diff --git a/op_builder/xpu/post_process_cpu_adam.sh b/op_builder/xpu/post_process_cpu_adam.sh
new file mode 100644
index 000000000000..96e1e74a67c9
--- /dev/null
+++ b/op_builder/xpu/post_process_cpu_adam.sh
@@ -0,0 +1,6 @@
+# delete including torch cuda headers
+find ./deepspeed/third-party/csrc -name "context.h" -exec sed -Ei "s:#include <ATen/cuda/CUDAContext.h>:// \0:g" {} +
+
+# fix cublas transpos flag to mkl's
+find ./deepspeed/third-party/csrc -name "context.h" -exec sed -i "s/CUBLAS_OP_T/oneapi::mkl::transpose::trans/g" {} +
+find ./deepspeed/third-party/csrc -name "context.h" -exec sed -i "s/CUBLAS_OP_N/oneapi::mkl::transpose::nontrans/g" {} +
diff --git a/op_builder/xpu/pre_process_cpu_adam.sh b/op_builder/xpu/pre_process_cpu_adam.sh
new file mode 100644
index 000000000000..eadb595230e4
--- /dev/null
+++ b/op_builder/xpu/pre_process_cpu_adam.sh
@@ -0,0 +1,27 @@
+# add at::cuda::getCurrentCUDAStream and at::cuda::getStreamFromPool
+patch ./build/csrc/includes/context.h << 'DIFF___'
+@@ -16,6 +16,23 @@
+ #include <dpct/rng_utils.hpp>
+
+ #include "gemm_test.h"
++#include <ipex.h>
++namespace at {
++  namespace cuda {
++    dpct::queue_ptr getCurrentCUDAStream() {
++      auto device_type = c10::DeviceType::XPU;
++      c10::impl::VirtualGuardImpl impl(device_type);
++      c10::Stream c10_stream = impl.getStream(c10::Device(device_type));
++      auto& queue = xpu::get_queue_from_stream(c10_stream);
++      return &queue;
++    }
++
++    dpct::queue_ptr getStreamFromPool() {
++      // not implemented
++      return nullptr;
++    }
++  }
++}
+
+ #define WARP_SIZE 32
+
+DIFF___