Revert "Fix handling of non-finite values in topk (#35253)" (#35582)

This reverts commit b12579da5398ff23b421332e21e18dc619a0b960.

This patch in-and-of itself looks fine, but it's causing some AMP tests to fail.

.circleci/cimodel/data/binary_build_data.py

@@ -34,8 +34,6 @@ def get_processor_arch_name(cuda_version):
 
 LINUX_PACKAGE_VARIANTS = OrderedDict(
     manywheel=[
-        "2.7m",
-        "2.7mu",
         "3.5m",
         "3.6m",
         "3.7m",

.circleci/config.yml

@@ -351,16 +351,16 @@ jobs:
           export id=$(docker run --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -t -d -w /var/lib/jenkins ${DOCKER_IMAGE})
 
           # TODO We may want to move the rebase logic to a separate step after checkout
-          # Rebase to master only if in xenial_py3_6_gcc5_4 case
-          if [[ "${CIRCLE_BRANCH}" != "master" && "${BUILD_ENVIRONMENT}" == *"gcc5"* ]]; then
-            echo "Merge master branch into $CIRCLE_BRANCH before build in environment $BUILD_ENVIRONMENT"
+          # Rebase to release/1.5 only if in xenial_py3_6_gcc5_4 case
+          if [[ "${CIRCLE_BRANCH}" != "release/1.5" && "${BUILD_ENVIRONMENT}" == *"gcc5"* ]]; then
+            echo "Merge release/1.5 branch into $CIRCLE_BRANCH before build in environment $BUILD_ENVIRONMENT"
             set -x
             git config --global user.email "circleci.ossci@gmail.com"
             git config --global user.name "CircleCI"
             git config remote.origin.url https://github.com/pytorch/pytorch.git
-            git config --add remote.origin.fetch +refs/heads/master:refs/remotes/origin/master
-            git fetch --tags --progress https://github.com/pytorch/pytorch.git +refs/heads/master:refs/remotes/origin/master --depth=100 --quiet
-            export GIT_MERGE_TARGET=`git log -n 1 --pretty=format:"%H" origin/master`
+            git config --add remote.origin.fetch +refs/heads/release/1.5:refs/remotes/origin/release/1.5
+            git fetch --tags --progress https://github.com/pytorch/pytorch.git +refs/heads/release/1.5:refs/remotes/origin/release/1.5 --depth=100 --quiet
+            export GIT_MERGE_TARGET=`git log -n 1 --pretty=format:"%H" origin/release/1.5`
             echo "GIT_MERGE_TARGET: " ${GIT_MERGE_TARGET}
             export GIT_COMMIT=${CIRCLE_SHA1}
             echo "GIT_COMMIT: " ${GIT_COMMIT}
@@ -369,7 +369,7 @@ jobs:
             git merge --allow-unrelated-histories --no-edit --no-ff ${GIT_MERGE_TARGET}
             set +x
           else
-            echo "Do NOT merge master branch into $CIRCLE_BRANCH in environment $BUILD_ENVIRONMENT"
+            echo "Do NOT merge release/1.5 branch into $CIRCLE_BRANCH in environment $BUILD_ENVIRONMENT"
           fi
 
           git submodule sync && git submodule update -q --init --recursive
@@ -2544,28 +2544,6 @@ workflows:
           libtorch_variant: "shared-with-deps"
           docker_image: "pytorch/pytorch-binary-docker-image-ubuntu16.04:latest"
 
-      - smoke_linux_test:
-          name: smoke_linux_manywheel_2_7m_cpu_devtoolset7_nightly
-          build_environment: "manywheel 2.7m cpu devtoolset7"
-          requires:
-            - setup
-            - update_s3_htmls_for_nightlies
-            - update_s3_htmls_for_nightlies_devtoolset7
-          filters:
-            branches:
-              only: postnightly
-          docker_image: "pytorch/manylinux-cuda102"
-      - smoke_linux_test:
-          name: smoke_linux_manywheel_2_7mu_cpu_devtoolset7_nightly
-          build_environment: "manywheel 2.7mu cpu devtoolset7"
-          requires:
-            - setup
-            - update_s3_htmls_for_nightlies
-            - update_s3_htmls_for_nightlies_devtoolset7
-          filters:
-            branches:
-              only: postnightly
-          docker_image: "pytorch/manylinux-cuda102"
       - smoke_linux_test:
           name: smoke_linux_manywheel_3_5m_cpu_devtoolset7_nightly
           build_environment: "manywheel 3.5m cpu devtoolset7"
@@ -2610,32 +2588,6 @@ workflows:
             branches:
               only: postnightly
           docker_image: "pytorch/manylinux-cuda102"
-      - smoke_linux_test:
-          name: smoke_linux_manywheel_2_7m_cu92_devtoolset7_nightly
-          build_environment: "manywheel 2.7m cu92 devtoolset7"
-          requires:
-            - setup
-            - update_s3_htmls_for_nightlies
-            - update_s3_htmls_for_nightlies_devtoolset7
-          filters:
-            branches:
-              only: postnightly
-          docker_image: "pytorch/manylinux-cuda92"
-          use_cuda_docker_runtime: "1"
-          resource_class: gpu.medium
-      - smoke_linux_test:
-          name: smoke_linux_manywheel_2_7mu_cu92_devtoolset7_nightly
-          build_environment: "manywheel 2.7mu cu92 devtoolset7"
-          requires:
-            - setup
-            - update_s3_htmls_for_nightlies
-            - update_s3_htmls_for_nightlies_devtoolset7
-          filters:
-            branches:
-              only: postnightly
-          docker_image: "pytorch/manylinux-cuda92"
-          use_cuda_docker_runtime: "1"
-          resource_class: gpu.medium
       - smoke_linux_test:
           name: smoke_linux_manywheel_3_5m_cu92_devtoolset7_nightly
           build_environment: "manywheel 3.5m cu92 devtoolset7"
@@ -2688,32 +2640,6 @@ workflows:
           docker_image: "pytorch/manylinux-cuda92"
           use_cuda_docker_runtime: "1"
           resource_class: gpu.medium
-      - smoke_linux_test:
-          name: smoke_linux_manywheel_2_7m_cu101_devtoolset7_nightly
-          build_environment: "manywheel 2.7m cu101 devtoolset7"
-          requires:
-            - setup
-            - update_s3_htmls_for_nightlies
-            - update_s3_htmls_for_nightlies_devtoolset7
-          filters:
-            branches:
-              only: postnightly
-          docker_image: "pytorch/manylinux-cuda101"
-          use_cuda_docker_runtime: "1"
-          resource_class: gpu.medium
-      - smoke_linux_test:
-          name: smoke_linux_manywheel_2_7mu_cu101_devtoolset7_nightly
-          build_environment: "manywheel 2.7mu cu101 devtoolset7"
-          requires:
-            - setup
-            - update_s3_htmls_for_nightlies
-            - update_s3_htmls_for_nightlies_devtoolset7
-          filters:
-            branches:
-              only: postnightly
-          docker_image: "pytorch/manylinux-cuda101"
-          use_cuda_docker_runtime: "1"
-          resource_class: gpu.medium
       - smoke_linux_test:
           name: smoke_linux_manywheel_3_5m_cu101_devtoolset7_nightly
           build_environment: "manywheel 3.5m cu101 devtoolset7"
@@ -2766,32 +2692,6 @@ workflows:
           docker_image: "pytorch/manylinux-cuda101"
           use_cuda_docker_runtime: "1"
           resource_class: gpu.medium
-      - smoke_linux_test:
-          name: smoke_linux_manywheel_2_7m_cu102_devtoolset7_nightly
-          build_environment: "manywheel 2.7m cu102 devtoolset7"
-          requires:
-            - setup
-            - update_s3_htmls_for_nightlies
-            - update_s3_htmls_for_nightlies_devtoolset7
-          filters:
-            branches:
-              only: postnightly
-          docker_image: "pytorch/manylinux-cuda102"
-          use_cuda_docker_runtime: "1"
-          resource_class: gpu.medium
-      - smoke_linux_test:
-          name: smoke_linux_manywheel_2_7mu_cu102_devtoolset7_nightly
-          build_environment: "manywheel 2.7mu cu102 devtoolset7"
-          requires:
-            - setup
-            - update_s3_htmls_for_nightlies
-            - update_s3_htmls_for_nightlies_devtoolset7
-          filters:
-            branches:
-              only: postnightly
-          docker_image: "pytorch/manylinux-cuda102"
-          use_cuda_docker_runtime: "1"
-          resource_class: gpu.medium
       - smoke_linux_test:
           name: smoke_linux_manywheel_3_5m_cu102_devtoolset7_nightly
           build_environment: "manywheel 3.5m cu102 devtoolset7"
@@ -3636,28 +3536,6 @@ workflows:
           filters:
             branches:
               only: postnightly
-      - binary_linux_build:
-          name: binary_linux_manywheel_2_7m_cpu_devtoolset7_nightly_build
-          build_environment: "manywheel 2.7m cpu devtoolset7"
-          requires:
-            - setup
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          docker_image: "pytorch/manylinux-cuda102"
-      - binary_linux_build:
-          name: binary_linux_manywheel_2_7mu_cpu_devtoolset7_nightly_build
-          build_environment: "manywheel 2.7mu cpu devtoolset7"
-          requires:
-            - setup
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          docker_image: "pytorch/manylinux-cuda102"
       - binary_linux_build:
           name: binary_linux_manywheel_3_5m_cpu_devtoolset7_nightly_build
           build_environment: "manywheel 3.5m cpu devtoolset7"
@@ -3702,28 +3580,6 @@ workflows:
             tags:
               only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
           docker_image: "pytorch/manylinux-cuda102"
-      - binary_linux_build:
-          name: binary_linux_manywheel_2_7m_cu92_devtoolset7_nightly_build
-          build_environment: "manywheel 2.7m cu92 devtoolset7"
-          requires:
-            - setup
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          docker_image: "pytorch/manylinux-cuda92"
-      - binary_linux_build:
-          name: binary_linux_manywheel_2_7mu_cu92_devtoolset7_nightly_build
-          build_environment: "manywheel 2.7mu cu92 devtoolset7"
-          requires:
-            - setup
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          docker_image: "pytorch/manylinux-cuda92"
       - binary_linux_build:
           name: binary_linux_manywheel_3_5m_cu92_devtoolset7_nightly_build
           build_environment: "manywheel 3.5m cu92 devtoolset7"
@@ -3768,28 +3624,6 @@ workflows:
             tags:
               only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
           docker_image: "pytorch/manylinux-cuda92"
-      - binary_linux_build:
-          name: binary_linux_manywheel_2_7m_cu101_devtoolset7_nightly_build
-          build_environment: "manywheel 2.7m cu101 devtoolset7"
-          requires:
-            - setup
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          docker_image: "pytorch/manylinux-cuda101"
-      - binary_linux_build:
-          name: binary_linux_manywheel_2_7mu_cu101_devtoolset7_nightly_build
-          build_environment: "manywheel 2.7mu cu101 devtoolset7"
-          requires:
-            - setup
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          docker_image: "pytorch/manylinux-cuda101"
       - binary_linux_build:
           name: binary_linux_manywheel_3_5m_cu101_devtoolset7_nightly_build
           build_environment: "manywheel 3.5m cu101 devtoolset7"
@@ -3834,28 +3668,6 @@ workflows:
             tags:
               only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
           docker_image: "pytorch/manylinux-cuda101"
-      - binary_linux_build:
-          name: binary_linux_manywheel_2_7m_cu102_devtoolset7_nightly_build
-          build_environment: "manywheel 2.7m cu102 devtoolset7"
-          requires:
-            - setup
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          docker_image: "pytorch/manylinux-cuda102"
-      - binary_linux_build:
-          name: binary_linux_manywheel_2_7mu_cu102_devtoolset7_nightly_build
-          build_environment: "manywheel 2.7mu cu102 devtoolset7"
-          requires:
-            - setup
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          docker_image: "pytorch/manylinux-cuda102"
       - binary_linux_build:
           name: binary_linux_manywheel_3_5m_cu102_devtoolset7_nightly_build
           build_environment: "manywheel 3.5m cu102 devtoolset7"
@@ -4706,30 +4518,6 @@ workflows:
 ##############################################################################
 # Nightly tests
 ##############################################################################
-      - binary_linux_test:
-          name: binary_linux_manywheel_2_7m_cpu_devtoolset7_nightly_test
-          build_environment: "manywheel 2.7m cpu devtoolset7"
-          requires:
-            - setup
-            - binary_linux_manywheel_2_7m_cpu_devtoolset7_nightly_build
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          docker_image: "pytorch/manylinux-cuda102"
-      - binary_linux_test:
-          name: binary_linux_manywheel_2_7mu_cpu_devtoolset7_nightly_test
-          build_environment: "manywheel 2.7mu cpu devtoolset7"
-          requires:
-            - setup
-            - binary_linux_manywheel_2_7mu_cpu_devtoolset7_nightly_build
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          docker_image: "pytorch/manylinux-cuda102"
       - binary_linux_test:
           name: binary_linux_manywheel_3_5m_cpu_devtoolset7_nightly_test
           build_environment: "manywheel 3.5m cpu devtoolset7"
@@ -4778,34 +4566,6 @@ workflows:
             tags:
               only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
           docker_image: "pytorch/manylinux-cuda102"
-      - binary_linux_test:
-          name: binary_linux_manywheel_2_7m_cu92_devtoolset7_nightly_test
-          build_environment: "manywheel 2.7m cu92 devtoolset7"
-          requires:
-            - setup
-            - binary_linux_manywheel_2_7m_cu92_devtoolset7_nightly_build
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          docker_image: "pytorch/manylinux-cuda92"
-          use_cuda_docker_runtime: "1"
-          resource_class: gpu.medium
-      - binary_linux_test:
-          name: binary_linux_manywheel_2_7mu_cu92_devtoolset7_nightly_test
-          build_environment: "manywheel 2.7mu cu92 devtoolset7"
-          requires:
-            - setup
-            - binary_linux_manywheel_2_7mu_cu92_devtoolset7_nightly_build
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          docker_image: "pytorch/manylinux-cuda92"
-          use_cuda_docker_runtime: "1"
-          resource_class: gpu.medium
       - binary_linux_test:
           name: binary_linux_manywheel_3_5m_cu92_devtoolset7_nightly_test
           build_environment: "manywheel 3.5m cu92 devtoolset7"
@@ -4862,34 +4622,6 @@ workflows:
           docker_image: "pytorch/manylinux-cuda92"
           use_cuda_docker_runtime: "1"
           resource_class: gpu.medium
-      - binary_linux_test:
-          name: binary_linux_manywheel_2_7m_cu101_devtoolset7_nightly_test
-          build_environment: "manywheel 2.7m cu101 devtoolset7"
-          requires:
-            - setup
-            - binary_linux_manywheel_2_7m_cu101_devtoolset7_nightly_build
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          docker_image: "pytorch/manylinux-cuda101"
-          use_cuda_docker_runtime: "1"
-          resource_class: gpu.medium
-      - binary_linux_test:
-          name: binary_linux_manywheel_2_7mu_cu101_devtoolset7_nightly_test
-          build_environment: "manywheel 2.7mu cu101 devtoolset7"
-          requires:
-            - setup
-            - binary_linux_manywheel_2_7mu_cu101_devtoolset7_nightly_build
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          docker_image: "pytorch/manylinux-cuda101"
-          use_cuda_docker_runtime: "1"
-          resource_class: gpu.medium
       - binary_linux_test:
           name: binary_linux_manywheel_3_5m_cu101_devtoolset7_nightly_test
           build_environment: "manywheel 3.5m cu101 devtoolset7"
@@ -4946,34 +4678,6 @@ workflows:
           docker_image: "pytorch/manylinux-cuda101"
           use_cuda_docker_runtime: "1"
           resource_class: gpu.medium
-      - binary_linux_test:
-          name: binary_linux_manywheel_2_7m_cu102_devtoolset7_nightly_test
-          build_environment: "manywheel 2.7m cu102 devtoolset7"
-          requires:
-            - setup
-            - binary_linux_manywheel_2_7m_cu102_devtoolset7_nightly_build
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          docker_image: "pytorch/manylinux-cuda102"
-          use_cuda_docker_runtime: "1"
-          resource_class: gpu.medium
-      - binary_linux_test:
-          name: binary_linux_manywheel_2_7mu_cu102_devtoolset7_nightly_test
-          build_environment: "manywheel 2.7mu cu102 devtoolset7"
-          requires:
-            - setup
-            - binary_linux_manywheel_2_7mu_cu102_devtoolset7_nightly_build
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          docker_image: "pytorch/manylinux-cuda102"
-          use_cuda_docker_runtime: "1"
-          resource_class: gpu.medium
       - binary_linux_test:
           name: binary_linux_manywheel_3_5m_cu102_devtoolset7_nightly_test
           build_environment: "manywheel 3.5m cu102 devtoolset7"
@@ -5772,30 +5476,6 @@ workflows:
       #      - binary_linux_libtorch_2.7m_cu90_build
 
       # Nightly uploads
-      - binary_linux_upload:
-          name: binary_linux_manywheel_2_7m_cpu_devtoolset7_nightly_upload
-          build_environment: "manywheel 2.7m cpu devtoolset7"
-          requires:
-            - setup
-            - binary_linux_manywheel_2_7m_cpu_devtoolset7_nightly_test
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          context: org-member
-      - binary_linux_upload:
-          name: binary_linux_manywheel_2_7mu_cpu_devtoolset7_nightly_upload
-          build_environment: "manywheel 2.7mu cpu devtoolset7"
-          requires:
-            - setup
-            - binary_linux_manywheel_2_7mu_cpu_devtoolset7_nightly_test
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          context: org-member
       - binary_linux_upload:
           name: binary_linux_manywheel_3_5m_cpu_devtoolset7_nightly_upload
           build_environment: "manywheel 3.5m cpu devtoolset7"
@@ -5844,30 +5524,6 @@ workflows:
             tags:
               only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
           context: org-member
-      - binary_linux_upload:
-          name: binary_linux_manywheel_2_7m_cu92_devtoolset7_nightly_upload
-          build_environment: "manywheel 2.7m cu92 devtoolset7"
-          requires:
-            - setup
-            - binary_linux_manywheel_2_7m_cu92_devtoolset7_nightly_test
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          context: org-member
-      - binary_linux_upload:
-          name: binary_linux_manywheel_2_7mu_cu92_devtoolset7_nightly_upload
-          build_environment: "manywheel 2.7mu cu92 devtoolset7"
-          requires:
-            - setup
-            - binary_linux_manywheel_2_7mu_cu92_devtoolset7_nightly_test
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          context: org-member
       - binary_linux_upload:
           name: binary_linux_manywheel_3_5m_cu92_devtoolset7_nightly_upload
           build_environment: "manywheel 3.5m cu92 devtoolset7"
@@ -5916,30 +5572,6 @@ workflows:
             tags:
               only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
           context: org-member
-      - binary_linux_upload:
-          name: binary_linux_manywheel_2_7m_cu101_devtoolset7_nightly_upload
-          build_environment: "manywheel 2.7m cu101 devtoolset7"
-          requires:
-            - setup
-            - binary_linux_manywheel_2_7m_cu101_devtoolset7_nightly_test
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          context: org-member
-      - binary_linux_upload:
-          name: binary_linux_manywheel_2_7mu_cu101_devtoolset7_nightly_upload
-          build_environment: "manywheel 2.7mu cu101 devtoolset7"
-          requires:
-            - setup
-            - binary_linux_manywheel_2_7mu_cu101_devtoolset7_nightly_test
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          context: org-member
       - binary_linux_upload:
           name: binary_linux_manywheel_3_5m_cu101_devtoolset7_nightly_upload
           build_environment: "manywheel 3.5m cu101 devtoolset7"
@@ -5988,30 +5620,6 @@ workflows:
             tags:
               only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
           context: org-member
-      - binary_linux_upload:
-          name: binary_linux_manywheel_2_7m_cu102_devtoolset7_nightly_upload
-          build_environment: "manywheel 2.7m cu102 devtoolset7"
-          requires:
-            - setup
-            - binary_linux_manywheel_2_7m_cu102_devtoolset7_nightly_test
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          context: org-member
-      - binary_linux_upload:
-          name: binary_linux_manywheel_2_7mu_cu102_devtoolset7_nightly_upload
-          build_environment: "manywheel 2.7mu cu102 devtoolset7"
-          requires:
-            - setup
-            - binary_linux_manywheel_2_7mu_cu102_devtoolset7_nightly_test
-          filters:
-            branches:
-              only: nightly
-            tags:
-              only: /v[0-9]+(\.[0-9]+)*-rc[0-9]+/
-          context: org-member
       - binary_linux_upload:
           name: binary_linux_manywheel_3_5m_cu102_devtoolset7_nightly_upload
           build_environment: "manywheel 3.5m cu102 devtoolset7"

.circleci/scripts/binary_populate_env.sh

@@ -2,6 +2,19 @@
 set -eux -o pipefail
 export TZ=UTC
 
+tagged_version() {
+  # Grabs version from either the env variable CIRCLE_TAG
+  # or the pytorch git described version
+  GIT_DESCRIBE="git --git-dir ${workdir}/pytorch/.git describe"
+  if [[ -n "${CIRCLE_TAG:-}" ]]; then
+    echo "${CIRCLE_TAG}"
+  elif ${GIT_DESCRIBE} --exact --tags >/dev/null; then
+    ${GIT_DESCRIBE} --tags
+  else
+    return 1
+  fi
+}
+
 # We need to write an envfile to persist these variables to following
 # steps, but the location of the envfile depends on the circleci executor
 if [[ "$(uname)" == Darwin ]]; then
@@ -47,15 +60,17 @@ export DATE="$(date -u +%Y%m%d)"
 #TODO: We should be pulling semver version from the base version.txt
 BASE_BUILD_VERSION="1.5.0.dev$DATE"
 # Change BASE_BUILD_VERSION to git tag when on a git tag
-if git describe --tags --exact >/dev/null 2>/dev/null; then
+# Use 'git -C' to make doubly sure we're in the correct directory for checking
+# the git tag
+if tagged_version >/dev/null; then
   # Switch upload folder to 'test/' if we are on a tag
   PIP_UPLOAD_FOLDER='test/'
   # Grab git tag, remove prefixed v and remove everything after -
   # Used to clean up tags that are for release candidates like v1.5.0-rc1
   # Turns tag v1.5.0-rc1 -> v1.5.0
-  BASE_BUILD_VERSION="$(git describe --tags | sed -e 's/^v//' -e 's/-.*$//')"
+  BASE_BUILD_VERSION="$(tagged_version | sed -e 's/^v//' -e 's/-.*$//')"
 fi
-if [[ "$(uname)" == 'Darwin' ]] || [[ "$DESIRED_CUDA" == "cu101" ]] || [[ "$PACKAGE_TYPE" == conda ]]; then
+if [[ "$(uname)" == 'Darwin' ]] || [[ "$DESIRED_CUDA" == "cu102" ]] || [[ "$PACKAGE_TYPE" == conda ]]; then
   export PYTORCH_BUILD_VERSION="${BASE_BUILD_VERSION}"
 else
   export PYTORCH_BUILD_VERSION="${BASE_BUILD_VERSION}+$DESIRED_CUDA"

.circleci/verbatim-sources/pytorch-job-specs.yml

@@ -20,16 +20,16 @@ jobs:
           export id=$(docker run --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -t -d -w /var/lib/jenkins ${DOCKER_IMAGE})
 
           # TODO We may want to move the rebase logic to a separate step after checkout
-          # Rebase to master only if in xenial_py3_6_gcc5_4 case
-          if [[ "${CIRCLE_BRANCH}" != "master" && "${BUILD_ENVIRONMENT}" == *"gcc5"* ]]; then
-            echo "Merge master branch into $CIRCLE_BRANCH before build in environment $BUILD_ENVIRONMENT"
+          # Rebase to release/1.5 only if in xenial_py3_6_gcc5_4 case
+          if [[ "${CIRCLE_BRANCH}" != "release/1.5" && "${BUILD_ENVIRONMENT}" == *"gcc5"* ]]; then
+            echo "Merge release/1.5 branch into $CIRCLE_BRANCH before build in environment $BUILD_ENVIRONMENT"
             set -x
             git config --global user.email "circleci.ossci@gmail.com"
             git config --global user.name "CircleCI"
             git config remote.origin.url https://github.com/pytorch/pytorch.git
-            git config --add remote.origin.fetch +refs/heads/master:refs/remotes/origin/master
-            git fetch --tags --progress https://github.com/pytorch/pytorch.git +refs/heads/master:refs/remotes/origin/master --depth=100 --quiet
-            export GIT_MERGE_TARGET=`git log -n 1 --pretty=format:"%H" origin/master`
+            git config --add remote.origin.fetch +refs/heads/release/1.5:refs/remotes/origin/release/1.5
+            git fetch --tags --progress https://github.com/pytorch/pytorch.git +refs/heads/release/1.5:refs/remotes/origin/release/1.5 --depth=100 --quiet
+            export GIT_MERGE_TARGET=`git log -n 1 --pretty=format:"%H" origin/release/1.5`
             echo "GIT_MERGE_TARGET: " ${GIT_MERGE_TARGET}
             export GIT_COMMIT=${CIRCLE_SHA1}
             echo "GIT_COMMIT: " ${GIT_COMMIT}
@@ -38,7 +38,7 @@ jobs:
             git merge --allow-unrelated-histories --no-edit --no-ff ${GIT_MERGE_TARGET}
             set +x
           else
-            echo "Do NOT merge master branch into $CIRCLE_BRANCH in environment $BUILD_ENVIRONMENT"
+            echo "Do NOT merge release/1.5 branch into $CIRCLE_BRANCH in environment $BUILD_ENVIRONMENT"
           fi
 
           git submodule sync && git submodule update -q --init --recursive

.jenkins/pytorch/build.sh

@@ -167,7 +167,7 @@ fi
 
 # Patch required to build xla
 if [[ "${BUILD_ENVIRONMENT}" == *xla* ]]; then
-  git clone --recursive https://github.com/pytorch/xla.git
+  git clone --recursive -b r1.5 https://github.com/pytorch/xla.git
   ./xla/scripts/apply_patches.sh
 fi
 

.jenkins/pytorch/macos-common.sh

@@ -18,7 +18,7 @@ if [ ! -d "${WORKSPACE_DIR}/miniconda3" ]; then
 fi
 export PATH="${WORKSPACE_DIR}/miniconda3/bin:$PATH"
 source ${WORKSPACE_DIR}/miniconda3/bin/activate
-retry conda install -y mkl mkl-include numpy pyyaml setuptools cmake cffi ninja
+retry conda install -y mkl mkl-include numpy pyyaml=5.3 setuptools=46.0.0 cmake cffi ninja
 
 # The torch.hub tests make requests to GitHub.
 #

aten/src/ATen/Utils.h

@@ -16,14 +16,6 @@
 #include <numeric>
 #include <memory>
 
-#if defined(__clang__)
-#define __ubsan_ignore_float_divide_by_zero__ __attribute__((no_sanitize("float-divide-by-zero")))
-#define __ubsan_ignore_vptr__ __attribute__((no_sanitize("vptr")))
-#else
-#define __ubsan_ignore_float_divide_by_zero__
-#define __ubsan_ignore_vptr__
-#endif
-
 #define AT_DISALLOW_COPY_AND_ASSIGN(TypeName) \
   TypeName(const TypeName&) = delete; \
   void operator=(const TypeName&) = delete

aten/src/ATen/core/custom_class.cpp

@@ -20,6 +20,10 @@ void registerCustomClass(at::ClassTypePtr class_type) {
 }
 
 at::ClassTypePtr getCustomClass(const std::string& name) {
+  // BC hack so we can upgrade a binary internally
+  if (name == "__torch__.torch.classes.SentencePiece") {
+    return getCustomClass("__torch__.torch.classes.fb.SentencePiece");
+  }
   return customClasses().count(name) ? customClasses()[name] : nullptr;
 }
 

aten/src/ATen/cpu/vec256/vec256_base.h

@@ -15,6 +15,7 @@
 #include <c10/util/math_compat.h>
 #include <ATen/native/cpu/zmath.h>
 #include <c10/util/TypeCast.h>
+#include <c10/macros/Macros.h>
 
 #if defined(__GNUC__)
 #define __at_align32__ __attribute__((aligned(32)))

aten/src/ATen/cudnn/Descriptors.h

@@ -145,7 +145,7 @@ private:
 
 std::ostream& operator<<(std::ostream & out, const TensorDescriptor& d);
 
-class FilterDescriptor
+class TORCH_CUDA_API FilterDescriptor
   : public Descriptor<cudnnFilterStruct,
                       &cudnnCreateFilterDescriptor,
                       &cudnnDestroyFilterDescriptor>

aten/src/ATen/native/BinaryOps.cpp

@@ -138,6 +138,10 @@ Tensor true_divide(const Tensor& self, const Tensor& divisor) {
   return iter.output();
 }
 
+Tensor& true_divide_(Tensor& self, const Tensor& divisor) {
+  return native::true_divide_out(self, self, divisor);
+}
+
 Tensor& floor_divide_out(Tensor& result, const Tensor& self, const Tensor& other) {
   auto iter = TensorIterator::binary_op(result, self, other,
     /*check_mem_overlap=*/true);
@@ -731,7 +735,11 @@ Tensor& fmod_(Tensor& self, Scalar other) {
 }
 
 Tensor true_divide(const Tensor& self, Scalar divisor) {
-  return at::true_divide(self, wrapped_scalar_tensor(divisor)); // redispatch!
+  return self.true_divide(wrapped_scalar_tensor(divisor)); // redispatch!
+}
+
+Tensor& true_divide_(Tensor& self, Scalar divisor) {
+  return self.true_divide_(wrapped_scalar_tensor(divisor)); // redispatch!
 }
 
 }

aten/src/ATen/native/TensorConversions.cpp

@@ -33,7 +33,7 @@ static inline Tensor to_impl(const Tensor& self, const TensorOptions& options, b
     if (self.is_non_overlapping_and_dense()) {
       // Copy all strides
       auto r = at::empty_strided(self.sizes(), self.strides(), options.memory_format(c10::nullopt));
-      r.copy_(self);
+      r.copy_(self, non_blocking);
       return r;
     } else {
       memory_format = self.suggest_memory_format();

aten/src/ATen/native/TensorShape.cpp

@@ -98,6 +98,15 @@ Tensor & _cat_out_cpu(Tensor& result, TensorList tensors, int64_t dim) {
         "output memory locations. Found overlap in input tensor ", i);
   }
 
+  // Dtypes should be the same
+  const auto first_in_cat = tensors[0];
+  for (int64_t i = 1; i < tensors.size(); i++) {
+    TORCH_CHECK(first_in_cat.dtype() == tensors[i].dtype(),
+              "Expected object of scalar type ", first_in_cat.dtype(),
+              " but got scalar type ", tensors[i].dtype(),
+              " for sequence element ", i, ".");
+  }
+
   auto should_skip = [](const Tensor& t) { return t.numel() == 0 && t.dim() == 1; };
   for (auto const &tensor : tensors) {
     if (should_skip(tensor)) {

aten/src/ATen/native/cpu/BinaryOpsKernel.cpp

@@ -7,6 +7,7 @@
 #include <ATen/native/TensorIterator.h>
 #include <ATen/native/BinaryOps.h>
 #include <ATen/native/cpu/Loops.h>
+#include <c10/macros/Macros.h>
 
 namespace at { namespace native {
 namespace {

aten/src/ATen/native/cuda/BinaryArithmeticKernel.cu

@@ -4,7 +4,7 @@
 #include <ATen/native/cuda/zmath.cuh>
 #include <ATen/native/TensorIterator.h>
 #include <ATen/native/BinaryOps.h>
-
+#include <c10/macros/Macros.h>
 
 // NOTE: CUDA on Windows requires that the enclosing function
 // of a __device__ lambda not have internal linkage.

aten/src/ATen/native/cuda/Shape.cu

@@ -307,6 +307,15 @@ Tensor& cat_out_cuda(Tensor& out, TensorList inputs, int64_t dimension) {
                 "tensor ", i);
   }
 
+  // Dtypes should be the same
+  const auto first_in_cat = inputs[0];
+  for (int64_t i = 1; i < inputs.size(); i++) {
+    TORCH_CHECK(first_in_cat.dtype() == inputs[i].dtype(),
+              "Expected object of scalar type ", first_in_cat.dtype(),
+              " but got scalar type ", inputs[i].dtype(),
+              " for sequence element ", i, ".");
+  }
+
   for (int i = 0; i < inputs.size(); i++)
   {
     if (should_skip(inputs[i])) {
@@ -325,6 +334,12 @@ Tensor& cat_out_cuda(Tensor& out, TensorList inputs, int64_t dimension) {
   TORCH_CHECK(inputs.size() > 0, "invalid number of inputs ", inputs.size());
   TORCH_CHECK(dimension >= 0, "invalid dimension ", dimension);
 
+  for (const Tensor& t: inputs) {
+    TORCH_CHECK(t.device() == notSkippedTensor->device(),
+                "All input tensors must be on the same device. Received ",
+                t.device(), " and ", notSkippedTensor->device());
+  }
+
   c10::MemoryFormat memory_format = compute_output_memory_format(inputs);
 
   std::vector<int64_t> size(notSkippedTensor->sizes().vec());
@@ -355,17 +370,11 @@ Tensor& cat_out_cuda(Tensor& out, TensorList inputs, int64_t dimension) {
   // 4. The number of dimensions is <= 4
   // 5. All input tensors are contiguous (output tensor may be non-contig)
   // 6. All input tensors can use 32-bit indexing
-  // 7. All input tensors are on the same device
 
   const bool all32BitIndexable = std::all_of(inputs.begin(), inputs.end(),
     [] (const Tensor& t) {
       return at::cuda::detail::canUse32BitIndexMath(t);
     });
-  Device firstDevice = notSkippedTensor->device();
-  const bool allSameDevice = std::all_of(inputs.begin(), inputs.end(),
-    [firstDevice](const Tensor& t) {
-      return t.device() == firstDevice;
-    });
   const bool allContiguous = std::all_of(inputs.begin(), inputs.end(),
     [=](const Tensor& t) {
       return !t.defined() || t.is_contiguous(memory_format);
@@ -375,8 +384,7 @@ Tensor& cat_out_cuda(Tensor& out, TensorList inputs, int64_t dimension) {
       out.dim() <= CAT_ARRAY_MAX_INPUT_DIMS &&
       at::cuda::detail::canUse32BitIndexMath(out) &&
       allContiguous &&
-      all32BitIndexable &&
-      allSameDevice) {
+      all32BitIndexable) {
 
     AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND3(
         at::ScalarType::Half, at::ScalarType::Bool, at::ScalarType::BFloat16,

aten/src/ATen/native/native_functions.yaml

@@ -2872,7 +2872,7 @@
 
 - func: true_divide.Tensor(Tensor self, Tensor other) -> Tensor
   use_c10_dispatcher: full
-  variants: function
+  variants: function, method
   dispatch:
     CPU: true_divide
     CUDA: true_divide
@@ -2880,6 +2880,15 @@
     SparseCUDA: true_divide_sparse
   supports_named_tensor: True
 
+- func: true_divide_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
+  variants: method
+  dispatch:
+    CPU: true_divide_
+    CUDA: true_divide_
+    SparseCPU: true_divide_sparse_
+    SparseCUDA: true_divide_sparse_
+  supports_named_tensor: True
+
 - func: true_divide.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
     CPU: true_divide_out
@@ -2890,7 +2899,11 @@
 
 - func: true_divide.Scalar(Tensor self, Scalar other) -> Tensor
   use_c10_dispatcher: full
-  variants: function
+  variants: function, method
+  supports_named_tensor: True
+
+- func: true_divide_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)
+  variants: method
   supports_named_tensor: True
 
 - func: trunc(Tensor self) -> Tensor

aten/src/ATen/native/sparse/SparseTensorMath.cpp

@@ -272,6 +272,10 @@ SparseTensor& true_divide_out_sparse_scalar(
   return true_divide_out_sparse_zerodim(result, dividend, wrapped_scalar_tensor(divisor));
 }
 
+Tensor& true_divide_sparse_(Tensor& self, const Tensor& divisor) {
+  return true_divide_out_sparse_zerodim(self, self, divisor);
+}
+
 // --------------------------------------------------------------------
 // floor_divide(SparseTensor, Scalar)
 // --------------------------------------------------------------------

aten/src/ATen/native/xnnpack/RegisterOpContextClass.cpp

@@ -14,7 +14,7 @@ namespace xnnpack {
 namespace {
 torch::jit::class_<XNNPackLinearOpContext> register_xnnpack_linear_op_context_class() {
   static auto register_linear_op_context_class =
-      torch::jit::class_<XNNPackLinearOpContext>("XNNPackLinearOpContext")
+      torch::jit::class_<XNNPackLinearOpContext>("xnnpack", "XNNPackLinearOpContext")
           .def_pickle(
               [](const c10::intrusive_ptr<XNNPackLinearOpContext>& op_context)
                   -> SerializationTypeLinearPrePack { // __getstate__
@@ -38,7 +38,7 @@ torch::jit::class_<XNNPackLinearOpContext> register_xnnpack_linear_op_context_cl
 
 torch::jit::class_<XNNPackConv2dOpContext> register_xnnpack_conv2d_op_context_class() {
   static auto register_conv2d_op_context_class =
-      torch::jit::class_<XNNPackConv2dOpContext>("XNNPackConv2dOpContext")
+      torch::jit::class_<XNNPackConv2dOpContext>("xnnpack", "XNNPackConv2dOpContext")
           .def_pickle(
               [](const c10::intrusive_ptr<XNNPackConv2dOpContext>& op_context)
                   -> SerializationTypeConv2dPrePack { // __getstate__
@@ -74,25 +74,25 @@ static auto registry =
   // Registering under _xnnpack namespace for now. As we add more backend requiring similar functionality
   // We can refactor the code and use a better namespace.
     torch::RegisterOperators()
-        .op("_xnnpack::linear_prepack(Tensor W, Tensor? B=None) -> __torch__.torch.classes.XNNPackLinearOpContext",
+        .op("_xnnpack::linear_prepack(Tensor W, Tensor? B=None) -> __torch__.torch.classes.xnnpack.XNNPackLinearOpContext",
             torch::RegisterOperators::options()
             .aliasAnalysis(at::AliasAnalysisKind::PURE_FUNCTION)
             .kernel<internal::linear::LinearPrePack>(
                 DispatchKey::CPUTensorId))
-        .op("_xnnpack::linear_packed(Tensor X, __torch__.torch.classes.XNNPackLinearOpContext W_prepack) -> Tensor Y",
+        .op("_xnnpack::linear_packed(Tensor X, __torch__.torch.classes.xnnpack.XNNPackLinearOpContext W_prepack) -> Tensor Y",
             torch::RegisterOperators::options()
             .aliasAnalysis(at::AliasAnalysisKind::PURE_FUNCTION)
             .kernel<internal::linear::LinearPacked>(
                 DispatchKey::CPUTensorId))
         .op("_xnnpack::conv2d_prepack(Tensor W, Tensor? B, int[2] stride, "
             "int[2] padding, int[2] dilation, int groups) "
-            "-> __torch__.torch.classes.XNNPackConv2dOpContext",
+            "-> __torch__.torch.classes.xnnpack.XNNPackConv2dOpContext",
             torch::RegisterOperators::options()
             .aliasAnalysis(at::AliasAnalysisKind::PURE_FUNCTION)
             .kernel<internal::convolution2d::Conv2dPrePack>(
                 DispatchKey::CPUTensorId))
         .op("_xnnpack::conv2d_packed(Tensor X, "
-            "__torch__.torch.classes.XNNPackConv2dOpContext W_prepack) -> Tensor Y",
+            "__torch__.torch.classes.xnnpack.XNNPackConv2dOpContext W_prepack) -> Tensor Y",
             torch::RegisterOperators::options()
             .aliasAnalysis(at::AliasAnalysisKind::PURE_FUNCTION)
             .kernel<internal::convolution2d::Conv2dPacked>(

aten/src/TH/THGeneral.h.in

@@ -69,12 +69,6 @@
 # define TH_UNUSED
 #endif
 
-#if defined(__clang__)
-#define __ubsan_ignore_float_divide_by_zero__ __attribute__((no_sanitize("float-divide-by-zero")))
-#else
-#define __ubsan_ignore_float_divide_by_zero__
-#endif
-
 #ifndef M_PI
 # define M_PI 3.14159265358979323846
 #endif

c10/macros/Macros.h

@@ -23,6 +23,14 @@
 
 #include "c10/macros/Export.h"
 
+#if defined(__clang__)
+  #define __ubsan_ignore_float_divide_by_zero__ __attribute__((no_sanitize("float-divide-by-zero")))
+  #define __ubsan_ignore_float_cast_overflow__ __attribute__((no_sanitize("float-cast-overflow")))
+#else
+  #define __ubsan_ignore_float_divide_by_zero__
+  #define __ubsan_ignore_float_cast_overflow__
+#endif
+
 // Disable the copy and assignment operator for a class. Note that this will
 // disable the usage of the class in std containers.
 #define C10_DISABLE_COPY_AND_ASSIGN(classname) \

c10/util/Exception.cpp

@@ -66,24 +66,44 @@ void Error::AppendMessage(const std::string& new_msg) {
 namespace Warning {
 
 namespace {
-  WarningHandler* getHandler() {
+  WarningHandler* getBaseHandler() {
     static WarningHandler base_warning_handler_ = WarningHandler();
     return &base_warning_handler_;
   };
-  static thread_local WarningHandler* warning_handler_ = getHandler();
+
+  class ThreadWarningHandler {
+    public:
+      ThreadWarningHandler() = delete;
+
+      static WarningHandler* get_handler() {
+        if (!warning_handler_) {
+          warning_handler_ = getBaseHandler();
+        }
+        return warning_handler_;
+      }
+
+      static void set_handler(WarningHandler* handler) {
+        warning_handler_ = handler;
+      }
+
+    private:
+      static thread_local WarningHandler* warning_handler_;
+  };
+
+  thread_local WarningHandler* ThreadWarningHandler::warning_handler_ = nullptr;
 
 }
 
 void warn(SourceLocation source_location, const std::string& msg) {
-  warning_handler_->process(source_location, msg);
+  ThreadWarningHandler::get_handler()->process(source_location, msg);
 }
 
 void set_warning_handler(WarningHandler* handler) noexcept(true) {
-  warning_handler_ = handler;
+  ThreadWarningHandler::set_handler(handler);
 }
 
 WarningHandler* get_warning_handler() noexcept(true) {
-  return warning_handler_;
+  return ThreadWarningHandler::get_handler();
 }
 
 } // namespace Warning

c10/util/TypeCast.h

@@ -67,7 +67,7 @@ struct maybe_real<true, src_t> {
 
 template <typename dest_t, typename src_t>
 struct static_cast_with_inter_type {
-  C10_HOST_DEVICE static inline dest_t apply(src_t src) {
+  C10_HOST_DEVICE __ubsan_ignore_float_cast_overflow__ static inline dest_t apply(src_t src) {
     constexpr bool real = needs_real<dest_t, src_t>::value;
     return static_cast<dest_t>(
       static_cast<inter_copy_type_t<dest_t>>(maybe_real<real, src_t>::apply(src)));

docs/source/distributed.rst

@@ -395,6 +395,8 @@ of 16
 .. autofunction:: all_gather_multigpu
 
 
+.. _distributed-launch:
+
 Launch utility
 --------------
 

docs/source/notes/cuda.rst

@@ -306,20 +306,30 @@ to overlap data transfers with computation.
 You can make the :class:`~torch.utils.data.DataLoader` return batches placed in
 pinned memory by passing ``pin_memory=True`` to its constructor.
 
-.. _cuda-nn-dataparallel-instead:
+.. _cuda-nn-ddp-instead:
 
-Use nn.DataParallel instead of multiprocessing
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+Use nn.parallel.DistributedDataParallel instead of multiprocessing or nn.DataParallel
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Most use cases involving batched inputs and multiple GPUs should default to
-using :class:`~torch.nn.DataParallel` to utilize more than one GPU. Even with
-the GIL, a single Python process can saturate multiple GPUs.
-
-As of version 0.1.9, large numbers of GPUs (8+) might not be fully utilized.
-However, this is a known issue that is under active development. As always,
-test your use case.
+using :class:`~torch.nn.parallel.DistributedDataParallel` to utilize more
+than one GPU.
 
 There are significant caveats to using CUDA models with
 :mod:`~torch.multiprocessing`; unless care is taken to meet the data handling
 requirements exactly, it is likely that your program will have incorrect or
 undefined behavior.
+
+It is recommended to use :class:`~torch.nn.parallel.DistributedDataParallel`,
+instead of :class:`~torch.nn.DataParallel` to do multi-GPU training, even if
+there is only a single node.
+
+The difference between :class:`~torch.nn.parallel.DistributedDataParallel` and
+:class:`~torch.nn.DataParallel` is: :class:`~torch.nn.parallel.DistributedDataParallel`
+uses multiprocessing where a process is created for each GPU, while
+:class:`~torch.nn.DataParallel` uses multithreading. By using multiprocessing,
+each GPU has its dedicated process, this avoids the performance overhead caused
+by GIL of Python interpreter. 
+
+If you use :class:`~torch.nn.parallel.DistributedDataParallel`, you could use 
+`torch.distributed.launch` utility to launch your program, see :ref:`distributed-launch`.

docs/source/notes/multiprocessing.rst

@@ -45,7 +45,7 @@ the consumer process has references to the tensor, and the refcounting can not
 save you if the consumer process exits abnormally via a fatal signal. See
 :ref:`this section <multiprocessing-cuda-sharing-details>`.
 
-See also: :ref:`cuda-nn-dataparallel-instead`
+See also: :ref:`cuda-nn-ddp-instead`
 
 
 Best practices and tips

docs/source/tensor_attributes.rst

@@ -210,3 +210,25 @@ Example::
     (1, 5)
 
 For more information on ``torch.sparse_coo`` tensors, see :ref:`sparse-docs`.
+
+torch.memory_format
+------------
+
+.. class:: torch.memory_format
+
+A :class:`torch.memory_format` is an object representing the memory format on which a :class:`torch.Tensor` is
+or will be allocated.
+
+Possible values are:
+
+- ``torch.contiguous_format``:
+  Tensor is or will be  allocated in dense non-overlapping memory. Strides represented by values in decreasing order.
+
+- ``torch.channels_last``:
+  Tensor is or will be  allocated in dense non-overlapping memory. Strides represented by values in 
+  ``strides[0] > strides[2] > strides[3] > strides[1] == 1`` aka NHWC order. 
+
+- ``torch.preserve_format``:
+  Used in functions like `clone` to preserve the memory format of the input tensor. If input tensor is 
+  allocated in dense non-overlapping memory, the output tensor strides will be copied from the input. 
+  Otherwise output strides will follow ``torch.contiguous_format``

docs/source/tensor_view.rst

@@ -49,8 +49,10 @@ For reference, here’s a full list of view ops in PyTorch:
 
 - Basic slicing and indexing op, e.g. ``tensor[0, 2:, 1:7:2]`` returns a view of base ``tensor``, see note below.
 - :meth:`~torch.Tensor.as_strided`
+- :meth:`~torch.Tensor.detach`
 - :meth:`~torch.Tensor.diagonal`
 - :meth:`~torch.Tensor.expand`
+- :meth:`~torch.Tensor.expand_as`
 - :meth:`~torch.Tensor.narrow`
 - :meth:`~torch.Tensor.permute`
 - :meth:`~torch.Tensor.select`

docs/source/tensors.rst

@@ -495,6 +495,8 @@ view of a storage and defines numeric operations on it.
    .. automethod:: tril_
    .. automethod:: triu
    .. automethod:: triu_
+   .. automethod:: true_divide
+   .. automethod:: true_divide_
    .. automethod:: trunc
    .. automethod:: trunc_
    .. automethod:: type

setup.py

@@ -352,10 +352,10 @@ def build_deps():
 ################################################################################
 
 # the list of runtime dependencies required by this built package
-install_requires = []
+install_requires = ['future']
 
 if sys.version_info <= (2, 7):
-    install_requires += ['future', 'typing']
+    install_requires += ['typing']
 
 missing_pydep = '''
 Missing build dependency: Unable to `import {importname}`.

test/backward_compatibility/check_backward_compatibility.py

@@ -115,6 +115,10 @@ white_list = [
     ('aten::confirmed_by_owner', datetime.date(2020, 3, 17)),
     ('aten::owner', datetime.date(2020, 3, 27)),
     ('aten::owner_name', datetime.date(2020, 3, 27)),
+    ('_xnnpack::conv2d_packed', datetime.date(2020, 4, 2)),
+    ('_xnnpack::conv2d_prepack', datetime.date(2020, 4, 2)),
+    ('_xnnpack::linear_packed', datetime.date(2020, 4, 2)),
+    ('_xnnpack::linear_prepack', datetime.date(2020, 4, 2)),
 ]
 
 
@@ -176,6 +180,9 @@ if __name__ == '__main__':
             line = f.readline()
             if not line:
                 break
+            if "torch.classes" in line:
+                # TODO Fix type __torch__.torch.classes.xxx
+                continue
 
             s = parse_schema(line.strip())
             slist = new_schema_dict.get(s.name, [])

test/cpp/api/functional.cpp

@@ -293,7 +293,7 @@ TEST_F(FunctionalTest, MultiLabelSoftMarginLossWeightedNoReduction) {
   auto input = torch::tensor({{0., 2., 2., 0.}, {2., 1., 0., 1.}}, torch::dtype(torch::kFloat).requires_grad(true));
   auto target = torch::tensor({{0., 0., 1., 0.}, {1., 0., 1., 1.}}, torch::kFloat);
   auto weight = torch::tensor({0.1, 0.6, 0.4, 0.8}, torch::kFloat);
-  auto options = F::MultiLabelSoftMarginLossFuncOptions().reduction(torch::kNone).weight(weight);
+  auto options = F::MultilabelSoftMarginLossFuncOptions().reduction(torch::kNone).weight(weight);
   auto output =
       F::multilabel_soft_margin_loss(input, target, options);
   auto expected = torch::tensor({0.4876902, 0.3321295}, torch::kFloat);
@@ -1875,7 +1875,7 @@ TEST_F(FunctionalTest, Interpolate) {
     // 1D interpolation
     auto input = torch::ones({1, 1, 2});
     auto options = F::InterpolateFuncOptions()
-                       .size({4})
+                       .size(std::vector<int64_t>({4}))
                        .mode(torch::kNearest);
     auto output = F::interpolate(input, options);
     auto expected = torch::ones({1, 1, 4});
@@ -1889,7 +1889,7 @@ TEST_F(FunctionalTest, Interpolate) {
       for (const auto scale_factor : {0.5, 1.5, 2.0}) {
         auto input = torch::ones({1, 1, 2, 2});
         auto options = F::InterpolateFuncOptions()
-                           .scale_factor({scale_factor, scale_factor})
+                           .scale_factor(std::vector<double>({scale_factor, scale_factor}))
                            .mode(torch::kBilinear)
                            .align_corners(align_corners);
         auto output = F::interpolate(input, options);
@@ -1908,7 +1908,7 @@ TEST_F(FunctionalTest, Interpolate) {
         auto input = torch::ones({1, 1, 2, 2, 2});
         auto options =
             F::InterpolateFuncOptions()
-                .scale_factor({scale_factor, scale_factor, scale_factor})
+                .scale_factor(std::vector<double>({scale_factor, scale_factor, scale_factor}))
                 .mode(torch::kTrilinear)
                 .align_corners(align_corners);
         auto output = F::interpolate(input, options);
@@ -1924,13 +1924,13 @@ TEST_F(FunctionalTest, Interpolate) {
   {
     auto input = torch::randn({3, 2, 2});
     ASSERT_THROWS_WITH(
-        F::interpolate(input[0], F::InterpolateFuncOptions().size({4, 4})),
+        F::interpolate(input[0], F::InterpolateFuncOptions().size(std::vector<int64_t>({4, 4}))),
         "Input Error: Only 3D, 4D and 5D input Tensors supported (got 2D) "
         "for the modes: nearest | linear | bilinear | bicubic | trilinear (got kNearest)");
     ASSERT_THROWS_WITH(
         F::interpolate(
             torch::reshape(input, {1, 1, 1, 3, 2, 2}),
-            F::InterpolateFuncOptions().size({1, 1, 1, 3, 4, 4})),
+            F::InterpolateFuncOptions().size(std::vector<int64_t>({1, 1, 1, 3, 4, 4}))),
         "Input Error: Only 3D, 4D and 5D input Tensors supported (got 6D) "
         "for the modes: nearest | linear | bilinear | bicubic | trilinear (got kNearest)");
     ASSERT_THROWS_WITH(
@@ -1939,12 +1939,12 @@ TEST_F(FunctionalTest, Interpolate) {
     ASSERT_THROWS_WITH(
         F::interpolate(
             input,
-            F::InterpolateFuncOptions().size({3, 4, 4}).scale_factor({0.5})),
+            F::InterpolateFuncOptions().size(std::vector<int64_t>({3, 4, 4})).scale_factor(std::vector<double>({0.5}))),
         "only one of size or scale_factor should be defined");
     ASSERT_THROWS_WITH(
-        F::interpolate(input, F::InterpolateFuncOptions().scale_factor({3, 2})),
+        F::interpolate(input, F::InterpolateFuncOptions().scale_factor(std::vector<double>({3, 2}))),
         "scale_factor shape must match input shape. "
-        "Input is 1D, scale_factor size is 2");
+        "Input is 1D, scale_factor size is [3, 2]");
     ASSERT_THROWS_WITH(
         F::interpolate(
             input,
@@ -2328,9 +2328,15 @@ TEST_F(FunctionalTest, AlphaDropout) {
   auto input_std = input.std();
 
   for (const auto rate : {0.2, 0.5, 0.8}) {
-    auto output = F::alpha_dropout(input, F::AlphaDropoutFuncOptions().p(rate).training(false));
-    ASSERT_TRUE(torch::allclose(input_mean, output.mean(), 0.1));
-    ASSERT_TRUE(torch::allclose(input_std, output.std(), 0.1));
+    for (const auto inplace : {false, true}) {
+      auto input_ = input.clone();
+      auto output = F::alpha_dropout(input_, F::AlphaDropoutFuncOptions().p(rate).training(false).inplace(inplace));
+      ASSERT_TRUE(torch::allclose(input_mean, output.mean(), 0.1));
+      ASSERT_TRUE(torch::allclose(input_std, output.std(), 0.1));
+      if (inplace) {
+        ASSERT_TRUE(torch::allclose(input_, output));
+      }
+    }
   }
   auto output = F::detail::alpha_dropout(input, 0.5, false, false);
   ASSERT_TRUE(torch::allclose(input_mean, output.mean(), 0.1));
@@ -2343,9 +2349,15 @@ TEST_F(FunctionalTest, FeatureAlphaDropout) {
   auto input_std = input.std();
 
   for (const auto rate : {0.2, 0.5, 0.8}) {
-    auto output = F::feature_alpha_dropout(input, F::FeatureAlphaDropoutFuncOptions().p(rate).training(false));
-    ASSERT_TRUE(torch::allclose(input_mean, output.mean(), 0.1));
-    ASSERT_TRUE(torch::allclose(input_std, output.std(), 0.1));
+    for (const auto inplace : {false, true}) {
+      auto input_ = input.clone();
+      auto output = F::feature_alpha_dropout(input_, F::FeatureAlphaDropoutFuncOptions().p(rate).training(false).inplace(inplace));
+      ASSERT_TRUE(torch::allclose(input_mean, output.mean(), 0.1));
+      ASSERT_TRUE(torch::allclose(input_std, output.std(), 0.1));
+      if (inplace) {
+        ASSERT_TRUE(torch::allclose(input_, output));
+      }
+    }
   }
   auto output = F::feature_alpha_dropout(input);
   ASSERT_TRUE(torch::allclose(input_mean, output.mean(), 0.1));

test/cpp/api/modules.cpp

@@ -1300,54 +1300,81 @@ TEST_F(ModulesTest, FeatureAlphaDropout) {
 }
 
 TEST_F(ModulesTest, Dropout) {
-  Dropout dropout(0.5);
-  torch::Tensor x = torch::ones(100, torch::requires_grad());
-  torch::Tensor y = dropout(x);
+  for (const auto inplace : {false, true}) {
+    Dropout dropout(DropoutOptions(0.5).inplace(inplace));
+    torch::Tensor x = torch::ones(100);
+    if (!inplace) {
+      x.requires_grad_(true);
+    }
+    torch::Tensor y = dropout(x);
 
-  y.backward(torch::ones_like(y));
-  ASSERT_EQ(y.ndimension(), 1);
-  ASSERT_EQ(y.size(0), 100);
-  ASSERT_LT(y.sum().item<float>(), 130); // Probably
-  ASSERT_GT(y.sum().item<float>(), 70); // Probably
+    ASSERT_EQ(y.ndimension(), 1);
+    ASSERT_EQ(y.size(0), 100);
+    ASSERT_LT(y.sum().item<float>(), 130); // Probably
+    ASSERT_GT(y.sum().item<float>(), 70); // Probably
+    if (inplace) {
+      ASSERT_TRUE(y.allclose(x));
+    } else {
+      y.backward(torch::ones_like(y));
+    }
 
-  dropout->eval();
-  y = dropout(x);
-  ASSERT_EQ(y.sum().item<float>(), 100);
+    dropout->eval();
+    y = dropout(torch::ones(100));
+    ASSERT_EQ(y.sum().item<float>(), 100);
+  }
 }
 
 TEST_F(ModulesTest, Dropout2d) {
-  Dropout2d dropout(0.5);
-  torch::Tensor x = torch::ones({10, 10}, torch::requires_grad());
-  torch::Tensor y = dropout(x);
+  for (const auto inplace : {false, true}) {
+    Dropout2d dropout(Dropout2dOptions(0.5).inplace(inplace));
+    torch::Tensor x = torch::ones({10, 10});
+    if (!inplace) {
+      x.requires_grad_(true);
+    }
+    torch::Tensor y = dropout(x);
 
-  y.backward(torch::ones_like(y));
-  ASSERT_EQ(y.ndimension(), 2);
-  ASSERT_EQ(y.size(0), 10);
-  ASSERT_EQ(y.size(1), 10);
-  ASSERT_LT(y.sum().item<float>(), 130); // Probably
-  ASSERT_GT(y.sum().item<float>(), 70); // Probably
+    ASSERT_EQ(y.ndimension(), 2);
+    ASSERT_EQ(y.size(0), 10);
+    ASSERT_EQ(y.size(1), 10);
+    ASSERT_LT(y.sum().item<float>(), 130); // Probably
+    ASSERT_GT(y.sum().item<float>(), 70); // Probably
+    if (inplace) {
+      ASSERT_TRUE(y.allclose(x));
+    } else {
+      y.backward(torch::ones_like(y));
+    }
 
-  dropout->eval();
-  y = dropout(x);
-  ASSERT_EQ(y.sum().item<float>(), 100);
+    dropout->eval();
+    y = dropout(torch::ones({10, 10}));
+    ASSERT_EQ(y.sum().item<float>(), 100);
+  }
 }
 
 TEST_F(ModulesTest, Dropout3d) {
-  Dropout3d dropout(0.5);
-  torch::Tensor x = torch::ones({4, 5, 5}, torch::requires_grad());
-  torch::Tensor y = dropout(x);
+  for (const auto inplace : {false, true}) {
+    Dropout3d dropout(Dropout3dOptions(0.5).inplace(inplace));
+    torch::Tensor x = torch::ones({4, 5, 5});
+    if (!inplace) {
+      x.requires_grad_(true);
+    }
+    torch::Tensor y = dropout(x);
 
-  y.backward(torch::ones_like(y));
-  ASSERT_EQ(y.ndimension(), 3);
-  ASSERT_EQ(y.size(0), 4);
-  ASSERT_EQ(y.size(1), 5);
-  ASSERT_EQ(y.size(1), 5);
-  ASSERT_LT(y.sum().item<float>(), 130); // Probably
-  ASSERT_GT(y.sum().item<float>(), 70); // Probably
+    ASSERT_EQ(y.ndimension(), 3);
+    ASSERT_EQ(y.size(0), 4);
+    ASSERT_EQ(y.size(1), 5);
+    ASSERT_EQ(y.size(1), 5);
+    ASSERT_LT(y.sum().item<float>(), 130); // Probably
+    ASSERT_GT(y.sum().item<float>(), 70); // Probably
+    if (inplace) {
+      ASSERT_TRUE(y.allclose(x));
+    } else {
+      y.backward(torch::ones_like(y));
+    }
 
-  dropout->eval();
-  y = dropout(x);
-  ASSERT_EQ(y.sum().item<float>(), 100);
+    dropout->eval();
+    y = dropout(torch::ones({4, 5, 5}));
+    ASSERT_EQ(y.sum().item<float>(), 100);
+  }
 }
 
 TEST_F(ModulesTest, Parameters) {
@@ -2147,38 +2174,58 @@ TEST_F(ModulesTest, PairwiseDistance) {
 TEST_F(ModulesTest, ELU) {
   const auto size = 3;
   for (const auto alpha : {0.0, 0.42, 1.0, 4.2, 42.42}) {
-    ELU model {ELUOptions().alpha(alpha)};
-    auto x = torch::linspace(-10.0, 10.0, size * size * size);
-    x.resize_({size, size, size}).set_requires_grad(true);
-    auto y = model(x);
-    torch::Tensor s = y.sum();
+    for (const auto inplace : {false, true}) {
+      ELU model {ELUOptions().alpha(alpha).inplace(inplace)};
+      auto x = torch::linspace(-10.0, 10.0, size * size * size);
+      x.resize_({size, size, size});
+      if (!inplace) {
+        x.requires_grad_(true);
+      }
+      auto x_orig = x.clone();
+      auto y = model(x);
+      torch::Tensor s = y.sum();
 
-    s.backward();
-    ASSERT_EQ(s.ndimension(), 0);
+      ASSERT_EQ(s.ndimension(), 0);
 
-    ASSERT_EQ(y.ndimension(), 3);
-    ASSERT_EQ(y.sizes(), std::vector<int64_t>({size, size, size}));
-    auto y_exp = torch::max(torch::zeros_like(x), x) +
-                 torch::min(torch::zeros_like(x), alpha * (torch::exp(x) - 1.0));
-    ASSERT_TRUE(torch::allclose(y, y_exp));
+      ASSERT_EQ(y.ndimension(), 3);
+      ASSERT_EQ(y.sizes(), std::vector<int64_t>({size, size, size}));
+      auto y_exp = torch::max(torch::zeros_like(x_orig), x_orig) +
+                   torch::min(torch::zeros_like(x_orig), alpha * (torch::exp(x_orig) - 1.0));
+      ASSERT_TRUE(torch::allclose(y, y_exp));
+      if (inplace) {
+        ASSERT_TRUE(torch::allclose(x, y_exp));
+      } else {
+        s.backward();
+      }
+    }
   }
 }
 
 TEST_F(ModulesTest, SELU) {
-  SELU model;
-  auto input = torch::randn({5, 5}, torch::requires_grad());
-  auto output = model->forward(input);
-  const double scale = 1.0507009873554804934193349852946;
-  const double alpha = 1.6732632423543772848170429916717;
-  auto zero = torch::zeros_like(input);
-  auto expected = scale *
-      (torch::max(zero, input) +
-       torch::min(zero, alpha * (torch::exp(input) - 1)));
-  auto s = output.sum();
-  s.backward();
+  for (const auto inplace : {false, true}) {
+    SELU model(inplace);
+    auto input = torch::randn({5, 5});
+    if (!inplace) {
+      input.requires_grad_(true);
+    }
+    auto input_orig = input.clone();
+    auto output = model->forward(input);
+    const double scale = 1.0507009873554804934193349852946;
+    const double alpha = 1.6732632423543772848170429916717;
+    auto zero = torch::zeros_like(input);
+    auto expected = scale *
+        (torch::max(zero, input_orig) +
+         torch::min(zero, alpha * (torch::exp(input_orig) - 1)));
+    auto s = output.sum();
 
-  ASSERT_EQ(s.ndimension(), 0);
-  ASSERT_TRUE(output.allclose(expected));
+    ASSERT_EQ(s.ndimension(), 0);
+    ASSERT_TRUE(output.allclose(expected));
+    if (inplace) {
+      ASSERT_TRUE(input.allclose(expected));
+    } else {
+      s.backward();
+    }
+  }
 }
 
 TEST_F(ModulesTest, Hardshrink) {
@@ -2192,7 +2239,6 @@ TEST_F(ModulesTest, Hardshrink) {
 
     s.backward();
     ASSERT_EQ(s.ndimension(), 0);
-
     ASSERT_EQ(y.ndimension(), 3);
     ASSERT_EQ(y.sizes(), std::vector<int64_t>({size, size, size}));
     auto y_exp = (x.abs() > lambda) * x;
@@ -2204,21 +2250,30 @@ TEST_F(ModulesTest, Hardtanh) {
   const auto size = 3;
   for (const auto min_val : {-4.2, -1.0, -0.42, 0.0}) {
     for (const auto max_val : {0.42, 1.0, 4.2}) {
-      Hardtanh model {HardtanhOptions().min_val(min_val).max_val(max_val)};
-      auto x = torch::linspace(-10.0, 10.0, size * size * size);
-      x.resize_({size, size, size}).set_requires_grad(true);
-      auto y = model(x);
-      torch::Tensor s = y.sum();
+      for (const auto inplace : {false, true}) {
+        Hardtanh model {HardtanhOptions().min_val(min_val).max_val(max_val).inplace(inplace)};
+        auto x = torch::linspace(-10.0, 10.0, size * size * size);
+        x.resize_({size, size, size});
+        if (!inplace) {
+          x.requires_grad_(true);
+        }
+        auto x_orig = x.clone();
+        auto y = model(x);
+        torch::Tensor s = y.sum();
 
-      s.backward();
-      ASSERT_EQ(s.ndimension(), 0);
-
-      ASSERT_EQ(y.ndimension(), 3);
-      ASSERT_EQ(y.sizes(), std::vector<int64_t>({size, size, size}));
-      auto y_exp = (x < min_val) * min_val +
-                   ((x >= min_val) * (x <= max_val)) * x +
-                   (x > max_val) * max_val;
-      ASSERT_TRUE(torch::allclose(y, y_exp));
+        ASSERT_EQ(s.ndimension(), 0);
+        ASSERT_EQ(y.ndimension(), 3);
+        ASSERT_EQ(y.sizes(), std::vector<int64_t>({size, size, size}));
+        auto y_exp = (x_orig < min_val) * min_val +
+                     ((x_orig >= min_val) * (x_orig <= max_val)) * x_orig +
+                     (x_orig > max_val) * max_val;
+        ASSERT_TRUE(torch::allclose(y, y_exp));
+        if (inplace) {
+          ASSERT_TRUE(torch::allclose(x, y_exp));
+        } else {
+          s.backward();
+        }
+      }
     }
   }
 }
@@ -2238,20 +2293,29 @@ TEST_F(ModulesTest, HardtanhMinValGEMaxVal) {
 
 TEST_F(ModulesTest, LeakyReLU) {
   const auto size = 3;
-  for (const auto negative_slope : {0.0, 0.42, 1.0}) {
-    LeakyReLU model {LeakyReLUOptions().negative_slope(negative_slope)};
-    auto x = torch::linspace(-10.0, 10.0, size * size * size);
-    x.resize_({size, size, size}).set_requires_grad(true);
-    auto y = model(x);
-    torch::Tensor s = y.sum();
+  for (const auto inplace : {false, true}) {
+    for (const auto negative_slope : {0.0, 0.42, 1.0}) {
+      LeakyReLU model {LeakyReLUOptions().negative_slope(negative_slope).inplace(inplace)};
+      auto x = torch::linspace(-10.0, 10.0, size * size * size);
+      x.resize_({size, size, size});
+      if (!inplace) {
+        x.requires_grad_(true);
+      }
+      auto x_orig = x.clone();
+      auto y = model(x);
+      torch::Tensor s = y.sum();
 
-    s.backward();
-    ASSERT_EQ(s.ndimension(), 0);
-
-    ASSERT_EQ(y.ndimension(), 3);
-    ASSERT_EQ(y.sizes(), std::vector<int64_t>({size, size, size}));
-    auto y_exp = (x < 0) * x * negative_slope + (x >= 0) * x;
-    ASSERT_TRUE(torch::allclose(y, y_exp));
+      ASSERT_EQ(s.ndimension(), 0);
+      ASSERT_EQ(y.ndimension(), 3);
+      ASSERT_EQ(y.sizes(), std::vector<int64_t>({size, size, size}));
+      auto y_exp = (x_orig < 0) * x_orig * negative_slope + (x_orig >= 0) * x_orig;
+      ASSERT_TRUE(torch::allclose(y, y_exp));
+      if (inplace) {
+        ASSERT_TRUE(torch::allclose(x, y_exp));
+      } else {
+        s.backward();
+      }
+    }
   }
 }
 
@@ -2394,78 +2458,114 @@ TEST_F(ModulesTest, PReLU) {
 }
 
 TEST_F(ModulesTest, ReLU) {
-  const auto size = 3;
-  ReLU model;
-  auto x = torch::linspace(-10.0, 10.0, size * size * size);
-  x.resize_({size, size, size}).set_requires_grad(true);
-  auto y = model(x);
-  torch::Tensor s = y.sum();
+  for (const auto inplace : {false, true}) {
+    const auto size = 3;
+    ReLU model(inplace);
+    auto x = torch::linspace(-10.0, 10.0, size * size * size);
+    x.resize_({size, size, size});
+    if (!inplace) {
+      x.requires_grad_(true);
+    }
+    auto x_orig = x.clone();
+    auto y = model(x);
+    torch::Tensor s = y.sum();
 
-  s.backward();
-  ASSERT_EQ(s.ndimension(), 0);
-
-  ASSERT_EQ(y.ndimension(), 3);
-  ASSERT_EQ(y.sizes(), std::vector<int64_t>({size, size, size}));
-  auto y_exp = (x < 0) * 0 + (x >= 0) * x;
-  ASSERT_TRUE(torch::allclose(y, y_exp));
+    ASSERT_EQ(s.ndimension(), 0);
+    ASSERT_EQ(y.ndimension(), 3);
+    ASSERT_EQ(y.sizes(), std::vector<int64_t>({size, size, size}));
+    auto y_exp = (x_orig < 0) * 0 + (x_orig >= 0) * x_orig;
+    ASSERT_TRUE(torch::allclose(y, y_exp));
+    if (inplace) {
+      ASSERT_TRUE(torch::allclose(x, y_exp));
+    } else {
+      s.backward();
+    }
+  }
 }
 
 TEST_F(ModulesTest, ReLU6) {
-  const auto size = 3;
-  ReLU6 model;
-  auto x = torch::linspace(-10.0, 10.0, size * size * size);
-  x.resize_({size, size, size}).set_requires_grad(true);
-  auto y = model(x);
-  torch::Tensor s = y.sum();
+  for (const auto inplace : {false, true}) {
+    const auto size = 3;
+    ReLU6 model(inplace);
+    auto x = torch::linspace(-10.0, 10.0, size * size * size);
+    x.resize_({size, size, size});
+    if (!inplace) {
+      x.requires_grad_(true);
+    }
+    auto x_orig = x.clone();
+    auto y = model(x);
+    torch::Tensor s = y.sum();
 
-  s.backward();
-  ASSERT_EQ(s.ndimension(), 0);
-
-  ASSERT_EQ(y.ndimension(), 3);
-  ASSERT_EQ(y.sizes(), std::vector<int64_t>({size, size, size}));
-  auto y_exp = (x < 0) * 0 + ((x >= 0) * (x <= 6)) * x + (x > 6) * 6;
-  ASSERT_TRUE(torch::allclose(y, y_exp));
+    ASSERT_EQ(s.ndimension(), 0);
+    ASSERT_EQ(y.ndimension(), 3);
+    ASSERT_EQ(y.sizes(), std::vector<int64_t>({size, size, size}));
+    auto y_exp = (x_orig < 0) * 0 + ((x_orig >= 0) * (x_orig <= 6)) * x_orig + (x_orig > 6) * 6;
+    ASSERT_TRUE(torch::allclose(y, y_exp));
+    if (inplace) {
+      ASSERT_TRUE(torch::allclose(x, y_exp));
+    } else {
+      s.backward();
+    }
+  }
 }
 
 TEST_F(ModulesTest, RReLU) {
   const auto size = 3;
   for (const auto lower : {0.01, 0.1, 0.2}) {
     for (const auto upper : {0.3, 0.4, 0.5}) {
-      RReLU model {RReLUOptions().lower(lower).upper(upper)};
-      auto x = torch::linspace(-10.0, 10.0, size * size * size);
-      x.resize_({size, size, size}).set_requires_grad(true);
-      auto y = model(x);
-      torch::Tensor s = y.sum();
+      for (const auto inplace : {false, true}) {
+        RReLU model {RReLUOptions().lower(lower).upper(upper).inplace(inplace)};
+        auto x = torch::linspace(-10.0, 10.0, size * size * size);
+        x.resize_({size, size, size});
+        if (!inplace) {
+          x.requires_grad_(true);
+        }
+        auto x_orig = x.clone();
+        auto y = model(x);
+        torch::Tensor s = y.sum();
 
-      s.backward();
-      ASSERT_EQ(s.ndimension(), 0);
-
-      ASSERT_EQ(y.ndimension(), 3);
-      ASSERT_EQ(y.sizes(), std::vector<int64_t>({size, size, size}));
-      auto z = ((x >= 0) * (x == y) +
-        (x < 0) * (y >= x * upper) * (y <= lower * x)) * 1.0;
-      ASSERT_TRUE(torch::allclose(z, torch::ones_like(z)));
+        ASSERT_EQ(s.ndimension(), 0);
+        ASSERT_EQ(y.ndimension(), 3);
+        ASSERT_EQ(y.sizes(), std::vector<int64_t>({size, size, size}));
+        auto z = ((x_orig >= 0) * (x_orig == y) +
+          (x_orig < 0) * (y >= x_orig * upper) * (y <= lower * x_orig)) * 1.0;
+        ASSERT_TRUE(torch::allclose(z, torch::ones_like(z)));
+        if (inplace) {
+          ASSERT_TRUE(torch::allclose(x, y));
+        } else {
+          s.backward();
+        }
+      }
     }
   }
 }
 
 TEST_F(ModulesTest, CELU) {
   const auto size = 3;
-  for (const auto alpha : {0.42, 1.0, 4.2, 42.42}) {
-    CELU model {CELUOptions().alpha(alpha)};
-    auto x = torch::linspace(-10.0, 10.0, size * size * size);
-    x.resize_({size, size, size}).set_requires_grad(true);
-    auto y = model(x);
-    torch::Tensor s = y.sum();
+  for (const auto inplace : {false, true}) {
+    for (const auto alpha : {0.42, 1.0, 4.2, 42.42}) {
+      CELU model {CELUOptions().alpha(alpha).inplace(inplace)};
+      auto x = torch::linspace(-10.0, 10.0, size * size * size);
+      x.resize_({size, size, size});
+      if (!inplace) {
+        x.requires_grad_(true);
+      }
+      auto x_orig = x.clone();
+      auto y = model(x);
+      torch::Tensor s = y.sum();
 
-    s.backward();
-    ASSERT_EQ(s.ndimension(), 0);
-
-    ASSERT_EQ(y.ndimension(), 3);
-    ASSERT_EQ(y.sizes(), std::vector<int64_t>({size, size, size}));
-    auto y_exp = torch::max(torch::zeros_like(x), x) +
-        torch::min(torch::zeros_like(x), alpha * (torch::exp(x / alpha) - 1.0));
-    ASSERT_TRUE(torch::allclose(y, y_exp));
+      ASSERT_EQ(s.ndimension(), 0);
+      ASSERT_EQ(y.ndimension(), 3);
+      ASSERT_EQ(y.sizes(), std::vector<int64_t>({size, size, size}));
+      auto y_exp = torch::max(torch::zeros_like(x_orig), x_orig) +
+          torch::min(torch::zeros_like(x_orig), alpha * (torch::exp(x_orig / alpha) - 1.0));
+      ASSERT_TRUE(torch::allclose(y, y_exp));
+      if (inplace) {
+        ASSERT_TRUE(torch::allclose(x, y_exp));
+      } else {
+        s.backward();
+      }
+    }
   }
 }
 
@@ -2597,12 +2697,16 @@ TEST_F(ModulesTest, Threshold) {
         Threshold model {ThresholdOptions(threshold, value).inplace(inplace)};
         auto x = torch::linspace(-3.0, 3.0, 61);
         x.resize_({size, size, size});
-        auto y_exp = (x <= threshold) * value + (x > threshold) * x;
+        auto x_orig = x.clone();
+        auto y_exp = (x_orig <= threshold) * value + (x_orig > threshold) * x_orig;
         auto y = model(x);
 
         ASSERT_EQ(y.ndimension(), 3);
         ASSERT_EQ(y.sizes(), std::vector<int64_t>({size, size, size}));
         ASSERT_TRUE(torch::allclose(y, y_exp));
+        if (inplace) {
+          ASSERT_TRUE(torch::allclose(x, y_exp));
+        }
       }
     }
   }
@@ -2611,7 +2715,7 @@ TEST_F(ModulesTest, Threshold) {
 TEST_F(ModulesTest, Upsampling1D) {
   {
     Upsample model(UpsampleOptions()
-                       .size({4})
+                       .size(std::vector<int64_t>({4}))
                        .mode(torch::kNearest));
     auto input = torch::ones({1, 1, 2}, torch::requires_grad());
     auto output = model->forward(input);
@@ -2627,7 +2731,7 @@ TEST_F(ModulesTest, Upsampling1D) {
       // test float scale factor up & down sampling
       for (const auto scale_factor : {0.5, 1.5, 2.0}) {
         Upsample model(UpsampleOptions()
-                           .scale_factor({scale_factor})
+                           .scale_factor(std::vector<double>({scale_factor}))
                            .mode(torch::kLinear)
                            .align_corners(align_corners));
         auto input = torch::ones({1, 1, 2}, torch::requires_grad());
@@ -2646,7 +2750,7 @@ TEST_F(ModulesTest, Upsampling1D) {
   {
     // linear (1D) upsampling spatial invariance
     Upsample model(UpsampleOptions()
-                       .scale_factor({3})
+                       .scale_factor(std::vector<double>({3}))
                        .mode(torch::kLinear)
                        .align_corners(false));
     auto input = torch::zeros({1, 1, 9});
@@ -2661,7 +2765,7 @@ TEST_F(ModulesTest, Upsampling1D) {
 TEST_F(ModulesTest, Upsampling2D) {
   {
     Upsample model(UpsampleOptions()
-                       .size({4, 4})
+                       .size(std::vector<int64_t>({4, 4}))
                        .mode(torch::kNearest));
     auto input = torch::ones({1, 1, 2, 2}, torch::requires_grad());
     auto output = model->forward(input);
@@ -2677,7 +2781,7 @@ TEST_F(ModulesTest, Upsampling2D) {
       // test float scale factor up & down sampling
       for (const auto scale_factor : {0.5, 1.5, 2.0}) {
         Upsample model(UpsampleOptions()
-                           .scale_factor({scale_factor, scale_factor})
+                           .scale_factor(std::vector<double>({scale_factor, scale_factor}))
                            .mode(torch::kBilinear)
                            .align_corners(align_corners));
         auto input = torch::ones({1, 1, 2, 2}, torch::requires_grad());
@@ -2698,7 +2802,7 @@ TEST_F(ModulesTest, Upsampling2D) {
       // test float scale factor up & down sampling
       for (const auto scale_factor : {0.5, 1.5, 2.0}) {
         Upsample model(UpsampleOptions()
-                           .scale_factor({scale_factor, scale_factor})
+                           .scale_factor(std::vector<double>({scale_factor, scale_factor}))
                            .mode(torch::kBicubic)
                            .align_corners(align_corners));
         auto input = torch::ones({1, 1, 2, 2}, torch::requires_grad());
@@ -2719,7 +2823,7 @@ TEST_F(ModulesTest, Upsampling2D) {
 TEST_F(ModulesTest, Upsampling3D) {
   {
     Upsample model(UpsampleOptions()
-                       .size({4, 4, 4})
+                       .size(std::vector<int64_t>({4, 4, 4}))
                        .mode(torch::kNearest));
     auto input = torch::ones({1, 1, 2, 2, 2}, torch::requires_grad());
     auto output = model->forward(input);
@@ -2736,7 +2840,7 @@ TEST_F(ModulesTest, Upsampling3D) {
       for (const auto scale_factor : {0.5, 1.5, 2.0}) {
         Upsample model(
             UpsampleOptions()
-                .scale_factor({scale_factor, scale_factor, scale_factor})
+                .scale_factor(std::vector<double>({scale_factor, scale_factor, scale_factor}))
                 .mode(torch::kTrilinear)
                 .align_corners(align_corners));
         auto input = torch::ones({1, 1, 2, 2, 2}, torch::requires_grad());
@@ -3876,10 +3980,10 @@ TEST_F(ModulesTest, PrettyPrintConvTranspose) {
 
 TEST_F(ModulesTest, PrettyPrintUpsample) {
   ASSERT_EQ(
-      c10::str(Upsample(UpsampleOptions().size({2, 4, 4}))),
+      c10::str(Upsample(UpsampleOptions().size(std::vector<int64_t>({2, 4, 4})))),
       "torch::nn::Upsample(size=[2, 4, 4], mode=kNearest)");
   ASSERT_EQ(
-      c10::str(Upsample(UpsampleOptions().scale_factor({0.5, 1.5}).mode(torch::kBilinear))),
+      c10::str(Upsample(UpsampleOptions().scale_factor(std::vector<double>({0.5, 1.5})).mode(torch::kBilinear))),
       "torch::nn::Upsample(scale_factor=[0.5, 1.5], mode=kBilinear)");
 }
 
@@ -3987,15 +4091,27 @@ TEST_F(ModulesTest, PrettyPrintAdaptiveMaxPool) {
       c10::str(AdaptiveMaxPool2d(5)),
       "torch::nn::AdaptiveMaxPool2d(output_size=[5, 5])");
   ASSERT_EQ(
-      c10::str(AdaptiveMaxPool2d(std::vector<int64_t>{5, 6})),
+      c10::str(AdaptiveMaxPool2d(AdaptiveMaxPool2dOptions({5, 6}))),
       "torch::nn::AdaptiveMaxPool2d(output_size=[5, 6])");
+  ASSERT_EQ(
+      c10::str(AdaptiveMaxPool2d(AdaptiveMaxPool2dOptions({5, c10::nullopt}))),
+      "torch::nn::AdaptiveMaxPool2d(output_size=[5, None])");
+  ASSERT_EQ(
+      c10::str(AdaptiveMaxPool2d(AdaptiveMaxPool2dOptions({c10::nullopt, c10::nullopt}))),
+      "torch::nn::AdaptiveMaxPool2d(output_size=[None, None])");
 
   ASSERT_EQ(
       c10::str(AdaptiveMaxPool3d(5)),
       "torch::nn::AdaptiveMaxPool3d(output_size=[5, 5, 5])");
   ASSERT_EQ(
-      c10::str(AdaptiveMaxPool3d(std::vector<int64_t>{5, 6, 7})),
+      c10::str(AdaptiveMaxPool3d(AdaptiveMaxPool3dOptions({5, 6, 7}))),
       "torch::nn::AdaptiveMaxPool3d(output_size=[5, 6, 7])");
+  ASSERT_EQ(
+      c10::str(AdaptiveMaxPool3d(AdaptiveMaxPool3dOptions({5, c10::nullopt, 7}))),
+      "torch::nn::AdaptiveMaxPool3d(output_size=[5, None, 7])");
+  ASSERT_EQ(
+      c10::str(AdaptiveMaxPool3d(AdaptiveMaxPool3dOptions({c10::nullopt, c10::nullopt, c10::nullopt}))),
+      "torch::nn::AdaptiveMaxPool3d(output_size=[None, None, None])");
 }
 
 TEST_F(ModulesTest, PrettyPrintAdaptiveAvgPool) {
@@ -4007,15 +4123,27 @@ TEST_F(ModulesTest, PrettyPrintAdaptiveAvgPool) {
       c10::str(AdaptiveAvgPool2d(5)),
       "torch::nn::AdaptiveAvgPool2d(output_size=[5, 5])");
   ASSERT_EQ(
-      c10::str(AdaptiveAvgPool2d(std::vector<int64_t>{5, 6})),
+      c10::str(AdaptiveAvgPool2d(AdaptiveAvgPool2dOptions({5, 6}))),
       "torch::nn::AdaptiveAvgPool2d(output_size=[5, 6])");
+  ASSERT_EQ(
+      c10::str(AdaptiveAvgPool2d(AdaptiveAvgPool2dOptions({5, c10::nullopt}))),
+      "torch::nn::AdaptiveAvgPool2d(output_size=[5, None])");
+  ASSERT_EQ(
+      c10::str(AdaptiveAvgPool2d(AdaptiveAvgPool2dOptions({c10::nullopt, c10::nullopt}))),
+      "torch::nn::AdaptiveAvgPool2d(output_size=[None, None])");
 
   ASSERT_EQ(
       c10::str(AdaptiveAvgPool3d(5)),
       "torch::nn::AdaptiveAvgPool3d(output_size=[5, 5, 5])");
   ASSERT_EQ(
-      c10::str(AdaptiveAvgPool3d(std::vector<int64_t>{5, 6, 7})),
+      c10::str(AdaptiveAvgPool3d(AdaptiveAvgPool3dOptions({5, 6, 7}))),
       "torch::nn::AdaptiveAvgPool3d(output_size=[5, 6, 7])");
+  ASSERT_EQ(
+      c10::str(AdaptiveAvgPool3d(AdaptiveAvgPool3dOptions({5, c10::nullopt, 7}))),
+      "torch::nn::AdaptiveAvgPool3d(output_size=[5, None, 7])");
+  ASSERT_EQ(
+      c10::str(AdaptiveAvgPool3d(AdaptiveAvgPool3dOptions({c10::nullopt, c10::nullopt, c10::nullopt}))),
+      "torch::nn::AdaptiveAvgPool3d(output_size=[None, None, None])");
 }
 
 TEST_F(ModulesTest, PrettyPrintMaxUnpool) {

test/cpp/api/optim.cpp

@@ -26,7 +26,7 @@ bool test_optimizer_xor(Options options) {
       Linear(8, 1),
       Functional(torch::sigmoid));
 
-  const int64_t kBatchSize = 4;
+  const int64_t kBatchSize = 50;
   const int64_t kMaximumNumberOfEpochs = 3000;
 
   OptimizerClass optimizer(model->parameters(), options);
@@ -40,13 +40,21 @@ bool test_optimizer_xor(Options options) {
       inputs[i] = torch::randint(2, {2}, torch::kInt64);
       labels[i] = inputs[i][0].item<int64_t>() ^ inputs[i][1].item<int64_t>();
     }
-    inputs.set_requires_grad(true);
-    optimizer.zero_grad();
-    auto x = model->forward(inputs);
-    torch::Tensor loss = torch::binary_cross_entropy(x, labels);
-    loss.backward();
 
-    optimizer.step();
+    inputs.set_requires_grad(true);
+
+    auto step = [&](OptimizerClass& optimizer, Sequential model, torch::Tensor inputs, torch::Tensor labels) {
+      auto closure = [&]() {
+        optimizer.zero_grad();
+        auto x = model->forward(inputs);
+        auto loss = torch::binary_cross_entropy(x, labels);
+        loss.backward();
+        return loss;
+      };
+      return optimizer.step(closure);
+    };
+
+    torch::Tensor loss = step(optimizer, model, inputs, labels);
 
     running_loss = running_loss * 0.99 + loss.item<float>() * 0.01;
     if (epoch > kMaximumNumberOfEpochs) {
@@ -166,30 +174,66 @@ TEST(OptimTest, OptimizerAccessors) {
   optimizer_.state();
 }
 
-TEST(OptimTest, BasicInterface) {
+#define OLD_INTERFACE_WARNING_CHECK(func) \
+{ \
+  std::stringstream buffer;\
+  torch::test::CerrRedirect cerr_redirect(buffer.rdbuf());\
+  func;\
+  ASSERT_EQ(\
+    torch::test::count_substr_occurrences(\
+      buffer.str(),\
+      "will be removed"\
+    ),\
+  1);\
+}
+
+struct MyOptimizerOptions : public OptimizerCloneableOptions<MyOptimizerOptions> {
+  MyOptimizerOptions(double lr = 1.0) : lr_(lr) {};
+  TORCH_ARG(double, lr) = 1.0;
+};
+
+TEST(OptimTest, OldInterface) {
   struct MyOptimizer : Optimizer {
     using Optimizer::Optimizer;
     torch::Tensor step(LossClosure closure = nullptr) override { return {};}
+    explicit MyOptimizer(
+        std::vector<at::Tensor> params, MyOptimizerOptions defaults = {}) :
+          Optimizer({std::move(OptimizerParamGroup(params))}, std::make_unique<MyOptimizerOptions>(defaults)) {}
   };
   std::vector<torch::Tensor> parameters = {
       torch::ones({2, 3}), torch::zeros({2, 3}), torch::rand({2, 3})};
   {
     MyOptimizer optimizer(parameters);
-    ASSERT_EQ(optimizer.size(), parameters.size());
+    size_t size;
+    OLD_INTERFACE_WARNING_CHECK(size = optimizer.size());
+    ASSERT_EQ(size, parameters.size());
   }
   {
-    MyOptimizer optimizer;
-    ASSERT_EQ(optimizer.size(), 0);
-    optimizer.add_parameters(parameters);
-    ASSERT_EQ(optimizer.size(), parameters.size());
-    for (size_t p = 0; p < parameters.size(); ++p) {
-      ASSERT_TRUE(optimizer.parameters()[p].allclose(parameters[p]));
+    std::vector<at::Tensor> params;
+    MyOptimizer optimizer(params);
+
+    size_t size;
+    OLD_INTERFACE_WARNING_CHECK(size = optimizer.size());
+    ASSERT_EQ(size, 0);
+
+    OLD_INTERFACE_WARNING_CHECK(optimizer.add_parameters(parameters));
+
+    OLD_INTERFACE_WARNING_CHECK(size = optimizer.size());
+    ASSERT_EQ(size, parameters.size());
+
+    std::vector<torch::Tensor> params_;
+    OLD_INTERFACE_WARNING_CHECK(params_ = optimizer.parameters());
+    for (size_t p = 0; p < size; ++p) {
+      ASSERT_TRUE(params_[p].allclose(parameters[p]));
     }
   }
   {
     Linear linear(3, 4);
     MyOptimizer optimizer(linear->parameters());
-    ASSERT_EQ(optimizer.size(), linear->parameters().size());
+
+    size_t size;
+    OLD_INTERFACE_WARNING_CHECK(size = optimizer.size());
+    ASSERT_EQ(size, linear->parameters().size());
   }
 }
 
@@ -198,6 +242,11 @@ TEST(OptimTest, XORConvergence_SGD) {
       SGDOptions(0.1).momentum(0.9).nesterov(true).weight_decay(1e-6)));
 }
 
+TEST(OptimTest, XORConvergence_LBFGS) {
+  ASSERT_TRUE(test_optimizer_xor<LBFGS>(LBFGSOptions(1.0)));
+  ASSERT_TRUE(test_optimizer_xor<LBFGS>(LBFGSOptions(1.0).line_search_fn("strong_wolfe")));
+}
+
 TEST(OptimTest, XORConvergence_Adagrad) {
   ASSERT_TRUE(test_optimizer_xor<Adagrad>(
       AdagradOptions(1.0).weight_decay(1e-6).lr_decay(1e-3)));
@@ -375,7 +424,7 @@ TEST(OptimTest, AddParameter_LBFGS) {
   }
 
   LBFGS optimizer(std::vector<torch::Tensor>{}, 1.0);
-  optimizer.add_parameters(parameters);
+  OLD_INTERFACE_WARNING_CHECK(optimizer.add_parameters(parameters));
 
   optimizer.step([]() { return torch::tensor(1); });
 

test/cpp/api/serialize.cpp

@@ -64,7 +64,7 @@ void is_optimizer_state_equal(
 }
 
 template <typename OptimizerClass, typename DerivedOptimizerOptions, typename DerivedOptimizerParamState>
-void test_serialize_optimizer(DerivedOptimizerOptions options) {
+void test_serialize_optimizer(DerivedOptimizerOptions options, bool only_has_global_state = false) {
   auto model1 = Linear(5, 2);
   auto model2 = Linear(5, 2);
   auto model3 = Linear(5, 2);
@@ -125,9 +125,11 @@ void test_serialize_optimizer(DerivedOptimizerOptions options) {
   auto& optim3_2_state = optim3_2.state();
   auto& optim3_state = optim3.state();
 
-  // optim3_2 and optim1 should have param_groups and state of size 1 and 2 respectively
+  // optim3_2 and optim1 should have param_groups and state of size 1 and state_size respectively
   ASSERT_TRUE(optim3_2_param_groups.size() == 1);
-  ASSERT_TRUE(optim3_2_state.size() == 2);
+  // state_size = 2 for all optimizers except LBFGS as LBFGS only maintains one global state
+  int state_size = only_has_global_state ? 1 : 2;
+  ASSERT_TRUE(optim3_2_state.size() == state_size);
 
   // optim3_2 and optim1 should have param_groups and state of same size
   ASSERT_TRUE(optim3_2_param_groups.size() == optim3_param_groups.size());
@@ -668,39 +670,16 @@ TEST(SerializeTest, Optim_RMSprop) {
 }
 
 TEST(SerializeTest, Optim_LBFGS) {
-  auto options = LBFGSOptions();
+  test_serialize_optimizer<LBFGS, LBFGSOptions, LBFGSParamState>(LBFGSOptions(), true);
+  // bc compatibility check
   auto model1 = Linear(5, 2);
-  auto model2 = Linear(5, 2);
-  auto model3 = Linear(5, 2);
-
-  // Models 1, 2, 3 will have the same parameters.
-  auto model_tempfile = c10::make_tempfile();
-  torch::save(model1, model_tempfile.name);
-  torch::load(model2, model_tempfile.name);
-  torch::load(model3, model_tempfile.name);
-
-  auto param1 = model1->named_parameters();
-  auto param2 = model2->named_parameters();
-  auto param3 = model3->named_parameters();
-  for (const auto& p : param1) {
-    ASSERT_TRUE(p->allclose(param2[p.key()]));
-    ASSERT_TRUE(param2[p.key()].allclose(param3[p.key()]));
-  }
-  // Make some optimizers
-  auto optim1 = LBFGS(
-      {torch::optim::OptimizerParamGroup(model1->parameters())}, options);
-  auto optim2 = LBFGS(
-      model2->parameters(), options);
-  auto optim2_2 = LBFGS(
-      model2->parameters(), options);
-  auto optim3 = LBFGS(
-      model3->parameters(), options);
-  auto optim3_2 = LBFGS(
-      model3->parameters(), options);
+  auto model1_params = model1->parameters();
+  // added a tensor for lazy init check - when all params do not have entry in buffers
+  model1_params.emplace_back(torch::randn({2,3}));
+  auto optim1 = torch::optim::LBFGS(model1_params, torch::optim::LBFGSOptions());
 
   auto x = torch::ones({10, 5});
-
-  auto step = [&x](torch::optim::LossClosureOptimizer& optimizer, Linear model) {
+  auto step = [&x](torch::optim::Optimizer& optimizer, Linear model) {
     optimizer.zero_grad();
     auto y = model->forward(x).sum();
     y.backward();
@@ -708,56 +687,47 @@ TEST(SerializeTest, Optim_LBFGS) {
     optimizer.step(closure);
   };
 
-  // Do 2 steps of model1
-  step(optim1, model1);
   step(optim1, model1);
 
-  // Do 2 steps of model 2 without saving the optimizer
-  step(optim2, model2);
-  step(optim2_2, model2);
+  at::Tensor d, t, H_diag, prev_flat_grad, prev_loss;
+  std::deque<at::Tensor> old_dirs, old_stps;
 
-  // Do 1 step of model 3
-  step(optim3, model3);
+  const auto& params_ = optim1.param_groups()[0].params();
+  auto key_ = c10::guts::to_string(params_[0].unsafeGetTensorImpl());
+  const auto& optim1_state = static_cast<const LBFGSParamState&>(*(optim1.state().at(key_).get()));
+  d = optim1_state.d();
+  t = at::tensor(optim1_state.t());
+  H_diag = optim1_state.H_diag();
+  prev_flat_grad = optim1_state.prev_flat_grad();
+  prev_loss = at::tensor(optim1_state.prev_loss());
+  old_dirs = optim1_state.old_dirs();
 
-  // save the optimizer
-  auto optim_tempfile = c10::make_tempfile();
-  torch::save(optim3, optim_tempfile.name);
-  torch::load(optim3_2, optim_tempfile.name);
+  // write buffers to the file
+  auto optim_tempfile_old_format = c10::make_tempfile();
+  torch::serialize::OutputArchive output_archive;
+  output_archive.write("d", d, /*is_buffer=*/true);
+  output_archive.write("t", t, /*is_buffer=*/true);
+  output_archive.write("H_diag", H_diag, /*is_buffer=*/true);
+  output_archive.write("prev_flat_grad", prev_flat_grad, /*is_buffer=*/true);
+  output_archive.write("prev_loss", prev_loss, /*is_buffer=*/true);
+  write_tensors_to_archive(output_archive, "old_dirs", old_dirs);
+  write_tensors_to_archive(output_archive, "old_stps", old_stps);
+  output_archive.save_to(optim_tempfile_old_format.name);
 
-  auto& optim3_2_param_groups = optim3_2.param_groups();
-  auto& optim3_param_groups = optim3.param_groups();
-  auto& optim3_2_state = optim3_2.state();
-  auto& optim3_state = optim3.state();
+  auto optim1_2 = LBFGS(model1_params, torch::optim::LBFGSOptions());
+  OLD_SERIALIZATION_LOGIC_WARNING_CHECK(torch::load, optim1_2, optim_tempfile_old_format.name);
 
-  // LBFGS only supports 1 param_group
-  // optim3_2 and optim1 should have param_groups of size 1
-  ASSERT_TRUE(optim3_param_groups.size() == 1);
-  ASSERT_TRUE(optim3_2_param_groups.size() == 1);
-  // LBFGS only maintains one global state
-  ASSERT_TRUE(optim3_2_state.size() == 1);
-  ASSERT_TRUE(optim3_state.size() == 1);
+  const auto& params1_2_ = optim1_2.param_groups()[0].params();
+  auto param_key = c10::guts::to_string(params1_2_[0].unsafeGetTensorImpl());
+  auto& optim1_2_state = static_cast<LBFGSParamState&>(*(optim1_2.state().at(param_key).get()));
 
-  // checking correctness of serialization logic for optimizer.param_groups_ and optimizer.state_
-  for (int i = 0; i < optim3_2_param_groups.size(); i++) {
-    is_optimizer_param_group_equal<LBFGSOptions>(
-      optim3_2_param_groups[i], optim3_param_groups[i]);
-    is_optimizer_state_equal<LBFGSParamState>(optim3_2_state, optim3_state);
-  }
+  // old LBFGS didn't track func_evals, n_iter, ro, al values
+  optim1_2_state.func_evals(optim1_state.func_evals());
+  optim1_2_state.n_iter(optim1_state.n_iter());
+  optim1_2_state.ro(optim1_state.ro());
+  optim1_2_state.al(optim1_state.al());
 
-  // Do step2 for model 3
-  step(optim3_2, model3);
-
-  param1 = model1->named_parameters();
-  param2 = model2->named_parameters();
-  param3 = model3->named_parameters();
-  for (const auto& p : param1) {
-    const auto& name = p.key();
-    // Model 1 and 3 should be the same
-    ASSERT_TRUE(
-        param1[name].norm().item<float>() == param3[name].norm().item<float>());
-    ASSERT_TRUE(
-        param1[name].norm().item<float>() != param2[name].norm().item<float>());
-  }
+  is_optimizer_state_equal<LBFGSParamState>(optim1.state(), optim1_2.state());
 }
 
 TEST(SerializeTest, XOR_CUDA) {

test/cpp/jit/test_class_import.cpp

@@ -138,7 +138,7 @@ void testClassDerive() {
 static const auto torchbindSrc = R"JIT(
 class FooBar1234(Module):
   __parameters__ = []
-  f : __torch__.torch.classes._TorchScriptTesting_StackString
+  f : __torch__.torch.classes._TorchScriptTesting._StackString
   training : bool
   def forward(self: __torch__.FooBar1234) -> str:
     return (self.f).top()

test/cpp/jit/test_custom_class.cpp

@@ -66,7 +66,7 @@ struct PickleTester : torch::CustomClassHolder {
   std::vector<int64_t> vals;
 };
 
-static auto test = torch::class_<Foo>("_TorchScriptTesting_Foo")
+static auto test = torch::class_<Foo>("_TorchScriptTesting", "_Foo")
                        .def(torch::init<int64_t, int64_t>())
                        // .def(torch::init<>())
                        .def("info", &Foo::info)
@@ -75,7 +75,9 @@ static auto test = torch::class_<Foo>("_TorchScriptTesting_Foo")
                        .def("combine", &Foo::combine);
 
 static auto testStack =
-    torch::class_<MyStackClass<std::string>>("_TorchScriptTesting_StackString")
+    torch::class_<MyStackClass<std::string>>(
+        "_TorchScriptTesting",
+        "_StackString")
         .def(torch::init<std::vector<std::string>>())
         .def("push", &MyStackClass<std::string>::push)
         .def("pop", &MyStackClass<std::string>::pop)
@@ -101,7 +103,7 @@ static auto testStack =
 // clang-format on
 
 static auto testPickle =
-    torch::class_<PickleTester>("_TorchScriptTesting_PickleTester")
+    torch::class_<PickleTester>("_TorchScriptTesting", "_PickleTester")
         .def(torch::init<std::vector<int64_t>>())
         .def_pickle(
             [](c10::intrusive_ptr<PickleTester> self) { // __getstate__
@@ -127,10 +129,10 @@ at::Tensor take_an_instance(const c10::intrusive_ptr<PickleTester>& instance) {
 
 torch::RegisterOperators& register_take_instance() {
   static auto instance_registry = torch::RegisterOperators().op(
-  torch::RegisterOperators::options()
-      .schema(
-          "_TorchScriptTesting::take_an_instance(__torch__.torch.classes._TorchScriptTesting_PickleTester x) -> Tensor Y")
-      .catchAllKernel<decltype(take_an_instance), &take_an_instance>());
+      torch::RegisterOperators::options()
+          .schema(
+              "_TorchScriptTesting::take_an_instance(__torch__.torch.classes._TorchScriptTesting._PickleTester x) -> Tensor Y")
+          .catchAllKernel<decltype(take_an_instance), &take_an_instance>());
   return instance_registry;
 }
 
@@ -146,7 +148,7 @@ void testTorchbindIValueAPI() {
   auto custom_class_obj = make_custom_class<MyStackClass<std::string>>(
       std::vector<std::string>{"foo", "bar"});
   m.define(R"(
-    def forward(self, s : __torch__.torch.classes._TorchScriptTesting_StackString):
+    def forward(self, s : __torch__.torch.classes._TorchScriptTesting._StackString):
       return s.pop(), s
   )");
 

test/cpp/jit/test_lite_interpreter.cpp

@@ -343,7 +343,8 @@ void testLiteInterpreterBuiltinFunction() {
 namespace {
 static auto reg =
     torch::jit::class_<TorchBindLiteInterpreterTestStruct>(
-        "_TorchScriptTesting_LiteInterpreterTest")
+        "_TorchScriptTesting",
+        "_LiteInterpreterTest")
         .def("get", &TorchBindLiteInterpreterTestStruct::get)
         .def_pickle(
             // __getattr__

test/custom_operator/test_custom_classes.py

@@ -35,19 +35,19 @@ class TestCustomOperators(unittest.TestCase):
 
     def test_no_return_class(self):
         def f():
-            val = torch.classes._TorchScriptTesting_Foo(5, 3)
+            val = torch.classes._TorchScriptTesting._Foo(5, 3)
             return val.info()
         self.assertEqual(*test_equality(f, lambda x: x))
 
     def test_constructor_with_args(self):
         def f():
-            val = torch.classes._TorchScriptTesting_Foo(5, 3)
+            val = torch.classes._TorchScriptTesting._Foo(5, 3)
             return val
         self.assertEqual(*test_equality(f, lambda x: x.info()))
 
     def test_function_call_with_args(self):
         def f():
-            val = torch.classes._TorchScriptTesting_Foo(5, 3)
+            val = torch.classes._TorchScriptTesting._Foo(5, 3)
             val.increment(1)
             return val
 
@@ -55,7 +55,7 @@ class TestCustomOperators(unittest.TestCase):
 
     def test_function_method_wrong_type(self):
         def f():
-            val = torch.classes._TorchScriptTesting_Foo(5, 3)
+            val = torch.classes._TorchScriptTesting._Foo(5, 3)
             val.increment("asdf")
             return val
 
@@ -65,8 +65,8 @@ class TestCustomOperators(unittest.TestCase):
     @unittest.skip("We currently don't support passing custom classes to custom methods.")
     def test_input_class_type(self):
         def f():
-            val = torch.classes._TorchScriptTesting_Foo(1, 2)
-            val2 = torch.classes._TorchScriptTesting_Foo(2, 3)
+            val = torch.classes._TorchScriptTesting._Foo(1, 2)
+            val2 = torch.classes._TorchScriptTesting._Foo(2, 3)
             val.combine(val2)
             return val
 
@@ -74,14 +74,14 @@ class TestCustomOperators(unittest.TestCase):
 
     def test_stack_string(self):
         def f():
-            val = torch.classes._TorchScriptTesting_StackString(["asdf", "bruh"])
+            val = torch.classes._TorchScriptTesting._StackString(["asdf", "bruh"])
             return val.pop()
         self.assertEqual(*test_equality(f, lambda x: x))
 
     def test_stack_push_pop(self):
         def f():
-            val = torch.classes._TorchScriptTesting_StackString(["asdf", "bruh"])
-            val2 = torch.classes._TorchScriptTesting_StackString(["111", "222"])
+            val = torch.classes._TorchScriptTesting._StackString(["asdf", "bruh"])
+            val2 = torch.classes._TorchScriptTesting._StackString(["111", "222"])
             val.push(val2.pop())
             return val.pop() + val2.pop()
         self.assertEqual(*test_equality(f, lambda x: x))

test/onnx/expect/TestOperators.test_acos.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_add_broadcast.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_add_left_broadcast.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_add_size1_broadcast.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_add_size1_right_broadcast.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_add_size1_singleton_broadcast.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_addconstant.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     output: "1"

test/onnx/expect/TestOperators.test_addmm.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_arange_dynamic.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_argmax.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_asin.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_at_op.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "x"

test/onnx/expect/TestOperators.test_atan.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_avg_pool2d.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_baddbmm.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "1"

test/onnx/expect/TestOperators.test_basic.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_batchnorm.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "input"

test/onnx/expect/TestOperators.test_batchnorm_1d.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "input"

test/onnx/expect/TestOperators.test_batchnorm_noaffine.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     output: "4"

test/onnx/expect/TestOperators.test_batchnorm_onnx_irv4.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "input"

test/onnx/expect/TestOperators.test_batchnorm_training.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "input"

test/onnx/expect/TestOperators.test_bitshift.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     output: "4"

test/onnx/expect/TestOperators.test_c2_op.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_chunk.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_clip.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_clip_max.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_clip_min.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_concat2.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_conv.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "input"

test/onnx/expect/TestOperators.test_conv_onnx_irv4.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "input"

test/onnx/expect/TestOperators.test_conv_onnx_irv4_opset8.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "input"

test/onnx/expect/TestOperators.test_convtranspose.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_cos.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_cumsum.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     output: "1"

test/onnx/expect/TestOperators.test_det.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_dict.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "1"

test/onnx/expect/TestOperators.test_dict_str.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     output: "1"

test/onnx/expect/TestOperators.test_dim.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     output: "1"

test/onnx/expect/TestOperators.test_dropout.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "x"

test/onnx/expect/TestOperators.test_elu.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "input"

test/onnx/expect/TestOperators.test_embedding_bags.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "weight"

test/onnx/expect/TestOperators.test_empty_like.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_empty_like_opset7.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_equal.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "x"

test/onnx/expect/TestOperators.test_erf.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_exp.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_expand.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     output: "1"

test/onnx/expect/TestOperators.test_flatten.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     output: "1"

test/onnx/expect/TestOperators.test_flatten2D.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_fmod.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_frobenius_norm.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "x"

test/onnx/expect/TestOperators.test_full.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_full_like.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_gather.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     output: "2"

test/onnx/expect/TestOperators.test_gather_opset11.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "0"

test/onnx/expect/TestOperators.test_ge.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "x"

test/onnx/expect/TestOperators.test_gelu.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     output: "1"

test/onnx/expect/TestOperators.test_gt.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "x"

test/onnx/expect/TestOperators.test_hardtanh.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     input: "input"

test/onnx/expect/TestOperators.test_implicit_expand.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     output: "1"

test/onnx/expect/TestOperators.test_index.expect

@@ -1,6 +1,6 @@
 ir_version: 6
 producer_name: "pytorch"
-producer_version: "1.4"
+producer_version: "1.5"
 graph {
   node {
     output: "1"