Eval bug: llama_memory_seq_rm returns true but doesn’t restore the recurrent state when rolling back entire last decode

该报错出现在 llama.cpp 的 recurrent/hybrid 模型上,当调用 llama_memory_seq_rm 回滚整批最后一段 decode(且 n_rs_seq > 0 )时,函数返回 true 但没有恢复 recurrent state,导致后续解码结果与“未回滚”不一致。优先




![unfuse_lora(components=[...]) desyncs _merged_adapters/num_fused_loras from actual per-component PEFT merge state](https://www.chat-gpts.plus/wp-content/uploads/2026/08/14214-8825d7a6-768x403.jpg)



