The llama.cpp project released build b10857, which includes a fix for Vulkan-Hpp handle usage on 32-bit platforms. This update addresses type safety issues where Vulkan non-dispatchable handles like VkBuffer are represented as uint64_t.
- Adds the operator<< for vk::Buffer to allow direct streaming into std::ostream for debug and memory logs.
- Passes vk::Buffer directly to Vulkan-Hpp CommandBuffer::copyBuffer APIs instead of casting it to VkBuffer.
This change ensures correct handling of Vulkan objects on 32-bit architectures, preventing potential errors in logging and buffer copy operations.