Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-03 11:35:48 +02:00
Code Issues Packages Projects Releases Wiki Activity

lots of tweaks for ace step

Browse Source
This commit is contained in:
Concedo
2026-03-11 23:57:52 +08:00
parent ecc4865244
commit 211d4fe632
8 changed files with 144 additions and 476 deletions
Download Patch File Download Diff File Expand all files Collapse all files
embd_res/kcpp_docs.embd
+6 -4
View File
@@ -1685,14 +1685,16 @@
"vocal_language": "en",
"task_type": "text2music",
"seed": 622315,
"thinking": false,
"lm_temperature": 0.85,
"lm_cfg_scale": 2.0,
"lm_top_p": 0.9,
"lm_temperature": 1.0,
"lm_cfg_scale": 3.0,
"lm_top_p": 0.95,
"lm_top_k": 40,
"lm_rep_pen": 1.03,
"lm_negative_prompt": "",
"inference_steps": 8,
"guidance_scale": 1.0,
"shift": 3.0,
"rewrite_caption": true,
"audio_codes": ""
},
"schema": {
embd_res/kcpp_musicui.embd
+12 -4
View File
@@ -67,7 +67,7 @@ label{font-size:12px;color:var(--muted);}
}
.compact-row{
display:grid;
grid-template-columns:repeat(auto-fit,minmax(100px,1fr));
grid-template-columns:repeat(auto-fit,minmax(80px,1fr));
gap:8px;
}
button{
@@ -173,7 +173,8 @@ input[type="checkbox"] {
<div class="panel">
<h2>Song Setup</h2>
<label>Caption</label>
<label>Caption</label> <div style="float:right"><input id="rewrite_caption" type="checkbox" style="width: auto;vertical-align: middle;" checked><label>Rewrite Caption</label></div>
<input id="caption" placeholder="Describe the song">
<div style="margin-top:10px">
@@ -196,7 +197,12 @@ input[type="checkbox"] {
<div class="compact-row" style="margin-top:10px">
<div><label>Temp</label><input id="lm_temperature" type="number" step="0.01"></div>
<div><label>CFG</label><input id="lm_cfg_scale" type="number" step="0.1"></div>
<div><label>Top P</label><input id="lm_top_p" type="number" step="0.01"></div>
<div><label>Top-P</label><input id="lm_top_p" type="number" step="0.01"></div>
<div><label>Top-K</label><input id="lm_top_k" type="number"></div>
<div><label>RepPen</label><input id="lm_rep_pen" type="number"></div>
<div><label>Codes Top-P</label><input id="codes_top_p" type="number"></div>
<div><label>Codes Top-K</label><input id="codes_top_k" type="number"></div>
<div><label>Codes Temp</label><input id="codes_temperature" type="number"></div>
<div><label>Steps</label><input id="inference_steps" type="number"></div>
<div><label>Guidance</label><input id="guidance_scale" type="number"></div>
<div><label>Shift</label><input id="shift" type="number"></div>
@@ -306,7 +312,8 @@ function toggleAdvanced(){
function getFormData(){
const ids=["caption","lyrics","bpm","duration","keyscale","timesignature",
"vocal_language","seed","lm_temperature","lm_cfg_scale","lm_top_p","inference_steps",
"vocal_language","seed","lm_temperature","lm_cfg_scale","lm_top_p","lm_top_k","lm_rep_pen","inference_steps",
"codes_top_p","codes_top_k","codes_temperature",
"guidance_scale","shift","audio_codes"];
const data={};
ids.forEach(id=>{
@@ -317,6 +324,7 @@ function getFormData(){
});
data["stereo"] = (document.getElementById("stereo").checked ? true : false);
data["gen_codes"] = (document.getElementById("gen_codes").checked ? true : false);
data["rewrite_caption"] = (document.getElementById("rewrite_caption").checked ? true : false);
return data;
}
embd_res/klite.embd
+19 -190
View File
@@ -3548,7 +3548,7 @@ Current version indicated by LITEVER below.
const favivon_normal =rootStyles.getPropertyValue('--img_favicon_normal').match(/url\("(.*)"\)/)[1];
// 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"];
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 instructstartplaceholder = "\n{{[INPUT]}}\n";
@@ -4138,7 +4138,6 @@ Current version indicated by LITEVER below.
const alltalk_voices_endpoint = "/api/voices";
const alltalk_rvc_voices_endpoint = "/api/rvcvoices";
const pollinations_img_endpoint = "https://image.pollinations.ai/prompt";
const pollinations_text_endpoint = "https://text.pollinations.ai/openai";
const dummy_api_key = "kobo";
@@ -4169,7 +4168,6 @@ Current version indicated by LITEVER below.
const XTTS_ID = 3;
const ALLTALK_ID = 4;
const OAI_TTS_ID = 5;
const POLLINATIONS_TTS_ID = 6;
const HD_RES_PX = 768;
const VHD_RES_PX = 960;
@@ -4362,7 +4360,7 @@ Current version indicated by LITEVER below.
prev_custom_endpoint_type: 0, //show a reconnect box to custom endpoint if needed. 0 is horde, otherwise its dropdown value+1
prev_custom_endpoint_model: "", //we may not be able to match, but set it if we do
prev_custom_endpoint_ischatcmpl: true,
generate_images_mode: (localflag?0:1), //0=off, 1=horde, 2=a1111, 3=dalle, 4=comfy, 5=pollinations
generate_images_mode: (localflag?0:1), //0=off, 1=horde, 2=a1111, 3=dalle, 4=comfy
autoscroll: true, //automatically scroll to bottom on render
printer_view: false, //automatically scroll to bottom on render
@@ -9076,71 +9074,6 @@ Current version indicated by LITEVER below.
},true);
}
function generate_pollinations_image(req_payload, autoappend)
{
let splits = req_payload.prompt.split("###");
let prompt = splits[0].trim();
let negprompt = (splits.length > 1 ? splits[1] : "");
const pollinations_params = new URLSearchParams({
model:req_payload.models[0],
seed:Math.floor(Math.random() * 99999999),
width: req_payload.params.width,
height: req_payload.params.height,
nologo: true,
private: true,
referrer: "koboldai"
});
let gen_endpoint = `${pollinations_img_endpoint}/${encodeURIComponent(prompt)}?${pollinations_params.toString()}`;
console.log(gen_endpoint);
let imgid = "PollAIimg"+(Math.floor(10000 + Math.random() * 90000)).toString();
let nimgtag = "[<|p|" + imgid + "|p|>]";
if (localsettings.img_newturn) {
if(localsettings.opmode == 4)
{
nimgtag = wrap_newgen_instruct_format(nimgtag,false);
}
else if(localsettings.opmode == 3)
{
nimgtag = wrap_newgen_chat_format(nimgtag);
}
}
if(autoappend)
{
gametext_arr.push(nimgtag);
}
image_db[imgid] = { done: false, queue: "Generating", result: "", prompt:prompt, poll_category:0 };
image_db[imgid].aspect = (req_payload.params.width>=req_payload.params.height*2?5:(req_payload.params.height>=req_payload.params.width*2?4:(req_payload.params.width>req_payload.params.height?2:(req_payload.params.width<req_payload.params.height?1:0))));
image_db[imgid].imsource = 0; //0=generated,1=uploaded
image_db[imgid].imrefid = "";
image_db[imgid].type = 0; //0=image, 1=audio
fetch(gen_endpoint, {
method: 'GET',
})
.then((response) => {
return response.blob(); // Convert the response into a Blob
})
.then((finalimg) => {
const reader = new FileReader();
reader.onloadend = () => {
let origImg = reader.result;
let imgres = localsettings.img_allowhd?VHD_RES_PX:NO_HD_RES_PX;
compressImage(origImg, (newDataUri) => {
image_db[imgid].done = true;
image_db[imgid].result = newDataUri;
}, false, imgres);
};
reader.readAsDataURL(finalimg);
}).catch((error) => {
console.log("Generation Error: " + error);
msgbox("Image Generation Failed!\n\nCould not generate image with Pollinations.ai, maybe you are rate limited. Try again later.\n");
});
return imgid;
}
function set_horde_key()
{
inputBox("Enter AI Horde API Key.\n\nThe same key is used for image and text generation in AI Horde.","AI Horde API Key",localsettings.my_api_key,"Input AI Horde API Key", ()=>{
@@ -9154,7 +9087,7 @@ Current version indicated by LITEVER below.
function set_dalle_key()
{
inputBox("Enter DALL-E API Key.\n\nNote: DALL-E is known to rephrase and rewrite submitted image prompts before generating, for censorship purposes. There is nothing KoboldAI Lite can do about that. ","DALL-E API Key",localsettings.saved_dalle_key,"Input DALL-E API Key", ()=>{
inputBox("Enter OpenAI Compatible Image API Key.\n\nNote: DALL-E is known to rephrase and rewrite submitted image prompts before generating, for censorship purposes. There is nothing KoboldAI Lite can do about that. ","OpenAI Compatible Image API Key",localsettings.saved_dalle_key,"Input OpenAI Compatible Image API Key", ()=>{
let userinput = getInputBoxValue();
userinput = userinput.trim();
if (userinput != null && userinput!="") {
@@ -9164,7 +9097,7 @@ Current version indicated by LITEVER below.
}
function set_dalle_url()
{
inputBox("Enter DALL-E API URL.\n\nNote: DALL-E is known to rephrase and rewrite submitted image prompts before generating, for censorship purposes. There is nothing KoboldAI Lite can do about that. ","DALL-E API URL",localsettings.saved_dalle_url,"Input DALL-E API URL", ()=>{
inputBox("Enter OpenAI Compatible Image API URL.\n\nNote: DALL-E is known to rephrase and rewrite submitted image prompts before generating, for censorship purposes. There is nothing KoboldAI Lite can do about that. ","OpenAI Compatible Image API URL",localsettings.saved_dalle_url,"Input OpenAI Compatible Image API URL", ()=>{
let userinput = getInputBoxValue();
userinput = userinput.trim();
if (userinput != null && userinput!="") {
@@ -9176,7 +9109,7 @@ Current version indicated by LITEVER below.
}
function set_dalle_model()
{
inputBox("Enter DALL-E API Model Identifier.","DALL-E API Model Identifier",localsettings.saved_dalle_model,"Input DALL-E Model Identifier", ()=>{
inputBox("Enter OpenAI Compatible Image API Model Identifier.","OpenAI Compatible Image API Model Identifier",localsettings.saved_dalle_model,"Input OpenAI Compatible Image Model Identifier", ()=>{
let userinput = getInputBoxValue();
userinput = userinput.trim();
if (userinput != null && userinput!="") {
@@ -15518,6 +15451,10 @@ Current version indicated by LITEVER below.
mainmenu_untab(true);
document.getElementById("settingscontainer").classList.remove("hidden");
display_settings_tab(current_settings_tab_idx);
if(localsettings.generate_images_mode>=5){localsettings.generate_images_mode=0;} //todo: temp hack to remove pollinations images
if(localsettings.tts_mode>=6){localsettings.tts_mode=0;} //todo: temp hack to remove pollinations tts
document.getElementById("max_context_length").value = document.getElementById("max_context_length_slide").value = localsettings.max_context_length;
document.getElementById("max_length").value = document.getElementById("max_length_slide").value = localsettings.max_length;
document.getElementById("temperature").value = document.getElementById("temperature_slide").value = localsettings.temperature;
@@ -16864,7 +16801,6 @@ Current version indicated by LITEVER below.
document.getElementById("generate_images_dalle_container").classList.add("hidden");
document.getElementById("generate_images_local_model_container").classList.add("hidden");
document.getElementById("generate_images_comfy_container").classList.add("hidden");
document.getElementById("generate_images_pollinations_container").classList.add("hidden");
if(document.getElementById("generate_images_mode").value==1){
document.getElementById("generate_images_model_container").classList.remove("hidden");
if(!image_models_fetched)
@@ -16883,9 +16819,6 @@ Current version indicated by LITEVER below.
}else if(document.getElementById("generate_images_mode").value==4){
document.getElementById("generate_images_comfy_container").classList.remove("hidden");
connect_to_comfyui(silent);
}else if(document.getElementById("generate_images_mode").value==5)
{
document.getElementById("generate_images_pollinations_container").classList.remove("hidden");
}
}
@@ -18633,7 +18566,7 @@ Current version indicated by LITEVER below.
}
let speakprompt = "Enter phrase to speak. Currently selected TTS will be used.";
if(ssval==XTTS_ID || ssval==ALLTALK_ID || ssval==OAI_TTS_ID || ssval==KCPP_TTS_ID || ssval==POLLINATIONS_TTS_ID)
if(ssval==XTTS_ID || ssval==ALLTALK_ID || ssval==OAI_TTS_ID || ssval==KCPP_TTS_ID)
{
speakprompt = `Enter phrase to speak. Currently selected TTS will be used.<br>
<div style="display: inline-flex;">
@@ -18669,7 +18602,6 @@ Current version indicated by LITEVER below.
{
document.getElementById("xtts_container").classList.add("hidden");
document.getElementById("oai_tts_container").classList.add("hidden");
document.getElementById("pollinations_tts_container").classList.add("hidden");
document.getElementById("alltalk_specific_controls").classList.add("hidden");
document.getElementById("kcpp_tts_container").classList.add("hidden");
document.getElementById("webbrowser_tts_container").classList.add("hidden");
@@ -18702,11 +18634,6 @@ Current version indicated by LITEVER below.
document.getElementById("oai_tts_container").classList.remove("hidden");
document.getElementById("savenarrationsdiv").classList.remove("hidden");
document.getElementById("narratestreamingdiv").classList.remove("hidden");
}else if(selectedTTS == POLLINATIONS_TTS_ID)
{
document.getElementById("pollinations_tts_container").classList.remove("hidden");
document.getElementById("savenarrationsdiv").classList.remove("hidden");
document.getElementById("narratestreamingdiv").classList.remove("hidden");
}
else if(selectedTTS == KCPP_TTS_ID) {
document.getElementById("kcpp_tts_container").classList.remove("hidden");
@@ -18839,12 +18766,13 @@ Current version indicated by LITEVER below.
function tts_speak(text, do_download=false, do_embed_tts=false, is_test=false)
{
let ssval = localsettings.tts_mode;
let streamallowed = (ssval==XTTS_ID || ssval==ALLTALK_ID || ssval==OAI_TTS_ID || ssval==KCPP_TTS_ID || ssval==POLLINATIONS_TTS_ID);
let streamallowed = (ssval==XTTS_ID || ssval==ALLTALK_ID || ssval==OAI_TTS_ID || ssval==KCPP_TTS_ID);
if(streamallowed && localsettings.tts_stream && !is_test)
{
//split into chunks by newline
tts_speak_output_queue = [];
tts_speak_input_queue = text.split("\n").map(chunk => chunk.trim()).filter(chunk => chunk.length > 0);
let text2 = text.replaceAll(". ", ".\n").replaceAll("! ", "!\n").replaceAll("? ", "?\n");
tts_speak_input_queue = text2.split("\n").map(chunk => chunk.trim()).filter(chunk => chunk.length > 0);
tts_queue_embed = do_embed_tts;
tts_queue_download = do_download;
}
@@ -18928,12 +18856,11 @@ Current version indicated by LITEVER below.
}
if(ssval==XTTS_ID || ssval==ALLTALK_ID || ssval==OAI_TTS_ID || ssval==KCPP_TTS_ID || ssval==POLLINATIONS_TTS_ID) //xtts api server
if(ssval==XTTS_ID || ssval==ALLTALK_ID || ssval==OAI_TTS_ID || ssval==KCPP_TTS_ID) //xtts api server
{
let is_xtts = (ssval==XTTS_ID);
let is_oai_tts = (ssval==OAI_TTS_ID);
let is_kcpp_tts = (ssval==KCPP_TTS_ID);
let is_pollinations_tts = (ssval==POLLINATIONS_TTS_ID);
const audioContext = new (window.AudioContext || window.webkitAudioContext)();
let audiofile_ref = null;
@@ -19019,62 +18946,6 @@ Current version indicated by LITEVER below.
console.log("XTTS Speak Error: " + error);
});
}
else if(is_pollinations_tts)
{
const pollinations_params = new URLSearchParams({
model:"openai-audio",
voice:document.getElementById("pollinations_voices").value,
private: true,
referrer: "koboldai"
});
const speechprompt = `Please narrate the following text:\n\n${text}`;
let gen_endpoint = `${pollinations_text_endpoint}/${encodeURIComponent(speechprompt)}?${pollinations_params.toString()}`;
fetch(gen_endpoint, {
method: 'GET',
})
.then(response => response.arrayBuffer())
.then(data => {
audiofile_ref = data.slice(0);
return audioContext.decodeAudioData(data);
})
.then(decodedData => {
if(do_download)
{
tts_download(audiofile_ref);
}
if(do_embed_tts)
{
tts_embed_audio(audiofile_ref);
}
tts_is_processing = false;
if(append_to_outqueue)
{
tts_speak_output_queue.push(decodedData);
}
else
{
const playSound = audioContext.createBufferSource();
playSound.buffer = decodedData;
playSound.connect(audioContext.destination);
tts_is_playing = true;
update_submit_button(false);
playSound.start(audioContext.currentTime);
playSound.onended = function() {
setTimeout(() => {
tts_is_playing = false;
update_submit_button(false);
console.log("Audio finished playing");
},300);
};
}
}).catch((error) => {
tts_is_playing = false;
tts_is_processing = false;
console.log("Pollinations Speak Error: " + error);
});
}
else if(xtts_is_connected)
{
if(is_xtts)
@@ -21668,7 +21539,7 @@ Current version indicated by LITEVER below.
{
if(localsettings.saved_dalle_key=="" || localsettings.saved_dalle_url=="")
{
msgbox("Error: A valid DALL-E URL and Key is required to generate images with DALL-E.\nThis is usually the same as your OpenAI API key, but can be customized in settings.","Invalid DALL-E Key");
msgbox("Error: A valid OpenAI Compatible Image URL and Key is required to generate images with DALL-E.\nThis is usually the same as your OpenAI API key, but can be customized in settings.","Invalid OpenAI Compatible Image Key");
}
else
{
@@ -21716,12 +21587,6 @@ Current version indicated by LITEVER below.
genimg_payload.models = [desired_model];
imgid = generate_comfy_image(genimg_payload, autoappend);
}
else if(localsettings.generate_images_mode==5) //pollinations
{
let desired_model = document.getElementById("generate_images_pollinations_model").value;
genimg_payload.models = [desired_model];
imgid = generate_pollinations_image(genimg_payload, autoappend);
}
return imgid;
}
@@ -22848,7 +22713,8 @@ Current version indicated by LITEVER below.
console.log("Replacing with Image: " + matchstr);
gametext_arr[i] = gametext_arr[i].replace(matchstr, newstr);
//default to llava if supported, and image is self uploaded
let desiredvismode = ((image_db[key].imsource==1 && ((is_using_kcpp_with_vision() && image_db[key].type==0) || (is_using_kcpp_with_audio() && image_db[key].type==1)))?3:0);
let isoaichat = (document.getElementById("useoaichatcompl").checked && custom_oai_key!="");
let desiredvismode = ((image_db[key].imsource==1 && (((is_using_kcpp_with_vision() || isoaichat) && image_db[key].type==0) || ((is_using_kcpp_with_audio() || isoaichat) && image_db[key].type==1)))?3:0);
completed_imgs_meta[metaid] = JSON.parse(JSON.stringify(default_imgs_meta));
completed_imgs_meta[metaid].prompt = image_db[key].prompt;
completed_imgs_meta[metaid].visionmode = desiredvismode;
@@ -29711,9 +29577,8 @@ Current version indicated by LITEVER below.
<option value="0">[Disabled]</option>
<option value="1">AI Horde</option>
<option value="2">KCPP / Forge / A1111</option>
<option value="3">OpenAI DALL-E</option>
<option value="3">OpenAI Compatible Img</option>
<option value="4">ComfyUI</option>
<option value="5">Pollinations.ai</option>
</select>
</div>
</div>
@@ -29772,16 +29637,6 @@ Current version indicated by LITEVER below.
</tr></table>
</div>
<div id="generate_images_pollinations_container" class="settinglabel hidden settingsbox">
<div class="settinglabel" style="display: flex; width: 100%">
<div class="justifyleft">Model</div>
<select title="Select Image Model" class="form-control push-right" id="generate_images_pollinations_model">
<option value="flux" selected>flux</option>
<option value="turbo">turbo</option>
</select>
</div>
</div>
<div>
<div class="settinglabel">
<div class="justifyleft">Add positive prompt</div>
@@ -29901,7 +29756,6 @@ Current version indicated by LITEVER below.
<option value="3">XTTS API Server</option>
<option value="4">AllTalk API Server</option>
<option value="5">OpenAI-Compat. API Server</option>
<option value="6">PollinationsAI TTS API</option>
</select>
</div>
<button id="manual_tts" type="button" class="bg_green btn btn-primary" style="margin-left: 2px;" onclick="test_tts()">Test</button>
@@ -29972,25 +29826,6 @@ Current version indicated by LITEVER below.
</div>
</div>
</div>
<div id="pollinations_tts_container" class="hidden">
<div class="settinglabel">
<div class="justifyleft">Voice </div>
<div class="push-right">
<select class="form-control" id="pollinations_voices" style="margin:0px 0 0;">
<option value="alloy">alloy</option>
<option value="ash">ash</option>
<option value="ballad">ballad</option>
<option value="coral">coral</option>
<option value="echo">echo</option>
<option value="fable">fable</option>
<option value="nova" selected>nova</option>
<option value="onyx">onyx</option>
<option value="sage">sage</option>
<option value="shimmer">shimmer</option>
</select>
</div>
</div>
</div>
<div id="kcpp_tts_container" class="hidden">
<div class="color_red hidden justifyright" id="nokcpptts">KoboldCpp TTS Unavailable</div>
<div class="settinglabel">
@@ -30988,13 +30823,7 @@ Current version indicated by LITEVER below.
<option style="display:none;" class="custom_model_option" value="custom">[Custom]</option>
</select>
<select title="Pollinations AI Model Selection" style="padding:4px;display:inline;width:calc(100% - 200px)" class="form-control hidden" id="custom_pollinations_model" onchange="oai_model_change(true)">
<option value="bidara">bidara</option>
<option value="chickytutor">chickytutor</option>
<option value="gemini">gemini</option>
<option value="midijourney">midijourney</option>
<option value="mistral">mistral</option>
<option value="openai">openai</option>
<option value="openai-fast">openai-fast</option>
<option value="openai-fast" selected>openai-fast</option>
<option style="display:none;" class="custom_model_option" value="custom">[Custom]</option>
</select>
<select title="Nvidia NIM AI Model Selection" style="padding:4px;display:inline;width:calc(100% - 200px)" class="form-control hidden" id="custom_nvidianim_model" onchange="oai_model_change(true)">
expose.h
+1
View File
@@ -348,6 +348,7 @@ struct music_generation_inputs
const bool is_planner_mode = false; //if true, generate codes, else, generate diffusion music
const bool stereo = false;
const bool gen_codes = false;
const bool rewrite_caption = true;
const char * input_json = nullptr;
};
struct music_generation_outputs
koboldcpp.py
+3
View File
@@ -462,6 +462,7 @@ class music_generation_inputs(ctypes.Structure):
_fields_ = [("is_planner_mode", ctypes.c_bool),
("stereo", ctypes.c_bool),
("gen_codes", ctypes.c_bool),
("rewrite_caption", ctypes.c_bool),
("input_json", ctypes.c_char_p)]
class music_generation_outputs(ctypes.Structure):
@@ -2492,6 +2493,7 @@ def music_generate_codes(genparams):
inputs.is_planner_mode = True
inputs.stereo = genparams.get('stereo', False)
inputs.gen_codes = genparams.get('gen_codes', False)
inputs.rewrite_caption = genparams.get('rewrite_caption', True)
inputs.input_json = input_json.encode("UTF-8")
ret = handle.music_generate(inputs)
outstr = ""
@@ -2507,6 +2509,7 @@ def music_generate_audio(genparams):
inputs.is_planner_mode = False
inputs.stereo = genparams.get('stereo', False)
inputs.gen_codes = genparams.get('gen_codes', False)
inputs.rewrite_caption = genparams.get('rewrite_caption', True)
inputs.input_json = input_json.encode("UTF-8")
ret = handle.music_generate(inputs)
outstr = ""
otherarch/acestep/ace-qwen3.cpp
+78 -273
View File
@@ -173,6 +173,7 @@ static std::mt19937 acestep_lm_rng;
static bool acestep_lm_dbg = false;
static std::vector<int32_t> forced_tokens;
static std::vector<int> caption_tokens = std::vector<int>(); //will be filled with caption tokens
const int rep_pen_range = 48;
//
// CoT parsing (extract metadata + lyrics from LLM Phase1 output)
@@ -481,9 +482,11 @@ struct MetadataFSM {
};
State state = DISABLED;
int caption_newline_counter = 0;
int name_pos = 0;
std::vector<int> value_acc;
bool enabled = false;
bool rewrite_caption = false;
std::vector<int> bpm_name, caption_name, duration_name;
std::vector<int> keyscale_name, language_name, timesig_name;
@@ -513,7 +516,7 @@ struct MetadataFSM {
}
}
void init(BPETokenizer & bpe, int vsize) {
void init(BPETokenizer & bpe, int vsize, bool rew_caption) {
vocab_size = vsize;
auto nl = bpe_encode(&bpe, "\n", false);
newline_tok = nl.empty() ? -1 : nl[0];
@@ -531,10 +534,10 @@ struct MetadataFSM {
for (int v = 30; v <= 300; v++) vals.push_back(std::to_string(v));
build_value_tree(bpe, bpm_tree, "bpm:", vals);
}
// Duration 10-300
// Duration 50-450
{
std::vector<std::string> vals;
for (int v = 10; v <= 300; v++) vals.push_back(std::to_string(v));
for (int v = 50; v <= 450; v++) vals.push_back(std::to_string(v));
build_value_tree(bpe, duration_tree, "duration:", vals);
}
// Keyscale
@@ -574,12 +577,15 @@ struct MetadataFSM {
enabled = true;
state = BPM_NAME;
name_pos = 0;
caption_newline_counter = 0;
rewrite_caption = rew_caption;
value_acc.clear();
}
void reset() {
state = BPM_NAME;
name_pos = 0;
caption_newline_counter = 0;
value_acc.clear();
}
@@ -707,7 +713,27 @@ struct MetadataFSM {
}
if (state == CAPTION_VALUE) {
if (token == newline_tok) {
if(token==newline_tok && !rewrite_caption)
{
//not changing caption, so end it now
caption_newline_counter = 0;
state = DURATION_NAME;
name_pos = 0;
value_acc.clear();
}
else if (token == newline_tok) {
// caption_newline_counter += 1;
// if(caption_newline_counter>=3) //ingest 3 newlines for the caption, making it longer
// {
caption_newline_counter = 0;
state = DURATION_NAME;
name_pos = 0;
value_acc.clear();
// }
}
else if(token == 17021) // fallback for duration token, immediately go next
{
caption_newline_counter = 0;
state = DURATION_NAME;
name_pos = 0;
value_acc.clear();
@@ -779,7 +805,7 @@ const std::vector<int> think_chain = {271,2,15953,2216,198}; // "\n# Lyric\n"
static std::vector<std::string> generate_phase1_batch(
Qwen3LM * m, BPETokenizer * bpe,
const std::vector<int> & prompt_tokens,
int max_new_tokens, float temperature, float top_p,
int max_new_tokens, float temperature, float top_p, int top_k, float rep_pen,
long long base_seed, int N,
MetadataFSM * fsm_template,
bool lyrics_mode,
@@ -791,6 +817,8 @@ static std::vector<std::string> generate_phase1_batch(
bool use_cfg = cfg_scale > 1.0f && uncond_tokens && !uncond_tokens->empty();
forced_tokens.clear();
std::vector<int32_t> quicklastntoks;
// KV sets: cond [0..N-1], uncond [N..2N-1] if CFG
for (int i = 0; i < N; i++) qw3lm_reset_kv(m, i);
if (use_cfg)
@@ -846,7 +874,11 @@ static std::vector<std::string> generate_phase1_batch(
if (fsm_template && fsm_template->enabled)
seqs[i].fsm.apply_mask(lg.data());
int tok = kcpp_quick_sample(lg.data(),V,std::vector<int32_t>(),1.03f,top_p,40,temperature,acestep_lm_rng);
int tok = kcpp_quick_sample(lg.data(),V,quicklastntoks,rep_pen,top_p,top_k,temperature,acestep_lm_rng);
quicklastntoks.push_back(tok);
if (quicklastntoks.size()>rep_pen_range) {
quicklastntoks.erase(quicklastntoks.begin());
}
if (tok == TOKEN_IM_END) {
seqs[i].done = true;
@@ -890,8 +922,6 @@ static std::vector<std::string> generate_phase1_batch(
for (int i = 0; i < N; i++)
if (seqs[i].done) n_active--;
std::vector<int32_t> quicklastntoks;
for (int step = 0; step < max_new_tokens && n_active > 0; step++) {
for (int i = 0; i < N; i++)
tokens[i] = seqs[i].last_token;
@@ -938,9 +968,9 @@ static std::vector<std::string> generate_phase1_batch(
if (v != TOKEN_IM_END) lc[v] = -1e9f;
}
int tok = kcpp_quick_sample(lc,V,quicklastntoks,1.03f,top_p,40,temperature,acestep_lm_rng);
int tok = kcpp_quick_sample(lc,V,quicklastntoks,rep_pen,top_p,top_k,temperature,acestep_lm_rng);
quicklastntoks.push_back(tok);
if (quicklastntoks.size()>32) {
if (quicklastntoks.size()>rep_pen_range) {
quicklastntoks.erase(quicklastntoks.begin());
}
@@ -1008,13 +1038,15 @@ static std::vector<std::string> generate_phase1_batch(
// Returns N code strings. Seeds = base_seed + 0, 1, ..., N-1.
static std::vector<std::string> run_phase2_batch(
Qwen3LM * m, BPETokenizer & bpe, const std::vector<AcePrompt> & aces,
float temperature, float top_p, long long base_seed, int N,
float temperature, float top_p, int top_k, float rep_pen, long long base_seed, int N,
float cfg_scale, const char * negative_prompt) {
int V = m->cfg.vocab_size;
bool use_cfg = cfg_scale > 1.0f;
bool shared_prompt = ((int)aces.size() == 1);
std::vector<int32_t> quicklastntoks;
// Build per-element prompts
std::vector<std::vector<int>> prompts(N), unconds(N);
int max_tokens = 0;
@@ -1101,7 +1133,12 @@ static std::vector<std::string> 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<int32_t>(),1.03f,top_p,40,temperature,acestep_lm_rng);
int tok = kcpp_quick_sample(lg.data(),V,quicklastntoks,rep_pen,top_p,top_k,temperature,acestep_lm_rng);
quicklastntoks.push_back(tok);
if (quicklastntoks.size()>rep_pen_range) {
quicklastntoks.erase(quicklastntoks.begin());
}
seqs[i].last_token = tok;
if (tok == TOKEN_IM_END) {
@@ -1139,7 +1176,6 @@ static std::vector<std::string> run_phase2_batch(
for (int i = 0; i < N; i++)
if (seqs[i].done) n_active--;
std::vector<int32_t> 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++)
@@ -1173,9 +1209,9 @@ static std::vector<std::string> 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,quicklastntoks,1.03f,top_p,40,temperature,acestep_lm_rng);
int tok = kcpp_quick_sample(lc,V,quicklastntoks,rep_pen,top_p,top_k,temperature,acestep_lm_rng);
quicklastntoks.push_back(tok);
if (quicklastntoks.size()>32) {
if (quicklastntoks.size()>rep_pen_range) {
quicklastntoks.erase(quicklastntoks.begin());
}
seqs[i].last_token = tok;
@@ -1236,253 +1272,6 @@ static void usage(const char * prog) {
, prog);
}
/*
int main(int argc, char ** argv) {
const char * model_path = nullptr;
const char * request_path = nullptr;
int max_seq = 8192;
int batch_size = 1;
bool use_fsm = true;
const char * dump_logits = nullptr;
const char * dump_tokens = nullptr;
if (argc < 2) {
usage(argv[0]);
return 1;
}
for (int i = 1; i < argc; i++) {
if (!strcmp(argv[i], "--model") && i + 1 < argc)
model_path = argv[++i];
else if (!strcmp(argv[i], "--request") && i + 1 < argc)
request_path = argv[++i];
else if (!strcmp(argv[i], "--max-seq") && i + 1 < argc)
max_seq = atoi(argv[++i]);
else if (!strcmp(argv[i], "--batch") && i + 1 < argc)
batch_size = atoi(argv[++i]);
else if (!strcmp(argv[i], "--no-fsm"))
use_fsm = false;
else if (!strcmp(argv[i], "--dump-logits") && i + 1 < argc)
dump_logits = argv[++i];
else if (!strcmp(argv[i], "--dump-tokens") && i + 1 < argc)
dump_tokens = argv[++i];
else if (!strcmp(argv[i], "--help") || !strcmp(argv[i], "-h")) {
usage(argv[0]);
return 0;
}
else {
fprintf(stderr, "Unknown option: %s\n", argv[i]);
usage(argv[0]);
return 1;
}
}
if (!model_path) {
fprintf(stderr, "ERROR: --model required\n");
usage(argv[0]); return 1;
}
if (!request_path) {
fprintf(stderr, "ERROR: --request required\n");
usage(argv[0]); return 1;
}
// Read request JSON
AceRequest req;
if (!request_parse(&req, request_path)) return 1;
request_dump(&req, stderr);
if (req.caption.empty()) {
fprintf(stderr, "ERROR: caption is empty in %s\n", request_path);
return 1;
}
// Resolve seed
long long seed = req.seed;
if (seed < 0) {
std::random_device rd;
seed = (int64_t)rd() << 32 | rd();
if (seed < 0) seed = -seed; // keep positive
}
req.seed = seed;
// Generation params from request
float temperature = req.lm_temperature;
float top_p = req.lm_top_p;
int top_k = req.lm_top_k;
float cfg_scale = req.lm_cfg_scale;
const char * neg_prompt = req.lm_negative_prompt.c_str();
Timer t_total;
// Load BPE tokenizer from model GGUF
BPETokenizer bpe;
if (!load_bpe_from_gguf(&bpe, model_path)) return 1;
// Load model
int n_kv_sets = (cfg_scale > 1.0f) ? 2 * batch_size : batch_size;
Timer t_load;
Qwen3LM model;
if (!qw3lm_load(&model, model_path, max_seq, n_kv_sets)) return 1;
double load_ms = t_load.ms();
// FSM
MetadataFSM fsm;
if (use_fsm) fsm.init(bpe, model.cfg.vocab_size);
// Copy request -> AcePrompt (internal LLM struct)
AcePrompt ace = {};
ace.caption = req.caption;
ace.lyrics = req.lyrics;
ace.duration = req.duration;
ace.bpm = req.bpm;
ace.keyscale = req.keyscale;
ace.timesignature = req.timesignature;
ace.vocal_language = req.vocal_language;
bool user_has_codes = !req.audio_codes.empty();
bool need_lm_codes = req.thinking && !user_has_codes;
bool is_simple = ace.lyrics.empty() &&
ace.bpm <= 0 && ace.duration <= 0 &&
ace.keyscale.empty() && ace.timesignature.empty();
std::vector<int> prompt;
std::vector<AcePrompt> aces; // populated by Phase 1 (simple or partial)
// Preprocessor: simple mode generates lyrics + metas from caption
if (is_simple) {
fprintf(stderr, "[Simple] Inspiration\n");
const char * sys =
"# Instruction\n"
"Expand the user's input into a more detailed"
" and specific musical description:\n";
std::string user_msg = ace.caption + "\n\ninstrumental: "
+ std::string(req.instrumental ? "true" : "false");
prompt = build_custom_prompt(bpe, sys, user_msg.c_str());
// FSM: reset then optionally force language (shared for both paths)
fsm.reset();
if (use_fsm && ace.vocal_language != "unknown" && !ace.vocal_language.empty())
fsm.force_language(bpe, ace.vocal_language);
// Phase 1: N lyrics + metadata generations (always batched, N=batch_size)
fprintf(stderr, "[Simple] %zu tokens, N=%d, seeds: %lld..%lld\n",
prompt.size(), batch_size, seed, seed + batch_size - 1);
auto phase1_texts = generate_phase1_batch(
&model, &bpe, prompt, 2048, temperature, 1.0f, 0,
seed, batch_size, use_fsm ? &fsm : nullptr, true);
parse_phase1_into_aces(phase1_texts, ace, aces, seed, "Simple", true);
for (int i = 0; i < batch_size; i++) qw3lm_reset_kv(&model, i);
}
// Re-evaluate after possible simple enrichment
const AcePrompt & ace_ref = aces.empty() ? ace : aces[0];
bool has_all_metas = (ace_ref.bpm > 0 && ace_ref.duration > 0 &&
!ace_ref.keyscale.empty() && !ace_ref.timesignature.empty());
if (!has_all_metas) {
// Partial-metas: Phase 1 with CFG to fill missing fields
prompt = build_lm_prompt(bpe, ace);
std::vector<int> uncond;
if (cfg_scale > 1.0f)
uncond = build_lm_prompt_uncond(bpe, ace, neg_prompt);
fprintf(stderr, "[Partial] %zu tokens, CFG: %.2f, N=%d, seeds: %lld..%lld\n",
prompt.size(), cfg_scale, batch_size, seed, seed + batch_size - 1);
fsm.reset();
auto phase1_texts = generate_phase1_batch(
&model, &bpe, prompt, 2048, temperature, top_p, top_k,
seed, batch_size, use_fsm ? &fsm : nullptr, false,
cfg_scale, uncond.empty() ? nullptr : &uncond, true);
parse_phase1_into_aces(phase1_texts, ace, aces, seed, "Partial", false);
for (int i = 0; i < 2 * batch_size; i++) qw3lm_reset_kv(&model, i);
}
// Guarantee aces is populated (all-metas: single shared ace for prefill optimization)
if (aces.empty()) aces = {ace};
// Debug: dump tokens/logits
if (need_lm_codes && (dump_logits || dump_tokens)) {
std::string cot = build_cot_yaml(aces[0]);
auto dbg_prompt = build_lm_prompt_with_cot(bpe, aces[0], cot);
if (dump_tokens) {
FILE * f = fopen(dump_tokens, "w");
if (f) {
for (size_t j = 0; j < dbg_prompt.size(); j++)
fprintf(f, "%s%d", j ? "," : "", dbg_prompt[j]);
fprintf(f, "\n");
fclose(f);
fprintf(stderr, "[Debug] Tokens -> %s (%zu)\n",
dump_tokens, dbg_prompt.size());
}
}
if (dump_logits) {
std::vector<float> dbg_logits(model.cfg.vocab_size);
qw3lm_forward(&model, dbg_prompt.data(), (int)dbg_prompt.size(), 0, dbg_logits.data());
FILE * f = fopen(dump_logits, "wb");
if (f) {
fwrite(dbg_logits.data(), sizeof(float), model.cfg.vocab_size, f);
fclose(f);
fprintf(stderr, "[Debug] Logits -> %s (%d floats, argmax=%d)\n",
dump_logits, model.cfg.vocab_size,
(int)(std::max_element(dbg_logits.begin(), dbg_logits.end()) - dbg_logits.begin()));
}
qw3lm_reset_kv(&model, 0);
}
}
// Phase 2: generate audio codes (always batched, N=batch_size)
std::vector<std::string> batch_codes(batch_size);
if (need_lm_codes) {
batch_codes = run_phase2_batch(&model, bpe, aces,
temperature, top_p, top_k, seed, batch_size, cfg_scale, neg_prompt);
} else {
fprintf(stderr, "[Skip] %s, no code generation\n",
user_has_codes ? "user codes present" : "thinking=false");
}
// Write N output files: request0.json, request1.json, ...
{
std::string base(request_path);
std::string ext = ".json";
size_t dot = base.rfind('.');
if (dot != std::string::npos) { ext = base.substr(dot); base = base.substr(0, dot); }
for (int b = 0; b < batch_size; b++) {
AceRequest rr = req;
const AcePrompt & a = aces[b < (int)aces.size() ? b : 0];
rr.caption = a.caption;
rr.lyrics = a.lyrics;
rr.bpm = a.bpm;
rr.duration = a.duration;
rr.keyscale = a.keyscale;
rr.timesignature = a.timesignature;
rr.vocal_language = a.vocal_language;
if (!batch_codes[b].empty()) rr.audio_codes = batch_codes[b];
rr.seed = seed + b;
char path[512];
snprintf(path, sizeof(path), "%s%d%s", base.c_str(), b, ext.c_str());
request_write(&rr, path);
fprintf(stderr, "[Output] Wrote %s\n", path);
}
}
fprintf(stderr, "[Ace-Qwen3] Load %.0f | Total %.0fms | seed=%lld\n",
load_ms, t_total.ms(), seed);
qw3lm_free(&model);
return 0;
}
*/
//kcpp stuff
static Qwen3LM acestep_llm;
@@ -1538,11 +1327,13 @@ std::string acestep_prepare_request(const music_generation_inputs inputs)
}
}
bool rewrite_caption = inputs.rewrite_caption;
const int batch_size = 1;
bool use_fsm = true;
MetadataFSM fsm;
if (use_fsm) {
fsm.init(acestep_bpe, acestep_llm.cfg.vocab_size);
fsm.init(acestep_bpe, acestep_llm.cfg.vocab_size,rewrite_caption);
}
// Read request and set essentials
@@ -1570,8 +1361,13 @@ std::string acestep_prepare_request(const music_generation_inputs inputs)
float temperature = req.lm_temperature;
float top_p = req.lm_top_p;
int top_k = req.lm_top_k;
float rep_pen = req.lm_rep_pen;
float cfg_scale = req.lm_cfg_scale;
float phase1cfg = 1.0f;
const char * neg_prompt = req.lm_negative_prompt.c_str();
float codes_temperature = req.codes_temperature;
float codes_top_p = req.codes_top_p;
int codes_top_k = req.codes_top_k;
// Copy request -> AcePrompt (internal LLM struct)
AcePrompt ace = {};
@@ -1592,10 +1388,10 @@ std::string acestep_prepare_request(const music_generation_inputs inputs)
std::vector<AcePrompt> aces; // populated by Phase 1 (simple or partial)
caption_tokens.clear();
// if(ace.caption!="")
// {
// caption_tokens = bpe_encode(&acestep_bpe, ace.caption+"\n", false);
// }
if(ace.caption!="" && !rewrite_caption)
{
caption_tokens = bpe_encode(&acestep_bpe, ace.caption+"\n", false);
}
// Preprocessor: simple mode generates lyrics + metas from caption
if (is_simple) {
@@ -1619,7 +1415,7 @@ std::string acestep_prepare_request(const music_generation_inputs inputs)
prompt.size(), batch_size, seed, seed + batch_size - 1);
auto phase1_texts = generate_phase1_batch(
&acestep_llm, &acestep_bpe, prompt, 2048, temperature, top_p,
&acestep_llm, &acestep_bpe, prompt, 2048, temperature, top_p, top_k, rep_pen,
seed, batch_size, use_fsm ? &fsm : nullptr, true);
parse_phase1_into_aces(phase1_texts, ace, aces, seed, "Simple", true);
@@ -1636,17 +1432,17 @@ std::string acestep_prepare_request(const music_generation_inputs inputs)
// Partial-metas: Phase 1 with CFG to fill missing fields
prompt = build_lm_prompt(acestep_bpe, ace);
std::vector<int> uncond;
if (cfg_scale > 1.0f)
if (phase1cfg > 1.0f)
uncond = build_lm_prompt_uncond(acestep_bpe, ace, neg_prompt);
fprintf(stderr, "[Partial] %zu tokens, CFG: %.2f, N=%d, seeds: %lld..%lld\n",
prompt.size(), cfg_scale, batch_size, seed, seed + batch_size - 1);
prompt.size(), phase1cfg, batch_size, seed, seed + batch_size - 1);
fsm.reset();
auto phase1_texts = generate_phase1_batch(
&acestep_llm, &acestep_bpe, prompt, 2048, temperature, top_p,
&acestep_llm, &acestep_bpe, prompt, 2048, temperature, top_p, top_k, rep_pen,
seed, batch_size, use_fsm ? &fsm : nullptr, false,
cfg_scale, uncond.empty() ? nullptr : &uncond, true);
phase1cfg, uncond.empty() ? nullptr : &uncond, true);
parse_phase1_into_aces(phase1_texts, ace, aces, seed, "Partial", false);
@@ -1663,7 +1459,7 @@ std::string acestep_prepare_request(const music_generation_inputs inputs)
std::vector<std::string> batch_codes(batch_size);
if (need_lm_codes) {
batch_codes = run_phase2_batch(&acestep_llm, acestep_bpe, aces,
temperature, top_p, seed, batch_size, cfg_scale, neg_prompt);
temperature, top_p, top_k, rep_pen, seed, batch_size, cfg_scale, neg_prompt);
} else {
fprintf(stderr, "[Skip] %s, no code generation\n",
user_has_codes ? "user codes present" : "thinking=false");
@@ -1693,6 +1489,10 @@ std::string acestep_prepare_request(const music_generation_inputs inputs)
rr.keyscale = rr.keyscale.substr(prefix_erase.size()); // Returns a new string starting after the prefix
}
std::string delimiter = "durationduration";
size_t ddpos = rr.caption.find(delimiter);
rr.caption = ((ddpos != std::string::npos) ? rr.caption.substr(0, ddpos) : rr.caption);
//now convert to string
std::ostringstream oss;
oss << "{\n";
@@ -1712,6 +1512,11 @@ std::string acestep_prepare_request(const music_generation_inputs inputs)
oss << " \"lm_temperature\": " << std::fixed << std::setprecision(2) << rr.lm_temperature << ",\n";
oss << " \"lm_cfg_scale\": " << std::fixed << std::setprecision(1) << rr.lm_cfg_scale << ",\n";
oss << " \"lm_top_p\": " << std::fixed << std::setprecision(2) << rr.lm_top_p << ",\n";
oss << " \"lm_top_k\": " << std::fixed << std::setprecision(1) << rr.lm_top_k << ",\n";
oss << " \"codes_temperature\": " << std::fixed << std::setprecision(2) << rr.codes_temperature << ",\n";
oss << " \"codes_top_p\": " << std::fixed << std::setprecision(2) << rr.codes_top_p << ",\n";
oss << " \"codes_top_k\": " << std::fixed << std::setprecision(1) << rr.codes_top_k << ",\n";
oss << " \"lm_rep_pen\": " << std::fixed << std::setprecision(2) << rr.lm_rep_pen << ",\n";
oss << " \"lm_negative_prompt\": \"" << json_escape(rr.lm_negative_prompt) << "\",\n";
oss << " \"inference_steps\": " << rr.inference_steps << ",\n";
oss << " \"guidance_scale\": " << std::fixed << std::setprecision(1) << rr.guidance_scale << ",\n";
otherarch/acestep/request.cpp
+20 -5
View File
@@ -23,10 +23,14 @@ void request_init(AceRequest * r) {
r->task_type = "text2music";
r->seed = -1;
r->thinking = false;
r->lm_temperature = 1.0f;
r->lm_temperature = 0.85f;
r->lm_cfg_scale = 3.0f;
r->lm_top_p = 0.95f;
r->lm_top_k = 0;
r->lm_top_p = 0.9f;
r->lm_top_k = 50;
r->codes_temperature = 1.0f;
r->codes_top_p = 0.99f;
r->codes_top_k = 1000;
r->lm_rep_pen = 1.03f;
r->lm_negative_prompt = "";
r->audio_codes = "";
r->inference_steps = 8;
@@ -241,11 +245,16 @@ bool request_parse_from_str(AceRequest * r, std::string json) {
else if (k == "lm_temperature") r->lm_temperature = (float)atof(v.c_str());
else if (k == "lm_cfg_scale") r->lm_cfg_scale = (float)atof(v.c_str());
else if (k == "lm_top_p") r->lm_top_p = (float)atof(v.c_str());
else if (k == "lm_rep_pen") r->lm_rep_pen = (float)atof(v.c_str());
else if (k == "lm_top_k") r->lm_top_k = atoi(v.c_str());
else if (k == "inference_steps") r->inference_steps = atoi(v.c_str());
else if (k == "guidance_scale") r->guidance_scale = (float)atof(v.c_str());
else if (k == "shift") r->shift = (float)atof(v.c_str());
else if (k == "codes_temperature") r->codes_temperature = (float)atof(v.c_str());
else if (k == "codes_top_p") r->codes_top_p = (float)atof(v.c_str());
else if (k == "codes_top_k") r->codes_top_k = atoi(v.c_str());
// bools
else if (k == "thinking") r->thinking = (v == "true");
else if (k == "instrumental") r->instrumental = (v == "true");
@@ -279,7 +288,11 @@ bool request_write(const AceRequest * r, const char * path) {
fprintf(f, " \"lm_temperature\": %.2f,\n", r->lm_temperature);
fprintf(f, " \"lm_cfg_scale\": %.1f,\n", r->lm_cfg_scale);
fprintf(f, " \"lm_top_p\": %.2f,\n", r->lm_top_p);
fprintf(f, " \"lm_rep_pen\": %.2f,\n", r->lm_rep_pen);
fprintf(f, " \"lm_top_k\": %d,\n", r->lm_top_k);
fprintf(f, " \"codes_temperature\": %.2f,\n", r->codes_temperature);
fprintf(f, " \"codes_top_p\": %.2f,\n", r->codes_top_p);
fprintf(f, " \"codes_top_k\": %d,\n", r->codes_top_k);
fprintf(f, " \"lm_negative_prompt\": \"%s\",\n", json_escape(r->lm_negative_prompt).c_str());
fprintf(f, " \"inference_steps\": %d,\n", r->inference_steps);
fprintf(f, " \"guidance_scale\": %.1f,\n", r->guidance_scale);
@@ -302,8 +315,10 @@ void request_dump(const AceRequest * r, FILE * f) {
fprintf(f, " bpm=%d dur=%.0f key=%s ts=%s lang=%s\n",
r->bpm, r->duration, r->keyscale.c_str(),
r->timesignature.c_str(), r->vocal_language.c_str());
fprintf(f, " lm: temp=%.2f cfg=%.1f top_p=%.2f top_k=%d\n",
r->lm_temperature, r->lm_cfg_scale, r->lm_top_p, r->lm_top_k);
fprintf(f, " lm: temp=%.2f cfg=%.1f top_p=%.2f top_k=%d rep_pen=%d\n",
r->lm_temperature, r->lm_cfg_scale, r->lm_top_p, r->lm_top_k, r->lm_rep_pen);
fprintf(f, " codes: temp=%.2f top_p=%.2f top_k=%d \n",
r->lm_temperature, r->lm_top_p, r->lm_top_k);
fprintf(f, " dit: steps=%d guidance=%.1f shift=%.1f\n",
r->inference_steps, r->guidance_scale, r->shift);
fprintf(f, " audio_codes: %s\n",
otherarch/acestep/request.h
+5
View File
@@ -32,8 +32,13 @@ struct AceRequest {
float lm_cfg_scale; // 2.0
float lm_top_p; // 0.9
int lm_top_k; // 0 = disabled (matches Python None)
float lm_rep_pen; // 1.03
std::string lm_negative_prompt; // "NO USER INPUT"
int codes_top_k;
float codes_top_p;
float codes_temperature;
// codes (Python-compatible string: "3101,11837,27514,...")
// empty = text2music (silence context), non-empty = cover mode
std::string audio_codes; // ""