Update install_cuda.sh

.ci/docker/common/install_cuda.sh

@@ -129,7 +129,7 @@ function install_129 {
 }
 
 function install_128 {
-  CUDNN_VERSION=9.8.0.87
+  CUDNN_VERSION=9.10.2.21
   echo "Installing CUDA 12.8.1 and cuDNN ${CUDNN_VERSION} and NVSHMEM and NCCL and cuSparseLt-0.7.1"
   # install CUDA 12.8.1 in the same container
   install_cuda 12.8.1 cuda_12.8.1_570.124.06_linux

.ci/pytorch/smoke_test/smoke_test.py

@@ -272,6 +272,18 @@ def smoke_test_cuda(
         torch_cudnn_version = cudnn_to_version_str(torch.backends.cudnn.version())
         print(f"Torch cuDNN version: {torch_cudnn_version}")
 
+        torch_cudnn_compile_version = torch._C._cudnn.getCompileVersion()
+        print(f"Torch cuDNN compile-time version: {torch_cudnn_compile_version}")
+        torch_cudnn_runtime_version = tuple(
+            [int(x) for x in torch_cudnn_version.split(".")]
+        )
+        if torch_cudnn_runtime_version != torch_cudnn_compile_version:
+            raise RuntimeError(
+                "cuDNN runtime version doesn't match comple version. "
+                f"Loaded: {torch_cudnn_runtime_version} "
+                f"Expected: {torch_cudnn_compile_version}"
+            )
+
         if sys.platform in ["linux", "linux2"]:
             torch_nccl_version = ".".join(str(v) for v in torch.cuda.nccl.version())
             print(f"Torch nccl; version: {torch_nccl_version}")