@dvai-bridge/ios-llama-core 4.0.0 → 4.0.2

package/ios/Sources/DVAILlamaCoreObjC/include/LlamaCppBridge.h

@@ -1,101 +1,101 @@
-#import <Foundation/Foundation.h>
-
-NS_ASSUME_NONNULL_BEGIN
-
-/// Objective-C++ bridge to llama.cpp. Wraps the C API for use from Swift via the
-/// `DVAICapacitorLlamaObjC` module. Owns the `llama_model` and `llama_context`
-/// for the lifetime of a load/unload cycle.
-@interface LlamaCppBridge : NSObject
-
-@property (nonatomic, readonly, getter=isLoaded) BOOL loaded;
-@property (nonatomic, readonly, copy, nullable) NSString *currentModelPath;
-
-- (instancetype)init;
-
-- (BOOL)loadModelAtPath:(NSString *)path
-             mmprojPath:(nullable NSString *)mmprojPath
-              gpuLayers:(int)gpuLayers
-            contextSize:(int)contextSize
-                threads:(int)threads
-          embeddingMode:(BOOL)embeddingMode
-                  error:(NSError **)error;
-
-- (void)unload;
-
-- (NSString *)versionString;
-
-/// Greedy-sample `maxTokens` tokens from the loaded model for the given prompt.
-/// `temperature` and `topP` are accepted but currently ignored — this entry point
-/// uses a deterministic greedy sampler for Task 30. Future work (Task 36) will
-/// honour temperature/top-p via additional sampler-chain stages.
-- (nullable NSString *)completePrompt:(NSString *)prompt
-                            maxTokens:(int)maxTokens
-                          temperature:(float)temperature
-                                 topP:(float)topP
-                                error:(NSError **)error;
-
-/// Compute an embedding vector for the given text. Requires the model to have
-/// been loaded with `embeddingMode:YES`; otherwise the returned values are
-/// undefined / not meaningful (the handler layer is responsible for the 400
-/// short-circuit before we get here). Returns the per-dimension floats (length
-/// == llama_n_embd(model)) wrapped as `NSNumber` doubles.
-- (nullable NSArray<NSNumber *> *)embedding:(NSString *)text
-                                      error:(NSError **)error;
-
-/// Whether a multimodal projector (mmproj) has been loaded for this bridge.
-@property (nonatomic, readonly, getter=isMmprojLoaded) BOOL mmprojLoaded;
-
-/// Load a multimodal projector (mmproj). The main model must already be
-/// loaded — the projector is always paired with a text model. Phase 2A Pass 2
-/// wires the real `mtmd_init_from_file()` call.
-///
-/// `useGPU` controls `mtmd_context_params.use_gpu`. Pass `YES` for production
-/// (Metal-accelerated CLIP / vision-encoder pass on real iPhone). Pass `NO`
-/// in environments where Metal allocation can't accommodate the projector's
-/// position-embedding tensor — most commonly the iOS Simulator, where
-/// `_xpc_shmem_create_with_prot` aborts on tensors >~ 60 MiB. The bridge
-/// uses YES by default. Returns NO on failure (with `error` populated).
-- (BOOL)loadMmprojAtPath:(NSString *)mmprojPath
-                  useGPU:(BOOL)useGPU
-                   error:(NSError **)error;
-
-/// Convenience wrapper that defaults `useGPU:YES`. Most callers want this.
-- (BOOL)loadMmprojAtPath:(NSString *)mmprojPath
-                   error:(NSError **)error;
-
-/// Unload the multimodal projector. Safe to call when nothing is loaded
-/// (idempotent). Frees the underlying mtmd_context.
-- (void)unloadMmproj;
-
-/// Apply a chat template via `llama_chat_apply_template`. If `templateOverride`
-/// is nil/empty, the model's bundled `tokenizer.chat_template` (via
-/// `llama_model_chat_template`) is used. `messages` is an array of
-/// `@{ @"role": @"...", @"content": @"..." }` dicts. Returns the rendered
-/// prompt string with role markers; multimodal callers should pre-populate
-/// `<__media__>` markers in the content fields where image/audio bytes will
-/// splice in.
-- (nullable NSString *)applyChatTemplate:(nullable NSString *)templateOverride
-                                messages:(NSArray<NSDictionary<NSString *, NSString *> *> *)messages
-                            addAssistant:(BOOL)addAssistant
-                                   error:(NSError **)error;
-
-/// Multimodal completion. `prompt` must contain N `<__media__>` markers
-/// matching the count of `media`. Bytes are auto-detected as image or audio
-/// via `mtmd_helper_bitmap_init_from_buf` (magic bytes); the caller must
-/// supply media in declaration order (i.e. the same order the markers
-/// appear in `prompt`). Returns the generated text. Throws on tokenization
-/// / eval / decode failures.
-- (nullable NSString *)completeMultimodalPrompt:(NSString *)prompt
-                                          media:(NSArray<NSData *> *)mediaInOrder
-                                      maxTokens:(int)maxTokens
-                                    temperature:(float)temperature
-                                           topP:(float)topP
-                                          error:(NSError **)error;
-
-/// Returns YES if the loaded model declares an audio encoder (via
-/// `mtmd_support_audio()`). Always NO when no mmproj is loaded.
-- (BOOL)hasAudioEncoder;
-
-@end
-
-NS_ASSUME_NONNULL_END
+#import <Foundation/Foundation.h>
+
+NS_ASSUME_NONNULL_BEGIN
+
+/// Objective-C++ bridge to llama.cpp. Wraps the C API for use from Swift via the
+/// `DVAICapacitorLlamaObjC` module. Owns the `llama_model` and `llama_context`
+/// for the lifetime of a load/unload cycle.
+@interface LlamaCppBridge : NSObject
+
+@property (nonatomic, readonly, getter=isLoaded) BOOL loaded;
+@property (nonatomic, readonly, copy, nullable) NSString *currentModelPath;
+
+- (instancetype)init;
+
+- (BOOL)loadModelAtPath:(NSString *)path
+             mmprojPath:(nullable NSString *)mmprojPath
+              gpuLayers:(int)gpuLayers
+            contextSize:(int)contextSize
+                threads:(int)threads
+          embeddingMode:(BOOL)embeddingMode
+                  error:(NSError **)error;
+
+- (void)unload;
+
+- (NSString *)versionString;
+
+/// Greedy-sample `maxTokens` tokens from the loaded model for the given prompt.
+/// `temperature` and `topP` are accepted but currently ignored — this entry point
+/// uses a deterministic greedy sampler for Task 30. Future work (Task 36) will
+/// honour temperature/top-p via additional sampler-chain stages.
+- (nullable NSString *)completePrompt:(NSString *)prompt
+                            maxTokens:(int)maxTokens
+                          temperature:(float)temperature
+                                 topP:(float)topP
+                                error:(NSError **)error;
+
+/// Compute an embedding vector for the given text. Requires the model to have
+/// been loaded with `embeddingMode:YES`; otherwise the returned values are
+/// undefined / not meaningful (the handler layer is responsible for the 400
+/// short-circuit before we get here). Returns the per-dimension floats (length
+/// == llama_n_embd(model)) wrapped as `NSNumber` doubles.
+- (nullable NSArray<NSNumber *> *)embedding:(NSString *)text
+                                      error:(NSError **)error;
+
+/// Whether a multimodal projector (mmproj) has been loaded for this bridge.
+@property (nonatomic, readonly, getter=isMmprojLoaded) BOOL mmprojLoaded;
+
+/// Load a multimodal projector (mmproj). The main model must already be
+/// loaded — the projector is always paired with a text model. Phase 2A Pass 2
+/// wires the real `mtmd_init_from_file()` call.
+///
+/// `useGPU` controls `mtmd_context_params.use_gpu`. Pass `YES` for production
+/// (Metal-accelerated CLIP / vision-encoder pass on real iPhone). Pass `NO`
+/// in environments where Metal allocation can't accommodate the projector's
+/// position-embedding tensor — most commonly the iOS Simulator, where
+/// `_xpc_shmem_create_with_prot` aborts on tensors >~ 60 MiB. The bridge
+/// uses YES by default. Returns NO on failure (with `error` populated).
+- (BOOL)loadMmprojAtPath:(NSString *)mmprojPath
+                  useGPU:(BOOL)useGPU
+                   error:(NSError **)error;
+
+/// Convenience wrapper that defaults `useGPU:YES`. Most callers want this.
+- (BOOL)loadMmprojAtPath:(NSString *)mmprojPath
+                   error:(NSError **)error;
+
+/// Unload the multimodal projector. Safe to call when nothing is loaded
+/// (idempotent). Frees the underlying mtmd_context.
+- (void)unloadMmproj;
+
+/// Apply a chat template via `llama_chat_apply_template`. If `templateOverride`
+/// is nil/empty, the model's bundled `tokenizer.chat_template` (via
+/// `llama_model_chat_template`) is used. `messages` is an array of
+/// `@{ @"role": @"...", @"content": @"..." }` dicts. Returns the rendered
+/// prompt string with role markers; multimodal callers should pre-populate
+/// `<__media__>` markers in the content fields where image/audio bytes will
+/// splice in.
+- (nullable NSString *)applyChatTemplate:(nullable NSString *)templateOverride
+                                messages:(NSArray<NSDictionary<NSString *, NSString *> *> *)messages
+                            addAssistant:(BOOL)addAssistant
+                                   error:(NSError **)error;
+
+/// Multimodal completion. `prompt` must contain N `<__media__>` markers
+/// matching the count of `media`. Bytes are auto-detected as image or audio
+/// via `mtmd_helper_bitmap_init_from_buf` (magic bytes); the caller must
+/// supply media in declaration order (i.e. the same order the markers
+/// appear in `prompt`). Returns the generated text. Throws on tokenization
+/// / eval / decode failures.
+- (nullable NSString *)completeMultimodalPrompt:(NSString *)prompt
+                                          media:(NSArray<NSData *> *)mediaInOrder
+                                      maxTokens:(int)maxTokens
+                                    temperature:(float)temperature
+                                           topP:(float)topP
+                                          error:(NSError **)error;
+
+/// Returns YES if the loaded model declares an audio encoder (via
+/// `mtmd_support_audio()`). Always NO when no mmproj is loaded.
+- (BOOL)hasAudioEncoder;
+
+@end
+
+NS_ASSUME_NONNULL_END

package/ios/Tests/DVAILlamaCoreTests/AudioDecoderTest.swift

@@ -1,46 +1,46 @@
-import XCTest
-@testable import DVAILlamaCore
-
-final class AudioDecoderTest: XCTestCase {
-    func testPCM16PassThrough() async throws {
-        let pcm = try Data(contentsOf: audioFixtureURL("pcm16-1s-16khz-mono.bin"))
-        let result = try await AudioDecoder.decode(data: pcm, format: .pcm16)
-        XCTAssertEqual(result.count, pcm.count)
-    }
-
-    func testWavToPCM() async throws {
-        let wav = try Data(contentsOf: audioFixtureURL("wav-1s-16khz-mono.wav"))
-        let result = try await AudioDecoder.decode(data: wav, format: .wav)
-        // 1s @ 16kHz mono PCM16 = 32000 bytes (allow ±5%).
-        XCTAssertGreaterThan(result.count, 30000)
-        XCTAssertLessThan(result.count, 34000)
-    }
-
-    func testM4AToPCM() async throws {
-        let m4a = try Data(contentsOf: audioFixtureURL("m4a-1s.m4a"))
-        let result = try await AudioDecoder.decode(data: m4a, format: .m4a)
-        // M4A AAC 1s decodes to ~32000 bytes; AAC priming may shave a few hundred samples.
-        XCTAssertGreaterThan(result.count, 25000)
-        XCTAssertLessThan(result.count, 36000)
-    }
-
-    /// Walks up from this test source file until it finds the repo-root
-    /// `fixtures/` directory. We don't bundle the fixtures into the SwiftPM
-    /// test target's resources because they're shared with Android and Node
-    /// tests.
-    private func fixturesURL() -> URL {
-        var dir = URL(fileURLWithPath: #file).deletingLastPathComponent()
-        while !FileManager.default.fileExists(atPath: dir.appendingPathComponent("fixtures").path) {
-            let parent = dir.deletingLastPathComponent()
-            if parent.path == dir.path {
-                fatalError("fixtures dir not found walking up from \(#file)")
-            }
-            dir = parent
-        }
-        return dir.appendingPathComponent("fixtures")
-    }
-
-    private func audioFixtureURL(_ name: String) -> URL {
-        fixturesURL().appendingPathComponent("audio").appendingPathComponent(name)
-    }
-}
+import XCTest
+@testable import DVAILlamaCore
+
+final class AudioDecoderTest: XCTestCase {
+    func testPCM16PassThrough() async throws {
+        let pcm = try Data(contentsOf: audioFixtureURL("pcm16-1s-16khz-mono.bin"))
+        let result = try await AudioDecoder.decode(data: pcm, format: .pcm16)
+        XCTAssertEqual(result.count, pcm.count)
+    }
+
+    func testWavToPCM() async throws {
+        let wav = try Data(contentsOf: audioFixtureURL("wav-1s-16khz-mono.wav"))
+        let result = try await AudioDecoder.decode(data: wav, format: .wav)
+        // 1s @ 16kHz mono PCM16 = 32000 bytes (allow ±5%).
+        XCTAssertGreaterThan(result.count, 30000)
+        XCTAssertLessThan(result.count, 34000)
+    }
+
+    func testM4AToPCM() async throws {
+        let m4a = try Data(contentsOf: audioFixtureURL("m4a-1s.m4a"))
+        let result = try await AudioDecoder.decode(data: m4a, format: .m4a)
+        // M4A AAC 1s decodes to ~32000 bytes; AAC priming may shave a few hundred samples.
+        XCTAssertGreaterThan(result.count, 25000)
+        XCTAssertLessThan(result.count, 36000)
+    }
+
+    /// Walks up from this test source file until it finds the repo-root
+    /// `fixtures/` directory. We don't bundle the fixtures into the SwiftPM
+    /// test target's resources because they're shared with Android and Node
+    /// tests.
+    private func fixturesURL() -> URL {
+        var dir = URL(fileURLWithPath: #file).deletingLastPathComponent()
+        while !FileManager.default.fileExists(atPath: dir.appendingPathComponent("fixtures").path) {
+            let parent = dir.deletingLastPathComponent()
+            if parent.path == dir.path {
+                fatalError("fixtures dir not found walking up from \(#file)")
+            }
+            dir = parent
+        }
+        return dir.appendingPathComponent("fixtures")
+    }
+
+    private func audioFixtureURL(_ name: String) -> URL {
+        fixturesURL().appendingPathComponent("audio").appendingPathComponent(name)
+    }
+}