diff --git a/CMakeLists.txt b/CMakeLists.txt
index ad9342580cdf6..c76a942089188 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -571,6 +571,9 @@ if(MSVC)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /Zc:__cplusplus")
   set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Xcompiler  /Zc:__cplusplus")
 
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /Zc:lambda")
+  set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Xcompiler /Zc:lambda")
+
   set(CMAKE_NINJA_CMCLDEPS_RC OFF)
   if(MSVC_Z7_OVERRIDE)
     # CMake set debug flags to use /Z7
diff --git a/aten/CMakeLists.txt b/aten/CMakeLists.txt
index d787d0850ab67..11d3ea4acb8e7 100644
--- a/aten/CMakeLists.txt
+++ b/aten/CMakeLists.txt
@@ -50,7 +50,6 @@ set(ATen_PUBLIC_HIP_DEPENDENCY_LIBS)
 set(ATEN_INSTALL_BIN_SUBDIR "bin" CACHE PATH "ATen install binary subdirectory")
 set(ATEN_INSTALL_LIB_SUBDIR "lib" CACHE PATH "ATen install library subdirectory")
 set(ATEN_INSTALL_INCLUDE_SUBDIR "include" CACHE PATH "ATen install include subdirectory")
-set(MEM_EFF_ATTENTION_CUDA_SOURCES)
 
 set(TH_LINK_STYLE STATIC)
 set(TH_CPU_INCLUDE
@@ -149,5 +148,4 @@ set(ATen_CUDA_DEPENDENCY_LIBS ${ATen_CUDA_DEPENDENCY_LIBS} PARENT_SCOPE)
 set(ATen_HIP_DEPENDENCY_LIBS ${ATen_HIP_DEPENDENCY_LIBS} PARENT_SCOPE)
 set(ATen_CORE_TEST_SRCS ${ATen_CORE_TEST_SRCS} PARENT_SCOPE)
 set(FLASH_ATTENTION_CUDA_SOURCES ${FLASH_ATTENTION_CUDA_SOURCES} PARENT_SCOPE)
-set(MEM_EFF_ATTENTION_CUDA_SOURCES ${MEM_EFF_ATTENTION_CUDA_SOURCES} PARENT_SCOPE)
 set(ATen_ATTENTION_KERNEL_SRCS ${ATen_ATTENTION_KERNEL_SRCS} PARENT_SCOPE)
diff --git a/aten/src/ATen/CMakeLists.txt b/aten/src/ATen/CMakeLists.txt
index 0d90f8a34e404..b5e3314255f92 100644
--- a/aten/src/ATen/CMakeLists.txt
+++ b/aten/src/ATen/CMakeLists.txt
@@ -248,7 +248,6 @@ if(USE_MEM_EFF_ATTENTION)
   list(APPEND native_transformers_cuda_cu ${mem_eff_attention_cuda_cu})
   list(APPEND native_transformers_cuda_cu ${mem_eff_attention_cuda_kernels_cu})
   list(APPEND native_transformers_cuda_cpp ${mem_eff_attention_cuda_cpp})
-  list(APPEND MEM_EFF_ATTENTION_CUDA_SOURCES ${native_transformers_cuda_cu} ${mem_eff_attention_cuda_cu} ${mem_eff_attention_cuda_kernels_cu})
   list(APPEND ATen_ATTENTION_KERNEL_SRCS ${mem_eff_attention_cuda_kernels_cu})
 endif()
 
@@ -914,5 +913,4 @@ set(ATen_CUDA_DEPENDENCY_LIBS ${ATen_CUDA_DEPENDENCY_LIBS} PARENT_SCOPE)
 set(ATen_XPU_DEPENDENCY_LIBS ${ATen_XPU_DEPENDENCY_LIBS} PARENT_SCOPE)
 set(ATen_HIP_DEPENDENCY_LIBS ${ATen_HIP_DEPENDENCY_LIBS} PARENT_SCOPE)
 set(FLASH_ATTENTION_CUDA_SOURCES ${FLASH_ATTENTION_CUDA_SOURCES} PARENT_SCOPE)
-set(MEM_EFF_ATTENTION_CUDA_SOURCES ${MEM_EFF_ATTENTION_CUDA_SOURCES} PARENT_SCOPE)
 set(ATen_ATTENTION_KERNEL_SRCS ${ATen_ATTENTION_KERNEL_SRCS} PARENT_SCOPE)
diff --git a/aten/src/ATen/native/cpu/ReduceOpsKernel.cpp b/aten/src/ATen/native/cpu/ReduceOpsKernel.cpp
index c06731dfc718c..d5d1f7b692b2d 100644
--- a/aten/src/ATen/native/cpu/ReduceOpsKernel.cpp
+++ b/aten/src/ATen/native/cpu/ReduceOpsKernel.cpp
@@ -54,8 +54,8 @@ static inline void cpu_cum_base_kernel(const Tensor& result,
     .add_const_input(self)
     .build();
 
-  auto result_dim_stride = ensure_nonempty_stride(result, dim);
-  auto self_dim_stride = ensure_nonempty_stride(self, dim);
+  const int64_t result_dim_stride = ensure_nonempty_stride(result, dim);
+  const int64_t self_dim_stride = ensure_nonempty_stride(self, dim);
 
   auto loop = [&](char** data, const int64_t* strides, int64_t n) {
     auto* result_data_bytes = data[0];
@@ -82,8 +82,8 @@ static void cumsum_cpu_kernel(const Tensor& result, const Tensor& self, int64_t
 
   AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND2(kBFloat16, kHalf, self.scalar_type(), "cumsum_out_cpu", [&] {
     cpu_cum_base_kernel<scalar_t>(result, self, wrap_dim, [&] (
-      scalar_t* result_data, auto result_dim_stride,
-      const scalar_t* self_data, auto self_dim_stride, scalar_t init_val) {
+      scalar_t* result_data, int64_t result_dim_stride,
+      const scalar_t* self_data, int64_t self_dim_stride, scalar_t init_val) {
         // NOLINTNEXTLINE(bugprone-signed-char-misuse)
         auto cum_number = (at::acc_type<scalar_t, false>)init_val;
         for (const auto i : c10::irange(self_dim_size)) {
@@ -101,8 +101,8 @@ static void cumprod_cpu_kernel(const Tensor& result, const Tensor& self, int64_t
 
   AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND2(kBFloat16, kHalf, self.scalar_type(), "cumprod_out_cpu", [&] {
     cpu_cum_base_kernel<scalar_t>(result, self, wrap_dim, [&] (
-      scalar_t* result_data, auto result_dim_stride,
-      const scalar_t* self_data, auto self_dim_stride, scalar_t init_val) {
+      scalar_t* result_data, int64_t result_dim_stride,
+      const scalar_t* self_data, int64_t self_dim_stride, scalar_t init_val) {
         // NOLINTNEXTLINE(bugprone-signed-char-misuse)
         auto cum_number = (at::acc_type<scalar_t, false>)init_val;
         for (const auto i : c10::irange(self_dim_size)) {
@@ -120,8 +120,8 @@ static void logcumsumexp_cpu_kernel(Tensor& result, const Tensor& self, int64_t
 
   AT_DISPATCH_FLOATING_AND_COMPLEX_TYPES_AND2(kBFloat16, kHalf, self.scalar_type(), "logcumsumexp_out_cpu", [&] {
     cpu_cum_base_kernel<scalar_t>(result, self, wrap_dim, [&] (
-      scalar_t* result_data, auto result_dim_stride,
-      const scalar_t* self_data, auto self_dim_stride, scalar_t init_val) {
+      scalar_t* result_data, int64_t result_dim_stride,
+      const scalar_t* self_data, int64_t self_dim_stride, scalar_t init_val) {
         using accscalar_t = at::acc_type<scalar_t, false>;
         auto cum_number = (accscalar_t)init_val;
         for (const auto i : c10::irange(self_dim_size)) {
diff --git a/aten/src/ATen/native/cpu/TensorCompareKernel.cpp b/aten/src/ATen/native/cpu/TensorCompareKernel.cpp
index 2c52a61fc553a..eb8e8a7c85505 100644
--- a/aten/src/ATen/native/cpu/TensorCompareKernel.cpp
+++ b/aten/src/ATen/native/cpu/TensorCompareKernel.cpp
@@ -77,7 +77,7 @@ static inline void compare_base_kernel(const Tensor& result1, const Tensor& resu
     bool keepdim,
     const func_t& f) {
 
-  auto self_dim_stride = ensure_nonempty_stride(self, dim);
+  const int64_t self_dim_stride = ensure_nonempty_stride(self, dim);
 
   auto loop = [&](char** data, const int64_t* strides, int64_t n) {
     auto* result1_data_bytes = data[0];
@@ -104,12 +104,12 @@ static void min_kernel_impl(
     const Tensor& self,
     int64_t dim,
     bool keepdim) {
-  int64_t self_dim_size = ensure_nonempty_size(self, dim);
+  const int64_t self_dim_size = ensure_nonempty_size(self, dim);
 
   AT_DISPATCH_ALL_TYPES_AND3(ScalarType::Half, ScalarType::BFloat16, ScalarType::Bool, self.scalar_type(), "min_cpu", [&] {
     compare_base_kernel<scalar_t>(result, indice, self, dim, keepdim, [&] (
       scalar_t* result_data, int64_t* indice_data,
-      const scalar_t* self_data, auto self_dim_stride) {
+      const scalar_t* self_data, int64_t self_dim_stride) {
         using value_t = typename c10::scalar_value_type<scalar_t>::type;
         value_t (*zabs_)(scalar_t) = zabs<scalar_t, value_t>;
         scalar_t min_number = c10::load(self_data);
@@ -137,12 +137,12 @@ static void max_kernel_impl(
     const Tensor& self,
     int64_t dim,
     bool keepdim) {
-  int64_t self_dim_size = ensure_nonempty_size(self, dim);
+  const int64_t self_dim_size = ensure_nonempty_size(self, dim);
 
   AT_DISPATCH_ALL_TYPES_AND3(ScalarType::Half, ScalarType::BFloat16, ScalarType::Bool, self.scalar_type(), "max_cpu", [&] {
     compare_base_kernel<scalar_t>(result, indice, self, dim, keepdim, [&] (
       scalar_t* result_data, int64_t* indice_data,
-      const scalar_t* self_data, auto self_dim_stride) {
+      const scalar_t* self_data, int64_t self_dim_stride) {
         using value_t = typename c10::scalar_value_type<scalar_t>::type;
         value_t (*zabs_)(scalar_t) = zabs<scalar_t, value_t>;
         scalar_t max_number = c10::load(self_data);
@@ -171,7 +171,7 @@ static void aminmax_kernel(
     Tensor& min_result,
     Tensor& max_result) {
   auto wrap_dim = maybe_wrap_dim(dim, self.dim());
-  int64_t self_dim_size = ensure_nonempty_size(self, wrap_dim);
+  const int64_t self_dim_size = ensure_nonempty_size(self, wrap_dim);
 
   TORCH_CHECK(min_result.scalar_type() == self.scalar_type() && max_result.scalar_type() == self.scalar_type(),
     "Expect min and max dtype ", self.scalar_type(),
@@ -189,7 +189,7 @@ static void aminmax_kernel(
   AT_DISPATCH_ALL_TYPES_AND3(ScalarType::Bool, ScalarType::BFloat16, ScalarType::Half, self.scalar_type(), "aminmax_cpu", [&] {
     compare_base_kernel<scalar_t, scalar_t>(min_result, max_result, self, wrap_dim, keepdim, [&] (
       scalar_t* min_result_data, scalar_t* max_result_data,
-      const scalar_t* self_data, auto self_dim_stride) {
+      const scalar_t* self_data, int64_t self_dim_stride) {
         scalar_t min_number = c10::load(self_data);
         scalar_t max_number = min_number;
         for (const auto i : c10::irange(self_dim_size)) {
diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
index 10191edee9cd6..0689b74995007 100644
--- a/caffe2/CMakeLists.txt
+++ b/caffe2/CMakeLists.txt
@@ -2005,18 +2005,4 @@ if(BUILD_TEST)
     endforeach()
   endif()
 endif()
-
-if(MSVC)
-  # This is used to enable the conforming lambda processor in MSVC
-  # Which allows us to capture constexpr in lambdas
-  # Note that this will be turned on by default for std=c++20 and above
-  # This should be applied globally when https://github.com/pytorch/pytorch/issues/92600 is fixed
-  foreach(tmp ${MEM_EFF_ATTENTION_CUDA_SOURCES})
-    # MEM_EFF_ATTENTION_CUDA is populated in pytorch/aten/src/ATen/CMakeLists.txt
-    # We iterate over these files, updating paths and adding the compile flag
-    FILE(RELATIVE_PATH tmp_path "${PROJECT_SOURCE_DIR}" "${tmp}")
-    SET(tmp_path "../${tmp_path}")
-    set_source_files_properties(${tmp_path} PROPERTIES COMPILE_FLAGS "-Xcompiler /Zc:lambda")
-  endforeach()
-endif()
 endif()