diff --git a/.github/scripts/validate_binaries.sh b/.github/scripts/validate_binaries.sh
index 85ad0de47..62fb21d99 100755
--- a/.github/scripts/validate_binaries.sh
+++ b/.github/scripts/validate_binaries.sh
@@ -49,6 +49,24 @@ elif [[ ${MATRIX_CHANNEL} = 'release' ]]; then
     export PYTORCH_URL="https://download.pytorch.org/whl/${CUDA_VERSION}"
 fi
 
+
+echo "CU_VERSION: ${CUDA_VERSION}"
+echo "MATRIX_CHANNEL: ${MATRIX_CHANNEL}"
+echo "CONDA_ENV: ${CONDA_ENV}"
+
+if [[ $CUDA_VERSION = cu* ]]; then
+    # Setting LD_LIBRARY_PATH fixes the runtime error with fbgemm_gpu not
+    # being able to locate libnvrtc.so
+    echo "[NOVA] Setting LD_LIBRARY_PATH ..."
+    conda env config vars set -p ${CONDA_ENV}  \
+        LD_LIBRARY_PATH="/usr/local/lib:${CUDA_HOME}/lib64:${CONDA_ENV}/lib:${LD_LIBRARY_PATH}"
+else
+    echo "[NOVA] Setting LD_LIBRARY_PATH ..."
+    conda env config vars set -p ${CONDA_ENV}  \
+        LD_LIBRARY_PATH="/usr/local/lib:${CONDA_ENV}/lib:${LD_LIBRARY_PATH}"
+fi
+
+
 # install pytorch
 # switch back to conda once torch nightly is fixed
 # if [[ ${MATRIX_GPU_ARCH_TYPE} = 'cuda' ]]; then
diff --git a/.github/workflows/validate-binaries.yml b/.github/workflows/validate-binaries.yml
index 248857214..98d69d721 100644
--- a/.github/workflows/validate-binaries.yml
+++ b/.github/workflows/validate-binaries.yml
@@ -1,6 +1,12 @@
 name: Validate binaries
 
 on:
+  pull_request:
+    paths-ignore:
+      - "docs/*"
+      - "third_party/*"
+      - .gitignore
+      - "*.md"
   workflow_call:
     inputs:
       channel: