diff --git a/.circleci/config.yml b/.circleci/config.yml
index d2adfb00592..a06d379c2d1 100644
--- a/.circleci/config.yml
+++ b/.circleci/config.yml
@@ -214,6 +214,11 @@ jobs:
     steps:
       - checkout_merge
       - designate_upload_channel
+      - run:
+          name: _HACK_ Install CUDA compatible cmath
+          no_output_timeout: 1m
+          command: |
+              powershell .circleci/scripts/vs_install_cmath.ps1
       - run:
           name: Build conda packages
           no_output_timeout: 20m
@@ -241,6 +246,11 @@ jobs:
     steps:
       - checkout_merge
       - designate_upload_channel
+      - run:
+          name: _HACK_ Install CUDA compatible cmath
+          no_output_timeout: 1m
+          command: |
+              powershell .circleci/scripts/vs_install_cmath.ps1
       - run:
           name: Build wheel packages
           command: |
@@ -548,6 +558,11 @@ jobs:
     steps:
       - checkout
       - designate_upload_channel
+      - run:
+          name: _HACK_ Install CUDA compatible cmath
+          no_output_timeout: 1m
+          command: |
+              powershell .circleci/scripts/vs_install_cmath.ps1
       - run:
           name: Generate cache key
           # This will refresh cache on Sundays, nightly build should generate new cache.
@@ -589,6 +604,11 @@ jobs:
     steps:
       - checkout
       - designate_upload_channel
+      - run:
+          name: _HACK_ Install CUDA compatible cmath
+          no_output_timeout: 1m
+          command: |
+              powershell .circleci/scripts/vs_install_cmath.ps1
       - run:
           name: Generate cache key
           # This will refresh cache on Sundays, nightly build should generate new cache.
@@ -718,6 +738,11 @@ jobs:
     steps:
       - checkout_merge
       - designate_upload_channel
+      - run:
+          name: _HACK_ Install CUDA compatible cmath
+          no_output_timeout: 1m
+          command: |
+              powershell .circleci/scripts/vs_install_cmath.ps1
       - run:
           command: |
             set -ex
@@ -731,6 +756,11 @@ jobs:
     steps:
       - checkout_merge
       - designate_upload_channel
+      - run:
+          name: _HACK_ Install CUDA compatible cmath
+          no_output_timeout: 1m
+          command: |
+              powershell .circleci/scripts/vs_install_cmath.ps1
       - run:
           command: |
             set -ex
diff --git a/.circleci/config.yml.in b/.circleci/config.yml.in
index f80062c53c7..d2c8d0f48be 100644
--- a/.circleci/config.yml.in
+++ b/.circleci/config.yml.in
@@ -214,6 +214,11 @@ jobs:
     steps:
       - checkout_merge
       - designate_upload_channel
+      - run:
+          name: _HACK_ Install CUDA compatible cmath
+          no_output_timeout: 1m
+          command: |
+              powershell .circleci/scripts/vs_install_cmath.ps1
       - run:
           name: Build conda packages
           no_output_timeout: 20m
@@ -241,6 +246,11 @@ jobs:
     steps:
       - checkout_merge
       - designate_upload_channel
+      - run:
+          name: _HACK_ Install CUDA compatible cmath
+          no_output_timeout: 1m
+          command: |
+              powershell .circleci/scripts/vs_install_cmath.ps1
       - run:
           name: Build wheel packages
           command: |
@@ -548,6 +558,11 @@ jobs:
     steps:
       - checkout
       - designate_upload_channel
+      - run:
+          name: _HACK_ Install CUDA compatible cmath
+          no_output_timeout: 1m
+          command: |
+              powershell .circleci/scripts/vs_install_cmath.ps1
       - run:
           name: Generate cache key
           # This will refresh cache on Sundays, nightly build should generate new cache.
@@ -589,6 +604,11 @@ jobs:
     steps:
       - checkout
       - designate_upload_channel
+      - run:
+          name: _HACK_ Install CUDA compatible cmath
+          no_output_timeout: 1m
+          command: |
+              powershell .circleci/scripts/vs_install_cmath.ps1
       - run:
           name: Generate cache key
           # This will refresh cache on Sundays, nightly build should generate new cache.
@@ -718,6 +738,11 @@ jobs:
     steps:
       - checkout_merge
       - designate_upload_channel
+      - run:
+          name: _HACK_ Install CUDA compatible cmath
+          no_output_timeout: 1m
+          command: |
+              powershell .circleci/scripts/vs_install_cmath.ps1
       - run:
           command: |
             set -ex
@@ -731,6 +756,11 @@ jobs:
     steps:
       - checkout_merge
       - designate_upload_channel
+      - run:
+          name: _HACK_ Install CUDA compatible cmath
+          no_output_timeout: 1m
+          command: |
+              powershell .circleci/scripts/vs_install_cmath.ps1
       - run:
           command: |
             set -ex
diff --git a/.circleci/scripts/vs_install_cmath.ps1 b/.circleci/scripts/vs_install_cmath.ps1
new file mode 100644
index 00000000000..c2998eba252
--- /dev/null
+++ b/.circleci/scripts/vs_install_cmath.ps1
@@ -0,0 +1,5 @@
+$CMATH_DOWNLOAD_LINK = "https://raw.githubusercontent.com/microsoft/STL/12c684bba78f9b032050526abdebf14f58ca26a3/stl/inc/cmath"
+$VC14_28_INSTALL_PATH="C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.28.29910\include"
+
+curl.exe --retry 3 -kL $CMATH_DOWNLOAD_LINK --output "$home\cmath"
+Move-Item -Path "$home\cmath" -Destination "$VC14_28_INSTALL_PATH" -Force
diff --git a/packaging/pkg_helpers.bash b/packaging/pkg_helpers.bash
index c5e55423c79..dfa0a9ae476 100644
--- a/packaging/pkg_helpers.bash
+++ b/packaging/pkg_helpers.bash
@@ -35,10 +35,7 @@ setup_cuda() {
   export WHEEL_DIR=""
   # Wheel builds need suffixes (but not if they're on OS X, which never has suffix)
   if [[ "$BUILD_TYPE" == "wheel" ]] && [[ "$(uname)" != Darwin ]]; then
-    # The default CUDA has no suffix
-    if [[ "$CU_VERSION" != "cu102" ]]; then
-      export PYTORCH_VERSION_SUFFIX="+$CU_VERSION"
-    fi
+    export PYTORCH_VERSION_SUFFIX="+$CU_VERSION"
     # Match the suffix scheme of pytorch, unless this package does not have
     # CUDA builds (in which case, use default)
     if [[ -z "$NO_CUDA_PACKAGE" ]]; then