feat!: Updating versions of CUDA, cuDNN, TensorRT and PyTorch

BREAKING CHANGE: PyTorch version has been bumped to 1.8.0 Default CUDA version is CUDA 11.1 TensorRT version is TensorRT 7.2.3.4 cuDNN version is now cuDNN 8.1 Signed-off-by: Naren Dasan <[email protected]> Signed-off-by: Naren Dasan <[email protected]>
pytorch · Mar 6, 2021 · 71c4dcb · 71c4dcb
1 parent 6bb9fbf
commit 71c4dcb
Show file tree

Hide file tree

Showing 14 changed files with 75 additions and 75 deletions.
diff --git a/README.md b/README.md
@@ -73,11 +73,11 @@ torch.jit.save(trt_ts_module, "trt_torchscript_module.ts")
 ### Dependencies
 These are the following dependencies used to verify the testcases. TRTorch can work with other versions, but the tests are not guaranteed to pass.
 
-- Bazel 3.7.0
-- Libtorch 1.7.x (built with CUDA 11.0)
-- CUDA 11.0 (10.2 on Jetson)
-- cuDNN 8
-- TensorRT 7.2
+- Bazel 4.0.0
+- Libtorch 1.8.0 (built with CUDA 11.1)
+- CUDA 11.1 (10.2 on Jetson)
+- cuDNN 8.1
+- TensorRT 7.2.3
 
 ## Prebuilt Binaries and Wheel files
 

diff --git a/WORKSPACE b/WORKSPACE
@@ -35,7 +35,7 @@ git_repository(
 # CUDA should be installed on the system locally
 new_local_repository(
     name = "cuda",
-    path = "/usr/local/cuda-11.0/",
+    path = "/usr/local/cuda-11.1/",
     build_file = "@//third_party/cuda:BUILD",
 )
 
@@ -53,16 +53,16 @@ http_archive(
     name = "libtorch",
     build_file = "@//third_party/libtorch:BUILD",
     strip_prefix = "libtorch",
-    sha256 = "117f6dd65b7267839197397edd0b10fd2900b0f291e3e54b0b800caefc31bcb6",
-    urls = ["https://download.pytorch.org/libtorch/cu110/libtorch-cxx11-abi-shared-with-deps-1.7.1%2Bcu110.zip"],
+    sha256 = "62a2c06761c32576b30f5884240cf675b937945d929e4b13cc776de8d9c2236c",
+    urls = ["https://download.pytorch.org/libtorch/cu111/libtorch-cxx11-abi-shared-with-deps-1.8.0%2Bcu111.zip"],
 )
 
 http_archive(
     name = "libtorch_pre_cxx11_abi",
     build_file = "@//third_party/libtorch:BUILD",
     strip_prefix = "libtorch",
-    sha256 = "c77f926afd55d7e860ec9c7abc992c25be77c89771c3ec6fcc13ea42f07d46df",
-    urls = ["https://download.pytorch.org/libtorch/cu110/libtorch-shared-with-deps-1.7.1%2Bcu110.zip"],
+    sha256 = "1c8b0c0883dd17f5ce952d42ec5f7f0cc7ceb370307535cee26a66c10419f1f6",
+    urls = ["https://download.pytorch.org/libtorch/cu111/libtorch-shared-with-deps-1.8.0%2Bcu111.zip"],
 )
 
 # Download these tarballs manually from the NVIDIA website
@@ -71,18 +71,18 @@ http_archive(
 
 http_archive(
     name = "cudnn",
-    urls = ["https://developer.nvidia.com/compute/machine-learning/cudnn/secure/8.0.5/11.0_20201106/cudnn-11.0-linux-x64-v8.0.5.39.tgz",],
+    urls = ["https://developer.nvidia.com/compute/machine-learning/cudnn/secure/8.1.1.33/11.2_20210301/cudnn-11.2-linux-x64-v8.1.1.33.tgz",],
     build_file = "@//third_party/cudnn/archive:BUILD",
-    sha256 = "4e16ee7895deb4a8b1c194b812ba49586ef7d26902051401d3717511898a9b73",
+    sha256 = "98a8784e92862f20018d20c281b30d4a0cd951f93694f6433ccf4ae9c502ba6a",
     strip_prefix = "cuda"
 )
 
 http_archive(
     name = "tensorrt",
-    urls = ["https://developer.nvidia.com/compute/machine-learning/tensorrt/secure/7.2.2/tars/TensorRT-7.2.2.3.Ubuntu-18.04.x86_64-gnu.cuda-11.0.cudnn8.0.tar.gz",],
+    urls = ["https://developer.nvidia.com/compute/machine-learning/tensorrt/secure/7.2.3/tars/TensorRT-7.2.3.4.Ubuntu-18.04.x86_64-gnu.cuda-11.1.cudnn8.1.tar.gz",],
     build_file = "@//third_party/tensorrt/archive:BUILD",
-    strip_prefix = "TensorRT-7.2.2.3",
-    sha256 = "b5c325e38e1d92ce1ce92ca8b54ede9c224bf128c9a53eb0b9022f1ee4313ee0"
+    strip_prefix = "TensorRT-7.2.3.4",
+    sha256 = "d3a1f478e304b48878604fac70ce7920fece71f9cac62f925c9c59c197f5d087"
 )
 
 ####################################################################################

diff --git a/core/util/trt_util.cpp b/core/util/trt_util.cpp
@@ -295,7 +295,7 @@ nvinfer1::DataType toTRTDataType(at::ScalarType t) {
 }
 
 c10::optional<nvinfer1::DataType> toTRTDataType(caffe2::TypeMeta dtype) {
-  if (auto t = c10::tryTypeMetaToScalarType(dtype)) {
+  if (auto t = c10::optTypeMetaToScalarType(dtype)) {
     return toTRTDataType(t.value());
   } else {
     return {};

diff --git a/cpp/api/include/trtorch/trtorch.h b/cpp/api/include/trtorch/trtorch.h
@@ -23,7 +23,7 @@ struct Module;
 } // namespace torch
 
 namespace c10 {
-enum class DeviceType : int16_t;
+enum class DeviceType : int8_t;
 enum class ScalarType : int8_t;
 template <class>
 class ArrayRef;

diff --git a/py/requirements.txt b/py/requirements.txt
@@ -1 +1 @@
-torch==1.7.1
+torch==1.8.0
diff --git a/py/setup.py b/py/setup.py
@@ -204,7 +204,7 @@ def run(self):
       long_description=long_description,
       ext_modules=ext_modules,
       install_requires=[
-          'torch>=1.7.0,<1.8.0',
+          'torch>=1.8.0+cu111,<1.9.0',
       ],
       setup_requires=[],
       cmdclass={

diff --git a/tests/core/conversion/converters/converter_test.bzl b/tests/core/conversion/converters/converter_test.bzl
@@ -12,5 +12,5 @@ def converter_test(name, visibility=None):
             ":use_pre_cxx11_abi":  ["@libtorch_pre_cxx11_abi//:libtorch"],
             "//conditions:default":  ["@libtorch//:libtorch"],
         }),
-        timeout="short"
+        timeout="moderate"
     )
diff --git a/tests/core/conversion/converters/test_activation.cpp b/tests/core/conversion/converters/test_activation.cpp
@@ -66,28 +66,28 @@ TEST(Converters, ATenTanhConvertsCorrectly) {
 
 // TODO: Seems like the IR parser is not handling negative numbers well, need to
 // follow up with the PyTorch Team
-// TEST(Converters, ATenHardTanhConvertsCorrectly) {
-//     const auto graph = R"IR(
-//       graph(%0 : Tensor):
-//         %1 : float = prim::Constant[value=-1.0]()
-//         %2 : float = prim::Constant[value=1.0]()
-//         %3 : Tensor = aten::hardtanh(%0, %1, %2)
-//         return (%3))IR";
+TEST(Converters, ATenHardTanhConvertsCorrectly) {
+  const auto graph = R"IR(
+    graph(%0 : Tensor):
+      %1 : float = prim::Constant[value=-1.0]()
+      %2 : float = prim::Constant[value=1.0]()
+      %3 : Tensor = aten::hardtanh(%0, %1, %2)
+      return (%3))IR";
 
-//     auto g = std::make_shared<torch::jit::Graph>();
-//     torch::jit::script::parseIR(graph, &*g);
+  auto g = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(graph, &*g);
 
-//     auto in = at::randint(-5, 5, {5}, {at::kCUDA});
-//     auto params = trtorch::core::conversion::get_named_params(g->inputs(),
-//     {}); auto jit_results = trtorch::tests::util::RunGraph(g, params, {in});
+  auto in = at::randint(-5, 5, {5}, {at::kCUDA});
+  auto params = trtorch::core::conversion::get_named_params(g->inputs(), {});
+  auto jit_results = trtorch::tests::util::RunGraph(g, params, {in});
 
-//     in = at::clone(in);
-//     params = trtorch::core::conversion::get_named_params(g->inputs(), {});
-//     auto trt_results = trtorch::tests::util::RunGraphEngine(g, params, {in});
+  in = at::clone(in);
+  params = trtorch::core::conversion::get_named_params(g->inputs(), {});
+  auto trt_results = trtorch::tests::util::RunGraphEngine(g, params, {in});
 
-//     ASSERT_TRUE(trtorch::tests::util::almostEqual(jit_results[0],
-//     trt_results[0], 2e-6));
-// }
+  ASSERT_TRUE(trtorch::tests::util::almostEqual(jit_results[0],
+  trt_results[0], 2e-6));
+}
 
 TEST(Converters, ATenHardTanhCustomRangeConvertsCorrectly) {
   const auto graph = R"IR(
@@ -114,7 +114,7 @@ TEST(Converters, ATenHardTanhCustomRangeConvertsCorrectly) {
 TEST(Converters, ATenPReLUConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(1:1)):
+            %1 : Float(1, strides=[1])):
         %3 : Tensor = aten::prelu(%0, %1)
         return (%3))IR";
 
@@ -137,7 +137,7 @@ TEST(Converters, ATenPReLUConvertsCorrectly) {
 TEST(Converters, ATenPReLUMultiChannelConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(10:1)):
+            %1 : Float(10, strides=[1])):
         %3 : Tensor = aten::prelu(%0, %1)
         return (%3))IR";
 

diff --git a/tests/core/conversion/converters/test_batch_norm.cpp b/tests/core/conversion/converters/test_batch_norm.cpp
@@ -7,10 +7,10 @@
 TEST(Converters, ATenBatchNormConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1: Float(5:1),
-            %2: Float(5:1),
-            %3: Float(5:1),
-            %4: Float(5:1)):
+            %1: Float(5, strides=[1]),
+            %2: Float(5, strides=[1]),
+            %3: Float(5, strides=[1]),
+            %4: Float(5, strides=[1])):
         %5 : bool = prim::Constant[value=0]()
         %6 : float = prim::Constant[value=1.0000000000000001e-05]()
         %7 : float = prim::Constant[value=0.10000000000000001]()

diff --git a/tests/core/conversion/converters/test_conv_deconv.cpp b/tests/core/conversion/converters/test_conv_deconv.cpp
@@ -39,8 +39,8 @@ void conv_test_helper(std::string graph_ir) {
 TEST(Converters, ATenConvolutionConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(8:45, 3:15, 5:5, 5:1),
-            %2 : Float(8:1)):
+            %1 : Float(8, 3, 5, 5, strides=[45, 15, 5, 1]),
+            %2 : Float(8)):
         %3 : int = prim::Constant[value=1]()
         %4 : int = prim::Constant[value=0]()
         %5 : int = prim::Constant[value=1]()
@@ -81,7 +81,7 @@ TEST(Converters, ATenConvolutionConvertsCorrectly) {
 TEST(Converters, ATenConvolutionNoBiasConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(4:9, 1:9, 3:3, 3:1)):
+            %1 : Float(4, 1, 3, 3, strides=[9, 9, 3, 1])):
         %2 : None = prim::Constant()
         %3 : int = prim::Constant[value=1]()
         %4 : int = prim::Constant[value=0]()
@@ -119,8 +119,8 @@ TEST(Converters, ATenConvolutionNoBiasConvertsCorrectly) {
 TEST(Converters, ATenConvolutionWithStrideConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(4:27, 3:9, 3:3, 3:1),
-            %2 : Float(4:1)):
+            %1 : Float(4, 3, 3, 3, strides=[27, 9, 3, 1]),
+            %2 : Float(4)):
         %3 : int = prim::Constant[value=3]()
         %4 : int = prim::Constant[value=0]()
         %5 : int = prim::Constant[value=1]()
@@ -162,8 +162,8 @@ TEST(Converters, ATenConvolutionWithStrideConvertsCorrectly) {
 TEST(Converters, ATenConvolutionWithPaddingConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(4:48, 3:16, 4:4, 4:1),
-            %2 : Float(4:1)):
+            %1 : Float(4, 3, 4, 4, strides=[48, 16, 4, 1]),
+            %2 : Float(4)):
         %3 : int = prim::Constant[value=1]()
         %4 : int = prim::Constant[value=2]()
         %5 : int = prim::Constant[value=1]()
@@ -205,8 +205,8 @@ TEST(Converters, ATenConvolutionWithPaddingConvertsCorrectly) {
 TEST(Converters, ATenConvolution3dConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(32:81, 3:27, 3:9, 3:3, 3:1),
-            %2 : Float(32:1)):
+            %1 : Float(32, 3, 3, 3, 3, strides=[81, 27, 9, 3, 1]),
+            %2 : Float(32)):
         %sv : int = prim::Constant[value=1]()
         %s : int[] = prim::ListConstruct(%sv, %sv, %sv)
         %pv : int = prim::Constant[value=0]()
@@ -247,7 +247,7 @@ TEST(Converters, ATenConvolution3dConvertsCorrectly) {
 TEST(Converters, ATenConvolution3dNoBiasConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(32:81, 3:27, 3:9, 3:3, 3:1)):
+            %1 : Float(32, 3, 3, 3, 3, strides=[81, 27, 9, 3, 1])):
         %bias : None = prim::Constant()
         %sv : int = prim::Constant[value=1]()
         %s : int[] = prim::ListConstruct(%sv, %sv, %sv)
@@ -285,8 +285,8 @@ TEST(Converters, ATenConvolution3dNoBiasConvertsCorrectly) {
 TEST(Converters, ATenConvolution3dWithPaddingConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(32:81, 3:27, 3:9, 3:3, 3:1),
-            %2 : Float(32:1)):
+            %1 : Float(32, 3, 3, 3, 3, strides=[81, 27, 9, 3, 1]),
+            %2 : Float(32)):
         %sv : int = prim::Constant[value=1]()
         %s : int[] = prim::ListConstruct(%sv, %sv, %sv)
         %pv : int = prim::Constant[value=1]()
@@ -327,8 +327,8 @@ TEST(Converters, ATenConvolution3dWithPaddingConvertsCorrectly) {
 TEST(Converters, ATenConvolution3dWithStrideDilationConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(32:81, 3:27, 3:9, 3:3, 3:1),
-            %2 : Float(32:1)):
+            %1 : Float(32, 3, 3, 3, 3, strides=[81, 27, 9, 3, 1]),
+            %2 : Float(32)):
         %sv : int = prim::Constant[value=2]()
         %s : int[] = prim::ListConstruct(%sv, %sv, %sv)
         %pv : int = prim::Constant[value=1]()
@@ -369,8 +369,8 @@ TEST(Converters, ATenConvolution3dWithStrideDilationConvertsCorrectly) {
 TEST(Converters, ATenConvTransposeConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(8:27, 3:9, 3:3, 3:1),
-            %2 : Float(8:1)):
+            %1 : Float(8, 3, 3, 3, strides=[27, 9, 3, 1]),
+            %2 : Float(8)):
         %3 : int = prim::Constant[value=1]()
         %4 : int = prim::Constant[value=0]()
         %5 : int = prim::Constant[value=1]()
@@ -411,7 +411,7 @@ TEST(Converters, ATenConvTransposeConvertsCorrectly) {
 TEST(Converters, ATenConvTransposeNoBiasConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(4:9, 1:9, 3:3, 3:1)):
+            %1 : Float(4, 1, 3, 3, strides=[9, 9, 3, 1])):
         %2 : None = prim::Constant()
         %3 : int = prim::Constant[value=1]()
         %4 : int = prim::Constant[value=0]()
@@ -449,8 +449,8 @@ TEST(Converters, ATenConvTransposeNoBiasConvertsCorrectly) {
 TEST(Converters, ATenConvTransposeWithStrideConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(4:27, 3:9, 3:3, 3:1),
-            %2 : Float(4:1)):
+            %1 : Float(4, 3, 3, 3, strides=[27, 9, 3, 1]),
+            %2 : Float(4)):
         %3 : int = prim::Constant[value=3]()
         %4 : int = prim::Constant[value=0]()
         %5 : int = prim::Constant[value=1]()
@@ -492,8 +492,8 @@ TEST(Converters, ATenConvTransposeWithStrideConvertsCorrectly) {
 TEST(Converters, ATenConvTransposeWithPaddingConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(4:48, 3:16, 4:4, 4:1),
-            %2 : Float(4:1)):
+            %1 : Float(4, 3, 4, 4, strides=[48, 16, 4, 1]),
+            %2 : Float(4)):
         %3 : int = prim::Constant[value=1]()
         %4 : int = prim::Constant[value=2]()
         %5 : int = prim::Constant[value=1]()
@@ -535,8 +535,8 @@ TEST(Converters, ATenConvTransposeWithPaddingConvertsCorrectly) {
 TEST(Converters, ATenConvolutionWithGroupConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(8:48, 1:16, 2:4, 2:1),
-            %2 : Float(8:1)):
+            %1 : Float(8, 1, 2, 2, strides=[48, 16, 4, 1]),
+            %2 : Float(8)):
         %3 : int = prim::Constant[value=1]()
         %4 : int = prim::Constant[value=2]()
         %5 : int = prim::Constant[value=1]()
@@ -578,8 +578,8 @@ TEST(Converters, ATenConvolutionWithGroupConvertsCorrectly) {
 TEST(Converters, ATenConvTransposeWithGroupConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(8:56, 4:16, 3:3, 3:1),
-            %2 : Float(16:1)):
+            %1 : Float(8, 4, 3, 3, strides=[56, 16, 3, 1]),
+            %2 : Float(16)):
         %3 : int = prim::Constant[value=1]()
         %4 : int = prim::Constant[value=1]()
         %5 : int = prim::Constant[value=1]()

diff --git a/tests/core/conversion/converters/test_linear.cpp b/tests/core/conversion/converters/test_linear.cpp
@@ -7,7 +7,7 @@
 TEST(Converters, ATenLinearNoBiasConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(3:2, 2:1)):
+            %1 : Float(3, 2, strides=[2, 1])):
         %2 : None = prim::Constant()
         %3 : Tensor = aten::linear(%0, %1, %2)
         return (%3))IR";
@@ -33,8 +33,8 @@ TEST(Converters, ATenLinearNoBiasConvertsCorrectly) {
 TEST(Converters, ATenLinearBiasConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(2:3, 3:1),
-            %2 : Float(2:1)):
+            %1 : Float(2, 3, strides=[3, 1]),
+            %2 : Float(2)):
         %3 : Tensor = aten::linear(%0, %1, %2)
         return (%3))IR";
 

diff --git a/tests/core/conversion/converters/test_select.cpp b/tests/core/conversion/converters/test_select.cpp
@@ -88,7 +88,7 @@ TEST(Converters, ATenNarrowStartScalarConvertsCorrectly) {
 
 TEST(Converters, ATenEmbeddingConvertsCorrectly) {
   const auto graph = R"IR(
-      graph(%1 : Tensor, %emb_weight : Float(10:3, 3:1)):
+      graph(%1 : Tensor, %emb_weight : Float(10, 3, strides=[3, 1])):
             %2 : bool = prim::Constant[value=0]()
             %3 : int = prim::Constant[value=-1]()
             %5 : Tensor = aten::embedding(%emb_weight, %1, %3, %2, %2)

diff --git a/tests/core/conversion/converters/test_stack.cpp b/tests/core/conversion/converters/test_stack.cpp
@@ -31,7 +31,7 @@ TEST(Converters, ATenStackPureTensorConvertsCorrectly) {
 TEST(Converters, ATenStackDiffTensorConvertsCorrectly) {
   const auto graph = R"IR(
       graph(%0 : Tensor,
-            %1 : Float(4:16, 4:4, 4:1)):
+            %1 : Float(4, 4, 4, strides=[16, 4, 1])):
         %2 : Tensor[] = prim::ListConstruct(%0, %1)
         %3 : int = prim::Constant[value=1]()
         %4 : Tensor = aten::stack(%2, %3)

diff --git a/tests/py/requirements.txt b/tests/py/requirements.txt
@@ -1 +1 @@
-torchvision==0.8.2
+torchvision==0.9.0