Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-04-09 10:31:45 +02:00
Code Issues Packages Projects Releases Wiki Activity

updated lite, modify henk fix to allow triggering on missing close only

Browse Source
This commit is contained in:
Concedo
2026-04-06 23:41:45 +08:00
parent a395af65db
commit e991bc044e
2 changed files with 80 additions and 56 deletions
Download Patch File Download Diff File Expand all files Collapse all files

134
embd_res/klite.embd
View File

@@ -12,7 +12,7 @@ Current version indicated by LITEVER below.
-->
<head>
<script id="init-config">
const LITEVER = 326;
const LITEVER = 327;
const urlParams = new URLSearchParams(window.location.search);
var localflag = urlParams.get('local'); //this will be replaced automatically in embedded kcpp
const STORAGE_PREFIX = (localflag?"e_":"")+"kaihordewebui_";
@@ -3558,6 +3558,7 @@ Current version indicated by LITEVER below.
// whitelisted auto selected horde model names
const defaultmodels = ["gpt4all","supercot","pygmalion-6","pygmalion-v8","pygmalion-2","hermes","airoboros","chrono","wizard","mantis","vicuna","manticore","alpaca","myth","xwin","spicyboros","mlewd","mxlewd","westlake","anubis","skyfall","llama2","llama3","llama-2","llama-3-","llama-3.","mistral","maid","mixtral","estopia","fighter","fimbul","euryale","nemo","gemma","lunaris","stheno","magnum","cydonia","qwen2.5-32b","behemoth","exaone","glm4","glm-4","tutu","deepseek","tlacuilo","rocinante","-14B","-32B","-27B","-35B"];
const ignoredmodels = ["tinyllama","debug-","-1b","-270m"]; //blacklisted model names
const hardcoded_think_closers = ["</think>","<channel|>"];
const instructstartplaceholder = "\n{{[INPUT]}}\n";
const instructendplaceholder = "\n{{[OUTPUT]}}\n";
@@ -5868,14 +5869,14 @@ Current version indicated by LITEVER below.
const safeStop = escapeRegExp(localsettings.stop_thinking_tag);
return `${safeStart}([\\s\\S]*?)${safeStop}`;
}
function get_orphaned_thinking_regex()
function get_orphaned_thinking_regex(tag)
{
if(!localsettings.start_thinking_tag || !localsettings.stop_thinking_tag)
if(!tag)
{
return "";
}
const safeStart = escapeRegExp(get_instructendplaceholder());
const safeStop = escapeRegExp(localsettings.stop_thinking_tag);
const safeStop = escapeRegExp(tag);
return `${safeStart}([\\s\\S]*?)${safeStop}`;
}
@@ -6095,12 +6096,12 @@ Current version indicated by LITEVER below.
let promptEtIndex = prompt.lastIndexOf(et);
if(prompt && promptEtIndex!=-1 && localsettings.instruct_gentag)
{
prompt = replaceLast(prompt,et,localsettings.instruct_gentag);
prompt = replaceLast(prompt,et,localsettings.instruct_gentag.replaceAll("\\n", "\n"));
replaceDone = true;
}
if(!replaceDone && memory && memoryEtIndex!=-1 && localsettings.instruct_gentag)
{
memory = replaceLast(memory,et,localsettings.instruct_gentag);
memory = replaceLast(memory,et,localsettings.instruct_gentag.replaceAll("\\n", "\n"));
replaceDone = true;
}
}
@@ -7475,7 +7476,7 @@ Current version indicated by LITEVER below.
pending_toolcall_objs.push(pending_toolcall_last_obj);
pending_toolcall_last_obj = null;
}
if((synchro_polled_response=="" || (synchro_polled_response.includes("<think>") && synchro_polled_response.includes("</think>"))) && last_stop_reason=="tool_calls" && pending_toolcall_objs!=null && pending_toolcall_objs.length>0)
if(last_stop_reason=="tool_calls" && pending_toolcall_objs!=null && pending_toolcall_objs.length>0)
{
//a tool call was triggered instead
synchro_polled_resptoolcalls = [];
@@ -9569,10 +9570,7 @@ Current version indicated by LITEVER below.
}
if (localsettings.eos_ban_mode == 0) {
if (localsettings.opmode == 1) {
return true; //story mode always ban
}
else if (localsettings.opmode == 3 && !localsettings.allow_continue_chat) {
if (localsettings.opmode == 3 && !localsettings.allow_continue_chat) {
return false; //chat mode always unban unless cont allowed
}
else if (!input_was_empty) //if user input is not empty, ALWAYS unban EOS.
@@ -17979,13 +17977,23 @@ Current version indicated by LITEVER below.
}
else
{
if(localsettings.handle_mismatched_think && !inputtxt.includes(localsettings.start_thinking_tag) && inputtxt.includes(localsettings.stop_thinking_tag))
if(localsettings.opmode==4 && localsettings.handle_mismatched_think && !inputtxt.includes(localsettings.start_thinking_tag))
{
let chunks = inputtxt.split(localsettings.stop_thinking_tag);
if(chunks.length==2) //exactly 2 parts
let orphans = hardcoded_think_closers.slice();
orphans.push(localsettings.stop_thinking_tag);
for(let t=0;t<orphans.length;++t)
{
matches.push(chunks[0]);
inputtxt = `%ExpandBtn%${chunks[1]}`;
let currtag = orphans[t];
if(currtag && inputtxt.includes(currtag))
{
let chunks = inputtxt.split(currtag);
if(chunks.length==2) //exactly 2 parts
{
matches.push(chunks[0]);
inputtxt = `%ExpandBtn%${chunks[1]}`;
break;
}
}
}
}
}
@@ -18001,7 +18009,9 @@ Current version indicated by LITEVER below.
let matchiter = 0;
inputtxt = inputtxt.replace(/%ExpandBtn%(?:\n{0,2})?/g, function (m) {
let curr = matches[matchiter];
let expandedhtml = `<span><button type="button" title="Show Thoughts" class="btn btn-primary" style="font-size:12px;padding:2px 2px;" onclick="toggle_hide_thinking(this)">Show Thoughts (${curr.length} characters)</button><span class="color_lightgreen hidden"><br>${escape_html(curr)}\n</span><br></span>`;
curr = escape_html(curr);
curr = unescape_html_alternate(curr);
let expandedhtml = `<span><button type="button" title="Show Thoughts" class="btn btn-primary" style="font-size:12px;padding:2px 2px;" onclick="toggle_hide_thinking(this)">Show Thoughts (${curr.length} characters)</button><span class="color_lightgreen hidden"><br>${(curr)}\n</span><br></span>`;
if(!curr || curr.trim()=="" || curr.trim()=="/nothink" || curr.trim()==`${localsettings.start_thinking_tag}${localsettings.stop_thinking_tag}`)
{
expandedhtml = `<span><button type="button" title="No Thoughts" class="btn btn-primary" style="font-size:12px;padding:2px 2px;" onclick="">No Thoughts</button><br></span>`;
@@ -19840,11 +19850,21 @@ Current version indicated by LITEVER below.
truncated_context = truncated_context.replace(pat, "");
//fallback, check for orphaned think tags and remove those as well
if(localsettings.opmode==4 && localsettings.stop_thinking_tag && truncated_context.includes(localsettings.stop_thinking_tag))
if(localsettings.opmode==4)
{
let endmatcher = get_instructendplaceholder();
let pat = new RegExp(get_orphaned_thinking_regex(), "gm");
truncated_context = truncated_context.replace(pat, endmatcher);
let orphans = hardcoded_think_closers.slice();
orphans.push(localsettings.stop_thinking_tag);
for(let t=0;t<orphans.length;++t)
{
let currtag = orphans[t];
if(localsettings.opmode==4 && currtag && truncated_context.includes(currtag))
{
let endmatcher = get_instructendplaceholder();
let pat = new RegExp(get_orphaned_thinking_regex(currtag), "gm");
truncated_context = truncated_context.replace(pat, endmatcher);
break;
}
}
}
}
else if(localsettings.strip_thinking_mode==1) //strip except recent
@@ -19857,11 +19877,21 @@ Current version indicated by LITEVER below.
let pat = new RegExp(get_thinking_regex(), "gm"); // Remove all thinking before splitpoint
beforePart = beforePart.replace(pat, "");
//fallback, check for orphaned think tags and remove those as well
if(localsettings.opmode==4 && localsettings.stop_thinking_tag && beforePart.includes(localsettings.stop_thinking_tag))
if(localsettings.opmode==4)
{
let pat = new RegExp(get_orphaned_thinking_regex(), "gm");
beforePart = beforePart.replace(pat, endmatcher);
//fallback, check for orphaned think tags and remove those as well
let orphans = hardcoded_think_closers.slice();
orphans.push(localsettings.stop_thinking_tag);
for(let t=0;t<orphans.length;++t)
{
let currtag = orphans[t];
if(localsettings.opmode==4 && currtag && beforePart.includes(currtag))
{
let pat = new RegExp(get_orphaned_thinking_regex(currtag), "gm");
beforePart = beforePart.replace(pat, endmatcher);
break;
}
}
}
truncated_context = beforePart + afterPart; // Combine and done
@@ -20524,22 +20554,8 @@ Current version indicated by LITEVER below.
{
seqs.push(me + "\:");
}
if(localsettings.chatopponent!="")
{
//NOTE: we do not add our opponents name as a stop sequence IF thinking is forced
//the model just gets too confused and is likely to repeat the tag.
if(localsettings.think_injected!=1)
{
let m_opps = localsettings.chatopponent.split("||$||");
for(let i=0;i<m_opps.length;++i)
{
if(m_opps[i] && m_opps[i].trim()!="")
{
seqs.push(m_opps[i] + "\:");
}
}
}
}
//NOTE: we do not add our opponents name as a stop sequence
//the model just gets too confused and is likely to repeat the tag.
}
}
//special case for GPT-OSS, never use it as a stop sequence or stuff gets messed
@@ -20979,6 +20995,14 @@ Current version indicated by LITEVER below.
mhistory.splice(1, 1);
}
//if the final turn is the assistant chatname only, we shift that into the user turn, that helps support models that reject prefills
if(mhistory.length>2 && !mhistory[mhistory.length-1].myturn && mhistory[mhistory.length-2].myturn &&
mhistory[mhistory.length-2].msg && mhistory[mhistory.length-1].msg == `${localsettings.chatopponent}:`)
{
mhistory[mhistory.length-2].msg += `\n(Continue ${localsettings.chatopponent}'s reply')`;
mhistory.pop();
}
if(mhistory.length>0 && mainoaibody.length>1 && mainoaibody[1].type && mainoaibody[1].type!="text")
{
mhistory[0].msg = [{type: 'text', text: mhistory[0].msg}];
@@ -21109,7 +21133,7 @@ Current version indicated by LITEVER below.
//pollinations always uses the exact same url for text gen regardless
if(targetep.toLowerCase().includes("text.pollinations.ai"))
{
targetep = apply_proxy_url(pollinations_text_endpoint);
targetep = apply_proxy_url(pollinations_text_endpoint,true);
oai_payload.private = true;
oai_payload.referrer = "koboldai";
oai_payload.seed = Math.floor(Math.random() * 99999999);
@@ -22480,7 +22504,7 @@ Current version indicated by LITEVER below.
//allow trim incomplete sentences
//do not trim if instruct mode AND stop token reached
let donottrim = ((localsettings.opmode == 4||localsettings.opmode == 3) && last_stop_reason=="stop");
let donottrim = (last_stop_reason=="stop");
if (!donottrim && localsettings.trimsentences == true) {
//also, to prevent a trim from bisecting a chat name, if a response contains a chatname, do not trim
donottrim = false;
@@ -22732,11 +22756,11 @@ Current version indicated by LITEVER below.
gentxt = gentxt.slice(et2.length);
found = gentxt.indexOf(et2);
}
if (found != -1) //if found, truncate to it
{
splitresponse = gentxt.split(et2);
gentxt = splitresponse[0];
}
// if (found != -1) //if found, truncate to it
// {
// splitresponse = gentxt.split(et2);
// gentxt = splitresponse[0];
// }
}
}
}
@@ -29783,13 +29807,13 @@ Current version indicated by LITEVER below.
<div style="display:flex;width:100%;">
<div class="settinglabel settingcell">
<div title="Sampler Order" class="justifyleft" style="width:100%">Sampler Order <span class="helpicon">?<span class="helptext">
<div class="justifyleft" style="width:100%">Sampler Order <span class="helpicon">?<span class="helptext">
The order by which all 7 samplers are applied, separated by commas. 0=top_k, 1=top_a, 2=top_p, 3=tfs, 4=typ, 5=temp, 6=rep_pen</span></span></div>
<div class="justifyleft" style="width:100%;">
<input title="Sampler Order List" class="settinglabel miniinput" type="text" placeholder="CSV" value="" id="sampler_order" onblur="validate_samplers()"></div>
</div>
<div class="settinglabel settingcell">
<div title="Top N Sigma" class="justifyleft" style="width:100%">T.NSigma <span class="helpicon">?<span class="helptext">
<div class="justifyleft" style="width:100%">T.NSigma <span class="helpicon">?<span class="helptext">
Top N-Sigma. 0 to deactivate. Range 0 to 5. Not supported on Horde.</span></span></div>
<div id="nsigmasupporteddiv" style="display:flex;">
<div class="justifyleft" style="width:100%">
@@ -29831,7 +29855,7 @@ Current version indicated by LITEVER below.
</div>
</div>
<div class="settinglabel settingcell">
<div title="XTC" class="justifyleft" style="width:100%">XTC <span class="helpicon">?<span class="helptext">
<div class="justifyleft" style="width:100%">XTC <span class="helpicon">?<span class="helptext">
Enables Exclude Top Choices (XTC) Sampling. May not be available depending on backend, not supported on Horde.</span></span></div>
<div class="justifyleft" style="width:100%">
<div id="xtcsupporteddiv" style="display:flex;">
@@ -29853,7 +29877,7 @@ Current version indicated by LITEVER below.
<div style="display:flex;width:100%;">
<div class="settinglabel settingcell">
<div title="Adaptive-P" class="justifyleft" style="width:100%">Adapt.P. <span class="helpicon">?<span class="helptext">
<div class="justifyleft" style="width:100%">Adapt.P. <span class="helpicon">?<span class="helptext">
Enables Adaptive-P Sampling. Target ranges 0.0 to 1.0, 0 to deactivate. Decay ranges from 0.0 to 1.0. Not supported on Horde.</span></span></div>
<div class="justifyleft" style="width:100%">
<div id="adaptpsupporteddiv" style="display:flex;">
@@ -29872,7 +29896,7 @@ Current version indicated by LITEVER below.
</div>
</div>
<div class="settinglabel settingcell">
<div title="Mirostat" class="justifyleft" style="width:100%">Mirostat <span class="helpicon">?<span class="helptext">
<div class="justifyleft" style="width:100%">Mirostat <span class="helpicon">?<span class="helptext">
Replaces your samplers with mirostat, an alternative sampling method. May not be available depending on backend, not supported on Horde.</span></span></div>
<div class="justifyleft" style="width:100%">
<div id="mirosupporteddiv" style="display:flex;">
@@ -31450,8 +31474,8 @@ Current version indicated by LITEVER below.
<div>
<div style="vertical-align: middle;">
<div>
<span class="color_red">Also Reset Memory and World Info? </span>
<input type="checkbox" id="new_session_erase_memory" style="vertical-align: top;">
<span class="color_red"><b>Also Reset Memory and World Info? </b></span>
<input type="checkbox" id="new_session_erase_memory" style="vertical-align: text-bottom;">
</div>
</div>
</div>

2
gpttype_adapter.cpp
View File

@@ -3865,7 +3865,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
const bool has_close = fullbody.find(channel_close) != std::string::npos;
// If neither opening nor closing tag is present anywhere, prepend both
if (!has_open && !has_close) {
if (!has_close) { //edit: even if we have open, lacking a close can break some cases. Add it.
addedmemory = channel_prefix + addedmemory;
}
}