llama.cpp found its Vulkan flash-attention dequant path only ever engaged when the KV cache was full

GitHub·medium signalXBlueskyLinkedInCopy link

Build b10781, published 2026-09-03T09:57Z, fixes PR #28190: the FA dequant path skipped engagement because of an `nb[3]` stride check the shader never reads when `ne[3] == 1`. Cache views carry the full-buffer stride there, so the old check collapsed to `n_kv == kv_size` and the optimized path only turned on with the cache completely full. Anyone benchmarking Vulkan flash attention on llama.cpp before this build was measuring the unoptimized path for nearly every real workload. Build b10782 separately unblocks CUDA graph optimization for multi-GPU under `GGML_CUDA_GRAPH_OPT=1`.