vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

revert auto rope scaling for already-ropetuned models - just use their values

Browse source
This commit is contained in:
Concedo 2023-11-24 14:20:36 +08:00
parent 56a5fa7a60
commit eb42c73953
3 changed files with 70 additions and 46 deletions
Download patch file Download diff file Expand all files Collapse all files

24
gpttype_adapter.cpp
View file

@ -941,19 +941,19 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
llamamodel->hparams.rope_freq_scale_train!=1.0f ||
llamamodel->hparams.rope_scaling_type_train==2)
{
float ropemultiplier = 1.0f;
if(llamamodel->hparams.rope_scaling_type_train!=2 &&
llamamodel->hparams.n_ctx_train > 2048 && clamped_max_context_length > llamamodel->hparams.n_ctx_train)
{
ropemultiplier = (float)llamamodel->hparams.n_ctx_train / (float)clamped_max_context_length;
llama_ctx_params.rope_freq_base = rope_freq_base = llamamodel->hparams.rope_freq_base_train;
llama_ctx_params.rope_freq_scale = rope_freq_scale = ropemultiplier * llamamodel->hparams.rope_freq_scale_train;
printf("Automatic RoPE Scaling: Using (scale:%.3f, base:%.1f).\n", rope_freq_scale, rope_freq_base);
}
else
{
// float ropemultiplier = 1.0f;
// if(llamamodel->hparams.rope_scaling_type_train!=2 &&
// llamamodel->hparams.n_ctx_train > 2048 && clamped_max_context_length > llamamodel->hparams.n_ctx_train)
// {
// ropemultiplier = (float)llamamodel->hparams.n_ctx_train / (float)clamped_max_context_length;
// llama_ctx_params.rope_freq_base = rope_freq_base = llamamodel->hparams.rope_freq_base_train;
// llama_ctx_params.rope_freq_scale = rope_freq_scale = ropemultiplier * llamamodel->hparams.rope_freq_scale_train;
// printf("Automatic RoPE Scaling: Using (scale:%.3f, base:%.1f).\n", rope_freq_scale, rope_freq_base);
// }
// else
// {
printf("Automatic RoPE Scaling: Using model internal value.\n");
}
//}
}
else
{

90
klite.embd
View file

@ -6,7 +6,7 @@ It requires no dependencies, installation or setup.
Just copy this single static HTML file anywhere and open it in a browser, or from a webserver.
Please go to https://github.com/LostRuins/lite.koboldai.net for updates on Kobold Lite.
Kobold Lite is under the AGPL v3.0 License unless otherwise exempted. Please do not remove this line.
Current version: 96
Current version: 97
-Concedo
-->
@ -1654,18 +1654,9 @@ Current version: 96
margin: 8px 0 8px;
}
.cht_inp_hold input {
border: medium none;
color: #bebebe;
font-size: 15px;
min-height: 36px;
/* width: 100%; */
outline:none;
}
.cht_inp
.cht_inp_bg
{
display: inline-block;
width: calc(100% - 84px);
background: #86868638 none repeat scroll 0 0;
margin-top: 8px;
@ -1673,8 +1664,23 @@ Current version: 96
border-radius: 16px;
padding-left: 10px;
padding-right: 10px;
padding-top: 7px;
}
.cht_inp.shorter
.cht_inp_bg_inner
{
width: 100%;
resize: none;
overflow-y:scroll;
overflow-x:hidden;
background: #00000000 none repeat scroll 0 0;
border: medium none;
color: #bebebe;
font-size: 15px;
/* min-height: 36px; */
/* width: 100%; */
outline:none;
}
.cht_inp_bg.shorter
{
width: calc(100% - 114px);
}
@ -3238,7 +3244,7 @@ Current version: 96
var image_db = {}; //stores a dictionary of pending images
var completed_imgs_meta = {}; //stores temp info on completed images like alt text
//key is ID, body is {done:false,queue:10,result:""}
var stablemodels = []; //stored as {name,count}
var stablemodels = [{"name": "stable_diffusion","count": 1}]; //stored as {name,count}
var custom_kobold_endpoint = ""; //if set, does not use horde. Instead, attempts to use this sync endpoint
var custom_oai_endpoint = "";
var custom_oai_key = ""; //if set, uses the OpenAI API to generate
@ -3599,11 +3605,18 @@ Current version: 96
.setAttribute("content","width=device-width, initial-scale=1, maximum-scale=1");
}
//fix for copy paste text in firefox
//fix for copy paste text in firefox, and also to prevent pasting rich text
{
document.getElementById("gametext").addEventListener("paste", function(e) {
e.preventDefault();
var text = (e.originalEvent || e).clipboardData.getData('text/plain');
let text = e.clipboardData
? (e.originalEvent || e).clipboardData.getData('text/plain')
: // For IE
window.clipboardData
? window.clipboardData.getData('Text')
: '';
text = escapeHtml(text);
text = text.replace(/\r?\n/g, '<br>');
document.execCommand("insertHTML", false, text);
});
@ -7081,9 +7094,9 @@ Current version: 96
{
let sdmodelshtml = "";
for (var i = 0; i < stablemodels.length; ++i) {
sdmodelshtml += "<option value=\"" + stablemodels[i].name + " (" + stablemodels[i].count + ")\">";
sdmodelshtml += "<option value=\"" + stablemodels[i].name + "\" "+(stablemodels[i].name==localsettings.generate_images_model?"selected":"")+">" + stablemodels[i].name + " (" + stablemodels[i].count + ")</option>";
}
document.getElementById("sdmodels").innerHTML = sdmodelshtml;
document.getElementById("generate_images_model").innerHTML = sdmodelshtml;
}
function toggle_preset() {
@ -7118,13 +7131,10 @@ Current version: 96
break;
}
}
if (!matched && inputmodel != "*") {
if (!matched) {
document.getElementById("generate_images_model").value = defaultsettings.generate_images_model;
}
}
function clear_sd_model() {
document.getElementById("generate_images_model").value = "";
}
function validate_samplers(savesetting = false) {
let samplerstr = document.getElementById("sampler_order").value;
@ -7665,6 +7675,7 @@ Current version: 96
current_anotetemplate = "[Author's note: <|>]";
document.getElementById("input_text").value = "";
document.getElementById("cht_inp").value = "";
chat_resize_input();
image_db = {};
completed_imgs_meta = {};
localsettings.adventure_is_action = false;
@ -7851,7 +7862,7 @@ Current version: 96
let id_to_cancel = pending_response_id;
//flush any streaming text first
if(is_using_custom_ep() && pending_response_id!="" && synchro_pending_stream!="")
if(is_using_custom_ep() && pending_response_id != "" && (synchro_pending_stream != "" || synchro_polled_response != ""))
{
//apply a short delay of 1s before button reenables
allow_reenable_submitbtn_timestamp = performance.now() + 500;
@ -7859,7 +7870,10 @@ Current version: 96
update_submit_button(true);
}, 1000);
synchro_polled_response = synchro_pending_stream;
if(synchro_pending_stream!="")
{
synchro_polled_response = synchro_pending_stream;
}
poll_in_progress = false;
horde_poll_nearly_completed = false;
poll_pending_response();
@ -10219,10 +10233,10 @@ Current version: 96
}
if (localsettings.opmode == 3 && localsettings.chatopponent != "") {
document.getElementById("chat_btnmode_chat").classList.remove("hidden");
document.getElementById("cht_inp").classList.add("shorter");
document.getElementById("cht_inp_bg").classList.add("shorter");
} else {
document.getElementById("chat_btnmode_chat").classList.add("hidden");
document.getElementById("cht_inp").classList.remove("shorter");
document.getElementById("cht_inp_bg").classList.remove("shorter");
}
// Show the 'AI is typing' message if an answer is pending, and prevent the 'send button' from being clicked again.
@ -10412,12 +10426,21 @@ Current version: 96
}
}
}
function chat_resize_input()
{
//resize chat inp
let textarea = document.getElementById("cht_inp");
let numberOfLineBreaks = (textarea.value.match(/\n/g) || []).length;
numberOfLineBreaks = numberOfLineBreaks>4?4:numberOfLineBreaks;
textarea.rows = numberOfLineBreaks+1;
}
function chat_submit_generation()
{
//easy solution is to just pump the text into the main box and submit it
document.getElementById("input_text").value = document.getElementById("cht_inp").value;
submit_generation();
document.getElementById("cht_inp").value = "";
chat_resize_input();
}
function chat_toggle_actionmenu()
{
@ -11430,15 +11453,16 @@ Current version: 96
<div class="cht_inp_hold_outer">
<div class="cht_inp_hold">
<button onclick="show_groupchat_select()" id="chat_btnmode_chat" class="chat_btnmode_chat hidden" type="button"></button>
<input id="cht_inp" type="text" name="chtchtinp" class="cht_inp" role="presentation" autocomplete="noppynop" spellcheck="true" placeholder="Type a message" value="" oninput="update_submit_button()" onkeypress="return chat_handle_typing(event)"/>
<div id="cht_inp_bg" class="cht_inp_bg">
<textarea class="cht_inp_bg_inner" id="cht_inp" type="text" name="chtchtinp" role="presentation" autocomplete="noppynop" spellcheck="true" rows="1" wrap="off" placeholder="Type a message" value="" oninput="update_submit_button();chat_resize_input();" onkeypress="return chat_handle_typing(event)"/></textarea>
</div>
<button onclick="chat_submit_generation()" id="chat_msg_send_btn" class="chat_msg_send_btn" type="button"></button>
<button onclick="abort_generation()" id="chat_msg_send_btn_abort" class="hidden chat_msg_send_btn_abort" type="button"></button>
<button type="button" class="chat_msg_cust_btn" id="btn_chat_cust" onclick="chat_toggle_actionmenu()"></button>
</div>
</div>
<div class="lastreq" id="lastreq2" style="padding-top: 4px; color:#999999"><span class="color_gray">Avoid sending privacy sensitive information. <a href="#" onclick="explain_horde()">Click here for more info</a>.</span></div>
<div class="lastreq" id="lastreq2" style="padding-top: 2px; color:#999999"><span class="color_gray">Avoid sending privacy sensitive information. <a href="#" onclick="explain_horde()">Click here for more info</a>.</span></div>
</div>
</div>
@ -11701,6 +11725,7 @@ Current version: 96
<option value="claude-instant-v1">claude-instant-v1</option>
<option value="claude-instant-v1-100k">claude-instant-v1-100k</option>
<option value="claude-2">claude-2</option>
<option value="claude-2.1">claude-2.1</option>
<option value="claude-2.0">claude-2.0</option>
</select>
<input type="checkbox" id="claudeaddversion" onchange="" checked>
@ -12101,10 +12126,9 @@ Current version: 96
<option value="2">Local A1111</option>
<option value="3">OpenAI DALL-E</option>
</select>
<input list="sdmodels" class="form-control mdlpicker hidden" id="generate_images_model" style="font-size: 12px;height:20px;padding:2px;margin:0px 0 0;" placeholder="[Select Model]" value="" onblur="validate_sd_model()" onfocus="clear_sd_model()" title="Select a stable diffusion model to generate images with">
<datalist id="sdmodels">
<option value="stable_diffusion">
</datalist>
<select class="form-control" id="generate_images_model" style="font-size: 12px;height:20px;padding:2px;margin:0px 0 0;" onblur="validate_sd_model()" title="Select a stable diffusion model to generate images with">
</select>
<div id="generate_images_local_model_container" class="settinglabel hidden">
<select class="form-control" id="generate_images_local_model" style="height:20px;padding:0;margin:0px 0 0; width:calc(100% - 30px)">
<option value="">[None]</option>

2
koboldcpp.py
View file

@ -388,7 +388,7 @@ maxhordelen = 256
modelbusy = threading.Lock()
requestsinqueue = 0
defaultport = 5001
KcppVersion = "1.50.1"
KcppVersion = "1.51"
showdebug = True
showsamplerwarning = True
showmaxctxwarning = True