pytorch 1.7 support

Summary: CircleCI build configuration to support pytorch1.7, including binaries with cuda 11.0. Note that the default torch on pip is still on cuda 10.2, so I have left the `main` (non conda build) on cuda 10.2 with the existing driver.

Reviewed By: gkioxari

Differential Revision: D24623523

fbshipit-source-id: 59cfa1be06c16225f0f12ed336c07220e8a9a511

.circleci/config.in.yml

@@ -18,9 +18,7 @@ setupcuda: &setupcuda
     working_directory: ~/
     command: |
       # download and install nvidia drivers, cuda, etc
-      wget --no-verbose --no-clobber -P ~/nvidia-downloads 'https://s3.amazonaws.com/ossci-linux/nvidia_driver/NVIDIA-Linux-x86_64-430.40.run'
       wget --no-verbose --no-clobber -P ~/nvidia-downloads http://developer.download.nvidia.com/compute/cuda/10.2/Prod/local_installers/cuda_10.2.89_440.33.01_linux.run
-      sudo /bin/bash ~/nvidia-downloads/NVIDIA-Linux-x86_64-430.40.run --no-drm -q --ui=none
       sudo sh ~/nvidia-downloads/cuda_10.2.89_440.33.01_linux.run --silent
       echo "Done installing CUDA."
       pyenv versions
@@ -157,8 +155,8 @@ jobs:
           sudo apt-get update && sudo apt-get install -y nvidia-container-toolkit=${NVIDIA_CONTAINER_VERSION}
           sudo systemctl restart docker
 
-          DRIVER_FN="NVIDIA-Linux-x86_64-440.59.run"
-          wget "https://s3.amazonaws.com/ossci-linux/nvidia_driver/$DRIVER_FN"
+          DRIVER_FN="NVIDIA-Linux-x86_64-450.80.02.run"
+          wget "https://us.download.nvidia.com/XFree86/Linux-x86_64/450.80.02/$DRIVER_FN"
           sudo /bin/bash "$DRIVER_FN" -s --no-drm || (sudo cat /var/log/nvidia-installer.log && false)
           nvidia-smi
 
@@ -178,7 +176,7 @@ jobs:
           cd ${HOME}/project/
 
           export DOCKER_IMAGE=pytorch/conda-cuda
-          export VARS_TO_PASS="-e PYTHON_VERSION -e BUILD_VERSION -e PYTORCH_VERSION -e UNICODE_ABI -e CU_VERSION"
+          export VARS_TO_PASS="-e PYTHON_VERSION -e BUILD_VERSION -e PYTORCH_VERSION -e CU_VERSION"
 
           docker run --gpus all  --ipc=host -v $(pwd):/remote -w /remote ${VARS_TO_PASS} ${DOCKER_IMAGE} ./packaging/build_conda.sh
 
@@ -216,6 +214,11 @@ workflows:
           python_version: "3.7"
           pytorch_version: '1.6.0'
           cu_version: "cu102"
+      - binary_linux_conda_cuda:
+          name: testrun_conda_cuda_py37_cu110_pyt170
+          python_version: "3.7"
+          pytorch_version: '1.7.0'
+          cu_version: "cu110"
       - binary_macos_wheel:
           cu_version: cpu
           name: macos_wheel_py36_cpu

.circleci/config.yml

@@ -18,9 +18,7 @@ setupcuda: &setupcuda
     working_directory: ~/
     command: |
       # download and install nvidia drivers, cuda, etc
-      wget --no-verbose --no-clobber -P ~/nvidia-downloads 'https://s3.amazonaws.com/ossci-linux/nvidia_driver/NVIDIA-Linux-x86_64-430.40.run'
       wget --no-verbose --no-clobber -P ~/nvidia-downloads http://developer.download.nvidia.com/compute/cuda/10.2/Prod/local_installers/cuda_10.2.89_440.33.01_linux.run
-      sudo /bin/bash ~/nvidia-downloads/NVIDIA-Linux-x86_64-430.40.run --no-drm -q --ui=none
       sudo sh ~/nvidia-downloads/cuda_10.2.89_440.33.01_linux.run --silent
       echo "Done installing CUDA."
       pyenv versions
@@ -157,8 +155,8 @@ jobs:
           sudo apt-get update && sudo apt-get install -y nvidia-container-toolkit=${NVIDIA_CONTAINER_VERSION}
           sudo systemctl restart docker
 
-          DRIVER_FN="NVIDIA-Linux-x86_64-440.59.run"
-          wget "https://s3.amazonaws.com/ossci-linux/nvidia_driver/$DRIVER_FN"
+          DRIVER_FN="NVIDIA-Linux-x86_64-450.80.02.run"
+          wget "https://us.download.nvidia.com/XFree86/Linux-x86_64/450.80.02/$DRIVER_FN"
           sudo /bin/bash "$DRIVER_FN" -s --no-drm || (sudo cat /var/log/nvidia-installer.log && false)
           nvidia-smi
 
@@ -178,7 +176,7 @@ jobs:
           cd ${HOME}/project/
 
           export DOCKER_IMAGE=pytorch/conda-cuda
-          export VARS_TO_PASS="-e PYTHON_VERSION -e BUILD_VERSION -e PYTORCH_VERSION -e UNICODE_ABI -e CU_VERSION"
+          export VARS_TO_PASS="-e PYTHON_VERSION -e BUILD_VERSION -e PYTORCH_VERSION -e CU_VERSION"
 
           docker run --gpus all  --ipc=host -v $(pwd):/remote -w /remote ${VARS_TO_PASS} ${DOCKER_IMAGE} ./packaging/build_conda.sh
 
@@ -260,6 +258,21 @@ workflows:
           name: linux_conda_py36_cu102_pyt160
           python_version: '3.6'
           pytorch_version: 1.6.0
+      - binary_linux_conda:
+          cu_version: cu101
+          name: linux_conda_py36_cu101_pyt170
+          python_version: '3.6'
+          pytorch_version: 1.7.0
+      - binary_linux_conda:
+          cu_version: cu102
+          name: linux_conda_py36_cu102_pyt170
+          python_version: '3.6'
+          pytorch_version: 1.7.0
+      - binary_linux_conda:
+          cu_version: cu110
+          name: linux_conda_py36_cu110_pyt170
+          python_version: '3.6'
+          pytorch_version: 1.7.0
       - binary_linux_conda:
           cu_version: cu92
           name: linux_conda_py37_cu92_pyt14
@@ -315,6 +328,21 @@ workflows:
           name: linux_conda_py37_cu102_pyt160
           python_version: '3.7'
           pytorch_version: 1.6.0
+      - binary_linux_conda:
+          cu_version: cu101
+          name: linux_conda_py37_cu101_pyt170
+          python_version: '3.7'
+          pytorch_version: 1.7.0
+      - binary_linux_conda:
+          cu_version: cu102
+          name: linux_conda_py37_cu102_pyt170
+          python_version: '3.7'
+          pytorch_version: 1.7.0
+      - binary_linux_conda:
+          cu_version: cu110
+          name: linux_conda_py37_cu110_pyt170
+          python_version: '3.7'
+          pytorch_version: 1.7.0
       - binary_linux_conda:
           cu_version: cu92
           name: linux_conda_py38_cu92_pyt14
@@ -370,6 +398,21 @@ workflows:
           name: linux_conda_py38_cu102_pyt160
           python_version: '3.8'
           pytorch_version: 1.6.0
+      - binary_linux_conda:
+          cu_version: cu101
+          name: linux_conda_py38_cu101_pyt170
+          python_version: '3.8'
+          pytorch_version: 1.7.0
+      - binary_linux_conda:
+          cu_version: cu102
+          name: linux_conda_py38_cu102_pyt170
+          python_version: '3.8'
+          pytorch_version: 1.7.0
+      - binary_linux_conda:
+          cu_version: cu110
+          name: linux_conda_py38_cu110_pyt170
+          python_version: '3.8'
+          pytorch_version: 1.7.0
       - binary_linux_wheel:
           cu_version: cu101
           name: linux_wheel_py36_cu101_pyt160
@@ -395,6 +438,11 @@ workflows:
           python_version: "3.7"
           pytorch_version: '1.6.0'
           cu_version: "cu102"
+      - binary_linux_conda_cuda:
+          name: testrun_conda_cuda_py37_cu110_pyt170
+          python_version: "3.7"
+          pytorch_version: '1.7.0'
+          cu_version: "cu110"
       - binary_macos_wheel:
           cu_version: cpu
           name: macos_wheel_py36_cpu

.circleci/regenerate.py

@@ -19,6 +19,7 @@ CONDA_CUDA_VERSIONS = {
     "1.5.0": ["cu92", "cu101", "cu102"],
     "1.5.1": ["cu92", "cu101", "cu102"],
     "1.6.0": ["cu92", "cu101", "cu102"],
+    "1.7.0": ["cu101", "cu102", "cu110"],
 }
 
 
@@ -26,7 +27,7 @@ def workflows(prefix="", filter_branch=None, upload=False, indentation=6):
     w = []
     for btype in ["conda"]:
         for python_version in ["3.6", "3.7", "3.8"]:
-            for pytorch_version in ["1.4", "1.5.0", "1.5.1", "1.6.0"]:
+            for pytorch_version in ["1.4", "1.5.0", "1.5.1", "1.6.0", "1.7.0"]:
                 for cu_version in CONDA_CUDA_VERSIONS[pytorch_version]:
                     w += workflow_pair(
                         btype=btype,

packaging/pkg_helpers.bash

@@ -51,6 +51,17 @@ setup_cuda() {
 
   # Now work out the CUDA settings
   case "$CU_VERSION" in
+    cu110)
+      if [[ "$OSTYPE" == "msys" ]]; then
+        export CUDA_HOME="C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v11.0"
+      else
+        export CUDA_HOME=/usr/local/cuda-11.0/
+      fi
+      export FORCE_CUDA=1
+      # Hard-coding gencode flags is temporary situation until
+      # https://github.com/pytorch/pytorch/pull/23408 lands
+      export NVCC_FLAGS="-gencode=arch=compute_35,code=sm_35 -gencode=arch=compute_50,code=sm_50 -gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_80,code=sm_80 -gencode=arch=compute_50,code=compute_50"
+      ;;
     cu102)
       if [[ "$OSTYPE" == "msys" ]]; then
         export CUDA_HOME="C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v10.2"
@@ -244,6 +255,9 @@ setup_conda_cudatoolkit_constraint() {
     export CONDA_CUDATOOLKIT_CONSTRAINT=""
   else
     case "$CU_VERSION" in
+      cu110)
+        export CONDA_CUDATOOLKIT_CONSTRAINT="- cudatoolkit >=11.0,<11.1 # [not osx]"
+        ;;
       cu102)
         export CONDA_CUDATOOLKIT_CONSTRAINT="- cudatoolkit >=10.2,<10.3 # [not osx]"
         ;;