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

Refactor test timeouts and update dependencies for performance (#2361)

Sun, Xuehao1965bdf02fee
diff --git a/.azure-pipelines/scripts/cuda_unit_test/run_cuda_ut.sh b/.azure-pipelines/scripts/cuda_unit_test/run_cuda_ut.sh
index 394d2d9cc..e0413b795 100644
--- a/.azure-pipelines/scripts/cuda_unit_test/run_cuda_ut.sh
+++ b/.azure-pipelines/scripts/cuda_unit_test/run_cuda_ut.sh
@@ -85,7 +85,7 @@ function setup_basic_test_env() {
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 -U pytest-cov pytest-timeout
uv pip install kernels==0.16.1 # For sm120: https://github.com/huggingface/transformers/blob/v5.16.1/setup.py#L93
uv pip uninstall torch torchvision
uv pip install torch==2.14.0 torchvision torchao --index-url https://download.pytorch.org/whl/cu130
@@ -105,7 +105,7 @@ function run_pytest() {
echo "##[group]Running ${test_case}..."
# Record the test targets so a retry can rerun exactly these cases.
printf '%s\n' ${test_case} > "${ut_log_name%.log}.list"
- pytest -m "not skip_ci" --cov=auto_round --cov-report= --cov-append -vs \
+ pytest -m "not skip_ci" --timeout=600 --cov=auto_round --cov-report= --cov-append -vs \
--junitxml="${ut_log_name%.log}.xml" ${test_case} 2>&1 | tee ${ut_log_name}
echo "##[endgroup]"
}
@@ -179,7 +179,7 @@ function run_unit_test_llmc() {
cd "${BUILD_SOURCESDIRECTORY}" || exit 1
rm -rf /root/.venv
uv venv --python=3.14 /root/.venv
- uv pip install -U pytest-cov
+ uv pip install -U pytest-cov pytest-timeout
BUILD_TYPE="nightly" uv pip install \
-r test/integration/test_cuda/requirements_llmc.txt \
--extra-index-url https://download.pytorch.org/whl/cu130 \
@@ -203,7 +203,7 @@ function run_unit_test_sglang() {
cd "${BUILD_SOURCESDIRECTORY}" || exit 1
rm -rf /root/.venv
uv venv --python=3.14 /root/.venv
- uv pip install -U pytest-cov
+ uv pip install -U pytest-cov pytest-timeout
uv pip install -r test/integration/test_cuda/requirements_sglang.txt \
--prerelease=allow \
--extra-index-url https://download.pytorch.org/whl/cu130 \
@@ -228,7 +228,7 @@ function run_unit_test_vllm() {
cd "${BUILD_SOURCESDIRECTORY}" || exit 1
rm -rf /root/.venv
uv venv --python=3.14 /root/.venv
- uv pip install -U pytest-cov
+ uv pip install -U pytest-cov pytest-timeout
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
diff --git a/.azure-pipelines/scripts/ut/run_ark_ut.sh b/.azure-pipelines/scripts/ut/run_ark_ut.sh
index b59accd5c..648252724 100644
--- a/.azure-pipelines/scripts/ut/run_ark_ut.sh
+++ b/.azure-pipelines/scripts/ut/run_ark_ut.sh
@@ -8,7 +8,7 @@ function setup_environment() {
cd /auto-round/auto_round_extension/ark
uv pip install -r requirements.txt
uv pip install /auto-round/ark_wheel/*.whl --no-deps
- uv pip install pytest pandas
+ uv pip install pytest pytest-timeout pandas
uv pip list
echo "##[endgroup]"
@@ -25,7 +25,7 @@ function run_unit_test() {
echo "##[group]Running ark ${test_file}..."
local ut_log_name="${LOG_DIR}/unittest_ark_${test_basename}.log"
- COVERAGE_CORE=sysmon pytest -v ${test_file} 2>&1 | tee ${ut_log_name}
+ COVERAGE_CORE=sysmon pytest --timeout=600 -v ${test_file} 2>&1 | tee ${ut_log_name}
echo "##[endgroup]"
done
}
diff --git a/.azure-pipelines/scripts/ut/run_ut.sh b/.azure-pipelines/scripts/ut/run_ut.sh
index b9324b958..6fee9f508 100644
--- a/.azure-pipelines/scripts/ut/run_ut.sh
+++ b/.azure-pipelines/scripts/ut/run_ut.sh
@@ -27,7 +27,7 @@ function setup_environment() {
git clone -b master --quiet --single-branch https://github.com/ggml-org/llama.cpp.git && cd llama.cpp/gguf-py && uv pip install .
echo "Install unit report dependencies ..."
- uv pip install pytest-cov
+ uv pip install pytest-cov pytest-timeout
uv pip install -U chardet
echo "Install auto-round for unit tests ..."
@@ -65,7 +65,7 @@ function run_pytest() {
# Record the test targets so a retry can rerun exactly these cases.
printf '%s\n' ${test_case} > "${ut_log_name%.log}.list"
numactl --physcpubind="${NUMA_CPUSET:-0-15}" --membind="${NUMA_NODE:-0}" \
- pytest -m "not skip_ci" --cov=auto_round --cov-report= --cov-append -vs \
+ pytest -m "not skip_ci" --timeout=600 --cov=auto_round --cov-report= --cov-append -vs \
--junitxml="${ut_log_name%.log}.xml" ${test_case} 2>&1 | tee ${ut_log_name}
echo "##[endgroup]"
}
diff --git a/.azure-pipelines/scripts/ut/run_ut_cuda.sh b/.azure-pipelines/scripts/ut/run_ut_cuda.sh
index a43d550b7..39182b329 100644
--- a/.azure-pipelines/scripts/ut/run_ut_cuda.sh
+++ b/.azure-pipelines/scripts/ut/run_ut_cuda.sh
@@ -4,7 +4,9 @@
set -xe
CONDA_ENV_NAME="unittest_cuda"
-PYTHON_VERSION="3.14"
+PYTHON_VERSION="3.12"
+TORCH_VERSION="2.14.0"
+
REPO_PATH=$(git rev-parse --show-toplevel)
LOG_DIR=${REPO_PATH}/ut_log_dir
SUMMARY_LOG=${LOG_DIR}/results_summary.log
@@ -24,7 +26,7 @@ function create_conda_env() {
if conda info --envs | grep -q "^$CONDA_ENV_NAME\s"; then conda remove -n ${CONDA_ENV_NAME} --all -y; fi
conda create --quiet -n ${CONDA_ENV_NAME} python=${PYTHON_VERSION} setuptools -y
source activate ${CONDA_ENV_NAME} > /dev/null 2>&1
- conda install --quiet -c conda-forge git gxx=11.2.0 gcc=11.2.0 gdb sysroot_linux-64 libgcc uv -y
+ conda install --quiet -c conda-forge git gxx=13.4 gcc=13.4 gdb sysroot_linux-64 libgcc uv -y
export LD_PRELOAD=${CONDA_PREFIX}/lib/libstdc++.so.6
# install AutoRound
@@ -94,14 +96,14 @@ function run_unit_test() {
cd ${REPO_PATH}/test
rm -rf .coverage* *.xml *.html
- uv pip install torch==2.14.0 torchvision torchao --index-url https://download.pytorch.org/whl/cu130
+ uv pip install torch==${TORCH_VERSION} 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@master#subdirectory=gguf-py'
uv pip install -r unit/test_cuda/requirements.txt
uv pip install -r unit/test_cuda/requirements_diffusion.txt
uv pip install -U transformers chardet
uv pip uninstall torch torchvision
- uv pip install torch==2.14.0 torchvision torchao --index-url https://download.pytorch.org/whl/cu130
+ uv pip install torch==${TORCH_VERSION} torchvision torchao --index-url https://download.pytorch.org/whl/cu130
cd ${REPO_PATH} && uv pip install . && cd ${REPO_PATH}/test
pip list > ${LOG_DIR}/ut_pip_list.txt
@@ -130,9 +132,8 @@ function run_unit_test_vlm() {
cd ${REPO_PATH}/test
rm -rf .coverage* *.xml *.html
- uv pip install torch==2.14.0 torchvision --index-url https://download.pytorch.org/whl/cu130
+ uv pip install torch==${TORCH_VERSION} torchvision --index-url https://download.pytorch.org/whl/cu130
uv pip install git+https://github.com/haotian-liu/[email protected] --no-deps
- uv pip install flash-attn==2.8.3 --no-build-isolation
uv pip install -r unit/test_cuda/requirements_vlm.txt \
--extra-index-url https://download.pytorch.org/whl/cu130 \
--index-strategy unsafe-best-match
diff --git a/.azure-pipelines/scripts/ut/run_ut_hpu.sh b/.azure-pipelines/scripts/ut/run_ut_hpu.sh
index ebad33723..8a6a0f1a1 100644
--- a/.azure-pipelines/scripts/ut/run_ut_hpu.sh
+++ b/.azure-pipelines/scripts/ut/run_ut_hpu.sh
@@ -9,7 +9,7 @@ function setup_environment() {
export TZ='Asia/Shanghai'
export TQDM_MININTERVAL=60
export HF_HUB_DISABLE_PROGRESS_BARS=1
- pip install pytest-cov
+ pip install pytest-cov pytest-timeout
pip list
echo "##[endgroup]"
@@ -38,14 +38,14 @@ function run_unit_test() {
echo "##[group]Running ${test_file} in HPU lazy mode..."
local ut_log_name="${LOG_DIR}/unittest_lazy_${test_basename}.log"
- PT_HPU_LAZY_MODE=1 pytest --cov=auto_round \
+ PT_HPU_LAZY_MODE=1 pytest --timeout=600 --cov=auto_round \
--cov-report= --cov-append -vs \
--junitxml="${ut_log_name%.log}.xml" ${test_file} 2>&1 | tee ${ut_log_name}
echo "##[endgroup]"
echo "##[group]Running ${test_file} in HPU compile mode..."
local ut_log_name="${LOG_DIR}/unittest_compile_${test_basename}.log"
- PT_HPU_LAZY_MODE=0 pytest --mode compile --cov=auto_round \
+ PT_HPU_LAZY_MODE=0 pytest --mode compile --timeout=600 --cov=auto_round \
--cov-report= --cov-append -vs \
--junitxml="${ut_log_name%.log}.xml" ${test_file} 2>&1 | tee ${ut_log_name}
echo "##[endgroup]"
diff --git a/.azure-pipelines/scripts/ut/run_ut_xpu.sh b/.azure-pipelines/scripts/ut/run_ut_xpu.sh
index 0837c1f2b..cdba8617c 100644
--- a/.azure-pipelines/scripts/ut/run_ut_xpu.sh
+++ b/.azure-pipelines/scripts/ut/run_ut_xpu.sh
@@ -13,7 +13,7 @@ SESSION_TIMEOUT=600
function setup_environment() {
echo "##[group]set up UT env..."
echo "Install unit report dependencies ..."
- uv pip install pytest-cov
+ uv pip install pytest-cov pytest-timeout
uv pip install -U chardet
# Keep the GGUF conversion helpers in sync with the model conversion code.
@@ -56,7 +56,7 @@ function run_pytest() {
# Record the test targets so a retry can rerun exactly these cases.
printf '%s\n' ${test_case} > "${ut_log_name%.log}.list"
numactl --physcpubind="${NUMA_CPUSET:-0-27}" --membind="${NUMA_NODE:-0}" \
- pytest -m "not skip_ci" --cov=auto_round --cov-report= --cov-append -vs \
+ pytest -m "not skip_ci" --timeout=600 --cov=auto_round --cov-report= --cov-append -vs \
--junitxml="${ut_log_name%.log}.xml" ${test_case} 2>&1 | tee ${ut_log_name}
echo "##[endgroup]"
}
diff --git a/test/integration/test_cpu/test_inc_calib_dataset.py b/test/integration/test_cpu/test_inc_calib_dataset.py
index d707d36df..bfb401fbe 100644
--- a/test/integration/test_cpu/test_inc_calib_dataset.py
+++ b/test/integration/test_cpu/test_inc_calib_dataset.py
@@ -98,7 +98,6 @@ class TestLocalCalibDataset:
)
autoround.quantize()
- @pytest.mark.timeout(60)
def test_apply_chat_template(self, tiny_qwen_model_path):
model_name = tiny_qwen_model_path
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", trust_remote_code=True)
@@ -118,7 +117,6 @@ class TestLocalCalibDataset:
)
autoround.quantize()
- @pytest.mark.timeout(180)
def test_combine_dataset(self, tiny_qwen_model_path):
dataset = "NeelNanda/pile-10k" + "," + "madao33/new-title-chinese" + "," + "mbpp" + "," + "opencode-instruct"
bits, group_size, sym = 4, 128, True
@@ -134,7 +132,6 @@ class TestLocalCalibDataset:
)
autoround.quantize()
- @pytest.mark.timeout(60)
def test_combine_dataset2(self, tiny_opt_model_path):
dataset = "NeelNanda/pile-10k:num=256,mbpp:num=256"
bits, group_size, sym = 4, 128, True
diff --git a/test/integration/test_cpu/test_inc_static_attn.py b/test/integration/test_cpu/test_inc_static_attn.py
index e5b5c983c..1bfa10eef 100644
--- a/test/integration/test_cpu/test_inc_static_attn.py
+++ b/test/integration/test_cpu/test_inc_static_attn.py
@@ -10,7 +10,6 @@ from auto_round import AutoRound
deepseekv3_model_name = get_model_path("tflsxyy/DeepSeek-V3-bf16-4layers")
def test_deepseek_v2(tiny_deepseek_v2_model_path):
model_name = tiny_deepseek_v2_model_path
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=False)
@@ -51,7 +50,6 @@ def setup_deepseekv3():
return model, tokenizer, output_dir, config
def test_deepseek_v3(setup_deepseekv3):
model, tokenizer, output_dir, config = setup_deepseekv3
autoround = AutoRound(
diff --git a/test/integration/test_xpu/test_llmc_integration.py b/test/integration/test_xpu/test_llmc_integration.py
index 4920338c6..3e68bd144 100644
--- a/test/integration/test_xpu/test_llmc_integration.py
+++ b/test/integration/test_xpu/test_llmc_integration.py
@@ -75,7 +75,6 @@ w8a8_static_recipe_modifier = AutoRoundModifier(
@pytest.mark.skipif(torch.xpu.device_count() < 1, reason="test requires at least 1 XPU")
@pytest.mark.parametrize(
"recipe",
[
diff --git a/test/unit/common/advanced/test_low_precision_input_model.py b/test/unit/common/advanced/test_low_precision_input_model.py
index 2a8fbdd4f..aadfcd6f8 100644
--- a/test/unit/common/advanced/test_low_precision_input_model.py
+++ b/test/unit/common/advanced/test_low_precision_input_model.py
@@ -86,7 +86,6 @@ class TestCompressedTensor:
model.model.layers[0].mlp.up_proj.weight.dtype == torch.bfloat16
), "CompressedLinear layer was not converted to Linear"
- @pytest.mark.timeout(60)
def test_w4a16_to_mxfp4(self, tmp_path):
model = get_tiny_model(get_model_path(self.w4a16_model_path))
model.config.name_or_path = None # Clear the name_or_path to avoid MTP copying issues
diff --git a/test/unit/common/models/test_conv1d.py b/test/unit/common/models/test_conv1d.py
index 5a69516d3..b9c3cd2fe 100644
--- a/test/unit/common/models/test_conv1d.py
+++ b/test/unit/common/models/test_conv1d.py
@@ -42,7 +42,6 @@ class TestQuantizationConv1d:
shutil.rmtree("runs", ignore_errors=True)
@pytest.mark.parametrize("device", _AVAILABLE_DEVICES)
- @pytest.mark.timeout(300)
def test_quant(self, dataloader, device, micro_conv1d_model_path):
"""Quantize a Conv1D-based model, save it, reload on `device`, and run inference."""
if device != "cpu" and not is_gptqmodel_available():
diff --git a/test/unit/common/models/test_moe_model.py b/test/unit/common/models/test_moe_model.py
index e608c8437..91fcd7178 100644
--- a/test/unit/common/models/test_moe_model.py
+++ b/test/unit/common/models/test_moe_model.py
@@ -41,7 +41,6 @@ def count_modules_by_type(model, target_module_name_or_class):
return cnt
@pytest.mark.parametrize("scheme", ["MXFP4", "MXFP8"])
def test_gptoss(scheme, tiny_gpt_oss_model_path, tmp_path):
config = AutoConfig.from_pretrained(tiny_gpt_oss_model_path, trust_remote_code=True)
@@ -75,7 +74,6 @@ def test_gptoss(scheme, tiny_gpt_oss_model_path, tmp_path):
loaded_out = loaded_model(inp)
def test_llama4(tiny_llama4_model_path):
output_dir = "./tmp/test_quantized_llama4"
quantized_model, save_folder = quantize_model(
diff --git a/test/unit/common/quantization/test_asym.py b/test/unit/common/quantization/test_asym.py
index 0f6892057..2bb50eacd 100644
--- a/test/unit/common/quantization/test_asym.py
+++ b/test/unit/common/quantization/test_asym.py
@@ -79,7 +79,6 @@ class TestAutoRoundAsym:
# ------------------------------------------------------------------
# RTN path (iters=0): cheap, so runs unconditionally in CI on every device.
# ------------------------------------------------------------------
- @pytest.mark.timeout(60)
@pytest.mark.parametrize("device", _device_params(_AVAILABLE_DEVICES))
@pytest.mark.parametrize("format", _FORMATS)
def test_asym_format_rtn(self, tiny_opt_model_path, format, device):
@@ -239,7 +238,6 @@ class TestAutoRoundAsym:
# heavier still and stay skip_ci -- manual/nightly coverage only).
# ------------------------------------------------------------------
@requires_cuda
- @pytest.mark.timeout(120)
@pytest.mark.parametrize("group_size", [32])
def test_asym_group_size_tuning(self, tiny_opt_model_path, group_size, dataloader):
"""Tuned (iters=1) asym quantization works across group sizes."""
diff --git a/test/unit/common/quantization/test_mix_bits.py b/test/unit/common/quantization/test_mix_bits.py
index 50c267946..f2f7a68be 100644
--- a/test/unit/common/quantization/test_mix_bits.py
+++ b/test/unit/common/quantization/test_mix_bits.py
@@ -172,7 +172,6 @@ class TestAutoRound:
inputs = tokenizer(text, return_tensors="pt").to(model.device)
print(tokenizer.decode(model.generate(**inputs, max_new_tokens=50)[0]))
- @pytest.mark.timeout(60)
def test_mixed_MXFP_autoround_format_loading(self, dataloader):
layer_config = {
"k_proj": {"bits": 8, "act_bits": 8},
diff --git a/test/unit/common/quantization/test_mxfp_nvfp_llmc_export.py b/test/unit/common/quantization/test_mxfp_nvfp_llmc_export.py
index c35588c55..c37973b38 100644
--- a/test/unit/common/quantization/test_mxfp_nvfp_llmc_export.py
+++ b/test/unit/common/quantization/test_mxfp_nvfp_llmc_export.py
@@ -61,7 +61,6 @@ class TestAutoRoundFP:
moe = compressed_model.model.layers[1].mlp
assert hasattr(moe.experts[0].gate_proj.orig_layer, "act_max")
- @pytest.mark.timeout(60)
def test_nvfp4_moe_actmax_ar(self, tiny_deepseek_v2_model_path, dataloader):
model_name = tiny_deepseek_v2_model_path
layer_config = {
@@ -94,7 +93,6 @@ class TestAutoRoundFP:
and lm_head.weight_scale.dtype is torch.float8_e4m3fn
), "Illegal NVFP4 packing for lm_head layer"
- @pytest.mark.timeout(60)
def test_mxfp4_moe_ar(self, tiny_deepseek_v2_model_path, dataloader):
model_name = tiny_deepseek_v2_model_path
layer_config = {
diff --git a/test/unit/common/schemes/test_auto_scheme.py b/test/unit/common/schemes/test_auto_scheme.py
index f32ffff4b..837a96605 100644
--- a/test/unit/common/schemes/test_auto_scheme.py
+++ b/test/unit/common/schemes/test_auto_scheme.py
@@ -308,7 +308,6 @@ class TestAutoScheme:
# Exporting both integer and MXFP schemes loads and rewrites the tiny
# checkpoint twice; on CPU/XPU this can exceed the historical 70s limit.
- @pytest.mark.timeout(180)
def test_auto_scheme_export(self, micro_opt_model_path, tmp_path):
calibration_dataset = _make_local_calibration_dataset(tmp_path)
model_name = micro_opt_model_path
@@ -335,7 +334,6 @@ class TestAutoScheme:
assert mxfp_config["data_type"] == "mx_fp"
assert os.path.exists(os.path.join(int_model_path, "config.json"))
- @pytest.mark.timeout(120)
def test_gguf_user_fixed_embedding_budget(self, micro_qwen_model_path, tmp_path):
"""Regression test: a user-fixed embedding must be budget-priced at its fixed bits.
diff --git a/test/unit/common/schemes/test_scheme.py b/test/unit/common/schemes/test_scheme.py
index 1edc23e19..a0db4a45b 100644
--- a/test/unit/common/schemes/test_scheme.py
+++ b/test/unit/common/schemes/test_scheme.py
@@ -45,7 +45,6 @@ class TestAutoRound:
ar.post_init()
assert ar.bits == 2
- @pytest.mark.timeout(120)
def test_w4a16_mixed(self, micro_qwen_moe_model_path, dataloader):
layer_config = {
@@ -186,7 +185,6 @@ class TestAutoRound:
else:
assert device_list == [f"{major_type}:0", f"{major_type}:1"]
- @pytest.mark.timeout(60)
def test_set_scheme(self, tiny_qwen_model_path):
pytest.importorskip("sentencepiece")
ar = AutoRound(
diff --git a/test/unit/common/utils/test_alg_ext.py b/test/unit/common/utils/test_alg_ext.py
index 7e64bd570..a9fb4c9e5 100644
--- a/test/unit/common/utils/test_alg_ext.py
+++ b/test/unit/common/utils/test_alg_ext.py
@@ -12,7 +12,6 @@ def _local_calibration_dataset(tmp_path):
class TestAlgExt:
- @pytest.mark.timeout(60)
def test_gguf_q4_alg_ext(self, tiny_qwen_model_path, tmp_path):
"""Exercise the GGUF algorithm-extension path not covered by CUDA matrix tests."""
AutoRound(
@@ -28,7 +27,6 @@ class TestAlgExt:
def test_alg_ext_import(self):
from auto_round.algorithms.quantization.sign_roundv2 import SignRoundV2Quantizer
- @pytest.mark.timeout(60)
def test_nvfp4_alg_ext(self, tiny_opt_model_path, tmp_path):
"""Keep NVFP4 as the representative format not exercised by the CUDA matrix."""
AutoRound(
diff --git a/test/unit/common/utils/test_cli_usage.py b/test/unit/common/utils/test_cli_usage.py
index e25234641..39ab04796 100644
--- a/test/unit/common/utils/test_cli_usage.py
+++ b/test/unit/common/utils/test_cli_usage.py
@@ -41,7 +41,6 @@ class TestAutoRoundCmd:
def test_auto_round_cmd(self, monkeypatch):
_assert_cli_ok(monkeypatch, ["auto_round", "-h"])
- @pytest.mark.timeout(90)
def test_auto_round_cmd2(self, monkeypatch, tiny_opt_model_path, tmp_path):
calibration_path = tmp_path / "calibration.json"
calibration_path.write_text(json.dumps(["AutoRound CLI local calibration sample. " * 32]))
diff --git a/test/unit/common/utils/test_generation.py b/test/unit/common/utils/test_generation.py
index 6ff11e6dd..b5635face 100644
--- a/test/unit/common/utils/test_generation.py
+++ b/test/unit/common/utils/test_generation.py
@@ -59,7 +59,6 @@ class TestAutoRoundFormatGeneration:
print(res)
assert "!!!" not in res
- @pytest.mark.timeout(120)
def test_autoround_sym(self, dataloader):
for bits in [4]:
model = AutoModelForCausalLM.from_pretrained(self.model_name, torch_dtype="auto", trust_remote_code=True)
diff --git a/test/unit/test_ark/test_model.py b/test/unit/test_ark/test_model.py
index 6c35da38b..823da1e61 100644
--- a/test/unit/test_ark/test_model.py
+++ b/test/unit/test_ark/test_model.py
@@ -58,7 +58,6 @@ class TestAutoRoundARKBackend:
evaluate_accuracy(model, tokenizer, threshold=tar_acc, batch_size=32, limit=limit)
torch.xpu.empty_cache()
- @pytest.mark.timeout(60)
@pytest.mark.parametrize("format", ["auto_round", "auto_round:gptqmodel"])
@pytest.mark.parametrize("bits, group_size, sym", [(4, 128, True), (8, 128, True)])
@pytest.mark.parametrize("dtype", [torch.bfloat16])
@@ -66,7 +65,6 @@ class TestAutoRoundARKBackend:
def test_formats(self, format, bits, group_size, sym, dtype, device):
self.main_op(format, bits, group_size, sym, dtype, device)
- @pytest.mark.timeout(60)
@pytest.mark.parametrize("format", ["auto_round:auto_awq"])
@pytest.mark.parametrize("bits, group_size, sym", [(4, 32, True)])
@pytest.mark.parametrize("dtype", [torch.float16])
@@ -74,7 +72,6 @@ class TestAutoRoundARKBackend:
def test_awq_fp16(self, format, bits, group_size, sym, dtype, device):
self.main_op(format, bits, group_size, sym, dtype, device)
- @pytest.mark.timeout(300)
@pytest.mark.parametrize("format", ["auto_round"])
@pytest.mark.parametrize("bits, group_size, sym", [(2, 32, False)])
@pytest.mark.parametrize("dtype", [torch.bfloat16])
diff --git a/test/unit/test_cpu/algorithms/test_awq.py b/test/unit/test_cpu/algorithms/test_awq.py
index 72c7e6dc3..380d76661 100644
--- a/test/unit/test_cpu/algorithms/test_awq.py
+++ b/test/unit/test_cpu/algorithms/test_awq.py
@@ -51,7 +51,6 @@ class TestAWQNormalLLM:
assert cfg.awq_seqlen == 128
- @pytest.mark.timeout(90)
def test_awq_w4a16_quantize_and_inference(self, tiny_opt_model_path):
"""W4A16 AWQ quantization produces valid layer_config and the model can generate.
@@ -80,7 +79,6 @@ class TestAWQNormalLLM:
output = generate_prompt(model, tokenizer, device=device)
assert len(output) > 0, "Model should produce non-empty output"
- @pytest.mark.timeout(120)
def test_awq_w4a16_export_default_scheme(self, tiny_opt_model_path):
"""Default W4A16 scheme export: quantization_config has bits=4, group_size=128."""
ar = AutoRound(
@@ -130,7 +128,6 @@ class TestAWQNormalLLM:
assert qconfig["sym"] == sym
assert "auto-round" in qconfig["quant_method"]
- @pytest.mark.timeout(120)
@pytest.mark.parametrize("device", _AVAILABLE_DEVICES)
def test_awq_w4a16_round_trip(self, tiny_opt_model_path, device):
"""Quantize, save, reload on `device`, and generate -- exercises the full save/load round trip.
@@ -161,7 +158,6 @@ class TestAWQNonIntegerSchemes:
MXFP/NVFP scheme.
"""
- @pytest.mark.timeout(60)
@pytest.mark.parametrize("scheme", ["MXFP4", "NVFP4"])
def test_awq_non_integer_scheme_smoke(self, tiny_opt_model_path, scheme):
"""Algorithm/config correctness (bits/act_bits assignment) -- runs once on cpu.
@@ -735,7 +731,6 @@ class TestAWQMoE:
assert kwargs["position_embeddings"][1].shape == (1, 4, 8)
@requires_cuda
- @pytest.mark.timeout(420)
def test_awq_moe_quantized_layers_check(self, tiny_qwen_moe_model_path):
"""AWQ on MoE: expert layers should be quantized, gates/routers stay FP.
@@ -779,7 +774,6 @@ class TestAWQMoE:
# TODO: Investigate and fix the excessive test runtime instead of relying on an increased timeout.
@requires_cuda
- @pytest.mark.timeout(400)
def test_awq_moe_save_quant_config(self, tiny_qwen_moe_model_path):
"""AWQ MoE: saved quantization_config should be consistent and loadable.
diff --git a/test/unit/test_cpu/backends/test_torch_backend.py b/test/unit/test_cpu/backends/test_torch_backend.py
index 5ade5afe2..8665f6176 100644
--- a/test/unit/test_cpu/backends/test_torch_backend.py
+++ b/test/unit/test_cpu/backends/test_torch_backend.py
@@ -55,7 +55,6 @@ class TestTorchBackendFunctional:
@pytest.mark.parametrize("device", _AVAILABLE_DEVICES)
@pytest.mark.parametrize("iters", [0, 1], ids=["rtn", "tuning"])
@pytest.mark.parametrize("sym,group_size,format", _4BIT_CASES)
- @pytest.mark.timeout(120)
def test_torch_backend_4bit(self, tiny_opt_model_path, sym, group_size, format, iters, device):
"""4-bit (a)symmetric model, RTN or tuned, loads via the torch backend and generates text."""
ar = AutoRound(tiny_opt_model_path, bits=4, group_size=group_size, sym=sym, iters=iters, seqlen=2, nsamples=1)
diff --git a/test/unit/test_cpu/core/test_autoround.py b/test/unit/test_cpu/core/test_autoround.py
index 7fe6d3cf1..d639be085 100644
--- a/test/unit/test_cpu/core/test_autoround.py
+++ b/test/unit/test_cpu/core/test_autoround.py
@@ -36,7 +36,6 @@ class TestAutoRound:
def teardown_class(self):
shutil.rmtree("runs", ignore_errors=True)
- @pytest.mark.timeout(90)
def test_bits_setting(self, tiny_opt_model_path):
layer_config = {"model.decoder.layers.0.self_attn.k_proj": {"data_type": "mx_fp8", "group_size": 32}}
autoround = AutoRound(tiny_opt_model_path, iters=2, seqlen=2, nsamples=1, layer_config=layer_config)
@@ -45,7 +44,6 @@ class TestAutoRound:
if module.bits != 8:
raise ValueError(f"Expected bits to be 8, but got {module.bits}")
- @pytest.mark.timeout(90)
def test_layer_config(self, tiny_opt_model_path, dataloader):
model_name = tiny_opt_model_path
layer_config = {"self_attn": {"bits": 4, "data_type": "nv_fp", "act_bits": 16, "group_size": 16}}
@@ -109,7 +107,6 @@ class TestAutoRound:
autoround.quantize()
# TODO: Investigate and fix the excessive test runtime instead of relying on an increased timeout.
- @pytest.mark.timeout(220)
def test_mx_fp4(self, dataloader):
model_name = opt_name_or_path
bits, group_size, sym = 4, 32, False
@@ -158,7 +155,6 @@ class TestAutoRound:
)
autoround.quantize()
- @pytest.mark.timeout(60)
@pytest.mark.parametrize("bits", [2, 3, 4])
def test_g128(self, bits, dataloader):
model_name = opt_name_or_path
@@ -371,7 +367,6 @@ class TestAutoRound:
)
autoround.quantize()
- @pytest.mark.timeout(60)
def test_rtn(self, tiny_opt_model_path):
model_name = tiny_opt_model_path
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", trust_remote_code=True)
@@ -680,7 +675,6 @@ class TestAutoRound:
assert ar.optimizer == torch.optim.AdamW
assert ar.mllm
- @pytest.mark.timeout(60)
def test_attention_mask_in_dataset(self):
from transformers import AutoTokenizer
@@ -698,7 +692,6 @@ class TestAutoRound:
ar = AutoRound(model_name, iters=1, dataset=data, seqlen=8)
ar.quantize()
- @pytest.mark.timeout(60)
def test_attention_mask_via_tokenize_in_dataset(self):
from transformers import AutoTokenizer
diff --git a/test/unit/test_cpu/core/test_resume_integration.py b/test/unit/test_cpu/core/test_resume_integration.py
index 8bcedbb29..6396a65aa 100644
--- a/test/unit/test_cpu/core/test_resume_integration.py
+++ b/test/unit/test_cpu/core/test_resume_integration.py
@@ -41,7 +41,6 @@ def _clean_resume_env():
class TestResumeIntegration:
- @pytest.mark.timeout(60)
def test_resume_skips_already_completed_blocks(self, tiny_opt_model_path, tmp_path):
resume_dir = str(tmp_path / "resume")
# AR_DISK_STREAM_MODEL keeps low_cpu_mem_usage active for this plain
diff --git a/test/unit/test_cpu/export/test_export.py b/test/unit/test_cpu/export/test_export.py
index e2310fab0..059bd398e 100644
--- a/test/unit/test_cpu/export/test_export.py
+++ b/test/unit/test_cpu/export/test_export.py
@@ -202,7 +202,6 @@ class TestAutoRound:
inputs = tokenizer(text, return_tensors="pt").to(model.device)
print(tokenizer.decode(model.generate(**inputs, max_new_tokens=50)[0]))
- @pytest.mark.timeout(120)
@pytest.mark.parametrize("static_kv_dtype", ["fp8", "float16"])
def test_static_afp8_export(self, static_kv_dtype):
import os
@@ -395,7 +394,6 @@ class TestAutoRound:
assert quantization_config["static_attention_dtype"] == "fp8"
assert quantization_config["static_attention_granularity"] == "head"
- @pytest.mark.timeout(120)
def test_awq_lmhead_export(self, dataloader):
bits, sym, group_size = 4, False, 128
model_name = get_model_path("microsoft/phi-4")
@@ -430,7 +428,6 @@ class TestAutoRound:
assert isinstance(lm_head, WQLinear_GEMM), "Illegal AWQ quantization for lm_head layer"
- @pytest.mark.timeout(120)
def test_gptq_lmhead_export(self, dataloader):
bits, sym, group_size = 4, True, 128
# Note that, to save UT tuning time, the local model is intentionally kept lightweight, using only 2 hidden layers.
@@ -598,7 +595,6 @@ class TestAutoRound:
scale_dtype=ar.scale_dtype,
)
- @pytest.mark.timeout(480)
def test_autoawq_qwen3_vl_infer(self, dataloader):
model_path = get_model_path("Qwen/Qwen3-VL-2B-Instruct")
autoround = AutoRound(
diff --git a/test/unit/test_cpu/export/test_gguf_format.py b/test/unit/test_cpu/export/test_gguf_format.py
index ee2379d3f..9929ba594 100644
--- a/test/unit/test_cpu/export/test_gguf_format.py
+++ b/test/unit/test_cpu/export/test_gguf_format.py
@@ -120,7 +120,6 @@ class TestGGUF:
yield
shutil.rmtree(self.save_dir, ignore_errors=True)
- @pytest.mark.timeout(60)
def test_q4_0(self, tiny_qwen_model_path):
bits, group_size, sym = 4, 32, True
autoround = AutoRound(
@@ -154,7 +153,6 @@ class TestGGUF:
assert type(autoround).__name__ == "CompressionOrchestrator"
assert isinstance(autoround.quantize_config, OptimizedRTNConfig)
- @pytest.mark.timeout(60)
def test_func(self):
bits, group_size, sym = 4, 128, True
autoround = AutoRound(
@@ -172,7 +170,6 @@ class TestGGUF:
model = AutoModelForCausalLM.from_pretrained(quantized_model_path, gguf_file=gguf_file, device_map="auto")
eval_generated_prompt(model, self.tokenizer)
- @pytest.mark.timeout(120)
def test_q4_k_m(self, dataloader, tiny_qwen_model_path):
model_name = tiny_qwen_model_path
layer_config = {
@@ -212,7 +209,6 @@ class TestGGUF:
assert autoround.model.model.layers[0].mlp.gate_proj.bits == 8
assert autoround.layer_config["model.layers.0.mlp.gate_proj"]["mostly"] == "gguf:q8_0"
- @pytest.mark.timeout(360)
def test_all_format(self, monkeypatch, tiny_qwen_model_path):
model_name = tiny_qwen_model_path
# for gguf_format in ["gguf:q4_0", "gguf:q4_1", "gguf:q4_k_m", "gguf:q6_k"]:
@@ -243,7 +239,6 @@ class TestGGUF:
)
shutil.rmtree("../../tmp_autoround", ignore_errors=True)
- @pytest.mark.timeout(90)
def test_vlm_gguf(self, tiny_qwen_vl_model_path):
from auto_round import AutoRound
@@ -264,7 +259,6 @@ class TestGGUF:
else:
assert file_size < 270, f"file size {file_size} MB is too large for non-quantized mmproj-model.gguf"
- @pytest.mark.timeout(60)
def test_vlm_gguf_wo_quant_nontext_module(self, tiny_qwen_vl_model_path):
from auto_round import AutoRound
@@ -523,7 +517,6 @@ class TestGGUF:
assert ar.layer_config["model.language_model.embed_tokens"]["bits"] == 6
assert ar.layer_config["model.language_model.embed_tokens"]["super_bits"] == 8
- @pytest.mark.timeout(60)
def test_q2k_mixed(self, tiny_qwen_moe_model_path):
model_name = tiny_qwen_moe_model_path
autoround = AutoRound(
@@ -546,7 +539,6 @@ class TestGGUF:
tensor_types = {tensor.name: tensor.tensor_type.name for tensor in gguf_model.tensors}
assert tensor_types["blk.0.ffn_up_exps.weight"] == "Q2_K"
- @pytest.mark.timeout(60)
def test_q2k_mixed_keeps_only_three_dim_expert_weights_at_q2k(self, tiny_qwen_moe_model_path):
model_name = tiny_qwen_moe_model_path
autoround = AutoRound(
diff --git a/test/unit/test_cpu/export/test_llmc_format.py b/test/unit/test_cpu/export/test_llmc_format.py
index 8b77d29fd..3a080d38a 100644
--- a/test/unit/test_cpu/export/test_llmc_format.py
+++ b/test/unit/test_cpu/export/test_llmc_format.py
@@ -213,7 +213,6 @@ class TestLLMC:
# autoround.quantize()
# autoround.save_quantized("./saved", format="llm_compressor", inplace=True)
- @pytest.mark.timeout(120)
def test_llmcompressor_fp8(self, tmp_path):
## quantize the model
model_name = opt_name_or_path
@@ -296,7 +295,6 @@ class TestLLMC:
and quantization_config["ignore"] == ["lm_head"]
), f"Invalid MXFP8 quantization configuration: {quantization_config}"
- @pytest.mark.timeout(60)
def test_mxfp8_llmcompressor_kv_config(self, tiny_opt_model_path, tmp_path):
ar = AutoRound(
model=tiny_opt_model_path,
@@ -417,7 +415,6 @@ class TestLLMC:
assert q_scale.shape == torch.Size([compressed_model.config.num_attention_heads])
assert k_scale.ndim == 1
- @pytest.mark.timeout(60)
def test_mixed_precision_llmcompressor_format(self, tiny_opt_model_path, tmp_path):
scheme = AutoScheme(
avg_bits=7,
diff --git a/test/unit/test_cpu/models/test_audio_quantization.py b/test/unit/test_cpu/models/test_audio_quantization.py
index faa693b5d..a3ce9dbe9 100644
--- a/test/unit/test_cpu/models/test_audio_quantization.py
+++ b/test/unit/test_cpu/models/test_audio_quantization.py
@@ -61,7 +61,6 @@ class TestMiMoAudioQuantization:
has_quantlinear = any(m.__class__.__name__ == "QuantLinear" for m in quantized_model.modules())
assert has_quantlinear, "Quantized model should contain QuantLinear layers"
- @pytest.mark.timeout(90)
def test_quantize_with_tuning(self, tiny_mimo_audio_model_path, tmp_path):
"""Tuned MiMo-Audio quantization uses local calibration data and produces QuantLinear layers."""
import transformers
diff --git a/test/unit/test_cpu/models/test_diffusion.py b/test/unit/test_cpu/models/test_diffusion.py
index f27602ba5..bb2b077fa 100644
--- a/test/unit/test_cpu/models/test_diffusion.py
+++ b/test/unit/test_cpu/models/test_diffusion.py
@@ -141,7 +141,6 @@ def setup_flux():
return pipe, output_dir
def test_flux_saving(setup_flux):
pipe, output_dir = setup_flux
autoround = AutoRound(
@@ -158,7 +157,6 @@ def test_flux_saving(setup_flux):
shutil.rmtree(output_dir, ignore_errors=True)
def test_flux(setup_flux):
pipe, output_dir = setup_flux
autoround = AutoRound(
diff --git a/test/unit/test_cpu/models/test_mllm.py b/test/unit/test_cpu/models/test_mllm.py
index 4905680f6..3ea8c72d5 100644
--- a/test/unit/test_cpu/models/test_mllm.py
+++ b/test/unit/test_cpu/models/test_mllm.py
@@ -147,7 +147,6 @@ class TestAutoRound:
model = AutoModelForCausalLM.from_pretrained(opt_name_or_path, trust_remote_code=True)
assert is_pure_text_model(model)
- @pytest.mark.timeout(120)
def test_str_input(self):
tokenizer = AutoTokenizer.from_pretrained(self.model_name)
processor = AutoProcessor.from_pretrained(self.model_name, trust_remote_code=True)
diff --git a/test/unit/test_cpu/models/test_moe_alignment.py b/test/unit/test_cpu/models/test_moe_alignment.py
index 9146d9a0b..f15334d1b 100644
--- a/test/unit/test_cpu/models/test_moe_alignment.py
+++ b/test/unit/test_cpu/models/test_moe_alignment.py
@@ -26,7 +26,6 @@ def setup_deepseek_v2_lite():
return model, tokenizer, output_dir, config
def test_moe_scale_alignment_fp8_static(setup_deepseek_v2_lite):
"""Test that FP8_STATIC quantization unifies gate/up input scales across experts."""
# Enable MoE scale unification explicitly
diff --git a/test/unit/test_cpu/quantization/test_model_free.py b/test/unit/test_cpu/quantization/test_model_free.py
index b5a590e60..a4fd6308a 100644
--- a/test/unit/test_cpu/quantization/test_model_free.py
+++ b/test/unit/test_cpu/quantization/test_model_free.py
@@ -1471,7 +1471,6 @@ _TRANSFORMER_TENSORS = {
class TestKimiK25Int4Source:
"""Tests for _copy_metadata_files including subdirectory handling."""
- @pytest.mark.timeout(120)
def test_non_diffusion_copies_subfolders(self, tmp_path):
"""Non-diffusion model: subdirectories should be copied to output."""
model_dir = _make_model_dir(tmp_path, _SIMPLE_CONFIG, _SIMPLE_TENSORS)
@@ -1715,7 +1714,6 @@ class TestModelFreeAutoScheme:
moved to test/unit/test_cpu/utils/test_model_free_utils.py -> TestAutoSchemeHelpers.
"""
- @pytest.mark.timeout(120)
def test_e2e_int_auto_scheme(self, tmp_path, tiny_opt_model_path):
from auto_round import AutoScheme
diff --git a/test/unit/test_cpu/quantization/test_model_free_parity.py b/test/unit/test_cpu/quantization/test_model_free_parity.py
index 1b61405f0..4db582e61 100755
--- a/test/unit/test_cpu/quantization/test_model_free_parity.py
+++ b/test/unit/test_cpu/quantization/test_model_free_parity.py
@@ -478,7 +478,6 @@ def test_disable_model_free_opt_out(tiny_opt_model_path):
("plain_rtn", True, {"disable_opt_rtn": True}),
],
)
def test_mxfp_auto_scheme_quantization_config_parity(
tmp_path, tiny_opt_model_path, test_name, model_free_disable_opt_rtn, regular_kwargs
):
diff --git a/test/unit/test_cpu/utils/test_model_scope.py b/test/unit/test_cpu/utils/test_model_scope.py
index 45368afe6..d874d5b81 100644
--- a/test/unit/test_cpu/utils/test_model_scope.py
+++ b/test/unit/test_cpu/utils/test_model_scope.py
@@ -32,13 +32,11 @@ class TestModelScope:
if os.path.exists(self.cache_path):
shutil.rmtree(self.cache_path, ignore_errors=True)
- @pytest.mark.timeout(120)
def test_llm(self, dataloader):
model_name = save_tiny_model(get_model_path("Qwen/Qwen2.5-0.5B-Instruct"), self.tiny_model_path, num_layers=2)
autoround = AutoRound(model_name, platform="model_scope", scheme="w4a16", iters=0, seqlen=2, dataset=dataloader)
autoround.quantize_and_save(self.saved_path)
- @pytest.mark.timeout(480)
def test_mllm(self, dataloader):
model_name = save_tiny_model(
get_model_path("Qwen/Qwen2-VL-2B-Instruct"), self.tiny_model_path, num_layers=2, is_mllm=True
diff --git a/test/unit/test_cuda/algorithms/test_alg_ext.py b/test/unit/test_cuda/algorithms/test_alg_ext.py
index 3db2033fb..996d9f0df 100644
--- a/test/unit/test_cuda/algorithms/test_alg_ext.py
+++ b/test/unit/test_cuda/algorithms/test_alg_ext.py
@@ -67,7 +67,6 @@ class TestAlgExt:
# "overrides were applied (data_type was not yet 'int_asym_dq')."
# )
- @pytest.mark.timeout(90)
def test_int2_g64_asym_enable_alg_ext_keeps_config(self, tiny_qwen_model_path):
"""Regression test: asym int2/g64 keeps the requested tuning config."""
@@ -101,7 +100,6 @@ class TestAlgExt:
assert quantizer.enable_norm_bias_tuning is True
assert quantizer.enable_quanted_input is False
- @pytest.mark.timeout(120)
@pytest.mark.parametrize("scheme", ["MXFP4", "W2A16G64"])
def test_all_support_dtype(self, scheme, tiny_qwen_model_path, tmp_path, monkeypatch):
from auto_round.auto_scheme import AutoScheme
diff --git a/test/unit/test_cuda/algorithms/test_auto_scheme.py b/test/unit/test_cuda/algorithms/test_auto_scheme.py
index 194b5d12d..b18f6ba2e 100644
--- a/test/unit/test_cuda/algorithms/test_auto_scheme.py
+++ b/test/unit/test_cuda/algorithms/test_auto_scheme.py
@@ -38,7 +38,6 @@ class TestAutoScheme:
yield
shutil.rmtree("runs", ignore_errors=True)
- @pytest.mark.timeout(60)
def test_gguf_k_0(self, tiny_qwen_model_path, tmp_path, monkeypatch):
monkeypatch.setenv("AR_AUTO_SCHEME_NSAMPLES", "1")
monkeypatch.setenv("AR_AUTO_SCHEME_BATCH_SIZE", "1")
@@ -97,7 +96,6 @@ class TestAutoScheme:
# Due to the tiny model and embedding, the actual number of bits of gguf format will be larger than the target bits.
assert target_bits - 0.1 < avg_bits <= target_bits + 0.3
- @pytest.mark.timeout(150)
def test_shared_layers(self, tiny_opt_model_path):
model_name = tiny_opt_model_path
from transformers import AutoModelForCausalLM, AutoTokenizer
@@ -270,7 +268,6 @@ class TestAutoScheme:
print(avg_bits)
assert target_bits - 0.1 < avg_bits <= target_bits + 1e-3
- @pytest.mark.timeout(150)
def test_lm_head_and_mix_dtype(self, tiny_untied_qwen_model_path):
target_bits = 5
scheme = AutoScheme(avg_bits=target_bits, options=("MXFP4", "MXFP8"))
diff --git a/test/unit/test_cuda/algorithms/test_awq.py b/test/unit/test_cuda/algorithms/test_awq.py
index b207bb2a4..da1459920 100644
--- a/test/unit/test_cuda/algorithms/test_awq.py
+++ b/test/unit/test_cuda/algorithms/test_awq.py
@@ -52,7 +52,6 @@ class TestAWQNormalLLM:
assert cfg.awq_seqlen == 128
@pytest.mark.skip_ci(reason="Coverage: CPU-only AWQ algorithm smoke; run in common/nightly coverage")
- @pytest.mark.timeout(90)
def test_awq_w4a16_quantize_and_inference(self, tiny_opt_model_path):
"""W4A16 AWQ quantization produces valid layer_config and the model can generate.
@@ -132,7 +131,6 @@ class TestAWQNormalLLM:
assert qconfig["sym"] == sym
assert "auto-round" in qconfig["quant_method"]
- @pytest.mark.timeout(120)
@pytest.mark.parametrize("device", _AVAILABLE_DEVICES)
def test_awq_w4a16_round_trip(self, tiny_opt_model_path, device):
"""Quantize, save, reload on `device`, and generate -- exercises the full save/load round trip.
@@ -165,7 +163,6 @@ class TestAWQNonIntegerSchemes:
"""
@pytest.mark.skip_ci(reason="Coverage: CPU-only AWQ datatype smoke; run in common/nightly coverage")
- @pytest.mark.timeout(60)
@pytest.mark.parametrize("scheme", ["MXFP4", "NVFP4"])
def test_awq_non_integer_scheme_smoke(self, tiny_opt_model_path, scheme):
"""Algorithm/config correctness (bits/act_bits assignment) -- runs once on cpu.
@@ -740,7 +737,6 @@ class TestAWQMoE:
assert kwargs["position_embeddings"][1].shape == (1, 4, 8)
@requires_cuda
- @pytest.mark.timeout(420)
def test_awq_moe_quantized_layers_check(self, micro_qwen_moe_model_path):
"""AWQ on MoE: expert layers should be quantized, gates/routers stay FP.
@@ -784,7 +780,6 @@ class TestAWQMoE:
# TODO: Investigate and fix the excessive test runtime instead of relying on an increased timeout.
@requires_cuda
- @pytest.mark.timeout(400)
def test_awq_moe_save_quant_config(self, micro_qwen_moe_model_path):
"""AWQ MoE: saved quantization_config should be consistent and loadable.
diff --git a/test/unit/test_cuda/algorithms/test_block_runner.py b/test/unit/test_cuda/algorithms/test_block_runner.py
index 73f5eebca..976798e50 100644
--- a/test/unit/test_cuda/algorithms/test_block_runner.py
+++ b/test/unit/test_cuda/algorithms/test_block_runner.py
@@ -52,7 +52,6 @@ class TestBlockRunnerGpu:
yield
shutil.rmtree(self.save_dir, ignore_errors=True)
- @pytest.mark.timeout(180)
def test_3bit_asym_sign_sgd_reload(self, tiny_opt_model_path, dataloader):
autoround = AutoRound(
tiny_opt_model_path,
@@ -77,7 +76,6 @@ class TestBlockRunnerGpu:
@pytest.mark.skip_ci(
reason="Matrix: group_size=64 packing is covered by test_autoround_int_export; retain full Sign-SGD case weekly"
)
- @pytest.mark.timeout(180)
def test_group64_sign_sgd_reload(self, tiny_opt_model_path):
autoround = AutoRound(
tiny_opt_model_path,
@@ -97,7 +95,6 @@ class TestBlockRunnerGpu:
@pytest.mark.skip_ci(
reason="Coverage: symmetric INT export/reload is covered by test_autoround_int_export; retain tuning case weekly"
)
- @pytest.mark.timeout(180)
def test_4bit_sym_group128_reload(self, tiny_opt_model_path):
autoround = AutoRound(
tiny_opt_model_path,
@@ -117,7 +114,6 @@ class TestBlockRunnerGpu:
@pytest.mark.skip_ci(
reason="Coverage: asymmetric INT2 runtime is covered by the Triton backend smoke; retain full tuning case weekly"
)
- @pytest.mark.timeout(180)
def test_2bit_asym_group32_reload(self, tiny_opt_model_path):
autoround = AutoRound(
tiny_opt_model_path,
@@ -141,7 +137,6 @@ class TestBlockRunnerGpu:
@pytest.mark.skip_ci(
reason="Coverage: test_3bit_asym_sign_sgd_reload already verifies tuned reload and forward; retain longer convergence case weekly"
)
- @pytest.mark.timeout(180)
def test_sign_sgd_forward_after_reload(self, tiny_opt_model_path):
"""Sign-SGD quantized model must forward-pass on GPU after reload."""
autoround = AutoRound(
diff --git a/test/unit/test_cuda/algorithms/test_config_resolver.py b/test/unit/test_cuda/algorithms/test_config_resolver.py
index 0d55a5a94..194cfe0c7 100644
--- a/test/unit/test_cuda/algorithms/test_config_resolver.py
+++ b/test/unit/test_cuda/algorithms/test_config_resolver.py
@@ -30,7 +30,6 @@ class TestConfigResolverGpu:
yield
shutil.rmtree(self.save_dir, ignore_errors=True)
- @pytest.mark.timeout(180)
@pytest.mark.parametrize("scheme", ["W4A16", "W2A16", "W8A16", "MXFP4"])
def test_preset_scheme_export(self, tiny_opt_model_path, scheme):
autoround = AutoRound(tiny_opt_model_path, scheme=scheme, iters=0, disable_opt_rtn=True, nsamples=1, seqlen=16)
@@ -39,7 +38,6 @@ class TestConfigResolverGpu:
config = AutoConfig.from_pretrained(quantized_model_path, trust_remote_code=True)
assert getattr(config, "quantization_config", {}).get("quant_method") == "auto-round"
- @pytest.mark.timeout(180)
def test_layer_config_override(self, tiny_opt_model_path):
"""Per-layer override wins over the default scheme during config resolution."""
layer_config = {"model.decoder.layers.0.self_attn.q_proj": {"bits": 8}}
diff --git a/test/unit/test_cuda/algorithms/test_sign_sgd_pipeline.py b/test/unit/test_cuda/algorithms/test_sign_sgd_pipeline.py
index a5a49712a..b9989a8ff 100644
--- a/test/unit/test_cuda/algorithms/test_sign_sgd_pipeline.py
+++ b/test/unit/test_cuda/algorithms/test_sign_sgd_pipeline.py
@@ -31,7 +31,6 @@ class TestSignSgdPipeline:
yield
shutil.rmtree(self.save_dir, ignore_errors=True)
- @pytest.mark.timeout(180)
def test_sign_sgd_quantize_reload_forward(self, tiny_opt_model_path):
"""Run a real sign-SGD quantization on GPU, reload and forward-pass it."""
autoround = AutoRound(
@@ -56,7 +55,6 @@ class TestSignSgdPipeline:
out = model(input_ids)
assert out.logits.shape[0] == 1
- @pytest.mark.timeout(180)
def test_w4a8_quantize_reload(self, tiny_opt_model_path):
"""W4A8 with iters>0 exercises the activation-quantization hooks on GPU."""
autoround = AutoRound(
@@ -81,7 +79,6 @@ class TestSignSgdPipeline:
out = model(input_ids)
assert out.logits.shape[0] == 1
- @pytest.mark.timeout(180)
def test_sign_sgd_asym(self, tiny_opt_model_path):
"""Asymmetric sign-SGD quantization + export on GPU.
@@ -103,7 +100,6 @@ class TestSignSgdPipeline:
assert isinstance(compressed_model, torch.nn.Module)
assert os.path.isdir(quantized_model_path)
- @pytest.mark.timeout(180)
def test_sign_sgd_bits2_group32(self, tiny_opt_model_path):
"""2-bit, small-group sign-SGD exercises low-bit block optimization."""
autoround = AutoRound(
diff --git a/test/unit/test_cuda/auto_scheme/test_auto_scheme.py b/test/unit/test_cuda/auto_scheme/test_auto_scheme.py
index e6f8d1c3e..6420f2408 100644
--- a/test/unit/test_cuda/auto_scheme/test_auto_scheme.py
+++ b/test/unit/test_cuda/auto_scheme/test_auto_scheme.py
@@ -47,7 +47,6 @@ class TestAutoSchemeGpu:
yield
shutil.rmtree(self.save_dir, ignore_errors=True)
- @pytest.mark.timeout(240)
def test_auto_scheme_quantize_reload(self, tiny_qwen_model_path, tmp_path, monkeypatch):
monkeypatch.setenv("AR_AUTO_SCHEME_NSAMPLES", "1")
monkeypatch.setenv("AR_AUTO_SCHEME_BATCH_SIZE", "1")
@@ -76,7 +75,6 @@ class TestAutoSchemeGpu:
)
assert isinstance(model, torch.nn.Module)
- @pytest.mark.timeout(120)
def test_auto_scheme_avg_bits_in_range(self, tiny_qwen_model_path, tmp_path, monkeypatch):
"""The generated mixed scheme lands near the requested average bit-width."""
from auto_round.auto_scheme.utils import compute_avg_bits_for_model
@@ -99,7 +97,6 @@ class TestAutoSchemeGpu:
avg_bits, _ = compute_avg_bits_for_model(model, ignore_scale_zp_bits=True)
assert target - 0.3 < avg_bits <= target + 0.3
- @pytest.mark.timeout(240)
def test_auto_scheme_iters1_reload(self, tiny_qwen_model_path, tmp_path, monkeypatch):
"""Auto scheme with sign-SGD (iters>0) + reload."""
monkeypatch.setenv("AR_AUTO_SCHEME_NSAMPLES", "1")
diff --git a/test/unit/test_cuda/backends/test_exllamav2_backend.py b/test/unit/test_cuda/backends/test_exllamav2_backend.py
index 434e7db08..cf2677979 100644
--- a/test/unit/test_cuda/backends/test_exllamav2_backend.py
+++ b/test/unit/test_cuda/backends/test_exllamav2_backend.py
@@ -34,7 +34,6 @@ class TestAutoRoundexllamaBackend:
# @pytest.mark.skip_ci(reason="Only tiny model is suggested")
# @pytest.mark.skip_ci(reason="Time-consuming; Accuracy evaluation")
@require_gptqmodel
- @pytest.mark.timeout(240)
def test_gptqmodel_exllmav2_4bits_asym(self, dataloader):
model_path = get_model_path("facebook/opt-125m")
bits, group_size, sym = 4, 128, False
@@ -123,7 +122,6 @@ class TestAutoRoundexllamaBackend:
@pytest.mark.skip_ci(
reason="Backend/JIT: AWQ ExLlamaV2 matrix is covered in nightly; keep the native asym smoke in PR CI"
)
- @pytest.mark.timeout(90)
def test_gptqmodel_awq_exllamav2_4bits_asym(self, dataloader):
"""Test AWQ quantization with gptqmodel:awq_exllamav2 backend (bfloat16 inference)."""
model_path = get_model_path("facebook/opt-125m")
diff --git a/test/unit/test_cuda/backends/test_marlin_backend.py b/test/unit/test_cuda/backends/test_marlin_backend.py
index 44da32583..5b868c5ba 100644
--- a/test/unit/test_cuda/backends/test_marlin_backend.py
+++ b/test/unit/test_cuda/backends/test_marlin_backend.py
@@ -34,7 +34,6 @@ class TestAutoRoundMarlinBackend:
# Keep one CI test for marlin backend and skip others to save time.
# @pytest.mark.skip_ci(reason="Only tiny model is suggested")
# @pytest.mark.skip_ci(reason="Time-consuming; Accuracy evaluation")
- @pytest.mark.timeout(240)
def test_marlin_4bits_sym_with_zp_m_1(self, dataloader):
model = AutoModelForCausalLM.from_pretrained(self.model_name, torch_dtype="auto", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(self.model_name, trust_remote_code=True)
@@ -175,7 +174,6 @@ class TestAutoRoundMarlinBackend:
@pytest.mark.skip_ci(
reason="Backend/JIT: AWQ Marlin is a format/backend matrix case; standard Marlin smoke remains in PR CI"
)
- @pytest.mark.timeout(90)
def test_gptqmodel_awq_marlin_4bits_sym(self):
"""Test AWQ quantization with gptqmodel:awq_marlin backend (sym-only, float16)."""
model_path = get_model_path("facebook/opt-125m")
diff --git a/test/unit/test_cuda/backends/test_torch_backend_accuracy.py b/test/unit/test_cuda/backends/test_torch_backend_accuracy.py
index 82cf60b72..a28685106 100644
--- a/test/unit/test_cuda/backends/test_torch_backend_accuracy.py
+++ b/test/unit/test_cuda/backends/test_torch_backend_accuracy.py
@@ -45,7 +45,6 @@ class TestTorchBackendAccuracy:
def teardown_class(cls):
shutil.rmtree("runs", ignore_errors=True)
- @pytest.mark.timeout(90)
def test_torch_backend_accuracy_asym_rtn(self, dataloader):
"""RTN-quantized 4-bit asym model meets the accuracy threshold in both fp16 and bf16."""
model = AutoModelForCausalLM.from_pretrained(self.model_name, dtype="auto", trust_remote_code=True)
@@ -65,7 +64,6 @@ class TestTorchBackendAccuracy:
evaluate_accuracy(model, tokenizer, threshold=0.35, batch_size=16, limit=10)
torch.cuda.empty_cache()
- @pytest.mark.timeout(90)
def test_torch_backend_accuracy_sym_rtn(self, dataloader):
"""RTN-quantized 4-bit sym model meets the accuracy threshold."""
model = AutoModelForCausalLM.from_pretrained(self.model_name, dtype="auto", trust_remote_code=True)
@@ -82,7 +80,6 @@ class TestTorchBackendAccuracy:
evaluate_accuracy(model, tokenizer, threshold=0.28, batch_size=32, limit=1000)
torch.cuda.empty_cache()
- @pytest.mark.timeout(90)
def test_torch_backend_accuracy_asym_tuning(self, dataloader):
"""Tuned (iters=1) 4-bit asym model meets the accuracy threshold."""
model = AutoModelForCausalLM.from_pretrained(self.model_name, dtype="auto", trust_remote_code=True)
diff --git a/test/unit/test_cuda/backends/test_triton_backend.py b/test/unit/test_cuda/backends/test_triton_backend.py
index 6d689db5c..c597a17d3 100644
--- a/test/unit/test_cuda/backends/test_triton_backend.py
+++ b/test/unit/test_cuda/backends/test_triton_backend.py
@@ -22,7 +22,6 @@ class TestAutoRoundTritonBackend:
shutil.rmtree("runs", ignore_errors=True)
@require_greater_than_050
- @pytest.mark.timeout(120)
def test_tritonv2_2bits_asym(self, tiny_opt_model_path):
"""A tuned INT2 asymmetric export reloads and generates with Triton.
diff --git a/test/unit/test_cuda/calibration/test_calib_dataset.py b/test/unit/test_cuda/calibration/test_calib_dataset.py
index f3b665031..32ab6c5fd 100644
--- a/test/unit/test_cuda/calibration/test_calib_dataset.py
+++ b/test/unit/test_cuda/calibration/test_calib_dataset.py
@@ -5,7 +5,6 @@ from auto_round.calib_dataset import get_dataloader
class TestLocalCalibDataset:
- @pytest.mark.timeout(150)
def test_combine_dataset(self, tiny_opt_model_path, monkeypatch):
# Subprocess lifecycle is covered by test_calib_dataset_subprocess.py. The
# streaming sources below cannot reuse its cache, so avoid loading each twice.
diff --git a/test/unit/test_cuda/compressors/test_layer_config.py b/test/unit/test_cuda/compressors/test_layer_config.py
index 2cfb748b2..3a41adcfb 100644
--- a/test/unit/test_cuda/compressors/test_layer_config.py
+++ b/test/unit/test_cuda/compressors/test_layer_config.py
@@ -36,7 +36,6 @@ class TestLayerConfigGpu:
yield
shutil.rmtree(self.save_dir, ignore_errors=True)
- @pytest.mark.timeout(180)
def test_mixed_precision_quantize_reload(self, tiny_opt_model_path):
layer_config = {
"model.decoder.layers.0.self_attn.q_proj": {"bits": 8},
@@ -63,7 +62,6 @@ class TestLayerConfigGpu:
out = model(input_ids)
assert out.logits.shape[0] == 1
- @pytest.mark.timeout(180)
def test_layer_config_string_scheme(self, tiny_opt_model_path):
"""layer_config values given as preset scheme strings."""
layer_config = {
@@ -87,7 +85,6 @@ class TestLayerConfigGpu:
model = AutoModelForCausalLM.from_pretrained(quantized_model_path, device_map="cuda:0", trust_remote_code=True)
assert isinstance(model, torch.nn.Module)
- @pytest.mark.timeout(180)
def test_ignore_layers_not_quantized(self, tiny_opt_model_path):
"""A layer listed in ignore_layers must not be quantized."""
ignored = "model.decoder.layers.0.self_attn.q_proj"
@@ -111,7 +108,6 @@ class TestLayerConfigGpu:
q_proj = getattr(q_proj, part)
assert not hasattr(q_proj, "qweight"), "ignored layer was unexpectedly quantized"
- @pytest.mark.timeout(180)
def test_layer_config_bits2_8(self, tiny_opt_model_path):
"""Mixed 2-bit and 8-bit per-layer config."""
layer_config = {
diff --git a/test/unit/test_cuda/export/test_auto_round_format.py b/test/unit/test_cuda/export/test_auto_round_format.py
index 9eea2b23a..69559bc9c 100644
--- a/test/unit/test_cuda/export/test_auto_round_format.py
+++ b/test/unit/test_cuda/export/test_auto_round_format.py
@@ -177,7 +177,6 @@ class TestAutoRound:
eval_generated_prompt(model, tokenizer)
torch.cuda.empty_cache()
- @pytest.mark.timeout(120)
def test_fp8_block_fp8_format(self, tiny_qwen_model_path):
model_name = tiny_qwen_model_path
diff --git a/test/unit/test_cuda/export/test_autoround_int_export.py b/test/unit/test_cuda/export/test_autoround_int_export.py
index cc39e2b0f..dd5767970 100644
--- a/test/unit/test_cuda/export/test_autoround_int_export.py
+++ b/test/unit/test_cuda/export/test_autoround_int_export.py
@@ -54,7 +54,6 @@ class TestAutoroundIntExportGpu:
yield
shutil.rmtree(self.save_dir, ignore_errors=True)
- @pytest.mark.timeout(180)
@pytest.mark.parametrize("bits", [2, 4, 8])
@pytest.mark.parametrize("group_size", [32, 128])
def test_int_export_reload_forward(self, tiny_opt_model_path, bits, group_size):
@@ -78,7 +77,6 @@ class TestAutoroundIntExportGpu:
out = model(input_ids)
assert out.logits.shape[0] == 1
- @pytest.mark.timeout(180)
@pytest.mark.parametrize("group_size", [128])
def test_w8_asym_env_int_export_reload_forward(self, tiny_opt_model_path, group_size, monkeypatch):
"""AR_ALLOW_W8_ASYM=1 skips the native-format refusal; the packed
@@ -105,7 +103,6 @@ class TestAutoroundIntExportGpu:
out = model(input_ids)
assert out.logits.shape[0] == 1
- @pytest.mark.timeout(180)
@pytest.mark.parametrize("bits", [4]) # 8-bit asym is refused at construction
@pytest.mark.parametrize("group_size", [32, 128])
def test_asym_int_export_reload_forward(self, tiny_opt_model_path, bits, group_size):
@@ -130,7 +127,6 @@ class TestAutoroundIntExportGpu:
out = model(input_ids)
assert out.logits.shape[0] == 1
- @pytest.mark.timeout(180)
def test_int_export_group_size_64(self, tiny_opt_model_path):
"""group_size=64 exercises the non-power-of-128 packing path."""
autoround = AutoRound(
diff --git a/test/unit/test_cuda/export/test_fp8_schemes.py b/test/unit/test_cuda/export/test_fp8_schemes.py
index abe0b2a3d..64a077a9d 100644
--- a/test/unit/test_cuda/export/test_fp8_schemes.py
+++ b/test/unit/test_cuda/export/test_fp8_schemes.py
@@ -39,13 +39,11 @@ class TestFp8SchemesGpu:
break
assert found_fp8, "No fp8 weight found in the compressed model"
- @pytest.mark.timeout(180)
def test_fp8_block_export(self, tiny_qwen_model_path):
autoround = AutoRound(tiny_qwen_model_path, scheme="FP8_BLOCK", iters=0, disable_opt_rtn=True, seqlen=2)
compressed_model, _ = autoround.quantize_and_save(output_dir=self.save_dir, format="auto_round")
self._assert_fp8_weight(compressed_model)
- @pytest.mark.timeout(180)
def test_fp8_static_export(self, tiny_qwen_model_path):
autoround = AutoRound(tiny_qwen_model_path, scheme="FP8_STATIC", iters=0, disable_opt_rtn=True, seqlen=2)
compressed_model, quantized_model_path = autoround.quantize_and_save(
@@ -55,7 +53,6 @@ class TestFp8SchemesGpu:
model = AutoModelForCausalLM.from_pretrained(quantized_model_path, device_map="cuda:0", trust_remote_code=True)
assert isinstance(model, torch.nn.Module)
- @pytest.mark.timeout(180)
def test_fp8_block_forward(self, tiny_qwen_model_path):
"""Reloaded FP8_BLOCK model must run a forward pass on GPU."""
autoround = AutoRound(tiny_qwen_model_path, scheme="FP8_BLOCK", iters=0, disable_opt_rtn=True, seqlen=2)
@@ -74,7 +71,6 @@ class TestFp8SchemesGpu:
out = model(input_ids)
assert out.logits.shape[0] == 1
- @pytest.mark.timeout(180)
def test_fp8_plain_export(self, tiny_qwen_model_path):
"""Weight-only FP8 (FPW8A16, per-channel scale) export."""
autoround = AutoRound(
diff --git a/test/unit/test_cuda/export/test_gguf_format.py b/test/unit/test_cuda/export/test_gguf_format.py
index 3cd83f0d8..360f3e90a 100644
--- a/test/unit/test_cuda/export/test_gguf_format.py
+++ b/test/unit/test_cuda/export/test_gguf_format.py
@@ -105,7 +105,6 @@ class TestAutoRound:
@pytest.mark.skipif(not check_version("transformers>=5.2.0"), reason="requires transformers >= 5.2.0")
@pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA")
@require_gguf
- @pytest.mark.timeout(240)
def test_qwen35_moe_gguf(self, tiny_qwen35_moe_model_path, monkeypatch):
monkeypatch.setenv("AR_DISABLE_GGUF_MTP_EXPORT", "1")
self._export_qwen35_moe_gguf(tiny_qwen35_moe_model_path)
@@ -231,7 +230,6 @@ class TestAutoRound:
reason="Architecture: Gemma VLM GGUF export validates mmproj output and requires a large multimodal fixture"
)
@require_gguf
- @pytest.mark.timeout(120)
def test_vlm_gguf(self):
from test.helpers import save_tiny_model
@@ -309,7 +307,6 @@ class TestAutoRound:
reason="Matrix: The Q2_K_S first-eighth FFN-down dtype rule is covered by a dedicated GGUF dtype unit test"
)
@require_gguf
- @pytest.mark.timeout(90)
def test_q2_k_s_ffn_down_q4k(self):
"""Verify blk.0.ffn_down.weight is Q4_K in gguf:q2_k_s format.
Blocks where i_layer < n_layer/8 should use Q4_K instead of Q2_K for ffn_down."""
diff --git a/test/unit/test_cuda/export/test_mxfp_nvfp.py b/test/unit/test_cuda/export/test_mxfp_nvfp.py
index 892be0862..0b658e661 100644
--- a/test/unit/test_cuda/export/test_mxfp_nvfp.py
+++ b/test/unit/test_cuda/export/test_mxfp_nvfp.py
@@ -30,7 +30,6 @@ class TestMxfpNvfpExportGpu:
yield
shutil.rmtree(self.save_dir, ignore_errors=True)
- @pytest.mark.timeout(180)
@pytest.mark.parametrize("scheme", ["MXFP4", "NVFP4", "MXFP8", "MXFP4_RCEIL"])
def test_export_reload_forward(self, tiny_qwen_model_path, scheme):
autoround = AutoRound(
@@ -50,7 +49,6 @@ class TestMxfpNvfpExportGpu:
out = model(input_ids)
assert out.logits.shape[0] == 1
- @pytest.mark.timeout(180)
@pytest.mark.parametrize("scheme", ["MXFP4", "NVFP4"])
def test_export_qlinear_modules(self, tiny_qwen_model_path, scheme):
"""The reloaded model must contain the expected MXFP/NVFP quant-linear module."""
diff --git a/test/unit/test_cuda/export/test_nvfp_autoround_export.py b/test/unit/test_cuda/export/test_nvfp_autoround_export.py
index 79d382e6a..19c0e549d 100644
--- a/test/unit/test_cuda/export/test_nvfp_autoround_export.py
+++ b/test/unit/test_cuda/export/test_nvfp_autoround_export.py
@@ -30,7 +30,6 @@ class TestNvfpAutoroundExport:
yield
shutil.rmtree(self.save_dir, ignore_errors=True)
- @pytest.mark.timeout(180)
def test_nvfp4_autoround_export(self, tiny_opt_model_path):
autoround = AutoRound(
tiny_opt_model_path,
diff --git a/test/unit/test_cuda/models/test_fp8_model.py b/test/unit/test_cuda/models/test_fp8_model.py
index 73f64b223..9d8d53e6f 100644
--- a/test/unit/test_cuda/models/test_fp8_model.py
+++ b/test/unit/test_cuda/models/test_fp8_model.py
@@ -79,7 +79,6 @@ class TestAutoRound:
tokenizer = AutoTokenizer.from_pretrained(quantized_model_path)
generate_prompt(model, tokenizer)
- @pytest.mark.timeout(90)
def test_gguf_imatrix(self, mock_fp8_capable_device, tiny_fp8_qwen_model_path):
ar = AutoRound(tiny_fp8_qwen_model_path, iters=0)
_, quantized_model_path = ar.quantize_and_save(format="gguf:q2_k_s", output_dir=self.save_dir)
@@ -112,7 +111,6 @@ class TestAutoRound:
with patch("torch.cuda.get_device_capability", return_value=DEVICE_CAPABILITY): # revert DEVICE_CAPABILITY
evaluate_accuracy(folder, threshold=0.25)
- @pytest.mark.timeout(300)
def test_small_model_opt_rtn(self, mock_fp8_capable_device):
model_name = get_model_path("Qwen/Qwen3-0.6B-FP8")
ar = AutoRound(model=model_name, iters=0)
diff --git a/test/unit/test_cuda/models/test_moe_model.py b/test/unit/test_cuda/models/test_moe_model.py
index fe3e3e34f..7e9869f4e 100644
--- a/test/unit/test_cuda/models/test_moe_model.py
+++ b/test/unit/test_cuda/models/test_moe_model.py
@@ -17,7 +17,6 @@ def _calibration_data(seqlen):
@pytest.mark.skipif(not check_version("transformers>=5.2.0"), reason="requires transformers >= 5.2.0")
def test_qwen3_5_moe(tiny_qwen35_moe_text_model_path, tmp_path):
from transformers import Qwen3_5MoeForCausalLM
@@ -47,7 +46,6 @@ def test_qwen3_5_moe(tiny_qwen35_moe_text_model_path, tmp_path):
reason="Architecture: Conditional Qwen3.5 fused-MoE validates the large VLM fixture; retain in the full test tier"
)
@pytest.mark.skipif(not check_version("transformers>=5.2.0"), reason="requires transformers >= 5.2.0")
def test_qwen3_5_moe_conditional(tiny_qwen35_moe_model_path, tmp_path):
from transformers import Qwen3_5MoeForConditionalGeneration
diff --git a/test/unit/test_cuda/models/test_omni_model.py b/test/unit/test_cuda/models/test_omni_model.py
index 99559ead3..b4782436e 100644
--- a/test/unit/test_cuda/models/test_omni_model.py
+++ b/test/unit/test_cuda/models/test_omni_model.py
@@ -133,7 +133,6 @@ class TestQwen3OmniMoeQuantization:
yield
shutil.rmtree(self.save_dir, ignore_errors=True)
- @pytest.mark.timeout(180)
def test_quantize_and_reload(self, tiny_opt_model_path, tiny_qwen3_omni_moe_model_path):
"""Quantize, save, reload, and run Qwen3-Omni's text-only thinker path."""
autoround = _make_text_only_autoround(tiny_qwen3_omni_moe_model_path, tiny_opt_model_path)
diff --git a/test/unit/test_cuda/models/test_special_model.py b/test/unit/test_cuda/models/test_special_model.py
index aad08170e..69215b7e3 100644
--- a/test/unit/test_cuda/models/test_special_model.py
+++ b/test/unit/test_cuda/models/test_special_model.py
@@ -38,7 +38,6 @@ class TestSpecialModelHandlerGpu:
layers = get_predefined_ignore_layers(model)
assert any(".gate" in name for name in layers), f"No MoE gate ignore layers found: {layers}"
- @pytest.mark.timeout(240)
def test_moe_quantize_reload(self, tiny_qwen_moe_model_path):
autoround = AutoRound(
tiny_qwen_moe_model_path,
@@ -65,7 +64,6 @@ class TestSpecialModelHandlerGpu:
layers = get_predefined_ignore_layers(model)
assert layers == []
- @pytest.mark.timeout(240)
def test_moe_lm_head_not_quantized(self, tiny_qwen_moe_model_path):
"""lm_head is not in the quant block list and must stay unquantized after export."""
autoround = AutoRound(
@@ -85,7 +83,6 @@ class TestSpecialModelHandlerGpu:
if lm_head is not None:
assert not hasattr(lm_head, "qweight"), "lm_head was unexpectedly quantized"
- @pytest.mark.timeout(240)
def test_moe_quantize_iters_reload(self, tiny_qwen_moe_model_path):
"""MoE quantization with sign-SGD (iters>0) + reload."""
autoround = AutoRound(
diff --git a/test/unit/test_cuda/quantization/test_act_quantization.py b/test/unit/test_cuda/quantization/test_act_quantization.py
index 6143c6377..e1d7e7cbc 100644
--- a/test/unit/test_cuda/quantization/test_act_quantization.py
+++ b/test/unit/test_cuda/quantization/test_act_quantization.py
@@ -41,7 +41,6 @@ class TestActQuantizationGpu:
out = model(input_ids)
assert out.logits.shape[0] == 1
- @pytest.mark.timeout(180)
def test_w4a8_act_dynamic(self, tiny_opt_model_path):
autoround = AutoRound(
tiny_opt_model_path,
@@ -58,7 +57,6 @@ class TestActQuantizationGpu:
_, quantized_model_path = autoround.quantize_and_save(output_dir=self.save_dir, format="auto_round")
self._reload_and_forward(quantized_model_path)
- @pytest.mark.timeout(180)
def test_w8a8_act_dynamic(self, tiny_opt_model_path):
autoround = AutoRound(
tiny_opt_model_path,
@@ -76,7 +74,6 @@ class TestActQuantizationGpu:
_, quantized_model_path = autoround.quantize_and_save(output_dir=self.save_dir, format="auto_round")
self._reload_and_forward(quantized_model_path)
- @pytest.mark.timeout(180)
def test_w4a8_act_static(self, tiny_opt_model_path):
"""Static (non-dynamic) activation quantization path."""
autoround = AutoRound(
@@ -95,7 +92,6 @@ class TestActQuantizationGpu:
_, quantized_model_path = autoround.quantize_and_save(output_dir=self.save_dir, format="auto_round")
self._reload_and_forward(quantized_model_path)
- @pytest.mark.timeout(180)
def test_w4a8_act_group128(self, tiny_opt_model_path):
"""Wider act_group_size exercises a different act-hook grouping path."""
autoround = AutoRound(
@@ -113,7 +109,6 @@ class TestActQuantizationGpu:
_, quantized_model_path = autoround.quantize_and_save(output_dir=self.save_dir, format="auto_round")
self._reload_and_forward(quantized_model_path)
- @pytest.mark.timeout(180)
def test_w8a8_asym(self, tiny_opt_model_path):
"""Asymmetric W8A8 cannot be exported to the auto_round format.
@@ -136,7 +131,6 @@ class TestActQuantizationGpu:
seqlen=16,
)
- @pytest.mark.timeout(180)
def test_w4a8_reload_layer_type_diff_but_output_equal(self, tiny_opt_model_path):
"""Save path keeps wrappers; reload path materializes FakeActQuantLinear, outputs should match."""
old_offload = os.environ.get("AR_DISABLE_OFFLOAD")
diff --git a/test/unit/test_cuda/quantization/test_mxfp_nvfp.py b/test/unit/test_cuda/quantization/test_mxfp_nvfp.py
index e214bf48b..9b1675694 100644
--- a/test/unit/test_cuda/quantization/test_mxfp_nvfp.py
+++ b/test/unit/test_cuda/quantization/test_mxfp_nvfp.py
@@ -28,7 +28,6 @@ QMODULE_MAPPING = {
}
@pytest.mark.parametrize("scheme", testing_schemes)
@torch.inference_mode()
def test_e2e_quant_and_infer(scheme, tiny_qwen_model_path):
diff --git a/test/unit/test_cuda/quantization/test_torch_compile.py b/test/unit/test_cuda/quantization/test_torch_compile.py
index 6ea92ec5f..2110c3400 100644
--- a/test/unit/test_cuda/quantization/test_torch_compile.py
+++ b/test/unit/test_cuda/quantization/test_torch_compile.py
@@ -57,7 +57,6 @@ class TestTorchCompile:
shutil.rmtree(self.save_dir, ignore_errors=True)
@require_gguf
- @pytest.mark.timeout(90)
def test_gguf_q2ks_torch_compile_iters0(self, tiny_qwen_model_path):
"""Test GGUF Q2_K_S with torch.compile and iters=0 (RTN mode).
diff --git a/test/unit/test_cuda/requirements_vlm.txt b/test/unit/test_cuda/requirements_vlm.txt
index f9d195112..3539bf033 100644
--- a/test/unit/test_cuda/requirements_vlm.txt
+++ b/test/unit/test_cuda/requirements_vlm.txt
@@ -1,10 +1,8 @@
# git+https://github.com/haotian-liu/[email protected] --no-deps
# pip install git+https://github.com/deepseek-ai/DeepSeek-VL2.git --no-deps
# pip install -v git+https://github.com/casper-hansen/[email protected] --no-build-isolation
-# pip install flash-attn==2.8.3 --no-build-isolation
bitsandbytes
einops
-flash-attn
gptqmodel
lm-eval>=0.4.2,<0.5
optimum
diff --git a/test/unit/test_cuda/schemes/test_scheme_presets.py b/test/unit/test_cuda/schemes/test_scheme_presets.py
index f175acc24..9c1d50df3 100644
--- a/test/unit/test_cuda/schemes/test_scheme_presets.py
+++ b/test/unit/test_cuda/schemes/test_scheme_presets.py
@@ -30,7 +30,6 @@ class TestSchemePresetsGpu:
yield
shutil.rmtree(self.save_dir, ignore_errors=True)
- @pytest.mark.timeout(180)
@pytest.mark.parametrize("scheme", ["W4A16", "W2A16", "W8A16", "W3A16", "INT8"])
def test_scheme_preset_quantize_forward(self, tiny_opt_model_path, scheme):
autoround = AutoRound(tiny_opt_model_path, scheme=scheme, iters=0, disable_opt_rtn=True, nsamples=1, seqlen=16)
@@ -43,7 +42,6 @@ class TestSchemePresetsGpu:
out = model(input_ids)
assert out.logits.shape[0] == 1
- @pytest.mark.timeout(180)
@pytest.mark.parametrize(
("bits", "expected"),
[(4, 4), (8, 8)],
diff --git a/test/unit/test_cuda/transform/test_mxfp4_transform.py b/test/unit/test_cuda/transform/test_mxfp4_transform.py
index 7bd5b30cf..9c2e7aeb3 100644
--- a/test/unit/test_cuda/transform/test_mxfp4_transform.py
+++ b/test/unit/test_cuda/transform/test_mxfp4_transform.py
@@ -26,7 +26,6 @@ class TestAutoRound:
shutil.rmtree("./saved", ignore_errors=True)
shutil.rmtree("runs", ignore_errors=True)
- @pytest.mark.timeout(120)
def test_transform_mxfp4_quant_infer(self):
model_name = get_model_path("qwen/Qwen3-0.6B")
scheme = "MXFP4"
@@ -49,7 +48,6 @@ class TestAutoRound:
@pytest.mark.skip_ci(
reason="Matrix: Default MXFP4 transform covers CUDA export/reload; MXFP4 tuning remains in the full test tier"
)
- @pytest.mark.timeout(90)
def test_transform_mxfp4_tuning_quant_infer(self):
model_name = get_model_path("qwen/Qwen3-0.6B")
scheme = "MXFP4"
@@ -72,7 +70,6 @@ class TestAutoRound:
@pytest.mark.skip_ci(
reason="Matrix: Random Hadamard behavior has dedicated unit coverage; keep the CUDA MXFP4 E2E smoke deterministic"
)
- @pytest.mark.timeout(90)
def test_random_transform_mxfp4_quant_infer(self):
model_name = get_model_path("qwen/Qwen3-0.6B")
scheme = "MXFP4"
diff --git a/test/unit/test_cuda/transform/test_spinquant.py b/test/unit/test_cuda/transform/test_spinquant.py
index 815602752..1a6d0808f 100644
--- a/test/unit/test_cuda/transform/test_spinquant.py
+++ b/test/unit/test_cuda/transform/test_spinquant.py
@@ -308,7 +308,6 @@ class TestPipelineIntegration:
yield
shutil.rmtree(self.save_dir, ignore_errors=True)
- @pytest.mark.timeout(180)
def test_pipeline_quarot_string(self):
"""AutoRound(alg_configs=["rtn", "quarot"]) should work end-to-end."""
model_name = get_model_path("Qwen/Qwen3-0.6B")
@@ -325,7 +324,6 @@ class TestPipelineIntegration:
tokenizer = AutoTokenizer.from_pretrained(quantized_model_path)
generate_prompt(model, tokenizer)
- @pytest.mark.timeout(90)
def test_pipeline_spinquant_config(self):
"""AutoRound(alg_configs=["rtn", SpinQuantConfig(...)]) should work."""
model_name = get_model_path("Qwen/Qwen3-0.6B")
@@ -353,7 +351,6 @@ class TestPipelineIntegration:
@pytest.mark.skip_ci(
reason="Matrix: The SpinQuantConfig-object pipeline is already covered by test_pipeline_spinquant_config"
)
- @pytest.mark.timeout(90)
def test_pipeline_config_object(self):
"""Building a SpinQuantConfig directly (instead of the dict shorthand) and passing
it via ``alg_configs`` should work."""
diff --git a/test/unit/test_hpu/test_auto_round.py b/test/unit/test_hpu/test_auto_round.py
index fea794177..90850c123 100644
--- a/test/unit/test_hpu/test_auto_round.py
+++ b/test/unit/test_hpu/test_auto_round.py
@@ -31,7 +31,6 @@ def run_opt_125m_on_hpu():
@pytest.mark.skipif(not is_hpex_available(), reason="HPU is not supported")
@pytest.mark.skipif(not is_pytest_mode_lazy(), reason="Only for lazy mode")
def test_opt_125m_lazy_mode():
run_opt_125m_on_hpu()
diff --git a/test/unit/test_hpu/test_quant_fp8.py b/test/unit/test_hpu/test_quant_fp8.py
index fc9737dd7..73091a1ec 100644
--- a/test/unit/test_hpu/test_quant_fp8.py
+++ b/test/unit/test_hpu/test_quant_fp8.py
@@ -31,7 +31,6 @@ class TestAutoRound:
def check_nan_inf_in_tensor(self, tensor, name=""):
return torch.isnan(tensor).any() or torch.isinf(tensor).any()
- @pytest.mark.timeout(120)
@pytest.mark.parametrize("model_name", MODEL_LIST)
def test_small_model_rtn_generation(self, model_name):
ar = AutoRound(model_name, iters=0, scheme="FP8_STATIC", nsamples=16)
diff --git a/test/unit/test_hpu/test_static_attn.py b/test/unit/test_hpu/test_static_attn.py
index 2a06a8a44..6f2c5cacf 100644
--- a/test/unit/test_hpu/test_static_attn.py
+++ b/test/unit/test_hpu/test_static_attn.py
@@ -26,7 +26,6 @@ def setup_deepseekv2():
@pytest.mark.skipif(not is_hpex_available(), reason="HPU is not supported")
@pytest.mark.skipif(not is_pytest_mode_lazy(), reason="Only for lazy mode")
def test_deepseek_v2_on_hpu(setup_deepseekv2):
model, tokenizer, output_dir, config = setup_deepseekv2
autoround = AutoRound(
diff --git a/test/unit/test_xpu/algorithms/test_awq.py b/test/unit/test_xpu/algorithms/test_awq.py
index f4c880f8f..9bbb8cb1a 100644
--- a/test/unit/test_xpu/algorithms/test_awq.py
+++ b/test/unit/test_xpu/algorithms/test_awq.py
@@ -51,7 +51,6 @@ class TestAWQNormalLLM:
assert cfg.awq_seqlen == 128
- @pytest.mark.timeout(90)
def test_awq_w4a16_quantize_and_inference(self, tiny_opt_model_path):
"""W4A16 AWQ quantization produces valid layer_config and the model can generate.
@@ -80,7 +79,6 @@ class TestAWQNormalLLM:
output = generate_prompt(model, tokenizer, device=device)
assert len(output) > 0, "Model should produce non-empty output"
- @pytest.mark.timeout(120)
def test_awq_w4a16_export_default_scheme(self, tiny_opt_model_path):
"""Default W4A16 scheme export: quantization_config has bits=4, group_size=128."""
ar = AutoRound(
@@ -130,7 +128,6 @@ class TestAWQNormalLLM:
assert qconfig["sym"] == sym
assert "auto-round" in qconfig["quant_method"]
- @pytest.mark.timeout(120)
@pytest.mark.parametrize("device", _AVAILABLE_DEVICES)
def test_awq_w4a16_round_trip(self, tiny_opt_model_path, device):
"""Quantize, save, reload on `device`, and generate -- exercises the full save/load round trip.
@@ -163,7 +160,6 @@ class TestAWQNonIntegerSchemes:
MXFP/NVFP scheme.
"""
- @pytest.mark.timeout(60)
@pytest.mark.parametrize("scheme", ["MXFP4", "NVFP4"])
def test_awq_non_integer_scheme_smoke(self, tiny_opt_model_path, scheme):
"""Algorithm/config correctness (bits/act_bits assignment) -- runs once on cpu.
@@ -737,7 +733,6 @@ class TestAWQMoE:
assert kwargs["position_embeddings"][1].shape == (1, 4, 8)
@requires_cuda
- @pytest.mark.timeout(420)
def test_awq_moe_quantized_layers_check(self, tiny_qwen_moe_model_path):
"""AWQ on MoE: expert layers should be quantized, gates/routers stay FP.
@@ -781,7 +776,6 @@ class TestAWQMoE:
# TODO: Investigate and fix the excessive test runtime instead of relying on an increased timeout.
@requires_cuda
- @pytest.mark.timeout(400)
def test_awq_moe_save_quant_config(self, tiny_qwen_moe_model_path):
"""AWQ MoE: saved quantization_config should be consistent and loadable.
diff --git a/test/unit/test_xpu/backends/test_torch_backend.py b/test/unit/test_xpu/backends/test_torch_backend.py
index 5ade5afe2..8665f6176 100644
--- a/test/unit/test_xpu/backends/test_torch_backend.py
+++ b/test/unit/test_xpu/backends/test_torch_backend.py
@@ -55,7 +55,6 @@ class TestTorchBackendFunctional:
@pytest.mark.parametrize("device", _AVAILABLE_DEVICES)
@pytest.mark.parametrize("iters", [0, 1], ids=["rtn", "tuning"])
@pytest.mark.parametrize("sym,group_size,format", _4BIT_CASES)
- @pytest.mark.timeout(120)
def test_torch_backend_4bit(self, tiny_opt_model_path, sym, group_size, format, iters, device):
"""4-bit (a)symmetric model, RTN or tuned, loads via the torch backend and generates text."""
ar = AutoRound(tiny_opt_model_path, bits=4, group_size=group_size, sym=sym, iters=iters, seqlen=2, nsamples=1)
diff --git a/test/unit/test_xpu/test_autoround.py b/test/unit/test_xpu/test_autoround.py
index b12286616..da32e35a6 100644
--- a/test/unit/test_xpu/test_autoround.py
+++ b/test/unit/test_xpu/test_autoround.py
@@ -29,7 +29,6 @@ class TestAutoRoundXPU:
yield
shutil.rmtree(self.save_dir, ignore_errors=True)
- @pytest.mark.timeout(180)
def test_gptq_format(self, dataloader):
model_name = get_model_path("facebook/opt-125m")
model = AutoModelForCausalLM.from_pretrained(
@@ -126,7 +125,6 @@ class TestAutoRoundXPU:
shutil.rmtree(quantized_model_path, ignore_errors=True)
- @pytest.mark.timeout(120)
def test_vlm_model(self, dataloader):
scheme = "W4A16"
model_name = get_model_path("Qwen/Qwen2-VL-2B-Instruct")
@@ -194,7 +192,6 @@ class TestAutoRoundXPU:
)
print(output_text[0])
- @pytest.mark.timeout(90)
def test_quant_lm_head(self, dataloader):
bits, sym, group_size = 4, True, 128
# Note that, to save UT tuning time, the local model is intentionally kept lightweight, using only 2 hidden layers.