Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-18 16:55:14 +02:00
Code Issues Packages Projects Releases Wiki Activity

contextshift more forgiving

Browse Source
This commit is contained in:
Concedo
2024-02-10 20:49:21 +08:00
parent 0ec0055edc
commit 590af480ab
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
gpttype_adapter.cpp
+2 -2
View File
@@ -621,7 +621,7 @@ void PurgeMissingTokens(llama_context * ctx, std::vector<int> &current_context_t
//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/60); //in case the end text is slightly modified, be forgiving
const int SlackAllowance = 50 + (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*0.55)), ShortfallThreshold-SlackAllowance);
const int LCSTokThreshold = std::max(std::min((new_tokens_len - trimstart) - (genamt+SlackAllowance), (int)(nctx*0.5)), 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());