diff --git a/.eslintrc.js b/.eslintrc.js
index b09ff37d23..fd7eca7393 100644
--- a/.eslintrc.js
+++ b/.eslintrc.js
@@ -8,6 +8,7 @@ const VALID_CATEGORIES = [
   'Models - Image Embeddings',
   'Models - Image Generation',
   'Models - LLM',
+  'Models - LLM Multimodal',
   'Models - Object Detection',
   'Models - Instance Segmentation',
   'Models - Pose Estimation',
diff --git a/apps/llm/components/llmModels.ts b/apps/llm/components/llmModels.ts
index 1991578973..58b8c01d74 100644
--- a/apps/llm/components/llmModels.ts
+++ b/apps/llm/components/llmModels.ts
@@ -10,8 +10,11 @@ const llm = models.llm;
 export type LLMModelSources = LLMProps['model'];
 
 export const LLM_MODELS: ModelOption<LLMModelSources>[] = [
-  // Gemma4
-  { label: 'Gemma4 E2B', value: llm.gemma4_e2b() },
+  //Gemma 4
+  {
+    label: 'Gemma 4 E2B',
+    value: llm.gemma4_e2b(),
+  },
   // Llama 3.2
   {
     label: 'Llama 3.2 1B',
diff --git a/packages/react-native-executorch/android/libs/classes.jar b/packages/react-native-executorch/android/libs/classes.jar
index be5ec2ee7f..81eee28f71 100644
Binary files a/packages/react-native-executorch/android/libs/classes.jar and b/packages/react-native-executorch/android/libs/classes.jar differ
diff --git a/packages/react-native-executorch/common/runner/text_prefiller.cpp b/packages/react-native-executorch/common/runner/text_prefiller.cpp
index 8325efae8f..370ca5c7f4 100644
--- a/packages/react-native-executorch/common/runner/text_prefiller.cpp
+++ b/packages/react-native-executorch/common/runner/text_prefiller.cpp
@@ -18,10 +18,11 @@ namespace llm {
 
 TextPrefiller::TextPrefiller(TextDecoderRunner *text_decoder_runner,
                              bool use_kv_cache, bool enable_parallel_prefill,
-                             int64_t max_seq_len)
+                             int64_t max_seq_len, int32_t prefill_chunk_size)
     : text_decoder_runner_(text_decoder_runner), use_kv_cache_(use_kv_cache),
       enable_parallel_prefill_(enable_parallel_prefill),
-      max_seq_len_(max_seq_len > 0 ? max_seq_len : 128) {}
+      max_seq_len_(max_seq_len > 0 ? max_seq_len : 128),
+      prefill_chunk_size_(prefill_chunk_size) {}
 
 ::executorch::runtime::Result<uint64_t>
 TextPrefiller::prefill(std::vector<uint64_t> &prompt_tokens,
@@ -31,11 +32,10 @@ TextPrefiller::prefill(std::vector<uint64_t> &prompt_tokens,
     ET_CHECK_OK_OR_RETURN_ERROR(text_decoder_runner_->load());
   }
 
-  // Check if we need to chunk the prompt tokens
   int32_t num_prompt_tokens = prompt_tokens.size();
-  const int32_t chunk_size = static_cast<int32_t>(max_seq_len_);
+  int32_t chunk_size =
+      prefill_chunk_size_ > 0 ? prefill_chunk_size_ : max_seq_len_;
 
-  // If prompt tokens exceed chunk_size, we need to chunk them
   if (num_prompt_tokens > chunk_size) {
     uint64_t cur_token = 0;
     int num_tokens_to_process = 0;
diff --git a/packages/react-native-executorch/common/runner/text_prefiller.h b/packages/react-native-executorch/common/runner/text_prefiller.h
index 7929fe9c7f..b8cdb1b98c 100644
--- a/packages/react-native-executorch/common/runner/text_prefiller.h
+++ b/packages/react-native-executorch/common/runner/text_prefiller.h
@@ -19,8 +19,14 @@ namespace llm {
 
 class TextPrefiller {
 public:
+  // prefill_chunk_size: when > 0, the prompt is always processed in steps of
+  // this size (see prefill()). Set to the model's forward sequence-length cap
+  // for the MLX backend (its forward is exported with a sliding-window bound
+  // and one-shot prefill spikes Metal memory). Other backends (XNNPACK/CoreML)
+  // pass 0 → original one-shot behavior.
   TextPrefiller(TextDecoderRunner *text_decoder_runner, bool use_kv_cache,
-                bool enable_parallel_prefill, int64_t max_seq_len = 128);
+                bool enable_parallel_prefill, int64_t max_seq_len = 128,
+                int32_t prefill_chunk_size = 0);
 
   virtual ~TextPrefiller() = default;
   /**
@@ -70,6 +76,7 @@ class TextPrefiller {
   bool use_kv_cache_;
   bool enable_parallel_prefill_;
   int64_t max_seq_len_;
+  int32_t prefill_chunk_size_;
 };
 
 } // namespace llm
diff --git a/packages/react-native-executorch/common/runner/text_runner.cpp b/packages/react-native-executorch/common/runner/text_runner.cpp
index 348f0bfb82..bbcc0d8981 100644
--- a/packages/react-native-executorch/common/runner/text_runner.cpp
+++ b/packages/react-native-executorch/common/runner/text_runner.cpp
@@ -26,11 +26,24 @@ Error TextRunner::load_subcomponents() {
 
   Stats *stats_ptr = &stats_;
 
-  text_decoder_runner_ = std::make_unique<TextDecoderRunner>(
-      *module_, io_manager_.get(), config_);
+  text_decoder_runner_ =
+      std::make_unique<TextDecoderRunner>(*module_, io_manager_.get(), config_);
+
+  int32_t prefill_chunk_size = 0;
+  auto fwd_meta = module_->method_meta("forward");
+  if (fwd_meta.ok() && fwd_meta->uses_backend("MLXBackend")) {
+    auto input_meta = fwd_meta->input_tensor_meta(0);
+    if (input_meta.ok()) {
+      auto sizes = input_meta->sizes();
+      if (sizes.size() >= 2 && sizes[sizes.size() - 1] > 0) {
+        prefill_chunk_size = sizes[sizes.size() - 1];
+      }
+    }
+  }
+
   text_prefiller_ = std::make_unique<TextPrefiller>(
       text_decoder_runner_.get(), config_.enable_kv_cache,
-      config_.enable_dynamic_shape, config_.max_seq_len);
+      config_.enable_dynamic_shape, config_.max_seq_len, prefill_chunk_size);
   text_token_generator_ = std::make_unique<TextTokenGenerator>(
       tokenizer_.get(), text_decoder_runner_.get(), config_.enable_kv_cache,
       std::move(eos_ids_), stats_ptr, config_);
@@ -83,10 +96,12 @@ Error TextRunner::generate_internal(
   std::vector<uint64_t> prompt_tokens = encodeResult.get();
   int num_prompt_tokens = prompt_tokens.size();
 
-  // For dynamic-shape PTEs (Gemma4 iter*), get_max_seq_len is the per-call
-  // decoder chunk size (e.g. 128) and the true generation budget lives in
-  // get_max_context_len. Static-shape PTEs set both equal, so this collapses
-  // to the old behavior. Mirrors multimodal_prefiller.cpp:96.
+  // For dynamic-shape PTEs (e.g. Gemma4 MLX/Vulkan), get_max_seq_len is the
+  // per-call decoder chunk size (e.g. the sliding window) and the real
+  // generation budget lives in get_max_context_len. Static-shape PTEs set both
+  // equal, so this collapses to the old behavior. Without this the budget is
+  // computed from the small chunk size, so max_new_tokens can resolve to ~0 and
+  // generation ends immediately after prefill.
   const int32_t seq_cap = config_.enable_dynamic_shape
                               ? config_.max_context_length
                               : config_.max_seq_len;
diff --git a/packages/react-native-executorch/react-native-executorch.podspec b/packages/react-native-executorch/react-native-executorch.podspec
index 849759243f..daf0da27c7 100644
--- a/packages/react-native-executorch/react-native-executorch.podspec
+++ b/packages/react-native-executorch/react-native-executorch.podspec
@@ -62,6 +62,12 @@ Pod::Spec.new do |s|
 
   s.libraries = "z"
   s.ios.vendored_frameworks = "third-party/ios/ExecutorchLib.xcframework"
+
+  # NOTE: mlx.metallib (the MLX GPU kernels) is bundled INSIDE
+  # ExecutorchLib.framework, colocated with the binary that contains the MLX
+  # code. MLX's runtime loader resolves the metallib relative to that binary
+  # (via dladdr), so it must live next to it in the framework — not at the app
+  # bundle root.
   # Exclude file with tests to not introduce gtest dependency.
   # Do not include the headers from common/rnexecutorch/jsi/ as source files.
   # Xcode/Cocoapods leaks them to other pods that an app also depends on, so if
diff --git a/packages/react-native-executorch/src/constants/modelRegistry.ts b/packages/react-native-executorch/src/constants/modelRegistry.ts
index bd04a37883..c3cda78498 100644
--- a/packages/react-native-executorch/src/constants/modelRegistry.ts
+++ b/packages/react-native-executorch/src/constants/modelRegistry.ts
@@ -38,7 +38,7 @@ import { RnExecutorchErrorCode } from '../errors/ErrorCodes';
  * compile-time error.
  * @category Utils
  */
-export type Backend = 'xnnpack' | 'coreml' | 'vulkan' | 'qnn';
+export type Backend = 'xnnpack' | 'coreml' | 'vulkan' | 'qnn' | 'mlx';
 
 /**
  * Options for a `models` accessor call.
@@ -78,7 +78,7 @@ type ConfigOf<V> = Extract<
 >;
 type BackendsOf<V> = Extract<keyof V, Backend>;
 
-const BACKEND_ORDER: Backend[] = ['xnnpack', 'coreml', 'vulkan', 'qnn'];
+const BACKEND_ORDER: Backend[] = ['xnnpack', 'coreml', 'mlx', 'vulkan', 'qnn'];
 
 function firstBackend(variants: AnyVariantMap): Backend {
   for (const b of BACKEND_ORDER) {
@@ -181,6 +181,33 @@ function tts<C extends TextToSpeechModelConfig>(c: C): () => C {
 // Per-backend variant maps for models that ship more than one backend.
 // ─────────────────────────────────────────────────────────────────────────────
 
+const GEMMA4_E2B_VARIANTS = {
+  mlx: {
+    base: {
+      modelName: 'gemma4-e2b' as const,
+      modelSource: M.GEMMA4_E2B_MLX_MODEL,
+      tokenizerSource: M.GEMMA4_E2B_TOKENIZER,
+      tokenizerConfigSource: M.GEMMA4_E2B_TOKENIZER_CONFIG,
+    },
+  },
+  xnnpack: {
+    base: {
+      modelName: 'gemma4-e2b' as const,
+      modelSource: M.GEMMA4_E2B_XNNPACK_MODEL,
+      tokenizerSource: M.GEMMA4_E2B_TOKENIZER,
+      tokenizerConfigSource: M.GEMMA4_E2B_TOKENIZER_CONFIG,
+    },
+  },
+  vulkan: {
+    base: {
+      modelName: 'gemma4-e2b' as const,
+      modelSource: M.GEMMA4_E2B_VULKAN_MODEL,
+      tokenizerSource: M.GEMMA4_E2B_TOKENIZER,
+      tokenizerConfigSource: M.GEMMA4_E2B_TOKENIZER_CONFIG,
+    },
+  },
+};
+
 const EFFICIENTNET_V2_S_VARIANTS = {
   xnnpack: {
     base: {
@@ -496,7 +523,10 @@ export const models = {
       M.LFM2_5_1_2B_INSTRUCT_QUANTIZED
     ),
     bielik_v3_0_1_5b: pair(M.BIELIK_V3_0_1_5B, M.BIELIK_V3_0_1_5B_QUANTIZED),
-    gemma4_e2b: base(M.GEMMA4_E2B),
+    gemma4_e2b: variant(GEMMA4_E2B_VARIANTS, {
+      ios: 'mlx',
+      android: 'vulkan',
+    }),
     // Multimodal LLMs — same hook/module as plain LLMs, listed here so users
     // pick a model by capability ("LLM") rather than by modality.
     lfm2_5_vl_1_6b: base(M.LFM2_5_VL_1_6B_QUANTIZED),
diff --git a/packages/react-native-executorch/src/constants/modelUrls.ts b/packages/react-native-executorch/src/constants/modelUrls.ts
index b14372d145..8ff13f585c 100644
--- a/packages/react-native-executorch/src/constants/modelUrls.ts
+++ b/packages/react-native-executorch/src/constants/modelUrls.ts
@@ -124,34 +124,40 @@ export const QWEN3_0_6B_QUANTIZED = {
   generationConfig: QWEN3_GENERATION_CONFIG,
 } as const;
 
-// GEMMA 4
-const GEMMA4_E2B_XNNPACK = `${URL_PREFIX}-gemma-4/${PREVIOUS_VERSION_TAG}/e2b/xnnpack/gemma_4_e2b_xnnpack_8da4w.pte`;
-const GEMMA4_E2B_VULKAN = `${URL_PREFIX}-gemma-4/${PREVIOUS_VERSION_TAG}/e2b/vulkan/gemma_4_e2b_vulkan_8da4w.pte`;
-const GEMMA4_E2B_XNNPACK_MM = `${URL_PREFIX}-gemma-4-multimodal/${PREVIOUS_VERSION_TAG}/e2b/xnnpack/gemma_4_e2b_xnnpack_8da4w.pte`;
+// GEMMA 4 — separate HF repo; tokenizer files live at the e2b root and are
+// shared by all backend variants.
+const GEMMA4_E2B_PREFIX = `${URL_PREFIX}-gemma-4/${PREVIOUS_VERSION_TAG}/e2b`;
+export const GEMMA4_E2B_MLX_MODEL = `${GEMMA4_E2B_PREFIX}/mlx/gemma4_e2b_mlx_int4.pte`;
+export const GEMMA4_E2B_XNNPACK_MODEL = `${GEMMA4_E2B_PREFIX}/xnnpack/gemma_4_e2b_xnnpack_8da4w.pte`;
+export const GEMMA4_E2B_VULKAN_MODEL = `${GEMMA4_E2B_PREFIX}/vulkan/gemma_4_e2b_vulkan_8da4w.pte`;
+export const GEMMA4_E2B_TOKENIZER = `${GEMMA4_E2B_PREFIX}/tokenizer.json`;
+export const GEMMA4_E2B_TOKENIZER_CONFIG = `${GEMMA4_E2B_PREFIX}/tokenizer_config.json`;
+
+const GEMMA4_E2B_MODEL =
+  Platform.OS === `android` ? GEMMA4_E2B_VULKAN_MODEL : GEMMA4_E2B_MLX_MODEL;
+
+const GEMMA4_E2B_MLX_MM = `${URL_PREFIX}-gemma-4-multimodal/${PREVIOUS_VERSION_TAG}/e2b/mlx/gemma4_e2b_mlx_int4.pte`;
 const GEMMA4_E2B_VULKAN_MM = `${URL_PREFIX}-gemma-4-multimodal/${PREVIOUS_VERSION_TAG}/e2b/vulkan/gemma_4_e2b_vulkan_8da4w.pte`;
-const GEMMA4_TOKENIZER = `${URL_PREFIX}-gemma-4/${PREVIOUS_VERSION_TAG}/e2b/tokenizer.json`;
-const GEMMA4_TOKENIZER_CONFIG = `${URL_PREFIX}-gemma-4/${PREVIOUS_VERSION_TAG}/e2b/tokenizer_config.json`;
 
 /**
  * @category Models - LLM
  */
 export const GEMMA4_E2B = {
   modelName: 'gemma4-e2b',
-  modelSource:
-    Platform.OS === `android` ? GEMMA4_E2B_VULKAN : GEMMA4_E2B_XNNPACK,
-  tokenizerSource: GEMMA4_TOKENIZER,
-  tokenizerConfigSource: GEMMA4_TOKENIZER_CONFIG,
+  modelSource: GEMMA4_E2B_MODEL,
+  tokenizerSource: GEMMA4_E2B_TOKENIZER,
+  tokenizerConfigSource: GEMMA4_E2B_TOKENIZER_CONFIG,
 } as const;
 
 /**
- * @category Models - VLM
+ * @category Models - LLM Multimodal
  */
 export const GEMMA4_E2B_MM = {
   modelName: 'gemma4-e2b-multimodal',
   modelSource:
-    Platform.OS === `android` ? GEMMA4_E2B_VULKAN_MM : GEMMA4_E2B_XNNPACK_MM,
-  tokenizerSource: GEMMA4_TOKENIZER,
-  tokenizerConfigSource: GEMMA4_TOKENIZER_CONFIG,
+    Platform.OS === `android` ? GEMMA4_E2B_VULKAN_MM : GEMMA4_E2B_MLX_MM,
+  tokenizerSource: GEMMA4_E2B_TOKENIZER,
+  tokenizerConfigSource: GEMMA4_E2B_TOKENIZER_CONFIG,
   capabilities: ['vision', 'audio'],
   audioConfig: {
     samplesPerBlock: 7680,
diff --git a/packages/react-native-executorch/third-party/android/libs/executorch/arm64-v8a/libexecutorch.so b/packages/react-native-executorch/third-party/android/libs/executorch/arm64-v8a/libexecutorch.so
index 1e882b92fa..36cb0919c2 100644
Binary files a/packages/react-native-executorch/third-party/android/libs/executorch/arm64-v8a/libexecutorch.so and b/packages/react-native-executorch/third-party/android/libs/executorch/arm64-v8a/libexecutorch.so differ
diff --git a/packages/react-native-executorch/third-party/android/libs/executorch/x86_64/libexecutorch.so b/packages/react-native-executorch/third-party/android/libs/executorch/x86_64/libexecutorch.so
index 45efcf585d..7d39ee85a3 100644
Binary files a/packages/react-native-executorch/third-party/android/libs/executorch/x86_64/libexecutorch.so and b/packages/react-native-executorch/third-party/android/libs/executorch/x86_64/libexecutorch.so differ
diff --git a/packages/react-native-executorch/third-party/include/executorch/ExecuTorch.h b/packages/react-native-executorch/third-party/include/executorch/ExecuTorch.h
index 3a12a5ddba..d0ad6c2840 100644
--- a/packages/react-native-executorch/third-party/include/executorch/ExecuTorch.h
+++ b/packages/react-native-executorch/third-party/include/executorch/ExecuTorch.h
@@ -6,6 +6,8 @@
  * LICENSE file in the root directory of this source tree.
  */
 
+#import "ExecuTorchBackendOption.h"
+#import "ExecuTorchBackendOptionsMap.h"
 #import "ExecuTorchError.h"
 #import "ExecuTorchLog.h"
 #import "ExecuTorchModule.h"
diff --git a/packages/react-native-executorch/third-party/include/executorch/ExecuTorchModule.h b/packages/react-native-executorch/third-party/include/executorch/ExecuTorchModule.h
index 823e5cf5cb..51b8abfa68 100644
--- a/packages/react-native-executorch/third-party/include/executorch/ExecuTorchModule.h
+++ b/packages/react-native-executorch/third-party/include/executorch/ExecuTorchModule.h
@@ -6,6 +6,8 @@
  * LICENSE file in the root directory of this source tree.
  */
 
+#import "ExecuTorchBackendOption.h"
+#import "ExecuTorchBackendOptionsMap.h"
 #import "ExecuTorchValue.h"
 
 NS_ASSUME_NONNULL_BEGIN
@@ -198,6 +200,37 @@ NS_SWIFT_NAME(Module)
  */
 - (BOOL)load:(NSError **)error;
 
+/**
+ * Loads the module's program with per-delegate backend options.
+ *
+ * The receiver retains @c options for as long as the underlying program
+ * references it (lifetime tracked via ARC).
+ *
+ * @param options A `ExecuTorchBackendOptionsMap` containing per-delegate
+ *        load-time configuration, built once via
+ *        `[ExecuTorchBackendOptionsMap mapWithOptions:error:]`.
+ * @param verification The verification level to apply when loading the program.
+ * @param error A pointer to an NSError pointer that will be set if an error
+ * occurs.
+ * @return YES if the program was successfully loaded; otherwise, NO.
+ */
+- (BOOL)loadWithOptions:(ExecuTorchBackendOptionsMap *)options
+           verification:(ExecuTorchVerification)verification
+                  error:(NSError **)error NS_REFINED_FOR_SWIFT;
+
+/**
+ * Loads the module's program with per-delegate backend options using minimal
+ * verification.
+ *
+ * @param options A `ExecuTorchBackendOptionsMap` containing per-delegate
+ *        load-time configuration.
+ * @param error A pointer to an NSError pointer that will be set if an error
+ * occurs.
+ * @return YES if the program was successfully loaded; otherwise, NO.
+ */
+- (BOOL)loadWithOptions:(ExecuTorchBackendOptionsMap *)options
+                  error:(NSError **)error NS_REFINED_FOR_SWIFT;
+
 /**
  * Checks if the module is loaded.
  *
@@ -215,6 +248,19 @@ NS_SWIFT_NAME(Module)
 - (BOOL)loadMethod:(NSString *)methodName
              error:(NSError **)error NS_SWIFT_NAME(load(_:));
 
+/**
+ * Loads a specific method from the program with per-delegate backend options.
+ *
+ * @param methodName A string representing the name of the method to load.
+ * @param options A `ExecuTorchBackendOptionsMap` containing per-delegate
+ *        load-time configuration.
+ * @param error A pointer to an NSError pointer that is set if an error occurs.
+ * @return YES if the method was successfully loaded; otherwise, NO.
+ */
+- (BOOL)loadMethod:(NSString *)methodName
+           options:(ExecuTorchBackendOptionsMap *)options
+             error:(NSError **)error NS_REFINED_FOR_SWIFT;
+
 /**
  * Checks if a specific method is loaded.
  *
diff --git a/packages/react-native-executorch/third-party/include/executorch/extension/data_loader/buffer_data_loader.h b/packages/react-native-executorch/third-party/include/executorch/extension/data_loader/buffer_data_loader.h
index 38322aff43..b3744d0970 100644
--- a/packages/react-native-executorch/third-party/include/executorch/extension/data_loader/buffer_data_loader.h
+++ b/packages/react-native-executorch/third-party/include/executorch/extension/data_loader/buffer_data_loader.h
@@ -36,9 +36,10 @@ class BufferDataLoader final : public executorch::runtime::DataLoader {
        ET_UNUSED const DataLoader::SegmentInfo &segment_info) const override {
     size_t total_size;
     bool overflow = c10::add_overflows(offset, size, &total_size);
-    ET_CHECK_OR_RETURN_ERROR(!overflow && total_size <= size_, InvalidArgument,
-                             "offset %zu + size %zu > size_ %zu", offset, size,
-                             size_);
+    ET_CHECK_OR_RETURN_ERROR(
+        !overflow && total_size <= size_, InvalidArgument,
+        "offset %zu + size %zu > size_ %zu, or overflow detected", offset, size,
+        size_);
     return executorch::runtime::FreeableBuffer(data_ + offset, size,
                                                /*free_fn=*/nullptr);
   }
diff --git a/packages/react-native-executorch/third-party/include/executorch/extension/data_loader/mman.h b/packages/react-native-executorch/third-party/include/executorch/extension/data_loader/mman.h
index 788560c168..c35fd07f46 100644
--- a/packages/react-native-executorch/third-party/include/executorch/extension/data_loader/mman.h
+++ b/packages/react-native-executorch/third-party/include/executorch/extension/data_loader/mman.h
@@ -17,6 +17,7 @@
 
 #ifndef _WIN32
 
+#include <fcntl.h>
 #include <sys/mman.h>
 #include <unistd.h>
 
@@ -41,6 +42,34 @@ ET_INLINE off_t get_mmap_offset(size_t offset) {
   return static_cast<off_t>(offset);
 }
 
+/**
+ * Hint the kernel to prefetch pages eagerly and to optimize for sequential
+ * reads. Intended to reduce page-fault stutter during model initialization
+ * when the caller does not want to mlock the pages into RAM.
+ */
+ET_INLINE void madvise_pages_willneed_sequential(void *addr, size_t len) {
+  ::madvise(addr, len, MADV_WILLNEED);
+  ::madvise(addr, len, MADV_SEQUENTIAL);
+}
+
+/**
+ * On Apple platforms, schedule kernel read-ahead on the file descriptor itself
+ * via fcntl(F_RDADVISE). This is more aggressive than madvise for cold starts:
+ * it brings pages into the unified buffer cache so first-touch faults are
+ * serviced from RAM instead of storage. No-op on non-Apple POSIX platforms.
+ */
+ET_INLINE void fcntl_rdadvise_apple(int fd, size_t file_size) {
+#if defined(__APPLE__)
+  struct radvisory advice;
+  advice.ra_offset = 0;
+  advice.ra_count = static_cast<int>(file_size);
+  ::fcntl(fd, F_RDADVISE, &advice);
+#else
+  (void)fd;
+  (void)file_size;
+#endif
+}
+
 #else
 
 #define NOMINMAX
@@ -78,4 +107,21 @@ ET_INLINE uint64_t get_mmap_offset(size_t offset) {
   return static_cast<uint64_t>(offset);
 }
 
+/**
+ * No-op on Windows: there is no direct equivalent to madvise(MADV_WILLNEED |
+ * MADV_SEQUENTIAL) and the existing mman_windows shim does not implement one.
+ */
+ET_INLINE void madvise_pages_willneed_sequential(void *addr, size_t len) {
+  (void)addr;
+  (void)len;
+}
+
+/**
+ * No-op on Windows: F_RDADVISE is an Apple-specific fcntl command.
+ */
+ET_INLINE void fcntl_rdadvise_apple(int fd, size_t file_size) {
+  (void)fd;
+  (void)file_size;
+}
+
 #endif
diff --git a/packages/react-native-executorch/third-party/include/executorch/extension/data_loader/mmap_data_loader.h b/packages/react-native-executorch/third-party/include/executorch/extension/data_loader/mmap_data_loader.h
index cbc2ced2d9..1b648ce35f 100644
--- a/packages/react-native-executorch/third-party/include/executorch/extension/data_loader/mmap_data_loader.h
+++ b/packages/react-native-executorch/third-party/include/executorch/extension/data_loader/mmap_data_loader.h
@@ -38,6 +38,10 @@ class MmapDataLoader final : public executorch::runtime::DataLoader {
     UseMlock,
     /// Call `mlock()` on loaded pages, ignoring errors if it fails.
     UseMlockIgnoreErrors,
+    /// Use madvise(MADV_WILLNEED | MADV_SEQUENTIAL) instead of mlock.
+    /// Tells the kernel to prefetch pages eagerly and optimize for
+    /// sequential reads, without pinning them in RAM.
+    UseMadvise,
   };
 
   /**
diff --git a/packages/react-native-executorch/third-party/include/executorch/extension/data_loader/shared_ptr_data_loader.h b/packages/react-native-executorch/third-party/include/executorch/extension/data_loader/shared_ptr_data_loader.h
index dc0e179187..e9bb26322f 100644
--- a/packages/react-native-executorch/third-party/include/executorch/extension/data_loader/shared_ptr_data_loader.h
+++ b/packages/react-native-executorch/third-party/include/executorch/extension/data_loader/shared_ptr_data_loader.h
@@ -8,6 +8,7 @@
 
 #pragma once
 
+#include <c10/util/safe_numerics.h>
 #include <executorch/runtime/core/data_loader.h>
 #include <executorch/runtime/core/error.h>
 #include <executorch/runtime/core/result.h>
@@ -32,9 +33,12 @@ class SharedPtrDataLoader final : public executorch::runtime::DataLoader {
   executorch::runtime::Result<executorch::runtime::FreeableBuffer>
   load(size_t offset, size_t size,
        ET_UNUSED const DataLoader::SegmentInfo &segment_info) const override {
-    ET_CHECK_OR_RETURN_ERROR(offset + size <= size_, InvalidArgument,
-                             "offset %zu + size %zu > size_ %zu", offset, size,
-                             size_);
+    size_t total_size;
+    bool overflow = c10::add_overflows(offset, size, &total_size);
+    ET_CHECK_OR_RETURN_ERROR(
+        !overflow && total_size <= size_, InvalidArgument,
+        "offset %zu + size %zu > size_ %zu, or overflow detected", offset, size,
+        size_);
     return executorch::runtime::FreeableBuffer(
         static_cast<uint8_t *>(data_.get()) + offset, size,
         /*free_fn=*/nullptr);
diff --git a/packages/react-native-executorch/third-party/include/executorch/extension/module/module.h b/packages/react-native-executorch/third-party/include/executorch/extension/module/module.h
index 278f996ed2..9dadc846ee 100644
--- a/packages/react-native-executorch/third-party/include/executorch/extension/module/module.h
+++ b/packages/react-native-executorch/third-party/include/executorch/extension/module/module.h
@@ -14,6 +14,8 @@
 #include <unordered_set>
 #include <vector>
 
+#include <executorch/runtime/backend/backend_options_map.h>
+#include <executorch/runtime/backend/options.h>
 #include <executorch/runtime/executor/program.h>
 
 #ifdef USE_ATEN_LIB
@@ -25,6 +27,7 @@
 namespace executorch {
 namespace extension {
 
+using ET_RUNTIME_NAMESPACE::Kernel;
 using ET_RUNTIME_NAMESPACE::Method;
 using ET_RUNTIME_NAMESPACE::MethodMeta;
 using ET_RUNTIME_NAMESPACE::NamedDataMap;
@@ -51,6 +54,8 @@ class Module {
     MmapUseMlock,
     /// Use memory locking and ignore errors.
     MmapUseMlockIgnoreErrors,
+    /// Use mmap with madvise(MADV_WILLNEED | MADV_SEQUENTIAL) hints.
+    MmapUseMadvise,
   };
 
   /**
@@ -182,9 +187,18 @@ class Module {
   /**
    * Loads the program with per-delegate runtime options.
    *
-   * @param[in] backend_options A LoadBackendOptionsMap containing per-delegate
-   * load-time configuration options. The caller must ensure this object
-   * outlives any methods loaded with these options.
+   * The Module deep-copies `backend_options` into internal storage, so the
+   * caller may release the input (and any backing BackendOption arrays its
+   * Spans referenced) immediately after this call returns. Future lazy
+   * `load_method` calls (e.g. triggered by `forward`) consume the
+   * Module-owned copy.
+   *
+   * Transactional: on failure, the previously-installed backend options
+   * (if any) are left in place; the input is not committed.
+   *
+   * @param[in] backend_options A LoadBackendOptionsMap containing
+   * per-delegate load-time configuration options. Deep-copied into the
+   * Module on success; not retained on failure.
    * @param[in] verification The type of verification to do before returning
    * success.
    *
@@ -195,6 +209,21 @@ class Module {
        const Program::Verification verification =
            Program::Verification::Minimal);
 
+  /**
+   * Returns the deep-copied LoadBackendOptionsMap most recently installed
+   * via `load(LoadBackendOptionsMap, ...)`. The returned reference is owned
+   * by the Module and remains valid until the next call to
+   * `load(LoadBackendOptionsMap, ...)` or until the Module is destroyed.
+   *
+   * If `load(LoadBackendOptionsMap, ...)` has never been called, returns a
+   * default-constructed (empty, `size() == 0`) map.
+   *
+   * @returns Const reference to the Module-owned LoadBackendOptionsMap.
+   */
+  inline const LoadBackendOptionsMap &backend_options() const {
+    return backend_options_map_;
+  }
+
   /**
    * Checks if the program is loaded.
    *
@@ -246,7 +275,8 @@ class Module {
   load_method(const std::string &method_name,
               runtime::HierarchicalAllocator *planned_memory = nullptr,
               torch::executor::EventTracer *event_tracer = nullptr,
-              const LoadBackendOptionsMap *backend_options = nullptr);
+              const LoadBackendOptionsMap *backend_options = nullptr,
+              std::vector<Kernel> kernel_registry = {});
 
   ET_DEPRECATED ET_NODISCARD runtime::Error inline load_method(
       const std::string &method_name,
@@ -294,9 +324,10 @@ class Module {
   ET_NODISCARD inline runtime::Error
   load_forward(runtime::HierarchicalAllocator *planned_memory = nullptr,
                torch::executor::EventTracer *event_tracer = nullptr,
-               const LoadBackendOptionsMap *backend_options = nullptr) {
-    return load_method("forward", planned_memory, event_tracer,
-                       backend_options);
+               const LoadBackendOptionsMap *backend_options = nullptr,
+               std::vector<Kernel> kernel_registry = {}) {
+    return load_method("forward", planned_memory, event_tracer, backend_options,
+                       std::move(kernel_registry));
   }
 
   ET_DEPRECATED ET_NODISCARD inline runtime::Error
@@ -678,6 +709,7 @@ class Module {
     std::unique_ptr<PlannedMemory> planned_memory;
     std::unique_ptr<runtime::MemoryManager> memory_manager;
     std::unique_ptr<Method> method;
+    std::vector<Kernel> kernel_registry;
   };
 
   std::string file_path_;
@@ -693,7 +725,14 @@ class Module {
   std::unique_ptr<NamedDataMap> merged_data_map_;
   std::vector<std::vector<uint8_t>> shared_arenas_;
   ET_DEPRECATED std::vector<uint8_t> debug_buffer_;
-  const LoadBackendOptionsMap *backend_options_ = nullptr;
+  // Module-owned deep-copy of the backend options most recently installed
+  // via load(LoadBackendOptionsMap, ...). `backend_options_storage_` owns
+  // the per-backend BackendOption arrays; `backend_options_map_` is a
+  // LoadBackendOptionsMap whose Spans reference those owned arrays. An
+  // empty map (`size() == 0`) is observationally indistinguishable from
+  // "never set" by downstream consumers, so we don't track that bit.
+  std::vector<std::vector<runtime::BackendOption>> backend_options_storage_;
+  LoadBackendOptionsMap backend_options_map_;
   bool share_memory_arenas_;
 
   ET_NODISCARD runtime::Error
diff --git a/packages/react-native-executorch/third-party/include/executorch/extension/tensor/tensor_ptr.h b/packages/react-native-executorch/third-party/include/executorch/extension/tensor/tensor_ptr.h
index 3dd4e890d3..15321ae608 100644
--- a/packages/react-native-executorch/third-party/include/executorch/extension/tensor/tensor_ptr.h
+++ b/packages/react-native-executorch/third-party/include/executorch/extension/tensor/tensor_ptr.h
@@ -14,6 +14,7 @@
 #include <vector>
 
 #include <c10/macros/Macros.h>
+#include <c10/util/safe_numerics.h>
 #include <executorch/runtime/core/error.h>
 #include <executorch/runtime/core/exec_aten/exec_aten.h>
 #include <executorch/runtime/core/exec_aten/util/scalar_type_util.h>
@@ -105,13 +106,21 @@ make_tensor_ptr(std::vector<executorch::aten::SizesType> sizes,
                 executorch::aten::ScalarType type = deduced_type,
                 executorch::aten::TensorShapeDynamism dynamism =
                     executorch::aten::TensorShapeDynamism::DYNAMIC_BOUND) {
-  ET_CHECK_MSG(data.size() ==
-                   executorch::aten::compute_numel(sizes.data(), sizes.size()),
+  auto numel_result = executorch::aten::safe_numel(sizes.data(), sizes.size());
+  ET_CHECK_MSG(numel_result.ok(), "safe_numel failed: %d",
+               static_cast<int>(numel_result.error()));
+  ET_CHECK_MSG(data.size() == static_cast<size_t>(numel_result.get()),
                "Data size does not match tensor size.");
   if (type != deduced_type) {
     ET_CHECK_MSG(runtime::canCast(deduced_type, type),
                  "Cannot cast deduced type to specified type.");
-    std::vector<uint8_t> casted_data(data.size() * aten::elementSize(type));
+    size_t casted_bytes = 0;
+    ET_CHECK_MSG(!c10::mul_overflows(
+                     data.size(), static_cast<size_t>(aten::elementSize(type)),
+                     &casted_bytes),
+                 "casted_data size overflow: %zu elements * %zu bytes/element",
+                 data.size(), static_cast<size_t>(aten::elementSize(type)));
+    std::vector<uint8_t> casted_data(casted_bytes);
 
     // Create a minimal context for error handling in ET_SWITCH
     struct {
@@ -327,8 +336,11 @@ make_tensor_ptr(const executorch::aten::Tensor &tensor,
   const auto same_rank = sizes.size() == static_cast<size_t>(tensor.dim());
   const auto same_shape = same_rank && std::equal(sizes.begin(), sizes.end(),
                                                   tensor.sizes().begin());
-  const auto element_count =
-      executorch::aten::compute_numel(sizes.data(), sizes.size());
+  auto element_count_result =
+      executorch::aten::safe_numel(sizes.data(), sizes.size());
+  ET_CHECK_MSG(element_count_result.ok(), "safe_numel failed: %d",
+               static_cast<int>(element_count_result.error()));
+  const auto element_count = element_count_result.get();
   const auto parent_element_count = tensor.numel();
   ET_CHECK_MSG(
       element_count <= parent_element_count,
diff --git a/packages/react-native-executorch/third-party/include/executorch/kernels/optimized/Functions.h b/packages/react-native-executorch/third-party/include/executorch/kernels/optimized/Functions.h
index 1b9b4a44b3..cd934e7997 100644
--- a/packages/react-native-executorch/third-party/include/executorch/kernels/optimized/Functions.h
+++ b/packages/react-native-executorch/third-party/include/executorch/kernels/optimized/Functions.h
@@ -91,6 +91,12 @@ TORCH_API inline torch::executor::Tensor & gelu_outf(torch::executor::KernelRunt
 }
 
 
+// aten::grid_sampler_2d.out(Tensor input, Tensor grid, int interpolation_mode, int padding_mode, bool align_corners, *, Tensor(a!) out) -> Tensor(a!)
+TORCH_API inline torch::executor::Tensor & grid_sampler_2d_outf(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & input, const torch::executor::Tensor & grid, int64_t interpolation_mode, int64_t padding_mode, bool align_corners, torch::executor::Tensor & out) {
+    return ::torch::executor::native::opt_grid_sampler_2d_out(context, input, grid, interpolation_mode, padding_mode, align_corners, out);
+}
+
+
 // aten::le.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)
 TORCH_API inline torch::executor::Tensor & le_outf(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, const torch::executor::Scalar & other, torch::executor::Tensor & out) {
     return ::torch::executor::native::opt_le_scalar_out(context, self, other, out);
@@ -139,6 +145,12 @@ TORCH_API inline torch::executor::Tensor & sub_outf(torch::executor::KernelRunti
 }
 
 
+// aten::sum.IntList_out(Tensor self, int[1]? dim, bool keepdim=False, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)
+TORCH_API inline torch::executor::Tensor & sum_outf(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, torch::executor::optional<torch::executor::ArrayRef<int64_t>> dim, bool keepdim, torch::executor::optional<torch::executor::ScalarType> dtype, torch::executor::Tensor & out) {
+    return ::torch::executor::native::opt_sum_dim_out(context, self, dim, keepdim, dtype, out);
+}
+
+
 // aten::sub.Scalar_out(Tensor self, Scalar other, Scalar alpha=1, *, Tensor(a!) out) -> Tensor(a!)
 TORCH_API inline torch::executor::Tensor & sub_outf(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, const torch::executor::Scalar & other, const torch::executor::Scalar & alpha, torch::executor::Tensor & out) {
     return ::torch::executor::native::opt_sub_scalar_out(context, self, other, alpha, out);
diff --git a/packages/react-native-executorch/third-party/include/executorch/kernels/optimized/NativeFunctions.h b/packages/react-native-executorch/third-party/include/executorch/kernels/optimized/NativeFunctions.h
index 3e83af3688..6dff2dae23 100644
--- a/packages/react-native-executorch/third-party/include/executorch/kernels/optimized/NativeFunctions.h
+++ b/packages/react-native-executorch/third-party/include/executorch/kernels/optimized/NativeFunctions.h
@@ -42,6 +42,8 @@ torch::executor::Tensor & opt_exp_out(const torch::executor::Tensor & self, torc
 torch::executor::Tensor & opt_exp_out(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, torch::executor::Tensor & out);
 torch::executor::Tensor & opt_gelu_out(const torch::executor::Tensor & self, torch::executor::string_view approximate, torch::executor::Tensor & out);
 torch::executor::Tensor & opt_gelu_out(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, torch::executor::string_view approximate, torch::executor::Tensor & out);
+torch::executor::Tensor & opt_grid_sampler_2d_out(const torch::executor::Tensor & input, const torch::executor::Tensor & grid, int64_t interpolation_mode, int64_t padding_mode, bool align_corners, torch::executor::Tensor & out);
+torch::executor::Tensor & opt_grid_sampler_2d_out(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & input, const torch::executor::Tensor & grid, int64_t interpolation_mode, int64_t padding_mode, bool align_corners, torch::executor::Tensor & out);
 torch::executor::Tensor & opt_le_scalar_out(const torch::executor::Tensor & self, const torch::executor::Scalar & other, torch::executor::Tensor & out);
 torch::executor::Tensor & opt_le_scalar_out(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, const torch::executor::Scalar & other, torch::executor::Tensor & out);
 torch::executor::Tensor & opt_le_tensor_out(const torch::executor::Tensor & self, const torch::executor::Tensor & other, torch::executor::Tensor & out);
@@ -58,6 +60,8 @@ ::std::tuple<torch::executor::Tensor &,torch::executor::Tensor &,torch::executor
 ::std::tuple<torch::executor::Tensor &,torch::executor::Tensor &,torch::executor::Tensor &> opt_native_layer_norm_out(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & input, torch::executor::ArrayRef<int64_t> normalized_shape, const torch::executor::optional<torch::executor::Tensor> & weight, const torch::executor::optional<torch::executor::Tensor> & bias, double eps, torch::executor::Tensor & out0, torch::executor::Tensor & out1, torch::executor::Tensor & out2);
 torch::executor::Tensor & opt_sub_out(const torch::executor::Tensor & self, const torch::executor::Tensor & other, const torch::executor::Scalar & alpha, torch::executor::Tensor & out);
 torch::executor::Tensor & opt_sub_out(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, const torch::executor::Tensor & other, const torch::executor::Scalar & alpha, torch::executor::Tensor & out);
+torch::executor::Tensor & opt_sum_dim_out(const torch::executor::Tensor & self, torch::executor::optional<torch::executor::ArrayRef<int64_t>> dim, bool keepdim, torch::executor::optional<torch::executor::ScalarType> dtype, torch::executor::Tensor & out);
+torch::executor::Tensor & opt_sum_dim_out(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, torch::executor::optional<torch::executor::ArrayRef<int64_t>> dim, bool keepdim, torch::executor::optional<torch::executor::ScalarType> dtype, torch::executor::Tensor & out);
 torch::executor::Tensor & opt_sub_scalar_out(const torch::executor::Tensor & self, const torch::executor::Scalar & other, const torch::executor::Scalar & alpha, torch::executor::Tensor & out);
 torch::executor::Tensor & opt_sub_scalar_out(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, const torch::executor::Scalar & other, const torch::executor::Scalar & alpha, torch::executor::Tensor & out);
 torch::executor::Tensor & opt_where_out(const torch::executor::Tensor & condition, const torch::executor::Tensor & self, const torch::executor::Tensor & other, torch::executor::Tensor & out);
diff --git a/packages/react-native-executorch/third-party/include/executorch/kernels/portable/Functions.h b/packages/react-native-executorch/third-party/include/executorch/kernels/portable/Functions.h
index b1911d7ae2..c96f3daf4b 100644
--- a/packages/react-native-executorch/third-party/include/executorch/kernels/portable/Functions.h
+++ b/packages/react-native-executorch/third-party/include/executorch/kernels/portable/Functions.h
@@ -25,12 +25,24 @@ namespace executor {
 
 namespace aten {
 
+// aten::_adaptive_avg_pool2d.out(Tensor self, SymInt[2] output_size, *, Tensor(a!) out) -> Tensor(a!)
+TORCH_API inline torch::executor::Tensor & _adaptive_avg_pool2d_outf(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, torch::executor::ArrayRef<int64_t> output_size, torch::executor::Tensor & out) {
+    return ::torch::executor::native::_adaptive_avg_pool2d_out(context, self, output_size, out);
+}
+
+
 // aten::_cdist_forward.out(Tensor x1, Tensor x2, float p, int? compute_mode, *, Tensor(a!) out) -> Tensor(a!)
 TORCH_API inline torch::executor::Tensor & _cdist_forward_outf(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & x1, const torch::executor::Tensor & x2, double p, torch::executor::optional<int64_t> compute_mode, torch::executor::Tensor & out) {
     return ::torch::executor::native::_cdist_forward_out(context, x1, x2, p, compute_mode, out);
 }
 
 
+// aten::_conj_physical.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+TORCH_API inline torch::executor::Tensor & _conj_physical_outf(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, torch::executor::Tensor & out) {
+    return ::torch::executor::native::_conj_physical_out(context, self, out);
+}
+
+
 // aten::_log_softmax.out(Tensor self, int dim, bool half_to_float, *, Tensor(a!) out) -> Tensor(a!)
 TORCH_API inline torch::executor::Tensor & _log_softmax_outf(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, int64_t dim, bool half_to_float, torch::executor::Tensor & out) {
     return ::torch::executor::native::log_softmax_out(context, self, dim, half_to_float, out);
@@ -1201,6 +1213,12 @@ TORCH_API inline torch::executor::Tensor & var_outf(torch::executor::KernelRunti
 }
 
 
+// aten::var_mean.correction_out(Tensor self, int[1]? dim=None, *, Scalar? correction=None, bool keepdim=False, Tensor(a!) out0, Tensor(b!) out1) -> (Tensor(a!), Tensor(b!))
+TORCH_API inline ::std::tuple<torch::executor::Tensor &,torch::executor::Tensor &> var_mean_outf(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, torch::executor::optional<torch::executor::ArrayRef<int64_t>> dim, const torch::executor::optional<torch::executor::Scalar> & correction, bool keepdim, torch::executor::Tensor & out0, torch::executor::Tensor & out1) {
+    return ::torch::executor::native::var_mean_correction_out(context, self, dim, correction, keepdim, out0, out1);
+}
+
+
 // aten::var.out(Tensor self, int[1]? dim, bool unbiased=True, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
 TORCH_API inline torch::executor::Tensor & var_outf(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, torch::executor::optional<torch::executor::ArrayRef<int64_t>> dim, bool unbiased, bool keepdim, torch::executor::Tensor & out) {
     return ::torch::executor::native::var_out(context, self, dim, unbiased, keepdim, out);
diff --git a/packages/react-native-executorch/third-party/include/executorch/kernels/portable/NativeFunctions.h b/packages/react-native-executorch/third-party/include/executorch/kernels/portable/NativeFunctions.h
index c36896b7a1..e2a6cc3f52 100644
--- a/packages/react-native-executorch/third-party/include/executorch/kernels/portable/NativeFunctions.h
+++ b/packages/react-native-executorch/third-party/include/executorch/kernels/portable/NativeFunctions.h
@@ -20,8 +20,12 @@
 namespace torch {
 namespace executor {
 namespace native {
+torch::executor::Tensor & _adaptive_avg_pool2d_out(const torch::executor::Tensor & self, torch::executor::ArrayRef<int64_t> output_size, torch::executor::Tensor & out);
+torch::executor::Tensor & _adaptive_avg_pool2d_out(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, torch::executor::ArrayRef<int64_t> output_size, torch::executor::Tensor & out);
 torch::executor::Tensor & _cdist_forward_out(const torch::executor::Tensor & x1, const torch::executor::Tensor & x2, double p, torch::executor::optional<int64_t> compute_mode, torch::executor::Tensor & out);
 torch::executor::Tensor & _cdist_forward_out(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & x1, const torch::executor::Tensor & x2, double p, torch::executor::optional<int64_t> compute_mode, torch::executor::Tensor & out);
+torch::executor::Tensor & _conj_physical_out(const torch::executor::Tensor & self, torch::executor::Tensor & out);
+torch::executor::Tensor & _conj_physical_out(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, torch::executor::Tensor & out);
 torch::executor::Tensor & log_softmax_out(const torch::executor::Tensor & self, int64_t dim, bool half_to_float, torch::executor::Tensor & out);
 torch::executor::Tensor & log_softmax_out(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, int64_t dim, bool half_to_float, torch::executor::Tensor & out);
 ::std::tuple<torch::executor::Tensor &,torch::executor::Tensor &,torch::executor::Tensor &> _native_batch_norm_legit_out(const torch::executor::Tensor & input, const torch::executor::optional<torch::executor::Tensor> & weight, const torch::executor::optional<torch::executor::Tensor> & bias, torch::executor::Tensor & running_mean, torch::executor::Tensor & running_var, bool training, double momentum, double eps, torch::executor::Tensor & out, torch::executor::Tensor & save_mean, torch::executor::Tensor & save_invstd);
@@ -412,6 +416,8 @@ torch::executor::Tensor & upsample_nearest2d_vec_out(const torch::executor::Tens
 torch::executor::Tensor & upsample_nearest2d_vec_out(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & input, torch::executor::optional<torch::executor::ArrayRef<int64_t>> output_size, torch::executor::optional<torch::executor::ArrayRef<double>> scale_factors, torch::executor::Tensor & out);
 torch::executor::Tensor & var_correction_out(const torch::executor::Tensor & self, torch::executor::optional<torch::executor::ArrayRef<int64_t>> dim, const torch::executor::optional<torch::executor::Scalar> & correction, bool keepdim, torch::executor::Tensor & out);
 torch::executor::Tensor & var_correction_out(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, torch::executor::optional<torch::executor::ArrayRef<int64_t>> dim, const torch::executor::optional<torch::executor::Scalar> & correction, bool keepdim, torch::executor::Tensor & out);
+::std::tuple<torch::executor::Tensor &,torch::executor::Tensor &> var_mean_correction_out(const torch::executor::Tensor & self, torch::executor::optional<torch::executor::ArrayRef<int64_t>> dim, const torch::executor::optional<torch::executor::Scalar> & correction, bool keepdim, torch::executor::Tensor & out0, torch::executor::Tensor & out1);
+::std::tuple<torch::executor::Tensor &,torch::executor::Tensor &> var_mean_correction_out(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, torch::executor::optional<torch::executor::ArrayRef<int64_t>> dim, const torch::executor::optional<torch::executor::Scalar> & correction, bool keepdim, torch::executor::Tensor & out0, torch::executor::Tensor & out1);
 torch::executor::Tensor & var_out(const torch::executor::Tensor & self, torch::executor::optional<torch::executor::ArrayRef<int64_t>> dim, bool unbiased, bool keepdim, torch::executor::Tensor & out);
 torch::executor::Tensor & var_out(torch::executor::KernelRuntimeContext & context, const torch::executor::Tensor & self, torch::executor::optional<torch::executor::ArrayRef<int64_t>> dim, bool unbiased, bool keepdim, torch::executor::Tensor & out);
 torch::executor::Tensor & view_as_real_copy_out(const torch::executor::Tensor & self, torch::executor::Tensor & out);
diff --git a/packages/react-native-executorch/third-party/include/executorch/runtime/backend/backend_options_map.h b/packages/react-native-executorch/third-party/include/executorch/runtime/backend/backend_options_map.h
index 8ae9543c29..5acf6cca2d 100644
--- a/packages/react-native-executorch/third-party/include/executorch/runtime/backend/backend_options_map.h
+++ b/packages/react-native-executorch/third-party/include/executorch/runtime/backend/backend_options_map.h
@@ -11,6 +11,7 @@
 #include <executorch/runtime/backend/options.h>
 #include <executorch/runtime/core/error.h>
 #include <executorch/runtime/core/span.h>
+#include <executorch/runtime/platform/assert.h>
 
 #include <cstring>
 
@@ -168,6 +169,42 @@ class LoadBackendOptionsMap final {
    */
   size_t size() const { return size_; }
 
+  /**
+   * Non-owning view of a single (backend_id, options) entry, returned by
+   * entry_at(). The pointer / span are valid until the map is mutated or
+   * destroyed.
+   */
+  struct EntryView {
+    const char *backend_id = nullptr;
+    Span<const BackendOption> options;
+  };
+
+  /**
+   * Returns the (backend_id, options) entry at the given index for
+   * enumeration over the map's contents.
+   *
+   * @param index The entry index. Must be < size(); behavior is undefined
+   *     otherwise. Use this together with size() to walk every entry.
+   * @return EntryView referencing the entry's backend_id and options. The
+   *     view is valid until the next mutation of, or destruction of, this
+   *     map.
+   *
+   * Example:
+   * @code
+   *   for (size_t i = 0; i < map.size(); ++i) {
+   *     const auto entry = map.entry_at(i);
+   *     // use entry.backend_id and entry.options ...
+   *   }
+   * @endcode
+   */
+  EntryView entry_at(size_t index) const {
+    ET_DCHECK_MSG(index < size_, "entry_at index %zu out of bounds (size=%zu)",
+                  index, size_);
+    return EntryView{entries_[index].backend_id,
+                     Span<const BackendOption>(entries_[index].options.data(),
+                                               entries_[index].options.size())};
+  }
+
 private:
   static constexpr size_t kMaxBackends = 8;
   static constexpr size_t kMaxBackendIdLength = 64;
diff --git a/packages/react-native-executorch/third-party/include/executorch/runtime/core/array_ref.h b/packages/react-native-executorch/third-party/include/executorch/runtime/core/array_ref.h
index 01dd581663..867520137f 100644
--- a/packages/react-native-executorch/third-party/include/executorch/runtime/core/array_ref.h
+++ b/packages/react-native-executorch/third-party/include/executorch/runtime/core/array_ref.h
@@ -30,6 +30,7 @@
 #include <cstdint>
 
 #include <c10/util/irange.h>
+#include <c10/util/safe_numerics.h>
 #include <executorch/runtime/platform/assert.h>
 
 namespace executorch {
@@ -146,7 +147,8 @@ template <typename T> class ArrayRef final {
   /// slice(n, m) - Take M elements of the array starting at element N
   ArrayRef<T> slice(size_t N, size_t M) const {
     // cant slice longer then the array
-    ET_CHECK(N + M <= size());
+    size_t end = 0;
+    ET_CHECK(!c10::add_overflows(N, M, &end) && end <= size());
     return ArrayRef<T>(data() + N, M);
   }
 
diff --git a/packages/react-native-executorch/third-party/include/executorch/runtime/core/error.h b/packages/react-native-executorch/third-party/include/executorch/runtime/core/error.h
index 6500a6f907..c107dd43db 100644
--- a/packages/react-native-executorch/third-party/include/executorch/runtime/core/error.h
+++ b/packages/react-native-executorch/third-party/include/executorch/runtime/core/error.h
@@ -152,6 +152,7 @@ constexpr const char *to_string(const Error error) {
   case Error::RegistrationAlreadyRegistered:
     return "Error::RegistrationAlreadyRegistered";
   }
+  return "Error::Unknown";
 }
 
 } // namespace runtime
diff --git a/packages/react-native-executorch/third-party/include/executorch/runtime/core/evalue.h b/packages/react-native-executorch/third-party/include/executorch/runtime/core/evalue.h
index b3b0689ba7..2e8bb5389e 100644
--- a/packages/react-native-executorch/third-party/include/executorch/runtime/core/evalue.h
+++ b/packages/react-native-executorch/third-party/include/executorch/runtime/core/evalue.h
@@ -8,6 +8,7 @@
 
 #pragma once
 #include <executorch/runtime/core/exec_aten/exec_aten.h>
+#include <executorch/runtime/core/result.h>
 #include <executorch/runtime/core/tag.h>
 #include <executorch/runtime/platform/assert.h>
 
@@ -67,6 +68,29 @@ template <typename T> class BoxedEvalueList {
    */
   executorch::aten::ArrayRef<T> get() const;
 
+  /**
+   * Result-returning counterpart of get(). Validates each wrapped EValue's
+   * tag before materializing; returns Error::InvalidType if any element's
+   * tag does not match T and Error::InvalidState if any element pointer is
+   * null. Use this when materializing lists from untrusted .pte data so that
+   * a malformed program cannot force a process abort inside to<T>() /
+   * ET_CHECK.
+   */
+  Result<executorch::aten::ArrayRef<T>> tryGet() const;
+
+  /**
+   * Destroys the unwrapped elements without re-dereferencing wrapped_vals_.
+   * This is safe to call during EValue destruction because it does not
+   * dereference wrapped_vals_, which may point to EValues mutated by
+   * MoveCall instructions.
+   */
+  void destroy_elements() {
+    for (typename executorch::aten::ArrayRef<T>::size_type i = 0;
+         i < wrapped_vals_.size(); i++) {
+      unwrapped_vals_[i].~T();
+    }
+  }
+
 private:
   static EValue **checkWrappedVals(EValue **wrapped_vals, int size) {
     ET_CHECK_MSG(wrapped_vals != nullptr, "wrapped_vals cannot be null");
@@ -89,6 +113,10 @@ template <>
 executorch::aten::ArrayRef<std::optional<executorch::aten::Tensor>>
 BoxedEvalueList<std::optional<executorch::aten::Tensor>>::get() const;
 
+template <>
+Result<executorch::aten::ArrayRef<std::optional<executorch::aten::Tensor>>>
+BoxedEvalueList<std::optional<executorch::aten::Tensor>>::tryGet() const;
+
 // Aggregate typing system similar to IValue only slimmed down with less
 // functionality, no dependencies on atomic, and fewer supported types to better
 // suit embedded systems (ie no intrusive ptr)
@@ -165,6 +193,13 @@ struct EValue {
     return payload.copyable_union.as_int;
   }
 
+  Result<int64_t> tryToInt() const {
+    if (!isInt()) {
+      return Error::InvalidType;
+    }
+    return payload.copyable_union.as_int;
+  }
+
   /****** Double Type ******/
   /*implicit*/ EValue(double d) : tag(Tag::Double) {
     payload.copyable_union.as_double = d;
@@ -177,6 +212,13 @@ struct EValue {
     return payload.copyable_union.as_double;
   }
 
+  Result<double> tryToDouble() const {
+    if (!isDouble()) {
+      return Error::InvalidType;
+    }
+    return payload.copyable_union.as_double;
+  }
+
   /****** Bool Type ******/
   /*implicit*/ EValue(bool b) : tag(Tag::Bool) {
     payload.copyable_union.as_bool = b;
@@ -189,6 +231,13 @@ struct EValue {
     return payload.copyable_union.as_bool;
   }
 
+  Result<bool> tryToBool() const {
+    if (!isBool()) {
+      return Error::InvalidType;
+    }
+    return payload.copyable_union.as_bool;
+  }
+
   /****** Scalar Type ******/
   /// Construct an EValue using the implicit value of a Scalar.
   /*implicit*/ EValue(executorch::aten::Scalar s) {
@@ -224,6 +273,19 @@ struct EValue {
     }
   }
 
+  Result<executorch::aten::Scalar> tryToScalar() const {
+    if (isDouble()) {
+      return executorch::aten::Scalar(payload.copyable_union.as_double);
+    }
+    if (isInt()) {
+      return executorch::aten::Scalar(payload.copyable_union.as_int);
+    }
+    if (isBool()) {
+      return executorch::aten::Scalar(payload.copyable_union.as_bool);
+    }
+    return Error::InvalidType;
+  }
+
   /****** Tensor Type ******/
   /*implicit*/ EValue(executorch::aten::Tensor t) : tag(Tag::Tensor) {
     // When built in aten mode, at::Tensor has a non trivial constructor
@@ -270,6 +332,16 @@ struct EValue {
     return payload.as_tensor;
   }
 
+  // Returns a copy of the Tensor handle (one intrusive_ptr refcount bump in
+  // ATen mode; free in lean mode). Unlike toTensor()'s const& / & overloads,
+  // tryToTensor() cannot return a reference — Result<T> wraps by value.
+  Result<executorch::aten::Tensor> tryToTensor() const {
+    if (!isTensor()) {
+      return Error::InvalidType;
+    }
+    return payload.as_tensor;
+  }
+
   /****** String Type ******/
   /*implicit*/ EValue(executorch::aten::ArrayRef<char> *s) : tag(Tag::String) {
     ET_CHECK_MSG(s != nullptr, "ArrayRef<char> pointer cannot be null");
@@ -286,6 +358,17 @@ struct EValue {
                             payload.copyable_union.as_string_ptr->size());
   }
 
+  Result<std::string_view> tryToString() const {
+    if (!isString()) {
+      return Error::InvalidType;
+    }
+    if (payload.copyable_union.as_string_ptr == nullptr) {
+      return Error::InvalidState;
+    }
+    return std::string_view(payload.copyable_union.as_string_ptr->data(),
+                            payload.copyable_union.as_string_ptr->size());
+  }
+
   /****** Int List Type ******/
   /*implicit*/ EValue(BoxedEvalueList<int64_t> *i) : tag(Tag::ListInt) {
     ET_CHECK_MSG(i != nullptr,
@@ -302,6 +385,16 @@ struct EValue {
     return (payload.copyable_union.as_int_list_ptr)->get();
   }
 
+  Result<executorch::aten::ArrayRef<int64_t>> tryToIntList() const {
+    if (!isIntList()) {
+      return Error::InvalidType;
+    }
+    if (payload.copyable_union.as_int_list_ptr == nullptr) {
+      return Error::InvalidState;
+    }
+    return (payload.copyable_union.as_int_list_ptr)->tryGet();
+  }
+
   /****** Bool List Type ******/
   /*implicit*/ EValue(executorch::aten::ArrayRef<bool> *b)
       : tag(Tag::ListBool) {
@@ -318,6 +411,16 @@ struct EValue {
     return *(payload.copyable_union.as_bool_list_ptr);
   }
 
+  Result<executorch::aten::ArrayRef<bool>> tryToBoolList() const {
+    if (!isBoolList()) {
+      return Error::InvalidType;
+    }
+    if (payload.copyable_union.as_bool_list_ptr == nullptr) {
+      return Error::InvalidState;
+    }
+    return *(payload.copyable_union.as_bool_list_ptr);
+  }
+
   /****** Double List Type ******/
   /*implicit*/ EValue(executorch::aten::ArrayRef<double> *d)
       : tag(Tag::ListDouble) {
@@ -334,6 +437,16 @@ struct EValue {
     return *(payload.copyable_union.as_double_list_ptr);
   }
 
+  Result<executorch::aten::ArrayRef<double>> tryToDoubleList() const {
+    if (!isDoubleList()) {
+      return Error::InvalidType;
+    }
+    if (payload.copyable_union.as_double_list_ptr == nullptr) {
+      return Error::InvalidState;
+    }
+    return *(payload.copyable_union.as_double_list_ptr);
+  }
+
   /****** Tensor List Type ******/
   /*implicit*/ EValue(BoxedEvalueList<executorch::aten::Tensor> *t)
       : tag(Tag::ListTensor) {
@@ -351,6 +464,17 @@ struct EValue {
     return payload.copyable_union.as_tensor_list_ptr->get();
   }
 
+  Result<executorch::aten::ArrayRef<executorch::aten::Tensor>>
+  tryToTensorList() const {
+    if (!isTensorList()) {
+      return Error::InvalidType;
+    }
+    if (payload.copyable_union.as_tensor_list_ptr == nullptr) {
+      return Error::InvalidState;
+    }
+    return payload.copyable_union.as_tensor_list_ptr->tryGet();
+  }
+
   /****** List Optional Tensor Type ******/
   /*implicit*/ EValue(
       BoxedEvalueList<std::optional<executorch::aten::Tensor>> *t)
@@ -371,6 +495,17 @@ struct EValue {
     return payload.copyable_union.as_list_optional_tensor_ptr->get();
   }
 
+  Result<executorch::aten::ArrayRef<std::optional<executorch::aten::Tensor>>>
+  tryToListOptionalTensor() const {
+    if (!isListOptionalTensor()) {
+      return Error::InvalidType;
+    }
+    if (payload.copyable_union.as_list_optional_tensor_ptr == nullptr) {
+      return Error::InvalidState;
+    }
+    return payload.copyable_union.as_list_optional_tensor_ptr->tryGet();
+  }
+
   /****** ScalarType Type ******/
   executorch::aten::ScalarType toScalarType() const {
     ET_CHECK_MSG(isInt(), "EValue is not a ScalarType.");
@@ -378,6 +513,14 @@ struct EValue {
         payload.copyable_union.as_int);
   }
 
+  Result<executorch::aten::ScalarType> tryToScalarType() const {
+    if (!isInt()) {
+      return Error::InvalidType;
+    }
+    return static_cast<executorch::aten::ScalarType>(
+        payload.copyable_union.as_int);
+  }
+
   /****** MemoryFormat Type ******/
   executorch::aten::MemoryFormat toMemoryFormat() const {
     ET_CHECK_MSG(isInt(), "EValue is not a MemoryFormat.");
@@ -385,12 +528,27 @@ struct EValue {
         payload.copyable_union.as_int);
   }
 
+  Result<executorch::aten::MemoryFormat> tryToMemoryFormat() const {
+    if (!isInt()) {
+      return Error::InvalidType;
+    }
+    return static_cast<executorch::aten::MemoryFormat>(
+        payload.copyable_union.as_int);
+  }
+
   /****** Layout Type ******/
   executorch::aten::Layout toLayout() const {
     ET_CHECK_MSG(isInt(), "EValue is not a Layout.");
     return static_cast<executorch::aten::Layout>(payload.copyable_union.as_int);
   }
 
+  Result<executorch::aten::Layout> tryToLayout() const {
+    if (!isInt()) {
+      return Error::InvalidType;
+    }
+    return static_cast<executorch::aten::Layout>(payload.copyable_union.as_int);
+  }
+
   /****** Device Type ******/
   executorch::aten::Device toDevice() const {
     ET_CHECK_MSG(isInt(), "EValue is not a Device.");
@@ -399,12 +557,29 @@ struct EValue {
                                     -1);
   }
 
+  Result<executorch::aten::Device> tryToDevice() const {
+    if (!isInt()) {
+      return Error::InvalidType;
+    }
+    return executorch::aten::Device(static_cast<executorch::aten::DeviceType>(
+                                        payload.copyable_union.as_int),
+                                    -1);
+  }
+
   template <typename T> T to() &&;
   template <typename T>
   typename internal::evalue_to_const_ref_overload_return<T>::type to() const &;
   template <typename T>
   typename internal::evalue_to_ref_overload_return<T>::type to() &;
 
+  /**
+   * Result-returning equivalent of `to<T>()`. Tag mismatch returns
+   * `Error::InvalidType`; a null list/string payload returns
+   * `Error::InvalidState`. Specializations are defined below via
+   * `EVALUE_DEFINE_TRY_TO`.
+   */
+  template <typename T> Result<T> tryTo() const;
+
   /**
    * Converts the EValue to an optional object that can represent both T and
    * an uninitialized state.
@@ -416,6 +591,22 @@ struct EValue {
     return this->to<T>();
   }
 
+  /**
+   * Result-returning equivalent of `toOptional<T>()`. None maps to an empty
+   * optional; any other tag that doesn't match T propagates `tryTo<T>()`'s
+   * error (`Error::InvalidType`).
+   */
+  template <typename T> inline Result<std::optional<T>> tryToOptional() const {
+    if (this->isNone()) {
+      return std::optional<T>(std::nullopt);
+    }
+    auto r = this->tryTo<T>();
+    if (!r.ok()) {
+      return r.error();
+    }
+    return std::optional<T>(std::move(r.get()));
+  }
+
 private:
   // Pre cond: the payload value has had its destructor called
   void clearToNone() noexcept {
@@ -446,17 +637,10 @@ struct EValue {
       payload.as_tensor.~Tensor();
     } else if (isTensorList() &&
                payload.copyable_union.as_tensor_list_ptr != nullptr) {
-      // for (auto& tensor : toTensorList()) {
-      for (auto &tensor : payload.copyable_union.as_tensor_list_ptr->get()) {
-        tensor.~Tensor();
-      }
+      payload.copyable_union.as_tensor_list_ptr->destroy_elements();
     } else if (isListOptionalTensor() &&
                payload.copyable_union.as_list_optional_tensor_ptr != nullptr) {
-      // for (auto& optional_tensor : toListOptionalTensor()) {
-      for (auto &optional_tensor :
-           payload.copyable_union.as_list_optional_tensor_ptr->get()) {
-        optional_tensor.~optional();
-      }
+      payload.copyable_union.as_list_optional_tensor_ptr->destroy_elements();
     }
   }
 
@@ -532,6 +716,53 @@ EVALUE_DEFINE_TO(
     toListOptionalTensor)
 #undef EVALUE_DEFINE_TO
 
+#define EVALUE_DEFINE_TRY_TO(T, method_name)                                   \
+  template <> inline Result<T> EValue::tryTo<T>() const {                      \
+    return this->method_name();                                                \
+  }
+
+EVALUE_DEFINE_TRY_TO(executorch::aten::Scalar, tryToScalar)
+EVALUE_DEFINE_TRY_TO(int64_t, tryToInt)
+EVALUE_DEFINE_TRY_TO(bool, tryToBool)
+EVALUE_DEFINE_TRY_TO(double, tryToDouble)
+EVALUE_DEFINE_TRY_TO(std::string_view, tryToString)
+EVALUE_DEFINE_TRY_TO(executorch::aten::ScalarType, tryToScalarType)
+EVALUE_DEFINE_TRY_TO(executorch::aten::MemoryFormat, tryToMemoryFormat)
+EVALUE_DEFINE_TRY_TO(executorch::aten::Layout, tryToLayout)
+EVALUE_DEFINE_TRY_TO(executorch::aten::Device, tryToDevice)
+// Tensor and Optional Tensor
+EVALUE_DEFINE_TRY_TO(executorch::aten::Tensor, tryToTensor)
+EVALUE_DEFINE_TRY_TO(std::optional<executorch::aten::Tensor>,
+                     tryToOptional<executorch::aten::Tensor>)
+
+// IntList and Optional IntList
+EVALUE_DEFINE_TRY_TO(executorch::aten::ArrayRef<int64_t>, tryToIntList)
+EVALUE_DEFINE_TRY_TO(std::optional<executorch::aten::ArrayRef<int64_t>>,
+                     tryToOptional<executorch::aten::ArrayRef<int64_t>>)
+
+// DoubleList and Optional DoubleList
+EVALUE_DEFINE_TRY_TO(executorch::aten::ArrayRef<double>, tryToDoubleList)
+EVALUE_DEFINE_TRY_TO(std::optional<executorch::aten::ArrayRef<double>>,
+                     tryToOptional<executorch::aten::ArrayRef<double>>)
+
+// BoolList and Optional BoolList
+EVALUE_DEFINE_TRY_TO(executorch::aten::ArrayRef<bool>, tryToBoolList)
+EVALUE_DEFINE_TRY_TO(std::optional<executorch::aten::ArrayRef<bool>>,
+                     tryToOptional<executorch::aten::ArrayRef<bool>>)
+
+// TensorList and Optional TensorList
+EVALUE_DEFINE_TRY_TO(executorch::aten::ArrayRef<executorch::aten::Tensor>,
+                     tryToTensorList)
+EVALUE_DEFINE_TRY_TO(
+    std::optional<executorch::aten::ArrayRef<executorch::aten::Tensor>>,
+    tryToOptional<executorch::aten::ArrayRef<executorch::aten::Tensor>>)
+
+// List of Optional Tensor
+EVALUE_DEFINE_TRY_TO(
+    executorch::aten::ArrayRef<std::optional<executorch::aten::Tensor>>,
+    tryToListOptionalTensor)
+#undef EVALUE_DEFINE_TRY_TO
+
 template <typename T>
 executorch::aten::ArrayRef<T> BoxedEvalueList<T>::get() const {
   for (typename executorch::aten::ArrayRef<T>::size_type i = 0;
@@ -542,6 +773,22 @@ executorch::aten::ArrayRef<T> BoxedEvalueList<T>::get() const {
   return executorch::aten::ArrayRef<T>{unwrapped_vals_, wrapped_vals_.size()};
 }
 
+template <typename T>
+Result<executorch::aten::ArrayRef<T>> BoxedEvalueList<T>::tryGet() const {
+  for (typename executorch::aten::ArrayRef<T>::size_type i = 0;
+       i < wrapped_vals_.size(); i++) {
+    if (wrapped_vals_[i] == nullptr) {
+      return Error::InvalidState;
+    }
+    auto r = wrapped_vals_[i]->template tryTo<T>();
+    if (!r.ok()) {
+      return r.error();
+    }
+    unwrapped_vals_[i] = std::move(r.get());
+  }
+  return executorch::aten::ArrayRef<T>{unwrapped_vals_, wrapped_vals_.size()};
+}
+
 } // namespace runtime
 } // namespace executorch
 
diff --git a/packages/react-native-executorch/third-party/include/executorch/runtime/core/exec_aten/exec_aten.h b/packages/react-native-executorch/third-party/include/executorch/runtime/core/exec_aten/exec_aten.h
index b4a1b2721f..db5096a46a 100644
--- a/packages/react-native-executorch/third-party/include/executorch/runtime/core/exec_aten/exec_aten.h
+++ b/packages/react-native-executorch/third-party/include/executorch/runtime/core/exec_aten/exec_aten.h
@@ -8,7 +8,10 @@
 
 #pragma once
 
+#include <executorch/runtime/core/error.h>                 // @manual
+#include <executorch/runtime/core/result.h>                // @manual
 #include <executorch/runtime/core/tensor_shape_dynamism.h> // @manual
+#include <executorch/runtime/platform/assert.h>            // @manual
 #include <executorch/runtime/platform/compiler.h>
 #ifdef USE_ATEN_LIB
 #include <ATen/Tensor.h> // @manual
@@ -28,6 +31,7 @@
 #include <c10/util/quint2x4.h>      // @manual
 #include <c10/util/quint4x2.h>      // @manual
 #include <c10/util/quint8.h>        // @manual
+#include <c10/util/safe_numerics.h> // @manual
 #include <c10/util/string_view.h>   // @manual
 #include <torch/torch.h>
 #else                                                            // use executor
@@ -107,6 +111,25 @@ inline ssize_t compute_numel(const SizesType *sizes, ssize_t dim) {
       c10::multiply_integers(c10::ArrayRef<SizesType>(sizes, dim)));
 }
 
+inline ::executorch::runtime::Result<ssize_t> safe_numel(const SizesType *sizes,
+                                                         ssize_t dim) {
+  ET_CHECK_OR_RETURN_ERROR(dim == 0 || sizes != nullptr, InvalidArgument,
+                           "Sizes must be provided for non-scalar tensors");
+  ssize_t numel = 1;
+  for (ssize_t i = 0; i < dim; i++) {
+    ET_CHECK_OR_RETURN_ERROR(
+        sizes[i] >= 0, InvalidArgument,
+        "Size must be non-negative, got %zd at dimension %zd",
+        static_cast<ssize_t>(sizes[i]), i);
+    ssize_t next_numel;
+    ET_CHECK_OR_RETURN_ERROR(
+        !c10::mul_overflows(numel, static_cast<ssize_t>(sizes[i]), &next_numel),
+        InvalidArgument, "Overflow computing numel at dimension %zd", i);
+    numel = next_numel;
+  }
+  return numel;
+}
+
 #undef ET_PRI_TENSOR_SIZE
 #define ET_PRI_TENSOR_SIZE PRId64
 
@@ -153,6 +176,7 @@ using OptionalArrayRef =
 using OptionalIntArrayRef = OptionalArrayRef<int64_t>;
 
 using torch::executor::compute_numel;
+using torch::executor::safe_numel;
 
 #endif // Use ExecuTorch types
 
diff --git a/packages/react-native-executorch/third-party/include/executorch/runtime/core/hierarchical_allocator.h b/packages/react-native-executorch/third-party/include/executorch/runtime/core/hierarchical_allocator.h
index 9e2335501b..e061c9d182 100644
--- a/packages/react-native-executorch/third-party/include/executorch/runtime/core/hierarchical_allocator.h
+++ b/packages/react-native-executorch/third-party/include/executorch/runtime/core/hierarchical_allocator.h
@@ -9,6 +9,7 @@
 #pragma once
 
 #include <c10/util/irange.h>
+#include <c10/util/safe_numerics.h>
 
 #include <executorch/runtime/core/memory_allocator.h>
 #include <executorch/runtime/core/result.h>
@@ -56,17 +57,19 @@ class HierarchicalAllocator final {
                                                  size_t offset_bytes,
                                                  size_t size_bytes) {
     // Check for integer overflow in offset_bytes + size_bytes.
-    ET_CHECK_OR_RETURN_ERROR(size_bytes <= SIZE_MAX - offset_bytes,
-                             InvalidArgument,
-                             "Integer overflow in offset_bytes (%" ET_PRIsize_t
-                             ") + size_bytes (%" ET_PRIsize_t ")",
-                             offset_bytes, size_bytes);
+    size_t end_bytes = 0;
+    ET_CHECK_OR_RETURN_ERROR(
+        !c10::add_overflows(offset_bytes, size_bytes, &end_bytes),
+        InvalidArgument,
+        "Integer overflow in offset_bytes (%" ET_PRIsize_t
+        ") + size_bytes (%" ET_PRIsize_t ")",
+        offset_bytes, size_bytes);
     ET_CHECK_OR_RETURN_ERROR(memory_id < buffers_.size(), InvalidArgument,
                              "id %" PRIu32 " >= %" ET_PRIsize_t, memory_id,
                              buffers_.size());
     Span<uint8_t> buffer = buffers_[memory_id];
     ET_CHECK_OR_RETURN_ERROR(
-        offset_bytes + size_bytes <= buffer.size(), MemoryAllocationFailed,
+        end_bytes <= buffer.size(), MemoryAllocationFailed,
         "offset_bytes (%" ET_PRIsize_t ") + size_bytes (%" ET_PRIsize_t
         ") >= allocator size (%" ET_PRIsize_t ") "
         "for memory_id %" PRIu32,
diff --git a/packages/react-native-executorch/third-party/include/executorch/runtime/core/portable_type/device.h b/packages/react-native-executorch/third-party/include/executorch/runtime/core/portable_type/device.h
index c6a42c55f0..32c53e9570 100644
--- a/packages/react-native-executorch/third-party/include/executorch/runtime/core/portable_type/device.h
+++ b/packages/react-native-executorch/third-party/include/executorch/runtime/core/portable_type/device.h
@@ -26,7 +26,6 @@ enum class DeviceType : int8_t {
 constexpr size_t kNumDeviceTypes = 2;
 
 /// An index representing a specific device; e.g. GPU 0 vs GPU 1.
-/// -1 means the default/unspecified device for that type.
 using DeviceIndex = int8_t;
 
 /**
@@ -41,7 +40,7 @@ struct Device final {
 
   /// Constructs a new `Device` from a `DeviceType` and an optional device
   /// index.
-  /* implicit */ Device(DeviceType type, DeviceIndex index = -1)
+  /* implicit */ Device(DeviceType type, DeviceIndex index = 0)
       : type_(type), index_(index) {}
 
   /// Returns the type of device the tensor data resides on.
@@ -50,7 +49,7 @@ struct Device final {
   /// Returns true if the device is of CPU type.
   bool is_cpu() const noexcept { return type_ == DeviceType::CPU; }
 
-  /// Returns the device index, or -1 if default/unspecified.
+  /// Returns the device index.
   DeviceIndex index() const noexcept { return index_; }
 
   bool operator==(const Device &other) const noexcept {
@@ -63,7 +62,7 @@ struct Device final {
 
 private:
   DeviceType type_;
-  DeviceIndex index_ = -1;
+  DeviceIndex index_ = 0;
 };
 
 } // namespace etensor
diff --git a/packages/react-native-executorch/third-party/include/executorch/runtime/core/portable_type/tensor_impl.h b/packages/react-native-executorch/third-party/include/executorch/runtime/core/portable_type/tensor_impl.h
index 4f79070626..c841aef0c8 100644
--- a/packages/react-native-executorch/third-party/include/executorch/runtime/core/portable_type/tensor_impl.h
+++ b/packages/react-native-executorch/third-party/include/executorch/runtime/core/portable_type/tensor_impl.h
@@ -10,8 +10,11 @@
 
 #include <executorch/runtime/core/array_ref.h>
 #include <executorch/runtime/core/error.h>
+#include <executorch/runtime/core/portable_type/device.h>
 #include <executorch/runtime/core/portable_type/scalar_type.h>
+#include <executorch/runtime/core/result.h>
 #include <executorch/runtime/core/tensor_shape_dynamism.h>
+#include <executorch/runtime/platform/compiler.h>
 
 // Forward declaration of a helper that provides access to internal resizing
 // methods of TensorImpl. Real definition is in
@@ -99,11 +102,15 @@ class TensorImpl {
    * @param strides Strides of the tensor at each dimension. Must contain `dim`
    *     entries.
    * @param dynamism The mutability of the shape of the tensor.
+   * @param device_type The type of device where tensor data resides.
+   * @param device_index The device index for multi-device scenarios.
    */
   TensorImpl(ScalarType type, ssize_t dim, SizesType *sizes,
              void *data = nullptr, DimOrderType *dim_order = nullptr,
              StridesType *strides = nullptr,
-             TensorShapeDynamism dynamism = TensorShapeDynamism::STATIC);
+             TensorShapeDynamism dynamism = TensorShapeDynamism::STATIC,
+             DeviceType device_type = DeviceType::CPU,
+             DeviceIndex device_index = 0);
 
   /**
    * Returns the size of the tensor in bytes.
@@ -161,6 +168,15 @@ class TensorImpl {
   /// Returns the mutability of the shape of the tensor.
   TensorShapeDynamism shape_dynamism() const { return shape_dynamism_; }
 
+  /// Returns the device where tensor data resides.
+  Device device() const { return device_; }
+
+  /// Returns the type of device where tensor data resides.
+  DeviceType device_type() const { return device_.type(); }
+
+  /// Returns the device index, or 0 if default/unspecified.
+  DeviceIndex device_index() const { return device_.index(); }
+
   /// Returns a pointer of type T to the constant underlying data blob.
   template <typename T> inline const T *data() const {
     return static_cast<const T *>(data());
@@ -238,6 +254,9 @@ class TensorImpl {
 
   /// Specifies the mutability of the shape of the tensor.
   const TensorShapeDynamism shape_dynamism_;
+
+  /// Device where tensor data resides (CPU, CUDA, etc.)
+  Device device_;
 };
 
 /**
@@ -247,6 +266,16 @@ ssize_t compute_numel(
     const ::executorch::runtime::etensor::TensorImpl::SizesType *sizes,
     ssize_t dim);
 
+/**
+ * Compute the number of elements based on the sizes of a tensor.
+ * Returns Error::InvalidArgument if any intermediate multiplication would
+ * overflow ssize_t, or if a size is negative. Prefer this over compute_numel()
+ * for paths that can propagate an Error upward.
+ */
+::executorch::runtime::Result<ssize_t>
+safe_numel(const ::executorch::runtime::etensor::TensorImpl::SizesType *sizes,
+           ssize_t dim);
+
 /// Appropriate format specifier for the result of calling
 /// size(). Must be used instead of using zd directly to support ATen
 /// mode.
@@ -276,6 +305,7 @@ namespace executor {
 // TODO(T197294990): Remove these deprecated aliases once all users have moved
 // to the new `::executorch` namespaces.
 using ::executorch::runtime::etensor::compute_numel;
+using ::executorch::runtime::etensor::safe_numel;
 using ::executorch::runtime::etensor::TensorImpl;
 } // namespace executor
 } // namespace torch
diff --git a/packages/react-native-executorch/third-party/include/executorch/runtime/executor/method.h b/packages/react-native-executorch/third-party/include/executorch/runtime/executor/method.h
index aecf8c8fa0..dd66951ca5 100644
--- a/packages/react-native-executorch/third-party/include/executorch/runtime/executor/method.h
+++ b/packages/react-native-executorch/third-party/include/executorch/runtime/executor/method.h
@@ -23,6 +23,7 @@
 #include <executorch/runtime/executor/memory_manager.h>
 #include <executorch/runtime/executor/merged_data_map.h>
 #include <executorch/runtime/executor/method_meta.h>
+#include <executorch/runtime/kernel/operator_registry.h>
 #include <executorch/runtime/platform/compiler.h>
 
 // Forward declare flatbuffer types. This is a public header and must not
@@ -77,7 +78,7 @@ class Method final {
         merged_data_map_(std::move(rhs.merged_data_map_)),
         external_constants_(rhs.external_constants_),
         n_external_constants_(rhs.n_external_constants_),
-        init_state_(rhs.init_state_) {
+        kernel_registry_(rhs.kernel_registry_), init_state_(rhs.init_state_) {
     // Required: clear out fields that the dtor looks at, so that we don't free
     // anything twice.
     rhs.n_value_ = 0;
@@ -323,13 +324,15 @@ class Method final {
   };
 
   Method(const Program *program, MemoryManager *memory_manager,
-         EventTracer *event_tracer, MemoryAllocator *temp_allocator)
+         EventTracer *event_tracer, MemoryAllocator *temp_allocator,
+         Span<const Kernel> kernel_registry = {})
       : step_state_(), program_(program), memory_manager_(memory_manager),
         temp_allocator_(temp_allocator), serialization_plan_(nullptr),
         event_tracer_(event_tracer), n_value_(0), values_(nullptr),
         input_set_(nullptr), n_delegate_(0), delegates_(nullptr), n_chains_(0),
         chains_(nullptr), merged_data_map_(nullptr),
         external_constants_(nullptr), n_external_constants_(0),
+        kernel_registry_(kernel_registry),
         init_state_(InitializationState::Uninitialized) {}
 
   /// Static factory used by Program.
@@ -337,7 +340,8 @@ class Method final {
   load(executorch_flatbuffer::ExecutionPlan *s_plan, const Program *program,
        MemoryManager *memory_manager, EventTracer *event_tracer,
        const NamedDataMap *named_data_map,
-       const LoadBackendOptionsMap *backend_options = nullptr);
+       const LoadBackendOptionsMap *backend_options = nullptr,
+       Span<const Kernel> kernel_registry = {});
 
   /**
    * Initialize the method from its serialized representation.
@@ -382,6 +386,8 @@ class Method final {
   NamedData *external_constants_;
   size_t n_external_constants_ = 0;
 
+  Span<const Kernel> kernel_registry_;
+
   InitializationState init_state_;
 
   /**
diff --git a/packages/react-native-executorch/third-party/include/executorch/runtime/executor/method_meta.h b/packages/react-native-executorch/third-party/include/executorch/runtime/executor/method_meta.h
index 3d387c695a..2cacb49506 100644
--- a/packages/react-native-executorch/third-party/include/executorch/runtime/executor/method_meta.h
+++ b/packages/react-native-executorch/third-party/include/executorch/runtime/executor/method_meta.h
@@ -9,6 +9,7 @@
 #pragma once
 
 #include <executorch/runtime/core/exec_aten/exec_aten.h>
+#include <executorch/runtime/core/portable_type/device.h>
 #include <executorch/runtime/core/result.h>
 #include <executorch/runtime/core/span.h>
 #include <executorch/runtime/core/tag.h>
@@ -230,6 +231,19 @@ class MethodMeta final {
    */
   Result<int64_t> memory_planned_buffer_size(size_t index) const;
 
+  /**
+   * Get the device placement for the specified memory-planned buffer.
+   *
+   * For CPU-only programs (no non_const_buffer_device in the PTE), all buffers
+   * default to Device{CPU, 0}. For programs with device annotations, returns
+   * the device type and index that the buffer should be allocated on.
+   *
+   * @param[in] index The index of the buffer to look up (0-based, same
+   *     indexing as memory_planned_buffer_size()).
+   * @returns The Device on success, or an error on failure.
+   */
+  Result<etensor::Device> memory_planned_buffer_device(size_t index) const;
+
   /**
    * Check to see if a backend is used in this method.
    *
diff --git a/packages/react-native-executorch/third-party/include/executorch/runtime/executor/platform_memory_allocator.h b/packages/react-native-executorch/third-party/include/executorch/runtime/executor/platform_memory_allocator.h
index c80d274257..181181b923 100644
--- a/packages/react-native-executorch/third-party/include/executorch/runtime/executor/platform_memory_allocator.h
+++ b/packages/react-native-executorch/third-party/include/executorch/runtime/executor/platform_memory_allocator.h
@@ -12,6 +12,7 @@
 #include <cstdint>
 #include <stdio.h>
 
+#include <c10/util/safe_numerics.h>
 #include <executorch/runtime/core/memory_allocator.h>
 #include <executorch/runtime/platform/log.h>
 #include <executorch/runtime/platform/platform.h>
@@ -46,8 +47,17 @@ class PlatformMemoryAllocator final : public MemoryAllocator {
       return nullptr;
     }
 
-    // Allocate enough memory for the node, the data and the alignment bump.
-    size_t alloc_size = sizeof(AllocationNode) + size + alignment;
+    // Check for overflow before computing total allocation size.
+    // Allocate enough for the node, data, and alignment bump (at most
+    // alignment - 1 extra bytes to align the data pointer).
+    size_t alloc_size = 0;
+    if (c10::add_overflows(sizeof(AllocationNode), size, &alloc_size) ||
+        c10::add_overflows(alloc_size, alignment - 1, &alloc_size)) {
+      ET_LOG(Error, "Allocation size overflow: size %zu, alignment %zu", size,
+             alignment);
+      return nullptr;
+    }
+
     void *node_memory = runtime::pal_allocate(alloc_size);
 
     // If allocation failed, log message and return nullptr.
diff --git a/packages/react-native-executorch/third-party/include/executorch/runtime/executor/program.h b/packages/react-native-executorch/third-party/include/executorch/runtime/executor/program.h
index 59ae346f36..01dbc1fdfa 100644
--- a/packages/react-native-executorch/third-party/include/executorch/runtime/executor/program.h
+++ b/packages/react-native-executorch/third-party/include/executorch/runtime/executor/program.h
@@ -21,6 +21,7 @@
 #include <executorch/runtime/executor/method.h>
 #include <executorch/runtime/executor/method_meta.h>
 #include <executorch/runtime/executor/pte_data_map.h>
+#include <executorch/runtime/kernel/operator_registry.h>
 #include <executorch/runtime/platform/compiler.h>
 
 // Forward declare flatbuffer types. This is a public header and must not
@@ -148,7 +149,8 @@ class Program final {
   load_method(const char *method_name, MemoryManager *memory_manager,
               EventTracer *event_tracer = nullptr,
               const NamedDataMap *named_data_map = nullptr,
-              const LoadBackendOptionsMap *backend_options = nullptr) const;
+              const LoadBackendOptionsMap *backend_options = nullptr,
+              Span<const Kernel> kernel_registry = {}) const;
 
   /**
    * Gathers metadata for the named method.
diff --git a/packages/react-native-executorch/third-party/include/executorch/runtime/executor/tensor_parser.h b/packages/react-native-executorch/third-party/include/executorch/runtime/executor/tensor_parser.h
index e1fc971d36..9d18e7a351 100644
--- a/packages/react-native-executorch/third-party/include/executorch/runtime/executor/tensor_parser.h
+++ b/packages/react-native-executorch/third-party/include/executorch/runtime/executor/tensor_parser.h
@@ -92,8 +92,12 @@ parseListOptionalType(const flatbuffers::Vector<int32_t> *value_indices,
       ET_CHECK_OR_RETURN_ERROR(
           index >= 0 && static_cast<size_t>(index) < values_len, InvalidProgram,
           "Invalid value index %" PRId32 " for ListOptional", index);
+      auto optional_result = values[index].tryToOptional<T>();
+      if (!optional_result.ok()) {
+        return optional_result.error();
+      }
       new (&optional_tensor_list[output_idx])
-          std::optional<T>(values[index].toOptional<T>());
+          std::optional<T>(std::move(optional_result.get()));
       evalp_list[output_idx] = &values[static_cast<size_t>(index)];
     }
     output_idx++;
diff --git a/packages/react-native-executorch/third-party/include/executorch/runtime/kernel/operator_registry.h b/packages/react-native-executorch/third-party/include/executorch/runtime/kernel/operator_registry.h
index ddb648d863..bc6c6729d5 100644
--- a/packages/react-native-executorch/third-party/include/executorch/runtime/kernel/operator_registry.h
+++ b/packages/react-native-executorch/third-party/include/executorch/runtime/kernel/operator_registry.h
@@ -216,6 +216,15 @@ ::executorch::runtime::Result<OpFunction>
 get_op_function_from_registry(const char *name,
                               Span<const TensorMeta> meta_list = {});
 
+/**
+ * Returns the operator with a given name and TensorMeta list from the provided
+ * kernel list instead of the global registry.
+ */
+::executorch::runtime::Result<OpFunction>
+get_op_function_from_registry(const char *name,
+                              Span<const TensorMeta> meta_list,
+                              Span<const Kernel> kernel_list);
+
 /**
  * Returns all registered kernels.
  */
diff --git a/packages/react-native-executorch/third-party/include/executorch/runtime/platform/profiler.h b/packages/react-native-executorch/third-party/include/executorch/runtime/platform/profiler.h
index 6905d64341..9d5671980b 100644
--- a/packages/react-native-executorch/third-party/include/executorch/runtime/platform/profiler.h
+++ b/packages/react-native-executorch/third-party/include/executorch/runtime/platform/profiler.h
@@ -255,8 +255,7 @@ using ::executorch::runtime::track_allocator;
   } while (0)
 
 #define EXECUTORCH_BEGIN_PROF(name)                                            \
-  {                                                                            \
-  }
+  {}
 
 #define EXECUTORCH_END_PROF(token_id)                                          \
   do {                                                                         \
@@ -278,8 +277,7 @@ using ::executorch::runtime::track_allocator;
   memset(prof_result_test, 0, sizeof(::executorch::runtime::prof_result_t));
 
 #define EXECUTORCH_RESET_PROFILE_RESULTS()                                     \
-  {                                                                            \
-  }
+  {}
 
 #define EXECUTORCH_TRACK_ALLOCATOR(name) ((void)(name), -1)
 
diff --git a/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64-simulator/ExecutorchLib.framework/ExecutorchLib b/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64-simulator/ExecutorchLib.framework/ExecutorchLib
index f74ed53c6a..d081c6501b 100755
Binary files a/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64-simulator/ExecutorchLib.framework/ExecutorchLib and b/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64-simulator/ExecutorchLib.framework/ExecutorchLib differ
diff --git a/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64-simulator/ExecutorchLib.framework/Info.plist b/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64-simulator/ExecutorchLib.framework/Info.plist
index b2b2aa2478..4b7da1be30 100644
Binary files a/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64-simulator/ExecutorchLib.framework/Info.plist and b/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64-simulator/ExecutorchLib.framework/Info.plist differ
diff --git a/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64-simulator/ExecutorchLib.framework/mlx.metallib b/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64-simulator/ExecutorchLib.framework/mlx.metallib
new file mode 100644
index 0000000000..104cc44a31
Binary files /dev/null and b/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64-simulator/ExecutorchLib.framework/mlx.metallib differ
diff --git a/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64/ExecutorchLib.framework/ExecutorchLib b/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64/ExecutorchLib.framework/ExecutorchLib
index 61193b77ef..160890caab 100755
Binary files a/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64/ExecutorchLib.framework/ExecutorchLib and b/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64/ExecutorchLib.framework/ExecutorchLib differ
diff --git a/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64/ExecutorchLib.framework/Info.plist b/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64/ExecutorchLib.framework/Info.plist
index a6f2d4a5dc..38f26327ea 100644
Binary files a/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64/ExecutorchLib.framework/Info.plist and b/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64/ExecutorchLib.framework/Info.plist differ
diff --git a/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64/ExecutorchLib.framework/mlx.metallib b/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64/ExecutorchLib.framework/mlx.metallib
new file mode 100644
index 0000000000..104cc44a31
Binary files /dev/null and b/packages/react-native-executorch/third-party/ios/ExecutorchLib.xcframework/ios-arm64/ExecutorchLib.framework/mlx.metallib differ
diff --git a/packages/react-native-executorch/third-party/ios/ExecutorchLib/ExecutorchLib.xcodeproj/project.pbxproj b/packages/react-native-executorch/third-party/ios/ExecutorchLib/ExecutorchLib.xcodeproj/project.pbxproj
index a4d139dcaa..c51f7098ea 100644
--- a/packages/react-native-executorch/third-party/ios/ExecutorchLib/ExecutorchLib.xcodeproj/project.pbxproj
+++ b/packages/react-native-executorch/third-party/ios/ExecutorchLib/ExecutorchLib.xcodeproj/project.pbxproj
@@ -11,6 +11,7 @@
 		0E4A7F482D67549100D8DCBA /* MetalPerformanceShaders.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 0E4A7F452D67549100D8DCBA /* MetalPerformanceShaders.framework */; };
 		0E4A7F492D67549100D8DCBA /* MetalPerformanceShadersGraph.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 0E4A7F462D67549100D8DCBA /* MetalPerformanceShadersGraph.framework */; };
 		5576B4B92CEF970E005027B7 /* ETModel.mm in Sources */ = {isa = PBXBuildFile; fileRef = 5576B4B82CEF970C005027B7 /* ETModel.mm */; };
+		A1MLX0001RESBUILD0001 /* mlx.metallib in Resources */ = {isa = PBXBuildFile; fileRef = A1MLX0001FILEREF0001 /* mlx.metallib */; };
 		55EA2C572CB90E7D004315B3 /* Accelerate.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 55EA2C562CB90E7D004315B3 /* Accelerate.framework */; };
 		55EA2C592CB90E80004315B3 /* CoreML.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 55EA2C582CB90E80004315B3 /* CoreML.framework */; };
 		55EA2C5B2CB90E85004315B3 /* libsqlite3.tbd in Frameworks */ = {isa = PBXBuildFile; fileRef = 55EA2C5A2CB90E85004315B3 /* libsqlite3.tbd */; };
@@ -22,6 +23,7 @@
 		0E4A7F462D67549100D8DCBA /* MetalPerformanceShadersGraph.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = MetalPerformanceShadersGraph.framework; path = System/Library/Frameworks/MetalPerformanceShadersGraph.framework; sourceTree = SDKROOT; };
 		5576B4B62CEF9705005027B7 /* ETModel.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = ETModel.h; sourceTree = "<group>"; };
 		5576B4B82CEF970C005027B7 /* ETModel.mm */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.objcpp; path = ETModel.mm; sourceTree = "<group>"; };
+		A1MLX0001FILEREF0001 /* mlx.metallib */ = {isa = PBXFileReference; lastKnownFileType = archive.metal-library; path = mlx.metallib; sourceTree = "<group>"; };
 		55EA2C1C2CB90C22004315B3 /* ExecutorchLib.framework */ = {isa = PBXFileReference; explicitFileType = wrapper.framework; includeInIndex = 0; path = ExecutorchLib.framework; sourceTree = BUILT_PRODUCTS_DIR; };
 		55EA2C562CB90E7D004315B3 /* Accelerate.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Accelerate.framework; path = System/Library/Frameworks/Accelerate.framework; sourceTree = SDKROOT; };
 		55EA2C582CB90E80004315B3 /* CoreML.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = CoreML.framework; path = System/Library/Frameworks/CoreML.framework; sourceTree = SDKROOT; };
@@ -66,6 +68,7 @@
 			isa = PBXGroup;
 			children = (
 				55EA2C352CB90C7A004315B3 /* Exported */,
+				A1MLX0001FILEREF0001 /* mlx.metallib */,
 			);
 			path = ExecutorchLib;
 			sourceTree = "<group>";
@@ -153,6 +156,7 @@
 			isa = PBXResourcesBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
+				A1MLX0001RESBUILD0001 /* mlx.metallib in Resources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};
@@ -343,6 +347,8 @@
 					"-force_load",
 					"$(PROJECT_DIR)/../../../third-party/ios/libs/executorch/libbackend_mps_ios.a",
 					"-force_load",
+					"$(PROJECT_DIR)/../../../third-party/ios/libs/executorch/libbackend_mlx_ios.a",
+					"-force_load",
 					"$(PROJECT_DIR)/../../../third-party/ios/libs/executorch/libexecutorch_ios.a",
 					"-force_load",
 					"$(PROJECT_DIR)/../../../third-party/ios/libs/executorch/libexecutorch_llm_ios.a",
@@ -435,6 +441,8 @@
 					"-force_load",
 					"$(PROJECT_DIR)/../../../third-party/ios/libs/executorch/libbackend_mps_ios.a",
 					"-force_load",
+					"$(PROJECT_DIR)/../../../third-party/ios/libs/executorch/libbackend_mlx_ios.a",
+					"-force_load",
 					"$(PROJECT_DIR)/../../../third-party/ios/libs/executorch/libexecutorch_ios.a",
 					"-force_load",
 					"$(PROJECT_DIR)/../../../third-party/ios/libs/executorch/libexecutorch_llm_ios.a",
diff --git a/packages/react-native-executorch/third-party/ios/ExecutorchLib/ExecutorchLib/mlx.metallib b/packages/react-native-executorch/third-party/ios/ExecutorchLib/ExecutorchLib/mlx.metallib
new file mode 100644
index 0000000000..104cc44a31
Binary files /dev/null and b/packages/react-native-executorch/third-party/ios/ExecutorchLib/ExecutorchLib/mlx.metallib differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_coreml_ios.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_coreml_ios.a
index 4e2a80a220..796b9d6796 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_coreml_ios.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_coreml_ios.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_coreml_simulator.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_coreml_simulator.a
index c5b16beb50..8f113a2954 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_coreml_simulator.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_coreml_simulator.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_mlx_ios.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_mlx_ios.a
new file mode 100644
index 0000000000..d34e650af7
Binary files /dev/null and b/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_mlx_ios.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_mlx_simulator.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_mlx_simulator.a
new file mode 100644
index 0000000000..59999e734d
Binary files /dev/null and b/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_mlx_simulator.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_mps_ios.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_mps_ios.a
index ce63bfd904..a8bf9728a9 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_mps_ios.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_mps_ios.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_mps_simulator.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_mps_simulator.a
index 3679faaa46..5cce760012 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_mps_simulator.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_mps_simulator.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_xnnpack_ios.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_xnnpack_ios.a
index 4049c0a640..2093739b1a 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_xnnpack_ios.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_xnnpack_ios.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_xnnpack_simulator.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_xnnpack_simulator.a
index bd145d474f..360569eb71 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_xnnpack_simulator.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libbackend_xnnpack_simulator.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libexecutorch_ios.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libexecutorch_ios.a
index 221409be07..ebe540bbbf 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libexecutorch_ios.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libexecutorch_ios.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libexecutorch_llm_ios.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libexecutorch_llm_ios.a
index 40b74e9c7d..a312c92a1e 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libexecutorch_llm_ios.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libexecutorch_llm_ios.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libexecutorch_llm_simulator.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libexecutorch_llm_simulator.a
index bec9105acb..8435e5b35a 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libexecutorch_llm_simulator.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libexecutorch_llm_simulator.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libexecutorch_simulator.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libexecutorch_simulator.a
index 99449364f2..e5a58d122d 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libexecutorch_simulator.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libexecutorch_simulator.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_llm_ios.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_llm_ios.a
index 5217a54902..ce8352f8d7 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_llm_ios.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_llm_ios.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_llm_simulator.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_llm_simulator.a
index 9add6b355c..c8a6cd42a4 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_llm_simulator.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_llm_simulator.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_optimized_ios.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_optimized_ios.a
index 9cc9a42bd2..7b530c9e28 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_optimized_ios.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_optimized_ios.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_optimized_simulator.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_optimized_simulator.a
index cb06a86712..666ab17b07 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_optimized_simulator.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_optimized_simulator.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_quantized_ios.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_quantized_ios.a
index d54805247e..2aa82a12d6 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_quantized_ios.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_quantized_ios.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_quantized_simulator.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_quantized_simulator.a
index 41f06f64fb..567ca5555f 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_quantized_simulator.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_quantized_simulator.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_torchao_ios.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_torchao_ios.a
index 19db3e80f4..b0d3385e72 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_torchao_ios.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_torchao_ios.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_torchao_simulator.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_torchao_simulator.a
index af26c633e1..662a2ac890 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_torchao_simulator.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libkernels_torchao_simulator.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libthreadpool_ios.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libthreadpool_ios.a
index 8c14cf924b..4d59154358 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libthreadpool_ios.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libthreadpool_ios.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/libthreadpool_simulator.a b/packages/react-native-executorch/third-party/ios/libs/executorch/libthreadpool_simulator.a
index d6deeb5a00..1ec1dbe74e 100644
Binary files a/packages/react-native-executorch/third-party/ios/libs/executorch/libthreadpool_simulator.a and b/packages/react-native-executorch/third-party/ios/libs/executorch/libthreadpool_simulator.a differ
diff --git a/packages/react-native-executorch/third-party/ios/libs/executorch/mlx.metallib b/packages/react-native-executorch/third-party/ios/libs/executorch/mlx.metallib
new file mode 100644
index 0000000000..104cc44a31
Binary files /dev/null and b/packages/react-native-executorch/third-party/ios/libs/executorch/mlx.metallib differ