PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
Update dependencies in CI and installation scripts for cuda compatibility (#1825)
Sun, Xuehaobd0a1f1e393c
diff --git a/.azure-pipelines/scripts/cuda_unit_test/run_cuda_ut.sh b/.azure-pipelines/scripts/cuda_unit_test/run_cuda_ut.shindex 11cacae23..dd38a18c4 100644--- a/.azure-pipelines/scripts/cuda_unit_test/run_cuda_ut.sh+++ b/.azure-pipelines/scripts/cuda_unit_test/run_cuda_ut.sh@@ -62,14 +62,14 @@ function run_unit_test() {# install unit test dependenciesecho "##[group]set up UT env..."cd "${BUILD_SOURCESDIRECTORY}" || exit 1- uv pip install torch==2.11.0 torchvision --index-url https://download.pytorch.org/whl/cu128- uv pip install https://github.com/XuehaoSun/llama-cpp-python/releases/download/v0.3.16/llama_cpp_python-0.3.16-cp312-cp312-linux_x86_64.whl+ uv pip install torch==2.12.0 torchvision torchao --index-url https://download.pytorch.org/whl/cu130+ uv pip install https://github.com/XuehaoSun/llama-cpp-python/releases/download/v0.3.23/llama_cpp_python-0.3.23-py3-none-linux_x86_64.whluv pip install 'git+https://github.com/ggml-org/llama.cpp.git#subdirectory=gguf-py'uv pip install -r test/test_cuda/requirements.txtuv pip install -r test/test_cuda/requirements_diffusion.txtuv pip install -U transformersuv pip uninstall torch torchvision- uv pip install torch==2.11.0 torchvision --index-url https://download.pytorch.org/whl/cu128+ uv pip install torch==2.12.0 torchvision torchao --index-url https://download.pytorch.org/whl/cu130uv pip install .echo "##[endgroup]"diff --git a/.azure-pipelines/scripts/cuda_unit_test/runpod_manager.py b/.azure-pipelines/scripts/cuda_unit_test/runpod_manager.pyindex 21aa866c7..5e9ce69df 100644--- a/.azure-pipelines/scripts/cuda_unit_test/runpod_manager.py+++ b/.azure-pipelines/scripts/cuda_unit_test/runpod_manager.py@@ -12,6 +12,10 @@ TARGET_GPUS = ["NVIDIA L40S","NVIDIA RTX PRO 6000 Blackwell Server Edition","NVIDIA RTX PRO 6000 Blackwell Workstation Edition",+ "NVIDIA L40",+ "NVIDIA A100-SXM4-80GB",+ "NVIDIA H100 80GB HBM3",+ "NVIDIA H200",]DATA_CENTER_IDS = ["AP-JP-1",@@ -41,11 +45,11 @@ DATA_CENTER_IDS = ["US-TX-4","US-WA-1",]-DATA_CENTER_BAN_LIST = ["EUR-IS-1", "EUR-IS-2", "US-IL-1"]+DATA_CENTER_BAN_LIST = ["EUR-IS-1", "EUR-IS-2", "US-IL-1", "EU-CZ-1"]DATA_CENTER_SELECT_LIST = [dc for dc in DATA_CENTER_IDS if dc not in DATA_CENTER_BAN_LIST]REQUIRED_COUNT = 1-IMAGES_NAME = "xuehaosu/azure-agent:v0.1"-CUDA_VERSIONS = ["13.0", "12.9", "12.8", "12.7", "12.6", "12.5", "12.4", "12.3", "12.2", "12.1", "12.0", "11.8"]+CU130_IMAGES_NAME = "ghcr.io/xuehaosun/azure-agent:13.0"+CUDA_VERSIONS = ["13.0"]def run_create_pod(api_key, payload):@@ -102,15 +106,16 @@ def create_pod(args):"gpuTypePriority": "custom","name": args.name,"volumeInGb": 0,- "imageName": IMAGES_NAME,+ "imageName": CU130_IMAGES_NAME,}print(f"🚀 Creating pod: {args.name}...")data = run_create_pod(args.api_key, payload)if data:pod_id = data.get("id")+ location = data.get("machine", {}).get("location", "unknown")if pod_id:- print(f"✅ Pod created successfully! Pod ID: {pod_id}")+ print(f"✅ Pod created successfully! Pod ID: {pod_id}, Location: {location}")print(f" Status is: {data.get('desiredStatus')}")else:print("❌ Failed to create pod (no data returned).")diff --git a/.azure-pipelines/scripts/ut/constraints.txt b/.azure-pipelines/scripts/ut/constraints.txtindex 33092d9a8..6a7778a58 100644--- a/.azure-pipelines/scripts/ut/constraints.txt+++ b/.azure-pipelines/scripts/ut/constraints.txt@@ -1,2 +1,2 @@-torch==2.11.0-torchvision==0.26.0+torch==2.12.0+torchvision==0.27.0diff --git a/.azure-pipelines/scripts/ut/run_ut_cuda.sh b/.azure-pipelines/scripts/ut/run_ut_cuda.shindex e5fa7a245..976a5ea0f 100644--- a/.azure-pipelines/scripts/ut/run_ut_cuda.sh+++ b/.azure-pipelines/scripts/ut/run_ut_cuda.sh@@ -31,7 +31,7 @@ function create_conda_env() {export PATH=/usr/local/cuda/bin${PATH:+:${PATH}}export LD_LIBRARY_PATH=$(python -c "import site; print(site.getsitepackages()[0])")/nvidia/nvjitlink/lib:$LD_LIBRARY_PATHfi- uv pip install pytest-cov pytest-html cmake+ uv pip install pytest-cov pytest-html cmake requests}function print_test_results_table() {@@ -91,14 +91,14 @@ function run_unit_test() {cd ${REPO_PATH}/testrm -rf .coverage* *.xml *.html- uv pip install torch==2.11.0 torchvision --index-url https://download.pytorch.org/whl/cu128- uv pip install https://github.com/XuehaoSun/llama-cpp-python/releases/download/v0.3.16/llama_cpp_python-0.3.16-cp312-cp312-linux_x86_64.whl+ uv pip install torch==2.12.0 torchvision torchao --index-url https://download.pytorch.org/whl/cu126+ uv pip install https://github.com/XuehaoSun/llama-cpp-python/releases/download/v0.3.23/llama_cpp_python-0.3.23+cu128-py3-none-linux_x86_64.whluv pip install 'git+https://github.com/ggml-org/llama.cpp.git#subdirectory=gguf-py'uv pip install -r test_cuda/requirements.txtuv pip install -r test_cuda/requirements_diffusion.txtuv pip install -U transformersuv pip uninstall torch torchvision- uv pip install torch==2.11.0 torchvision --index-url https://download.pytorch.org/whl/cu128+ uv pip install torch==2.12.0 torchvision torchao --index-url https://download.pytorch.org/whl/cu126cd ${REPO_PATH} && uv pip install . && cd ${REPO_PATH}/testpip list > ${LOG_DIR}/ut_pip_list.txt@@ -131,11 +131,13 @@ function run_unit_test_vlm() {cd ${REPO_PATH}/testrm -rf .coverage* *.xml *.html- uv pip install torch==2.11.0 torchvision --index-url https://download.pytorch.org/whl/cu128+ uv pip install torch==2.12.0 torchvision --index-url https://download.pytorch.org/whl/cu126uv pip install git+https://github.com/haotian-liu/[email protected] --no-depsuv pip install -v git+https://github.com/casper-hansen/[email protected] --no-build-isolationuv pip install flash-attn==2.8.3 --no-build-isolation- uv pip install -r test_cuda/requirements_vlm.txt --extra-index-url https://download.pytorch.org/whl/cu128+ uv pip install -r test_cuda/requirements_vlm.txt \+ --extra-index-url https://download.pytorch.org/whl/cu126 \+ --index-strategy unsafe-best-matchcd ${REPO_PATH} && uv pip install . && cd ${REPO_PATH}/testpip list > ${LOG_DIR}/vlm_ut_pip_list.txt@@ -168,7 +170,7 @@ function run_unit_test_llmc() {cd ${REPO_PATH}/testrm -rf .coverage* *.xml *.html- BUILD_TYPE="nightly" uv pip install -r test_cuda/requirements_llmc.txt --extra-index-url https://download.pytorch.org/whl/cu128 --index-strategy unsafe-best-match+ BUILD_TYPE="nightly" uv pip install -r test_cuda/requirements_llmc.txt --extra-index-url https://download.pytorch.org/whl/cu126 --index-strategy unsafe-best-matchcd ${REPO_PATH} && uv pip install . && cd ${REPO_PATH}/testpip list > ${LOG_DIR}/llmc_ut_pip_list.txt@@ -200,7 +202,7 @@ function run_unit_test_sglang() {cd ${REPO_PATH}/testrm -rf .coverage* *.xml *.html- uv pip install -r test_cuda/requirements_sglang.txt --prerelease=allow+ uv pip install "sglang<0.5.11" --prerelease=allowcd ${REPO_PATH} && uv pip install . && cd ${REPO_PATH}/testpip list > ${LOG_DIR}/sglang_ut_pip_list.txt@@ -236,7 +238,7 @@ function run_unit_test_vllm() {vllm_version=$(curl -s https://api.github.com/repos/vllm-project/vllm/releases/latest | jq -r .tag_name | sed 's/^v//')uv pip install -r test_cuda/requirements_vllm.txt \--extra-index-url https://wheels.vllm.ai/${vllm_version}/cu129 \- --extra-index-url https://download.pytorch.org/whl/cu128 \+ --extra-index-url https://download.pytorch.org/whl/cu126 \--index-strategy unsafe-best-matchcd ${REPO_PATH} && uv pip install . && cd ${REPO_PATH}/testdiff --git a/test/test_cuda/integrations/test_sglang.py b/test/test_cuda/integrations/test_sglang.pyindex 65329d335..44aaf3af5 100644--- a/test/test_cuda/integrations/test_sglang.py+++ b/test/test_cuda/integrations/test_sglang.py@@ -1,5 +1,6 @@import gcimport json+import multiprocessing.resource_trackerimport shutilimport sysfrom pathlib import Path@@ -12,6 +13,20 @@ from auto_round import AutoRoundfrom ...helpers import get_model_path, qwen_name_or_path+# A patch to fix the Python `multiprocessing.ResourceTracker` [Errno 10] error.+_original_stop = multiprocessing.resource_tracker.ResourceTracker._stop+++def _patched_stop(self, *args, _orig=_original_stop, **kwargs):+ if _orig is not None:+ try:+ _orig(self, *args, **kwargs)+ except ChildProcessError:+ pass+++multiprocessing.resource_tracker.ResourceTracker._stop = _patched_stop+class TestAutoRound:model_name = qwen_name_or_pathdiff --git a/test/test_cuda/requirements_sglang.txt b/test/test_cuda/requirements_sglang.txtindex cfb861600..e21e7647e 100644--- a/test/test_cuda/requirements_sglang.txt+++ b/test/test_cuda/requirements_sglang.txt@@ -1 +1 @@-sglang < 0.5.11 # due to not support cu128\ No newline at end of file+sglang >= 0.5.11 # due to not support cu128\ No newline at end of file