From 846129f2ea355037dc9b13b4f1ac49ac0630dd28 Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Wed, 15 Oct 2025 21:18:56 -0700
Subject: [PATCH 01/21] platform layer for windows and linux compatibility

---
 backends/cuda/runtime/cuda_backend.cpp    | 39 +++++++-----
 backends/cuda/runtime/platform/platform.h | 76 +++++++++++++++++++++++
 2 files changed, 98 insertions(+), 17 deletions(-)
 create mode 100644 backends/cuda/runtime/platform/platform.h

diff --git a/backends/cuda/runtime/cuda_backend.cpp b/backends/cuda/runtime/cuda_backend.cpp
index 3fcd25a3d1d..5c09973ac5f 100644
--- a/backends/cuda/runtime/cuda_backend.cpp
+++ b/backends/cuda/runtime/cuda_backend.cpp
@@ -7,7 +7,6 @@
  */
 
 #include <cuda_runtime.h>
-#include <dlfcn.h>
 #include <executorch/runtime/backend/interface.h>
 #include <executorch/runtime/core/error.h>
 #include <executorch/runtime/core/evalue.h>
@@ -23,17 +22,23 @@
 // Include our shim layer headers
 #include <executorch/backends/aoti/aoti_delegate_handle.h>
 #include <executorch/backends/aoti/common_shims.h>
+#include <executorch/backends/cuda/runtime/platform/platform.h>
 #include <executorch/backends/cuda/runtime/shims/memory.h>
 #include <executorch/backends/cuda/runtime/utils.h>
 
 namespace executorch::backends::cuda {
 
-#define LOAD_SYMBOL(handle, member, name, so_handle)                        \
-  do {                                                                      \
-    handle->member = reinterpret_cast<name##Func>(dlsym(so_handle, #name)); \
-    ET_CHECK_OR_RETURN_ERROR(                                               \
-        handle->member != nullptr, AccessFailed, "Failed to load " #name);  \
-  } while (0)
+#define LOAD_SYMBOL(handle, member, name, so_handle)                 \
+  \ 
+  do {                                                               \
+    auto symbol_res = get_function(so_handle, #name);                \
+    \  
+    if (!symbol_res.ok()) {                                          \
+      return symbol_res.error();                                     \
+    }                                                                \
+    handle->member = reinterpret_cast<name##Func>(symbol_res.get()); \
+  }                                                                  \
+  while (0)
 
 using namespace std;
 using namespace aoti;
@@ -144,24 +149,23 @@ class ET_EXPERIMENTAL CudaBackend final
     // Finish writing the file to disk
     outfile.close();
 
-    // Load the ELF using dlopen
-    void* so_handle = dlopen(so_path.c_str(), RTLD_LAZY | RTLD_LOCAL);
-    ET_CHECK_OR_RETURN_ERROR(
-        so_handle != nullptr,
-        AccessFailed,
-        "Failed to load shared library: %s",
-        dlerror());
+    // Load the lib
+    Result<void*> lib_handle_res = load_library(so_path);
+    if (!lib_handle_res.ok()) {
+      return lib_handle_res.error();
+    }
+    void* lib_handle = lib_handle_res.get();
 
     processed->Free();
 
     // Create handle and load function pointers into it
     AOTIDelegateHandle* handle = new AOTIDelegateHandle();
-    handle->so_handle = so_handle;
+    handle->so_handle = lib_handle;
     handle->so_path = so_path.string();
 
     // Load function pointers specific to this handle's shared library
     ET_CHECK_OK_OR_RETURN_ERROR(
-        load_function_pointers_into_handle(so_handle, handle));
+        load_function_pointers_into_handle(lib_handle, handle));
 
     AOTInductorModelContainerHandle container_handle = nullptr;
 
@@ -332,8 +336,9 @@ class ET_EXPERIMENTAL CudaBackend final
     // AOTInductorModelContainerDelete(handle->container_handle);
 
     // Now close the shared library
+    auto err = Error::Ok;
     if (handle->so_handle != nullptr) {
-      dlclose(handle->so_handle);
+      err = close_library(handle->so_handle);
     }
 
     // Remove the temporary shared library file
diff --git a/backends/cuda/runtime/platform/platform.h b/backends/cuda/runtime/platform/platform.h
new file mode 100644
index 00000000000..526b46cb261
--- /dev/null
+++ b/backends/cuda/runtime/platform/platform.h
@@ -0,0 +1,76 @@
+
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <string>
+#include <executorch/runtime/core/error.h>
+
+#ifdef _WIN32
+#include <windows.h>
+#else
+#include <dlfcn.h>
+#endif
+
+namespace executorch {
+namespace backends {
+namespace cuda {
+
+Result<void*> load_library(const std::string& path) {
+#ifdef _WIN32
+  auto lib_handle = LoadLibrary(path.c_str());
+  if (hModule == NULL) {
+      ET_LOG(Error, "Failed to load %s with error: %lu", path.c_str(), GetLastError());
+      return Error::AccessFailed;
+  }
+
+#else
+  void* lib_handle = dlopen(path.c_str(), RTLD_LAZY | RTLD_LOCAL);
+  if (so_handle == nullptr) {
+    ET_LOG(Error, "Failed to load %s with error: %s", path.c_str(), dlerror());
+    return Error::AccessFailed;
+  }
+#endif
+  return (void*) lib_handle;
+}
+
+Error close_library(void* lib_handle) {
+#ifdef _WIN32
+    if (!FreeLibrary((HModule)lib_handle)) {
+        printf("FreeLibrary failed with error %lu\n", GetLastError());
+        return Error::Internal
+    } 
+#else
+  if (dlclose(lib_handle) != 0) {
+      ET_LOG(Error, "dlclose failed: %s\n", dlerror());
+      return Error::Internal;
+  }
+#endif
+  return Error::Ok;
+}
+
+Result<void*> get_function(void* lib_handle, const std::string& fn_name) {
+#ifdef _WIN32
+  auto fn = GetProcAddress((HModule)lib_handle, fn_name.c_str());
+  if (!fn) {
+      ET_LOG(Error, "Failed loading symbol %s with error %lu\n", fn_name.c_str(), GetLastError());
+      return Error::Internal;
+  }
+#else
+  auto fn = dlsym(lib_handle, fn_name.c_str());
+  if (fn == nullptr) {
+      ET_LOG(Error, "Failed loading symbol %s with error %s\n", fn_name.c_str(), dlerror());
+      return Error::Internal;
+  }
+#endif
+
+  return (void*)fn; // This I think is technically ub on windows. We should probably explicitly pack the bytes.
+}
+
+} // cuda
+} // backends
+} // executorch

From f3568971cc5f70502257bf63331e80be7a68343f Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Wed, 15 Oct 2025 21:21:54 -0700
Subject: [PATCH 02/21] lint

---
 backends/cuda/runtime/cuda_backend.cpp | 2 --
 1 file changed, 2 deletions(-)

diff --git a/backends/cuda/runtime/cuda_backend.cpp b/backends/cuda/runtime/cuda_backend.cpp
index 5c09973ac5f..d4788072094 100644
--- a/backends/cuda/runtime/cuda_backend.cpp
+++ b/backends/cuda/runtime/cuda_backend.cpp
@@ -29,10 +29,8 @@
 namespace executorch::backends::cuda {
 
 #define LOAD_SYMBOL(handle, member, name, so_handle)                 \
-  \ 
   do {                                                               \
     auto symbol_res = get_function(so_handle, #name);                \
-    \  
     if (!symbol_res.ok()) {                                          \
       return symbol_res.error();                                     \
     }                                                                \

From 3e9c66b9b57341add183904bcce1e51b5b294fa4 Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Wed, 15 Oct 2025 21:32:11 -0700
Subject: [PATCH 03/21] abstract process id

---
 backends/cuda/runtime/cuda_backend.cpp    |  6 +--
 backends/cuda/runtime/platform/platform.h | 60 ++++++++++++++++-------
 2 files changed, 44 insertions(+), 22 deletions(-)

diff --git a/backends/cuda/runtime/cuda_backend.cpp b/backends/cuda/runtime/cuda_backend.cpp
index d4788072094..c391297629d 100644
--- a/backends/cuda/runtime/cuda_backend.cpp
+++ b/backends/cuda/runtime/cuda_backend.cpp
@@ -11,7 +11,6 @@
 #include <executorch/runtime/core/error.h>
 #include <executorch/runtime/core/evalue.h>
 #include <executorch/runtime/core/exec_aten/util/tensor_util.h>
-#include <unistd.h>
 #include <cstdio>
 
 #include <filesystem>
@@ -35,8 +34,7 @@ namespace executorch::backends::cuda {
       return symbol_res.error();                                     \
     }                                                                \
     handle->member = reinterpret_cast<name##Func>(symbol_res.get()); \
-  }                                                                  \
-  while (0)
+  } while (0)
 
 using namespace std;
 using namespace aoti;
@@ -125,7 +123,7 @@ class ET_EXPERIMENTAL CudaBackend final
     // Generate dynamic temporary file path
     filesystem::path temp_dir = filesystem::temp_directory_path();
     filesystem::path so_path =
-        temp_dir / (so_blob_key + to_string(getpid()) + ".so");
+        temp_dir / (so_blob_key + to_string(get_process_id()) + ".so");
 
     // Create a temporary file
     ofstream outfile(so_path.c_str(), ios::binary);
diff --git a/backends/cuda/runtime/platform/platform.h b/backends/cuda/runtime/platform/platform.h
index 526b46cb261..aadd6d7173d 100644
--- a/backends/cuda/runtime/platform/platform.h
+++ b/backends/cuda/runtime/platform/platform.h
@@ -7,13 +7,14 @@
  * LICENSE file in the root directory of this source tree.
  */
 
-#include <string>
 #include <executorch/runtime/core/error.h>
+#include <string>
 
 #ifdef _WIN32
 #include <windows.h>
 #else
 #include <dlfcn.h>
+#include <unistd.h>
 #endif
 
 namespace executorch {
@@ -23,31 +24,35 @@ namespace cuda {
 Result<void*> load_library(const std::string& path) {
 #ifdef _WIN32
   auto lib_handle = LoadLibrary(path.c_str());
-  if (hModule == NULL) {
-      ET_LOG(Error, "Failed to load %s with error: %lu", path.c_str(), GetLastError());
-      return Error::AccessFailed;
+  if (lib_handle == NULL) {
+    ET_LOG(
+        Error,
+        "Failed to load %s with error: %lu",
+        path.c_str(),
+        GetLastError());
+    return Error::AccessFailed;
   }
 
 #else
   void* lib_handle = dlopen(path.c_str(), RTLD_LAZY | RTLD_LOCAL);
-  if (so_handle == nullptr) {
+  if (lib_handle == nullptr) {
     ET_LOG(Error, "Failed to load %s with error: %s", path.c_str(), dlerror());
     return Error::AccessFailed;
   }
 #endif
-  return (void*) lib_handle;
+  return (void*)lib_handle;
 }
 
 Error close_library(void* lib_handle) {
 #ifdef _WIN32
-    if (!FreeLibrary((HModule)lib_handle)) {
-        printf("FreeLibrary failed with error %lu\n", GetLastError());
-        return Error::Internal
-    } 
+  if (!FreeLibrary((HMODULE)lib_handle)) {
+    printf("FreeLibrary failed with error %lu\n", GetLastError());
+    return Error::Internal;
+  }
 #else
   if (dlclose(lib_handle) != 0) {
-      ET_LOG(Error, "dlclose failed: %s\n", dlerror());
-      return Error::Internal;
+    ET_LOG(Error, "dlclose failed: %s\n", dlerror());
+    return Error::Internal;
   }
 #endif
   return Error::Ok;
@@ -55,20 +60,39 @@ Error close_library(void* lib_handle) {
 
 Result<void*> get_function(void* lib_handle, const std::string& fn_name) {
 #ifdef _WIN32
-  auto fn = GetProcAddress((HModule)lib_handle, fn_name.c_str());
+  auto fn = GetProcAddress((HMODULE)lib_handle, fn_name.c_str());
   if (!fn) {
-      ET_LOG(Error, "Failed loading symbol %s with error %lu\n", fn_name.c_str(), GetLastError());
-      return Error::Internal;
+    ET_LOG(
+        Error,
+        "Failed loading symbol %s with error %lu\n",
+        fn_name.c_str(),
+        GetLastError());
+    return Error::Internal;
   }
 #else
   auto fn = dlsym(lib_handle, fn_name.c_str());
   if (fn == nullptr) {
-      ET_LOG(Error, "Failed loading symbol %s with error %s\n", fn_name.c_str(), dlerror());
-      return Error::Internal;
+    ET_LOG(
+        Error,
+        "Failed loading symbol %s with error %s\n",
+        fn_name.c_str(),
+        dlerror());
+    return Error::Internal;
   }
 #endif
 
-  return (void*)fn; // This I think is technically ub on windows. We should probably explicitly pack the bytes.
+  return (void*)fn; // This I think is technically ub on windows. We should
+                    // probably explicitly pack the bytes.
+}
+
+int32_t get_process_id(void* lib_handle) {
+#ifdef _WIN32
+  return GetCurrentProcessId();
+#else
+  return getpid();
+#endif
+}
+
 }
 
 } // cuda

From d39a36eaa4157d76274f77004a10fca978746805 Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Wed, 15 Oct 2025 22:07:12 -0700
Subject: [PATCH 04/21] missing import

---
 backends/aoti/utils.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/backends/aoti/utils.h b/backends/aoti/utils.h
index b24fcaac864..80abe663fa2 100644
--- a/backends/aoti/utils.h
+++ b/backends/aoti/utils.h
@@ -15,6 +15,7 @@
 #include <executorch/runtime/platform/log.h>
 #include <cstddef>
 #include <cstdint>
+#include <vector>
 
 namespace executorch {
 namespace backends {

From caf8495c337f06b2b614d3ea9c2810f4a97bfeaf Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Wed, 15 Oct 2025 22:12:36 -0700
Subject: [PATCH 05/21] missing import and namespacing

---
 backends/cuda/runtime/platform/platform.h | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/backends/cuda/runtime/platform/platform.h b/backends/cuda/runtime/platform/platform.h
index aadd6d7173d..9391c198535 100644
--- a/backends/cuda/runtime/platform/platform.h
+++ b/backends/cuda/runtime/platform/platform.h
@@ -8,6 +8,7 @@
  */
 
 #include <executorch/runtime/core/error.h>
+#include <executorch/runtime/core/result.h>
 #include <string>
 
 #ifdef _WIN32
@@ -21,7 +22,7 @@ namespace executorch {
 namespace backends {
 namespace cuda {
 
-Result<void*> load_library(const std::string& path) {
+executorch::runtime::Result<void*> load_library(const std::string& path) {
 #ifdef _WIN32
   auto lib_handle = LoadLibrary(path.c_str());
   if (lib_handle == NULL) {
@@ -43,7 +44,7 @@ Result<void*> load_library(const std::string& path) {
   return (void*)lib_handle;
 }
 
-Error close_library(void* lib_handle) {
+executorch::runtime::Error close_library(void* lib_handle) {
 #ifdef _WIN32
   if (!FreeLibrary((HMODULE)lib_handle)) {
     printf("FreeLibrary failed with error %lu\n", GetLastError());
@@ -58,7 +59,7 @@ Error close_library(void* lib_handle) {
   return Error::Ok;
 }
 
-Result<void*> get_function(void* lib_handle, const std::string& fn_name) {
+executorch::runtime::Result<void*> get_function(void* lib_handle, const std::string& fn_name) {
 #ifdef _WIN32
   auto fn = GetProcAddress((HMODULE)lib_handle, fn_name.c_str());
   if (!fn) {

From 2fa126aaf050f4ae62c2fc732185e2e16a315f45 Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Wed, 15 Oct 2025 22:15:21 -0700
Subject: [PATCH 06/21] more namespace errors

---
 backends/cuda/runtime/platform/platform.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/backends/cuda/runtime/platform/platform.h b/backends/cuda/runtime/platform/platform.h
index 9391c198535..898110cf84f 100644
--- a/backends/cuda/runtime/platform/platform.h
+++ b/backends/cuda/runtime/platform/platform.h
@@ -38,7 +38,7 @@ executorch::runtime::Result<void*> load_library(const std::string& path) {
   void* lib_handle = dlopen(path.c_str(), RTLD_LAZY | RTLD_LOCAL);
   if (lib_handle == nullptr) {
     ET_LOG(Error, "Failed to load %s with error: %s", path.c_str(), dlerror());
-    return Error::AccessFailed;
+    return executorch::runtime::Error::AccessFailed;
   }
 #endif
   return (void*)lib_handle;
@@ -48,7 +48,7 @@ executorch::runtime::Error close_library(void* lib_handle) {
 #ifdef _WIN32
   if (!FreeLibrary((HMODULE)lib_handle)) {
     printf("FreeLibrary failed with error %lu\n", GetLastError());
-    return Error::Internal;
+    return executorch::runtime::Error::Internal;
   }
 #else
   if (dlclose(lib_handle) != 0) {
@@ -56,7 +56,7 @@ executorch::runtime::Error close_library(void* lib_handle) {
     return Error::Internal;
   }
 #endif
-  return Error::Ok;
+  return executorch::runtime::Error::Ok;
 }
 
 executorch::runtime::Result<void*> get_function(void* lib_handle, const std::string& fn_name) {
@@ -68,7 +68,7 @@ executorch::runtime::Result<void*> get_function(void* lib_handle, const std::str
         "Failed loading symbol %s with error %lu\n",
         fn_name.c_str(),
         GetLastError());
-    return Error::Internal;
+    return executorch::runtime::Error::Internal;
   }
 #else
   auto fn = dlsym(lib_handle, fn_name.c_str());

From 36b67105f16258207a9a6270c8db58607ff47f5c Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Wed, 15 Oct 2025 22:16:14 -0700
Subject: [PATCH 07/21] namespacing hell

---
 backends/cuda/runtime/platform/platform.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/backends/cuda/runtime/platform/platform.h b/backends/cuda/runtime/platform/platform.h
index 898110cf84f..355c6fcd7eb 100644
--- a/backends/cuda/runtime/platform/platform.h
+++ b/backends/cuda/runtime/platform/platform.h
@@ -31,7 +31,7 @@ executorch::runtime::Result<void*> load_library(const std::string& path) {
         "Failed to load %s with error: %lu",
         path.c_str(),
         GetLastError());
-    return Error::AccessFailed;
+    return executorch::runtime::Error::AccessFailed;
   }
 
 #else

From 4c0215c1f0d0b1ca15d918e348305376c003132a Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Wed, 15 Oct 2025 22:18:19 -0700
Subject: [PATCH 08/21] erroneous bracket

---
 backends/cuda/runtime/platform/platform.h | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/backends/cuda/runtime/platform/platform.h b/backends/cuda/runtime/platform/platform.h
index 355c6fcd7eb..a9d574cc365 100644
--- a/backends/cuda/runtime/platform/platform.h
+++ b/backends/cuda/runtime/platform/platform.h
@@ -13,7 +13,7 @@
 
 #ifdef _WIN32
 #include <windows.h>
-#else
+#else // Posix
 #include <dlfcn.h>
 #include <unistd.h>
 #endif
@@ -94,8 +94,6 @@ int32_t get_process_id(void* lib_handle) {
 #endif
 }
 
-}
-
 } // cuda
 } // backends
 } // executorch

From a685328b3d3ddceb49ca004f61891072890d71e9 Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Wed, 15 Oct 2025 22:25:51 -0700
Subject: [PATCH 09/21] no std::string for path

---
 backends/cuda/runtime/platform/platform.h | 22 ++++++++++++----------
 1 file changed, 12 insertions(+), 10 deletions(-)

diff --git a/backends/cuda/runtime/platform/platform.h b/backends/cuda/runtime/platform/platform.h
index a9d574cc365..a93db08dcee 100644
--- a/backends/cuda/runtime/platform/platform.h
+++ b/backends/cuda/runtime/platform/platform.h
@@ -22,22 +22,22 @@ namespace executorch {
 namespace backends {
 namespace cuda {
 
-executorch::runtime::Result<void*> load_library(const std::string& path) {
+executorch::runtime::Result<void*> load_library(const char* path) {
 #ifdef _WIN32
-  auto lib_handle = LoadLibrary(path.c_str());
+  auto lib_handle = LoadLibrary(path);
   if (lib_handle == NULL) {
     ET_LOG(
         Error,
         "Failed to load %s with error: %lu",
-        path.c_str(),
+        path,
         GetLastError());
     return executorch::runtime::Error::AccessFailed;
   }
 
 #else
-  void* lib_handle = dlopen(path.c_str(), RTLD_LAZY | RTLD_LOCAL);
+  void* lib_handle = dlopen(path, RTLD_LAZY | RTLD_LOCAL);
   if (lib_handle == nullptr) {
-    ET_LOG(Error, "Failed to load %s with error: %s", path.c_str(), dlerror());
+    ET_LOG(Error, "Failed to load %s with error: %s", path, dlerror());
     return executorch::runtime::Error::AccessFailed;
   }
 #endif
@@ -59,7 +59,9 @@ executorch::runtime::Error close_library(void* lib_handle) {
   return executorch::runtime::Error::Ok;
 }
 
-executorch::runtime::Result<void*> get_function(void* lib_handle, const std::string& fn_name) {
+executorch::runtime::Result<void*> get_function(
+    void* lib_handle,
+    const std::string& fn_name) {
 #ifdef _WIN32
   auto fn = GetProcAddress((HMODULE)lib_handle, fn_name.c_str());
   if (!fn) {
@@ -86,7 +88,7 @@ executorch::runtime::Result<void*> get_function(void* lib_handle, const std::str
                     // probably explicitly pack the bytes.
 }
 
-int32_t get_process_id(void* lib_handle) {
+int32_t get_process_id() {
 #ifdef _WIN32
   return GetCurrentProcessId();
 #else
@@ -94,6 +96,6 @@ int32_t get_process_id(void* lib_handle) {
 #endif
 }
 
-} // cuda
-} // backends
-} // executorch
+} // namespace cuda
+} // namespace backends
+} // namespace executorch

From 2469b441e340358e9f1fd7118dbc80d49ddfed74 Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Wed, 15 Oct 2025 22:35:27 -0700
Subject: [PATCH 10/21] remove posix_memalign

---
 backends/cuda/runtime/cuda_backend.cpp    |  2 +-
 backends/cuda/runtime/platform/platform.h | 24 ++++++++++++++++++-----
 backends/cuda/runtime/shims/memory.cpp    | 11 +++--------
 3 files changed, 23 insertions(+), 14 deletions(-)

diff --git a/backends/cuda/runtime/cuda_backend.cpp b/backends/cuda/runtime/cuda_backend.cpp
index c391297629d..b677ef459fe 100644
--- a/backends/cuda/runtime/cuda_backend.cpp
+++ b/backends/cuda/runtime/cuda_backend.cpp
@@ -146,7 +146,7 @@ class ET_EXPERIMENTAL CudaBackend final
     outfile.close();
 
     // Load the lib
-    Result<void*> lib_handle_res = load_library(so_path);
+    Result<void*> lib_handle_res = load_library(so_path.c_str());
     if (!lib_handle_res.ok()) {
       return lib_handle_res.error();
     }
diff --git a/backends/cuda/runtime/platform/platform.h b/backends/cuda/runtime/platform/platform.h
index a93db08dcee..8e4890677a7 100644
--- a/backends/cuda/runtime/platform/platform.h
+++ b/backends/cuda/runtime/platform/platform.h
@@ -12,10 +12,12 @@
 #include <string>
 
 #ifdef _WIN32
+#include <malloc.h>
 #include <windows.h>
 #else // Posix
 #include <dlfcn.h>
 #include <unistd.h>
+#include <cstdlib>
 #endif
 
 namespace executorch {
@@ -26,11 +28,7 @@ executorch::runtime::Result<void*> load_library(const char* path) {
 #ifdef _WIN32
   auto lib_handle = LoadLibrary(path);
   if (lib_handle == NULL) {
-    ET_LOG(
-        Error,
-        "Failed to load %s with error: %lu",
-        path,
-        GetLastError());
+    ET_LOG(Error, "Failed to load %s with error: %lu", path, GetLastError());
     return executorch::runtime::Error::AccessFailed;
   }
 
@@ -96,6 +94,22 @@ int32_t get_process_id() {
 #endif
 }
 
+void* aligned_alloc(size_t alignment, size_t size) {
+#ifdef _WIN32
+  return _aligned_malloc(size, alignment);
+#else
+  return std::aligned_alloc(alignment, size);
+#endif
+}
+
+void aligned_free(void* ptr) {
+#ifdef _WIN32
+  _aligned_free(ptr);
+#else
+  std::free(ptr);
+#endif
+}
+
 } // namespace cuda
 } // namespace backends
 } // namespace executorch
diff --git a/backends/cuda/runtime/shims/memory.cpp b/backends/cuda/runtime/shims/memory.cpp
index fe8ccf07281..811db422475 100644
--- a/backends/cuda/runtime/shims/memory.cpp
+++ b/backends/cuda/runtime/shims/memory.cpp
@@ -13,7 +13,6 @@
 #include <executorch/backends/cuda/runtime/utils.h>
 #include <executorch/runtime/platform/log.h>
 #include <cstdint>
-#include <cstdlib> // For posix_memalign
 #include <memory>
 #include <unordered_map>
 #include <unordered_set>
@@ -230,15 +229,11 @@ AOTITorchError aoti_torch_empty_strided(
         cudaMallocAsync(&ptr, static_cast<size_t>(nbytes), cudaStreamDefault));
   } else if (device_type == static_cast<int32_t>(SupportedDevices::CPU)) {
     // Ensure 16-byte alignment for CPU memory to match CUDA requirements
-    int result = posix_memalign(&ptr, 16, nbytes);
-    ET_CHECK_OR_RETURN_ERROR(
-        result == 0,
-        MemoryAllocationFailed,
-        "Failed to allocate aligned CPU memory");
+    &ptr = aligned_alloc(16, nbytes);
     ET_CHECK_OR_RETURN_ERROR(
         ptr != nullptr,
         MemoryAllocationFailed,
-        "Failed to call posix_memalign");
+        "Failed to allocate aligned CPU memory");
   } else {
     ET_CHECK_OR_RETURN_ERROR(
         false,
@@ -339,7 +334,7 @@ AOTITorchError aoti_torch_delete_tensor_object(Tensor* tensor) {
                 Internal,
                 "Expected host memory but got managed!")
             // This is CPU memory - free immediately
-            free(data_ptr);
+            aligned_free(data_ptr);
             data_ptr = nullptr;
           }
 

From 8610ba36206855b6b12ac53be6ff8bc909137b1e Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Wed, 15 Oct 2025 22:44:35 -0700
Subject: [PATCH 11/21] more path iteration

---
 backends/cuda/runtime/cuda_backend.cpp    |  4 ++--
 backends/cuda/runtime/platform/platform.h | 10 ++++++----
 backends/cuda/runtime/shims/memory.cpp    |  1 +
 3 files changed, 9 insertions(+), 6 deletions(-)

diff --git a/backends/cuda/runtime/cuda_backend.cpp b/backends/cuda/runtime/cuda_backend.cpp
index b677ef459fe..e61b03ee8e6 100644
--- a/backends/cuda/runtime/cuda_backend.cpp
+++ b/backends/cuda/runtime/cuda_backend.cpp
@@ -126,7 +126,7 @@ class ET_EXPERIMENTAL CudaBackend final
         temp_dir / (so_blob_key + to_string(get_process_id()) + ".so");
 
     // Create a temporary file
-    ofstream outfile(so_path.c_str(), ios::binary);
+    ofstream outfile(so_path, ios::binary);
 
     // Write the ELF buffer to the temporary file
     ET_LOG(
@@ -146,7 +146,7 @@ class ET_EXPERIMENTAL CudaBackend final
     outfile.close();
 
     // Load the lib
-    Result<void*> lib_handle_res = load_library(so_path.c_str());
+    Result<void*> lib_handle_res = load_library(so_path);
     if (!lib_handle_res.ok()) {
       return lib_handle_res.error();
     }
diff --git a/backends/cuda/runtime/platform/platform.h b/backends/cuda/runtime/platform/platform.h
index 8e4890677a7..a5e9ce333a4 100644
--- a/backends/cuda/runtime/platform/platform.h
+++ b/backends/cuda/runtime/platform/platform.h
@@ -24,18 +24,20 @@ namespace executorch {
 namespace backends {
 namespace cuda {
 
-executorch::runtime::Result<void*> load_library(const char* path) {
+executorch::runtime::Result<void*> load_library(
+    const std::filesystem::path& path) {
 #ifdef _WIN32
-  auto lib_handle = LoadLibrary(path);
+  std::string utf8 = p.u8string();
+  auto lib_handle = LoadLibrary(utf8.c_str());
   if (lib_handle == NULL) {
-    ET_LOG(Error, "Failed to load %s with error: %lu", path, GetLastError());
+    ET_LOG(Error, "Failed to load %s with error: %lu", utf8, GetLastError());
     return executorch::runtime::Error::AccessFailed;
   }
 
 #else
   void* lib_handle = dlopen(path, RTLD_LAZY | RTLD_LOCAL);
   if (lib_handle == nullptr) {
-    ET_LOG(Error, "Failed to load %s with error: %s", path, dlerror());
+    ET_LOG(Error, "Failed to load %s with error: %s", path.c_str(), dlerror());
     return executorch::runtime::Error::AccessFailed;
   }
 #endif
diff --git a/backends/cuda/runtime/shims/memory.cpp b/backends/cuda/runtime/shims/memory.cpp
index 811db422475..5fbb9e92f67 100644
--- a/backends/cuda/runtime/shims/memory.cpp
+++ b/backends/cuda/runtime/shims/memory.cpp
@@ -8,6 +8,7 @@
 
 #include <executorch/backends/aoti/common_shims.h>
 #include <executorch/backends/aoti/utils.h>
+#include <executorch/backends/cuda/runtime/platform/platform.h>
 #include <executorch/backends/cuda/runtime/shims/memory.h>
 #include <executorch/backends/cuda/runtime/shims/tensor_attribute.h>
 #include <executorch/backends/cuda/runtime/utils.h>

From c30c33b3c486d6047dcf2ab509189c562a2778fc Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Wed, 15 Oct 2025 22:50:07 -0700
Subject: [PATCH 12/21] missing include

---
 backends/cuda/runtime/platform/platform.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/backends/cuda/runtime/platform/platform.h b/backends/cuda/runtime/platform/platform.h
index a5e9ce333a4..fc01beecb34 100644
--- a/backends/cuda/runtime/platform/platform.h
+++ b/backends/cuda/runtime/platform/platform.h
@@ -10,6 +10,7 @@
 #include <executorch/runtime/core/error.h>
 #include <executorch/runtime/core/result.h>
 #include <string>
+#include <filesystem>
 
 #ifdef _WIN32
 #include <malloc.h>

From 917361a1ebf8588b2fb4d52772b2e26f9bab22a2 Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Wed, 15 Oct 2025 22:51:16 -0700
Subject: [PATCH 13/21] drop reference

---
 backends/cuda/runtime/platform/platform.h | 2 +-
 backends/cuda/runtime/shims/memory.cpp    | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/backends/cuda/runtime/platform/platform.h b/backends/cuda/runtime/platform/platform.h
index fc01beecb34..9c05afa5aac 100644
--- a/backends/cuda/runtime/platform/platform.h
+++ b/backends/cuda/runtime/platform/platform.h
@@ -28,7 +28,7 @@ namespace cuda {
 executorch::runtime::Result<void*> load_library(
     const std::filesystem::path& path) {
 #ifdef _WIN32
-  std::string utf8 = p.u8string();
+  std::string utf8 = path.u8string();
   auto lib_handle = LoadLibrary(utf8.c_str());
   if (lib_handle == NULL) {
     ET_LOG(Error, "Failed to load %s with error: %lu", utf8, GetLastError());
diff --git a/backends/cuda/runtime/shims/memory.cpp b/backends/cuda/runtime/shims/memory.cpp
index 5fbb9e92f67..5d30d3124d9 100644
--- a/backends/cuda/runtime/shims/memory.cpp
+++ b/backends/cuda/runtime/shims/memory.cpp
@@ -230,7 +230,7 @@ AOTITorchError aoti_torch_empty_strided(
         cudaMallocAsync(&ptr, static_cast<size_t>(nbytes), cudaStreamDefault));
   } else if (device_type == static_cast<int32_t>(SupportedDevices::CPU)) {
     // Ensure 16-byte alignment for CPU memory to match CUDA requirements
-    &ptr = aligned_alloc(16, nbytes);
+    ptr = aligned_alloc(16, nbytes);
     ET_CHECK_OR_RETURN_ERROR(
         ptr != nullptr,
         MemoryAllocationFailed,

From 45039a469b1dfbfe61abafeb035525f358c598bd Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Wed, 15 Oct 2025 23:24:05 -0700
Subject: [PATCH 14/21] lint

---
 backends/cuda/runtime/platform/platform.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/backends/cuda/runtime/platform/platform.h b/backends/cuda/runtime/platform/platform.h
index 9c05afa5aac..e29bed36cf8 100644
--- a/backends/cuda/runtime/platform/platform.h
+++ b/backends/cuda/runtime/platform/platform.h
@@ -9,8 +9,8 @@
 
 #include <executorch/runtime/core/error.h>
 #include <executorch/runtime/core/result.h>
-#include <string>
 #include <filesystem>
+#include <string>
 
 #ifdef _WIN32
 #include <malloc.h>

From 25ce94e56f0f8ddc3d98ff79c9ae3757d0765e92 Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Thu, 16 Oct 2025 10:00:26 -0700
Subject: [PATCH 15/21] linux fix

---
 backends/cuda/runtime/platform/platform.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/backends/cuda/runtime/platform/platform.h b/backends/cuda/runtime/platform/platform.h
index e29bed36cf8..b479d69d4ef 100644
--- a/backends/cuda/runtime/platform/platform.h
+++ b/backends/cuda/runtime/platform/platform.h
@@ -36,7 +36,7 @@ executorch::runtime::Result<void*> load_library(
   }
 
 #else
-  void* lib_handle = dlopen(path, RTLD_LAZY | RTLD_LOCAL);
+  void* lib_handle = dlopen(path.c_str(), RTLD_LAZY | RTLD_LOCAL);
   if (lib_handle == nullptr) {
     ET_LOG(Error, "Failed to load %s with error: %s", path.c_str(), dlerror());
     return executorch::runtime::Error::AccessFailed;
@@ -54,7 +54,7 @@ executorch::runtime::Error close_library(void* lib_handle) {
 #else
   if (dlclose(lib_handle) != 0) {
     ET_LOG(Error, "dlclose failed: %s\n", dlerror());
-    return Error::Internal;
+    return executorch::runtime::Error::Internal;
   }
 #endif
   return executorch::runtime::Error::Ok;

From 7d701a36cf09474c6a44d9179aa05322ea574519 Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Thu, 16 Oct 2025 10:39:22 -0700
Subject: [PATCH 16/21] more linux fix

---
 backends/cuda/runtime/platform/platform.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/backends/cuda/runtime/platform/platform.h b/backends/cuda/runtime/platform/platform.h
index b479d69d4ef..5bb91a67879 100644
--- a/backends/cuda/runtime/platform/platform.h
+++ b/backends/cuda/runtime/platform/platform.h
@@ -81,7 +81,7 @@ executorch::runtime::Result<void*> get_function(
         "Failed loading symbol %s with error %s\n",
         fn_name.c_str(),
         dlerror());
-    return Error::Internal;
+    return executorch::runtime::Error::Internal;
   }
 #endif
 

From 7af462377893aa64f95bb97a80ddd30ab2c8b2b1 Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Thu, 16 Oct 2025 11:45:11 -0700
Subject: [PATCH 17/21] forgot pragma once

---
 backends/cuda/runtime/platform/platform.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/backends/cuda/runtime/platform/platform.h b/backends/cuda/runtime/platform/platform.h
index 5bb91a67879..e6edec8bfe6 100644
--- a/backends/cuda/runtime/platform/platform.h
+++ b/backends/cuda/runtime/platform/platform.h
@@ -7,6 +7,8 @@
  * LICENSE file in the root directory of this source tree.
  */
 
+#pragma once
+
 #include <executorch/runtime/core/error.h>
 #include <executorch/runtime/core/result.h>
 #include <filesystem>

From 362a5c0b1d506a8c1d2db48b843aeb6ee110d58f Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Thu, 16 Oct 2025 12:46:53 -0700
Subject: [PATCH 18/21] ok give up on header only

---
 backends/cuda/CMakeLists.txt                |   1 +
 backends/cuda/runtime/platform/platform.cpp | 120 ++++++++++++++++++++
 backends/cuda/runtime/platform/platform.h   |  94 +--------------
 extension/module/module.cpp                 |  75 +++++++++---
 4 files changed, 187 insertions(+), 103 deletions(-)
 create mode 100644 backends/cuda/runtime/platform/platform.cpp

diff --git a/backends/cuda/CMakeLists.txt b/backends/cuda/CMakeLists.txt
index 8285fc0d582..d790051eb0c 100644
--- a/backends/cuda/CMakeLists.txt
+++ b/backends/cuda/CMakeLists.txt
@@ -39,6 +39,7 @@ set(_aoti_cuda_sources
     runtime/cuda_backend.cpp runtime/shims/memory.cpp
     runtime/shims/tensor_attribute.cpp runtime/guard.cpp
     runtime/shims/cuda_guard.cpp runtime/shims/int4mm.cu
+    runtime/platform/platform.cpp
 )
 add_library(aoti_cuda STATIC ${_aoti_cuda_sources})
 target_include_directories(
diff --git a/backends/cuda/runtime/platform/platform.cpp b/backends/cuda/runtime/platform/platform.cpp
new file mode 100644
index 00000000000..14a3ff2081f
--- /dev/null
+++ b/backends/cuda/runtime/platform/platform.cpp
@@ -0,0 +1,120 @@
+
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <executorch/backends/cuda/runtime/platform/platform.h>
+#include <executorch/runtime/core/error.h>
+#include <executorch/runtime/core/result.h>
+#include <filesystem>
+#include <string>
+
+#ifdef _WIN32
+#include <malloc.h>
+#include <windows.h>
+#else // Posix
+#include <dlfcn.h>
+#include <unistd.h>
+#include <cstdlib>
+#endif
+
+namespace executorch {
+namespace backends {
+namespace cuda {
+
+executorch::runtime::Result<void*> load_library(
+    const std::filesystem::path& path) {
+#ifdef _WIN32
+  std::string utf8 = path.u8string();
+  auto lib_handle = LoadLibrary(utf8.c_str());
+  if (lib_handle == NULL) {
+    ET_LOG(Error, "Failed to load %s with error: %lu", utf8.c_str(), GetLastError());
+    return executorch::runtime::Error::AccessFailed;
+  }
+
+#else
+  std::string path_str = path.string();
+  void* lib_handle = dlopen(path_str.c_str(), RTLD_LAZY | RTLD_LOCAL);
+  if (lib_handle == nullptr) {
+    ET_LOG(Error, "Failed to load %s with error: %s", path_str.c_str(), dlerror());
+    return executorch::runtime::Error::AccessFailed;
+  }
+#endif
+  return (void*)lib_handle;
+}
+
+executorch::runtime::Error close_library(void* lib_handle) {
+#ifdef _WIN32
+  if (!FreeLibrary((HMODULE)lib_handle)) {
+    printf("FreeLibrary failed with error %lu\n", GetLastError());
+    return executorch::runtime::Error::Internal;
+  }
+#else
+  if (dlclose(lib_handle) != 0) {
+    ET_LOG(Error, "dlclose failed: %s\n", dlerror());
+    return executorch::runtime::Error::Internal;
+  }
+#endif
+  return executorch::runtime::Error::Ok;
+}
+
+executorch::runtime::Result<void*> get_function(
+    void* lib_handle,
+    const std::string& fn_name) {
+#ifdef _WIN32
+  auto fn = GetProcAddress((HMODULE)lib_handle, fn_name.c_str());
+  if (!fn) {
+    ET_LOG(
+        Error,
+        "Failed loading symbol %s with error %lu\n",
+        fn_name.c_str(),
+        GetLastError());
+    return executorch::runtime::Error::Internal;
+  }
+#else
+  auto fn = dlsym(lib_handle, fn_name.c_str());
+  if (fn == nullptr) {
+    ET_LOG(
+        Error,
+        "Failed loading symbol %s with error %s\n",
+        fn_name.c_str(),
+        dlerror());
+    return executorch::runtime::Error::Internal;
+  }
+#endif
+
+  return (void*)fn; // This I think is technically ub on windows. We should
+                    // probably explicitly pack the bytes.
+}
+
+int32_t get_process_id() {
+#ifdef _WIN32
+  return GetCurrentProcessId();
+#else
+  return getpid();
+#endif
+}
+
+void* aligned_alloc(size_t alignment, size_t size) {
+#ifdef _WIN32
+  return _aligned_malloc(size, alignment);
+#else
+  return std::aligned_alloc(alignment, size);
+#endif
+}
+
+void aligned_free(void* ptr) {
+#ifdef _WIN32
+  _aligned_free(ptr);
+#else
+  std::free(ptr);
+#endif
+}
+
+} // namespace cuda
+} // namespace backends
+} // namespace executorch
diff --git a/backends/cuda/runtime/platform/platform.h b/backends/cuda/runtime/platform/platform.h
index e6edec8bfe6..00f278ef85e 100644
--- a/backends/cuda/runtime/platform/platform.h
+++ b/backends/cuda/runtime/platform/platform.h
@@ -14,106 +14,24 @@
 #include <filesystem>
 #include <string>
 
-#ifdef _WIN32
-#include <malloc.h>
-#include <windows.h>
-#else // Posix
-#include <dlfcn.h>
-#include <unistd.h>
-#include <cstdlib>
-#endif
-
 namespace executorch {
 namespace backends {
 namespace cuda {
 
 executorch::runtime::Result<void*> load_library(
-    const std::filesystem::path& path) {
-#ifdef _WIN32
-  std::string utf8 = path.u8string();
-  auto lib_handle = LoadLibrary(utf8.c_str());
-  if (lib_handle == NULL) {
-    ET_LOG(Error, "Failed to load %s with error: %lu", utf8, GetLastError());
-    return executorch::runtime::Error::AccessFailed;
-  }
-
-#else
-  void* lib_handle = dlopen(path.c_str(), RTLD_LAZY | RTLD_LOCAL);
-  if (lib_handle == nullptr) {
-    ET_LOG(Error, "Failed to load %s with error: %s", path.c_str(), dlerror());
-    return executorch::runtime::Error::AccessFailed;
-  }
-#endif
-  return (void*)lib_handle;
-}
+    const std::filesystem::path& path);
 
-executorch::runtime::Error close_library(void* lib_handle) {
-#ifdef _WIN32
-  if (!FreeLibrary((HMODULE)lib_handle)) {
-    printf("FreeLibrary failed with error %lu\n", GetLastError());
-    return executorch::runtime::Error::Internal;
-  }
-#else
-  if (dlclose(lib_handle) != 0) {
-    ET_LOG(Error, "dlclose failed: %s\n", dlerror());
-    return executorch::runtime::Error::Internal;
-  }
-#endif
-  return executorch::runtime::Error::Ok;
-}
+executorch::runtime::Error close_library(void* lib_handle);
 
 executorch::runtime::Result<void*> get_function(
     void* lib_handle,
-    const std::string& fn_name) {
-#ifdef _WIN32
-  auto fn = GetProcAddress((HMODULE)lib_handle, fn_name.c_str());
-  if (!fn) {
-    ET_LOG(
-        Error,
-        "Failed loading symbol %s with error %lu\n",
-        fn_name.c_str(),
-        GetLastError());
-    return executorch::runtime::Error::Internal;
-  }
-#else
-  auto fn = dlsym(lib_handle, fn_name.c_str());
-  if (fn == nullptr) {
-    ET_LOG(
-        Error,
-        "Failed loading symbol %s with error %s\n",
-        fn_name.c_str(),
-        dlerror());
-    return executorch::runtime::Error::Internal;
-  }
-#endif
-
-  return (void*)fn; // This I think is technically ub on windows. We should
-                    // probably explicitly pack the bytes.
-}
+    const std::string& fn_name);
 
-int32_t get_process_id() {
-#ifdef _WIN32
-  return GetCurrentProcessId();
-#else
-  return getpid();
-#endif
-}
+int32_t get_process_id();
 
-void* aligned_alloc(size_t alignment, size_t size) {
-#ifdef _WIN32
-  return _aligned_malloc(size, alignment);
-#else
-  return std::aligned_alloc(alignment, size);
-#endif
-}
+void* aligned_alloc(size_t alignment, size_t size);
 
-void aligned_free(void* ptr) {
-#ifdef _WIN32
-  _aligned_free(ptr);
-#else
-  std::free(ptr);
-#endif
-}
+void aligned_free(void* ptr);
 
 } // namespace cuda
 } // namespace backends
diff --git a/extension/module/module.cpp b/extension/module/module.cpp
index 9de77bcbc79..c4517c24a3b 100644
--- a/extension/module/module.cpp
+++ b/extension/module/module.cpp
@@ -50,19 +50,41 @@ runtime::Result<std::unique_ptr<runtime::DataLoader>> make_data_loader(
   std::unique_ptr<runtime::DataLoader> data_loader;
   switch (mode) {
     case Module::LoadMode::File:
-      data_loader = ET_UNWRAP_UNIQUE(FileDataLoader::from(file_path.c_str()));
+      auto res = FileDataLoader::from(file_path.c_str());
+      if (!res.ok()) {
+        return res.error();
+      }
+      data_loader = std::make_unique<std::remove_reference_t<decltype(*res)>>(
+          std::move(*res));
       break;
     case Module::LoadMode::Mmap:
-      data_loader = ET_UNWRAP_UNIQUE(MmapDataLoader::from(
-          file_path.c_str(), MmapDataLoader::MlockConfig::NoMlock));
+      auto res_mmap = MmapDataLoader::from(
+          file_path.c_str(), MmapDataLoader::MlockConfig::NoMlock);
+      if (!res_mmap.ok()) {
+        return res_mmap.error();
+      }
+      data_loader =
+          std::make_unique<std::remove_reference_t<decltype(*res_mmap)>>(
+              std::move(*res_mmap));
       break;
     case Module::LoadMode::MmapUseMlock:
-      data_loader = ET_UNWRAP_UNIQUE(MmapDataLoader::from(file_path.c_str()));
+      auto res_mlock = MmapDataLoader::from(file_path.c_str());
+      if (!res_mlock.ok()) {
+        return res_mlock.error();
+      }
+      data_loader =
+          std::make_unique<std::remove_reference_t<decltype(*res_mlock)>>(
+              std::move(*res_mlock));
       break;
     case Module::LoadMode::MmapUseMlockIgnoreErrors:
-      data_loader = ET_UNWRAP_UNIQUE(MmapDataLoader::from(
-          file_path.c_str(),
-          MmapDataLoader::MlockConfig::UseMlockIgnoreErrors));
+      auto res_mlock_ignore = MmapDataLoader::from(
+          file_path.c_str(), MmapDataLoader::MlockConfig::UseMlockIgnoreErrors);
+      if (!res_mlock_ignore.ok()) {
+        return res_mlock_ignore.error();
+      }
+      data_loader = std::make_unique<
+          std::remove_reference_t<decltype(*res_mlock_ignore)>>(
+          std::move(*res_mlock_ignore));
       break;
   }
   return data_loader;
@@ -165,8 +187,15 @@ runtime::Error Module::load(const Program::Verification verification) {
 
     if (data_map_loaders_.size() > 0) {
       for (auto i = 0; i < data_map_loaders_.size(); ++i) {
-        named_data_maps_.push_back(ET_UNWRAP_UNIQUE(
-            FlatTensorDataMap::load(data_map_loaders_[i].get())));
+        auto res_flat_tensor =
+            FlatTensorDataMap::load(data_map_loaders_[i].get());
+        if (!res_flat_tensor.ok()) {
+          return res_flat_tensor.error();
+        }
+        named_data_maps_.push_back(
+            std::make_unique<
+                std::remove_reference_t<decltype(*res_flat_tensor)>>(
+                std::move(*res_flat_tensor)));
       }
 
       // Extract raw pointers from unique_ptrs to pass to MergedDataMap::load()
@@ -175,13 +204,23 @@ runtime::Error Module::load(const Program::Verification verification) {
       for (const auto& data_map : named_data_maps_) {
         raw_data_maps.push_back(data_map.get());
       }
-      merged_data_map_ = ET_UNWRAP_UNIQUE(
-          MergedDataMap::load(runtime::Span<const NamedDataMap*>(
-              raw_data_maps.data(), raw_data_maps.size())));
+      auto res_merged = MergedDataMap::load(runtime::Span<const NamedDataMap*>(
+          raw_data_maps.data(), raw_data_maps.size()));
+      if (!res_merged.ok()) {
+        return res_merged.error();
+      }
+      merged_data_map_ =
+          std::make_unique<std::remove_reference_t<decltype(*res_merged)>>(
+              std::move(*res_merged));
     }
 
+    auto res_program = Program::load(data_loader_.get(), verification);
+    if (!res_program.ok()) {
+      return res_program.error();
+    }
     auto program =
-        ET_UNWRAP_UNIQUE(Program::load(data_loader_.get(), verification));
+        std::make_unique<std::remove_reference_t<decltype(*res_program)>>(
+            std::move(*res_program));
     program_ = std::shared_ptr<Program>(
         program.release(), [](Program* pointer) { delete pointer; });
   }
@@ -237,11 +276,17 @@ runtime::Error Module::load_method(
     }
     method_holder.memory_manager = std::make_unique<runtime::MemoryManager>(
         memory_allocator_.get(), planned_memory, temp_allocator_.get());
-    method_holder.method = ET_UNWRAP_UNIQUE(program_->load_method(
+    auto res_method = program_->load_method(
         method_name.c_str(),
         method_holder.memory_manager.get(),
         event_tracer ? event_tracer : this->event_tracer(),
-        merged_data_map_.get()));
+        merged_data_map_.get());
+    if (!res_method.ok()) {
+      return res_method.error();
+    }
+    method_holder.method =
+        std::make_unique<std::remove_reference_t<decltype(*res_method)>>(
+            std::move(*res_method));
     methods_.emplace(method_name, std::move(method_holder));
   }
   return runtime::Error::Ok;

From 6c0a0a5d2e8708cb301ca04c671da2c241334517 Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Thu, 16 Oct 2025 12:56:33 -0700
Subject: [PATCH 19/21] issue with history lead to module change being included

---
 extension/module/module.cpp | 75 ++++++++-----------------------------
 1 file changed, 15 insertions(+), 60 deletions(-)

diff --git a/extension/module/module.cpp b/extension/module/module.cpp
index c4517c24a3b..9de77bcbc79 100644
--- a/extension/module/module.cpp
+++ b/extension/module/module.cpp
@@ -50,41 +50,19 @@ runtime::Result<std::unique_ptr<runtime::DataLoader>> make_data_loader(
   std::unique_ptr<runtime::DataLoader> data_loader;
   switch (mode) {
     case Module::LoadMode::File:
-      auto res = FileDataLoader::from(file_path.c_str());
-      if (!res.ok()) {
-        return res.error();
-      }
-      data_loader = std::make_unique<std::remove_reference_t<decltype(*res)>>(
-          std::move(*res));
+      data_loader = ET_UNWRAP_UNIQUE(FileDataLoader::from(file_path.c_str()));
       break;
     case Module::LoadMode::Mmap:
-      auto res_mmap = MmapDataLoader::from(
-          file_path.c_str(), MmapDataLoader::MlockConfig::NoMlock);
-      if (!res_mmap.ok()) {
-        return res_mmap.error();
-      }
-      data_loader =
-          std::make_unique<std::remove_reference_t<decltype(*res_mmap)>>(
-              std::move(*res_mmap));
+      data_loader = ET_UNWRAP_UNIQUE(MmapDataLoader::from(
+          file_path.c_str(), MmapDataLoader::MlockConfig::NoMlock));
       break;
     case Module::LoadMode::MmapUseMlock:
-      auto res_mlock = MmapDataLoader::from(file_path.c_str());
-      if (!res_mlock.ok()) {
-        return res_mlock.error();
-      }
-      data_loader =
-          std::make_unique<std::remove_reference_t<decltype(*res_mlock)>>(
-              std::move(*res_mlock));
+      data_loader = ET_UNWRAP_UNIQUE(MmapDataLoader::from(file_path.c_str()));
       break;
     case Module::LoadMode::MmapUseMlockIgnoreErrors:
-      auto res_mlock_ignore = MmapDataLoader::from(
-          file_path.c_str(), MmapDataLoader::MlockConfig::UseMlockIgnoreErrors);
-      if (!res_mlock_ignore.ok()) {
-        return res_mlock_ignore.error();
-      }
-      data_loader = std::make_unique<
-          std::remove_reference_t<decltype(*res_mlock_ignore)>>(
-          std::move(*res_mlock_ignore));
+      data_loader = ET_UNWRAP_UNIQUE(MmapDataLoader::from(
+          file_path.c_str(),
+          MmapDataLoader::MlockConfig::UseMlockIgnoreErrors));
       break;
   }
   return data_loader;
@@ -187,15 +165,8 @@ runtime::Error Module::load(const Program::Verification verification) {
 
     if (data_map_loaders_.size() > 0) {
       for (auto i = 0; i < data_map_loaders_.size(); ++i) {
-        auto res_flat_tensor =
-            FlatTensorDataMap::load(data_map_loaders_[i].get());
-        if (!res_flat_tensor.ok()) {
-          return res_flat_tensor.error();
-        }
-        named_data_maps_.push_back(
-            std::make_unique<
-                std::remove_reference_t<decltype(*res_flat_tensor)>>(
-                std::move(*res_flat_tensor)));
+        named_data_maps_.push_back(ET_UNWRAP_UNIQUE(
+            FlatTensorDataMap::load(data_map_loaders_[i].get())));
       }
 
       // Extract raw pointers from unique_ptrs to pass to MergedDataMap::load()
@@ -204,23 +175,13 @@ runtime::Error Module::load(const Program::Verification verification) {
       for (const auto& data_map : named_data_maps_) {
         raw_data_maps.push_back(data_map.get());
       }
-      auto res_merged = MergedDataMap::load(runtime::Span<const NamedDataMap*>(
-          raw_data_maps.data(), raw_data_maps.size()));
-      if (!res_merged.ok()) {
-        return res_merged.error();
-      }
-      merged_data_map_ =
-          std::make_unique<std::remove_reference_t<decltype(*res_merged)>>(
-              std::move(*res_merged));
+      merged_data_map_ = ET_UNWRAP_UNIQUE(
+          MergedDataMap::load(runtime::Span<const NamedDataMap*>(
+              raw_data_maps.data(), raw_data_maps.size())));
     }
 
-    auto res_program = Program::load(data_loader_.get(), verification);
-    if (!res_program.ok()) {
-      return res_program.error();
-    }
     auto program =
-        std::make_unique<std::remove_reference_t<decltype(*res_program)>>(
-            std::move(*res_program));
+        ET_UNWRAP_UNIQUE(Program::load(data_loader_.get(), verification));
     program_ = std::shared_ptr<Program>(
         program.release(), [](Program* pointer) { delete pointer; });
   }
@@ -276,17 +237,11 @@ runtime::Error Module::load_method(
     }
     method_holder.memory_manager = std::make_unique<runtime::MemoryManager>(
         memory_allocator_.get(), planned_memory, temp_allocator_.get());
-    auto res_method = program_->load_method(
+    method_holder.method = ET_UNWRAP_UNIQUE(program_->load_method(
         method_name.c_str(),
         method_holder.memory_manager.get(),
         event_tracer ? event_tracer : this->event_tracer(),
-        merged_data_map_.get());
-    if (!res_method.ok()) {
-      return res_method.error();
-    }
-    method_holder.method =
-        std::make_unique<std::remove_reference_t<decltype(*res_method)>>(
-            std::move(*res_method));
+        merged_data_map_.get()));
     methods_.emplace(method_name, std::move(method_holder));
   }
   return runtime::Error::Ok;

From 20b92d57f3e865e9bfd92d09b2496e67e2426e1b Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Thu, 16 Oct 2025 14:01:02 -0700
Subject: [PATCH 20/21] lint

---
 backends/cuda/runtime/platform/platform.cpp | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/backends/cuda/runtime/platform/platform.cpp b/backends/cuda/runtime/platform/platform.cpp
index 14a3ff2081f..5264dcbd03a 100644
--- a/backends/cuda/runtime/platform/platform.cpp
+++ b/backends/cuda/runtime/platform/platform.cpp
@@ -32,7 +32,11 @@ executorch::runtime::Result<void*> load_library(
   std::string utf8 = path.u8string();
   auto lib_handle = LoadLibrary(utf8.c_str());
   if (lib_handle == NULL) {
-    ET_LOG(Error, "Failed to load %s with error: %lu", utf8.c_str(), GetLastError());
+    ET_LOG(
+        Error,
+        "Failed to load %s with error: %lu",
+        utf8.c_str(),
+        GetLastError());
     return executorch::runtime::Error::AccessFailed;
   }
 
@@ -40,7 +44,8 @@ executorch::runtime::Result<void*> load_library(
   std::string path_str = path.string();
   void* lib_handle = dlopen(path_str.c_str(), RTLD_LAZY | RTLD_LOCAL);
   if (lib_handle == nullptr) {
-    ET_LOG(Error, "Failed to load %s with error: %s", path_str.c_str(), dlerror());
+    ET_LOG(
+        Error, "Failed to load %s with error: %s", path_str.c_str(), dlerror());
     return executorch::runtime::Error::AccessFailed;
   }
 #endif

From 56734d28d2b4bcd165da7c3121c83b5de5bcd752 Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@fb.com>
Date: Thu, 16 Oct 2025 15:05:42 -0700
Subject: [PATCH 21/21] cmake lint

---
 backends/cuda/CMakeLists.txt | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/backends/cuda/CMakeLists.txt b/backends/cuda/CMakeLists.txt
index d790051eb0c..8b94351d469 100644
--- a/backends/cuda/CMakeLists.txt
+++ b/backends/cuda/CMakeLists.txt
@@ -36,9 +36,12 @@ find_package_torch()
 
 # CUDA-specific AOTI functionality
 set(_aoti_cuda_sources
-    runtime/cuda_backend.cpp runtime/shims/memory.cpp
-    runtime/shims/tensor_attribute.cpp runtime/guard.cpp
-    runtime/shims/cuda_guard.cpp runtime/shims/int4mm.cu
+    runtime/cuda_backend.cpp
+    runtime/shims/memory.cpp
+    runtime/shims/tensor_attribute.cpp
+    runtime/guard.cpp
+    runtime/shims/cuda_guard.cpp
+    runtime/shims/int4mm.cu
     runtime/platform/platform.cpp
 )
 add_library(aoti_cuda STATIC ${_aoti_cuda_sources})