context shift even more lenient

gpttype_adapter.cpp

@@ -620,8 +620,8 @@ void PurgeMissingTokens(llama_context * ctx, std::vector<int> &current_context_t
     //if passed, save beginning of LCQ from old ctx as p1
     //remove all tokens from old ctx between p0 and p1, updating both arrays and kv, then continue as normal
 
-    const int ShortfallThreshold = 200 + (nctx/20); //dont trigger shifting if the distance between trimstart and currhead < this
-    const int SlackAllowance = 50 + (nctx/50); //in case the end text is slightly modified, be forgiving
+    const int ShortfallThreshold = 200 + (nctx/30); //dont trigger shifting if the distance between trimstart and currhead < this
+    const int SlackAllowance = 60 + (nctx/50); //in case the end text is slightly modified, be forgiving
 
     int trimstart = 0;
     int new_tokens_len = new_context_tokens.size();
@@ -650,7 +650,7 @@ void PurgeMissingTokens(llama_context * ctx, std::vector<int> &current_context_t
     }
 
     //at least this many tokens need to match, otherwise don't bother trimming
-    const int LCSTokThreshold = std::max(std::min((new_tokens_len - trimstart) - (genamt+SlackAllowance), (int)(nctx*0.5)), ShortfallThreshold-SlackAllowance);
+    const int LCSTokThreshold = std::max(std::min((new_tokens_len - trimstart) - (genamt+SlackAllowance), (int)(nctx*0.45)), ShortfallThreshold-SlackAllowance);
 
     auto curr_ctx_without_memory = std::vector<int>(current_context_tokens.begin() + trimstart, current_context_tokens.end());
     auto new_ctx_without_memory = std::vector<int>(new_context_tokens.begin() + trimstart, new_context_tokens.end());