llama.cpp

fix: speculative decoding broken on hybrid SSM/MoE (Qwen3.5 MoE)

#20075

Closed