Merge branch 'main' into recurrent-fla

Author: Gasoonjia

.ci/scripts/export_model_artifact.sh

@@ -358,7 +358,7 @@ if [ "$MODEL_NAME" = "voxtral_realtime" ]; then
   STREAMING_ARG=""
   PREPROCESSOR_ARGS="--feature_size 128 --output_file ${OUTPUT_DIR}/preprocessor.pte"
   if [ "$USE_STREAMING" = "true" ]; then
-    STREAMING_ARG="--streaming"
+    STREAMING_ARG="--streaming --sliding-window 2048"
     PREPROCESSOR_ARGS="$PREPROCESSOR_ARGS --streaming"
   else
     PREPROCESSOR_ARGS="$PREPROCESSOR_ARGS --stack_output --max_audio_len 300"

.github/workflows/android-release-artifacts.yml

@@ -165,7 +165,7 @@ jobs:
       contents: read
     steps:
       - name: configure aws credentials
-        uses: aws-actions/configure-aws-credentials@v1.7.0
+        uses: aws-actions/configure-aws-credentials@v4
         with:
           role-to-assume: arn:aws:iam::308535385114:role/gha_executorch_upload-frameworks-android
           aws-region: us-east-1

.github/workflows/apple.yml

@@ -239,7 +239,7 @@ jobs:
           python-version: '3.11'
           cache: pip
       - name: configure aws credentials
-        uses: aws-actions/configure-aws-credentials@v1.7.0
+        uses: aws-actions/configure-aws-credentials@v4
         with:
           role-to-assume: arn:aws:iam::308535385114:role/gha_executorch_upload-frameworks-ios
           aws-region: us-east-1

.lintrunner.toml

@@ -223,6 +223,7 @@ exclude_patterns = [
     '**/*.gif',
     'extension/llm/tokenizers',
     'extension/llm/tokenizers/**',
+    'backends/cadence/utils/FACTO',
     'examples/cuda',
     'kernels/portable',
     # File contains @generated

backends/aoti/utils.h

@@ -15,6 +15,7 @@
 #include <executorch/runtime/platform/log.h>
 #include <cstddef>
 #include <cstdint>
+#include <utility>
 #include <vector>
 
 namespace executorch {
@@ -163,6 +164,27 @@ inline bool is_contiguous_tensor(
   return true;
 }
 
+// Scope guard: invokes a callable on destruction. Equivalent to
+// std::scope_exit (C++23 <scope>), which is not available in C++17/20.
+template <typename F>
+class ScopeGuard {
+ public:
+  static_assert(
+      noexcept(std::declval<F&>()()),
+      "ScopeGuard callable must be noexcept to avoid std::terminate "
+      "if it throws during stack unwinding");
+
+  explicit ScopeGuard(F&& fn) : fn_(std::move(fn)) {}
+  ~ScopeGuard() noexcept {
+    fn_();
+  }
+  ScopeGuard(const ScopeGuard&) = delete;
+  ScopeGuard& operator=(const ScopeGuard&) = delete;
+
+ private:
+  F fn_;
+};
+
 } // namespace aoti
 } // namespace backends
 } // namespace executorch

backends/apple/coreml/runtime/delegate/ETCoreMLCacheProtocol.h

@@ -0,0 +1,61 @@
+//
+// ETCoreMLCacheProtocol.h
+//
+// Copyright © 2024 Apple Inc. All rights reserved.
+//
+// Please refer to the license found in the LICENSE file in the root directory of the source tree.
+
+#import <Foundation/Foundation.h>
+
+NS_ASSUME_NONNULL_BEGIN
+
+/// Protocol defining the interface for CoreML model caching.
+///
+/// This protocol abstracts the cache implementation
+@protocol ETCoreMLCache <NSObject>
+
+/// Returns the URL of a cached model if it exists and is valid, otherwise nil.
+///
+/// @param identifier The unique identifier for the cached model.
+/// @param error On failure, error is filled with the failure information.
+/// @return The URL to the cached model bundle, or nil if not found or invalid.
+///
+/// @warning The returned URL may become invalid before the caller uses it if another
+/// process deletes or replaces the cached model. Callers MUST handle MLModel load
+/// failures gracefully by treating them as cache misses and recompiling.
+- (nullable NSURL*)cachedModelURLForIdentifier:(NSString*)identifier error:(NSError**)error;
+
+/// Stores a compiled model in the cache. Returns the cached URL on success.
+///
+/// @param compiledModelURL The URL of the compiled model bundle to cache. Must exist.
+/// @param identifier The unique identifier for this model.
+/// @param error On failure, error is filled with the failure information.
+/// @return The URL of the cached model, or nil on failure.
+- (nullable NSURL*)storeModelAtURL:(NSURL*)compiledModelURL withIdentifier:(NSString*)identifier error:(NSError**)error;
+
+/// Removes a specific cached model.
+///
+/// @param identifier The unique identifier for the cached model to remove.
+/// @param error On failure, error is filled with the failure information.
+/// @return YES if the model was removed or didn't exist. Returns NO only on I/O errors.
+- (BOOL)removeCachedModelWithIdentifier:(NSString*)identifier error:(NSError**)error;
+
+/// Clears the entire cache, including all cached models.
+///
+/// @param error On failure, error is filled with the failure information.
+/// @return YES if the cache was purged successfully, otherwise NO.
+- (BOOL)purgeAndReturnError:(NSError**)error;
+
+/// Returns a temp URL where intermediate files can be written during compilation.
+/// This is guaranteed to be on the same filesystem as the cache, ensuring atomic moves.
+///
+/// @param error On failure, error is filled with the failure information.
+/// @return A temp URL where intermediate files can be written, or nil on failure.
+///
+/// @note The temp URL is unique (UUID-based) to avoid conflicts.
+/// @note Temp entries are cleaned up automatically after 24 hours.
+- (nullable NSURL*)temporaryDirectoryWithError:(NSError**)error;
+
+@end
+
+NS_ASSUME_NONNULL_END