diff --git a/.ci/docker/ci_commit_pins/pytorch.txt b/.ci/docker/ci_commit_pins/pytorch.txt
index 242371cbebe..401a0594d98 100644
--- a/.ci/docker/ci_commit_pins/pytorch.txt
+++ b/.ci/docker/ci_commit_pins/pytorch.txt
@@ -1 +1 @@
-release/2.12
+release/2.13
diff --git a/.ci/docker/common/install_pytorch.sh b/.ci/docker/common/install_pytorch.sh
index 3c80d093ab2..c55e3ac878d 100755
--- a/.ci/docker/common/install_pytorch.sh
+++ b/.ci/docker/common/install_pytorch.sh
@@ -16,6 +16,21 @@ install_domains() {
   pip_install --no-build-isolation --user "git+https://github.com/pytorch/vision.git@${TORCHVISION_VERSION}"
 }
 
+configure_pytorch_compiler() {
+  local cxx_version
+  cxx_version=$(/usr/bin/c++ --version | head -n1)
+  if [[ "${cxx_version}" == *clang* ]]; then
+    local clang_major
+    clang_major=$(/usr/bin/c++ -dumpversion | cut -d. -f1)
+    if [[ "${clang_major}" =~ ^[0-9]+$ && "${clang_major}" -lt 16 ]] &&
+      command -v gcc >/dev/null &&
+      command -v g++ >/dev/null; then
+      export CC=gcc
+      export CXX=g++
+    fi
+  fi
+}
+
 install_pytorch_and_domains() {
   git clone https://github.com/pytorch/pytorch.git
 
@@ -33,6 +48,7 @@ install_pytorch_and_domains() {
   if [[ -n "${PYTORCH_BUILD_MAX_JOBS:-}" ]]; then
     export MAX_JOBS="${PYTORCH_BUILD_MAX_JOBS}"
   fi
+  configure_pytorch_compiler
   # Then build and install PyTorch
   conda_run python setup.py bdist_wheel
   pip_install "$(echo dist/*.whl)"
@@ -40,7 +56,7 @@ install_pytorch_and_domains() {
   # Grab the pinned audio and vision commits from PyTorch
   TORCHAUDIO_VERSION=release/2.11
   export TORCHAUDIO_VERSION
-  TORCHVISION_VERSION=release/0.27
+  TORCHVISION_VERSION=release/0.28
   export TORCHVISION_VERSION
 
   install_domains
diff --git a/.ci/scripts/setup-windows.ps1 b/.ci/scripts/setup-windows.ps1
index 2cf28e0cdfe..a38d43bbfd8 100644
--- a/.ci/scripts/setup-windows.ps1
+++ b/.ci/scripts/setup-windows.ps1
@@ -14,6 +14,10 @@ conda activate et
 # Install test dependencies
 pip install -r .ci/docker/requirements-ci.txt
 
+# PyTorch 2.13 headers use designated initializers in c10, which require C++20
+# with MSVC.
+$env:CMAKE_ARGS = "$env:CMAKE_ARGS -DCMAKE_CXX_STANDARD=20"
+
 # The Windows CI image ships CUDA toolkits on PATH, so install_executorch
 # (setup.py) auto-enables EXECUTORCH_BUILD_CUDA whenever the detected nvcc
 # version is in SUPPORTED_CUDA_VERSIONS. CPU-only jobs install CPU torch, so a
diff --git a/.ci/scripts/test_model_e2e_windows.ps1 b/.ci/scripts/test_model_e2e_windows.ps1
index 73a78c92112..5a2d8eccf00 100644
--- a/.ci/scripts/test_model_e2e_windows.ps1
+++ b/.ci/scripts/test_model_e2e_windows.ps1
@@ -172,22 +172,35 @@ try {
             "-DCUDAToolkit_ROOT=$env:CUDA_HOME"
         )
     }
+    $cmakeCommonArgs = @("-DCMAKE_CXX_STANDARD=20")
     Write-Host "::endgroup::"
 
     Write-Host "::group::Build ExecuTorch (CUDA)"
     $numCores = [Math]::Max([Environment]::ProcessorCount - 1, 1)
+    $cmakeOut = Join-Path -Path $executorchRoot -ChildPath "cmake-out"
+    $executorchCmakeDir = Join-Path -Path $cmakeOut -ChildPath "lib\cmake\ExecuTorch"
+    $executorchConfig = Join-Path -Path $executorchCmakeDir -ChildPath "executorch-config.cmake"
     # EXECUTORCH_BUILD_EXTENSION_IMAGE defaults OFF in the preset, so it must be
     # enabled explicitly here (matching the Makefile CUDA target used on Linux).
     # The dinov2 runner links the installed extension_image.lib; without this the
     # main install never builds it and dinov2_runner fails to link (LNK1181).
-    cmake --preset llm-release-cuda -DEXECUTORCH_BUILD_EXTENSION_IMAGE=ON @cmakeCudaArgs
-    cmake --build cmake-out --target install --config Release -j $numCores
+    cmake --preset llm-release-cuda -DEXECUTORCH_BUILD_EXTENSION_IMAGE=ON @cmakeCommonArgs @cmakeCudaArgs
+    cmake --build $cmakeOut --config Release -j $numCores
+    cmake --install $cmakeOut --config Release
+    if (-not (Test-Path -Path $executorchConfig -PathType Leaf)) {
+        throw "ExecuTorch CMake package config not found after install: $executorchConfig"
+    }
     Write-Host "::endgroup::"
 
     Write-Host "::group::Build $runnerTarget"
+    $cmakePackageArgs = @(
+        "-DCMAKE_FIND_ROOT_PATH=$cmakeOut",
+        "-DCMAKE_PREFIX_PATH=$cmakeOut;$executorchCmakeDir",
+        "-Dexecutorch_DIR=$executorchCmakeDir"
+    )
     Push-Location (Join-Path -Path $executorchRoot -ChildPath "examples\models\$runnerPath")
     try {
-        cmake --preset $runnerPreset @cmakeCudaArgs
+        cmake --preset $runnerPreset @cmakeCommonArgs @cmakePackageArgs @cmakeCudaArgs
         cmake --build (Join-Path -Path $executorchRoot -ChildPath "cmake-out\examples\models\$runnerPath") --target $runnerTarget --config Release -j $numCores
     }
     finally {
diff --git a/.ci/scripts/utils.sh b/.ci/scripts/utils.sh
index b312d0ede83..234e162e48e 100644
--- a/.ci/scripts/utils.sh
+++ b/.ci/scripts/utils.sh
@@ -107,7 +107,7 @@ install_pytorch_and_domains() {
   local torch_release=$(cat version.txt)
   # Download key must match the upload key below (basename of dist/*.whl,
   # which always carries setup.py's resolved +gitHASH). Branch-ref pins
-  # like `release/2.12` would otherwise produce `+gitrelease` here and
+  # like `release/2.13` would otherwise produce `+gitrelease` here and
   # never hit the cache.
   local torch_short_hash=$(git rev-parse --short=7 HEAD)
   local torch_wheel_path="cached_artifacts/pytorch/executorch/pytorch_wheels/${system_name}/${python_version}"
@@ -178,7 +178,7 @@ install_pytorch_and_domains() {
   # Grab the pinned audio and vision commits from PyTorch
   TORCHAUDIO_VERSION=release/2.11
   export TORCHAUDIO_VERSION
-  TORCHVISION_VERSION=release/0.27
+  TORCHVISION_VERSION=release/0.28
   export TORCHVISION_VERSION
 
   install_domains
diff --git a/.github/workflows/mlx.yml b/.github/workflows/mlx.yml
index 40ebabbcf28..0de68134c4e 100644
--- a/.github/workflows/mlx.yml
+++ b/.github/workflows/mlx.yml
@@ -158,10 +158,10 @@ jobs:
           --prompt-len 4 \
           --max-new-tokens 5 2>&1)
         echo "$OUTPUT"
-        if echo "$OUTPUT" | grep -q "Generated token ids: \[167, 94, 253, 88, 227\]"; then
+        if echo "$OUTPUT" | grep -q "Generated token ids: \[167, 180, 167, 180, 167\]"; then
           echo "Success: Qwen 3.5 MoE MLX export + inference completed with expected output"
         else
-          echo "Failed: unexpected output (expected [167, 94, 253, 88, 227])"
+          echo "Failed: unexpected output (expected [167, 180, 167, 180, 167])"
           exit 1
         fi
         echo "::endgroup::"
diff --git a/backends/aoti/common_shims.cpp b/backends/aoti/common_shims.cpp
index 49d445e955d..f3a34a09987 100644
--- a/backends/aoti/common_shims.cpp
+++ b/backends/aoti/common_shims.cpp
@@ -154,6 +154,11 @@ AOTITorchError aoti_torch_get_dim(Tensor* tensor, int64_t* ret_dim) {
   return Error::Ok;
 }
 
+AOTITorchError aoti_torch_get_numel(Tensor* tensor, int64_t* ret_numel) {
+  *ret_numel = static_cast<int64_t>(tensor->numel());
+  return Error::Ok;
+}
+
 // Device and layout utility functions
 int32_t aoti_torch_device_type_cpu() {
   // Let's say cpu is 0 for ET as well
diff --git a/backends/aoti/common_shims.h b/backends/aoti/common_shims.h
index 869aece5a3b..d057279e22a 100644
--- a/backends/aoti/common_shims.h
+++ b/backends/aoti/common_shims.h
@@ -59,6 +59,9 @@ aoti_torch_get_device_index(Tensor* tensor, int32_t* ret_device_index);
 AOTI_SHIM_EXPORT AOTITorchError
 aoti_torch_get_dim(Tensor* tensor, int64_t* ret_dim);
 
+AOTI_SHIM_EXPORT AOTITorchError
+aoti_torch_get_numel(Tensor* tensor, int64_t* ret_numel);
+
 // Utility functions for device and layout information
 AOTI_SHIM_EXPORT int32_t aoti_torch_device_type_cpu();
 AOTI_SHIM_EXPORT int32_t aoti_torch_layout_strided();
diff --git a/backends/aoti/common_shims_slim.cpp b/backends/aoti/common_shims_slim.cpp
index f4b50400778..c8c7408aa62 100644
--- a/backends/aoti/common_shims_slim.cpp
+++ b/backends/aoti/common_shims_slim.cpp
@@ -60,6 +60,14 @@ AOTITorchError aoti_torch_get_dim(Tensor* tensor, int64_t* ret_dim) {
   return Error::Ok;
 }
 
+AOTITorchError aoti_torch_get_numel(Tensor* tensor, int64_t* ret_numel) {
+  if (tensor == nullptr || ret_numel == nullptr) {
+    return Error::InvalidArgument;
+  }
+  *ret_numel = static_cast<int64_t>(tensor->numel());
+  return Error::Ok;
+}
+
 int32_t aoti_torch_layout_strided() {
   // Slimtensor only support strided layout, the return value will always be 0,
   // a.k.a at::Layout::Strided;
diff --git a/backends/aoti/common_shims_slim.h b/backends/aoti/common_shims_slim.h
index 3ec867bcf6e..c5a5cab9413 100644
--- a/backends/aoti/common_shims_slim.h
+++ b/backends/aoti/common_shims_slim.h
@@ -48,6 +48,9 @@ aoti_torch_get_dtype(Tensor* tensor, int32_t* ret_dtype);
 AOTI_SHIM_EXPORT AOTITorchError
 aoti_torch_get_dim(Tensor* tensor, int64_t* ret_dim);
 
+AOTI_SHIM_EXPORT AOTITorchError
+aoti_torch_get_numel(Tensor* tensor, int64_t* ret_numel);
+
 AOTI_SHIM_EXPORT int32_t aoti_torch_layout_strided();
 
 // ============================================================
diff --git a/backends/aoti/tests/test_common_shims.cpp b/backends/aoti/tests/test_common_shims.cpp
index 0fd1b057f99..b2dc2c01dc2 100644
--- a/backends/aoti/tests/test_common_shims.cpp
+++ b/backends/aoti/tests/test_common_shims.cpp
@@ -241,6 +241,21 @@ TEST_F(CommonShimsTest, ScalarTensor) {
   error = aoti_torch_get_sizes(tensor_0d, &sizes_ptr2);
   EXPECT_EQ(error, Error::Ok);
   EXPECT_EQ(sizes_ptr, sizes_ptr2);
+
+  int64_t numel = -1;
+  error = aoti_torch_get_numel(tensor_0d, &numel);
+  EXPECT_EQ(error, Error::Ok);
+  EXPECT_EQ(numel, 1);
+}
+
+TEST_F(CommonShimsTest, GetNumel) {
+  auto tensor = create_tracked_tensor({2, 3, 4});
+
+  int64_t numel = -1;
+  AOTITorchError error = aoti_torch_get_numel(tensor, &numel);
+
+  EXPECT_EQ(error, Error::Ok);
+  EXPECT_EQ(numel, 24);
 }
 
 // Test large tensor dimensions
@@ -266,6 +281,11 @@ TEST_F(CommonShimsTest, LargeTensorDimensions) {
   EXPECT_EQ(strides_ptr[1], 120000);
   EXPECT_EQ(strides_ptr[2], 400);
   EXPECT_EQ(strides_ptr[3], 1);
+
+  int64_t numel = -1;
+  error = aoti_torch_get_numel(tensor, &numel);
+  EXPECT_EQ(error, Error::Ok);
+  EXPECT_EQ(numel, 2400000000);
 }
 
 // Test that cleanup_tensor_metadata clears the cache
diff --git a/backends/aoti/tests/test_common_shims_slim.cpp b/backends/aoti/tests/test_common_shims_slim.cpp
index 94319c6f94d..2dd8475a4d8 100644
--- a/backends/aoti/tests/test_common_shims_slim.cpp
+++ b/backends/aoti/tests/test_common_shims_slim.cpp
@@ -289,6 +289,50 @@ void runGetDimTest(slim_c10::DeviceType device_type) {
   }
 }
 
+void runGetNumelTest(slim_c10::DeviceType device_type) {
+  slim_c10::Device device(device_type, 0);
+
+  // Test 0D tensor (scalar)
+  {
+    std::vector<int64_t> sizes = {};
+    std::vector<int64_t> strides = {};
+
+    Tensor* tensor = new Tensor(slim::empty_strided(
+        slim::makeArrayRef(sizes),
+        slim::makeArrayRef(strides),
+        slim_c10::ScalarType::Float,
+        device));
+
+    int64_t ret_numel = -1;
+    AOTITorchError error = aoti_torch_get_numel(tensor, &ret_numel);
+
+    EXPECT_EQ(error, Error::Ok);
+    EXPECT_EQ(ret_numel, 1);
+
+    delete tensor;
+  }
+
+  // Test 3D tensor
+  {
+    std::vector<int64_t> sizes = {2, 3, 4};
+    std::vector<int64_t> strides = calculateContiguousStrides(sizes);
+
+    Tensor* tensor = new Tensor(slim::empty_strided(
+        slim::makeArrayRef(sizes),
+        slim::makeArrayRef(strides),
+        slim_c10::ScalarType::Float,
+        device));
+
+    int64_t ret_numel = -1;
+    AOTITorchError error = aoti_torch_get_numel(tensor, &ret_numel);
+
+    EXPECT_EQ(error, Error::Ok);
+    EXPECT_EQ(ret_numel, 24);
+
+    delete tensor;
+  }
+}
+
 // ============================================================================
 // Storage & Device Property Tests
 // ============================================================================
@@ -400,6 +444,10 @@ TEST_F(CommonShimsSlimTest, GetDim_CPU) {
   runGetDimTest(slim_c10::DeviceType::CPU);
 }
 
+TEST_F(CommonShimsSlimTest, GetNumel_CPU) {
+  runGetNumelTest(slim_c10::DeviceType::CPU);
+}
+
 TEST_F(CommonShimsSlimTest, GetStorageOffset_CPU) {
   runGetStorageOffsetTest(slim_c10::DeviceType::CPU);
 }
@@ -456,6 +504,13 @@ TEST_F(CommonShimsSlimTest, GetDim_CUDA) {
   runGetDimTest(slim_c10::DeviceType::CUDA);
 }
 
+TEST_F(CommonShimsSlimTest, GetNumel_CUDA) {
+  if (!isCudaAvailable()) {
+    GTEST_SKIP() << "CUDA not available";
+  }
+  runGetNumelTest(slim_c10::DeviceType::CUDA);
+}
+
 TEST_F(CommonShimsSlimTest, GetStorageOffset_CUDA) {
   if (!isCudaAvailable()) {
     GTEST_SKIP() << "CUDA not available";
@@ -495,6 +550,7 @@ TEST_F(CommonShimsSlimTest, NullTensorArgument) {
   int64_t* strides = nullptr;
   int32_t dtype = -1;
   int64_t dim = -1;
+  int64_t numel = -1;
 
   EXPECT_EQ(
       aoti_torch_get_data_ptr(nullptr, &data_ptr), Error::InvalidArgument);
@@ -502,6 +558,7 @@ TEST_F(CommonShimsSlimTest, NullTensorArgument) {
   EXPECT_EQ(aoti_torch_get_strides(nullptr, &strides), Error::InvalidArgument);
   EXPECT_EQ(aoti_torch_get_dtype(nullptr, &dtype), Error::InvalidArgument);
   EXPECT_EQ(aoti_torch_get_dim(nullptr, &dim), Error::InvalidArgument);
+  EXPECT_EQ(aoti_torch_get_numel(nullptr, &numel), Error::InvalidArgument);
 }
 
 TEST_F(CommonShimsSlimTest, NullReturnPointer) {
@@ -512,6 +569,7 @@ TEST_F(CommonShimsSlimTest, NullReturnPointer) {
   EXPECT_EQ(aoti_torch_get_strides(tensor, nullptr), Error::InvalidArgument);
   EXPECT_EQ(aoti_torch_get_dtype(tensor, nullptr), Error::InvalidArgument);
   EXPECT_EQ(aoti_torch_get_dim(tensor, nullptr), Error::InvalidArgument);
+  EXPECT_EQ(aoti_torch_get_numel(tensor, nullptr), Error::InvalidArgument);
 }
 
 // ============================================================================
@@ -534,6 +592,7 @@ TEST_F(CommonShimsSlimTest, ScalarTensor) {
   int64_t* ret_sizes = nullptr;
   int64_t* ret_strides = nullptr;
   int64_t ret_dim = -1;
+  int64_t ret_numel = -1;
 
   EXPECT_EQ(aoti_torch_get_sizes(tensor, &ret_sizes), Error::Ok);
   EXPECT_NE(ret_sizes, nullptr);
@@ -543,6 +602,9 @@ TEST_F(CommonShimsSlimTest, ScalarTensor) {
 
   EXPECT_EQ(aoti_torch_get_dim(tensor, &ret_dim), Error::Ok);
   EXPECT_EQ(ret_dim, 0);
+
+  EXPECT_EQ(aoti_torch_get_numel(tensor, &ret_numel), Error::Ok);
+  EXPECT_EQ(ret_numel, 1);
 }
 
 TEST_F(CommonShimsSlimTest, LargeTensor) {
@@ -559,6 +621,7 @@ TEST_F(CommonShimsSlimTest, LargeTensor) {
 
   int64_t* ret_sizes = nullptr;
   int64_t* ret_strides = nullptr;
+  int64_t ret_numel = -1;
 
   EXPECT_EQ(aoti_torch_get_sizes(tensor, &ret_sizes), Error::Ok);
   EXPECT_EQ(ret_sizes[0], 100);
@@ -569,6 +632,9 @@ TEST_F(CommonShimsSlimTest, LargeTensor) {
   EXPECT_EQ(ret_strides[0], 60000); // 200 * 300
   EXPECT_EQ(ret_strides[1], 300); // 300
   EXPECT_EQ(ret_strides[2], 1);
+
+  EXPECT_EQ(aoti_torch_get_numel(tensor, &ret_numel), Error::Ok);
+  EXPECT_EQ(ret_numel, 6000000);
 }
 
 TEST_F(CommonShimsSlimTest, ConsistentPointerReturn) {
diff --git a/backends/cuda/cuda_backend.py b/backends/cuda/cuda_backend.py
index ef746830bed..bf28077c62c 100644
--- a/backends/cuda/cuda_backend.py
+++ b/backends/cuda/cuda_backend.py
@@ -263,6 +263,13 @@ def _move_to_device_resize_kv(ep, location):
     return ep
 
 
+def _on_off_compile_spec_value(spec: CompileSpec) -> bool:
+    value = spec.value.decode("utf-8").upper()
+    if value not in ["ON", "OFF"]:
+        raise ValueError(f"Invalid {spec.key}: {value}. Expected 'ON' or 'OFF'.")
+    return value == "ON"
+
+
 @final
 @experimental(
     "This API and all of cuda backend related functionality are experimental."
@@ -496,12 +503,25 @@ def get_aoti_compile_options(
         # Parse compile_specs to check for platform
 
         platform = "linux"
+        emulate_precision_casts = True
+        max_autotune = True
+        autotune_at_compile_time = None
         shim_library_path = None
         for spec in compile_specs:
             if spec.key == "platform":
                 platform = spec.value.decode("utf-8")
-            if spec.key == "shim_library_path":
+            elif spec.key == "emulate_precision_casts":
+                emulate_precision_casts = _on_off_compile_spec_value(spec)
+            elif spec.key == "max_autotune":
+                max_autotune = _on_off_compile_spec_value(spec)
+            elif spec.key == "autotune_at_compile_time":
+                autotune_at_compile_time = _on_off_compile_spec_value(spec)
+            elif spec.key == "shim_library_path":
                 shim_library_path = spec.value.decode("utf-8")
+        options["emulate_precision_casts"] = emulate_precision_casts
+        options["max_autotune"] = max_autotune
+        if autotune_at_compile_time is not None:
+            options["triton.autotune_at_compile_time"] = autotune_at_compile_time
         # Add platform-specific options
 
         if platform == "windows":
diff --git a/backends/cuda/runtime/aoti_cuda_shims.lib b/backends/cuda/runtime/aoti_cuda_shims.lib
index 63c8d79bd7a..8bb03cc1c1e 100644
Binary files a/backends/cuda/runtime/aoti_cuda_shims.lib and b/backends/cuda/runtime/aoti_cuda_shims.lib differ
diff --git a/backends/cuda/tests/test_cuda_export.py b/backends/cuda/tests/test_cuda_export.py
index 83d792187d1..2bd3d4db109 100644
--- a/backends/cuda/tests/test_cuda_export.py
+++ b/backends/cuda/tests/test_cuda_export.py
@@ -16,6 +16,47 @@
 from torch.export import export
 
 
+class TestCudaBackendCompileOptions(unittest.TestCase):
+    def test_emulate_precision_casts_compile_spec(self):
+        options = CudaBackend.get_aoti_compile_options(
+            [CompileSpec(key="emulate_precision_casts", value=b"OFF")]
+        )
+
+        self.assertFalse(options["emulate_precision_casts"])
+
+    def test_invalid_emulate_precision_casts_compile_spec(self):
+        with self.assertRaisesRegex(ValueError, "Invalid emulate_precision_casts"):
+            CudaBackend.get_aoti_compile_options(
+                [CompileSpec(key="emulate_precision_casts", value=b"MAYBE")]
+            )
+
+    def test_max_autotune_compile_spec(self):
+        options = CudaBackend.get_aoti_compile_options(
+            [CompileSpec(key="max_autotune", value=b"OFF")]
+        )
+
+        self.assertFalse(options["max_autotune"])
+
+    def test_invalid_max_autotune_compile_spec(self):
+        with self.assertRaisesRegex(ValueError, "Invalid max_autotune"):
+            CudaBackend.get_aoti_compile_options(
+                [CompileSpec(key="max_autotune", value=b"MAYBE")]
+            )
+
+    def test_autotune_at_compile_time_compile_spec(self):
+        options = CudaBackend.get_aoti_compile_options(
+            [CompileSpec(key="autotune_at_compile_time", value=b"OFF")]
+        )
+
+        self.assertFalse(options["triton.autotune_at_compile_time"])
+
+    def test_invalid_autotune_at_compile_time_compile_spec(self):
+        with self.assertRaisesRegex(ValueError, "Invalid autotune_at_compile_time"):
+            CudaBackend.get_aoti_compile_options(
+                [CompileSpec(key="autotune_at_compile_time", value=b"MAYBE")]
+            )
+
+
 class TestCudaExport(unittest.TestCase):
     """Test CUDA export functionality for various operations using to_edge_transform_and_lower."""
 
diff --git a/examples/models/dinov2/CMakeLists.txt b/examples/models/dinov2/CMakeLists.txt
index a2af3002a35..8e83ee44873 100644
--- a/examples/models/dinov2/CMakeLists.txt
+++ b/examples/models/dinov2/CMakeLists.txt
@@ -7,7 +7,9 @@
 cmake_minimum_required(VERSION 3.24)
 project(dinov2_runner)
 
-set(CMAKE_CXX_STANDARD 17)
+if(NOT CMAKE_CXX_STANDARD)
+  set(CMAKE_CXX_STANDARD 17)
+endif()
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
 set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../../..)
diff --git a/examples/models/llama/install_requirements.sh b/examples/models/llama/install_requirements.sh
index 8a2fa25d244..7692306c50b 100755
--- a/examples/models/llama/install_requirements.sh
+++ b/examples/models/llama/install_requirements.sh
@@ -11,7 +11,18 @@
 # Install snakeviz for cProfile flamegraph
 # Install lm-eval for Model Evaluation with lm-evalution-harness.
 # Install safetensors to load safetensors checkpoints (currently adapter only).
-pip install hydra-core huggingface_hub tiktoken torchtune sentencepiece tokenizers snakeviz lm_eval==0.4.5 blobfile safetensors
+TORCHTUNE_COMMIT=3c1872ace149f03ef4ffec765d3f0a5fd0399497
+pip install \
+  hydra-core \
+  huggingface_hub \
+  tiktoken \
+  "torchtune @ git+https://github.com/pytorch/torchtune.git@${TORCHTUNE_COMMIT}" \
+  sentencepiece \
+  tokenizers \
+  snakeviz \
+  lm_eval==0.4.5 \
+  blobfile \
+  safetensors
 
 # Call the install helper for further setup
 python examples/models/llama/install_requirement_helper.py
diff --git a/examples/models/parakeet/CMakeLists.txt b/examples/models/parakeet/CMakeLists.txt
index a2b798de557..9449583c613 100644
--- a/examples/models/parakeet/CMakeLists.txt
+++ b/examples/models/parakeet/CMakeLists.txt
@@ -7,7 +7,9 @@
 cmake_minimum_required(VERSION 3.24)
 project(parakeet_runner)
 
-set(CMAKE_CXX_STANDARD 17)
+if(NOT CMAKE_CXX_STANDARD)
+  set(CMAKE_CXX_STANDARD 17)
+endif()
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
 set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../../..)
diff --git a/examples/models/qwen3_5_moe/export.py b/examples/models/qwen3_5_moe/export.py
index d2c3914c6a2..9485a5ee32c 100644
--- a/examples/models/qwen3_5_moe/export.py
+++ b/examples/models/qwen3_5_moe/export.py
@@ -1073,6 +1073,8 @@ def _export_cuda(model, config, args):
         "enable_dynamic_shape": True,
         "get_mutable_buffer_metadata": _mutable_buffer_metadata_json(model),
     }
+    # Avoid a PyTorch 2.13 Triton codegen bug in Qwen's fused cast/reduction
+    # compile-time autotune path while keeping the CUDA backend defaults unchanged.
     et_prog = to_edge_transform_and_lower(
         {"decode": decode_ep, "prefill": prefill_ep},
         partitioner={
@@ -1081,6 +1083,9 @@ def _export_cuda(model, config, args):
                     [
                         CudaBackend.generate_method_name_compile_spec("decode"),
                         CompileSpec("low_memory_mode", b"ON"),
+                        CompileSpec("emulate_precision_casts", b"OFF"),
+                        CompileSpec("max_autotune", b"OFF"),
+                        CompileSpec("autotune_at_compile_time", b"OFF"),
                     ]
                 )
             ],
@@ -1089,6 +1094,9 @@ def _export_cuda(model, config, args):
                     [
                         CudaBackend.generate_method_name_compile_spec("prefill"),
                         CompileSpec("low_memory_mode", b"ON"),
+                        CompileSpec("emulate_precision_casts", b"OFF"),
+                        CompileSpec("max_autotune", b"OFF"),
+                        CompileSpec("autotune_at_compile_time", b"OFF"),
                     ]
                 )
             ],
diff --git a/examples/models/sortformer/CMakeLists.txt b/examples/models/sortformer/CMakeLists.txt
index bdc0fbfbbe5..171eb726ed2 100644
--- a/examples/models/sortformer/CMakeLists.txt
+++ b/examples/models/sortformer/CMakeLists.txt
@@ -7,7 +7,9 @@
 cmake_minimum_required(VERSION 3.24)
 project(sortformer_runner)
 
-set(CMAKE_CXX_STANDARD 17)
+if(NOT CMAKE_CXX_STANDARD)
+  set(CMAKE_CXX_STANDARD 17)
+endif()
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
 set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../../..)
diff --git a/examples/models/voxtral_realtime/CMakeLists.txt b/examples/models/voxtral_realtime/CMakeLists.txt
index 28545f407ca..93fa9249cad 100644
--- a/examples/models/voxtral_realtime/CMakeLists.txt
+++ b/examples/models/voxtral_realtime/CMakeLists.txt
@@ -7,7 +7,9 @@
 cmake_minimum_required(VERSION 3.24)
 project(voxtral_realtime)
 
-set(CMAKE_CXX_STANDARD 17)
+if(NOT CMAKE_CXX_STANDARD)
+  set(CMAKE_CXX_STANDARD 17)
+endif()
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
 set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../../..)
diff --git a/extension/llm/export/nvfp4.py b/extension/llm/export/nvfp4.py
index feeb95f50a6..b0671d11173 100644
--- a/extension/llm/export/nvfp4.py
+++ b/extension/llm/export/nvfp4.py
@@ -161,7 +161,7 @@ def _exportable_nvfp4_transform(module: nn.Module, config: ExportableNVFP4Config
             f"NVFP4 requires weight dims divisible by 16, got {weight.shape}"
         )
 
-    per_tensor_scale = 1.0
+    per_tensor_scale = None
     if config.use_per_tensor_scale:
         tensor_amax = torch.max(torch.abs(weight))
         per_tensor_scale = per_tensor_amax_to_scale(tensor_amax)
@@ -173,7 +173,11 @@ def _exportable_nvfp4_transform(module: nn.Module, config: ExportableNVFP4Config
     qdata_u32 = qdata_packed.view(torch.uint32)
     scales_u8 = scales_fp8.view(torch.uint8)
 
-    pts = torch.tensor(per_tensor_scale, dtype=torch.float32)
+    pts = torch.as_tensor(
+        1.0 if per_tensor_scale is None else per_tensor_scale,
+        dtype=torch.float32,
+        device=weight.device,
+    )
     quantized_weight = ExportableNVFP4Tensor(
         qdata_u32,
         scales_u8,
diff --git a/install_requirements.py b/install_requirements.py
index 53204ffd3ee..94485804a42 100644
--- a/install_requirements.py
+++ b/install_requirements.py
@@ -15,6 +15,8 @@
 # The pip repository that hosts nightly torch packages.
 # This will be dynamically set based on CUDA availability and CUDA backend enabled/disabled.
 TORCH_URL_BASE = "https://download.pytorch.org/whl/test"
+TORCHAO_URL_BASE = "https://download.pytorch.org/whl/nightly"
+TORCHAO_NIGHTLY_VERSION = "0.18.0.dev20260715"
 
 # Since ExecuTorch often uses main-branch features of pytorch, only the nightly
 # pip versions will have the required features.
@@ -43,13 +45,15 @@ def install_requirements(use_pytorch_nightly):
 
     # Determine the appropriate PyTorch URL based on CUDA delegate status
     torch_url = determine_torch_url(TORCH_URL_BASE)
+    torchao_url = determine_torch_url(TORCHAO_URL_BASE)
 
     # pip packages needed by exir.
     TORCH_PACKAGE = [
         # Setting use_pytorch_nightly to false to test the pinned PyTorch commit. Note
         # that we don't need to set any version number there because they have already
         # been installed on CI before this step, so pip won't reinstall them
-        ("torch==2.12.0" if use_pytorch_nightly else "torch"),
+        ("torch==2.13.0" if use_pytorch_nightly else "torch"),
+        f"torchao=={TORCHAO_NIGHTLY_VERSION}",
     ]
 
     # Install the requirements for core ExecuTorch package.
@@ -66,44 +70,45 @@ def install_requirements(use_pytorch_nightly):
             *TORCH_PACKAGE,
             "--extra-index-url",
             torch_url,
+            "--extra-index-url",
+            torchao_url,
         ],
         check=True,
     )
 
-    LOCAL_REQUIREMENTS = [
-        "third-party/ao",  # We need the latest kernels for fast iteration, so not relying on pypi.
-    ] + (
-        [
-            "extension/llm/tokenizers",  # TODO(larryliu0820): Setup a pypi package for this.
-        ]
-        if sys.platform != "win32"
-        else []
-    )  # TODO(gjcomer): Re-enable when buildable on Windows.
-
-    # Install packages directly from local copy instead of pypi.
-    # This is usually not recommended.
-    new_env = os.environ.copy()
-    if ("EXECUTORCH_BUILD_KERNELS_TORCHAO" not in new_env) or (
-        new_env["EXECUTORCH_BUILD_KERNELS_TORCHAO"] == "0"
-    ):
-        new_env["USE_CPP"] = "0"
-    else:
-        assert new_env["EXECUTORCH_BUILD_KERNELS_TORCHAO"] == "1"
-        new_env["USE_CPP"] = "1"
-        new_env["CMAKE_POLICY_VERSION_MINIMUM"] = "3.5"
-    subprocess.run(
-        [
-            sys.executable,
-            "-m",
-            "pip",
-            "install",
-            # Without --no-build-isolation, setup.py can't find the torch module.
-            "--no-build-isolation",
-            *LOCAL_REQUIREMENTS,
-        ],
-        env=new_env,
-        check=True,
-    )
+    LOCAL_REQUIREMENTS = []
+    if os.environ.get("TORCHAO_BUILD_EXPERIMENTAL_MPS") == "1":
+        LOCAL_REQUIREMENTS.append("third-party/ao")
+    if sys.platform != "win32":
+        # TODO(larryliu0820): Setup a pypi package for this.
+        LOCAL_REQUIREMENTS.append("extension/llm/tokenizers")
+    # TODO(gjcomer): Re-enable when buildable on Windows.
+
+    if LOCAL_REQUIREMENTS:
+        # Install packages directly from local copy instead of pypi.
+        # This is usually not recommended.
+        new_env = os.environ.copy()
+        if ("EXECUTORCH_BUILD_KERNELS_TORCHAO" not in new_env) or (
+            new_env["EXECUTORCH_BUILD_KERNELS_TORCHAO"] == "0"
+        ):
+            new_env["USE_CPP"] = "0"
+        else:
+            assert new_env["EXECUTORCH_BUILD_KERNELS_TORCHAO"] == "1"
+            new_env["USE_CPP"] = "1"
+            new_env["CMAKE_POLICY_VERSION_MINIMUM"] = "3.5"
+        subprocess.run(
+            [
+                sys.executable,
+                "-m",
+                "pip",
+                "install",
+                # Without --no-build-isolation, setup.py can't find the torch module.
+                "--no-build-isolation",
+                *LOCAL_REQUIREMENTS,
+            ],
+            env=new_env,
+            check=True,
+        )
 
 
 def install_optional_example_requirements(use_pytorch_nightly):
@@ -112,7 +117,7 @@ def install_optional_example_requirements(use_pytorch_nightly):
 
     print("Installing torch domain libraries")
     DOMAIN_LIBRARIES = [
-        ("torchvision==0.27.0" if use_pytorch_nightly else "torchvision"),
+        ("torchvision==0.28.0" if use_pytorch_nightly else "torchvision"),
         ("torchaudio==2.11.0" if use_pytorch_nightly else "torchaudio"),
     ]
     # Then install domain libraries
diff --git a/requirements-examples.txt b/requirements-examples.txt
index f1579edcb14..1f36b7c6108 100644
--- a/requirements-examples.txt
+++ b/requirements-examples.txt
@@ -3,5 +3,5 @@
 datasets == 3.6.0 # 4.0.0 deprecates trust_remote_code and load scripts. For now pin to 3.6.0
 timm == 1.0.7
 torchsr == 1.0.4
-torchtune @ git+https://github.com/pytorch/torchtune.git@6f2aa7254458145f99d7004cbd6ebc8e53a06404
+torchtune @ git+https://github.com/pytorch/torchtune.git@3c1872ace149f03ef4ffec765d3f0a5fd0399497
 transformers == 5.0.0rc1
diff --git a/runtime/core/portable_type/c10/c10/util/BFloat16-math.h b/runtime/core/portable_type/c10/c10/util/BFloat16-math.h
index 8291cd74481..bce89e8acd9 100644
--- a/runtime/core/portable_type/c10/c10/util/BFloat16-math.h
+++ b/runtime/core/portable_type/c10/c10/util/BFloat16-math.h
@@ -181,7 +181,7 @@ template <
     typename T,
     typename std::enable_if_t<c10::is_reduced_floating_point_v<T>, int> = 0>
 inline T rsqrt(T a) {
-  return 1.0 / std::sqrt(float(a));
+  return 1.0f / std::sqrt(float(a));
 }
 template <
     typename T,
diff --git a/runtime/core/portable_type/c10/c10/util/complex_math.h b/runtime/core/portable_type/c10/c10/util/complex_math.h
index d369df50592..2b9bbea6c71 100644
--- a/runtime/core/portable_type/c10/c10/util/complex_math.h
+++ b/runtime/core/portable_type/c10/c10/util/complex_math.h
@@ -327,7 +327,7 @@ C10_HOST_DEVICE inline c10::complex<T> atanh(const c10::complex<T>& x) {
 template <typename T>
 C10_HOST_DEVICE inline c10::complex<T> log1p(const c10::complex<T>& z) {
 #if defined(__APPLE__) || defined(__MACOSX) || defined(__CUDACC__) || \
-    defined(__HIPCC__)
+    defined(__HIPCC__) || defined(__SYCL_DEVICE_ONLY__)
   // For Mac, the new implementation yielded a high relative error. Falling back
   // to the old version for now.
   // See https://github.com/numpy/numpy/pull/22611#issuecomment-1667945354
diff --git a/runtime/core/portable_type/c10/torch/headeronly/util/Half.h b/runtime/core/portable_type/c10/torch/headeronly/util/Half.h
index a9c0b166ba2..e5aa622656c 100644
--- a/runtime/core/portable_type/c10/torch/headeronly/util/Half.h
+++ b/runtime/core/portable_type/c10/torch/headeronly/util/Half.h
@@ -236,7 +236,7 @@ C10_HOST_DEVICE inline float fp16_ieee_to_fp32_value(uint16_t h) {
   /*
    * - Choose either results of conversion of input as a normalized number, or
    * as a denormalized number, depending on the input exponent. The variable
-   * two_w contains input exponent in bits 27-31, therefore if its smaller than
+   * two_w contains input exponent in bits 27-31, therefore if it's smaller than
    * 2**27, the input is either a denormal number, or zero.
    * - Combine the result of conversion of exponent and mantissa with the sign
    * of the input number.
diff --git a/third-party/ao b/third-party/ao
index 02105d46c61..4aa81011346 160000
--- a/third-party/ao
+++ b/third-party/ao
@@ -1 +1 @@
-Subproject commit 02105d46c61dc80a8c9d39d5836e827ba3af8439
+Subproject commit 4aa810113463571cabb7342d1a00a1f3de9c54f2
diff --git a/torch_pin.py b/torch_pin.py
index 0c5cd50fe6d..ca593b1ef05 100644
--- a/torch_pin.py
+++ b/torch_pin.py
@@ -1,2 +1,2 @@
-TORCH_VERSION = "2.12.0"
+TORCH_VERSION = "2.13.0"
 # NIGHTLY_VERSION = "dev20260318" Temporarily pinning to stable release candidate. Revert https://github.com/pytorch/executorch/pull/18287