💬 Reddit 13h ago
llama.cpp speculative checkpointing was merged
llama.cpp merged speculative checkpointing support achieving 0-50% speedup on coding tasks with optimized parameters, though performance varies by prompt repetition patterns and draft acceptance rates. The feature uses n-gram matching for speculative decoding with configurable draft token ranges.