Skip to content

LLM inference in C/C++

C++ ggml

vulkan: Add workaround for device free memory > total memory on Intel Windows (#29835)

Masato Nakasakacdce056a3e88
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index f996674f4c..1496dd47a8 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -15265,7 +15265,10 @@ void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total
*total += heap.size;
if (membudget_supported && i < budgetprops.heapUsage.size()) {
- *free += budgetprops.heapBudget[i] - budgetprops.heapUsage[i];
+ // Driver may report heapBudget > heap.size so we add a safeguard here
+ // See https://github.com/ggml-org/llama.cpp/issues/29277
+ const vk::DeviceSize budget = std::min(budgetprops.heapBudget[i], heap.size);
+ *free += budget > budgetprops.heapUsage[i] ? budget - budgetprops.heapUsage[i] : 0;
} else {
*free += heap.size;
}