llama.cpp b11530 Fixes Sampling Graph Reallocation Aborts
llama.cpp release b11530 updates the backend sampling logic to maintain a static computation graph topology across micro-batches. This change prevents execution aborts when running with the GGML_SCHED_NO_REALLOC scheduler flag. Maintaining a static computation graph is crucial for predictable memory management and high performance on hardware backends during LLM inference. By eliminating dynamic topology changes between memory allocation and decoding phases, inference execution remains stable without requiring runtime buffer reallocations. Previously, reservation built n_outputs_max_per_seq sampling chains while decode built only one per output row, changing graph topology and triggering GGML_SCHED_NO_REALLOC aborts. Samplers now construct n_outputs_max_per_seq chains consistently by assigning unselected ubatch slots to padding rows, allowing graph_max_nodes to count them accurately.
## BACKGROUND
llama.cpp is a widely used C/C++ framework for running Large Language Models efficiently on consumer hardware. During inference execution, computation steps are built as an operational graph where memory buffers can be dynamically or statically allocated by the GGML runtime scheduler.