diff --git a/.github/workflows/ci-linux.yml b/.github/workflows/ci-linux.yml
index e7dcb810e..17a122008 100644
--- a/.github/workflows/ci-linux.yml
+++ b/.github/workflows/ci-linux.yml
@@ -49,7 +49,10 @@ jobs:
           git remote add origin $REPO_ADDRESS
           git -c protocol.version=2 fetch --depth 1 origin $BRANCH_NAME
           git reset --hard FETCH_HEAD
-          git -c submodule."third_party/torch-mlir".update=none -c submodule."third_party/stablehlo".update=none -c submodule."src/runtime_src/core/common/aiebu".update=none submodule update --init --recursive --depth 1 --single-branch -j 10
+          git -c submodule."third_party/torch-mlir".update=none \
+              -c submodule."third_party/stablehlo".update=none \
+              -c submodule."src/runtime_src/core/common/aiebu".update=none \
+              submodule update --init --recursive --depth 1 --single-branch -j 10
           
       - name: Install deps
         run: |
@@ -61,6 +64,11 @@ jobs:
         run: |
           pip install "numpy<2" pyyaml "pybind11[global]==2.10.3" nanobind pytest
 
+      - name: Peano dep
+        run: |
+          bash build_tools/download_peano.sh
+          echo "PEANO_INSTALL_DIR=$PWD/llvm-aie" >> $GITHUB_ENV
+
       - name: Run Pytest
         run: |
           pytest build_tools/ci
diff --git a/.github/workflows/ci-windows.yml b/.github/workflows/ci-windows.yml
index 2fcc702e7..c1b696d0a 100644
--- a/.github/workflows/ci-windows.yml
+++ b/.github/workflows/ci-windows.yml
@@ -60,7 +60,10 @@ jobs:
           git remote add origin $REPO_ADDRESS
           git -c protocol.version=2 fetch --depth 1 origin $BRANCH_NAME
           git reset --hard FETCH_HEAD
-          git -c submodule."third_party/torch-mlir".update=none -c submodule."third_party/stablehlo".update=none -c submodule."src/runtime_src/core/common/aiebu".update=none submodule update --init --recursive --depth 1 --single-branch -j 10
+          git -c submodule."third_party/torch-mlir".update=none \
+              -c submodule."third_party/stablehlo".update=none \
+              -c submodule."src/runtime_src/core/common/aiebu".update=none \
+              submodule update --init --recursive --depth 1 --single-branch -j 10
 
       - name: Setup Cpp
         uses: aminya/setup-cpp@v1
@@ -86,6 +89,11 @@ jobs:
           key: ${{ env.CACHE_KEY }}
           restore-keys: windows-build-test-cpp-
 
+      - name: Peano dep
+        run: |
+          .\build_tools\download_peano.ps1
+          Add-Content -Path $env:GITHUB_ENV -Value "PEANO_INSTALL_DIR=$PWD\llvm-aie"
+
       - name: Build packages
         run: |
           $env:cache_dir = "${{ env.CACHE_DIR }}"
diff --git a/build_tools/build_test_cpp.ps1 b/build_tools/build_test_cpp.ps1
index f7c3b7db2..4670dd9d5 100644
--- a/build_tools/build_test_cpp.ps1
+++ b/build_tools/build_test_cpp.ps1
@@ -80,10 +80,18 @@ $CMAKE_ARGS = @(
     "-DIREE_TARGET_BACKEND_DEFAULTS=OFF"
     "-DIREE_TARGET_BACKEND_LLVM_CPU=ON"
     "-DIREE_CMAKE_PLUGIN_PATHS=$repo_root"
-    "-DIREE_EXTERNAL_HAL_DRIVERS=xrt"
+    "-DIREE_EXTERNAL_HAL_DRIVERS=xrt;xrt-lite"
     "-DIREE_BUILD_PYTHON_BINDINGS=ON"
+    # iree/runtime/src/iree/hal/cts/cts_test_base.h:173:24: error: unused variable 'device_buffer' [-Werror,-Wunused-variable]
+    "-DIREE_ENABLE_WERROR_FLAG=OFF"
 )
 
+$peano_install_dir = "$env:PEANO_INSTALL_DIR"
+if ($peano_install_dir -and (Test-Path "$peano_install_dir"))
+{
+    $CMAKE_ARGS += @("-DPEANO_INSTALL_DIR=$peano_install_dir")
+}
+
 if ($llvm_install_dir -and (Test-Path "$llvm_install_dir"))
 {
     echo "using existing llvm install @ $llvm_install_dir"
@@ -121,7 +129,7 @@ echo "-----"
 # better have git-bash installed...
 $env:Path = "C:\Program Files\Git\bin;$env:Path"
 pushd $build_dir
-& bash -l -c "ctest -R amd-aie --output-on-failure -j --repeat until-pass:5"
+& bash -l -c "ctest -R amd-aie -E driver/xrt-lite --output-on-failure -j --repeat until-pass:5"
 popd
 
 if ($llvm_install_dir -and (Test-Path "$llvm_install_dir"))
diff --git a/build_tools/build_test_cpp.sh b/build_tools/build_test_cpp.sh
index d04bf8bb4..857c81f44 100644
--- a/build_tools/build_test_cpp.sh
+++ b/build_tools/build_test_cpp.sh
@@ -31,7 +31,7 @@ mkdir -p "${cache_dir}/pip"
 python="$(which python)"
 echo "Using python: $python"
 
-if [[ "$OSTYPE" == "linux-gnu"* ]]; then
+if [[ "$OSTYPE" == "linux"* ]]; then
   export CMAKE_TOOLCHAIN_FILE="$this_dir/linux_default_toolchain.cmake"
   export CC=clang
   export CXX=clang++
@@ -61,54 +61,73 @@ echo '{
 }' > $iree_dir/CMakeUserPresets.json 
 
 cd $iree_dir
-CMAKE_ARGS="\
-  -GNinja \
-  -DCMAKE_BUILD_TYPE=Release \
-  -DCMAKE_INSTALL_PREFIX=$install_dir \
-  -DCMAKE_INSTALL_LIBDIR=lib \
-  -DIREE_ERROR_ON_MISSING_SUBMODULES=OFF \
-  -DIREE_ENABLE_ASSERTIONS=ON \
-  -DIREE_BUILD_SAMPLES=OFF \
-  -DIREE_BUILD_PYTHON_BINDINGS=ON \
-  -DIREE_BUILD_BINDINGS_TFLITE=OFF \
-  -DIREE_HAL_DRIVER_DEFAULTS=OFF \
-  -DIREE_HAL_DRIVER_LOCAL_SYNC=ON \
-  -DIREE_HAL_DRIVER_LOCAL_TASK=ON \
-  -DIREE_TARGET_BACKEND_DEFAULTS=OFF \
-  -DIREE_TARGET_BACKEND_LLVM_CPU=ON \
-  -DIREE_INPUT_TOSA=OFF \
-  -DIREE_INPUT_STABLEHLO=OFF \
-  -DIREE_INPUT_TORCH=OFF \
-  -DCMAKE_OBJECT_PATH_MAX=4096 \
-  -DIREE_CMAKE_PLUGIN_PATHS=$repo_root"
-
-if [ -d "${llvm_install_dir}" ]; then
-  CMAKE_ARGS="$CMAKE_ARGS \
-    -DIREE_BUILD_BUNDLED_LLVM=OFF \
-    -DClang_DIR=$llvm_install_dir/lib/cmake/clang \
-    -DLLD_DIR=$llvm_install_dir/lib/cmake/lld \
-    -DMLIR_DIR=$llvm_install_dir/lib/cmake/mlir \
-    -DLLVM_DIR=$llvm_install_dir/lib/cmake/llvm"
+CMAKE_ARGS=(
+  -GNinja
+  -DCMAKE_BUILD_TYPE=Release
+  -DCMAKE_INSTALL_PREFIX="$install_dir"
+  -DCMAKE_INSTALL_LIBDIR=lib
+  -DIREE_ERROR_ON_MISSING_SUBMODULES=OFF
+  -DIREE_ENABLE_ASSERTIONS=ON
+  -DIREE_BUILD_SAMPLES=OFF
+  -DIREE_BUILD_PYTHON_BINDINGS=ON
+  -DIREE_BUILD_BINDINGS_TFLITE=OFF
+  -DIREE_HAL_DRIVER_DEFAULTS=OFF
+  -DIREE_HAL_DRIVER_LOCAL_SYNC=ON
+  -DIREE_HAL_DRIVER_LOCAL_TASK=ON
+  -DIREE_TARGET_BACKEND_DEFAULTS=OFF
+  -DIREE_TARGET_BACKEND_LLVM_CPU=ON
+  -DIREE_INPUT_TOSA=OFF
+  -DIREE_INPUT_STABLEHLO=OFF
+  -DIREE_INPUT_TORCH=OFF
+  -DCMAKE_OBJECT_PATH_MAX=4096
+  -DIREE_CMAKE_PLUGIN_PATHS="$repo_root"
+  # iree/runtime/src/iree/hal/cts/cts_test_base.h:173:24: error: unused variable 'device_buffer' [-Werror,-Wunused-variable]
+  -DIREE_ENABLE_WERROR_FLAG=OFF
+)
+
+PEANO_INSTALL_DIR=${PEANO_INSTALL_DIR:-""}
+if [ "$PEANO_INSTALL_DIR" != "" ] && [ -d "$PEANO_INSTALL_DIR" ]; then
+  CMAKE_ARGS+=(-DPEANO_INSTALL_DIR="$PEANO_INSTALL_DIR")
 fi
 
-if [[ "$OSTYPE" == "linux-gnu"* ]]; then
-  cmake $CMAKE_ARGS \
-    -DCMAKE_EXE_LINKER_FLAGS_INIT="-fuse-ld=lld" \
-    -DCMAKE_SHARED_LINKER_FLAGS_INIT="-fuse-ld=lld" \
-    -DCMAKE_MODULE_LINKER_FLAGS_INIT="-fuse-ld=lld" \
-    -DCMAKE_C_COMPILER="${CC}" \
-    -DCMAKE_CXX_COMPILER="${CXX}" \
-    -DLLVM_TARGET_ARCH=X86 \
-    -DLLVM_TARGETS_TO_BUILD=X86 \
-    -DIREE_EXTERNAL_HAL_DRIVERS=xrt \
-    -S $iree_dir -B $build_dir
+if [ -d "$llvm_install_dir" ]; then
+  CMAKE_ARGS+=(
+    -DIREE_BUILD_BUNDLED_LLVM=OFF
+    -DClang_DIR="$llvm_install_dir/lib/cmake/clang"
+    -DLLD_DIR="$llvm_install_dir/lib/cmake/lld"
+    -DMLIR_DIR="$llvm_install_dir/lib/cmake/mlir"
+    -DLLVM_DIR="$llvm_install_dir/lib/cmake/llvm"
+  )
+fi
+
+if [[ "$OSTYPE" == "linux"* ]]; then
+  CMAKE_ARGS+=(
+    -DCMAKE_EXE_LINKER_FLAGS_INIT="-fuse-ld=lld"
+    -DCMAKE_SHARED_LINKER_FLAGS_INIT="-fuse-ld=lld"
+    -DCMAKE_MODULE_LINKER_FLAGS_INIT="-fuse-ld=lld"
+    -DCMAKE_C_COMPILER="${CC}"
+    -DCMAKE_CXX_COMPILER="${CXX}"
+    -DLLVM_TARGET_ARCH=X86
+    -DLLVM_TARGETS_TO_BUILD=X86
+    -DIREE_EXTERNAL_HAL_DRIVERS="xrt;xrt-lite"
+    -S
+    "$iree_dir"
+    -B
+    "$build_dir"
+  )
 elif [[ "$OSTYPE" == "darwin"* ]]; then
-  cmake $CMAKE_ARGS \
-    -DLLVM_TARGET_ARCH="X86;ARM" \
-    -DLLVM_TARGETS_TO_BUILD="X86;ARM" \
-    -S $iree_dir -B $build_dir
+  CMAKE_ARGS+=(
+    -DLLVM_TARGET_ARCH="X86;ARM"
+    -DLLVM_TARGETS_TO_BUILD="X86;ARM"
+    -S
+    "$iree_dir"
+    -B
+    "$build_dir"
+  )
 fi
 
+cmake "${CMAKE_ARGS[@]}"
+
 echo "Building all"
 echo "------------"
 cmake --build "$build_dir" -- -k 0
@@ -123,8 +142,8 @@ cmake --build "$build_dir" --target iree-install-dist
 
 echo "CTest"
 echo "-----"
-if [[ "$OSTYPE" == "linux-gnu"* ]]; then
-  ctest --test-dir "$build_dir" -R amd-aie --output-on-failure -j
+if [[ "$OSTYPE" == "linux"* ]]; then
+  ctest --test-dir "$build_dir" -R amd-aie -E "driver/xrt-lite" --output-on-failure -j
 elif [[ "$OSTYPE" == "darwin"* ]]; then
   ctest --test-dir "$build_dir" -R amd-aie -E "matmul_pack_peel_air_e2e|matmul_elementwise_pack_peel_air_e2e|conv_fill_spec_pad" --output-on-failure -j --repeat until-pass:5
 fi
diff --git a/build_tools/download_peano.ps1 b/build_tools/download_peano.ps1
index 89bd6808f..6589cc562 100644
--- a/build_tools/download_peano.ps1
+++ b/build_tools/download_peano.ps1
@@ -9,4 +9,7 @@ $ErrorActionPreference = 'Stop'
 $this_dir = Split-Path -Path $MyInvocation.MyCommand.Path -Parent
 $RELEASE = (Get-Content -Path "$this_dir/peano_commit.txt")
 pip download llvm_aie==$RELEASE -f https://github.com/Xilinx/llvm-aie/releases/expanded_assets/nightly
-Expand-Archive (Get-ChildItem -Filter llvm*.whl).FullName -DestinationPath $PWD.Path
+$peano = (Get-ChildItem -Filter llvm*.whl)
+$new_name = ($peano.Basename + ".zip")
+Rename-Item -Path $peano.Name -NewName $new_name
+Expand-Archive $new_name -DestinationPath $PWD.Path
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/amdxdna_accel.h b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/amdxdna_accel.h
index e7f52afc3..d4d18d6b3 100644
--- a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/amdxdna_accel.h
+++ b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/amdxdna_accel.h
@@ -132,11 +132,11 @@ struct amdxdna_cu_config {
  * @pad: MBZ
  * @cu_configs: Array of CU configurations of struct amdxdna_cu_config
  */
-struct amdxdna_hwctx_param_config_cu {
-  __u16 num_cus;
-  __u16 pad[3];
-  struct amdxdna_cu_config cu_configs[] __counted_by(num_cus);
-};
+// struct amdxdna_hwctx_param_config_cu {
+//   __u16 num_cus;
+//   __u16 pad[3];
+//   struct amdxdna_cu_config cu_configs[] __counted_by(num_cus);
+// };
 
 enum amdxdna_drm_config_hwctx_param {
   DRM_AMDXDNA_HWCTX_CONFIG_CU,
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/bo.cpp b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/bo.cpp
index 3076a386a..739456ccf 100644
--- a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/bo.cpp
+++ b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/bo.cpp
@@ -9,6 +9,7 @@
 #include <unistd.h>
 #include <x86intrin.h>
 
+#include "pcidev.h"
 #include "shim_debug.h"
 
 namespace {
@@ -16,7 +17,7 @@ namespace {
 uint32_t alloc_drm_bo(const shim_xdna::pdev& dev, amdxdna_bo_type type,
                       void* buf, size_t size) {
   amdxdna_drm_create_bo cbo = {
-      .type = type,
+      .type = static_cast<uint32_t>(type),
       .vaddr = reinterpret_cast<uintptr_t>(buf),
       .size = size,
   };
@@ -38,7 +39,7 @@ void get_drm_bo_info(const shim_xdna::pdev& dev, uint32_t boh,
 void* map_parent_range(size_t size) {
   auto p = ::mmap(0, size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS,
                   -1, 0);
-  if (!p) shim_err(errno, "mmap(len=%ld) failed", size);
+  if (!p) shim_xdna::shim_err(errno, "mmap(len=%ld) failed", size);
 
   return p;
 }
@@ -101,7 +102,7 @@ bool is_power_of_two(size_t x) { return (x > 0) && ((x & (x - 1)) == 0); }
 
 void* addr_align(void* p, size_t align) {
   if (!is_power_of_two(align))
-    shim_err(EINVAL, "Alignment 0x%lx is not power of two", align);
+    shim_xdna::shim_err(EINVAL, "Alignment 0x%lx is not power of two", align);
 
   return (void*)(((uintptr_t)p + align) & ~(align - 1));
 }
@@ -129,7 +130,8 @@ inline void clflush_data(const void* base, size_t offset, size_t len) {
 
   if (!cacheline_size) {
     long sz = sysconf(_SC_LEVEL1_DCACHE_LINESIZE);
-    if (sz <= 0) shim_err(EINVAL, "Invalid cache line size: %ld", sz);
+    if (sz <= 0)
+      shim_xdna::shim_err(EINVAL, "Invalid cache line size: %ld", sz);
     cacheline_size = sz;
   }
 
@@ -313,8 +315,7 @@ std::unique_ptr<shared_handle> bo::share() const {
 
 amdxdna_bo_type bo::get_type() const { return m_type; }
 
-bo::bo(const device& device, hw_ctx::slot_id ctx_id, size_t size,
-       uint64_t flags)
+bo::bo(const device& device, uint32_t ctx_id, size_t size, uint64_t flags)
     : bo(device, ctx_id, size, flags, flag_to_type(flags)) {
   if (m_type == AMDXDNA_BO_INVALID)
     shim_err(EINVAL, "Invalid BO flags: 0x%lx", flags);
@@ -323,8 +324,8 @@ bo::bo(const device& device, hw_ctx::slot_id ctx_id, size_t size,
 bo::bo(const device& device, size_t size, amdxdna_bo_type type)
     : bo(device, AMDXDNA_INVALID_CTX_HANDLE, size, 0, type) {}
 
-bo::bo(const device& device, hw_ctx::slot_id ctx_id, size_t size,
-       uint64_t flags, amdxdna_bo_type type)
+bo::bo(const device& device, uint32_t ctx_id, size_t size, uint64_t flags,
+       amdxdna_bo_type type)
     : m_pdev(device.get_pdev()),
       m_aligned_size(size),
       m_flags(flags),
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/bo.h b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/bo.h
index 2f513ae8d..ea163db45 100644
--- a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/bo.h
+++ b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/bo.h
@@ -5,12 +5,11 @@
 #define _BO_XDNA_H_
 
 #include <atomic>
+#include <map>
+#include <mutex>
 #include <string>
 
 #include "amdxdna_accel.h"
-#include "device.h"
-#include "hwctx.h"
-#include "pcidev.h"
 #include "shared.h"
 #include "shim_debug.h"
 
@@ -62,6 +61,9 @@ struct xcl_bo_flags {
   };
 };
 
+struct device;
+struct pdev;
+
 struct bo {
   // map_type - determines how a buffer is mapped
   enum class map_type { read, write };
@@ -87,10 +89,11 @@ struct bo {
     uint64_t kmhdl;  // kernel mode handle
   };
 
-  bo(const device& device, hw_ctx::slot_id ctx_id, size_t size, uint64_t flags,
+  using uint32_t = uint32_t;
+  bo(const device& device, uint32_t ctx_id, size_t size, uint64_t flags,
      amdxdna_bo_type type);
 
-  bo(const device& device, hw_ctx::slot_id ctx_id, size_t size, uint64_t flags);
+  bo(const device& device, uint32_t ctx_id, size_t size, uint64_t flags);
 
   bo(const device& device, shared_handle::export_handle ehdl);
 
@@ -171,7 +174,7 @@ struct bo {
 
   // Used when exclusively assigned to a HW context. By default, BO is shared
   // among all HW contexts.
-  hw_ctx::slot_id m_owner_ctx_id = AMDXDNA_INVALID_CTX_HANDLE;
+  uint32_t m_owner_ctx_id = AMDXDNA_INVALID_CTX_HANDLE;
 
   void bind_at(size_t pos, const bo* bh, size_t offset, size_t size);
 
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/device.cpp b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/device.cpp
index 20ec767ea..69676385c 100644
--- a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/device.cpp
+++ b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/device.cpp
@@ -3,18 +3,15 @@
 
 #include "device.h"
 
-#include <sys/syscall.h>
-#include <unistd.h>
-
-#include <any>
 #include <filesystem>
 
 #include "bo.h"
 #include "hwctx.h"
+#include "pcidev.h"
 
 namespace shim_xdna {
 
-device::device(const pdev& pdev, handle_type shim_handle)
+device::device(const pdev& pdev, void* shim_handle)
     : m_pdev(pdev), m_handle(shim_handle) {
   shim_debug("Created KMQ device (%s) ...", get_pdev().m_sysfs_name.c_str());
 }
@@ -24,16 +21,16 @@ device::~device() {
   m_pdev.close();
 }
 
-std::unique_ptr<hw_ctx> device::create_hw_context(
-    const device& dev, const hw_ctx::qos_type& qos) const {
-  return std::make_unique<hw_ctx>(dev, qos);
-}
+// std::unique_ptr<hw_ctx> device::create_hw_context(
+//     const device& dev, const hw_ctx::qos_type& qos) const {
+//   return std::make_unique<hw_ctx>(dev, qos);
+// }
 
-std::unique_ptr<bo> device::alloc_bo(void* userptr, hw_ctx::slot_id ctx_id,
+std::unique_ptr<bo> device::alloc_bo(void* userptr, uint32_t ctx_id,
                                      size_t size, uint64_t flags) {
   if (userptr) shim_not_supported_err("User ptr BO");
 
-  auto b = bo(this->m_pdev, ctx_id, size, flags);
+  auto b = bo(*this, ctx_id, size, flags);
   return std::make_unique<bo>(*this, ctx_id, size, flags);
 }
 
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/device.h b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/device.h
index 8db0f2227..ca4f295f8 100644
--- a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/device.h
+++ b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/device.h
@@ -5,13 +5,13 @@
 #define PCIE_DEVICE_LINUX_XDNA_H
 
 #include <map>
+#include <mutex>
 
-#include "pcidev.h"
+#include "shared.h"
 #include "shim_debug.h"
 
 namespace shim_xdna {
 
-typedef void* xclDeviceHandle;
 #define XRT_NULL_HANDLE NULL
 
 // cuidx_type - encode cuidx and domain
@@ -36,21 +36,22 @@ struct cuidx_type {
   using domain_index_type = uint16_t;
 };
 
-struct device {
-  // device index type
-  using id_type = unsigned int;
-  using slot_id = uint32_t;
-  using handle_type = xclDeviceHandle;
+struct hw_ctx;
+struct pdev;
+struct bo;
 
-  device(const pdev& pdev, handle_type shim_handle);
+struct device {
+  device(const pdev& pdev, void* shim_handle);
 
   ~device();
 
-  std::unique_ptr<bo> alloc_bo(void* userptr, hw_ctx::slot_id ctx_id,
-                               size_t size, uint64_t flags);
+  using qos_type = std::map<std::string, uint32_t>;
+  enum class access_mode : uint8_t { exclusive = 0, shared = 1 };
+  std::unique_ptr<bo> alloc_bo(void* userptr, uint32_t ctx_id, size_t size,
+                               uint64_t flags);
 
-  std::unique_ptr<hw_ctx> create_hw_context(const device& dev,
-                                            const hw_ctx::qos_type& qos) const;
+  // std::unique_ptr<hw_ctx> create_hw_context(const device& dev,
+  //                                           const qos_type& qos) const;
 
   std::unique_ptr<bo> import_bo(shared_handle::export_handle ehdl) const;
 
@@ -60,8 +61,8 @@ struct device {
 
   std::unique_ptr<bo> import_bo(pid_t, shared_handle::export_handle);
 
-  std::unique_ptr<hw_ctx> create_hw_context(const hw_ctx::qos_type& qos,
-                                            hw_ctx::access_mode mode) const;
+  std::unique_ptr<hw_ctx> create_hw_context(const qos_type& qos,
+                                            access_mode mode) const;
 
   std::vector<char> read_aie_mem(uint16_t col, uint16_t row, uint32_t offset,
                                  uint32_t size);
@@ -76,7 +77,7 @@ struct device {
 
   const pdev& m_pdev;  // The pcidev that this device object is derived from
   std::map<uint32_t, bo*> m_bo_map;
-  xclDeviceHandle m_handle = XRT_NULL_HANDLE;
+  void* m_handle = XRT_NULL_HANDLE;
 
   mutable std::mutex m_mutex;
 };
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/ert.h b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/ert.h
index ac5858db4..bce5d1623 100644
--- a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/ert.h
+++ b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/ert.h
@@ -447,7 +447,7 @@ struct config_sk_image_uuid {
   uint32_t num_cus;
   uint32_t sk_name[5];
   unsigned char     sk_uuid[16];
-  uint32_t slot_id;
+  uint32_t uint32_t;
 };
 
 /**
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/fence.cpp b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/fence.cpp
index 4af239eb1..850e4198a 100644
--- a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/fence.cpp
+++ b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/fence.cpp
@@ -6,6 +6,9 @@
 #include <limits>
 
 #include "amdxdna_accel.h"
+#include "fence.h"
+#include "hwctx.h"
+#include "pcidev.h"
 
 namespace {
 
@@ -138,15 +141,6 @@ fence::fence(const device& device, shared_handle::export_handle ehdl)
   shim_debug("Fence imported: %d@%ld", m_syncobj_hdl, m_state);
 }
 
-fence::fence(const fence& f)
-    : m_pdev(f.m_pdev),
-      m_import(f.share()),
-      m_syncobj_hdl(import_syncobj(m_pdev, m_import->get_export_handle())),
-      m_state{f.m_state},
-      m_signaled{f.m_signaled} {
-  shim_debug("Fence cloned: %d@%ld", m_syncobj_hdl, m_state);
-}
-
 fence::~fence() {
   shim_debug("Fence going away: %d@%ld", m_syncobj_hdl, m_state);
   destroy_syncobj(m_pdev, m_syncobj_hdl);
@@ -161,10 +155,6 @@ std::unique_ptr<shared_handle> fence::share() const {
 
 uint64_t fence::get_next_state() const { return m_state + 1; }
 
-std::unique_ptr<fence> fence::clone() const {
-  return std::make_unique<fence>(*this);
-}
-
 uint64_t fence::wait_next_state() const {
   std::lock_guard<std::mutex> guard(m_lock);
 
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/fence.h b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/fence.h
index 2b9067c1c..a5acd4c1d 100644
--- a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/fence.h
+++ b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/fence.h
@@ -6,13 +6,15 @@
 
 #include <mutex>
 
-#include "device.h"
-#include "hwctx.h"
 #include "shared.h"
 #include "shim_debug.h"
 
 namespace shim_xdna {
 
+struct device;
+struct hw_ctx;
+struct pdev;
+
 struct fence {
   using export_handle = shared_handle::export_handle;
   enum class access_mode : uint8_t { local, shared, process, hybrid };
@@ -23,8 +25,6 @@ struct fence {
 
   ~fence();
 
-  std::unique_ptr<fence> clone() const;
-
   std::unique_ptr<shared_handle> share() const;
 
   void wait(uint32_t timeout_ms) const;
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwctx.cpp b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwctx.cpp
index a10abf3ae..23dd3b728 100644
--- a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwctx.cpp
+++ b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwctx.cpp
@@ -4,30 +4,16 @@
 #include "hwctx.h"
 
 #include "bo.h"
+#include "device.h"
 #include "hwq.h"
+#include "pcidev.h"
 
-namespace {
-
-std::vector<uint8_t> get_pdi(const xrt_core::xclbin::aie_partition_obj& aie,
-                             uint16_t kernel_id) {
-  for (auto& pdi : aie.pdis) {
-    for (auto& cdo : pdi.cdo_groups) {
-      for (auto kid : cdo.kernel_ids) {
-        if (kid == kernel_id) return pdi.pdi;
-      }
-    }
-  }
-  shim_err(ENOENT, "PDI for kernel ID 0x%x not found", kernel_id);
-}
-
-}  // namespace
 namespace shim_xdna {
 
 hw_ctx::hw_ctx(const device& dev, const qos_type& qos, std::unique_ptr<hw_q> q)
     : m_device(dev), m_q(std::move(q)), m_doorbell(0), m_log_buf(nullptr) {
   shim_debug("Creating HW context...");
   init_qos_info(qos);
-  parse_xclbin(xclbin);
 }
 
 hw_ctx::~hw_ctx() {
@@ -35,9 +21,9 @@ hw_ctx::~hw_ctx() {
   shim_debug("Destroyed HW context (%d)...", m_handle);
 }
 
-hw_ctx::slot_id hw_ctx::get_slotidx() const { return m_handle; }
+uint32_t hw_ctx::get_slotidx() const { return m_handle; }
 
-void hw_ctx::set_slotidx(slot_id id) { m_handle = id; }
+void hw_ctx::set_slotidx(uint32_t id) { m_handle = id; }
 
 cuidx_type hw_ctx::open_cu_context(const std::string& cu_name) {
   for (uint32_t i = 0; i < m_cu_info.size(); i++) {
@@ -82,20 +68,6 @@ void hw_ctx::init_qos_info(const qos_type& qos) {
   }
 }
 
-void hw_ctx::print_xclbin_info() {
-  if (m_cu_info.empty()) {
-    shim_debug("CU INFO is empty");
-    return;
-  }
-
-  for (int idx = 0; idx < m_cu_info.size(); idx++) {
-    auto& e = m_cu_info[idx];
-    shim_debug("index=%d, name=%s, func=%d, pdi(p=%p, sz=%ld)", idx,
-               e.m_name.c_str(), e.m_func, e.m_pdi.data(), e.m_pdi.size());
-  }
-  shim_debug("OPs/cycle: %d", m_ops_per_cycle);
-}
-
 const device& hw_ctx::get_device() { return m_device; }
 
 const std::vector<hw_ctx::cu_info>& hw_ctx::get_cu_info() const {
@@ -107,10 +79,10 @@ void hw_ctx::create_ctx_on_device() {
   arg.qos_p = reinterpret_cast<uintptr_t>(&m_qos);
   arg.umq_bo = m_q->get_queue_bo();
   arg.max_opc = m_ops_per_cycle;
-  arg.num_tiles =
-      m_num_cols *
-      xrt_core::device_query<xrt_core::query::aie_tiles_stats>(&m_device)
-          .core_rows;
+  // arg.num_tiles =
+  //     m_num_cols *
+  //     xrt_core::device_query<xrt_core::query::aie_tiles_stats>(&m_device)
+  //         .core_rows;
   arg.log_buf_bo = m_log_bo
                        ? static_cast<bo*>(m_log_bo.get())->get_drm_bo_handle()
                        : AMDXDNA_INVALID_BO_HANDLE;
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwctx.h b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwctx.h
index 446d64f58..9fec8cc6f 100644
--- a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwctx.h
+++ b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwctx.h
@@ -7,18 +7,19 @@
 #include <map>
 
 #include "amdxdna_accel.h"
-#include "device.h"
+#include "shared.h"
 #include "shim_debug.h"
 
 namespace shim_xdna {
 
 struct hw_q;  // forward declaration
+struct device;
+struct bo;
+struct cuidx_type;
 
 struct hw_ctx {
   using qos_type = std::map<std::string, uint32_t>;
   enum class access_mode : uint8_t { exclusive = 0, shared = 1 };
-  using access_mode = access_mode;
-  using slot_id = uint32_t;
 
   hw_ctx(const device& dev, const qos_type& qos, std::unique_ptr<hw_q> q);
 
@@ -29,7 +30,7 @@ struct hw_ctx {
 
   void update_access_mode(access_mode) { shim_not_supported_err(__func__); }
 
-  slot_id get_slotidx() const;
+  uint32_t get_slotidx() const;
 
   hw_q* get_hw_queue();
 
@@ -57,7 +58,7 @@ struct hw_ctx {
 
   const std::vector<cu_info>& get_cu_info() const;
 
-  void set_slotidx(slot_id id);
+  void set_slotidx(uint32_t id);
 
   void set_doorbell(uint32_t db);
 
@@ -68,7 +69,7 @@ struct hw_ctx {
   void fini_log_buf();
 
   const device& m_device;
-  slot_id m_handle = AMDXDNA_INVALID_CTX_HANDLE;
+  uint32_t m_handle = AMDXDNA_INVALID_CTX_HANDLE;
   amdxdna_qos_info m_qos = {};
   std::vector<cu_info> m_cu_info;
   std::unique_ptr<hw_q> m_q;
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwq.cpp b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwq.cpp
index d52c2a8ac..25bc89ad0 100644
--- a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwq.cpp
+++ b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwq.cpp
@@ -6,6 +6,8 @@
 #include "bo.h"
 #include "ert.h"
 #include "fence.h"
+#include "hwctx.h"
+#include "pcidev.h"
 #include "shim_debug.h"
 
 namespace {
@@ -22,7 +24,7 @@ int wait_cmd(const shim_xdna::pdev &pdev, const shim_xdna::hw_ctx *ctx,
   auto boh = static_cast<shim_xdna::bo *>(cmd);
   auto id = boh->get_cmd_id();
 
-  shim_debug("Waiting for cmd (%ld)...", id);
+  shim_xdna::shim_debug("Waiting for cmd (%ld)...", id);
 
   amdxdna_drm_wait_cmd wcmd = {
       .hwctx = ctx->get_slotidx(),
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwq.h b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwq.h
index 30b3cfbf9..98442c49c 100644
--- a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwq.h
+++ b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/hwq.h
@@ -3,12 +3,15 @@
 
 #ifndef _HWQ_XDNA_H_
 #define _HWQ_XDNA_H_
-
-#include "fence.h"
-#include "hwctx.h"
-#include "shim_debug.h"
+#include <cstdint>
+#include <vector>
 
 namespace shim_xdna {
+struct device;
+struct bo;
+struct hw_ctx;
+struct pdev;
+struct fence;
 
 struct hw_q {
   hw_q(const device &device);
@@ -25,7 +28,7 @@ struct hw_q {
 
   void submit_signal(const fence *);
 
-  virtual void bind_hwctx(const hw_ctx *ctx) = 0;
+  void bind_hwctx(const hw_ctx *ctx);
 
   void unbind_hwctx();
 
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/pcidev.cpp b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/pcidev.cpp
index e057c61ac..f078d3e00 100644
--- a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/pcidev.cpp
+++ b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/pcidev.cpp
@@ -427,9 +427,8 @@ void* pdev::mmap(void* addr, size_t len, int prot, int flags,
 
 void pdev::munmap(void* addr, size_t len) const { ::munmap(addr, len); }
 
-std::shared_ptr<device> pdev::create_device(device::handle_type handle,
-                                            device::id_type id) const {
-  auto dev = std::make_shared<device>(*this, handle, id);
+std::shared_ptr<device> pdev::create_device(void* handle) const {
+  auto dev = std::make_shared<device>(*this, handle);
   // Alloc device memory on first device creation.
   // No locking is needed since driver will ensure only one heap BO is
   // created.
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/pcidev.h b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/pcidev.h
index a2cb858fe..a84fa646c 100644
--- a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/pcidev.h
+++ b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/pcidev.h
@@ -39,8 +39,7 @@ struct pdev {
 
   std::string get_subdev_path(const std::string& subdev, uint32_t idx) const;
 
-  std::shared_ptr<device> create_device(device::handle_type handle,
-                                        device::id_type id) const;
+  std::shared_ptr<device> create_device(void* handle) const;
 
   void ioctl(unsigned long cmd, void* arg) const;
 
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/shim_debug.cpp b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/shim_debug.cpp
index f57848458..d761a3995 100644
--- a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/shim_debug.cpp
+++ b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/shim_debug.cpp
@@ -8,7 +8,7 @@
 
 static std::recursive_mutex s_debug_mutex;
 
-namespace {
+namespace shim_xdna {
 struct debug_lock {
   std::lock_guard<std::recursive_mutex> m_lk;
   debug_lock();
@@ -32,4 +32,4 @@ void debugf(const char* format, ...) {
   vprintf(format, args);
   va_end(args);
 }
-}  // namespace
\ No newline at end of file
+}  // namespace shim_xdna
\ No newline at end of file
diff --git a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/shim_debug.h b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/shim_debug.h
index fbc88dc09..e37dc2b55 100644
--- a/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/shim_debug.h
+++ b/runtime/src/iree-amd-aie/driver/xrt-lite/shim/linux/kmq/shim_debug.h
@@ -9,11 +9,10 @@
 #include <cstdio>
 #include <memory>
 #include <string>
-#include <system_error>
 
 #include "llvm/Support/Error.h"
 
-namespace {
+namespace shim_xdna {
 
 void debugf(const char* format, ...);
 
@@ -52,6 +51,6 @@ void shim_info(const char* fmt, Args&&... args) {
   XRT_PRINTF(format.c_str(), getpid(), std::forward<Args>(args)...);
 }
 
-}  // namespace
+}  // namespace shim_xdna
 
 #endif  // SHIM_DEBUG_H