PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
Update CUDA UT to cu130 (#1992)
Sun, Xuehao525648235484
diff --git a/.azure-pipelines/scripts/ut/run_ut_cuda.sh b/.azure-pipelines/scripts/ut/run_ut_cuda.shindex 92f715c7d..087d18047 100644--- a/.azure-pipelines/scripts/ut/run_ut_cuda.sh+++ b/.azure-pipelines/scripts/ut/run_ut_cuda.sh@@ -31,7 +31,8 @@ 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 cmake requests+ uv pip install pytest-cov cmake requests ninja psutil+ uv cache prune}function print_test_results_table() {@@ -91,14 +92,14 @@ function run_unit_test() {cd ${REPO_PATH}/testrm -rf .coverage* *.xml *.html- uv pip install torch==2.12.0 torchvision torchao --index-url https://download.pytorch.org/whl/cu126- CMAKE_ARGS="-DGGML_CUDA=on -DLLAVA_BUILD=off" uv pip install llama-cpp-python+ uv pip install torch==2.12.1 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/cu130uv 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 transformers chardetuv pip uninstall torch torchvision- uv pip install torch==2.12.0 torchvision torchao --index-url https://download.pytorch.org/whl/cu126+ uv pip install torch==2.12.1 torchvision torchao --index-url https://download.pytorch.org/whl/cu130cd ${REPO_PATH} && uv pip install . && cd ${REPO_PATH}/testpip list > ${LOG_DIR}/ut_pip_list.txt@@ -126,12 +127,11 @@ function run_unit_test_vlm() {cd ${REPO_PATH}/testrm -rf .coverage* *.xml *.html- uv pip install torch==2.12.0 torchvision --index-url https://download.pytorch.org/whl/cu126+ uv pip install torch==2.12.1 torchvision --index-url https://download.pytorch.org/whl/cu130uv pip install git+https://github.com/haotian-liu/[email protected] --no-deps- uv pip install -v git+https://github.com/casper-hansen/[email protected] --no-build-isolationuv pip install flash-attn==2.8.3 --no-build-isolationuv pip install -r test_cuda/requirements_vlm.txt \- --extra-index-url https://download.pytorch.org/whl/cu126 \+ --extra-index-url https://download.pytorch.org/whl/cu130 \--index-strategy unsafe-best-matchuv pip install -U chardetcd ${REPO_PATH} && uv pip install . && cd ${REPO_PATH}/test@@ -161,7 +161,7 @@ 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/cu126 --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/cu130 --index-strategy unsafe-best-matchuv pip install -U chardetcd ${REPO_PATH} && uv pip install . && cd ${REPO_PATH}/test@@ -190,7 +190,12 @@ function run_unit_test_sglang() {cd ${REPO_PATH}/testrm -rf .coverage* *.xml *.html- uv pip install "sglang<0.5.11" --prerelease=allow+ uv pip install -r test_cuda/requirements_sglang.txt \+ --prerelease=allow \+ --extra-index-url https://download.pytorch.org/whl/cu130 \+ --index-strategy unsafe-best-match+ local flashinfer_version=$(uv pip show flashinfer-python 2>/dev/null | grep -i "^Version" | awk '{print $2}')+ uv pip install flashinfer-jit-cache==${flashinfer_version} --index-url https://flashinfer.ai/whl/cu130cd ${REPO_PATH} && uv pip install . && cd ${REPO_PATH}/testpip list > ${LOG_DIR}/sglang_ut_pip_list.txt@@ -219,11 +224,11 @@ function run_unit_test_vllm() {cd ${REPO_PATH}/testrm -rf .coverage* *.xml *.html- vllm_version=$(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/${vllm_version}/cu129 \- --extra-index-url https://download.pytorch.org/whl/cu126 \+ --extra-index-url https://download.pytorch.org/whl/cu130 \--index-strategy unsafe-best-match+ local flashinfer_version=$(uv pip show flashinfer-python 2>/dev/null | grep -i "^Version" | awk '{print $2}')+ uv pip install flashinfer-jit-cache==${flashinfer_version} --index-url https://flashinfer.ai/whl/cu130uv pip install -U chardetcd ${REPO_PATH} && uv pip install . && cd ${REPO_PATH}/testdiff --git a/.azure-pipelines/unit-test-cuda.yml b/.azure-pipelines/unit-test-cuda.ymlindex ba6baca35..bf66fdac9 100644--- a/.azure-pipelines/unit-test-cuda.yml+++ b/.azure-pipelines/unit-test-cuda.yml@@ -72,11 +72,8 @@ stages:script: |export PYTHONUNBUFFERED=1cd ${BUILD_SOURCESDIRECTORY}- if [ ${{ pair.value.PART }} -eq 0 ]; then- CUDA_VERSION="13.0"- else- CUDA_VERSION="any"- fi+ CUDA_VERSION="13.0"+python .azure-pipelines/scripts/cuda_unit_test/runpod_manager.py \--action create \--api_key $(RUNPOD_API_KEY) \diff --git a/test/test_cuda/models/test_fp8_model.py b/test/test_cuda/models/test_fp8_model.pyindex ce1ae47c9..1b98ab7d0 100644--- a/test/test_cuda/models/test_fp8_model.py+++ b/test/test_cuda/models/test_fp8_model.py@@ -90,7 +90,7 @@ class TestAutoRound:# inputs = tokenizer(text, return_tensors="pt").to(model.device)# print(tokenizer.decode(model.generate(**inputs, max_new_tokens=50)[0]))- @pytest.mark.skip_ci(reason="Triton issue; time-consuming")+ @pytest.mark.skip_ci(reason="time-consuming")def test_small_model_rtn(self, mock_fp8_capable_device):model_name = get_model_path("Qwen/Qwen3-0.6B-FP8")ar = AutoRound(model=model_name, iters=0, disable_opt_rtn=True)@@ -98,7 +98,7 @@ class TestAutoRound:with patch("torch.cuda.get_device_capability", return_value=DEVICE_CAPABILITY): # revert DEVICE_CAPABILITYevaluate_accuracy(folder, threshold=0.25)- @pytest.mark.skip_ci(reason="Triton issue; time-consuming")+ @pytest.mark.skip_ci(reason="time-consuming")def test_small_model_iters1(self, mock_fp8_capable_device):model_name = get_model_path("Qwen/Qwen3-0.6B-FP8")ar = AutoRound(model=model_name, iters=1)@@ -113,7 +113,7 @@ class TestAutoRound:with patch("torch.cuda.get_device_capability", return_value=DEVICE_CAPABILITY): # revert DEVICE_CAPABILITYevaluate_accuracy(folder, threshold=0.33)- @pytest.mark.skip_ci(reason="Triton issue; time-consuming")+ @pytest.mark.skip_ci(reason="time-consuming")def test_small_model_rtn_with_lm_head(self, mock_fp8_capable_device):model_name = get_model_path("Qwen/Qwen3-0.6B-FP8")layer_config = {"lm_head": {"bits": 4}}