metal : use autoreleasepool to avoid memory leaks (ggml-org#5437)

irbull · web-flow · commit f026f8120f97 · 2024-02-10T12:53:28.000+02:00
There appears to be a known memory leak when using the `MLTCommandBuffer`. It is suggested to use `@autoreleasepool` in [1,2] [1] https://developer.apple.com/forums/thread/662721 [2] https://forums.developer.apple.com/forums/thread/120931 This change-set wraps the `ggml_metal_graph_compute` in a `@autoreleasepool`. This commit addresses ggml-org#5436
diff --git a/ggml-metal.m b/ggml-metal.m
@@ -687,6 +687,7 @@ static bool ggml_metal_graph_compute(
         struct ggml_metal_context * ctx,
                struct ggml_cgraph * gf) {
 
+    @autoreleasepool {
     MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;
     edesc.dispatchType = MTLDispatchTypeSerial;
 
@@ -2272,6 +2273,7 @@ static bool ggml_metal_graph_compute(
         [[MTLCaptureManager sharedCaptureManager] stopCapture];
     }
 
+    }
     return true;
 }
 

Original file line number	Diff line number	Diff line change
`@@ -687,6 +687,7 @@ static bool ggml_metal_graph_compute(`
`687`	`687`	`struct ggml_metal_context * ctx,`
`688`	`688`	`struct ggml_cgraph * gf) {`
`689`	`689`
	`690`	`+ @autoreleasepool {`
`690`	`691`	`MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;`
`691`	`692`	`edesc.dispatchType = MTLDispatchTypeSerial;`
`692`	`693`
`@@ -2272,6 +2273,7 @@ static bool ggml_metal_graph_compute(`
`2272`	`2273`	`[[MTLCaptureManager sharedCaptureManager] stopCapture];`
`2273`	`2274`	`}`
`2274`	`2275`
	`2276`	`+ }`
`2275`	`2277`	`return true;`
`2276`	`2278`	`}`
`2277`	`2279`