chore: Upgrade TensorRT version to TRT 10 EA (#2699) (#2774)

Co-authored-by: Evan Li <[email protected]>
Co-authored-by: George S <[email protected]>

Author: 3 people

.github/scripts/install-torch-tensorrt.sh

@@ -5,6 +5,13 @@ source ${BUILD_ENV_FILE}
 ${CONDA_RUN} ${PIP_INSTALL_TORCH} torchvision
 ${CONDA_RUN} python -m pip install pyyaml mpmath==1.3.0
 export TRT_VERSION=$(${CONDA_RUN} python -c "import versions; versions.tensorrt_version()")
-${CONDA_RUN} python -m pip install /opt/torch-tensorrt-builds/torch_tensorrt*+${CU_VERSION}*.whl tensorrt~=${TRT_VERSION} tensorrt-bindings~=${TRT_VERSION} --extra-index-url=https://pypi.ngc.nvidia.com
+
+# Install TensorRT manually
+wget -q -P /opt/torch-tensorrt-builds/ https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/10.0.0/TensorRT-10.0.0.6.Linux.x86_64-gnu.cuda-12.4.tar.gz
+tar -xzf /opt/torch-tensorrt-builds/TensorRT-10.0.0.6.Linux.x86_64-gnu.cuda-12.4.tar.gz -C /opt/torch-tensorrt-builds/
+python -m pip install /opt/torch-tensorrt-builds/TensorRT-10.0.0.6/python/tensorrt-10.0.0b6-cp${PYTHON_VERSION//./}-none-linux_x86_64.whl
+
+# Install Torch-TensorRT
+${CONDA_RUN} python -m pip install /opt/torch-tensorrt-builds/torch_tensorrt*+${CU_VERSION}*.whl
 
 echo -e "Running test script";

.github/workflows/build-test.yml

@@ -15,7 +15,7 @@ on:
 
 jobs:
   generate-matrix:
-    uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@main
+    uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@release/2.3
     with:
       package-type: wheel
       os: linux
@@ -36,11 +36,11 @@ jobs:
           - repository: pytorch/tensorrt
             pre-script: packaging/pre_build_script.sh
             env-var-script: packaging/env_vars.txt
-            post-script: ""
-            smoke-test-script: ""
+            post-script: packaging/post_build_script.sh
+            smoke-test-script: packaging/smoke_test_script.sh
             package-name: torch_tensorrt
     name: Build torch-tensorrt whl package
-    uses: pytorch/test-infra/.github/workflows/build_wheels_linux.yml@main
+    uses: pytorch/test-infra/.github/workflows/build_wheels_linux.yml@release/2.3
     with:
       repository: ${{ matrix.repository }}
       ref: ""
@@ -64,7 +64,8 @@ jobs:
           - repository: pytorch/tensorrt
             package-name: torch_tensorrt
             pre-script: packaging/pre_build_script.sh
-    uses: pytorch/tensorrt/.github/workflows/linux-test.yml@main
+            post-script: packaging/post_build_script.sh
+    uses: pytorch/tensorrt/.github/workflows/linux-test.yml@release/2.3
     with:
       job-name: tests-py-torchscript-fe
       repository: "pytorch/tensorrt"
@@ -76,9 +77,11 @@ jobs:
       script: |
         export USE_HOST_DEPS=1
         export LD_LIBRARY_PATH=/usr/lib64:$LD_LIBRARY_PATH
+        export LD_LIBRARY_PATH=/opt/torch-tensorrt-builds/TensorRT-10.0.0.6/lib:$LD_LIBRARY_PATH
         pushd .
         cd tests/modules
-        ${CONDA_RUN} python -m pip install --pre -r requirements.txt --use-deprecated=legacy-resolver
+        # Don't use requirements.txt here as it contains tensorrt and torch which should have been installed by now.
+        ${CONDA_RUN} python -m pip install numpy packaging pyyaml transformers timm pybind11==2.6.2
         ${CONDA_RUN} python hub.py
         popd
         pushd .
@@ -99,7 +102,8 @@ jobs:
           - repository: pytorch/tensorrt
             package-name: torch_tensorrt
             pre-script: packaging/pre_build_script.sh
-    uses: pytorch/tensorrt/.github/workflows/linux-test.yml@main
+            post-script: packaging/post_build_script.sh
+    uses: pytorch/tensorrt/.github/workflows/linux-test.yml@release/2.3
     with:
       job-name: tests-py-dynamo-converters
       repository: "pytorch/tensorrt"
@@ -110,6 +114,7 @@ jobs:
       pre-script: ${{ matrix.pre-script }}
       script: |
         export USE_HOST_DEPS=1
+        export LD_LIBRARY_PATH=/opt/torch-tensorrt-builds/TensorRT-10.0.0.6/lib:$LD_LIBRARY_PATH
         pushd .
         cd tests/py/dynamo
         ${CONDA_RUN} python -m pip install --pre pytest-xdist timm transformers parameterized expecttest==0.1.6 --use-deprecated=legacy-resolver
@@ -126,7 +131,8 @@ jobs:
           - repository: pytorch/tensorrt
             package-name: torch_tensorrt
             pre-script: packaging/pre_build_script.sh
-    uses: pytorch/tensorrt/.github/workflows/linux-test.yml@main
+            post-script: packaging/post_build_script.sh
+    uses: pytorch/tensorrt/.github/workflows/linux-test.yml@release/2.3
     with:
       job-name: tests-py-dynamo-fe
       repository: "pytorch/tensorrt"
@@ -137,6 +143,7 @@ jobs:
       pre-script: ${{ matrix.pre-script }}
       script: |
         export USE_HOST_DEPS=1
+        export LD_LIBRARY_PATH=/opt/torch-tensorrt-builds/TensorRT-10.0.0.6/lib:$LD_LIBRARY_PATH
         pushd .
         cd tests/py/dynamo
         ${CONDA_RUN} python -m pip install --pre pytest timm transformers parameterized expecttest==0.1.6 --use-deprecated=legacy-resolver
@@ -154,7 +161,8 @@ jobs:
           - repository: pytorch/tensorrt
             package-name: torch_tensorrt
             pre-script: packaging/pre_build_script.sh
-    uses: pytorch/tensorrt/.github/workflows/linux-test.yml@main
+            post-script: packaging/post_build_script.sh
+    uses: pytorch/tensorrt/.github/workflows/linux-test.yml@release/2.3
     with:
       job-name: tests-py-dynamo-serde
       repository: "pytorch/tensorrt"
@@ -165,6 +173,7 @@ jobs:
       pre-script: ${{ matrix.pre-script }}
       script: |
         export USE_HOST_DEPS=1
+        export LD_LIBRARY_PATH=/opt/torch-tensorrt-builds/TensorRT-10.0.0.6/lib:$LD_LIBRARY_PATH
         pushd .
         cd tests/py/dynamo
         ${CONDA_RUN} python -m pip install --pre pytest timm transformers parameterized expecttest==0.1.6 --use-deprecated=legacy-resolver
@@ -181,7 +190,8 @@ jobs:
           - repository: pytorch/tensorrt
             package-name: torch_tensorrt
             pre-script: packaging/pre_build_script.sh
-    uses: pytorch/tensorrt/.github/workflows/linux-test.yml@main
+            post-script: packaging/post_build_script.sh
+    uses: pytorch/tensorrt/.github/workflows/linux-test.yml@release/2.3
     with:
       job-name: tests-py-torch-compile-be
       repository: "pytorch/tensorrt"
@@ -192,6 +202,7 @@ jobs:
       pre-script: ${{ matrix.pre-script }}
       script: |
         export USE_HOST_DEPS=1
+        export LD_LIBRARY_PATH=/opt/torch-tensorrt-builds/TensorRT-10.0.0.6/lib:$LD_LIBRARY_PATH
         pushd .
         cd tests/py/dynamo
         ${CONDA_RUN} python -m pip install --pre pytest-xdist timm transformers parameterized expecttest==0.1.6 --use-deprecated=legacy-resolver
@@ -209,7 +220,8 @@ jobs:
           - repository: pytorch/tensorrt
             package-name: torch_tensorrt
             pre-script: packaging/pre_build_script.sh
-    uses: pytorch/tensorrt/.github/workflows/linux-test.yml@main
+            post-script: packaging/post_build_script.sh
+    uses: pytorch/tensorrt/.github/workflows/linux-test.yml@release/2.3
     with:
       job-name: tests-py-dynamo-core
       repository: "pytorch/tensorrt"
@@ -220,6 +232,7 @@ jobs:
       pre-script: ${{ matrix.pre-script }}
       script: |
         export USE_HOST_DEPS=1
+        export LD_LIBRARY_PATH=/opt/torch-tensorrt-builds/TensorRT-10.0.0.6/lib:$LD_LIBRARY_PATH
         pushd .
         cd tests/py/dynamo
         ${CONDA_RUN} python -m pip install --pre pytest-xdist timm transformers parameterized expecttest==0.1.6 --use-deprecated=legacy-resolver
@@ -249,6 +262,7 @@ jobs:
       pre-script: ${{ matrix.pre-script }}
       script: |
         export USE_HOST_DEPS=1
+        export LD_LIBRARY_PATH=/opt/torch-tensorrt-builds/TensorRT-10.0.0.6/lib:$LD_LIBRARY_PATH
         pushd .
         cd tests/py/core
         ${CONDA_RUN} python -m pip install --pre pytest-xdist timm transformers parameterized expecttest==0.1.6 --use-deprecated=legacy-resolver

README.md

@@ -119,7 +119,7 @@ These are the following dependencies used to verify the testcases. Torch-TensorR
 - Libtorch 2.4.0.dev (latest nightly) (built with CUDA 12.1)
 - CUDA 12.1
 - cuDNN 8.9.5
-- TensorRT 8.6.1
+- TensorRT 10.0.0.6
 
 ## Prebuilt Binaries and Wheel files
 

core/conversion/converters/converter_util.cpp

@@ -39,6 +39,12 @@ nvinfer1::ITensor* addPadding(
   }
 }
 
+nvinfer1::ITensor* getShapeOutput(ConversionCtx* ctx, nvinfer1::ITensor* input_tensor, const std::string& name) {
+  nvinfer1::ITensor* input_shape = ctx->net->addShape(*input_tensor)->getOutput(0);
+  input_shape = castITensor(ctx, input_shape, nvinfer1::DataType::kINT32, name);
+  return input_shape;
+}
+
 nvinfer1::ITensor* addUnpadding(
     ConversionCtx* ctx,
     const torch::jit::Node* n,
@@ -134,7 +140,7 @@ nvinfer1::ILayer* add_elementwise(
       }
       auto otherStaticShapeMask = tensor_to_const(ctx, thOtherStaticShapeMask);
       auto otherDynamicShapeMask = tensor_to_const(ctx, thOtherDynamicShapeMask);
-      auto selfShape = ctx->net->addShape(*self)->getOutput(0);
+      nvinfer1::ITensor* selfShape = getShapeOutput(ctx, self, std::string(name + "_shape_cast").c_str());
       // size of dynamic dimension of other need to the same as that of
       // corresponding dimension of self
       auto otherDynamicShape =
@@ -348,7 +354,6 @@ nvinfer1::ITensor* normalize_indices(
   auto neg_itensor = tensor_to_const(ctx, neg);
   // find the indices that = -1
   auto signs = clamp(ctx, indices, neg_itensor, zero_itensor, "clamp layer for " + name);
-
   // get the inputDim value where indices == -1, else 0
   auto mul = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kPROD, signs, input_dim, "prod layer for " + name);
   TORCHTRT_CHECK(mul, "Unable to create mul layer in normalize_indices");

core/conversion/converters/converter_util.h

@@ -62,6 +62,9 @@ nvinfer1::ITensor* castITensor(
     nvinfer1::DataType dtype,
     const std::string& layer_name_prefix = "");
 
+// Get the shape of the input tensor and cast it to INT32 type
+nvinfer1::ITensor* getShapeOutput(ConversionCtx* ctx, nvinfer1::ITensor* input_tensor, const std::string& name = "");
+
 // Freeze an at::Tensor in a IConstant layer
 nvinfer1::ITensor* tensor_to_const(ConversionCtx* ctx, at::Tensor t, const std::string& name = std::string());
 

core/conversion/converters/impl/chunk.cpp

@@ -17,7 +17,6 @@ auto cat_registrations TORCHTRT_UNUSED = RegisterNodeConversionPatterns()
               auto chunks = args[1].unwrapToInt();
               auto dim = args[2].unwrapToInt();
               bool dynamic_shape = ctx->input_is_dynamic;
-              int size = in->getDimensions().nbDims;
               int maxDim = static_cast<int32_t>(in->getDimensions().d[dim]);
 
               c10::ListTypePtr lt = n->output()->type()->expect<c10::ListType>();
@@ -41,9 +40,6 @@ auto cat_registrations TORCHTRT_UNUSED = RegisterNodeConversionPatterns()
               size_.nbDims = nbdims;
               stride_.nbDims = nbdims;
 
-              int startIdx = 0;
-              int endIdx = maxDim;
-
               for (int i = 0; i < nbdims; i++) {
                 start_.d[i] = 0;
                 size_.d[i] = 0;

core/conversion/converters/impl/constant_pad.cpp

@@ -55,18 +55,15 @@ auto constant_pad_registrations TORCHTRT_UNUSED = RegisterNodeConversionPatterns
            util::toDims(c10::IntArrayRef(stride)));
        TORCHTRT_CHECK(slice_layer, "Unable to create slice layer from node: " << *n);
        slice_layer->setName((util::node_info(n) + "_slice").c_str());
-       slice_layer->setMode(nvinfer1::SliceMode::kFILL);
+       slice_layer->setMode(nvinfer1::SampleMode::kFILL);
        slice_layer->setInput(4, *value_itensor);
 
        if (ctx->input_is_dynamic) {
          // build the size using inetwork layers
-         auto shape_layer = ctx->net->addShape(*in);
-         TORCHTRT_CHECK(shape_layer, "Unable to create shape layer from node: " << *n);
-         shape_layer->setName((util::node_info(n) + "_shape").c_str());
          auto total_padding_itensor = tensor_to_const(ctx, torch::tensor(total_padding, torch::kInt32));
-
-         auto add_layer = ctx->net->addElementWise(
-             *shape_layer->getOutput(0), *total_padding_itensor, nvinfer1::ElementWiseOperation::kSUM);
+         nvinfer1::ITensor* shapeOutput = getShapeOutput(ctx, in, (util::node_info(n) + "_shape").c_str());
+         auto add_layer =
+             ctx->net->addElementWise(*shapeOutput, *total_padding_itensor, nvinfer1::ElementWiseOperation::kSUM);
          TORCHTRT_CHECK(add_layer, "Unable to create add layer from node: " << *n);
          add_layer->setName((util::node_info(n) + "_add").c_str());
          slice_layer->setInput(2, *add_layer->getOutput(0));

core/conversion/converters/impl/conv_deconv.cpp

@@ -33,7 +33,7 @@ nvinfer1::ILayer* add_bias_layer(
     nvinfer1::Dims& input_dims,
     nvinfer1::Dims& output_padding,
     Weights& bias) {
-  nvinfer1::ITensor* input_shape = ctx->net->addShape(*input_tensor)->getOutput(0);
+  nvinfer1::ITensor* input_shape = getShapeOutput(ctx, input_tensor, std::string("bias_shape_cast").c_str());
   // Add padding layer
   nvinfer1::ITensor* start;
   nvinfer1::ITensor* totalPadding;
@@ -61,7 +61,7 @@ nvinfer1::ILayer* add_bias_layer(
   auto* sliceLayer = ctx->net->addSlice(*input_tensor, dummy, dummy, stride);
   sliceLayer->setInput(1, *start);
   sliceLayer->setInput(2, *size);
-  sliceLayer->setMode(nvinfer1::SliceMode::kFILL);
+  sliceLayer->setMode(nvinfer1::SampleMode::kFILL);
   nvinfer1::ITensor* slice_output = sliceLayer->getOutput(0);
 
   nvinfer1::Dims constantDims;
@@ -146,9 +146,9 @@ bool add_conv_deconv(ConversionCtx* ctx, const torch::jit::Node* n, args& args)
       // TensorRT expects nbSpatialDims = 2 or 3
       filter_dim = util::unsqueezeDims(filter_dim, filter_dim.nbDims, 1, false);
       // Reshape input dimensions
-      in = addPadding(ctx, n, in, 4);
+      in = addPadding(ctx, n, in, 4, true, true, std::string(util::node_info(n) + "_input_shuffle"));
       LOG_DEBUG("Reshaping input dimensions to: " << in->getDimensions());
-      kernel = addPadding(ctx, n, kernel, 4);
+      kernel = addPadding(ctx, n, kernel, 4, true, true, std::string(util::node_info(n) + "_kernel_shuffle"));
       LOG_DEBUG("Reshaping kernel dimensions to: " << kernel->getDimensions());
       if (transposed) {
         num_output_maps = kernel_dims.d[1];
@@ -194,7 +194,7 @@ bool add_conv_deconv(ConversionCtx* ctx, const torch::jit::Node* n, args& args)
       nvinfer1::IConvolutionLayer* convLayer =
           ctx->net->addConvolutionNd(*in, num_output_maps, filter_dim, kernel_weights, bias.data);
       convLayer->setStrideNd(stride);
-      convLayer->setPaddingMode(nvinfer1::PaddingMode::kCAFFE_ROUND_DOWN);
+      convLayer->setPaddingMode(nvinfer1::PaddingMode::kEXPLICIT_ROUND_DOWN);
       convLayer->setPaddingNd(padding);
       convLayer->setPostPadding(out_padding);
       convLayer->setDilationNd(dilation);
@@ -291,11 +291,9 @@ bool add_conv_deconv(ConversionCtx* ctx, const torch::jit::Node* n, args& args)
     // shape of convolution's weight: [out, in/groups, ...]
     auto conv = ctx->net->addConvolutionNd(*in, w.shape.d[0], w.kernel_shape, w.data, bias.data);
     TORCHTRT_CHECK(conv, "Unable to create convolution layer from node: " << *n);
-
     conv->setStrideNd(stride);
-    conv->setPaddingMode(nvinfer1::PaddingMode::kCAFFE_ROUND_DOWN);
+    conv->setPaddingMode(nvinfer1::PaddingMode::kEXPLICIT_ROUND_DOWN);
     conv->setPaddingNd(padding);
-    conv->setPostPadding(out_padding);
     conv->setDilationNd(dilation);
     conv->setNbGroups(groups);
     new_layer = conv;

core/conversion/converters/impl/cumsum.cpp

@@ -36,7 +36,7 @@ auto cumsum_registrations TORCHTRT_UNUSED = RegisterNodeConversionPatterns().pat
          torch::Tensor axis = torch::tensor(input_dims.d[dim], torch::kInt32);
          tripLimit = tensor_to_const(ctx, axis);
        } else {
-         nvinfer1::ITensor* inpShape = ctx->net->addShape(*in)->getOutput(0);
+         nvinfer1::ITensor* inpShape = getShapeOutput(ctx, in);
          torch::Tensor dimValue = torch::tensor(dim, torch::kInt32);
          nvinfer1::ITensor* axis = tensor_to_const(ctx, dimValue);
          tripLimit = ctx->net->addGather(*inpShape, *axis, 0)->getOutput(0);

core/conversion/converters/impl/expand.cpp

@@ -19,11 +19,11 @@ nvinfer1::ITensor* concat(int max_rank, int old_rank, ConversionCtx* ctx, nvinfe
   if (max_rank - old_rank > 0) {
     torch::Tensor thOne = torch::tensor(std::vector<int32_t>(max_rank - old_rank, 1), torch::kInt32);
     auto one_tensor = tensor_to_const(ctx, thOne);
-    auto in_shape_tensor = ctx->net->addShape(*tensor)->getOutput(0);
+    auto in_shape_tensor = getShapeOutput(ctx, tensor);
     nvinfer1::ITensor* const args[2] = {one_tensor, in_shape_tensor};
     return ctx->net->addConcatenation(args, 2)->getOutput(0);
   } else { // max_rank - old_rank == 0
-    return ctx->net->addShape(*tensor)->getOutput(0);
+    return getShapeOutput(ctx, tensor);
   }
 }
 
@@ -221,8 +221,7 @@ auto expand_registrations TORCHTRT_UNUSED =
                auto targetDims = targetTensor->getDimensions();
                LOG_DEBUG("(expand_as layer) Expand input from " << input_dims << " to " << targetDims);
                if (ctx->input_is_dynamic) {
-                 return add_expand_dynamic(
-                     ctx, n, in, ctx->net->addShape(*targetTensor)->getOutput(0), targetDims, false);
+                 return add_expand_dynamic(ctx, n, in, getShapeOutput(ctx, targetTensor), targetDims, false);
                } else {
                  return add_expand(ctx, n, in, targetDims);
                }
@@ -357,7 +356,7 @@ auto expand_registrations TORCHTRT_UNUSED =
                if (ctx->input_is_dynamic) {
                  auto start_tensor = tensor_to_const(ctx, torch::tensor(start_vec, torch::kInt32));
 
-                 auto expand_output_shape = ctx->net->addShape(*expand->getOutput(0))->getOutput(0);
+                 auto expand_output_shape = getShapeOutput(ctx, expand->getOutput(0));
                  std::vector<int64_t> repeat_const_vec(repeat_shape_dims.nbDims, 1);
                  repeat_const_vec[dim + 1] = repeats;
                  auto repeat_const = tensor_to_const(ctx, torch::tensor(repeat_const_vec, torch::kInt32));