Merge pull request #3181 from SillyTavern/tc-split-generic Add generic text completion API type (100% OAI compatible)

8811010c7396911e6b035a4862d3c963e935dd69

Cohee <18619528+Cohee1207@users.noreply.github.com>

Signed
Parents: 5c1a9a300313f662f4a24a0a33e2078a5c4204dc e15e6dc3bd19ec35e1fd59a089dfc7f853fbf21d
11 files changed, +117 -12Ignore whitespace
public/index.html+30 -10
@@ -1235,7 +1235,8 @@
12351235 <input class="neo-range-slider" type="range" id="temp_textgenerationwebui" name="volume" min="0.0" max="5.0" step="0.01" x-setting-id="temp">
12361236 <input class="neo-range-input" type="number" min="0.0" max="5.0" step="0.01" data-for="temp_textgenerationwebui" id="temp_counter_textgenerationwebui">
12371237 </div>
1238- <div class="alignitemscenter flex-container flexFlowColumn flexBasis30p flexGrow flexShrink gap0">
1238+ <!-- Note: "except" mode = show for all BUT types in data-tg-type -->
1239+ <div data-tg-type-mode="except" data-tg-type="generic" class="alignitemscenter flex-container flexFlowColumn flexBasis30p flexGrow flexShrink gap0">
12391240 <small>
12401241 <span data-i18n="Top K">Top K</span>
12411242 <div class="fa-solid fa-circle-info opacity50p" title="Top K sets a maximum amount of top tokens that can be chosen from.&#13;E.g Top K is 20, this means only the 20 highest ranking tokens will be kept (regardless of their probabilities being diverse or limited).&#13;Set to 0 (or -1, depending on your backend) to disable." data-i18n="[title]Top_K_desc"></div>
@@ -1251,7 +1252,7 @@
12511252 <input class="neo-range-slider" type="range" id="top_p_textgenerationwebui" name="volume" min="0" max="1" step="0.01">
12521253 <input class="neo-range-input" type="number" min="0" max="1" step="0.01" data-for="top_p_textgenerationwebui" id="top_p_counter_textgenerationwebui">
12531254 </div>
12541255 <div data-tg-type-mode="except" data-tg-type="generic" class="alignitemscenter flex-container flexFlowColumn flexBasis30p flexGrow flexShrink gap0">
12551256 <small>
12561257 <span data-i18n="Typical P">Typical P</span>
12571258 <div class="fa-solid fa-circle-info opacity50p" title="Typical P Sampling prioritizes tokens based on their deviation from the average entropy of the set.&#13;It maintains tokens whose cumulative probability is close to a predefined threshold (e.g., 0.5), emphasizing those with average information content.&#13;Set to 1.0 to disable." data-i18n="[title]Typical_P_desc"></div>
@@ -1259,7 +1260,7 @@
12591260 <input class="neo-range-slider" type="range" id="typical_p_textgenerationwebui" name="volume" min="0" max="1" step="0.01">
12601261 <input class="neo-range-input" type="number" min="0" max="1" step="0.01" data-for="typical_p_textgenerationwebui" id="typical_p_counter_textgenerationwebui">
12611262 </div>
12621263 <div data-tg-type-mode="except" data-tg-type="generic" class="alignitemscenter flex-container flexFlowColumn flexBasis30p flexGrow flexShrink gap0">
12631264 <small>
12641265 <span data-i18n="Min P">Min P</span>
12651266 <div class="fa-solid fa-circle-info opacity50p" data-i18n="[title]Min_P_desc" title="Min P sets a base minimum probability. This is scaled according to the top token's probability.&#13;E.g If Top token is 80% probability, and Min P is 0.1, only tokens higher than 8% would be considered.&#13;Set to 0 to disable."></div>
@@ -1267,7 +1268,7 @@
12671268 <input class="neo-range-slider" type="range" id="min_p_textgenerationwebui" name="volume" min="0" max="1" step="0.001">
12681269 <input class="neo-range-input" type="number" min="0" max="1" step="0.001" data-for="min_p_textgenerationwebui" id="min_p_counter_textgenerationwebui">
12691270 </div>
12701271 <div data-tg-type-mode="except" data-tg-type="generic" class="alignitemscenter flex-container flexFlowColumn flexBasis30p flexGrow flexShrink gap0">
12711272 <small>
12721273 <span data-i18n="Top A">Top A</span>
12731274 <div class="fa-solid fa-circle-info opacity50p" title="Top A sets a threshold for token selection based on the square of the highest token probability.&#13;E.g if the Top-A value is 0.2 and the top token's probability is 50%, tokens with probabilities below 5% (0.2 * 0.5^2) are excluded.&#13;Set to 0 to disable." data-i18n="[title]Top_A_desc"></div>
@@ -1275,7 +1276,7 @@
12751276 <input class="neo-range-slider" type="range" id="top_a_textgenerationwebui" name="volume" min="0" max="1" step="0.01">
12761277 <input class="neo-range-input" type="number" min="0" max="1" step="0.01" data-for="top_a_textgenerationwebui" id="top_a_counter_textgenerationwebui">
12771278 </div>
12781279 <div data-tg-type-mode="except" data-tg-type="generic" class="alignitemscenter flex-container flexFlowColumn flexBasis30p flexGrow flexShrink gap0">
12791280 <small>
12801281 <span data-i18n="TFS">TFS</span>
12811282 <div class="fa-solid fa-circle-info opacity50p" data-i18n="[title]Tail_Free_Sampling_desc" title="Tail-Free Sampling (TFS) searches for a tail of low-probability tokens in the distribution,&#13;by analyzing the rate of change in token probabilities using derivatives. It retains tokens up to a threshold (e.g., 0.3) based on the normalized second derivative.&#13;The closer to 0, the more discarded tokens. Set to 1.0 to disable."></div>
@@ -1307,7 +1308,7 @@
13071308 <input class="neo-range-slider" type="range" id="eta_cutoff_textgenerationwebui" name="volume" min="0" max="20" step="0.01">
13081309 <input class="neo-range-input" type="number" min="0" max="20" step="0.01" data-for="eta_cutoff_textgenerationwebui" id="eta_cutoff_counter_textgenerationwebui">
13091310 </div>
13101311 <div data-tg-type-mode="except" data-tg-type="generic" class="alignitemscenter flex-container flexFlowColumn flexBasis30p flexGrow flexShrink gap0">
13111312 <small data-i18n="rep.pen">Repetition Penalty</small>
13121313 <input class="neo-range-slider" type="range" id="rep_pen_textgenerationwebui" name="volume" min="1" max="3" step="0.01">
13131314 <input class="neo-range-input" type="number" min="1" max="3" step="0.01" data-for="rep_pen_textgenerationwebui" id="rep_pen_counter_textgenerationwebui">
@@ -1569,7 +1570,7 @@
15691570 <small data-i18n="Ignore EOS Token">Ignore EOS Token</small>
15701571 <div class="fa-solid fa-circle-info opacity50p " data-i18n="[title]Ignore the EOS Token even if it generates." title="Ignore the EOS Token even if it generates."></div>
15711572 </label>
15721573 <label data-tg-type-mode="except" data-tg-type="generic" class="checkbox_label flexGrow flexShrink" for="skip_special_tokens_textgenerationwebui">
15731574 <input type="checkbox" id="skip_special_tokens_textgenerationwebui" />
15741575 <small data-i18n="Skip Special Tokens">Skip Special Tokens</small>
15751576 </label>
@@ -1594,14 +1595,14 @@
15941595 </label>
15951596 </div>
15961597 </div>
15971598 <div data-tg-type="mancer, ooba, koboldcpp, vllm, aphrodite, llamacpp, ollama, infermaticai, huggingface, generic" class="flex-container flexFlowColumn alignitemscenter flexBasis48p flexGrow flexShrink gap0">
15981599 <label>
15991600 <small data-i18n="Seed">Seed</small>
16001601 <div class="fa-solid fa-circle-info opacity50p " data-i18n="[title]Seed_desc" title="A random seed to use for deterministic and reproducable outputs. Set to -1 to use a random seed."></div>
16011602 </label>
16021603 <input type="number" id="seed_textgenerationwebui" class="text_pole textAlignCenter" min="-1" value="-1" />
16031604 </div>
16041605 <div data-tg-type-mode="except" data-tg-type="generic" id="banned_tokens_block_ooba" class="wide100p">
16051606 <hr class="width100p">
16061607 <h4 class="range-block-title justifyCenter">
16071608 <span data-i18n="Banned Tokens">Banned Tokens/Strings</span>
@@ -2187,10 +2188,10 @@
21872188 <div>
21882189 <h4 data-i18n="API Type">API Type</h4>
21892190 <select id="textgen_type">
2190- <option value="ooba" data-i18n="Default (completions compatible)">Default [OpenAI /completions compatible: oobabooga, LM Studio, etc.]</option>
21912191 <option value="aphrodite">Aphrodite</option>
21922192 <option value="dreamgen">DreamGen</option>
21932193 <option value="featherless">Featherless</option>
2194+ <option value="generic" data-i18n="Generic (OpenAI-compatible) [LM Studio, LiteLLM, etc.]">Generic (OpenAI-compatible) [LM Studio, LiteLLM, etc.]</option>
21942195 <option value="huggingface">HuggingFace (Inference Endpoint)</option>
21952196 <option value="infermaticai">InfermaticAI</option>
21962197 <option value="koboldcpp">KoboldCpp</option>
@@ -2199,6 +2200,7 @@
21992200 <option value="ollama">Ollama</option>
22002201 <option value="openrouter">OpenRouter</option>
22012202 <option value="tabby">TabbyAPI</option>
2203+ <option value="ooba">Text Generation WebUI (oobabooga)</option>
22022204 <option value="togetherai">TogetherAI</option>
22032205 <option value="vllm">vLLM</option>
22042206 </select>
@@ -2321,6 +2323,24 @@
23212323 </select>
23222324 </div>
23232325 </div>
2326+ <div data-tg-type="generic" class="flex-container flexFlowColumn">
2327+ <h4 data-i18n="API key (optional)">API key (optional)</h4>
2328+ <div class="flex-container">
2329+ <input id="api_key_generic" name="api_key_generic" class="text_pole flex1 wide100p" type="text" autocomplete="off">
2330+ <div title="Clear your API key" data-i18n="[title]Clear your API key" class="menu_button fa-solid fa-circle-xmark clear-api-key" data-key="api_key_generic">
2331+ </div>
2332+ </div>
2333+ <div data-for="api_key_generic" class="neutral_warning" data-i18n="For privacy reasons, your API key will be hidden after you reload the page.">
2334+ For privacy reasons, your API key will be hidden after you reload the page.
2335+ </div>
2336+ <div class="flex1">
2337+ <h4 data-i18n="Server url">Server URL</h4>
2338+ <small data-i18n="Example: 127.0.0.1:5000">Example: http://127.0.0.1:5000</small>
2339+ <input id="generic_api_url_text" name="generic_api_url" class="text_pole wide100p" value="" autocomplete="off" data-server-history="generic">
2340+ </div>
2341+ <datalist id="generic_model_fill"></datalist>
2342+ <input id="generic_model_textgenerationwebui" list="generic_model_fill" class="text_pole wide100p" placeholder="Model ID (optional)" data-i18n="[placeholder]Model ID (optional)" type="text">
2343+ </div>
23242344 <div data-tg-type="ooba" class="flex-container flexFlowColumn">
23252345 <div class="flex-container flexFlowColumn">
23262346 <a href="https://github.com/oobabooga/text-generation-webui" target="_blank">
public/script.js+5 -1
@@ -234,7 +234,7 @@ import {
234234import { getBackgrounds, initBackgrounds, loadBackgroundSettings, background_settings } from './scripts/backgrounds.js';
235235import { hideLoader, showLoader } from './scripts/loader.js';
236236import { BulkEditOverlay, CharacterContextMenu } from './scripts/BulkEditOverlay.js';
237237import { loadFeatherlessModels, loadMancerModels, loadOllamaModels, loadTogetherAIModels, loadInfermaticAIModels, loadOpenRouterModels, loadVllmModels, loadAphroditeModels, loadDreamGenModels, initTextGenModels, loadTabbyModels, loadGenericModels } from './scripts/textgen-models.js';
238238import { appendFileContent, hasPendingFileAttachment, populateFileAttachment, decodeStyleTags, encodeStyleTags, isExternalMediaAllowed, getCurrentEntityId, preserveNeutralChat, restoreNeutralChat } from './scripts/chats.js';
239239import { getPresetManager, initPresetManager } from './scripts/preset-manager.js';
240240import { evaluateMacros, getLastMessageId, initMacros } from './scripts/macros.js';
@@ -1221,6 +1221,9 @@ async function getStatusTextgen() {
12211221 } else if (textgen_settings.type === textgen_types.TABBY) {
12221222 loadTabbyModels(data?.data);
12231223 setOnlineStatus(textgen_settings.tabby_model || data?.result);
1224+ } else if (textgen_settings.type === textgen_types.GENERIC) {
1225+ loadGenericModels(data?.data);
1226+ setOnlineStatus(textgen_settings.generic_model || 'Connected');
12241227 } else {
12251228 setOnlineStatus(data?.result);
12261229 }
@@ -10042,6 +10045,7 @@ jQuery(async function () {
1004210045 { id: 'api_key_llamacpp', secret: SECRET_KEYS.LLAMACPP },
1004310046 { id: 'api_key_featherless', secret: SECRET_KEYS.FEATHERLESS },
1004410047 { id: 'api_key_huggingface', secret: SECRET_KEYS.HUGGINGFACE },
10048+ { id: 'api_key_generic', secret: SECRET_KEYS.GENERIC },
1004510049 ];
1004610050
1004710051 for (const key of keys) {
public/scripts/preset-manager.js+1 -0
@@ -585,6 +585,7 @@ class PresetManager {
585585 'openrouter_allow_fallbacks',
586586 'tabby_model',
587587 'derived',
588+ 'generic_model',
588589 ];
589590 const settings = Object.assign({}, getSettingsByApiId(this.apiId));
590591
public/scripts/secrets.js+2 -0
@@ -38,6 +38,7 @@ export const SECRET_KEYS = {
3838 NANOGPT: 'api_key_nanogpt',
3939 TAVILY: 'api_key_tavily',
4040 BFL: 'api_key_bfl',
41+ GENERIC: 'api_key_generic',
4142};
4243
4344const INPUT_MAP = {
@@ -71,6 +72,7 @@ const INPUT_MAP = {
7172 [SECRET_KEYS.HUGGINGFACE]: '#api_key_huggingface',
7273 [SECRET_KEYS.BLOCKENTROPY]: '#api_key_blockentropy',
7374 [SECRET_KEYS.NANOGPT]: '#api_key_nanogpt',
75+ [SECRET_KEYS.GENERIC]: '#api_key_generic',
7476};
7577
7678async function clearSecret() {
public/scripts/slash-commands.js+1 -0
@@ -3743,6 +3743,7 @@ function setBackgroundCallback(_, bg) {
37433743function getModelOptions(quiet) {
37443744 const nullResult = { control: null, options: null };
37453745 const modelSelectMap = [
3746+ { id: 'generic_model_textgenerationwebui', api: 'textgenerationwebui', type: textgen_types.GENERIC },
37463747 { id: 'custom_model_textgenerationwebui', api: 'textgenerationwebui', type: textgen_types.OOBA },
37473748 { id: 'model_togetherai_select', api: 'textgenerationwebui', type: textgen_types.TOGETHERAI },
37483749 { id: 'openrouter_model', api: 'textgenerationwebui', type: textgen_types.OPENROUTER },
public/scripts/textgen-models.js+18 -0
@@ -160,6 +160,24 @@ export async function loadInfermaticAIModels(data) {
160160 }
161161}
162162
163+export function loadGenericModels(data) {
164+ if (!Array.isArray(data)) {
165+ console.error('Invalid Generic models data', data);
166+ return;
167+ }
168+
169+ data.sort((a, b) => a.id.localeCompare(b.id));
170+ const dataList = $('#generic_model_fill');
171+ dataList.empty();
172+
173+ for (const model of data) {
174+ const option = document.createElement('option');
175+ option.value = model.id;
176+ option.text = model.id;
177+ dataList.append(option);
178+ }
179+}
180+
163181export async function loadDreamGenModels(data) {
164182 if (!Array.isArray(data)) {
165183 console.error('Invalid DreamGen models data', data);
public/scripts/textgen-settings.js+17 -0
@@ -33,9 +33,11 @@ export const textgen_types = {
3333 OPENROUTER: 'openrouter',
3434 FEATHERLESS: 'featherless',
3535 HUGGINGFACE: 'huggingface',
36+ GENERIC: 'generic',
3637};
3738
3839const {
40+ GENERIC,
3941 MANCER,
4042 VLLM,
4143 APHRODITE,
@@ -120,6 +122,7 @@ export const SERVER_INPUTS = {
120122 [textgen_types.LLAMACPP]: '#llamacpp_api_url_text',
121123 [textgen_types.OLLAMA]: '#ollama_api_url_text',
122124 [textgen_types.HUGGINGFACE]: '#huggingface_api_url_text',
125+ [textgen_types.GENERIC]: '#generic_api_url_text',
123126};
124127
125128const KOBOLDCPP_ORDER = [6, 0, 1, 3, 4, 2, 5];
@@ -205,6 +208,7 @@ const settings = {
205208 xtc_probability: 0,
206209 nsigma: 0.0,
207210 featherless_model: '',
211+ generic_model: '',
208212};
209213
210214export {
@@ -282,6 +286,7 @@ export const setting_names = [
282286 'xtc_threshold',
283287 'xtc_probability',
284288 'nsigma',
289+ 'generic_model',
285290];
286291
287292const DYNATEMP_BLOCK = document.getElementById('dynatemp_block_ooba');
@@ -833,7 +838,14 @@ jQuery(function () {
833838
834839function showTypeSpecificControls(type) {
835840 $('[data-tg-type]').each(function () {
841+ const mode = String($(this).attr('data-tg-type-mode') ?? '').toLowerCase().trim();
836842 const tgTypes = $(this).attr('data-tg-type').split(',').map(x => x.trim());
843+
844+ if (mode === 'except') {
845+ $(this)[tgTypes.includes(type) ? 'hide' : 'show']();
846+ return;
847+ }
848+
837849 for (const tgType of tgTypes) {
838850 if (tgType === type || tgType == 'all') {
839851 $(this).show();
@@ -1100,6 +1112,11 @@ export function getTextGenModel() {
11001112 return settings.custom_model;
11011113 }
11021114 break;
1115+ case GENERIC:
1116+ if (settings.generic_model) {
1117+ return settings.generic_model;
1118+ }
1119+ break;
11031120 case MANCER:
11041121 return settings.mancer_model;
11051122 case TOGETHERAI:
src/additional-headers.js+14 -0
@@ -172,6 +172,19 @@ function getHuggingFaceHeaders(directories) {
172172 }) : {};
173173}
174174
175+/**
176+ * Gets the headers for the Generic text completion API.
177+ * @param {import('./users.js').UserDirectoryList} directories
178+ * @returns {object} Headers for the request
179+ */
180+function getGenericHeaders(directories) {
181+ const apiKey = readSecret(directories, SECRET_KEYS.GENERIC);
182+
183+ return apiKey ? ({
184+ 'Authorization': `Bearer ${apiKey}`,
185+ }) : {};
186+}
187+
175188export function getOverrideHeaders(urlHost) {
176189 const requestOverrides = getConfigValue('requestOverrides', []);
177190 const overrideHeaders = requestOverrides?.find((e) => e.hosts?.includes(urlHost))?.headers;
@@ -214,6 +227,7 @@ export function setAdditionalHeadersByType(requestHeaders, type, server, directo
214227 [TEXTGEN_TYPES.LLAMACPP]: getLlamaCppHeaders,
215228 [TEXTGEN_TYPES.FEATHERLESS]: getFeatherlessHeaders,
216229 [TEXTGEN_TYPES.HUGGINGFACE]: getHuggingFaceHeaders,
230+ [TEXTGEN_TYPES.GENERIC]: getGenericHeaders,
217231 };
218232
219233 const getHeaders = headerGetters[type];
src/constants.js+19 -0
@@ -225,6 +225,7 @@ export const TEXTGEN_TYPES = {
225225 OPENROUTER: 'openrouter',
226226 FEATHERLESS: 'featherless',
227227 HUGGINGFACE: 'huggingface',
228+ GENERIC: 'generic',
228229};
229230
230231export const INFERMATICAI_KEYS = [
@@ -346,6 +347,24 @@ export const OLLAMA_KEYS = [
346347 'min_p',
347348];
348349
350+// https://platform.openai.com/docs/api-reference/completions
351+export const OPENAI_KEYS = [
352+ 'model',
353+ 'prompt',
354+ 'stream',
355+ 'temperature',
356+ 'top_p',
357+ 'frequency_penalty',
358+ 'presence_penalty',
359+ 'stop',
360+ 'seed',
361+ 'logit_bias',
362+ 'logprobs',
363+ 'max_tokens',
364+ 'n',
365+ 'best_of',
366+];
367+
349368export const AVATAR_WIDTH = 512;
350369export const AVATAR_HEIGHT = 768;
351370
src/endpoints/backends/text-completions.js+9 -1
@@ -13,6 +13,7 @@ import {
1313 VLLM_KEYS,
1414 DREAMGEN_KEYS,
1515 FEATHERLESS_KEYS,
16+ OPENAI_KEYS,
1617} from '../../constants.js';
1718import { forwardFetchResponse, trimV1, getConfigValue } from '../../util.js';
1819import { setAdditionalHeaders } from '../../additional-headers.js';
@@ -113,8 +114,8 @@ router.post('/status', jsonParser, async function (request, response) {
113114 let url = baseUrl;
114115 let result = '';
115116
116-
117117 switch (apiType) {
118+ case TEXTGEN_TYPES.GENERIC:
118119 case TEXTGEN_TYPES.OOBA:
119120 case TEXTGEN_TYPES.VLLM:
120121 case TEXTGEN_TYPES.APHRODITE:
@@ -287,6 +288,7 @@ router.post('/generate', jsonParser, async function (request, response) {
287288 let url = trimV1(baseUrl);
288289
289290 switch (request.body.api_type) {
291+ case TEXTGEN_TYPES.GENERIC:
290292 case TEXTGEN_TYPES.VLLM:
291293 case TEXTGEN_TYPES.FEATHERLESS:
292294 case TEXTGEN_TYPES.APHRODITE:
@@ -347,6 +349,12 @@ router.post('/generate', jsonParser, async function (request, response) {
347349 args.body = JSON.stringify(request.body);
348350 }
349351
352+ if (request.body.api_type === TEXTGEN_TYPES.GENERIC) {
353+ request.body = _.pickBy(request.body, (_, key) => OPENAI_KEYS.includes(key));
354+ if (Array.isArray(request.body.stop)) { request.body.stop = request.body.stop.slice(0, 4); }
355+ args.body = JSON.stringify(request.body);
356+ }
357+
350358 if (request.body.api_type === TEXTGEN_TYPES.OPENROUTER) {
351359 if (Array.isArray(request.body.provider) && request.body.provider.length > 0) {
352360 request.body.provider = {
src/endpoints/secrets.js+1 -0
@@ -50,6 +50,7 @@ export const SECRET_KEYS = {
5050 TAVILY: 'api_key_tavily',
5151 NANOGPT: 'api_key_nanogpt',
5252 BFL: 'api_key_bfl',
53+ GENERIC: 'api_key_generic',
5354};
5455
5556// These are the keys that are safe to expose, even if allowKeysExposure is false