Merge aeb970b into 80fd3b3

pytorch · Jul 11, 2022 · 49b629a · 49b629a
2 parents 80fd3b3 + aeb970b
commit 49b629a
Show file tree

Hide file tree

Showing 5 changed files with 49 additions and 20 deletions.
diff --git a/.github/workflows/runtime_nightly.yaml b/.github/workflows/runtime_nightly.yaml
@@ -3,6 +3,10 @@ name: Multipy runtime nightly release
 on:
   schedule:
     - cron: '0 2 * * *' # run at 2 AM UTC
+  pull_request:
+  push:
+    branches:
+      - main
 
 jobs:
   unittest:
@@ -11,6 +15,7 @@ jobs:
         python-version: [3.8]
         platform: [linux.2xlarge]
         abi: [0,1]
+        cuda: [0,1]
       fail-fast: false
     runs-on: ${{ matrix.platform }}
     steps:
@@ -27,6 +32,7 @@ jobs:
 
       - name: setup Path
         run: |
+          echo /usr/local/cuda-11.3/bin >> $GITHUB_PATH
           echo "/home/ec2-user/miniconda/bin" >> $GITHUB_PATH
           echo "CONDA=/home/ec2-user/miniconda" >> $GITHUB_PATH
 
@@ -75,7 +81,7 @@ jobs:
             conda run -n example_env python -m pip install torch torchvision torchaudio pathlib
             conda run -n example_env python generate_examples.py
 
-      - name: Build pytorch with ABI=${{ matrix.abi }}
+      - name: Build pytorch with ABI=${{ matrix.abi }} and USE_CUDA=${{ matrix.cuda }}
         shell: bash -l {0}
         env:
           PYTHON_VERSION: ${{ matrix.python-version }}
@@ -85,9 +91,10 @@ jobs:
               export TORCH_CXX_FLAGS="-D_GLIBCXX_USE_CXX11_ABI=${{ matrix.abi }}"
               cd multipy/runtime/third-party/pytorch
               export USE_DEPLOY=1
+              export USE_CUDA=${{ matrix.cuda }}
               conda run -n multipy_runtime_env python setup.py develop
 
-      - name: Build multipy runtime with ABI=${{ matrix.abi }}
+      - name: Build multipy runtime
         shell: bash -l {0}
         env:
           PYTHON_VERSION: ${{ matrix.python-version }}
@@ -98,22 +105,33 @@ jobs:
               conda run -n multipy_runtime_env cmake -DABI_EQUALS_1=${{ matrix.abi }} ..
               conda run -n multipy_runtime_env cmake --build . --config Release
 
-      - name: install files
+      - name: Install files
         shell: bash -l {0}
         env:
           PYTHON_VERSION: ${{ matrix.python-version }}
         run: |
               cd multipy/runtime/build
               conda run -n multipy_runtime_env cmake --install . --prefix "."
 
-      - name: Run unit tests with ABI=${{ matrix.abi }}
+      - name: Run unit tests with ABI=${{ matrix.abi }} with cuda=${{ matrix.cuda }}
         shell: bash -l {0}
         env:
           PYTHON_VERSION: ${{ matrix.python-version }}
         run: |
               cd multipy/runtime/build
               ./test_deploy
 
+      - name: Run unit tests with ABI=${{ matrix.abi }} with gpu
+        shell: bash -l {0}
+        env:
+          PYTHON_VERSION: ${{ matrix.python-version }}
+        run: |
+              if [[ ${{ matrix.cuda }} -eq 1 ]]
+              then
+              cd multipy/runtime/build
+              ./test_deploy_gpu
+              fi
+
       - name: create tarball [click me to get a list of files for the nightly release]
         shell: bash -l {0}
         env:
@@ -125,7 +143,7 @@ jobs:
       - name: Update nightly release
         uses: pyTooling/Actions/releaser@main
         with:
-          tag: nightly-runtime-abi-${{ matrix.abi }}
+          tag: nightly-runtime-abi-${{ matrix.abi }}-cuda-${{matrix.cuda}}
           rm: true
           token: ${{ secrets.GITHUB_TOKEN }}
           files: |

diff --git a/.github/workflows/runtime_tests.yaml b/.github/workflows/runtime_tests.yaml
@@ -13,6 +13,7 @@ jobs:
         python-version: [3.7, 3.8, 3.9]
         platform: [linux.2xlarge]
         abi: [0,1]
+        cuda: [0,1]
       fail-fast: false
     runs-on: ${{ matrix.platform }}
     steps:
@@ -30,6 +31,7 @@ jobs:
 
       - name: setup Path
         run: |
+          echo /usr/local/cuda-11.3/bin >> $GITHUB_PATH
           echo "/home/ec2-user/miniconda/bin" >> $GITHUB_PATH
           echo "CONDA=/home/ec2-user/miniconda" >> $GITHUB_PATH
 
@@ -78,7 +80,7 @@ jobs:
             conda run -n example_env python -m pip install torch torchvision torchaudio pathlib
             conda run -n example_env python generate_examples.py
 
-      - name: Build pytorch with ABI=${{ matrix.abi }}
+      - name: Build pytorch with ABI=${{ matrix.abi }} and USE_CUDA=${{ matrix.cuda }}
         shell: bash -l {0}
         env:
           PYTHON_VERSION: ${{ matrix.python-version }}
@@ -88,6 +90,7 @@ jobs:
               export TORCH_CXX_FLAGS="-D_GLIBCXX_USE_CXX11_ABI=${{ matrix.abi }}"
               cd multipy/runtime/third-party/pytorch
               export USE_DEPLOY=1
+              export USE_CUDA=${{ matrix.cuda }}
               conda run -n multipy_runtime_env python setup.py develop
 
       - name: Build multipy runtime
@@ -109,11 +112,20 @@ jobs:
               cd multipy/runtime/build
               conda run -n multipy_runtime_env cmake --install . --prefix "."
 
-      - name: Run unit tests with ABI=${{ matrix.abi }}
-        if: ${{ matrix.abi }} == 1
+      - name: Run unit tests with ABI=${{ matrix.abi }} with cuda=${{ matrix.cuda }}
         shell: bash -l {0}
         env:
           PYTHON_VERSION: ${{ matrix.python-version }}
         run: |
               cd multipy/runtime/build
               ./test_deploy
+      - name: Run unit tests with ABI=${{ matrix.abi }} with gpu
+        shell: bash -l {0}
+        env:
+          PYTHON_VERSION: ${{ matrix.python-version }}
+        run: |
+              if [[ ${{ matrix.cuda }} -eq 1 ]]
+              then
+              cd multipy/runtime/build
+              ./test_deploy_gpu
+              fi
diff --git a/README.md b/README.md
@@ -13,7 +13,9 @@ internally, please see the related [arXiv paper](https://arxiv.org/pdf/2104.0025
 
 ### Installing `multipy::runtime` **(recommended)**
 
-The C++ binaries (`libtorch_interpreter.so`,`libtorch_deploy.a`, `utils.cmake`), and the header files of `multipy::runtime` can be installed from our [nightly release](https://github.com/pytorch/multipy/releases/tag/nightly-runtime-abi-0). The ABI for the nightly release is 0. You can find a version of the release with ABI=1 [here](https://github.com/pytorch/multipy/releases/tag/nightly-runtime-abi-1).
+The C++ binaries (`libtorch_interpreter.so`,`libtorch_deploy.a`, `utils.cmake`), and the header files of `multipy::runtime` can be installed from our [nightly release](https://github.com/pytorch/multipy/releases/tag/nightly-runtime-abi-0-cuda-0) (uses the pre-cxx11 ABI). You can find a version of the release with the [cxx11 ABI](https://github.com/pytorch/multipy/releases/tag/nightly-runtime-abi-1-cuda-0.
+
+C++ binaries with cuda (11.3) support can also be found for the [pre-cxx11 ABI](https://github.com/pytorch/multipy/releases/tag/nightly-runtime-cuda-abi-0-cuda-1) and the [cxx11 ABI](https://github.com/pytorch/multipy/releases/tag/nightly-runtime-cuda-abi-1-cuda-1).
 
 ```
 wget https://github.com/pytorch/multipy/releases/download/nightly-runtime-abi-0/multipy_runtime.tar.gz

diff --git a/multipy/runtime/CMakeLists.txt b/multipy/runtime/CMakeLists.txt
@@ -86,14 +86,12 @@ target_link_libraries(test_deploy
 )
 target_include_directories(test_deploy PRIVATE ${CMAKE_SOURCE_DIR}/../..)
 
-#  LINK_DIRECTORIES("${PYTORCH_ROOT}/torch/lib")
-#  add_executable(test_deploy_gpu ${INTERPRETER_TEST_SOURCES_GPU})
-#  target_compile_definitions(test_deploy_gpu PUBLIC TEST_CUSTOM_LIBRARY)
-#  target_include_directories(test_deploy_gpu PRIVATE ${PYTORCH_ROOT}/torch)
-#  target_include_directories(test_deploy_gpu PRIVATE ${CMAKE_SOURCE_DIR}/../..)
-#  target_link_libraries(test_deploy_gpu
-#   PUBLIC "-Wl,--no-as-needed -rdynamic" gtest dl torch_deploy_interface c10 torch_cpu
-# )
+LINK_DIRECTORIES("${PYTORCH_ROOT}/torch/lib")
+add_executable(test_deploy_gpu ${INTERPRETER_TEST_SOURCES_GPU})
+target_compile_definitions(test_deploy_gpu PUBLIC TEST_CUSTOM_LIBRARY)
+target_include_directories(test_deploy_gpu PRIVATE ${PYTORCH_ROOT}/torch)
+target_include_directories(test_deploy_gpu PRIVATE ${CMAKE_SOURCE_DIR}/../..)
+target_link_libraries(test_deploy_gpu PUBLIC "-Wl,--no-as-needed -rdynamic" gtest dl torch_deploy_interface c10 torch_cpu)
 
 LINK_DIRECTORIES("${PYTORCH_ROOT}/torch/lib")
 add_library(test_deploy_lib SHARED test_deploy_lib.cpp)
@@ -119,7 +117,7 @@ target_link_libraries(interactive_embedded_interpreter
 )
 
 install(TARGETS test_deploy DESTINATION tests/bin)
-# install(TARGETS test_deploy_gpu DESTINATION tests/bin)
+install(TARGETS test_deploy_gpu DESTINATION tests/bin)
 
 install(TARGETS test_deploy DESTINATION tests/bin)
 

diff --git a/multipy/runtime/test_deploy_gpu.cpp b/multipy/runtime/test_deploy_gpu.cpp
@@ -63,8 +63,7 @@ TEST(TorchDeployGPUTest, SimpleModel) {
 
 TEST(TorchDeployGPUTest, UsesDistributed) {
   const auto model_filename = path(
-      "USES_DISTRIBUTED",
-      "torch/csrc/deploy/example/generated/uses_distributed");
+      "USES_DISTRIBUTED", "multipy/runtime/example/generated/uses_distributed");
   torch::deploy::InterpreterManager m(1);
   torch::deploy::Package p = m.loadPackage(model_filename);
   {