llama.cpp fixes graph scheduling when switching token and image batches

Graph input collection becomes stable across batch types

The llama.cpp b11254 development build collects all input tensors into graph_inputs after graph splitting. The release notes explain that switching between batches using different inputs, such as token and image batches, could previously change copied graph composition, trigger unnecessary scheduler re-reservations and later cause an abort. The change makes composition depend on which inputs exist. This prerelease fix is relevant to pipeline-parallel and multimodal inference workloads.

• github.com — ggml-org/llama.cpp b11254 release notes (opens in a new tab)