whisper.rn 0.5.3 → 0.5.5

package/cpp/ggml.c

@@ -61,13 +61,15 @@
 
 #define UNUSED WSP_GGML_UNUSED
 
+// Needed for wsp_ggml_fp32_to_bf16_row()
+#if defined(__AVX512BF16__)
 #if defined(_MSC_VER)
-#define m512bh(p) p
 #define m512i(p) p
 #else
-#define m512bh(p) (__m512bh)(p)
+#include <immintrin.h>
 #define m512i(p) (__m512i)(p)
-#endif
+#endif // defined(_MSC_VER)
+#endif // defined(__AVX512BF16__)
 
 #if defined(__linux__) || \
     defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || \
@@ -132,6 +134,13 @@ static void wsp_ggml_print_backtrace_symbols(void) {
     int nptrs = backtrace(trace, sizeof(trace)/sizeof(trace[0]));
     backtrace_symbols_fd(trace, nptrs, STDERR_FILENO);
 }
+#elif defined(__APPLE__)
+#include <execinfo.h>
+static void wsp_ggml_print_backtrace_symbols(void) {
+    void * trace[100];
+    int nptrs = backtrace(trace, sizeof(trace)/sizeof(trace[0]));
+    backtrace_symbols_fd(trace, nptrs, STDERR_FILENO);
+}
 #else
 static void wsp_ggml_print_backtrace_symbols(void) {
     // platform not supported
@@ -143,6 +152,20 @@ void wsp_ggml_print_backtrace(void) {
     if (WSP_GGML_NO_BACKTRACE) {
         return;
     }
+#if defined(__APPLE__)
+    // On macOS, fork+debugger attachment is problematic due to:
+    // 1. libdispatch "poisons" forked child processes
+    // 2. lldb has issues attaching to parent from forked child
+    // Use simple backtrace() instead to avoid Terminal.app crashes
+    const char * WSP_GGML_BACKTRACE_LLDB = getenv("WSP_GGML_BACKTRACE_LLDB");
+    if (!WSP_GGML_BACKTRACE_LLDB) {
+        fprintf(stderr, "WARNING: Using native backtrace. Set WSP_GGML_BACKTRACE_LLDB for more info.\n");
+        fprintf(stderr, "WARNING: WSP_GGML_BACKTRACE_LLDB may cause native MacOS Terminal.app to crash.\n");
+        fprintf(stderr, "See: https://github.com/ggml-org/llama.cpp/pull/17869\n");
+        wsp_ggml_print_backtrace_symbols();
+        return;
+    }
+#endif
 #if defined(__linux__)
     FILE * f = fopen("/proc/self/status", "r");
     size_t size = 0;
@@ -998,6 +1021,7 @@ static const char * WSP_GGML_OP_NAME[WSP_GGML_OP_COUNT] = {
     "ARANGE",
     "TIMESTEP_EMBEDDING",
     "ARGSORT",
+    "TOP_K",
     "LEAKY_RELU",
     "TRI",
     "FILL",
@@ -1031,7 +1055,7 @@ static const char * WSP_GGML_OP_NAME[WSP_GGML_OP_COUNT] = {
     "GLU",
 };
 
-static_assert(WSP_GGML_OP_COUNT == 94, "WSP_GGML_OP_COUNT != 94");
+static_assert(WSP_GGML_OP_COUNT == 95, "WSP_GGML_OP_COUNT != 95");
 
 static const char * WSP_GGML_OP_SYMBOL[WSP_GGML_OP_COUNT] = {
     "none",
@@ -1106,6 +1130,7 @@ static const char * WSP_GGML_OP_SYMBOL[WSP_GGML_OP_COUNT] = {
     "arange(start, stop, step)",
     "timestep_embedding(timesteps, dim, max_period)",
     "argsort(x)",
+    "top_k(x)",
     "leaky_relu(x)",
     "tri(x)",
     "fill(x, c)",
@@ -1139,7 +1164,7 @@ static const char * WSP_GGML_OP_SYMBOL[WSP_GGML_OP_COUNT] = {
     "glu(x)",
 };
 
-static_assert(WSP_GGML_OP_COUNT == 94, "WSP_GGML_OP_COUNT != 94");
+static_assert(WSP_GGML_OP_COUNT == 95, "WSP_GGML_OP_COUNT != 95");
 
 static_assert(WSP_GGML_OP_POOL_COUNT == 2, "WSP_GGML_OP_POOL_COUNT != 2");
 
@@ -4897,6 +4922,8 @@ static struct wsp_ggml_tensor * wsp_ggml_interpolate_impl(
         int64_t               ne3,
         uint32_t              mode) {
     WSP_GGML_ASSERT((mode & 0xFF) < WSP_GGML_SCALE_MODE_COUNT);
+    // TODO: implement antialias for modes other than bilinear
+    WSP_GGML_ASSERT(!(mode & WSP_GGML_SCALE_FLAG_ANTIALIAS) || (mode & 0xFF) == WSP_GGML_SCALE_MODE_BILINEAR);
 
     struct wsp_ggml_tensor * result = wsp_ggml_new_tensor_4d(ctx, a->type, ne0, ne1, ne2, ne3);
 
@@ -4951,6 +4978,18 @@ struct wsp_ggml_tensor * wsp_ggml_pad(
     return wsp_ggml_pad_ext(ctx, a, 0, p0, 0, p1, 0, p2, 0, p3);
 }
 
+// wsp_ggml_pad_circular
+
+struct wsp_ggml_tensor * wsp_ggml_pad_circular(
+        struct wsp_ggml_context * ctx,
+        struct wsp_ggml_tensor  * a,
+        int                   p0,
+        int                   p1,
+        int                   p2,
+        int                   p3) {
+    return wsp_ggml_pad_ext_circular(ctx, a, 0, p0, 0, p1, 0, p2, 0, p3);
+}
+
 struct wsp_ggml_tensor * wsp_ggml_pad_ext(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
@@ -4977,6 +5016,7 @@ struct wsp_ggml_tensor * wsp_ggml_pad_ext(
     wsp_ggml_set_op_params_i32(result, 5, rp2);
     wsp_ggml_set_op_params_i32(result, 6, lp3);
     wsp_ggml_set_op_params_i32(result, 7, rp3);
+    wsp_ggml_set_op_params_i32(result, 8, 0); // not circular by default
 
 
     result->op     = WSP_GGML_OP_PAD;
@@ -4985,6 +5025,25 @@ struct wsp_ggml_tensor * wsp_ggml_pad_ext(
     return result;
 }
 
+// wsp_ggml_pad_ext_circular
+
+struct wsp_ggml_tensor * wsp_ggml_pad_ext_circular(
+        struct wsp_ggml_context * ctx,
+        struct wsp_ggml_tensor  * a,
+        int                  lp0,
+        int                  rp0,
+        int                  lp1,
+        int                  rp1,
+        int                  lp2,
+        int                  rp2,
+        int                  lp3,
+        int                  rp3
+        ) {
+    struct wsp_ggml_tensor * result = wsp_ggml_pad_ext(ctx, a, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3);
+    wsp_ggml_set_op_params_i32(result, 8, 1); // circular
+    return result;
+}
+
 // wsp_ggml_pad_reflect_1d
 
 struct wsp_ggml_tensor * wsp_ggml_pad_reflect_1d(
@@ -5044,28 +5103,6 @@ struct wsp_ggml_tensor * wsp_ggml_roll(
     return result;
 }
 
-// wsp_ggml_arange
-
-struct wsp_ggml_tensor * wsp_ggml_arange(
-        struct wsp_ggml_context * ctx,
-        float                 start,
-        float                 stop,
-        float                 step) {
-    WSP_GGML_ASSERT(stop > start);
-
-    const int64_t steps = (int64_t) ceilf((stop - start) / step);
-
-    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor_1d(ctx, WSP_GGML_TYPE_F32, steps);
-
-    wsp_ggml_set_op_params_f32(result, 0, start);
-    wsp_ggml_set_op_params_f32(result, 1, stop);
-    wsp_ggml_set_op_params_f32(result, 2, step);
-
-    result->op = WSP_GGML_OP_ARANGE;
-
-    return result;
-}
-
 // wsp_ggml_timestep_embedding
 
 struct wsp_ggml_tensor * wsp_ggml_timestep_embedding(
@@ -5147,6 +5184,7 @@ struct wsp_ggml_tensor * wsp_ggml_argsort(
         struct wsp_ggml_tensor   * a,
         enum wsp_ggml_sort_order   order) {
     WSP_GGML_ASSERT(a->ne[0] <= INT32_MAX);
+
     struct wsp_ggml_tensor * result = wsp_ggml_new_tensor(ctx, WSP_GGML_TYPE_I32, WSP_GGML_MAX_DIMS, a->ne);
 
     wsp_ggml_set_op_params_i32(result, 0, (int32_t) order);
@@ -5157,9 +5195,9 @@ struct wsp_ggml_tensor * wsp_ggml_argsort(
     return result;
 }
 
-// wsp_ggml_top_k
+// wsp_ggml_argsort_top_k
 
-struct wsp_ggml_tensor * wsp_ggml_top_k(
+struct wsp_ggml_tensor * wsp_ggml_argsort_top_k(
         struct wsp_ggml_context * ctx,
         struct wsp_ggml_tensor  * a,
         int                   k) {
@@ -5175,6 +5213,44 @@ struct wsp_ggml_tensor * wsp_ggml_top_k(
     return result;
 }
 
+// wsp_ggml_top_k
+
+struct wsp_ggml_tensor * wsp_ggml_top_k(
+        struct wsp_ggml_context * ctx,
+        struct wsp_ggml_tensor  * a,
+        int                   k) {
+    WSP_GGML_ASSERT(a->ne[0] >= k);
+
+    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor_4d(ctx, WSP_GGML_TYPE_I32, k, a->ne[1], a->ne[2], a->ne[3]);
+
+    result->op     = WSP_GGML_OP_TOP_K;
+    result->src[0] = a;
+
+    return result;
+}
+
+// wsp_ggml_arange
+
+struct wsp_ggml_tensor * wsp_ggml_arange(
+        struct wsp_ggml_context * ctx,
+        float                 start,
+        float                 stop,
+        float                 step) {
+    WSP_GGML_ASSERT(stop > start);
+
+    const int64_t steps = (int64_t) ceilf((stop - start) / step);
+
+    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor_1d(ctx, WSP_GGML_TYPE_F32, steps);
+
+    wsp_ggml_set_op_params_f32(result, 0, start);
+    wsp_ggml_set_op_params_f32(result, 1, stop);
+    wsp_ggml_set_op_params_f32(result, 2, step);
+
+    result->op = WSP_GGML_OP_ARANGE;
+
+    return result;
+}
+
 // wsp_ggml_flash_attn_ext
 
 struct wsp_ggml_tensor * wsp_ggml_flash_attn_ext(
@@ -5194,8 +5270,6 @@ struct wsp_ggml_tensor * wsp_ggml_flash_attn_ext(
 
     if (mask) {
         WSP_GGML_ASSERT(wsp_ggml_is_contiguous(mask));
-        WSP_GGML_ASSERT(mask->ne[1] >= WSP_GGML_PAD(q->ne[1], WSP_GGML_KQ_MASK_PAD) &&
-                "the Flash-Attention kernel requires the mask to be padded to WSP_GGML_KQ_MASK_PAD and at least n_queries big");
         //WSP_GGML_ASSERT(wsp_ggml_can_repeat_rows(mask, qk));
 
         WSP_GGML_ASSERT(q->ne[2] % mask->ne[2] == 0);
@@ -7502,6 +7576,11 @@ size_t wsp_ggml_wsp_quantize_chunk(
 
 ////////////////////////////////////////////////////////////////////////////////
 
+void wsp_ggml_log_get(wsp_ggml_log_callback * log_callback, void ** user_data) {
+    *log_callback = g_logger_state.log_callback;
+    *user_data    = g_logger_state.log_callback_user_data;
+}
+
 void wsp_ggml_log_set(wsp_ggml_log_callback log_callback, void * user_data) {
     g_logger_state.log_callback = log_callback ? log_callback : wsp_ggml_log_callback_default;
     g_logger_state.log_callback_user_data = user_data;

package/cpp/ggml.h

@@ -204,6 +204,10 @@
 #    define WSP_GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
 #endif
 
+#if defined(_WIN32) && !defined(_WIN32_WINNT)
+#    define _WIN32_WINNT 0x0A00
+#endif
+
 #include <stdbool.h>
 #include <stddef.h>
 #include <stdint.h>
@@ -230,6 +234,11 @@
 
 #if UINTPTR_MAX == 0xFFFFFFFF
     #define WSP_GGML_MEM_ALIGN 4
+#elif defined(__EMSCRIPTEN__)
+// emscripten uses max_align_t == 8, so we need WSP_GGML_MEM_ALIGN == 8 for 64-bit wasm.
+// (for 32-bit wasm, the first conditional is true and WSP_GGML_MEM_ALIGN stays 4.)
+// ref: https://github.com/ggml-org/llama.cpp/pull/18628
+    #define WSP_GGML_MEM_ALIGN 8
 #else
     #define WSP_GGML_MEM_ALIGN 16
 #endif
@@ -530,6 +539,7 @@ extern "C" {
         WSP_GGML_OP_ARANGE,
         WSP_GGML_OP_TIMESTEP_EMBEDDING,
         WSP_GGML_OP_ARGSORT,
+        WSP_GGML_OP_TOP_K,
         WSP_GGML_OP_LEAKY_RELU,
         WSP_GGML_OP_TRI,
         WSP_GGML_OP_FILL,
@@ -2147,7 +2157,8 @@ extern "C" {
     };
 
     enum wsp_ggml_scale_flag {
-        WSP_GGML_SCALE_FLAG_ALIGN_CORNERS = (1 << 8)
+        WSP_GGML_SCALE_FLAG_ALIGN_CORNERS = (1 << 8),
+        WSP_GGML_SCALE_FLAG_ANTIALIAS     = (1 << 9),
     };
 
     // interpolate
@@ -2190,6 +2201,15 @@ extern "C" {
             int                  p2,
             int                  p3);
 
+    // pad each dimension with values on the other side of the torus (looping around)
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_pad_circular(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            int                   p0,
+            int                   p1,
+            int                   p2,
+            int                   p3);
+
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_pad_ext(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
@@ -2203,6 +2223,19 @@ extern "C" {
             int                  rp3
             );
 
+    // pad each dimension with values on the other side of the torus (looping around)
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_pad_ext_circular(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            int                   lp0,
+            int                   rp0,
+            int                   lp1,
+            int                   rp1,
+            int                   lp2,
+            int                   rp2,
+            int                   lp3,
+            int                   rp3);
+
     // pad each dimension with reflection: [a, b, c, d] -> [b, a, b, c, d, c]
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_pad_reflect_1d(
             struct wsp_ggml_context * ctx,
@@ -2258,25 +2291,30 @@ extern "C" {
             struct wsp_ggml_tensor  * a,
             enum wsp_ggml_sort_order  order);
 
-    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_arange(
+    // similar to wsp_ggml_top_k but implemented as `argsort` + `view`
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_argsort_top_k(
             struct wsp_ggml_context * ctx,
-            float                 start,
-            float                 stop,
-            float                 step);
+            struct wsp_ggml_tensor  * a,
+            int                   k);
 
     // top k elements per row
+    // note: the resulting top k indices are in no particular order
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_top_k(
             struct wsp_ggml_context * ctx,
             struct wsp_ggml_tensor  * a,
             int                   k);
 
-#define WSP_GGML_KQ_MASK_PAD 64
+    WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_arange(
+            struct wsp_ggml_context * ctx,
+            float                 start,
+            float                 stop,
+            float                 step);
 
-    // q:    [n_embd_k, n_batch,     n_head,    ne3 ]
-    // k:    [n_embd_k, n_kv,        n_head_kv, ne3 ]
-    // v:    [n_embd_v, n_kv,        n_head_kv, ne3 ] !! not transposed !!
-    // mask: [n_kv,     n_batch_pad, ne32,      ne33] !! n_batch_pad = WSP_GGML_PAD(n_batch, WSP_GGML_KQ_MASK_PAD) !!
-    // res:  [n_embd_v, n_head,      n_batch,   ne3 ] !! permuted !!
+    // q:    [n_embd_k, n_batch, n_head,    ne3 ]
+    // k:    [n_embd_k, n_kv,    n_head_kv, ne3 ]
+    // v:    [n_embd_v, n_kv,    n_head_kv, ne3 ] !! not transposed !!
+    // mask: [n_kv,     n_batch, ne32,      ne33]
+    // res:  [n_embd_v, n_head,  n_batch,   ne3 ] !! permuted !!
     //
     // broadcast:
     //   n_head % n_head_kv == 0
@@ -2582,7 +2620,8 @@ extern "C" {
 
     // Set callback for all future logging events.
     // If this is not called, or NULL is supplied, everything is output on stderr.
-    WSP_GGML_API void wsp_ggml_log_set(wsp_ggml_log_callback log_callback, void * user_data);
+    WSP_GGML_API void wsp_ggml_log_get(wsp_ggml_log_callback * log_callback, void ** user_data);
+    WSP_GGML_API void wsp_ggml_log_set(wsp_ggml_log_callback   log_callback, void *  user_data);
 
     WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_set_zero(struct wsp_ggml_tensor * tensor);
 

package/cpp/jsi/RNWhisperJSI.cpp

@@ -574,6 +574,7 @@ void installJSIBindings(
                                     if (code == 0) {
                                         auto resultObj = Object(runtime);
                                         resultObj.setProperty(runtime, "code", Value(code));
+                                        resultObj.setProperty(runtime, "language", String::createFromUtf8(runtime, whisper_lang_str(whisper_full_lang_id(context))));
                                         resultObj.setProperty(runtime, "result", String::createFromUtf8(runtime, createFullTextFromSegments(context, 0)));
                                         resultObj.setProperty(runtime, "segments", createSegmentsArray(runtime, context, 0));
                                         resolvePtr->call(runtime, resultObj);

package/cpp/whisper.cpp

@@ -2505,7 +2505,7 @@ static struct wsp_ggml_cgraph * whisper_build_graph_decoder(
 
     const float KQscale = pow(float(n_state_head), -0.25);
 
-    struct wsp_ggml_tensor * KQ_mask = wsp_ggml_new_tensor_3d(ctx0, WSP_GGML_TYPE_F32, n_kv, WSP_GGML_PAD(n_tokens, WSP_GGML_KQ_MASK_PAD), 1);
+    struct wsp_ggml_tensor * KQ_mask = wsp_ggml_new_tensor_3d(ctx0, WSP_GGML_TYPE_F32, n_kv, n_tokens, 1);
     wsp_ggml_set_name(KQ_mask, "KQ_mask");
     wsp_ggml_set_input(KQ_mask);
 
@@ -2929,7 +2929,7 @@ static bool whisper_decode_internal(
                     }
                 }
 
-                for (int i = n_tokens; i < WSP_GGML_PAD(n_tokens, WSP_GGML_KQ_MASK_PAD); ++i) {
+                for (int i = n_tokens; i < n_tokens; ++i) {
                     for (int j = 0; j < n_kv; ++j) {
                         data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY;
                     }
@@ -6030,6 +6030,19 @@ static inline bool should_split_on_word(const char * txt, bool split_on_word) {
     return txt[0] == ' ';
 }
 
+// Count UTF-8 characters (not bytes) in a string
+static int utf8_len(const char * str) {
+    int count = 0;
+    while (*str) {
+        // Skip continuation bytes (10xxxxxx)
+        if ((*str & 0xC0) != 0x80) {
+            count++;
+        }
+        str++;
+    }
+    return count;
+}
+
 static void whisper_exp_compute_token_level_timestamps_dtw(
             struct whisper_context * ctx,
               struct whisper_state * state,
@@ -6058,7 +6071,7 @@ static int whisper_wrap_segment(struct whisper_context & ctx, struct whisper_sta
         }
 
         const auto txt = whisper_token_to_str(&ctx, token.id);
-        const int cur = strlen(txt);
+        const int cur = utf8_len(txt);  // Use UTF-8 character count instead of byte count
 
         if (acc + cur > max_len && i > 0 && should_split_on_word(txt, split_on_word)) {
             state.result_all.back().text = std::move(text);
@@ -6697,7 +6710,7 @@ static bool whisper_vad(
             }
 
             segment_start_samples = std::min(segment_start_samples, n_samples - 1);
-            segment_end_samples = std::min(segment_end_samples, n_samples);
+            segment_end_samples = std::min(segment_end_samples, n_samples - 1);
             int segment_length = segment_end_samples - segment_start_samples;
             if (segment_length > 0) {
                 whisper_state::vad_segment_info segment;

package/ios/CMakeLists.txt

@@ -16,6 +16,8 @@ add_definitions(
     -DWSP_GGML_USE_ACCELERATE
     -DWSP_GGML_USE_METAL
     -DWSP_GGML_METAL_USE_BF16
+    -DWHISPER_USE_COREML
+    -DWHISPER_COREML_ALLOW_FALLBACK
 )
 
 if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64;x86_64")
@@ -31,6 +33,14 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64")
     )
 endif ()
 
+# CoreML source files
+set(SOURCE_FILES_COREML
+    ${SOURCE_DIR}/coreml/whisper-encoder.mm
+    ${SOURCE_DIR}/coreml/whisper-encoder-impl.m
+    ${SOURCE_DIR}/coreml/whisper-decoder-impl.m
+    ${SOURCE_DIR}/coreml/whisper-compat.m
+)
+
 # Define public headers
 set(PUBLIC_HEADERS
     ${SOURCE_DIR}/rn-whisper.h
@@ -69,6 +79,7 @@ add_library(rnwhisper SHARED
     ${SOURCE_DIR}/rn-whisper.cpp
     ${SOURCE_DIR}/rn-audioutils.cpp
     ${SOURCE_FILES_ARCH}
+    ${SOURCE_FILES_COREML}
 )
 
 # Setup include directories
@@ -76,6 +87,7 @@ target_include_directories(rnwhisper
     PUBLIC
         $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../cpp>
         $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../cpp/ggml-cpu>
+        $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../cpp/coreml>
         $<INSTALL_INTERFACE:include>
 )
 
@@ -85,6 +97,7 @@ target_link_libraries(rnwhisper PRIVATE
     "-framework Foundation"
     "-framework Metal"
     "-framework MetalKit"
+    "-framework CoreML"
 )
 
 # Set properties for framework
@@ -96,5 +109,12 @@ set_target_properties(rnwhisper PROPERTIES
     FRAMEWORK_VERSION 1.0.0
     VERSION 1.0.0
     PUBLIC_HEADER "${PUBLIC_HEADERS}"
-    XCODE_ATTRIBUTE_CLANG_ENABLE_OBJC_ARC NO
+    XCODE_ATTRIBUTE_CLANG_ENABLE_OBJC_ARC YES
+)
+
+# Disable ARC for Metal Objective-C files (they don't support ARC)
+set_source_files_properties(
+    ${SOURCE_DIR}/ggml-metal/ggml-metal-context.m
+    ${SOURCE_DIR}/ggml-metal/ggml-metal-device.m
+    PROPERTIES COMPILE_FLAGS "-fno-objc-arc"
 )

package/ios/RNWhisperContext.mm

@@ -612,6 +612,11 @@ struct rnwhisper_segments_callback_data {
     NSMutableDictionary *result = [[NSMutableDictionary alloc] init];
     result[@"result"] = text;
     result[@"segments"] = segments;
+    int lang_id = whisper_full_lang_id(self->ctx);
+    const char *lang_str = whisper_lang_str(lang_id);
+    if (lang_str != nullptr) {
+        result[@"language"] = [NSString stringWithUTF8String:lang_str];
+    }
     return result;
 }
 

package/ios/rnwhisper.xcframework/ios-arm64/rnwhisper.framework/Headers/ggml-alloc.h

@@ -53,7 +53,14 @@ WSP_GGML_API void           wsp_ggml_gallocr_free(wsp_ggml_gallocr_t galloc);
 // call with a worst-case graph to avoid buffer reallocations
 // not strictly required for single buffer usage: wsp_ggml_gallocr_alloc_graph will reallocate the buffers automatically if needed
 // returns false if the buffer allocation failed
+// wsp_ggml_gallocr_resrve_n_size writes the buffer sizes per galloc buffer that would be allocated by wsp_ggml_gallocr_reserve_n to sizes
 WSP_GGML_API bool wsp_ggml_gallocr_reserve(wsp_ggml_gallocr_t galloc, struct wsp_ggml_cgraph * graph);
+WSP_GGML_API void wsp_ggml_gallocr_reserve_n_size(
+    wsp_ggml_gallocr_t galloc,
+    struct wsp_ggml_cgraph * graph,
+    const int * node_buffer_ids,
+    const int * leaf_buffer_ids,
+    size_t * sizes);
 WSP_GGML_API bool wsp_ggml_gallocr_reserve_n(
     wsp_ggml_gallocr_t galloc,
     struct wsp_ggml_cgraph * graph,
@@ -68,6 +75,8 @@ WSP_GGML_API size_t wsp_ggml_gallocr_get_buffer_size(wsp_ggml_gallocr_t galloc,
 
 // Utils
 // Create a buffer and allocate all the tensors in a wsp_ggml_context
+// wsp_ggml_backend_alloc_ctx_tensors_from_buft_size returns the size of the buffer that would be allocated by wsp_ggml_backend_alloc_ctx_tensors_from_buft
+WSP_GGML_API size_t                       wsp_ggml_backend_alloc_ctx_tensors_from_buft_size(struct wsp_ggml_context * ctx, wsp_ggml_backend_buffer_type_t buft);
 WSP_GGML_API struct wsp_ggml_backend_buffer * wsp_ggml_backend_alloc_ctx_tensors_from_buft(struct wsp_ggml_context * ctx, wsp_ggml_backend_buffer_type_t buft);
 WSP_GGML_API struct wsp_ggml_backend_buffer * wsp_ggml_backend_alloc_ctx_tensors(struct wsp_ggml_context * ctx, wsp_ggml_backend_t backend);
 

package/ios/rnwhisper.xcframework/ios-arm64/rnwhisper.framework/Headers/ggml-backend-impl.h

@@ -144,7 +144,7 @@ extern "C" {
         // device description: short informative description of the device, could be the model name
         const char * (*get_description)(wsp_ggml_backend_dev_t dev);
 
-        // device memory in bytes
+        // device memory in bytes: 0 bytes to indicate no memory to report
         void         (*get_memory)(wsp_ggml_backend_dev_t dev, size_t * free, size_t * total);
 
         // device type

package/ios/rnwhisper.xcframework/ios-arm64/rnwhisper.framework/Headers/ggml-backend.h

@@ -307,6 +307,7 @@ extern "C" {
     WSP_GGML_API void                 wsp_ggml_backend_sched_free(wsp_ggml_backend_sched_t sched);
 
     // Initialize backend buffers from a measure graph
+    WSP_GGML_API void                 wsp_ggml_backend_sched_reserve_size(wsp_ggml_backend_sched_t sched, struct wsp_ggml_cgraph * measure_graph, size_t * sizes);
     WSP_GGML_API bool                 wsp_ggml_backend_sched_reserve(wsp_ggml_backend_sched_t sched, struct wsp_ggml_cgraph * measure_graph); // returns success
 
     WSP_GGML_API int                  wsp_ggml_backend_sched_get_n_backends(wsp_ggml_backend_sched_t sched);
@@ -357,7 +358,7 @@ extern "C" {
     typedef bool (*wsp_ggml_backend_eval_callback)(int node_index, struct wsp_ggml_tensor * t1, struct wsp_ggml_tensor * t2, void * user_data);
 
     // Compare the output of two backends
-    WSP_GGML_API bool wsp_ggml_backend_compare_graph_backend(wsp_ggml_backend_t backend1, wsp_ggml_backend_t backend2, struct wsp_ggml_cgraph * graph, wsp_ggml_backend_eval_callback callback, void * user_data, struct wsp_ggml_tensor * test_node);
+    WSP_GGML_API bool wsp_ggml_backend_compare_graph_backend(wsp_ggml_backend_t backend1, wsp_ggml_backend_t backend2, struct wsp_ggml_cgraph * graph, wsp_ggml_backend_eval_callback callback, void * user_data, struct wsp_ggml_tensor const * const * test_nodes, size_t num_test_nodes);
 
     // Tensor initialization
     WSP_GGML_API enum wsp_ggml_status wsp_ggml_backend_tensor_alloc(wsp_ggml_backend_buffer_t buffer, struct wsp_ggml_tensor * tensor, void * addr);

package/ios/rnwhisper.xcframework/ios-arm64/rnwhisper.framework/Headers/ggml-cpu.h

@@ -99,6 +99,7 @@ extern "C" {
     WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_sme        (void);
     // other
     WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_riscv_v    (void);
+    WSP_GGML_BACKEND_API int wsp_ggml_cpu_get_rvv_vlen   (void);  // risc-v vector length in bytes
     WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_vsx        (void);
     WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_vxe        (void);
     WSP_GGML_BACKEND_API int wsp_ggml_cpu_has_wasm_simd  (void);

package/ios/rnwhisper.xcframework/ios-arm64/rnwhisper.framework/Headers/ggml-impl.h

@@ -24,10 +24,6 @@
 #include <arm_neon.h>
 #endif
 
-#if defined(__F16C__)
-#include <immintrin.h>
-#endif
-
 #ifdef __cplusplus
 extern "C" {
 #endif