llama.cpp fixes a Vulkan graph-optimizer bug that silently produced wrong tokens on AMD and NVIDIA with nothing logged

GitHub·high signalXBlueskyLinkedInCopy link

Build b10677 (2026-08-28) fixes `ggml_vk_graph_optimize`, where `is_src_of` did not treat two views of one tensor as dependent, so the optimizer reordered nodes across aliased reads and writes. The maintainers describe the result as silently wrong tokens under greedy decoding, different output on every server start, and invalid speculative-decoding acceptance, hitting Qwen3.8's recurrent state and any model with view-aliased state; CUDA was unaffected. A second commit stops treating NONE/RESHAPE/TRANSPOSE/VIEW/PERMUTE no-op nodes as aliasing dependencies, which only cost the optimizer reordering freedom.