Add marks nightly and weakly for torch tests (#2092)

### Changes

- Add marks `nightly` and `weakly` for tests.
- Mark sanity tests as `nightly`
- Split `test_functions.TestParametrized` to fast for precommit and long
for nightly
- Time of torch precommit reduced from 60 to 40 mins
- Set `xfail` for sanity tests with `--mode train` in case of segment
fault.
Sporadic segment fault reproduced on torch>=2.0.0 on call `backward`
function.

### Related tickets

119128

Makefile

@@ -10,6 +10,10 @@ ifdef DATA
 	DATA_ARG := --data $(DATA)
 endif
 
+ifdef WEEKLY_MODELS
+	WEEKLY_MODELS_ARG := --weekly-models $(WEEKLY_MODELS)
+endif
+
 install-pre-commit:
 	pip install pre-commit==3.2.2
 
@@ -124,7 +128,13 @@ install-torch-dev: install-torch-test install-pre-commit install-pylint
 	pip install -r examples/post_training_quantization/torch/ssd300_vgg16/requirements.txt
 
 test-torch:
-	pytest ${COVERAGE_ARGS} tests/common tests/torch --junitxml ${JUNITXML_PATH} $(DATA_ARG)
+	pytest ${COVERAGE_ARGS} tests/common tests/torch -m "not weekly and not nightly" --junitxml ${JUNITXML_PATH} $(DATA_ARG)
+
+test-torch-nightly:
+	pytest ${COVERAGE_ARGS} tests/torch -m nightly --junitxml ${JUNITXML_PATH} $(DATA_ARG)
+
+test-torch-weekly:
+	pytest ${COVERAGE_ARGS} tests/torch -m weekly --junitxml ${JUNITXML_PATH} $(DATA_ARG) ${WEEKLY_MODELS_ARG}
 
 COMMON_PYFILES := $(shell python3 tools/collect_pylint_input_files_for_backend.py common)
 pylint-torch:

tests/torch/nas/test_sanity_sample.py

@@ -96,6 +96,7 @@ def fixture_nas_desc(request, dataset_dir):
     return desc.finalize(dataset_dir)
 
 
+@pytest.mark.nightly
 def test_e2e_supernet_training(nas_desc: NASSampleTestDescriptor, tmp_path, mocker):
     validator = nas_desc.get_validator()
     args = validator.get_default_args(tmp_path)

tests/torch/pytest.ini

@@ -5,5 +5,7 @@ markers =
     convert
     train
     install
+    nightly
+    weekly
 python_files = test_*
 xfail_strict = true

tests/torch/quantization/test_functions.py

@@ -171,16 +171,16 @@ def skip_if_half_on_cpu(is_fp16, use_cuda):
 def check_quant_moved(test_input, test_val, ref_val, quant_len, input_low, input_range, is_fp16, rtol, atol=1e-10):
     """
     Checks values in `test_val` are inside of closest quant and
-    values in `test_val` and `ref_val` elementwise eather equal with given rtol/atol or
+    values in `test_val` and `ref_val` elementwise either equal with given rtol/atol or
     values differ by correspondent `quant_len` +- rtol.
 
     :param test_input: Input of a quantizer.
     :param test_val: Given test value.
     :param ref_val: Given reference value.
-    :param quant_len: Lenghts of quants in quantizers
+    :param quant_len: Length of quants in quantizers
         (for each channel in case per channel quantization).
-    :param atol: Absolute tollerance.
-    :param rtol: Relative tollerance.
+    :param atol: Absolute tolerance.
+    :param rtol: Relative tolerance.
     """
 
     def to_tensor(a):
@@ -214,15 +214,10 @@ def check_outputs_for_quantization_functions(test_val: torch.Tensor, ref_val: np
     PTTensorListComparator.check_equal(test_val, ref_val, rtol, atol)
 
 
-@pytest.mark.parametrize(
-    "input_size",
-    [[1, 48, 112, 112], [1, 96, 28, 28], [1, 288, 14, 14], [16, 96, 112, 112], [16, 192, 28, 28], [16, 576, 14, 14]],
-    ids=idfn,
-)
 @pytest.mark.parametrize("bits", (8, 4), ids=("8bit", "4bit"))
 @pytest.mark.parametrize("scale_mode", ["single_scale", "per_channel_scale"])
 @pytest.mark.parametrize("is_fp16", (True, False), ids=("fp16", "fp32"))
-class TestParametrized:
+class BaseParametrized:
     class TestSymmetric:
         @staticmethod
         def generate_scale(input_size, scale_mode, is_weights, is_fp16, fixed=None):
@@ -523,12 +518,12 @@ def test_quantize_asymmetric_backward(self, _seed, input_size, bits, use_cuda, i
             if is_fp16:
                 # This is needed to make scale == 1 to prevent
                 # quant movement on forward pass in FP16 precision.
-                # In case scale != 1., not precice scale multiplication in FP16
+                # In case scale != 1., not precise scale multiplication in FP16
                 # could lead to big deviations, so even if an input point
                 # lies in safe range (far from middles of quants) after a scaling
                 # it could end up in the middle of a quant. It happens mostly
-                # when target quant > 150 because in real life scenarious quantization range
-                # usualy less than 2 ** quantization bits,
+                # when target quant > 150 because in real life scenarios quantization range
+                # usually less than 2 ** quantization bits,
                 # so input is small and scale is big, small FP16 input multiplies big fp16 scale,
                 # deviation is significant.
                 fixed = {}
@@ -589,6 +584,21 @@ def test_quantize_asymmetric_backward(self, _seed, input_size, bits, use_cuda, i
             check_outputs_for_quantization_functions(test_grads, ref_grads, rtol=1e-2 if is_fp16 else 1e-3)
 
 
+@pytest.mark.parametrize("input_size", [[1, 16, 64, 64], [4, 16, 16, 16]], ids=idfn)
+class TestParametrizedFast(BaseParametrized):
+    pass
+
+
+@pytest.mark.nightly
+@pytest.mark.parametrize(
+    "input_size",
+    [[1, 48, 112, 112], [1, 96, 28, 28], [1, 288, 14, 14], [16, 96, 112, 112], [16, 192, 28, 28], [16, 576, 14, 14]],
+    ids=idfn,
+)
+class TestParametrizedLong(BaseParametrized):
+    pass
+
+
 @pytest.mark.parametrize("device", ["cuda", "cpu"])
 def test_mapping_to_zero(quantization_mode, device):
     torch.manual_seed(42)

tests/torch/quantization/test_sanity_sample.py

@@ -272,6 +272,7 @@ def fixture_precision_desc(request, dataset_dir):
     return desc.finalize(dataset_dir)
 
 
+@pytest.mark.nightly
 def test_precision_init(precision_desc: PrecisionTestCaseDescriptor, tmp_path, mocker):
     validator = precision_desc.get_validator()
     args = validator.get_default_args(tmp_path)
@@ -351,6 +352,7 @@ def fixture_export_desc(request):
     return desc.finalize()
 
 
+@pytest.mark.nightly
 @pytest.mark.parametrize(
     ("extra_args", "is_export_called"),
     (({}, False), ({"-m": ["export", "train"]}, True)),