mirror of
https://github.com/SillyTavern/SillyTavern.git
synced 2025-06-05 21:59:27 +02:00
Implement Token Probabilities UI using logprobs
This commit is contained in:
khanon
@ -1132,13 +1132,15 @@ export function initRossMods() {
|
||||
.not('#right-nav-panel')
|
||||
.not('#floatingPrompt')
|
||||
.not('#cfgConfig')
|
||||
.not("#logprobsViewer")
|
||||
.is(':visible')) {
|
||||
let visibleDrawerContent = $('.drawer-content:visible')
|
||||
.not('#WorldInfo')
|
||||
.not('#left-nav-panel')
|
||||
.not('#right-nav-panel')
|
||||
.not('#floatingPrompt')
|
||||
.not('#cfgConfig');
|
||||
.not('#cfgConfig')
|
||||
.not("#logprobsViewer");
|
||||
$(visibleDrawerContent).parent().find('.drawer-icon').trigger('click');
|
||||
return;
|
||||
}
|
||||
@ -1158,6 +1160,11 @@ export function initRossMods() {
|
||||
return;
|
||||
}
|
||||
|
||||
if ($('#logprobsViewer').is(':visible')) {
|
||||
$('#logprobsViewerClose').trigger('click');
|
||||
return;
|
||||
}
|
||||
|
||||
if ($('#left-nav-panel').is(':visible') &&
|
||||
$(LPanelPin).prop('checked') === false) {
|
||||
$('#leftNavDrawerIcon').trigger('click');
|
||||
|
||||
466
public/scripts/logprobs.js
Normal file
466
public/scripts/logprobs.js
Normal file
@ -0,0 +1,466 @@
|
||||
import {
|
||||
animation_duration,
|
||||
callPopup,
|
||||
chat,
|
||||
cleanUpMessage,
|
||||
event_types,
|
||||
eventSource,
|
||||
Generate,
|
||||
getGeneratingApi,
|
||||
is_send_press,
|
||||
} from '../script.js';
|
||||
import { debounce, delay, getStringHash } from './utils.js';
|
||||
import { decodeTextTokens, getTokenizerBestMatch } from './tokenizers.js';
|
||||
import { power_user } from './power-user.js';
|
||||
|
||||
const TINTS = 4;
|
||||
const MAX_MESSAGE_LOGPROBS = 100;
|
||||
|
||||
/**
|
||||
* Tuple of a candidate token and its logarithm of probability of being chosen
|
||||
* @typedef {[string, number]} Candidate - (token, logprob)
|
||||
*/
|
||||
|
||||
/**
|
||||
* Logprob data for a single message
|
||||
* @typedef {Object} MessageLogprobData
|
||||
* @property {number} created - timestamp of when the message was generated
|
||||
* @property {number} hash - hash of the message object
|
||||
* @property {number} messageId - ID of the source message
|
||||
* @property {number} swipeId - ID of the source swipe on the source message
|
||||
* @property {string} api - API used to generate the message
|
||||
* @property {TokenLogprobs[]} messageLogprobs Logprob data for each token, by
|
||||
* its index in the message
|
||||
* @property {string | null} continueFrom - the 'continue' prefix used to
|
||||
* generate the message, if any
|
||||
*/
|
||||
|
||||
/**
|
||||
* Logprob data for a single token
|
||||
* @typedef {Object} TokenLogprobs
|
||||
* @property {string} token - A token generated by the model
|
||||
* @property {Candidate[]} topLogprobs - Array of top candidate tokens
|
||||
*/
|
||||
|
||||
let state = {
|
||||
/** @type {TokenLogprobs | null} */
|
||||
selectedTokenLogprobs: null,
|
||||
/** @type {Map<number, MessageLogprobData>} */
|
||||
messageLogprobs: new Map(),
|
||||
};
|
||||
|
||||
/**
|
||||
* renderAlternativeTokensView renders the Token Probabilities UI and all
|
||||
* subviews with the active message's logprobs data. If the message has no token
|
||||
* logprobs, a zero-state is rendered.
|
||||
*/
|
||||
function renderAlternativeTokensView() {
|
||||
const view = $('#logprobs_generation_output');
|
||||
if (!view.is(':visible')) {
|
||||
return;
|
||||
}
|
||||
view.empty();
|
||||
state.selectedTokenLogprobs = null;
|
||||
renderTopLogprobs();
|
||||
|
||||
const { messageLogprobs, continueFrom } = getActiveMessageLogprobData() || {};
|
||||
if (!messageLogprobs?.length) {
|
||||
const emptyState = $('<div></div>');
|
||||
const msg = power_user.request_token_probabilities
|
||||
? 'No token probabilities available for the current message.'
|
||||
: `<span>Enable <b>Request token probabilities</b> in the User Settings menu to use this feature.</span>`;
|
||||
emptyState.html(msg);
|
||||
emptyState.addClass('logprobs_empty_state');
|
||||
view.append(emptyState);
|
||||
return;
|
||||
}
|
||||
|
||||
const prefix = continueFrom || '';
|
||||
const tokenSpans = [];
|
||||
|
||||
if (prefix) {
|
||||
const prefixSpan = $('<span></span>');
|
||||
prefixSpan.text(prefix);
|
||||
prefixSpan.html(prefixSpan.html().replace(/\n/g, '<br>'));
|
||||
prefixSpan.addClass('logprobs_output_prefix');
|
||||
prefixSpan.attr('title', 'Select to reroll the last \'Continue\' generation');
|
||||
prefixSpan.click(onPrefixClicked);
|
||||
addKeyboardProps(prefixSpan);
|
||||
tokenSpans.push(...withVirtualWhitespace(prefix, prefixSpan));
|
||||
}
|
||||
|
||||
messageLogprobs.forEach((tokenData, i) => {
|
||||
const { token } = tokenData;
|
||||
const span = $('<span></span>');
|
||||
const text = toVisibleWhitespace(token);
|
||||
span.text(text);
|
||||
span.addClass('logprobs_output_token');
|
||||
span.addClass('logprobs_tint_' + (i % TINTS));
|
||||
span.click(() => onSelectedTokenChanged(tokenData, span));
|
||||
addKeyboardProps(span);
|
||||
tokenSpans.push(...withVirtualWhitespace(token, span));
|
||||
});
|
||||
|
||||
view.append(tokenSpans);
|
||||
|
||||
// scroll past long prior context
|
||||
if (prefix) {
|
||||
view.find('.logprobs_output_token').first()[0].scrollIntoView();
|
||||
}
|
||||
}
|
||||
|
||||
function addKeyboardProps(element) {
|
||||
element.attr('role', 'button');
|
||||
element.attr('tabindex', '0');
|
||||
element.keydown(function (e) {
|
||||
if (e.key === 'Enter' || e.key === ' ') {
|
||||
element.click();
|
||||
}
|
||||
});
|
||||
}
|
||||
|
||||
/**
|
||||
* renderTopLogprobs renders the top logprobs subview with the currently
|
||||
* selected token highlighted. If no token is selected, the subview is hidden.
|
||||
*/
|
||||
function renderTopLogprobs() {
|
||||
const view = $('.logprobs_candidate_list');
|
||||
const hint = $('#logprobs_top_logprobs_hint').hide();
|
||||
view.empty();
|
||||
|
||||
if (!state.selectedTokenLogprobs) {
|
||||
return;
|
||||
}
|
||||
|
||||
const { token: selectedToken, topLogprobs } = state.selectedTokenLogprobs;
|
||||
|
||||
let sum = 0;
|
||||
const nodes = [];
|
||||
const candidates = topLogprobs
|
||||
.sort(([, logA], [, logB]) => logB - logA)
|
||||
.map(([text, log]) => {
|
||||
const probability = Math.exp(log);
|
||||
sum += probability;
|
||||
return [text, probability, log];
|
||||
});
|
||||
candidates.push(['<others>', 1 - sum, 0]);
|
||||
|
||||
let matched = false;
|
||||
for (const [token, probability, log] of candidates) {
|
||||
const container = $('<button class="flex-container flexFlowColumn logprobs_top_candidate"></button>');
|
||||
|
||||
if (token === selectedToken) {
|
||||
matched = true;
|
||||
container.addClass('selected');
|
||||
}
|
||||
|
||||
const tokenText = $('<span></span>').text(`${toVisibleWhitespace(token)}`);
|
||||
const percentText = $('<span></span>').text(`${(probability * 100).toFixed(2)}%`);
|
||||
container.append(tokenText, percentText);
|
||||
container.attr('title', `logarithm: ${log}`);
|
||||
addKeyboardProps(container);
|
||||
if (token !== '<others>') {
|
||||
container.click(() => onAlternativeClicked(state.selectedTokenLogprobs, token));
|
||||
} else {
|
||||
container.prop('disabled', true);
|
||||
}
|
||||
nodes.push(container);
|
||||
}
|
||||
|
||||
// Highlight the <others> node if the selected token was not included in the
|
||||
// top logprobs
|
||||
if (!matched) {
|
||||
nodes[nodes.length - 1].css('background-color', 'rgba(255, 0, 0, 0.1)');
|
||||
}
|
||||
|
||||
view.append(nodes);
|
||||
}
|
||||
|
||||
/**
|
||||
* onSelectedTokenChanged is called when the user clicks on a token in the
|
||||
* token output view. It updates the selected token state and re-renders the
|
||||
* top logprobs view, or deselects the token if it was already selected.
|
||||
* @param {TokenLogprobs} logprobs - logprob data for the selected token
|
||||
* @param {Element} span - target span node that was clicked
|
||||
*/
|
||||
function onSelectedTokenChanged(logprobs, span) {
|
||||
$('.logprobs_output_token.selected').removeClass('selected');
|
||||
if (state.selectedTokenLogprobs === logprobs) {
|
||||
state.selectedTokenLogprobs = null;
|
||||
} else {
|
||||
state.selectedTokenLogprobs = logprobs;
|
||||
$(span).addClass('selected');
|
||||
}
|
||||
renderTopLogprobs();
|
||||
}
|
||||
|
||||
/**
|
||||
* onAlternativeClicked is called when the user clicks on an alternative token
|
||||
* in the top logprobs view. It will create a new swipe message and prefill it
|
||||
* with all text up to the selected token, followed by the chosen alternative.
|
||||
* Then it requests a `continue` completion from the model with the new prompt.
|
||||
* @param {TokenLogprobs} tokenLogprobs - logprob data for selected alternative
|
||||
* @param {string} alternative - selected alternative token's text
|
||||
*/
|
||||
function onAlternativeClicked(tokenLogprobs, alternative) {
|
||||
if (!checkGenerateReady()) {
|
||||
return;
|
||||
}
|
||||
|
||||
if (getGeneratingApi() === 'openai') {
|
||||
return callPopup(`<h3>Feature unavailable</h3><p>Due to API limitations, rerolling a token is not supported with OpenAI. Try switching to a different API.</p>`, 'text');
|
||||
}
|
||||
|
||||
const { messageLogprobs, continueFrom } = getActiveMessageLogprobData();
|
||||
const replaceIndex = messageLogprobs.findIndex(x => x === tokenLogprobs);
|
||||
|
||||
const tokens = messageLogprobs.slice(0, replaceIndex + 1).map(({ token }) => token);
|
||||
tokens[replaceIndex] = alternative;
|
||||
|
||||
const prefix = continueFrom || '';
|
||||
const prompt = prefix + tokens.join('');
|
||||
const messageId = chat.length - 1;
|
||||
createSwipe(messageId, prompt);
|
||||
|
||||
$('.swipe_right:last').click(); // :see_no_evil:
|
||||
|
||||
Generate('continue').then(_ => void _);
|
||||
}
|
||||
|
||||
/**
|
||||
* onPrefixClicked is called when the user clicks on the carried-over prefix
|
||||
* in the token output view. It allows them to reroll the last 'continue'
|
||||
* completion with none of the output generated from it, in case they don't
|
||||
* like the results.
|
||||
*/
|
||||
function onPrefixClicked() {
|
||||
if (!checkGenerateReady()) {
|
||||
return;
|
||||
}
|
||||
|
||||
const { continueFrom } = getActiveMessageLogprobData();
|
||||
const messageId = chat.length - 1;
|
||||
const prefix = continueFrom || '';
|
||||
createSwipe(messageId, prefix);
|
||||
$('.swipe_right:last').click();
|
||||
Generate('continue').then(_ => void _);
|
||||
}
|
||||
|
||||
function checkGenerateReady() {
|
||||
if (is_send_press) {
|
||||
toastr.warning(`Please wait for the current generation to complete.`);
|
||||
return false;
|
||||
}
|
||||
return true;
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* onToggleLogprobsPanel is called when the user performs an action that toggles
|
||||
* the logprobs view, such as clicking the Token Probabilities menu item or the
|
||||
* close button.
|
||||
*/
|
||||
function onToggleLogprobsPanel() {
|
||||
const logprobsViewer = $('#logprobsViewer');
|
||||
|
||||
// largely copied from CFGScale toggle
|
||||
if (logprobsViewer.css('display') === 'none') {
|
||||
logprobsViewer.addClass('resizing');
|
||||
logprobsViewer.css('display', 'flex');
|
||||
logprobsViewer.css('opacity', 0.0);
|
||||
renderAlternativeTokensView();
|
||||
logprobsViewer.transition({
|
||||
opacity: 1.0,
|
||||
duration: animation_duration,
|
||||
}, async function () {
|
||||
await delay(50);
|
||||
logprobsViewer.removeClass('resizing');
|
||||
});
|
||||
} else {
|
||||
logprobsViewer.addClass('resizing');
|
||||
logprobsViewer.transition({
|
||||
opacity: 0.0,
|
||||
duration: animation_duration,
|
||||
},
|
||||
async function () {
|
||||
await delay(50);
|
||||
logprobsViewer.removeClass('resizing');
|
||||
});
|
||||
setTimeout(function () {
|
||||
logprobsViewer.hide();
|
||||
}, animation_duration);
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* createSwipe appends a new swipe to the target chat message with the given
|
||||
* text.
|
||||
* @param {number} messageId - target chat message ID
|
||||
* @param {string} prompt - initial prompt text which will be continued
|
||||
*/
|
||||
function createSwipe(messageId, prompt) {
|
||||
// need to call `cleanUpMessage` on our new prompt, because we were working
|
||||
// with raw model output and our new prompt is missing trimming/macro replacements
|
||||
const cleanedPrompt = cleanUpMessage(prompt, false, false);
|
||||
|
||||
const msg = chat[messageId];
|
||||
const newSwipeInfo = {
|
||||
send_date: msg.send_date,
|
||||
gen_started: msg.gen_started,
|
||||
gen_finished: msg.gen_finished,
|
||||
extra: { ...structuredClone(msg.extra), from_logprobs: new Date().getTime() },
|
||||
};
|
||||
|
||||
msg.swipes = msg.swipes || [];
|
||||
msg.swipe_info = msg.swipe_info || [];
|
||||
|
||||
// Add our new swipe, then make sure the active swipe is the one just before
|
||||
// it. The call to `swipe_right` will switch to it immediately.
|
||||
msg.swipes.push(cleanedPrompt);
|
||||
msg.swipe_info.push(newSwipeInfo);
|
||||
msg.swipe_id = Math.max(0, msg.swipes.length - 2);
|
||||
}
|
||||
|
||||
/**
|
||||
* toVisibleWhitespace receives input text and replaces spaces with · and
|
||||
* newlines with ↵.
|
||||
* @param {string} input
|
||||
* @returns {string}
|
||||
*/
|
||||
function toVisibleWhitespace(input) {
|
||||
return input.replace(/ /g, '·').replace(/\n/g, '↵');
|
||||
}
|
||||
|
||||
/**
|
||||
* withVirtualWhitespace inserts line breaks and a zero-width space before and
|
||||
* after the span node if its token begins or ends with whitespace in order to
|
||||
* allow text to wrap despite whitespace characters being replaced with a dot.
|
||||
* @param {string} text - token text being evaluated for whitespace
|
||||
* @param {Element} span - target span node to be wrapped
|
||||
* @returns {Element[]} array of nodes to be appended to the DOM
|
||||
*/
|
||||
function withVirtualWhitespace(text, span) {
|
||||
const result = [span];
|
||||
if (text.match(/^\s/)) {
|
||||
result.unshift(document.createTextNode('\u200b'));
|
||||
}
|
||||
if (text.match(/\s$/)) {
|
||||
result.push($(document.createTextNode('\u200b')));
|
||||
}
|
||||
// line breaks are trickier. we don't currently handle consecutive line
|
||||
// breaks or line breaks occuring in between non-whitespace characters, but
|
||||
// tokenizers generally don't produce those anyway.
|
||||
|
||||
// matches leading line break, at least one character, and trailing line break
|
||||
if (text.match(/^\n(?:.|\n)+\n$/)) {
|
||||
result.unshift($('<br>'));
|
||||
result.push($('<br>'));
|
||||
} else if (text.match(/^\n/)) {
|
||||
result.unshift($('<br>'));
|
||||
} else if (text.match(/\n$/)) {
|
||||
result.push($('<br>'));
|
||||
}
|
||||
return result;
|
||||
}
|
||||
|
||||
/**
|
||||
* saveLogprobsForActiveMessage receives an array of TokenLogprobs objects
|
||||
* representing the top logprobs for each token in a message and associates it
|
||||
* with the active message.
|
||||
*
|
||||
* **Ensure the active message has been updated and rendered before calling
|
||||
* this function or the logprobs data will be saved to the wrong message.**
|
||||
* @param {TokenLogprobs[]} logprobs - array of logprobs data for each token
|
||||
* @param {string | null} continueFrom - for 'continue' generations, the prompt
|
||||
*/
|
||||
export function saveLogprobsForActiveMessage(logprobs, continueFrom) {
|
||||
convertTokenIdLogprobsToText(logprobs);
|
||||
|
||||
const msgId = chat.length - 1;
|
||||
/** @type {MessageLogprobData} */
|
||||
const data = {
|
||||
created: new Date().getTime(),
|
||||
api: getGeneratingApi(),
|
||||
messageId: msgId,
|
||||
swipeId: chat[msgId].swipe_id,
|
||||
messageLogprobs: logprobs,
|
||||
continueFrom,
|
||||
hash: getMessageHash(chat[msgId]),
|
||||
}
|
||||
|
||||
state.messageLogprobs.set(data.hash, data);
|
||||
|
||||
// Clean up old logprobs data
|
||||
const oldLogprobs = Array.from(state.messageLogprobs.values())
|
||||
.sort((a, b) => b.created - a.created)
|
||||
.slice(MAX_MESSAGE_LOGPROBS);
|
||||
for (const oldData of oldLogprobs) {
|
||||
state.messageLogprobs.delete(oldData.hash);
|
||||
}
|
||||
}
|
||||
|
||||
function getMessageHash(message) {
|
||||
// We don't use the swipe ID as a hash component because it's not stable,
|
||||
// deleting a swipe will change the ID of all subsequent swipes.
|
||||
const hashParams = {
|
||||
name: message.name,
|
||||
mid: chat.indexOf(message),
|
||||
text: message.mes,
|
||||
};
|
||||
return getStringHash(JSON.stringify(hashParams));
|
||||
}
|
||||
|
||||
/**
|
||||
* getActiveMessageLogprobData returns the logprobs data for the active chat
|
||||
* message.
|
||||
* @returns {MessageLogprobData || null}
|
||||
*/
|
||||
function getActiveMessageLogprobData() {
|
||||
const hash = getMessageHash(chat[chat.length - 1]);
|
||||
return state.messageLogprobs.get(hash) || null;
|
||||
}
|
||||
|
||||
/**
|
||||
* convertLogprobTokenIdsToText mutates the given logprobs data's topLogprobs
|
||||
* field keyed by token text instead of token ID. This is only necessary for
|
||||
* APIs which only return token IDs in their logprobs data; for others this
|
||||
* function is a no-op.
|
||||
* @param {TokenLogprobs[]} input - logprobs data with numeric token IDs
|
||||
*/
|
||||
function convertTokenIdLogprobsToText(input) {
|
||||
const api = getGeneratingApi();
|
||||
if (api !== 'novel') {
|
||||
return input;
|
||||
}
|
||||
|
||||
const tokenizerId = getTokenizerBestMatch(api);
|
||||
|
||||
// Flatten unique token IDs across all logprobs
|
||||
const tokenIds = Array.from(new Set(input.flatMap(logprobs =>
|
||||
logprobs.topLogprobs.map(([token]) => token).concat(logprobs.token)
|
||||
)));
|
||||
|
||||
// Submit token IDs to tokenizer to get token text, then build ID->text map
|
||||
const { chunks } = decodeTextTokens(tokenizerId, tokenIds);
|
||||
const tokenIdText = new Map(tokenIds.map((id, i) => [id, chunks[i]]));
|
||||
|
||||
// Fixup logprobs data with token text
|
||||
input.forEach(logprobs => {
|
||||
logprobs.token = tokenIdText.get(logprobs.token);
|
||||
logprobs.topLogprobs = logprobs.topLogprobs.map(([token, logprob]) =>
|
||||
[tokenIdText.get(token), logprob]
|
||||
);
|
||||
});
|
||||
}
|
||||
|
||||
export function initLogprobs() {
|
||||
const debouncedRender = debounce(renderAlternativeTokensView, 250);
|
||||
$('#logprobsViewerClose').click(onToggleLogprobsPanel);
|
||||
$('#option_toggle_logprobs').click(onToggleLogprobsPanel);
|
||||
eventSource.on(event_types.CHAT_CHANGED, debouncedRender);
|
||||
eventSource.on(event_types.CHARACTER_MESSAGE_RENDERED, debouncedRender);
|
||||
eventSource.on(event_types.IMPERSONATE_READY, debouncedRender);
|
||||
eventSource.on(event_types.MESSAGE_DELETED, debouncedRender);
|
||||
eventSource.on(event_types.MESSAGE_EDITED, debouncedRender);
|
||||
eventSource.on(event_types.MESSAGE_SWIPED, debouncedRender);
|
||||
}
|
||||
@ -416,10 +416,7 @@ export function getNovelGenerationData(finalPrompt, settings, maxLength, isImper
|
||||
cfgValues.negativePrompt = (getCfgPrompt(cfgValues.guidanceScale, true))?.value;
|
||||
}
|
||||
|
||||
const clio = nai_settings.model_novel.includes('clio');
|
||||
const kayra = nai_settings.model_novel.includes('kayra');
|
||||
|
||||
const tokenizerType = kayra ? tokenizers.NERD2 : (clio ? tokenizers.NERD : tokenizers.NONE);
|
||||
const tokenizerType = getTokenizerTypeForModel(nai_settings.model_novel);
|
||||
const stopSequences = (tokenizerType !== tokenizers.NONE)
|
||||
? getStoppingStrings(isImpersonate, isContinue)
|
||||
.map(t => getTextTokens(tokenizerType, t))
|
||||
@ -471,6 +468,7 @@ export function getNovelGenerationData(finalPrompt, settings, maxLength, isImper
|
||||
'return_full_text': false,
|
||||
'prefix': prefix,
|
||||
'order': nai_settings.order || settings.order || default_order,
|
||||
'num_logprobs': power_user.request_token_probabilities ? 10 : undefined,
|
||||
};
|
||||
}
|
||||
|
||||
@ -491,6 +489,16 @@ function selectPrefix(selected_prefix, finalPrompt) {
|
||||
return 'vanilla';
|
||||
}
|
||||
|
||||
function getTokenizerTypeForModel(model) {
|
||||
if (model.includes('clio')) {
|
||||
return tokenizers.NERD;
|
||||
}
|
||||
if (model.includes('kayra')) {
|
||||
return tokenizers.NERD2;
|
||||
}
|
||||
return tokenizers.NONE;
|
||||
}
|
||||
|
||||
// Sort the samplers by the order array
|
||||
function sortItemsByOrder(orderArray) {
|
||||
console.debug('Preset samplers order: ' + orderArray);
|
||||
@ -540,9 +548,7 @@ function calculateLogitBias() {
|
||||
return [];
|
||||
}
|
||||
|
||||
const clio = nai_settings.model_novel.includes('clio');
|
||||
const kayra = nai_settings.model_novel.includes('kayra');
|
||||
const tokenizerType = kayra ? tokenizers.NERD2 : (clio ? tokenizers.NERD : tokenizers.NONE);
|
||||
const tokenizerType = getTokenizerTypeForModel(nai_settings.model_novel);
|
||||
|
||||
/**
|
||||
* Creates a bias object for Novel AI
|
||||
@ -624,11 +630,68 @@ export async function generateNovelWithStreaming(generate_data, signal) {
|
||||
text += data.token;
|
||||
}
|
||||
|
||||
yield { text, swipes: [] };
|
||||
yield { text, swipes: [], logprobs: parseNovelAILogprobs(data.logprobs) };
|
||||
}
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* A single token's ID.
|
||||
* @typedef {[number]} TokenIdEntry
|
||||
*/
|
||||
/**
|
||||
* A single token's log probabilities. The first element is before repetition
|
||||
* penalties and samplers are applied, the second is after.
|
||||
* @typedef {[number, number]} LogprobsEntry
|
||||
*/
|
||||
/**
|
||||
* Combination of token ID and its corresponding log probabilities.
|
||||
* @typedef {[TokenIdEntry, LogprobsEntry]} TokenLogprobTuple
|
||||
*/
|
||||
/**
|
||||
* Represents all logprob data for a single token, including its
|
||||
* before, after, and the ultimately selected token.
|
||||
* @typedef {Object} NAITokenLogprobs
|
||||
* @property {TokenLogprobTuple[]} chosen - always length 1
|
||||
* @property {TokenLogprobTuple[]} before - always `top_logprobs` length
|
||||
* @property {TokenLogprobTuple[]} after - maybe less than `top_logprobs` length
|
||||
*/
|
||||
/**
|
||||
* parseNovelAILogprobs converts a logprobs object returned from the NovelAI API
|
||||
* for a single token into a TokenLogprobs object used by the Token Probabilities
|
||||
* feature.
|
||||
* @param {NAITokenLogprobs} data - NAI logprobs object for one token
|
||||
* @returns {import('logprobs.js').TokenLogprobs | null} converted logprobs
|
||||
*/
|
||||
export function parseNovelAILogprobs(data) {
|
||||
if (!data) {
|
||||
return null;
|
||||
}
|
||||
const befores = data.before.map(([[tokenId], [before, _]]) => [tokenId, before]);
|
||||
const afters = data.after.map(([[tokenId], [_, after]]) => [tokenId, after]);
|
||||
|
||||
// Find any tokens in `befores` that are missing from `afters`. Then add
|
||||
// them with a logprob of -Infinity (0% probability)
|
||||
const notInAfter = befores
|
||||
.filter(([id]) => !afters.some(([aid]) => aid === id))
|
||||
.map(([id]) => [id, -Infinity])
|
||||
const merged = afters.concat(notInAfter);
|
||||
|
||||
// Add the chosen token to `merged` if it's not already there. This can
|
||||
// happen if the chosen token was not among the top 10 most likely ones.
|
||||
const [[chosenId], [_, chosenAfter]] = data.chosen[0];
|
||||
if (!merged.some(([id]) => id === chosenId)) {
|
||||
merged.push([chosenId, chosenAfter]);
|
||||
}
|
||||
|
||||
// nb: returned logprobs are provided alongside token IDs, not decoded text.
|
||||
// We don't want to send an API call for every streaming tick to decode the
|
||||
// text so we will use the IDs instead and bulk decode them in
|
||||
// StreamingProcessor. JSDoc typechecking may complain about this, but it's
|
||||
// intentional.
|
||||
return { token: chosenId, topLogprobs: merged };
|
||||
}
|
||||
|
||||
$('#nai_preamble_textarea').on('input', function () {
|
||||
nai_settings.preamble = String($('#nai_preamble_textarea').val());
|
||||
saveSettingsDebounced();
|
||||
|
||||
@ -63,6 +63,7 @@ import {
|
||||
formatInstructModeSystemPrompt,
|
||||
} from './instruct-mode.js';
|
||||
import { isMobile } from './RossAscends-mods.js';
|
||||
import { saveLogprobsForActiveMessage } from './logprobs.js';
|
||||
|
||||
export {
|
||||
openai_messages_count,
|
||||
@ -1534,6 +1535,7 @@ async function sendOpenAIRequest(type, messages, signal) {
|
||||
const isImpersonate = type === 'impersonate';
|
||||
const isContinue = type === 'continue';
|
||||
const stream = oai_settings.stream_openai && !isQuiet && !isScale && !isAI21 && !(isGoogle && oai_settings.google_model.includes('bison'));
|
||||
const useLogprobs = !!power_user.request_token_probabilities;
|
||||
|
||||
if (isTextCompletion && isOpenRouter) {
|
||||
messages = convertChatCompletionToInstruct(messages, type);
|
||||
@ -1601,6 +1603,11 @@ async function sendOpenAIRequest(type, messages, signal) {
|
||||
generate_data['proxy_password'] = oai_settings.proxy_password;
|
||||
}
|
||||
|
||||
// Add logprobs request (currently OpenAI only, max 5 on their side)
|
||||
if (useLogprobs && isOAI) {
|
||||
generate_data['logprobs'] = 5;
|
||||
}
|
||||
|
||||
if (isClaude) {
|
||||
generate_data['top_k'] = Number(oai_settings.top_k_openai);
|
||||
generate_data['exclude_assistant'] = oai_settings.exclude_assistant;
|
||||
@ -1689,8 +1696,9 @@ async function sendOpenAIRequest(type, messages, signal) {
|
||||
const rawData = isSSEStream ? value.data : utf8Decoder.decode(value, { stream: true });
|
||||
if (isSSEStream && rawData === '[DONE]') return;
|
||||
tryParseStreamingError(response, rawData);
|
||||
text += getStreamingReply(JSON.parse(rawData));
|
||||
yield { text, swipes: [] };
|
||||
const parsed = JSON.parse(rawData);
|
||||
text += getStreamingReply(parsed);
|
||||
yield { text, swipes: [], logprobs: parseChatCompletionLogprobs(parsed) };
|
||||
}
|
||||
};
|
||||
}
|
||||
@ -1705,6 +1713,13 @@ async function sendOpenAIRequest(type, messages, signal) {
|
||||
throw new Error(data);
|
||||
}
|
||||
|
||||
if (type !== 'quiet') {
|
||||
const logprobs = parseChatCompletionLogprobs(data);
|
||||
// Delay is required to allow the active message to be updated to
|
||||
// the one we are generating (happens right after sendOpenAIRequest)
|
||||
delay(1).then(() => saveLogprobsForActiveMessage(logprobs, null));
|
||||
}
|
||||
|
||||
return !isTextCompletion ? data.choices[0]['message']['content'] : data.choices[0]['text'];
|
||||
}
|
||||
}
|
||||
@ -1719,6 +1734,88 @@ function getStreamingReply(data) {
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* parseChatCompletionLogprobs converts the response data returned from a chat
|
||||
* completions-like source into an array of TokenLogprobs found in the response.
|
||||
* @param {Object} data - response data from a chat completions-like source
|
||||
* @returns {import('logprobs.js').TokenLogprobs[] | null} converted logprobs
|
||||
*/
|
||||
function parseChatCompletionLogprobs(data) {
|
||||
if (!data) {
|
||||
return null;
|
||||
}
|
||||
|
||||
switch (oai_settings.chat_completion_source) {
|
||||
case chat_completion_sources.OPENAI:
|
||||
if (!data.choices?.length) {
|
||||
return null;
|
||||
}
|
||||
// OpenAI Text Completion API is treated as a chat completion source
|
||||
// by SillyTavern, hence its presence in this function.
|
||||
return textCompletionModels.includes(oai_settings.openai_model)
|
||||
? parseOpenAITextLogprobs(data.choices[0]?.logprobs)
|
||||
: parseOpenAIChatLogprobs(data.choices[0]?.logprobs);
|
||||
default:
|
||||
// implement other chat completion sources here
|
||||
}
|
||||
return null;
|
||||
}
|
||||
|
||||
/**
|
||||
* parseOpenAIChatLogprobs receives a `logprobs` response from OpenAI's chat
|
||||
* completion API and converts into the structure used by the Token Probabilities
|
||||
* view.
|
||||
* @param {{content: { token: string, logprob: number, top_logprobs: { token: string, logprob: number }[] }[]}} logprobs
|
||||
* @returns {import('logprobs.js').TokenLogprobs[] | null} converted logprobs
|
||||
*/
|
||||
function parseOpenAIChatLogprobs(logprobs) {
|
||||
const { content } = logprobs ?? {};
|
||||
|
||||
if (!Array.isArray(content)) {
|
||||
return null;
|
||||
}
|
||||
|
||||
/** @type {({ token: string, logprob: number }) => [string, number]} */
|
||||
const toTuple = (x) => [x.token, x.logprob];
|
||||
|
||||
return content.map(({ token, logprob, top_logprobs }) => {
|
||||
// Add the chosen token to top_logprobs if it's not already there, then
|
||||
// convert to a list of [token, logprob] pairs
|
||||
const chosenTopToken = top_logprobs.some((top) => token === top.token);
|
||||
const topLogprobs = chosenTopToken
|
||||
? top_logprobs.map(toTuple)
|
||||
: [...top_logprobs.map(toTuple), [token, logprob]];
|
||||
return { token, topLogprobs };
|
||||
});
|
||||
}
|
||||
|
||||
/**
|
||||
* parseOpenAITextLogprobs receives a `logprobs` response from OpenAI's text
|
||||
* completion API and converts into the structure used by the Token Probabilities
|
||||
* view.
|
||||
* @param {{tokens: string[], token_logprobs: number[], top_logprobs: { token: string, logprob: number }[][]}} logprobs
|
||||
* @returns {import('logprobs.js').TokenLogprobs[] | null} converted logprobs
|
||||
*/
|
||||
function parseOpenAITextLogprobs(logprobs) {
|
||||
const { tokens, token_logprobs, top_logprobs } = logprobs ?? {};
|
||||
|
||||
if (!Array.isArray(tokens)) {
|
||||
return null;
|
||||
}
|
||||
|
||||
return tokens.map((token, i) => {
|
||||
// Add the chosen token to top_logprobs if it's not already there, then
|
||||
// convert to a list of [token, logprob] pairs
|
||||
const topLogprobs = top_logprobs[i] ? Object.entries(top_logprobs[i]) : [];
|
||||
const chosenTopToken = topLogprobs.some(([topToken]) => token === topToken);
|
||||
if (!chosenTopToken) {
|
||||
topLogprobs.push([token, token_logprobs[i]]);
|
||||
}
|
||||
return { token, topLogprobs };
|
||||
});
|
||||
}
|
||||
|
||||
|
||||
function handleWindowError(err) {
|
||||
const text = parseWindowError(err);
|
||||
toastr.error(text, 'Window.ai returned an error');
|
||||
|
||||
@ -164,6 +164,7 @@ let power_user = {
|
||||
auto_fix_generated_markdown: true,
|
||||
send_on_enter: send_on_enter_options.AUTO,
|
||||
console_log_prompts: false,
|
||||
request_token_probabilities: false,
|
||||
render_formulas: false,
|
||||
allow_name1_display: false,
|
||||
allow_name2_display: false,
|
||||
@ -1453,6 +1454,7 @@ function loadPowerUserSettings(settings, data) {
|
||||
$(`#example_messages_behavior option[value="${getExampleMessagesBehavior()}"]`).prop('selected', true);
|
||||
|
||||
$('#console_log_prompts').prop('checked', power_user.console_log_prompts);
|
||||
$('#request_token_probabilities').prop('checked', power_user.request_token_probabilities);
|
||||
$('#auto_fix_generated_markdown').prop('checked', power_user.auto_fix_generated_markdown);
|
||||
$('#auto_scroll_chat_to_bottom').prop('checked', power_user.auto_scroll_chat_to_bottom);
|
||||
$('#bogus_folders').prop('checked', power_user.bogus_folders);
|
||||
@ -2952,6 +2954,11 @@ $(document).ready(() => {
|
||||
saveSettingsDebounced();
|
||||
});
|
||||
|
||||
$('#request_token_probabilities').on('input', function () {
|
||||
power_user.request_token_probabilities = !!$(this).prop('checked');
|
||||
saveSettingsDebounced();
|
||||
});
|
||||
|
||||
$('#auto_scroll_chat_to_bottom').on('input', function () {
|
||||
power_user.auto_scroll_chat_to_bottom = !!$(this).prop('checked');
|
||||
saveSettingsDebounced();
|
||||
|
||||
@ -354,8 +354,8 @@ function trimTokensCallback(arg, value) {
|
||||
}
|
||||
|
||||
const sliceTokens = direction === 'start' ? textTokens.slice(0, limit) : textTokens.slice(-limit);
|
||||
const decodedText = decodeTextTokens(tokenizerId, sliceTokens);
|
||||
return decodedText;
|
||||
const { text } = decodeTextTokens(tokenizerId, sliceTokens);
|
||||
return text;
|
||||
} catch (error) {
|
||||
console.warn('WARN: Tokenization failed for /trimtokens command, returning original', error);
|
||||
return value;
|
||||
|
||||
@ -10,10 +10,7 @@ import {
|
||||
} from '../script.js';
|
||||
import { BIAS_CACHE, createNewLogitBiasEntry, displayLogitBias, getLogitBiasListResult } from './logit-bias.js';
|
||||
|
||||
import {
|
||||
power_user,
|
||||
registerDebugFunction,
|
||||
} from './power-user.js';
|
||||
import { power_user, registerDebugFunction } from './power-user.js';
|
||||
import EventSourceStream from './sse-stream.js';
|
||||
import { SENTENCEPIECE_TOKENIZERS, TEXTGEN_TOKENIZERS, getTextTokens, tokenizers } from './tokenizers.js';
|
||||
import { getSortableDelay, onlyUnique } from './utils.js';
|
||||
@ -675,6 +672,8 @@ async function generateTextGenWithStreaming(generate_data, signal) {
|
||||
|
||||
return async function* streamData() {
|
||||
let text = '';
|
||||
/** @type {import('logprobs.js').TokenLogprobs | null} */
|
||||
let logprobs = null;
|
||||
const swipes = [];
|
||||
while (true) {
|
||||
const { done, value } = await reader.read();
|
||||
@ -689,14 +688,44 @@ async function generateTextGenWithStreaming(generate_data, signal) {
|
||||
const swipeIndex = data.choices[0].index - 1;
|
||||
swipes[swipeIndex] = (swipes[swipeIndex] || '') + data.choices[0].text;
|
||||
} else {
|
||||
text += data?.choices?.[0]?.text || data?.content || '';
|
||||
const newText = data?.choices?.[0]?.text || data?.content || '';
|
||||
text += newText;
|
||||
logprobs = parseTextgenLogprobs(newText, data.choices[0]?.logprobs);
|
||||
}
|
||||
|
||||
yield { text, swipes };
|
||||
yield { text, swipes, logprobs };
|
||||
}
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* parseTextgenLogprobs converts a logprobs object returned from a textgen API
|
||||
* for a single token into a TokenLogprobs object used by the Token
|
||||
* Probabilities feature.
|
||||
* @param {string} token - the text of the token that the logprobs are for
|
||||
* @param {Object} logprobs - logprobs object returned from the API
|
||||
* @returns {import('logprobs.js').TokenLogprobs | null} - converted logprobs
|
||||
*/
|
||||
function parseTextgenLogprobs(token, logprobs) {
|
||||
if (!logprobs) {
|
||||
return null;
|
||||
}
|
||||
|
||||
switch (settings.type) {
|
||||
case OOBA: {
|
||||
/** @type {Record<string, number>[]} */
|
||||
const topLogprobs = logprobs.top_logprobs;
|
||||
if (!topLogprobs?.length) {
|
||||
return null;
|
||||
}
|
||||
const candidates = Object.entries(topLogprobs[0]);
|
||||
return { token, topLogprobs: candidates };
|
||||
}
|
||||
default:
|
||||
return null;
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Parses errors in streaming responses and displays them in toastr.
|
||||
* @param {Response} response - Response from the server.
|
||||
@ -769,6 +798,7 @@ export function getTextGenGenerationData(finalPrompt, maxTokens, isImpersonate,
|
||||
'model': getModel(),
|
||||
'max_new_tokens': maxTokens,
|
||||
'max_tokens': maxTokens,
|
||||
'logprobs': power_user.request_token_probabilities ? 10: undefined,
|
||||
'temperature': settings.dynatemp ? (settings.min_temp + settings.max_temp) / 2 : settings.temp,
|
||||
'top_p': settings.top_p,
|
||||
'typical_p': settings.typical_p,
|
||||
|
||||
@ -669,9 +669,11 @@ function getTextTokensFromKoboldAPI(str) {
|
||||
* Calls the underlying tokenizer model to decode token ids to text.
|
||||
* @param {string} endpoint API endpoint.
|
||||
* @param {number[]} ids Array of token ids
|
||||
* @returns {({ text: string, chunks?: string[] })} Decoded token text as a single string and individual chunks (if available).
|
||||
*/
|
||||
function decodeTextTokensFromServer(endpoint, ids) {
|
||||
let text = '';
|
||||
let chunks = [];
|
||||
jQuery.ajax({
|
||||
async: false,
|
||||
type: 'POST',
|
||||
@ -681,9 +683,10 @@ function decodeTextTokensFromServer(endpoint, ids) {
|
||||
contentType: 'application/json',
|
||||
success: function (data) {
|
||||
text = data.text;
|
||||
chunks = data.chunks;
|
||||
},
|
||||
});
|
||||
return text;
|
||||
return { text, chunks };
|
||||
}
|
||||
|
||||
/**
|
||||
@ -725,6 +728,7 @@ export function getTextTokens(tokenizerType, str) {
|
||||
* Decodes token ids to text using the server API.
|
||||
* @param {number} tokenizerType Tokenizer type.
|
||||
* @param {number[]} ids Array of token ids
|
||||
* @returns {({ text: string, chunks?: string[] })} Decoded token text as a single string and individual chunks (if available).
|
||||
*/
|
||||
export function decodeTextTokens(tokenizerType, ids) {
|
||||
// Currently, neither remote API can decode, but this may change in the future. Put this guard here to be safe
|
||||
@ -734,12 +738,12 @@ export function decodeTextTokens(tokenizerType, ids) {
|
||||
const tokenizerEndpoints = TOKENIZER_URLS[tokenizerType];
|
||||
if (!tokenizerEndpoints) {
|
||||
console.warn('Unknown tokenizer type', tokenizerType);
|
||||
return [];
|
||||
return { text: '', chunks: [] };
|
||||
}
|
||||
let endpointUrl = tokenizerEndpoints.decode;
|
||||
if (!endpointUrl) {
|
||||
console.warn('This tokenizer type does not support decoding', tokenizerType);
|
||||
return [];
|
||||
return { text: '', chunks: [] };
|
||||
}
|
||||
if (tokenizerType === tokenizers.OPENAI) {
|
||||
endpointUrl += `?model=${getTokenizerModel()}`;
|
||||
|
||||
Reference in New Issue
Block a user