diff --git a/.ci/scripts/test_llama.sh b/.ci/scripts/test_llama.sh
index e1098455470..5e5ed588a2d 100644
--- a/.ci/scripts/test_llama.sh
+++ b/.ci/scripts/test_llama.sh
@@ -51,6 +51,9 @@ UPLOAD_DIR="${UPLOAD_DIR:-}"
 # Default PT2E_QUANTIZE to empty string if not set
 PT2E_QUANTIZE="${PT2E_QUANTIZE:-}"
 
+# Default CMake Build Type to release mode
+CMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE:-Release}
+
 if [[ $# -lt 4 ]]; then # Assuming 4 mandatory args
     echo "Expecting atleast 4 positional arguments"
     echo "Usage: [...]"
@@ -143,7 +146,7 @@ cmake_install_executorch_libraries() {
     rm -rf cmake-out
     retry cmake \
         -DCMAKE_INSTALL_PREFIX=cmake-out \
-        -DCMAKE_BUILD_TYPE=Debug \
+        -DCMAKE_BUILD_TYPE="$CMAKE_BUILD_TYPE" \
         -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
         -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
         -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
@@ -157,7 +160,7 @@ cmake_install_executorch_libraries() {
         -DQNN_SDK_ROOT="$QNN_SDK_ROOT" \
         -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" \
         -Bcmake-out .
-    cmake --build cmake-out -j9 --target install --config Debug
+    cmake --build cmake-out -j9 --target install --config "$CMAKE_BUILD_TYPE"
 }
 
 cmake_build_llama_runner() {
@@ -165,14 +168,14 @@ cmake_build_llama_runner() {
     dir="examples/models/llama"
     retry cmake \
         -DCMAKE_INSTALL_PREFIX=cmake-out \
-        -DCMAKE_BUILD_TYPE=Debug \
+        -DCMAKE_BUILD_TYPE="$CMAKE_BUILD_TYPE" \
         -DEXECUTORCH_BUILD_KERNELS_CUSTOM="$CUSTOM" \
         -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
         -DEXECUTORCH_BUILD_XNNPACK="$XNNPACK" \
         -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" \
         -Bcmake-out/${dir} \
         ${dir}
-    cmake --build cmake-out/${dir} -j9 --config Debug
+    cmake --build cmake-out/${dir} -j9 --config "$CMAKE_BUILD_TYPE"
 
 }
 
diff --git a/.ci/scripts/test_llava.sh b/.ci/scripts/test_llava.sh
index 1057fa8f4ac..a30143d8951 100644
--- a/.ci/scripts/test_llava.sh
+++ b/.ci/scripts/test_llava.sh
@@ -8,11 +8,11 @@
 set -exu
 # shellcheck source=/dev/null
 
-BUILD_TYPE=${1:-Debug}
 TARGET_OS=${2:-Native}
 BUILD_DIR=${3:-cmake-out}
+CMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE:-Release}
 
-echo "Building with BUILD_TYPE: $BUILD_TYPE, TARGET_OS: $TARGET_OS, BUILD_DIR: $BUILD_DIR"
+echo "Building with CMAKE_BUILD_TYPE: $CMAKE_BUILD_TYPE, TARGET_OS: $TARGET_OS, BUILD_DIR: $BUILD_DIR"
 
 if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
     PYTHON_EXECUTABLE=python3
@@ -32,7 +32,7 @@ if hash nproc &> /dev/null; then NPROC=$(nproc); fi
 
 EXECUTORCH_COMMON_CMAKE_ARGS="                      \
         -DCMAKE_INSTALL_PREFIX=${BUILD_DIR}         \
-        -DCMAKE_BUILD_TYPE=${BUILD_TYPE}            \
+        -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}            \
         -DEXECUTORCH_ENABLE_LOGGING=ON              \
         -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON      \
         -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
@@ -49,7 +49,7 @@ cmake_install_executorch_libraries() {
         ${EXECUTORCH_COMMON_CMAKE_ARGS} \
         -B${BUILD_DIR} .
 
-    cmake --build ${BUILD_DIR} -j${NPROC} --target install --config ${BUILD_TYPE}
+    cmake --build ${BUILD_DIR} -j${NPROC} --target install --config ${CMAKE_BUILD_TYPE}
 }
 
 cmake_install_executorch_libraries_for_android() {
@@ -59,14 +59,14 @@ cmake_install_executorch_libraries_for_android() {
         ${EXECUTORCH_COMMON_CMAKE_ARGS}                                         \
         -B${BUILD_DIR} .
 
-    cmake --build ${BUILD_DIR} -j${NPROC} --target install --config ${BUILD_TYPE}
+    cmake --build ${BUILD_DIR} -j${NPROC} --target install --config ${CMAKE_BUILD_TYPE}
 }
 
 
 LLAVA_COMMON_CMAKE_ARGS="                        \
         -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" \
         -DCMAKE_INSTALL_PREFIX=${BUILD_DIR}      \
-        -DCMAKE_BUILD_TYPE=${BUILD_TYPE}         \
+        -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}         \
         -DEXECUTORCH_BUILD_KERNELS_CUSTOM=ON     \
         -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON  \
         -DEXECUTORCH_BUILD_XNNPACK=ON"
@@ -81,7 +81,7 @@ cmake_build_llava_runner() {
         -B${BUILD_DIR}/${dir}             \
         ${dir}
 
-    cmake --build ${BUILD_DIR}/${dir} -j${NPROC} --config ${BUILD_TYPE}
+    cmake --build ${BUILD_DIR}/${dir} -j${NPROC} --config ${CMAKE_BUILD_TYPE}
 }
 
 
@@ -98,7 +98,7 @@ cmake_build_llava_runner_for_android() {
         -B${BUILD_DIR}/${dir}                                                   \
         ${dir}
 
-    cmake --build ${BUILD_DIR}/${dir} -j${NPROC} --config ${BUILD_TYPE}
+    cmake --build ${BUILD_DIR}/${dir} -j${NPROC} --config ${CMAKE_BUILD_TYPE}
 }
 
 # only export the one without custom op for now since it's
diff --git a/.github/workflows/trunk.yml b/.github/workflows/trunk.yml
index 07e0650cd12..741b17c2816 100644
--- a/.github/workflows/trunk.yml
+++ b/.github/workflows/trunk.yml
@@ -290,7 +290,7 @@ jobs:
   #       ${CONDA_RUN} python -m unittest examples.models.llava.test.test_llava
 
   #       # run e2e (export, tokenizer and runner)
-  #       PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/test_llava.sh Release
+  #       PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/test_llava.sh
 
   test-qnn-model:
     name: test-qnn-model
diff --git a/backends/xnnpack/third-party/XNNPACK b/backends/xnnpack/third-party/XNNPACK
index d5d572e46ed..4ea82e595b3 160000
--- a/backends/xnnpack/third-party/XNNPACK
+++ b/backends/xnnpack/third-party/XNNPACK
@@ -1 +1 @@
-Subproject commit d5d572e46ed3929fa3e67f6174192893943cf724
+Subproject commit 4ea82e595b36106653175dcb04b2aa532660d0d8
diff --git a/backends/xnnpack/third-party/xnnpack.buck.bzl b/backends/xnnpack/third-party/xnnpack.buck.bzl
index d2068661fea..6ce0316010d 100644
--- a/backends/xnnpack/third-party/xnnpack.buck.bzl
+++ b/backends/xnnpack/third-party/xnnpack.buck.bzl
@@ -42,7 +42,7 @@ def define_xnnpack():
             "XNNPACK/src/mutex.c",
             "XNNPACK/src/normalization.c",
             "XNNPACK/src/operator-utils.c",
-            "XNNPACK/src/packing.cc",
+            "XNNPACK/src/reference/packing.cc",
         ],
         headers = get_xnnpack_headers(),
         header_namespace = "",
@@ -67,7 +67,7 @@ def define_xnnpack():
     # @lint-ignore BUCKLINT: native and fb_native are explicitly forbidden in fbcode.
     native.cxx_library(
         name = "subgraph",
-        srcs = SUBGRAPH_SRCS,
+        srcs = SUBGRAPH_SRCS + ["XNNPACK/src/datatype.c"],
         compiler_flags = [
             "-Wno-error=missing-braces",  # required since the SGX toolchain does not have this by default
         ],
@@ -1076,6 +1076,8 @@ def define_xnnpack():
             "XNNPACK/src/configs/hardware-config.c",
             "XNNPACK/src/microparams-init.c",
             "XNNPACK/src/microkernel-utils.c",
+            "XNNPACK/src/reference/binary-elementwise.cc",
+            "XNNPACK/src/reference/unary-elementwise.cc",
         ],
         headers = get_xnnpack_headers(),
         exported_headers = {
diff --git a/backends/xnnpack/third-party/xnnpack_src_defs.bzl b/backends/xnnpack/third-party/xnnpack_src_defs.bzl
index 038b90acab0..8cb9affede3 100644
--- a/backends/xnnpack/third-party/xnnpack_src_defs.bzl
+++ b/backends/xnnpack/third-party/xnnpack_src_defs.bzl
@@ -17,24 +17,14 @@ def prod_srcs_for_arch_wrapper(arch):
     return define_xnnpack_build_src(prod_srcs)
 
 def get_xnnpack_headers():
-    # XNNPACK Headers in the path containing xnnpack/ or configs/
-    # do not contain the src/ path. However headers not in xnnpack/ or
-    # configs/ are prepend with the src/ path. This function helps us
-    # to correctly parse all the header files to the correct name
     src_headers = subdir_glob([
         ("XNNPACK/src", "**/*.h"),
     ])
-    fixed_headers = {}
-    for k, v in src_headers.items():
-        new_key = k
-        if not k.startswith("xnnpack") and not k.startswith("configs"):
-            new_key = "src/{}".format(k)
-        fixed_headers[new_key] = v
     include_headers = subdir_glob([
         ("XNNPACK/include", "*.h"),
     ])
 
-    return fixed_headers | include_headers
+    return src_headers | include_headers
 
 OPERATOR_SRCS = define_xnnpack_build_src(_OPERATOR_SRCS)
 SUBGRAPH_SRCS = define_xnnpack_build_src(_SUBGRAPH_SRCS)