metal : fix more leaks due to missing autoreleasepools (#27883)

* metal : fix more leaks due to missing autoreleasepools

* metal : rename variable

* metal : fix another missing pool warning

Co-authored-by: YiChen Lv <63285796+forforever73@users.noreply.github.com>

---------

Co-authored-by: YiChen Lv <63285796+forforever73@users.noreply.github.com>

ggml/src/ggml-metal/ggml-metal-context.m

@@ -69,6 +69,10 @@ struct ggml_metal {
     // extra command buffers for things like getting, setting and copying tensors
     NSMutableArray * cmd_bufs_ext;
 
+    // buffers to release after async Metal operations complete
+    // if Metal released them, it would do so on a Metal-internal thread without an autorelease pool, which could cause leaks
+    NSMutableArray * buf_refs;
+
     // the last command buffer queued into the Metal queue with operations relevant to the current Metal backend
     id<MTLCommandBuffer> cmd_buf_last;
 
@@ -179,6 +183,7 @@ ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
         }
 
         res->cmd_bufs_ext = [[NSMutableArray alloc] init];
+        res->buf_refs     = [[NSMutableArray alloc] init];
 
         res->cmd_buf_last = nil;
 
@@ -206,6 +211,11 @@ void ggml_metal_free(ggml_metal_t ctx) {
     [ctx->cmd_bufs_ext removeAllObjects];
     [ctx->cmd_bufs_ext release];
 
+    @autoreleasepool {
+        [ctx->buf_refs removeAllObjects];
+        [ctx->buf_refs release];
+    }
+
     if (ctx->pipelines_ext) {
         ggml_metal_pipelines_free(ctx->pipelines_ext);
         ctx->pipelines_ext = nil;
@@ -294,6 +304,10 @@ void ggml_metal_synchronize(ggml_metal_t ctx) {
 
         [ctx->cmd_bufs_ext removeAllObjects];
     }
+
+    @autoreleasepool {
+        [ctx->buf_refs removeAllObjects];
+    }
 }
 
 static struct ggml_metal_buffer_id ggml_metal_get_buffer_id(const struct ggml_tensor * t) {
@@ -337,6 +351,8 @@ void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor,
 
         [encoder endEncoding];
         [cmd_buf commit];
+
+        [ctx->buf_refs addObject:buf_src];
         [buf_src release];
 
         // do not wait here for completion
@@ -381,6 +397,8 @@ void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * te
 
         [encoder endEncoding];
         [cmd_buf commit];
+
+        [ctx->buf_refs addObject:buf_dst];
         [buf_dst release];
 
         // do not wait here for completion

ggml/src/ggml-metal/ggml-metal-device.m

@@ -1346,6 +1346,7 @@ ggml_metal_device_t ggml_metal_device_init(int device, int n_devices) {
 void ggml_metal_device_free(ggml_metal_device_t dev) {
     assert(dev != NULL);
 
+    @autoreleasepool {
         ggml_metal_rsets_free(dev->rsets);
 
         ggml_metal_library_free(dev->library);
@@ -1360,6 +1361,7 @@ void ggml_metal_device_free(ggml_metal_device_t dev) {
             [dev->mtl_device release];
             dev->mtl_device = nil;
         }
+    }
 
     free(dev);
 }
@@ -1446,12 +1448,14 @@ ggml_metal_event_t ggml_metal_device_event_init(ggml_metal_device_t dev) {
 }
 
 void ggml_metal_device_event_free(ggml_metal_device_t dev, ggml_metal_event_t ev) {
+    @autoreleasepool {
         id<MTLSharedEvent> event = ev->obj;
         [event release];
 
         free(ev);
 
         GGML_UNUSED(dev);
+    }
 }
 
 void ggml_metal_device_event_synchronize(ggml_metal_device_t dev, ggml_metal_event_t ev) {
@@ -2226,6 +2230,7 @@ ggml_metal_buffer_t ggml_metal_buffer_map(ggml_metal_device_t dev, void * ptr, s
 }
 
 void ggml_metal_buffer_free(ggml_metal_buffer_t buf) {
+    @autoreleasepool {
         ggml_metal_device_rsets_rm(buf->dev, buf->rset);
 
         for (int i = 0; i < buf->n_buffers; i++) {
@@ -2233,6 +2238,7 @@ void ggml_metal_buffer_free(ggml_metal_buffer_t buf) {
         }
 
         ggml_metal_buffer_rset_free(buf);
+    }
 
     if (buf->is_shared && buf->owned) {
 #if TARGET_OS_OSX