fix: Upgrade to PyTorch 2.0.1 Release Candidate + Other improvements

.circleci/config.yml

@@ -102,7 +102,7 @@ commands:
             sudo apt-get --purge remove "*nvidia*"
 
   install-cudnn:
-    description: "Install CUDNN 8.5.0"
+    description: "Install CUDNN 8.8.0"
     parameters:
       os:
         type: string
@@ -112,10 +112,10 @@ commands:
         default: "x86_64"
       cudnn-version:
         type: string
-        default: "8.5.0.96"
+        default: "8.8.0.121"
       cuda-version:
         type: string
-        default: "cuda11.7"
+        default: "cuda11.8"
     steps:
       - run:
           name: Install CUDNN
@@ -200,7 +200,7 @@ commands:
         default: "cuda11.8"
       cudnn-version:
         type: string
-        default: "8.5.0.96"
+        default: "8.8.0.121"
       trt-version-short:
         type: string
         default: "8.6.0"
@@ -252,7 +252,7 @@ commands:
           default: "8.6.0"
         cudnn-version-long:
           type: string
-          default: "8.5.0.96"
+          default: "8.8.0.121"
       steps:
         - run:
             name: Set up python environment
@@ -269,10 +269,10 @@ commands:
     parameters:
       torch-build:
         type: string
-        default: "2.0.0"
+        default: "2.0.1"
       torch-build-index:
         type: string
-        default: "https://download.pytorch.org/whl/cu118"
+        default: "https://download.pytorch.org/whl/test/cu118"
     steps:
       - run:
           name: Install Torch
@@ -474,6 +474,7 @@ commands:
       - run: mkdir -p /tmp/artifacts
       - run:
           name: Run core / C++ tests
+          no_output_timeout: 15m
           environment:
             LD_LIBRARY_PATH: "/opt/circleci/.pyenv/versions/3.9.4/lib/python3.9/site-packages/torch_tensorrt.libs:/home/circleci/project/bazel-project/external/libtorch_pre_cxx11_abi/lib/:/home/circleci/project/bazel-project/external/tensorrt/lib/:/usr/local/cuda-11.8/lib64/:$LD_LIBRARY_PATH"
           command: |
@@ -1205,10 +1206,10 @@ parameters:
   # Nightly platform config
   torch-build:
     type: string
-    default: "2.0.0"
+    default: "2.0.1"
   torch-build-index:
     type: string
-    default: "https://download.pytorch.org/whl/cu118"
+    default: "https://download.pytorch.org/whl/test/cu118"
   torch-build-legacy:
     type: string
     default: "1.13.1+cu117"
@@ -1217,7 +1218,7 @@ parameters:
     default: "https://download.pytorch.org/whl/cu117"
   cudnn-version:
     type: string
-    default: "8.5.0.96"
+    default: "8.8.0.121"
   trt-version-short:
     type: string
     default: "8.6.0"
@@ -1412,4 +1413,3 @@ workflows:
           trt-version-short: << pipeline.parameters.trt-version-short >>
           cudnn-version: << pipeline.parameters.cudnn-version >>
           python-version: << pipeline.parameters.python-version >>
-

README.md

@@ -31,12 +31,7 @@ In the case of building on top of a custom base container, you first must determ
 version of the PyTorch C++ ABI. If your source of PyTorch is pytorch.org, likely this is the pre-cxx11-abi in which case you must modify `//docker/dist-build.sh` to not build the
 C++11 ABI version of Torch-TensorRT.
 
-You can then build the container using:
-
-
-```bash
-docker build --build-arg BASE_IMG=<IMAGE> -f docker/Dockerfile -t torch_tensorrt:latest .
-```
+You can then build the container using the build command in the [docker README](docker/README.md#instructions)
 
 If you would like to build outside a docker container, please follow the section [Compiling Torch-TensorRT](#compiling-torch-tensorrt)
 
@@ -73,6 +68,7 @@ import torch_tensorrt
 ...
 
 trt_ts_module = torch_tensorrt.compile(torch_script_module,
+    # If the inputs to the module are plain Tensors, specify them via the `inputs` argument:
     inputs = [example_tensor, # Provide example tensor for input shape or...
         torch_tensorrt.Input( # Specify input object with shape and dtype
             min_shape=[1, 3, 224, 224],
@@ -81,6 +77,12 @@ trt_ts_module = torch_tensorrt.compile(torch_script_module,
             # For static size shape=[1, 3, 224, 224]
             dtype=torch.half) # Datatype of input tensor. Allowed options torch.(float|half|int8|int32|bool)
     ],
+
+    # For inputs containing tuples or lists of tensors, use the `input_signature` argument:
+    # Below, we have an input consisting of a Tuple of two Tensors (Tuple[Tensor, Tensor])
+    # input_signature = ( (torch_tensorrt.Input(shape=[1, 3, 224, 224], dtype=torch.half),
+    #                      torch_tensorrt.Input(shape=[1, 3, 224, 224], dtype=torch.half)), ),
+
     enabled_precisions = {torch.half}, # Run with FP16
 )
 
@@ -114,17 +116,17 @@ torch.jit.save(trt_ts_module, "trt_torchscript_module.ts") # save the TRT embedd
 These are the following dependencies used to verify the testcases. Torch-TensorRT can work with other versions, but the tests are not guaranteed to pass.
 
 - Bazel 5.2.0
-- Libtorch 2.0.0.dev20230103 (built with CUDA 11.7)
-- CUDA 11.7
-- cuDNN 8.5.0
-- TensorRT 8.5.1.7
+- Libtorch 2.0.1 (built with CUDA 11.8)
+- CUDA 11.8
+- cuDNN 8.8.0
+- TensorRT 8.6.0
 
 ## Prebuilt Binaries and Wheel files
 
 Releases: https://github.com/pytorch/TensorRT/releases
 
 ```
-pip install torch-tensorrt==1.2.0 --find-links https://github.com/pytorch/TensorRT/releases/expanded_assets/v1.2.0
+pip install torch-tensorrt
 ```
 
 ## Compiling Torch-TensorRT
@@ -245,7 +247,7 @@ A tarball with the include files and library can then be found in bazel-bin
 ### Running Torch-TensorRT on a JIT Graph
 
 > Make sure to add LibTorch to your LD_LIBRARY_PATH <br>
-> `export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$(pwd)/bazel-Torch-TensorRT/external/libtorch/lib`
+> `export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$(pwd)/bazel-TensorRT/external/libtorch/lib`
 
 ``` shell
 bazel run //cpp/bin/torchtrtc -- $(realpath <PATH TO GRAPH>) out.ts <input-size>

WORKSPACE

@@ -51,17 +51,17 @@ new_local_repository(
 http_archive(
     name = "libtorch",
     build_file = "@//third_party/libtorch:BUILD",
-    sha256 = "292b3f81e7c857fc102be93e2e44c40cdb4d8ef03d98121bc6af434c66e8490b",
+    sha256 = "c5174f18c0866421a5738d389aaea0c02f32a1a5be5f0747dc8dd0d96034c9b0",
     strip_prefix = "libtorch",
-    urls = ["https://download.pytorch.org/libtorch/cu118/libtorch-cxx11-abi-shared-with-deps-2.0.0%2Bcu118.zip"],
+    urls = ["https://download.pytorch.org/libtorch/test/cu118/libtorch-cxx11-abi-shared-with-deps-latest.zip"],
 )
 
 http_archive(
     name = "libtorch_pre_cxx11_abi",
     build_file = "@//third_party/libtorch:BUILD",
-    sha256 = "f3cbd7e9593f0c64b8671d02a21d562c98b60ef1abf5898c0ee9acfbc5a6b5d2",
+    sha256 = "cc19b398cf435e0e34f347ef90fc11c2a42703998330a9c4a9fb0d2291737df7",
     strip_prefix = "libtorch",
-    urls = ["https://download.pytorch.org/libtorch/cu118/libtorch-shared-with-deps-2.0.0%2Bcu118.zip"],
+    urls = ["https://download.pytorch.org/libtorch/test/cu118/libtorch-shared-with-deps-latest.zip"],
 )
 
 # Download these tarballs manually from the NVIDIA website
@@ -71,10 +71,10 @@ http_archive(
 http_archive(
     name = "cudnn",
     build_file = "@//third_party/cudnn/archive:BUILD",
-    sha256 = "5454a6fd94f008728caae9adad993c4e85ef36302e26bce43bea7d458a5e7b6d",
-    strip_prefix = "cudnn-linux-x86_64-8.5.0.96_cuda11-archive",
+    sha256 = "36fff137153ef73e6ee10bfb07f4381240a86fb9fb78ce372414b528cbab2293",
+    strip_prefix = "cudnn-linux-x86_64-8.8.0.121_cuda11-archive",
     urls = [
-        "https://developer.nvidia.com/compute/cudnn/secure/8.5.0/local_installers/11.7/cudnn-linux-x86_64-8.5.0.96_cuda11-archive.tar.xz",
+        "https://developer.download.nvidia.com/compute/cudnn/secure/8.8.0/local_installers/11.8/cudnn-linux-x86_64-8.8.0.121_cuda11-archive.tar.xz",
     ],
 )
 

docker/Dockerfile

@@ -1,40 +1,54 @@
 # Base image starts with CUDA
-ARG BASE_IMG=nvidia/cuda:11.7.1-devel-ubuntu20.04
+ARG BASE_IMG=nvidia/cuda:11.8.0-devel-ubuntu22.04
 FROM ${BASE_IMG} as base
 
+ARG TENSORRT_VERSION
+RUN test -n "$TENSORRT_VERSION" || (echo "No tensorrt version specified, please use --build-arg TENSORRT_VERSION=x.y.z to specify a version." && exit 1)
+ARG CUDNN_VERSION
+RUN test -n "$CUDNN_VERSION" || (echo "No cudnn version specified, please use --build-arg CUDNN_VERSION=x.y.z to specify a version." && exit 1)
+
+ARG PYTHON_VERSION=3.10
+ENV PYTHON_VERSION=${PYTHON_VERSION}
+
+ARG USE_CXX11_ABI
+ENV USE_CXX11=${USE_CXX11_ABI}
+ENV DEBIAN_FRONTEND=noninteractive
+
 # Install basic dependencies
 RUN apt-get update
-RUN DEBIAN_FRONTEND=noninteractive apt install -y build-essential manpages-dev wget zlib1g software-properties-common git
-RUN add-apt-repository ppa:deadsnakes/ppa
-RUN apt install -y python3.8 python3.8-distutils python3.8-dev
-RUN wget https://bootstrap.pypa.io/get-pip.py
-RUN ln -s /usr/bin/python3.8 /usr/bin/python
-RUN python get-pip.py
-RUN pip3 install wheel
-
-# Install Pytorch
-RUN pip3 install torch==2.0.0.dev20230103+cu117 torchvision==0.15.0.dev20230103+cu117 --extra-index-url https://download.pytorch.org/whl/nightly/cu117
-
-# Install CUDNN + TensorRT
-RUN wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/cuda-ubuntu2004.pin
-RUN mv cuda-ubuntu2004.pin /etc/apt/preferences.d/cuda-repository-pin-600
-RUN apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/7fa2af80.pub
+RUN apt install -y build-essential manpages-dev wget zlib1g software-properties-common git libssl-dev zlib1g-dev libbz2-dev libreadline-dev libsqlite3-dev wget ca-certificates curl llvm libncurses5-dev xz-utils tk-dev libxml2-dev libxmlsec1-dev libffi-dev liblzma-dev mecab-ipadic-utf8
+
+# Install PyEnv and desired Python version
+ENV HOME="/root"
+ENV PYENV_DIR="$HOME/.pyenv"
+ENV PATH="$PYENV_DIR/shims:$PYENV_DIR/bin:$PATH"
+RUN wget -L https://github.com/pyenv/pyenv-installer/raw/master/bin/pyenv-installer &&\
+    chmod 755 pyenv-installer &&\
+    bash pyenv-installer &&\
+    eval "$(pyenv init -)"
+
+RUN pyenv install -v ${PYTHON_VERSION}
+RUN pyenv global ${PYTHON_VERSION}
+
+# Install CUDNN + TensorRT + dependencies
+RUN wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-ubuntu2204.pin
+RUN mv cuda-ubuntu2204.pin /etc/apt/preferences.d/cuda-repository-pin-600
+RUN apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/7fa2af80.pub
 RUN apt-key adv --keyserver keyserver.ubuntu.com --recv-keys 536F8F1DE80F6A35
 RUN apt-key adv --keyserver keyserver.ubuntu.com --recv-keys A4B469963BF863CC
-RUN add-apt-repository "deb https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/ /"
+RUN add-apt-repository "deb https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/ /"
 RUN apt-get update
-RUN apt-get install -y libcudnn8=8.5.0* libcudnn8-dev=8.5.0*
+RUN apt-get install -y libcudnn8=${CUDNN_VERSION}* libcudnn8-dev=${CUDNN_VERSION}*
 
-RUN apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/3bf863cc.pub
-RUN add-apt-repository "deb https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/ /"
+RUN apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/3bf863cc.pub
+RUN add-apt-repository "deb https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/ /"
 RUN apt-get update
 
-RUN apt-get install -y libnvinfer8=8.5.1* libnvinfer-plugin8=8.5.1* libnvinfer-dev=8.5.1* libnvinfer-plugin-dev=8.5.1* libnvonnxparsers8=8.5.1-1* libnvonnxparsers-dev=8.5.1-1* libnvparsers8=8.5.1-1*  libnvparsers-dev=8.5.1-1*
+RUN apt-get install -y libnvinfer8=${TENSORRT_VERSION}* libnvinfer-plugin8=${TENSORRT_VERSION}* libnvinfer-dev=${TENSORRT_VERSION}* libnvinfer-plugin-dev=${TENSORRT_VERSION}* libnvonnxparsers8=${TENSORRT_VERSION}-1* libnvonnxparsers-dev=${TENSORRT_VERSION}-1* libnvparsers8=${TENSORRT_VERSION}-1*  libnvparsers-dev=${TENSORRT_VERSION}-1*
 
-# Setup Bazel
-ARG BAZEL_VERSION=5.2.0
-RUN wget -q https://github.com/bazelbuild/bazel/releases/download/${BAZEL_VERSION}/bazel-${BAZEL_VERSION}-linux-x86_64 -O /usr/bin/bazel \
- && chmod a+x /usr/bin/bazel
+# Setup Bazel via Bazelisk
+RUN wget -q https://github.com/bazelbuild/bazelisk/releases/download/v1.16.0/bazelisk-linux-amd64 -O /usr/bin/bazel &&\
+    chmod a+x /usr/bin/bazel
 
 # Build Torch-TensorRT in an auxillary container
 FROM base as torch-tensorrt-builder-base
@@ -43,19 +57,24 @@ ARG ARCH="x86_64"
 ARG TARGETARCH="amd64"
 
 RUN apt-get install -y python3-setuptools
-RUN apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/3bf863cc.pub
-RUN  apt-get update
+RUN apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/3bf863cc.pub
 
-RUN apt-get update && apt-get install -y --no-install-recommends locales ninja-build && rm -rf /var/lib/apt/lists/* && locale-gen en_US.UTF-8
+RUN apt-get update &&\
+    apt-get install -y --no-install-recommends locales ninja-build &&\
+    rm -rf /var/lib/apt/lists/* &&\
+    locale-gen en_US.UTF-8
 
 FROM torch-tensorrt-builder-base as torch-tensorrt-builder
 
 COPY . /workspace/torch_tensorrt/src
 WORKDIR /workspace/torch_tensorrt/src
 RUN cp ./docker/WORKSPACE.docker WORKSPACE
 
+# Symlink the path pyenv is using for python with the /opt directory for package sourcing
+RUN ln -s "`pyenv which python | xargs dirname | xargs dirname`/lib/python$PYTHON_VERSION/site-packages" "/opt/python3"
+
 # This script builds both libtorchtrt bin/lib/include tarball and the Python wheel, in dist/
-RUN ./docker/dist-build.sh
+RUN bash ./docker/dist-build.sh
 
 # Copy and install Torch-TRT into the main container
 FROM base as torch-tensorrt
@@ -64,13 +83,13 @@ COPY . /opt/torch_tensorrt
 COPY --from=torch-tensorrt-builder  /workspace/torch_tensorrt/src/py/dist/ .
 
 RUN cp /opt/torch_tensorrt/docker/WORKSPACE.docker /opt/torch_tensorrt/WORKSPACE
-RUN pip3 install *.whl && rm -fr /workspace/torch_tensorrt/py/dist/* *.whl
-
-# Install native tensorrt python package required by torch_tensorrt whl file
-RUN pip install tensorrt==8.5.1.7
+RUN pip install -r /opt/torch_tensorrt/py/requirements.txt
+RUN pip install tensorrt==${TENSORRT_VERSION}.*
+RUN pip install *.whl && rm -fr /workspace/torch_tensorrt/py/dist/* *.whl
 
 WORKDIR /opt/torch_tensorrt
-ENV LD_LIBRARY_PATH /usr/local/lib/python3.8/dist-packages/torch/lib:/usr/local/lib/python3.8/dist-packages/torch_tensorrt/lib:/usr/lib/x86_64-linux-gnu:${LD_LIBRARY_PATH}
-ENV PATH /usr/local/lib/python3.8/dist-packages/torch_tensorrt/bin:${PATH}
+
+ENV LD_LIBRARY_PATH /opt/python3/site-packages/torch/lib:/opt/python3/site-packages/torch_tensorrt/lib:/usr/lib/x86_64-linux-gnu:${LD_LIBRARY_PATH}
+ENV PATH /opt/python3/site-packages/torch_tensorrt/bin:${PATH}
 
 CMD /bin/bash

docker/README.md

@@ -2,22 +2,29 @@
 
 * Use `Dockerfile` to build a container which provides the exact development environment that our master branch is usually tested against.
 
-* `Dockerfile` currently uses the exact library versions (Torch, CUDA, CUDNN, TensorRT) listed in <a href="https://github.com/pytorch/TensorRT#dependencies">dependencies</a> to build Torch-TensorRT.
+* The `Dockerfile` currently uses <a href="https://github.com/bazelbuild/bazelisk">Bazelisk</a> to select the Bazel version, and uses the exact library versions of Torch and CUDA listed in <a href="https://github.com/pytorch/TensorRT#dependencies">dependencies</a>.
+  * The desired versions of CUDNN and TensorRT must be specified as build-args, with major, minor, and patch versions as in: `--build-arg TENSORRT_VERSION=a.b.c --build-arg CUDNN_VERSION=x.y.z`
+  * [**Optional**] The desired base image be changed by explicitly setting a base image, as in `--build-arg BASE_IMG=nvidia/cuda:11.8.0-devel-ubuntu22.04`, though this is optional
+  * [**Optional**] Additionally, the desired Python version can be changed by explicitly setting a version, as in `--build-arg PYTHON_VERSION=3.10`, though this is optional as well.
 
 * This `Dockerfile` installs `pre-cxx11-abi` versions of Pytorch and builds Torch-TRT using `pre-cxx11-abi` libtorch as well.
-Note: To install `cxx11_abi` version of Torch-TensorRT, enable `USE_CXX11=1` flag so that `dist-build.sh` can build it accordingly.
+
+Note: By default the container uses the `pre-cxx11-abi` version of Torch + Torch-TRT. If you are using a workflow that requires a build of PyTorch on the CXX11 ABI (e.g. using the PyTorch NGC containers as a base image), add the Docker build argument: `--build-arg USE_CXX11_ABI=1`
 
 ### Dependencies
 
 * Install nvidia-docker by following https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html#docker
 
 ### Instructions
 
+- The example below uses CUDNN 8.8.0 and TensorRT 8.6.0
+- See <a href="https://github.com/pytorch/TensorRT#dependencies">dependencies</a> for a list of current default dependencies.
+
 > From root of Torch-TensorRT repo
 
 Build:
 ```
-DOCKER_BUILDKIT=1 docker build -f docker/Dockerfile -t torch_tensorrt:latest .
+DOCKER_BUILDKIT=1 docker build --build-arg TENSORRT_VERSION=8.6.0 --build-arg CUDNN_VERSION=8.8.0 -f docker/Dockerfile -t torch_tensorrt:latest .
 ```
 
 Run:
@@ -38,4 +45,4 @@ bazel test //tests/core/conversion/converters:test_activation --compilation_mode
 
 ### Pytorch NGC containers
 
-We also ship Torch-TensorRT in <a href="https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch">Pytorch NGC containers </a>. Release notes for these containers can be found <a href="https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/index.html">here</a>. Check out `release/ngc/23.XX` branch of Torch-TensorRT for source code that gets shipped with `23.XX` version of Pytorch NGC container.
+We also ship Torch-TensorRT in <a href="https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch">Pytorch NGC containers </a>. Release notes for these containers can be found <a href="https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/index.html">here</a>. Check out `release/ngc/23.XX` branch of Torch-TensorRT for source code that gets shipped with `23.XX` version of Pytorch NGC container.

docker/WORKSPACE.docker

@@ -48,20 +48,16 @@ new_local_repository(
 # Tarballs and fetched dependencies (default - use in cases when building from precompiled bin and tarballs)
 #############################################################################################################
 
-http_archive(
+new_local_repository(
     name = "libtorch",
-    build_file = "@//third_party/libtorch:BUILD",
-    sha256 = "59b8b5e1954a86d50b79c13f06398d385b200da13e37a08ecf31d3c62e5ca127",
-    strip_prefix = "libtorch",
-    urls = ["https://download.pytorch.org/libtorch/nightly/cu117/libtorch-cxx11-abi-shared-with-deps-2.0.0.dev20230103%2Bcu117.zip"],
+    path = "/opt/python3/site-packages/torch/",
+    build_file = "third_party/libtorch/BUILD"
 )
 
-http_archive(
+new_local_repository(
     name = "libtorch_pre_cxx11_abi",
-    build_file = "@//third_party/libtorch:BUILD",
-    sha256 = "e260fc7476be89d1650953e8643e9f7363845f5a52de4bab87ac0e619c1f6ad4",
-    strip_prefix = "libtorch",
-    urls = ["https://download.pytorch.org/libtorch/nightly/cu117/libtorch-shared-with-deps-2.0.0.dev20230103%2Bcu117.zip"],
+    path = "/opt/python3/site-packages/torch/",
+    build_file = "third_party/libtorch/BUILD"
 )
 
 ####################################################################################