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

[CUDNN] Remove defunct cuDNN V8 API build flag (#120006)

Browse Source 
The flag basically does nothing following #95722

Let's see if the quantization tests break

CC @malfet @atalmanagement

Pull Request resolved: https://github.com/pytorch/pytorch/pull/120006
Approved by: https://github.com/malfet
This commit is contained in:
eqy
2024-06-03 22:42:02 +00:00
committed by PyTorch MergeBot
parent 0e7bd7fedd
commit ac568fc007
3 changed files with 23 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
cmake/Summary.cmake
View File

@ -71,7 +71,6 @@ function(caffe2_print_configuration_summary)
message(STATUS " Split CUDA : ${BUILD_SPLIT_CUDA}") message(STATUS " Split CUDA : ${BUILD_SPLIT_CUDA}")
message(STATUS " CUDA static link : ${CAFFE2_STATIC_LINK_CUDA}") message(STATUS " CUDA static link : ${CAFFE2_STATIC_LINK_CUDA}")
message(STATUS " USE_CUDNN : ${USE_CUDNN}") message(STATUS " USE_CUDNN : ${USE_CUDNN}")
message(STATUS " USE_EXPERIMENTAL_CUDNN_V8_API: ${USE_EXPERIMENTAL_CUDNN_V8_API}")
message(STATUS " USE_CUSPARSELT : ${USE_CUSPARSELT}") message(STATUS " USE_CUSPARSELT : ${USE_CUSPARSELT}")
message(STATUS " CUDA version : ${CUDA_VERSION}") message(STATUS " CUDA version : ${CUDA_VERSION}")
message(STATUS " USE_FLASH_ATTENTION : ${USE_FLASH_ATTENTION}") message(STATUS " USE_FLASH_ATTENTION : ${USE_FLASH_ATTENTION}")

1
defs.bzl
View File

@ -33,7 +33,6 @@ default_compiler_flags = [
"-DTH_INDEX_BASE=0", "-DTH_INDEX_BASE=0",
"-DMAGMA_V2", "-DMAGMA_V2",
"-DNO_CUDNN_DESTROY_HANDLE", "-DNO_CUDNN_DESTROY_HANDLE",
"-DUSE_EXPERIMENTAL_CUDNN_V8_API", # enable cudnn v8 api
"-DUSE_FBGEMM", "-DUSE_FBGEMM",
"-DUSE_PYTORCH_QNNPACK", "-DUSE_PYTORCH_QNNPACK",
# The dynamically loaded NVRTC trick doesn't work in fbcode, # The dynamically loaded NVRTC trick doesn't work in fbcode,

49
test/quantization/core/test_quantized_op.py
View File

@ -21,6 +21,7 @@ from hypothesis import strategies as st
import torch.testing._internal.hypothesis_utils as hu import torch.testing._internal.hypothesis_utils as hu
hu.assert_deadline_disabled() hu.assert_deadline_disabled()
from torch.testing._internal.common_cuda import SM80OrLater
from torch.testing._internal.common_utils import TestCase from torch.testing._internal.common_utils import TestCase
from torch.testing._internal.common_utils import IS_PPC, TEST_WITH_UBSAN, IS_MACOS, BUILD_WITH_CAFFE2, IS_SANDCASTLE from torch.testing._internal.common_utils import IS_PPC, TEST_WITH_UBSAN, IS_MACOS, BUILD_WITH_CAFFE2, IS_SANDCASTLE
from torch.testing._internal.common_quantization import skipIfNoFBGEMM, skipIfNoQNNPACK, skipIfNoONEDNN from torch.testing._internal.common_quantization import skipIfNoFBGEMM, skipIfNoQNNPACK, skipIfNoONEDNN
@ -31,10 +32,12 @@ from torch.testing._internal.common_quantized import (
qengine_is_onednn, qengine_is_onednn,
) )
from torch.ao.quantization import PerChannelMinMaxObserver from torch.ao.quantization import PerChannelMinMaxObserver
from torch.testing._internal.common_cuda import TEST_CUDNN, TEST_CUDA from torch.testing._internal.common_cuda import TEST_CUDNN, TEST_CUDNN_VERSION, TEST_CUDA
from torch.testing._internal.optests import opcheck from torch.testing._internal.optests import opcheck
import torch.backends.xnnpack import torch.backends.xnnpack
from torch.utils.cpp_extension import ROCM_HOME
from typing import Optional from typing import Optional
np_dtype = { np_dtype = {
@ -43,6 +46,8 @@ np_dtype = {
torch.qint32 : np.int32 torch.qint32 : np.int32
} }
TEST_ROCM = TEST_CUDA and torch.version.hip is not None and ROCM_HOME is not None
class PointwisePostOp(NamedTuple): class PointwisePostOp(NamedTuple):
binary_attr : str = "none" binary_attr : str = "none"
alpha : float = 1.0 alpha : float = 1.0
@ -905,9 +910,8 @@ class TestQuantizedOps(TestCase):
"""Tests the correctness of the cudnn add and add_relu op """Tests the correctness of the cudnn add and add_relu op
(Similar to test_qadd_relu_different_qparams, will probably merge in the future)""" (Similar to test_qadd_relu_different_qparams, will probably merge in the future)"""
@unittest.skipIf(not TEST_CUDNN, "cudnn is not enabled.") @unittest.skipIf(not TEST_CUDNN, "cudnn is not enabled.")
@unittest.skip("Local only - currently the test_qadd_relu_cudnn op is bulid " @unittest.skipIf(not SM80OrLater, "requires sm80 or later.")
"with USE_EXPERIMENTAL_CUDNN_V8_API, we can enable the test " @unittest.skipIf(TEST_ROCM, "not supported on rocm.")
"after it is built by default")
def test_qadd_relu_cudnn(self): def test_qadd_relu_cudnn(self):
dtype = torch.qint8 dtype = torch.qint8
add_relu = torch.ops.quantized.add_relu add_relu = torch.ops.quantized.add_relu
@ -940,9 +944,8 @@ class TestQuantizedOps(TestCase):
"""Tests the correctness of the cudnn add and add_relu op for nhwc format""" """Tests the correctness of the cudnn add and add_relu op for nhwc format"""
@unittest.skipIf(not TEST_CUDNN, "cudnn is not enabled.") @unittest.skipIf(not TEST_CUDNN, "cudnn is not enabled.")
@unittest.skip("Local only - currently the test_qadd_relu_cudnn_nhwc op is bulid " @unittest.skipIf(not SM80OrLater, "requires sm80 or later.")
"with USE_EXPERIMENTAL_CUDNN_V8_API, we can enable the test " @unittest.skipIf(TEST_ROCM, "not supported on rocm.")
"after it is built by default")
def test_qadd_relu_cudnn_nhwc(self): def test_qadd_relu_cudnn_nhwc(self):
dtype = torch.qint8 dtype = torch.qint8
add_relu = torch.ops.quantized.add_relu add_relu = torch.ops.quantized.add_relu
@ -1379,7 +1382,7 @@ class TestQuantizedOps(TestCase):
self.assertEqual(a_ref, a_hat.dequantize(), self.assertEqual(a_ref, a_hat.dequantize(),
msg="ops.quantized.max_pool1d results are off") msg="ops.quantized.max_pool1d results are off")
# TODO: merge this test with test_max_pool2d when USE_EXPERIMENTAL_CUDNN_V8_API flag is enabled in CI # TODO: merge this test with test_max_pool2d
"""Tests 2D cudnn max pool operation on quantized tensors.""" """Tests 2D cudnn max pool operation on quantized tensors."""
@given(X=hu.tensor(shapes=hu.array_shapes(min_dims=3, max_dims=4, @given(X=hu.tensor(shapes=hu.array_shapes(min_dims=3, max_dims=4,
min_side=1, max_side=10), min_side=1, max_side=10),
@ -1394,9 +1397,8 @@ class TestQuantizedOps(TestCase):
padding=st.integers(0, 2), padding=st.integers(0, 2),
ceil_mode=st.booleans()) ceil_mode=st.booleans())
@unittest.skipIf(not TEST_CUDNN, "cudnn is not enabled.") @unittest.skipIf(not TEST_CUDNN, "cudnn is not enabled.")
@unittest.skip("Local only - currently the qconv2d_cudnn op is bulid " @unittest.skipIf(TEST_CUDNN_VERSION <= 90100, "cuDNN maxpool2d mishandles -128 before v90100")
"with USE_EXPERIMENTAL_CUDNN_V8_API, we can enable the test " @unittest.skipIf(TEST_ROCM, "not supported on rocm.")
"after it is built by default")
def test_max_pool2d_cudnn(self, X, kernel, stride, dilation, padding, ceil_mode): def test_max_pool2d_cudnn(self, X, kernel, stride, dilation, padding, ceil_mode):
X, (scale, zero_point, torch_type) = X X, (scale, zero_point, torch_type) = X
assume(kernel // 2 >= padding) # Kernel cannot be overhanging! assume(kernel // 2 >= padding) # Kernel cannot be overhanging!
@ -4050,9 +4052,8 @@ class TestQuantizedLinear(TestCase):
use_channelwise=st.sampled_from([False])) # channelwise currently not supported for qlinear cudnn use_channelwise=st.sampled_from([False])) # channelwise currently not supported for qlinear cudnn
@skipIfNoFBGEMM @skipIfNoFBGEMM
@unittest.skipIf(not TEST_CUDNN, "cudnn is not enabled.") @unittest.skipIf(not TEST_CUDNN, "cudnn is not enabled.")
@unittest.skip("Local only - currently the qlinear_cudnn op is bulid " @unittest.skipIf(not SM80OrLater, "requires sm80 or later.")
"with USE_EXPERIMENTAL_CUDNN_V8_API, we can enable the test " @unittest.skipIf(TEST_ROCM, "not supported on rocm.")
"after it is built by default")
# TODO: check with yang regarding CUDNN flags # TODO: check with yang regarding CUDNN flags
def test_qlinear_cudnn(self, batch_size, input_channels, output_channels, use_bias, def test_qlinear_cudnn(self, batch_size, input_channels, output_channels, use_bias,
use_relu, use_multi_dim_input, use_channelwise): use_relu, use_multi_dim_input, use_channelwise):
@ -5427,9 +5428,8 @@ class TestQuantizedConv(TestCase):
use_channelwise=st.sampled_from([False])) use_channelwise=st.sampled_from([False]))
@skipIfNoFBGEMM @skipIfNoFBGEMM
@unittest.skipIf(not TEST_CUDNN, "cudnn is not enabled.") @unittest.skipIf(not TEST_CUDNN, "cudnn is not enabled.")
@unittest.skip("Local only - currently the qconv2d_cudnn op is bulid " @unittest.skipIf(not SM80OrLater, "requires sm80 or later.")
"with USE_EXPERIMENTAL_CUDNN_V8_API, we can enable the test " @unittest.skipIf(TEST_ROCM, "not supported on rocm.")
"after it is built by default")
def test_qconv2d_cudnn( def test_qconv2d_cudnn(
self, self,
batch_size, batch_size,
@ -5510,9 +5510,8 @@ class TestQuantizedConv(TestCase):
use_channelwise=st.sampled_from([False])) use_channelwise=st.sampled_from([False]))
@skipIfNoFBGEMM @skipIfNoFBGEMM
@unittest.skipIf(not TEST_CUDNN, "cudnn is not enabled.") @unittest.skipIf(not TEST_CUDNN, "cudnn is not enabled.")
@unittest.skip("Local only - currently the qconv2d_cudnn op is bulid " @unittest.skipIf(not SM80OrLater, "requires sm80 or later.")
"with USE_EXPERIMENTAL_CUDNN_V8_API, we can enable the test " @unittest.skipIf(TEST_ROCM, "not supported on rocm.")
"after it is built by default")
def test_qconv2d_relu_cudnn( def test_qconv2d_relu_cudnn(
self, self,
batch_size, batch_size,
@ -6245,9 +6244,8 @@ class TestQuantizedConv(TestCase):
use_channelwise=st.sampled_from([False])) use_channelwise=st.sampled_from([False]))
@skipIfNoFBGEMM @skipIfNoFBGEMM
@unittest.skipIf(not TEST_CUDNN, "cudnn is not enabled.") @unittest.skipIf(not TEST_CUDNN, "cudnn is not enabled.")
@unittest.skip("Local only - currently the qconv1d_cudnn op is bulid " @unittest.skipIf(not SM80OrLater, "requires sm80 or later.")
"with USE_EXPERIMENTAL_CUDNN_V8_API, we can enable the test " @unittest.skipIf(TEST_ROCM, "not supported on rocm.")
"after it is built by default")
def test_qconv1d_cudnn( def test_qconv1d_cudnn(
self, self,
batch_size, batch_size,
@ -6319,9 +6317,8 @@ class TestQuantizedConv(TestCase):
use_channelwise=st.sampled_from([False])) use_channelwise=st.sampled_from([False]))
@skipIfNoFBGEMM @skipIfNoFBGEMM
@unittest.skipIf(not TEST_CUDNN, "cudnn is not enabled.") @unittest.skipIf(not TEST_CUDNN, "cudnn is not enabled.")
@unittest.skip("Local only - currently the qconv1d_cudnn op is bulid " @unittest.skipIf(not SM80OrLater, "requires sm80 or later.")
"with USE_EXPERIMENTAL_CUDNN_V8_API, we can enable the test " @unittest.skipIf(TEST_ROCM, "not supported on rocm.")
"after it is built by default")
def test_qconv1d_relu_cudnn( def test_qconv1d_relu_cudnn(
self, self,
batch_size, batch_size,