Deprecation of CPU fallback for Vulkan GET_ROWS in ggml-org/llama.cpp

The latest release of ggml-org/llama.cpp introduces significant updates to the Vulkan backend, particularly addressing the handling of misaligned offsets in the GET_ROWS operation. According to the project's GitHub release notes, previous versions relied on a CPU fallback mechanism when misaligned offsets caused crashes during tensor operations, such as those encountered with models like Qwen3-TTS and Qwen3-VL. This fallback has now been deprecated, as the Vulkan backend has been enhanced to natively handle misaligned offsets. The update includes fixes to prevent truncation errors for quantized block types and ensures consistent handling across both UMA and discrete GPUs. Additionally, extensive backend tests were added, confirming the reliability of these changes across various tensor types and configurations.

Operators planning to upgrade should carefully evaluate the compatibility of their existing models, especially if they previously relied on the CPU fallback mechanism for misaligned offsets. Models that depend on older patterns of tensor alignment may encounter issues if the Vulkan backend's new handling does not align with their specific requirements. Testing models in a staging environment with Vulkan-enabled configurations is highly recommended to ensure smooth operation post-upgrade. This change reflects a broader pattern in AIOps projects, where reliance on fallback mechanisms is being reduced in favor of more robust native handling, emphasizing the importance of proactive compatibility checks during upgrades.