Merge pull request #389 from NVIDIA/1.8_update

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

Author: narendasan

.github/workflows/docgen.yml

@@ -22,7 +22,7 @@ jobs:
       - name: Run image
         run: docker run -it -d --name builder -v $GITHUB_WORKSPACE:/workspace -w /workspace docker.pkg.github.com/nvidia/trtorch/docgen:latest
       - name: Build python package
-        run: docker exec builder bash -c "cp docker/WORKSPACE.docker WORKSPACE && cd py && python3 setup.py install"
+        run: docker exec builder bash -c "cp docker/WORKSPACE.docs WORKSPACE && cd py && python3 setup.py install"
       - name: Generate new docs
         run: docker exec builder bash -c "cd docsrc && make html"
       - uses: stefanzweifel/git-auto-commit-action@v4

.github/workflows/linter.yml

@@ -19,7 +19,7 @@ jobs:
       - name: Run image
         run: |
           docker run -it -d --name cpplinter -e GITHUB_TOKEN=$GITHUB_TOKEN -v $GITHUB_WORKSPACE:/workspace -v $GITHUB_EVENT_PATH:/GITHUB_EVENT.json -w /workspace docker.pkg.github.com/nvidia/trtorch/docgen:latest
-          docker exec cpplinter bash -c "cp /workspace/docker/WORKSPACE.docker /workspace/WORKSPACE"
+          docker exec cpplinter bash -c "cp /workspace/docker/WORKSPACE.docs /workspace/WORKSPACE"
         env:
           GITHUB_TOKEN: ${{secrets.GITHUB_TOKEN}}
       - name: Install dependencies
@@ -41,7 +41,7 @@ jobs:
       - name: Run image
         run: |
           docker run -it -d --name pylinter -e GITHUB_TOKEN=$GITHUB_TOKEN -v $GITHUB_WORKSPACE:/workspace -v $GITHUB_EVENT_PATH:/GITHUB_EVENT.json -w /workspace docker.pkg.github.com/nvidia/trtorch/docgen:latest
-          docker exec pylinter bash -c "cp /workspace/docker/WORKSPACE.docker /workspace/WORKSPACE"
+          docker exec pylinter bash -c "cp /workspace/docker/WORKSPACE.docs /workspace/WORKSPACE"
         env:
           GITHUB_TOKEN: ${{secrets.GITHUB_TOKEN}}
       - name: Install dependencies

CONTRIBUTING.md

@@ -16,7 +16,7 @@ We use the PyTorch Slack for communication about core development, integration w
 
     - Use the built in linting tools to ensure that your code matches the style guidelines
       ```sh
-      # C++ Linting (After installing clang-format)
+      # C++ Linting (After installing clang-format [Version 9.0.0])
       # Print non-conforming sections of code
       bazel run //tools/linter:cpplint_diff -- //...
       # Modify code to conform with style guidelines

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
 

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"
 )
 
 ####################################################################################

core/conversion/converters/impl/plugins/interpolate_plugin.cpp

@@ -254,7 +254,8 @@ int InterpolatePlugin::enqueue(
     cudaStream_t stream) {
 #if NV_TENSORRT_MAJOR < 7 || (NV_TENSORRT_MAJOR == 7 && NV_TENSORRT_MINOR < 1)
   at::Tensor input = at::from_blob((void*)inputs[0], util::toVec(inputDesc->dims), [](void*) {}, tensor_options_);
-  at::Tensor output = at::from_blob(outputs[0], util::volume(outputDesc->dims), [](void*) {}, tensor_options_);
+  at::Tensor output = at::from_blob(
+      outputs[0], util::volume(outputDesc->dims), [](void*) {}, tensor_options_);
 
   at::cuda::CUDAStream torch_stream = at::cuda::getStreamFromPool();
   at::cuda::CUDAStreamGuard torch_guard(torch_stream);

core/conversion/evaluators/aten.cpp

@@ -396,6 +396,20 @@ auto aten_registrations TRTORCH_UNUSED =
                     EvalOptions().validSchemas({
                         "aten::numel(Tensor self) -> int",
                     })})
+        .evaluator({c10::Symbol::fromQualString("aten::t"),
+                    [](const torch::jit::Node* n, kwargs& args) -> c10::optional<torch::jit::IValue> {
+                      auto tensor_var = args.at(n->input(0));
+                      if (tensor_var.IValue()->isTensor()) {
+                        auto tensor = tensor_var.unwrapToTensor();
+                        return tensor.t();
+                      } else {
+                        TRTORCH_THROW_ERROR("Unimplemented data type for aten::t evaluator: ITensor");
+                        return {};
+                      }
+                    },
+                    EvalOptions().validSchemas({
+                        "aten::t(Tensor self) -> Tensor",
+                    })})
         .evaluator({c10::Symbol::fromQualString("aten::dim"),
                     [](const torch::jit::Node* n, kwargs& args) -> c10::optional<torch::jit::IValue> {
                       auto tensor_var = args.at(n->input(0));

core/lowering/lowering.cpp

@@ -36,7 +36,7 @@ void LowerGraph(std::shared_ptr<torch::jit::Graph>& g) {
   torch::jit::LowerAllTuples(g);
   passes::RemoveContiguous(g);
   passes::RemoveDropout(g);
-  passes::FuseFlattenLinear(g);
+  passes::LinearToAddMM(g);
   passes::Conv2DToConvolution(g);
   passes::Conv3DToConvolution(g);
   passes::FuseAddMMBranches(g);

core/lowering/passes/BUILD

@@ -17,7 +17,7 @@ cc_library(
         "conv3d_to_convolution.cpp",
         "exception_elimination.cpp",
         "fuse_addmm_branches.cpp",
-        "fuse_flatten_linear.cpp",
+        "linear_to_addmm.cpp",
         "remove_bn_dim_check.cpp",
         "remove_contiguous.cpp",
         "remove_dropout.cpp",

core/lowering/passes/fuse_flatten_linear.cpp renamed to core/lowering/passes/linear_to_addmm.cpp

@@ -7,29 +7,31 @@ namespace core {
 namespace lowering {
 namespace passes {
 
-void FuseFlattenLinear(std::shared_ptr<torch::jit::Graph>& graph) {
+void LinearToAddMM(std::shared_ptr<torch::jit::Graph>& graph) {
   // TensorRT implicitly adds a flatten layer infront of FC layers if necessary
   std::string flatten_linear_pattern = R"IR(
-        graph(%input, %6, %7, %weight, %bias):
-            %flat = aten::flatten(%input, %6, %7)
-            %res = aten::linear(%flat, %weight, %bias)
+        graph(%input, %weight, %bias):
+            %res = aten::linear(%input, %weight, %bias)
             return (%res))IR";
   std::string flatten_linear_bias_none_pattern = R"IR(
-        graph(%input, %6, %7, %weight):
-            %flat = aten::flatten(%input, %6, %7)
+        graph(%input, %weight):
             %bias: Tensor? = prim::Constant()
-            %res = aten::linear(%flat, %weight, %bias)
-            return (%res))IR";
-  std::string fused_linear = R"IR(
-        graph(%input, %6, %7, %weight, %bias):
             %res = aten::linear(%input, %weight, %bias)
             return (%res))IR";
 
+  std::string fused_linear = R"IR(
+        graph(%input, %weight_t, %bias):
+            %1: int = prim::Constant[value=1]()
+            %weight = aten::t(%weight_t)
+            %mm: Tensor = aten::matmul(%input, %weight)
+            %b_f: Tensor = trt::const(%bias)
+            %out: Tensor = aten::add_(%b_f, %mm, %1)
+            return (%out))IR";
   std::string fused_linear_bias_none = R"IR(
-        graph(%input, %6, %7, %weight):
-            %bias: Tensor? = prim::Constant()
-            %res = aten::linear(%input, %weight, %bias)
-            return (%res))IR";
+        graph(%input, %weight_t):
+            %weight = aten::t(%weight_t)
+            %mm: Tensor = aten::matmul(%input, %weight)
+            return (%mm))IR";
 
   torch::jit::SubgraphRewriter flatten_linear_to_linear;
   flatten_linear_to_linear.RegisterRewritePattern(flatten_linear_pattern, fused_linear);
@@ -38,7 +40,7 @@ void FuseFlattenLinear(std::shared_ptr<torch::jit::Graph>& graph) {
   torch::jit::SubgraphRewriter flatten_linear_bias_none_to_linear;
   flatten_linear_bias_none_to_linear.RegisterRewritePattern(flatten_linear_bias_none_pattern, fused_linear_bias_none);
   flatten_linear_bias_none_to_linear.runOnGraph(graph);
-  LOG_GRAPH("Post flatten linear: " << *graph);
+  LOG_GRAPH("Post linear to addmm: " << *graph);
 }
 
 } // namespace passes