metal : fix more leaks due to missing autoreleasepools

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 * bufs_release_later;
+
     // the last command buffer queued into the Metal queue with operations relevant to the current Metal backend
     id<MTLCommandBuffer> cmd_buf_last;
 
@@ -178,7 +182,8 @@ ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
             res->cmd_bufs[i].obj = nil;
         }
 
-        res->cmd_bufs_ext = [[NSMutableArray alloc] init];
+        res->cmd_bufs_ext       = [[NSMutableArray alloc] init];
+        res->bufs_release_later = [[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->bufs_release_later removeAllObjects];
+        [ctx->bufs_release_later 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->bufs_release_later 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->bufs_release_later 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->bufs_release_later addObject:buf_dst];
         [buf_dst release];
 
         // do not wait here for completion

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

@@ -1279,19 +1279,21 @@ 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);
 
-    ggml_metal_rsets_free(dev->rsets);
+    @autoreleasepool {
+        ggml_metal_rsets_free(dev->rsets);
 
-    ggml_metal_library_free(dev->library);
-    dev->library = NULL;
+        ggml_metal_library_free(dev->library);
+        dev->library = NULL;
 
-    if (dev->mtl_queue) {
-        [dev->mtl_queue release];
-        dev->mtl_queue = nil;
-    }
+        if (dev->mtl_queue) {
+            [dev->mtl_queue release];
+            dev->mtl_queue = nil;
+        }
 
-    if (dev->mtl_device) {
-        [dev->mtl_device release];
-        dev->mtl_device = nil;
+        if (dev->mtl_device) {
+            [dev->mtl_device release];
+            dev->mtl_device = nil;
+        }
     }
 
     free(dev);
@@ -2118,14 +2120,16 @@ 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) {
-    ggml_metal_device_rsets_rm(buf->dev, buf->rset);
+    @autoreleasepool {
+        ggml_metal_device_rsets_rm(buf->dev, buf->rset);
 
-    for (int i = 0; i < buf->n_buffers; i++) {
-        [buf->buffers[i].metal release];
+        for (int i = 0; i < buf->n_buffers; i++) {
+            [buf->buffers[i].metal release];
+        }
+
+        ggml_metal_buffer_rset_free(buf);
     }
 
-    ggml_metal_buffer_rset_free(buf);
-
     if (buf->is_shared && buf->owned) {
 #if TARGET_OS_OSX
         vm_deallocate((vm_map_t)mach_task_self(), (vm_address_t)buf->all_data, buf->all_size);