Merge branch 'main' into models/weight_update

Author: datumbox

.circleci/config.yml

@@ -268,9 +268,9 @@ def cmake_workflows(indentation=6):
         for device in device_types:
             job = {"name": f"cmake_{os_type}_{device}", "python_version": python_version}
 
-            job["cu_version"] = "cu102" if device == "gpu" else "cpu"
+            job["cu_version"] = "cu113" if device == "gpu" else "cpu"
             if device == "gpu" and os_type == "linux":
-                job["wheel_docker_image"] = "pytorch/manylinux-cuda102"
+                job["wheel_docker_image"] = "pytorch/manylinux-cuda113"
             jobs.append({f"cmake_{os_type}_{device}": job})
     return indent(indentation, jobs)
 

.circleci/regenerate.py

@@ -10,6 +10,11 @@ if(WITH_CUDA)
   add_definitions(-D__CUDA_NO_HALF_OPERATORS__)
   add_definitions(-DWITH_CUDA)
   set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --expt-relaxed-constexpr")
+  # CUDA-11.x can not be compiled using C++14 standard on Windows
+  string(REGEX MATCH "^[0-9]+" CUDA_MAJOR ${CMAKE_CUDA_COMPILER_VERSION})
+  if(${CUDA_MAJOR} GREATER 10 AND MSVC)
+    set(CMAKE_CXX_STANDARD 17)
+  endif()
 endif()
 
 find_package(Python3 COMPONENTS Development)

CMakeLists.txt

@@ -52,7 +52,6 @@ setup_cuda() {
       else
         export CUDA_HOME=/usr/local/cuda-11.3/
       fi
-      export FORCE_CUDA=1
       export TORCH_CUDA_ARCH_LIST="3.5;5.0+PTX;6.0;7.0;7.5;8.0;8.6"
       ;;
     cu112)
@@ -61,7 +60,6 @@ setup_cuda() {
       else
         export CUDA_HOME=/usr/local/cuda-11.2/
       fi
-      export FORCE_CUDA=1
       export TORCH_CUDA_ARCH_LIST="3.5;5.0+PTX;6.0;7.0;7.5;8.0;8.6"
       ;;
     cu111)
@@ -70,7 +68,6 @@ setup_cuda() {
       else
         export CUDA_HOME=/usr/local/cuda-11.1/
       fi
-      export FORCE_CUDA=1
       export TORCH_CUDA_ARCH_LIST="3.5;5.0+PTX;6.0;7.0;7.5;8.0;8.6"
       ;;
     cu110)
@@ -79,7 +76,6 @@ setup_cuda() {
       else
         export CUDA_HOME=/usr/local/cuda-11.0/
       fi
-      export FORCE_CUDA=1
       export TORCH_CUDA_ARCH_LIST="3.5;5.0+PTX;6.0;7.0;7.5;8.0"
       ;;
     cu102)
@@ -88,7 +84,6 @@ setup_cuda() {
       else
         export CUDA_HOME=/usr/local/cuda-10.2/
       fi
-      export FORCE_CUDA=1
       export TORCH_CUDA_ARCH_LIST="3.5;5.0+PTX;6.0;7.0;7.5"
       ;;
     cu101)
@@ -97,7 +92,6 @@ setup_cuda() {
       else
         export CUDA_HOME=/usr/local/cuda-10.1/
       fi
-      export FORCE_CUDA=1
       export TORCH_CUDA_ARCH_LIST="3.5;5.0+PTX;6.0;7.0;7.5"
       ;;
     cu100)
@@ -106,7 +100,6 @@ setup_cuda() {
       else
         export CUDA_HOME=/usr/local/cuda-10.0/
       fi
-      export FORCE_CUDA=1
       export TORCH_CUDA_ARCH_LIST="3.5;5.0+PTX;6.0;7.0;7.5"
       ;;
     cu92)
@@ -115,7 +108,6 @@ setup_cuda() {
       else
         export CUDA_HOME=/usr/local/cuda-9.2/
       fi
-      export FORCE_CUDA=1
       export TORCH_CUDA_ARCH_LIST="3.5;5.0+PTX;6.0;7.0"
       ;;
     cpu)
@@ -128,6 +120,11 @@ setup_cuda() {
       exit 1
       ;;
   esac
+  if [[ -n "$CUDA_HOME" ]]; then
+    # Adds nvcc binary to the search path so that CMake's `find_package(CUDA)` will pick the right one
+    export PATH="$CUDA_HOME/bin:$PATH"
+    export FORCE_CUDA=1
+  fi
 }
 
 # Populate build version if necessary, and add version suffix