frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-20 21:14:14 +08:00
Code Issues Packages Projects Releases Wiki Activity

Enable FBGEMM tests under UBSAN as well (#23570)

Browse Source 
Summary:
Enabling tests under UBSAN as well
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23570

Test Plan:
buck test mode/dev caffe2/test:quantized
```
Running 29 tests
Started new test run: https://our.intern.facebook.com/intern/testinfra/testrun/3940649677415136
      ✓ caffe2/test:quantized - test_qtensor (test_quantized_tensor.TestQuantizedTensor) 0.536 1/29 (passed)
      ✓ caffe2/test:quantized - test_qtensor_per_channel_affine (test_quantized_tensor.TestQuantizedTensor) 0.453 2/29 (passed)
      ✓ caffe2/test:quantized - test_qtensor_reshape (test_quantized_tensor.TestQuantizedTensor) 0.302 3/29 (passed)
      ✓ caffe2/test:quantized - test_qadd_relu_same_qparams (test_quantized.TestQuantizedOps) 0.332 4/29 (passed)
      ✓ caffe2/test:quantized - test_qtensor_view (test_quantized_tensor.TestQuantizedTensor) 0.351 5/29 (passed)
      ✓ caffe2/test:quantized - test_qadd_relu_different_qparams (test_quantized.TestQuantizedOps) 0.348 6/29 (passed)
      ✓ caffe2/test:quantized - test_qtensor_dequantize_linear (test_quantized_tensor.TestQuantizedTensor) 0.338 7/29 (passed)
      ✓ caffe2/test:quantized - test_qtensor_copy (test_quantized_tensor.TestQuantizedTensor) 0.267 8/29 (passed)
      ✓ caffe2/test:quantized - test_qtensor_clone (test_quantized_tensor.TestQuantizedTensor) 0.330 9/29 (passed)
      ✓ caffe2/test:quantized - test_qrelu (test_quantized.TestQuantizedOps) 1.774 10/29 (passed)
      ✓ caffe2/test:quantized - test_pool_api (test_nn_quantized.ModuleAPITest) 0.418 11/29 (passed)
      ✓ caffe2/test:quantized - test_qtensor_load_save (test_quantized_tensor.TestQuantizedTensor) 0.724 12/29 (passed)
      ✓ caffe2/test:quantized - test_relu_api (test_nn_quantized.FunctionalAPITest) 1.013 13/29 (passed)
      ✓ caffe2/test:quantized - test_qtensor_quant_dequant (test_quantized_tensor.TestQuantizedTensor) 1.055 14/29 (passed)
      ✓ caffe2/test:quantized - test_qtensor_permute (test_quantized_tensor.TestQuantizedTensor) 0.696 15/29 (passed)
      ✓ caffe2/test:quantized - test_qtensor_dtypes (test_quantized_tensor.TestQuantizedTensor) 0.841 16/29 (passed)
      ✓ caffe2/test:quantized - test_quant_dequant_api (test_nn_quantized.ModuleAPITest) 0.616 17/29 (passed)
      ✓ caffe2/test:quantized - test_qtensor_creation (test_quantized_tensor.TestQuantizedTensor) 0.698 18/29 (passed)
      ✓ caffe2/test:quantized - test_qconv (test_quantized.TestQuantizedConv) 4.743 19/29 (passed)
      ✓ caffe2/test:quantized - test_cat (test_quantized.TestQuantizedOps) 6.992 20/29 (passed)
      ✓ caffe2/test:quantized - test_linear_api (test_nn_quantized.ModuleAPITest) 8.970 21/29 (passed)
      ✓ caffe2/test:quantized - test_conv_api (test_quantized_conv.QuantizedConvTest) 9.403 22/29 (passed)
      ↷ caffe2/test:quantized - test_qnnpack_linear (test_quantized.TestQNNPackOps) 0.000 23/29 (skipped)
Test output:
> Skipped: QNNPACK does not play well with UBSAN at the moment, so we skip the test if we are in a UBSAN environment.
> test_qnnpack_linear (test_quantized.TestQNNPackOps) ... skipped 'QNNPACK does not play well with UBSAN at the moment, so we skip the test if we are in a UBSAN environment.'
>
> ----------------------------------------------------------------------
> Ran 1 test in 0.000s
>
> OK (skipped=1)
      ↷ caffe2/test:quantized - test_qnnpack_relu (test_quantized.TestQNNPackOps) 0.000 24/29 (skipped)
Test output:
> Skipped: QNNPACK does not play well with UBSAN at the moment, so we skip the test if we are in a UBSAN environment.
> test_qnnpack_relu (test_quantized.TestQNNPackOps) ... skipped 'QNNPACK does not play well with UBSAN at the moment, so we skip the test if we are in a UBSAN environment.'
>
> ----------------------------------------------------------------------
> Ran 1 test in 0.000s
>
> OK (skipped=1)
      ✓ caffe2/test:quantized - test_max_pool2d (test_quantized.TestQuantizedOps) 8.453 25/29 (passed)
      ✓ caffe2/test:quantized - test_qlinear_unpack (test_quantized.TestQuantizedLinear) 0.664 26/29 (passed)
      ✓ caffe2/test:quantized - test_qconv_unpack (test_quantized.TestQuantizedConv) 2.965 27/29 (passed)
      ✓ caffe2/test:quantized - test_qlinear (test_quantized.TestQuantizedLinear) 1.915 28/29 (passed)
      ✓ caffe2/test:quantized - test_conv_api (test_nn_quantized.ModuleAPITest) 60.804 29/29 (passed)
      ✓ caffe2/test:quantized - main 0.000 (passed)
Finished test run: https://our.intern.facebook.com/intern/testinfra/testrun/3940649677415136
Summary (total time 68.66s):
  PASS: 28
  FAIL: 0
  SKIP: 2
    caffe2/test:quantized - test_qnnpack_linear (test_quantized.TestQNNPackOps)
    caffe2/test:quantized - test_qnnpack_relu (test_quantized.TestQNNPackOps)
  FATAL: 0
  TIMEOUT: 0
  OMIT: 0
```

Reviewed By: jianyuh

Differential Revision: D16569166

Pulled By: dskhudia

fbshipit-source-id: 53522b4162eb1ebb35b408a1503d9664305c85b0
This commit is contained in:
Daya Khudia
2019-08-12 17:56:09 -07:00
committed by Facebook Github Bot
parent 71fd30e33b
commit f510409281
6 changed files with 35 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
test/test_jit.py
View File

@ -6586,10 +6586,9 @@ a")
a = A()
self.assertEqual(a.with_docstring.__doc__, 'test str')
@unittest.skipIf(TEST_WITH_UBSAN or not torch.fbgemm_is_cpu_supported(),
'Quantized RNN requires FBGEMM. FBGEMM does not play'
' well with UBSAN at the moment, so we skip the test if'
' we are in a UBSAN environment.')
@unittest.skipIf(not torch.fbgemm_is_cpu_supported(),
'Quantized RNN requires FBGEMM. FBGEMM is only optimized for CPUs'
' with instruction set support avx2 or newer.')
def test_rnn_cell_quantized(self):
d_in, d_hid = 2, 2
@ -6681,10 +6680,9 @@ a")
for out, ref_out in zip(outs, ref_outs):
torch.testing.assert_allclose(out, ref_out)
@unittest.skipIf(TEST_WITH_UBSAN or not torch.fbgemm_is_cpu_supported(),
'Quantized RNN requires FBGEMM. FBGEMM does not play'
' well with UBSAN at the moment, so we skip the test if'
' we are in a UBSAN environment.')
@unittest.skipIf(not torch.fbgemm_is_cpu_supported(),
'Quantized RNN requires FBGEMM. FBGEMM is only optimized for CPUs'
' with instruction set support avx2 or newer.')
def test_rnn_quantized(self):
d_in, d_hid = 2, 2
@ -11450,9 +11448,7 @@ a")
traced = torch.jit.trace(foo, torch.rand(3, 4), check_inputs=[(torch.rand(3, 4),)])
# These tests don't work because UBSAN has a false positive about accessing
# out of bounds on a dynamically sized struct internal to asmjit
if not TEST_WITH_UBSAN and torch.fbgemm_is_cpu_supported():
if torch.fbgemm_is_cpu_supported():
def test_quantization_modules(self):
K1, N1 = 2, 2
@ -14872,7 +14868,7 @@ class TestEndToEndHybridFrontendModels(JitTestCase):
def test_snli(self):
self._test_snli(self, device='cpu')
if not TEST_WITH_UBSAN and torch.fbgemm_is_cpu_supported():
if torch.fbgemm_is_cpu_supported():
def test_snli_quantized(self):
self._test_snli(self, device='cpu', quantized=True)
@ -15014,7 +15010,7 @@ class TestEndToEndHybridFrontendModels(JitTestCase):
def test_vae(self):
self._test_vae(self, device='cpu')
if not TEST_WITH_UBSAN and torch.fbgemm_is_cpu_supported():
if torch.fbgemm_is_cpu_supported():
def test_vae_quantized(self):
self._test_vae(self, device='cpu', quantized=True)

8
test/test_nn.py
View File

@ -36,7 +36,6 @@ from common_nn import NNTestCase, ModuleTest, CriterionTest, TestBase, \
module_tests, criterion_tests, loss_reference_fns, get_reduction, \
get_weight, smoothl1loss_reference, kldivloss_reference, \
ctcloss_reference, new_module_tests
from common_utils import TEST_WITH_UBSAN
from torch.nn import MultiheadAttention
@ -2594,10 +2593,9 @@ class TestNN(NNTestCase):
def test_softmax_backward_cuda(self):
self._test_softmax_backward(torch.device('cuda'))
@unittest.skipIf(TEST_WITH_UBSAN or not torch.fbgemm_is_cpu_supported(),
'Linear_FP16_weight requires FBGEMM. FBGEMM does not play'
' well with UBSAN at the moment, so we skip the test if'
' we are in a UBSAN environment.')
@unittest.skipIf(not torch.fbgemm_is_cpu_supported(),
'Linear_FP16_weight requires FBGEMM. FBGEMM is only optimized for CPUs'
' with instruction set support avx2 or newer.')
def test_fb_fc_packed(self):
X = np.random.rand(16, 16).astype(np.float32) - 0.5
W = np.random.rand(16, 16).astype(np.float32) - 0.5

20
test/test_quantization.py
View File

@ -6,7 +6,7 @@ import torch.nn.quantized as nnq
from torch.quantization import \
quantize, prepare, convert, prepare_qat, quantize_qat, fuse_modules
from common_utils import run_tests, TEST_WITH_UBSAN
from common_utils import run_tests
from common_quantization import QuantizationTestCase, SingleLayerLinearModel, \
SkipQuantModel, QuantStubModel, \
ModForFusion, ManualLinearQATModel, ManualConvLinearQATModel, test_only_eval_fn, test_only_train_fn
@ -14,10 +14,11 @@ from common_quantization import QuantizationTestCase, SingleLayerLinearModel, \
from common_quantization import AnnotatedTwoLayerLinearModel, AnnotatedNestedModel, \
AnnotatedSubNestedModel, AnnotatedCustomConfigNestedModel
@unittest.skipIf(TEST_WITH_UBSAN or not torch.fbgemm_is_cpu_supported(),
'Quantization requires FBGEMM. FBGEMM does not play'
' well with UBSAN at the moment, so we skip the test if'
' we are in a UBSAN environment.')
@unittest.skipIf(
not torch.fbgemm_is_cpu_supported(),
" Quantized operations require FBGEMM. FBGEMM is only optimized for CPUs"
" with instruction set support avx2 or newer.",
)
class PostTrainingQuantTest(QuantizationTestCase):
def test_single_layer(self):
r"""Quantize SingleLayerLinearModel which has one Linear module, make sure it is swapped
@ -240,10 +241,11 @@ class PostTrainingQuantTest(QuantizationTestCase):
model = quantize(QuantStubModel(), test_only_eval_fn, self.calib_data)
checkQuantized(model)
@unittest.skipIf(TEST_WITH_UBSAN or not torch.fbgemm_is_cpu_supported(),
'Quantization requires FBGEMM. FBGEMM does not play'
' well with UBSAN at the moment, so we skip the test if'
' we are in a UBSAN environment.')
@unittest.skipIf(
not torch.fbgemm_is_cpu_supported(),
" Quantized operations require FBGEMM. FBGEMM is only optimized for CPUs"
" with instruction set support avx2 or newer.",
)
class QuantizationAwareTrainingTest(QuantizationTestCase):
def test_manual(self):
model = ManualLinearQATModel()

14
test/test_quantized.py
View File

@ -312,10 +312,9 @@ class TestQuantizedOps(TestCase):
zero_point=zero_point)
@unittest.skipIf(
TEST_WITH_UBSAN or not torch.fbgemm_is_cpu_supported(),
" Quantized Linear requires FBGEMM. FBGEMM does not play"
" well with UBSAN at the moment, so we skip the test if"
" we are in a UBSAN environment.",
not torch.fbgemm_is_cpu_supported(),
" Quantized operations require FBGEMM. FBGEMM is only optimized for CPUs"
" with instruction set support avx2 or newer.",
)
class TestQuantizedLinear(unittest.TestCase):
"""Tests the correctness of the quantized linear and linear_relu op."""
@ -432,10 +431,9 @@ class TestQuantizedLinear(unittest.TestCase):
np.testing.assert_equal(W_q.q_zero_point(), W_q_origin.q_zero_point())
@unittest.skipIf(
TEST_WITH_UBSAN or not torch.fbgemm_is_cpu_supported(),
" Quantized convolution requires FBGEMM. FBGEMM does not play"
" well with UBSAN at the moment, so we skip the test if"
" we are in a UBSAN environment.",
not torch.fbgemm_is_cpu_supported(),
" Quantized operations require FBGEMM. FBGEMM is only optimized for CPUs"
" with instruction set support avx2 or newer.",
)
class TestQuantizedConv(unittest.TestCase):
"""Tests the correctness of quantized convolution op."""

11
test/test_quantized_conv.py
View File

@ -13,12 +13,13 @@ from hypothesis import strategies as st
import hypothesis_utils as hu
from common_quantized import _conv_output_shape
from common_utils import TestCase, run_tests, TEST_WITH_UBSAN
from common_utils import TestCase, run_tests
@unittest.skipIf(TEST_WITH_UBSAN or not torch.fbgemm_is_cpu_supported(),
'Quantization requires FBGEMM. FBGEMM does not play'
' well with UBSAN at the moment, so we skip the test if'
' we are in a UBSAN environment.')
@unittest.skipIf(
not torch.fbgemm_is_cpu_supported(),
" Quantized operations require FBGEMM. FBGEMM is only optimized for CPUs"
" with instruction set support avx2 or newer.",
)
class QuantizedConvTest(TestCase):
@given(X=hu.tensor_conv2d(min_batch=1, max_batch=3,
min_in_channels=1, max_in_channels=7,

1
ubsan.supp
View File

@ -1,3 +1,2 @@
vptr:libtorch.so
vptr:libcaffe2.so
bounds:asmjit::Zone::_alloc