diff --git a/gpttype_adapter.cpp b/gpttype_adapter.cpp index 55ba69a12..3d131a8ce 100644 --- a/gpttype_adapter.cpp +++ b/gpttype_adapter.cpp @@ -551,7 +551,7 @@ void ContextRewind(std::vector &embd, std::vector ¤t_context_tok } if(file_format == FileFormat::RWKV_1 || file_format==FileFormat::RWKV_2 || is_recurrent) { - if(!showed_rnn_warning) + if(!showed_rnn_warning && debugmode==1 && !is_quiet) { showed_rnn_warning = true; printf("\nWARNING: RNN models do not support context rewind!\n"); diff --git a/otherarch/acestep/ace-qwen3.cpp b/otherarch/acestep/ace-qwen3.cpp index 32e16cd90..87d500bd5 100644 --- a/otherarch/acestep/ace-qwen3.cpp +++ b/otherarch/acestep/ace-qwen3.cpp @@ -1086,7 +1086,7 @@ static std::vector run_phase2_batch( for (int v = 0; v < AUDIO_CODE_BASE; v++) if (v != TOKEN_IM_END) lg[v] = -1e9f; - int tok = kcpp_quick_sample(lg.data(),V,std::vector(),1.02f,top_p,30,temperature,acestep_lm_rng); + int tok = kcpp_quick_sample(lg.data(),V,std::vector(),1.03f,top_p,30,temperature,acestep_lm_rng); seqs[i].last_token = tok; if (tok == TOKEN_IM_END) { @@ -1124,6 +1124,7 @@ static std::vector run_phase2_batch( for (int i = 0; i < N; i++) if (seqs[i].done) n_active--; + std::vector quicklastntoks; for (int step = 0; step < max_tokens && n_active > 0; step++) { // Collect tokens (done sequences feed their last token, result ignored) for (int i = 0; i < N; i++) @@ -1157,7 +1158,11 @@ static std::vector run_phase2_batch( for (int v = 0; v < AUDIO_CODE_BASE; v++) if (v != TOKEN_IM_END) lc[v] = -1e9f; - int tok = kcpp_quick_sample(lc,V,std::vector(),1.02f,top_p,30,temperature,acestep_lm_rng); + int tok = kcpp_quick_sample(lc,V,quicklastntoks,1.03f,top_p,30,temperature,acestep_lm_rng); + quicklastntoks.push_back(tok); + if (quicklastntoks.size()>32) { + quicklastntoks.erase(quicklastntoks.begin()); + } seqs[i].last_token = tok; if (tok == TOKEN_IM_END) {