LLM inference in C/C++
C++ ggml
metal : fix memory leaks and cleanup logic in GPU capture path (#30890)
Niklas Wenzelea9e3446b5c8
diff --git a/ggml/src/ggml-metal/ggml-metal-context.m b/ggml/src/ggml-metal/ggml-metal-context.mindex e9064666f6..597c126ca3 100644--- a/ggml/src/ggml-metal/ggml-metal-context.m+++ b/ggml/src/ggml-metal/ggml-metal-context.m@@ -539,8 +539,14 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *descriptor.outputURL = [NSURL fileURLWithPath:path];NSError * error = nil;- if (![[MTLCaptureManager sharedCaptureManager] startCaptureWithDescriptor:descriptor error:&error]) {+ const bool ok = [[MTLCaptureManager sharedCaptureManager] startCaptureWithDescriptor:descriptor error:&error];++ [descriptor release];++ if (!ok) {GGML_LOG_ERROR("%s: error: unable to start capture '%s' (did you set METAL_CAPTURE_ENABLED=1 ?)\n", __func__, [[error localizedDescription] UTF8String]);+ [ctx->capture_scope release];+ ctx->capture_scope = nil;} else {[ctx->capture_scope beginScope];ctx->capture_started = true;@@ -600,6 +606,8 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *// enter here only when capturing in order to wait for all computation to finish// otherwise, we leave the graph to compute asynchronouslyif (use_capture && ctx->capture_started) {+ enum ggml_status res = GGML_STATUS_SUCCESS;+// wait for completion and check status of each command buffer// needed to detect if the device ran out-of-memory for example (#1881){@@ -613,11 +621,11 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *GGML_LOG_INFO("error: %s\n", [[cmd_buf error].localizedDescription UTF8String]);}- return GGML_STATUS_FAILED;+ res = GGML_STATUS_FAILED;}}- for (int i = 0; i < n_cb; ++i) {+ for (int i = 0; i < n_cb && res == GGML_STATUS_SUCCESS; ++i) {id<MTLCommandBuffer> cmd_buf = ctx->cmd_bufs[i].obj;[cmd_buf waitUntilCompleted];@@ -627,8 +635,8 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *if (status == MTLCommandBufferStatusError) {GGML_LOG_INFO("error: %s\n", [[cmd_buf error].localizedDescription UTF8String]);}-- return GGML_STATUS_FAILED;+ res = GGML_STATUS_FAILED;+ break;}id<MTLCommandBuffer> next_buffer = (i + 1 < n_cb ? ctx->cmd_bufs[i + 1].obj : nil);@@ -643,7 +651,8 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *if (ctx->abort_callback && ctx->abort_callback(ctx->abort_callback_data)) {GGML_LOG_INFO("%s: command buffer %d aborted", __func__, i);- return GGML_STATUS_ABORTED;+ res = GGML_STATUS_ABORTED;+ break;}[next_buffer commit];@@ -652,7 +661,14 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *[ctx->capture_scope endScope];[[MTLCaptureManager sharedCaptureManager] stopCapture];+ [ctx->capture_scope release];+ ctx->capture_scope = nil;+ctx->capture_started = false;++ if (res != GGML_STATUS_SUCCESS) {+ return res;+ }}}