The llama.cpp project released version b10839, which resolves hard crashes in the Vulkan backend caused by misaligned offsets during tensor row retrieval operations. Previously, models like Qwen3-TTS and Qwen3-VL would fail when using `ggml_view` with non-zero view offsets because the shader asserted on alignment violations relative to `minStorageBufferOffsetAlignment`. This update implements native support for aligned offsets in both quantized and non-quantized paths, eliminating the need for CPU fallbacks.
Key technical changes include:
- Binding buffer offsets to an aligned position near the view offset and passing adjusted misalignment via push constants to prevent truncation errors.
- Adding a `use_view_offs` parameter to `ggml_vk_tensor_subbuffer` to correctly handle physical offsets on both UMA and discrete GPUs.
- Introducing `ggml_vk_get_adjusted_misalign` to find the smallest valid misalignment that satisfies storage buffer alignment requirements.
- Removing the previous defensive CPU fallback in `supports_op()` since the Vulkan backend now handles these cases natively.
- Adding comprehensive backend test coverage for non-zero view offsets across multiple data types, with all 223 GET_ROWS tests passing on NVIDIA RTX 5060 Ti.
This fix ensures stable execution of models relying on complex tensor views and KV cache slices on Vulkan hardware without requiring CPU offloading.