LLM inference in C/C++
C++ ggml
metal : fix memory query under low-memory conditions (#27701)
Mads Marquart8c0b9cd04ac7
diff --git a/ggml/src/ggml-metal/ggml-metal-device.m b/ggml/src/ggml-metal/ggml-metal-device.mindex ef81084241..e20a4e8916 100644--- a/ggml/src/ggml-metal/ggml-metal-device.m+++ b/ggml/src/ggml-metal/ggml-metal-device.m@@ -1471,8 +1471,10 @@ void ggml_metal_device_event_synchronize(ggml_metal_device_t dev, ggml_metal_evevoid ggml_metal_device_get_memory(ggml_metal_device_t dev, size_t * free, size_t * total) {if (@available(macOS 10.12, iOS 16.0, *)) {- *total = dev->mtl_device.recommendedMaxWorkingSetSize;- *free = *total - dev->mtl_device.currentAllocatedSize;+ *total = dev->mtl_device.recommendedMaxWorkingSetSize;+ size_t cur = dev->mtl_device.currentAllocatedSize;+ // it's possible to allocate more than `recommendedMaxWorkingSetSize`+ *free = *total > cur ? *total - cur : 0;} else {*free = 0;*total = 0;