save_quantized_as_itrex
+
+ return save_quantized_as_itrex(*args, **kwargs)
+
+
+@register_format("itrex_xpu")
+def _save_quantized_as_itrex_xpu(*args, **kwargs):
+ from auto_round.export.export_to_itrex.export import save_quantized_as_itrex_xpu
+
+ return save_quantized_as_itrex_xpu(*args, **kwargs)
+
+
+@register_format("auto_round")
+def _save_quantized_as_autoround(*args, **kwargs):
+ from auto_round.export.export_to_autoround.export import save_quantized_as_autoround
+
+ return save_quantized_as_autoround(*args, **kwargs)
+
+
+@register_format("auto_awq")
+def _save_quantized_as_autoawq(*args, **kwargs):
+ from auto_round.export.export_to_awq.export import save_quantized_as_autoawq
+
+ return save_quantized_as_autoawq(*args, **kwargs)
diff --git a/auto_round/export/export_to_autogptq/export.py b/auto_round/export/export_to_autogptq/export.py
index 2ca99b123..6e8e1c90b 100644
--- a/auto_round/export/export_to_autogptq/export.py
+++ b/auto_round/export/export_to_autogptq/export.py
@@ -114,7 +114,6 @@ def pack_layer(name, model, layer_config, backend, pbar):
pbar.update(1)
-@register_format("auto_gptq")
def save_quantized_as_autogptq(output_dir, inplace=True, backend="auto_gptq:exllamav2",
**kwargs):
"""Export the model to autogptq format to easily leverage cuda kernel."""
diff --git a/auto_round/export/export_to_autoround/export.py b/auto_round/export/export_to_autoround/export.py
index aa0ce85a3..f8e769e2c 100644
--- a/auto_round/export/export_to_autoround/export.py
+++ b/auto_round/export/export_to_autoround/export.py
@@ -153,7 +153,6 @@ def pack_layer(name, model, layer_config, backend, pbar):
pbar.update(1)
-@register_format("auto_round")
def save_quantized_as_autoround(output_dir, inplace=True, backend="auto_round:exllamav2", **kwargs):
"""
Saves a quantized model in the auto-round format.
diff --git a/auto_round/export/export_to_awq/export.py b/auto_round/export/export_to_awq/export.py
index d191f3442..bba2423a1 100644
--- a/auto_round/export/export_to_awq/export.py
+++ b/auto_round/export/export_to_awq/export.py
@@ -77,7 +77,6 @@ def pack_layer(name, model, layer_config, backend, pbar):
pbar.update(1)
-@register_format("auto_awq")
def save_quantized_as_autoawq(output_dir, inplace=True, **kwargs):
"""Export the model to autogptq format to easily leverage cuda kernel."""
model = kwargs["model"]
diff --git a/auto_round/export/export_to_itrex/export.py b/auto_round/export/export_to_itrex/export.py
index 950d9d8e0..d0bbb5618 100644
--- a/auto_round/export/export_to_itrex/export.py
+++ b/auto_round/export/export_to_itrex/export.py
@@ -60,7 +60,6 @@ def quant_weight_w_scale(weight, scale, zp, group_size=-1, device="cpu"):
return int_weight
-@register_format("itrex")
def save_quantized_as_itrex(output_dir, inplace=True, **kwargs):
"""Save configure file and weights for CPU backend inference."""
model = kwargs["model"]
@@ -107,7 +106,7 @@ def save_quantized_as_itrex(output_dir, inplace=True, **kwargs):
return compressed_model
-@register_format("itrex_xpu")
+
def save_quantized_as_itrex_xpu(output_dir, inplace=True, **kwargs):
"""Save configure file and weights for XPU backend inference."""
model = kwargs["model"]
diff --git a/auto_round_extension/ipex/__init__.py b/auto_round_extension/ipex/__init__.py
index e69de29bb..2d32570d3 100644
--- a/auto_round_extension/ipex/__init__.py
+++ b/auto_round_extension/ipex/__init__.py
@@ -0,0 +1,6 @@
+from auto_round_extension.ipex.qlinear_ipex_awq import QuantLinear as IpexAWQQuantLinear
+from auto_round_extension.ipex.qlinear_ipex_gptq import (
+ QuantLinear as IpexGPTQQuantLinear,
+)
+
+ipex_qlinear_classes = (IpexAWQQuantLinear, IpexGPTQQuantLinear)
diff --git a/auto_round_extension/qbits/__init__.py b/auto_round_extension/qbits/__init__.py
index e69de29bb..416f1dfed 100644
--- a/auto_round_extension/qbits/__init__.py
+++ b/auto_round_extension/qbits/__init__.py
@@ -0,0 +1,6 @@
+from auto_round_extension.qbits.qlinear_qbits import QuantLinear as QBitsQuantLinear
+from auto_round_extension.qbits.qlinear_qbits_gptq import (
+ QuantLinear as QBitsGPTQQuantLinear,
+)
+
+qbits_qlinear_classes = (QBitsQuantLinear, QBitsGPTQQuantLinear)
diff --git a/requirements-hpu.txt b/requirements-hpu.txt
new file mode 100644
index 000000000..e3f9cc9db
--- /dev/null
+++ b/requirements-hpu.txt
@@ -0,0 +1,11 @@
+accelerate
+datasets
+py-cpuinfo
+sentencepiece
+transformers
+numpy < 2.0
+tqdm
+packaging
+pillow
+numba
+tbb
\ No newline at end of file
diff --git a/setup.py b/setup.py
index fac699380..cfef6a99b 100644
--- a/setup.py
+++ b/setup.py
@@ -3,7 +3,7 @@ from io import open
import os
from setuptools import find_packages, setup
import sys
-
+from functools import lru_cache
os.environ["CC"] = "g++"
os.environ["CXX"] = "g++"
try:
@@ -18,6 +18,19 @@ version = __version__
BUILD_CUDA_EXT = int(os.environ.get('BUILD_CUDA_EXT', '1')) == 1
PYPI_RELEASE = os.environ.get('PYPI_RELEASE', None)
+BUILD_HPU_ONLY = os.environ.get('BUILD_HPU_ONLY', '0') == '1'
+
+@lru_cache(None)
+def is_hpu_available():
+ try:
+ import habana_frameworks.torch.core as htcore # pylint: disable=E0401
+ return True
+ except ImportError:
+ return False
+
+if is_hpu_available():
+ # When HPU is available, we build HPU only by default
+ BUILD_HPU_ONLY = True
def is_cpu_env():
try:
@@ -167,9 +180,54 @@ if BUILD_CUDA_EXT:
"cmdclass": {'build_ext': cpp_extension.BuildExtension}
}
+PKG_INSTALL_CFG = {
+ # overall installation config, pip install neural-compressor
+ "auto_round": {
+ "project_name": "auto_round",
+ "include_packages": find_packages(
+ include=[
+ "auto_round",
+ "auto_round.*",
+ "auto_round_extension",
+ "auto_round_extension.*",
+ ],
+ ),
+ "install_requires": fetch_requirements("requirements.txt"),
+ "extras_require": {
+ "hpu": fetch_requirements("requirements-hpu.txt"),
+ },
+ },
+ "auto_round_hpu": {
+ "project_name": "auto_round_hpu",
+ "include_packages": find_packages(
+ include=["auto_round", "auto_round.*"],
+ exclude=[
+ "auto_round.export.export_to_autogptq",
+ "auto_round.export.export_to_awq",
+ ],
+ ),
+ "install_requires": fetch_requirements("requirements-hpu.txt"),
+ },
+}
+
if __name__ == "__main__":
+ # There are two ways to install hpu-only package:
+ # 1. pip install setup.py hpu.
+ # 2. Within the gaudi docker where the HPU is available, we install the hpu package by default.
+ cfg_key = "auto_round"
+ if "hpu" in sys.argv:
+ sys.argv.remove("hpu")
+ cfg_key = "auto_round_hpu"
+ if BUILD_HPU_ONLY:
+ cfg_key = "auto_round_hpu"
+
+ project_name = PKG_INSTALL_CFG[cfg_key].get("project_name")
+ include_packages = PKG_INSTALL_CFG[cfg_key].get("include_packages", {})
+ install_requires = PKG_INSTALL_CFG[cfg_key].get("install_requires", [])
+ extras_require = PKG_INSTALL_CFG[cfg_key].get("extras_require", {})
+
setup(
- name="auto_round",
+ name=project_name,
author="Intel AIPT Team",
version=version,
@@ -179,10 +237,11 @@ if __name__ == "__main__":
keywords="quantization,auto-around,LLM,SignRound",
license="Apache 2.0",
url="https://github.com/intel/auto-round",
- packages=find_packages(),
+ packages=include_packages,
include_dirs=include_dirs,
##include_package_data=False,
- install_requires=fetch_requirements("requirements.txt"),
+ install_requires=install_requires,
+ extras_require=extras_require,
python_requires=">=3.7.0",
classifiers=[
"Intended Audience :: Science/Research",
@@ -192,5 +251,5 @@ if __name__ == "__main__":
],
include_package_data=True,
package_data={"": ["mllm/templates/*.json"]},
- **additional_setup_kwargs
+ **additional_setup_kwargs,
)
diff --git a/test/test_auto_round_hpu_only.py b/test/test_auto_round_hpu_only.py
new file mode 100644
index 000000000..0e9e0680f
--- /dev/null
+++ b/test/test_auto_round_hpu_only.py
@@ -0,0 +1,3 @@
+def test_import():
+ from auto_round import AutoRound
+ from auto_round.export.export_to_itrex.export import save_quantized_as_itrex, WeightOnlyLinear
\ No newline at end of file
diff --git a/test/test_autoround_export_to_itrex.py b/test/test_autoround_export_to_itrex.py
index e2b1fe1e5..502b4d2c6 100644
--- a/test/test_autoround_export_to_itrex.py
+++ b/test/test_autoround_export_to_itrex.py
@@ -81,7 +81,7 @@ class TestAutoroundExport(unittest.TestCase):
self.assertTrue(torch.all(torch.isclose(out4[0], out5[0], atol=1e-3)))
def test_config(self):
- from auto_round.export import QuantConfig
+ from auto_round.export.export_to_itrex import QuantConfig
config = QuantConfig.from_pretrained("TheBloke/Llama-2-7B-Chat-GPTQ")
config.save_pretrained("quantization_config_dir")