rixty/SillyTavern
1
0
Fork 0
mirror of https://github.com/SillyTavern/SillyTavern.git synced 2025-06-05 21:59:27 +02:00
Code Issues Packages Projects Releases Wiki Activity

Go team dropdown

Browse Source
This commit is contained in:
Cohee
2025-04-23 00:38:28 +03:00
parent a95056db40
commit bee3cee740
5 changed files with 117 additions and 105 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
public/css/toggle-dependent.css
View File

@ -498,7 +498,3 @@ label[for="trim_spaces"]:not(:has(input:checked)) small {
#banned_tokens_block_ooba:not(:has(#send_banned_tokens_textgenerationwebui:checked)) #banned_tokens_controls_ooba {
filter: brightness(0.5);
}
#thinking_budget_controls:not(:has(#enable_thinking:checked)) .range-block:has(#thinking_budget) {
filter: brightness(0.5);
}

35
public/index.html
View File

@ -2034,32 +2034,6 @@
</span>
</div>
</div>
<div id="thinking_budget_container" data-source="makersuite">
<div id="thinking_budget_controls">
<div class="range-block">
<label for="enable_thinking" class="checkbox_label widthFreeExpand">
<input id="enable_thinking" type="checkbox" />
<span data-i18n="Enable thinking">Enable thinking</span>
</label>
<div class="toggle-description justifyLeft marginBot5" data-i18n="thinking_budget_desc">
Must be enabled for Thinking Budget to take effect.
</div>
</div>
<div class="range-block">
<div class="range-block-title" data-i18n="Thinking Budget (Tokens)">
Thinking Budget (tokens)
</div>
<div class="range-block-range-and-counter">
<div class="range-block-range">
<input id="thinking_budget" type="range" min="0" max="24576" step="1">
</div>
<div class="range-block-counter">
<input id="thinking_budget_counter" type="number" min="0" max="24576" step="1" data-for="thinking_budget">
</div>
</div>
</div>
</div>
</div>
<div class="range-block" data-source="deepseek,openrouter,custom,claude,xai">
<label for="openai_show_thoughts" class="checkbox_label widthFreeExpand">
<input id="openai_show_thoughts" type="checkbox" />
@ -2074,16 +2048,19 @@
</span>
</div>
</div>
<div class="flex-container flexFlowColumn wide100p textAlignCenter marginTop10" data-source="openai,custom,claude,xai">
<div class="flex-container oneline-dropdown" title="Constrains effort on reasoning for reasoning models.&#10;Currently supported values are low, medium, and high.&#10;Reducing reasoning effort can result in faster responses and fewer tokens used on reasoning in a response." data-i18n="[title]Constrains effort on reasoning for reasoning models.">
<div class="flex-container flexFlowColumn wide100p textAlignCenter marginTop10" data-source="openai,custom,claude,xai,makersuite">
<div class="flex-container oneline-dropdown" title="Constrains effort on reasoning for reasoning models.&#10;Reducing reasoning effort can result in faster responses and fewer tokens used on reasoning in a response." data-i18n="[title]Constrains effort on reasoning for reasoning models.">
<label for="openai_reasoning_effort">
<span data-i18n="Reasoning Effort">Reasoning Effort</span>
<i data-source="claude" class="opacity50p fa-solid fa-circle-info" title="Allocates a portion of the response length for thinking (low: 10%, medium: 25%, high: 50%), but minimum 1024 tokens."></i>
<i data-source="claude,makersuite" class="opacity50p fa-solid fa-circle-info" title="Allocates a portion of the response length for thinking (low: 10%, medium: 25%, high: 50%), but minimum 1024 tokens."></i>
</label>
<select id="openai_reasoning_effort">
<option data-i18n="openai_reasoning_effort_auto" value="auto">Unset (model-default)</option>
<option data-i18n="openai_reasoning_effort_minimum" value="min">Mininum (model-defined)</option>
<option data-i18n="openai_reasoning_effort_low" value="low">Low</option>
<option data-i18n="openai_reasoning_effort_medium" value="medium">Medium</option>
<option data-i18n="openai_reasoning_effort_high" value="high">High</option>
<option data-i18n="openai_reasoning_effort_maximum" value="max">Maximum (model-defined)</option>
</select>
</div>
</div>

102
public/scripts/openai.js
View File

@ -165,11 +165,6 @@ const textCompletionModels = [
'code-search-ada-code-001',
];
// One more models list to maintain, yay
const thinkingBudgetModels = [
'gemini-2.5-flash-preview-04-17',
];
let biasCache = undefined;
export let model_list = [];
@ -221,6 +216,15 @@ const openrouter_middleout_types = {
OFF: 'off',
};
export const reasoning_effort_types = {
auto: 'auto',
low: 'low',
medium: 'medium',
high: 'high',
min: 'min',
max: 'max',
};
const sensitiveFields = [
'reverse_proxy',
'proxy_password',
@ -311,8 +315,6 @@ export const settingsToUpdate = {
n: ['#n_openai', 'n', false],
bypass_status_check: ['#openai_bypass_status_check', 'bypass_status_check', true],
request_images: ['#openai_request_images', 'request_images', true],
enable_thinking: ['#enable_thinking', 'enable_thinking', true],
thinking_budget: ['#thinking_budget', 'thinking_budget', false],
};
const default_settings = {
@ -389,11 +391,9 @@ const default_settings = {
continue_postfix: continue_postfix_types.SPACE,
custom_prompt_post_processing: custom_prompt_post_processing_types.NONE,
show_thoughts: true,
reasoning_effort: 'medium',
reasoning_effort: reasoning_effort_types.medium,
enable_web_search: false,
request_images: false,
enable_thinking: true,
thinking_budget: 1000,
seed: -1,
n: 1,
};
@ -472,11 +472,9 @@ const oai_settings = {
continue_postfix: continue_postfix_types.SPACE,
custom_prompt_post_processing: custom_prompt_post_processing_types.NONE,
show_thoughts: true,
reasoning_effort: 'medium',
reasoning_effort: reasoning_effort_types.medium,
enable_web_search: false,
request_images: false,
enable_thinking: true,
thinking_budget: 1000,
seed: -1,
n: 1,
};
@ -1948,6 +1946,35 @@ async function sendAltScaleRequest(messages, logit_bias, signal, type) {
return data.output;
}
function getReasoningEffort() {
// Do not set the field. Let the model decide.
if (oai_settings.reasoning_effort === reasoning_effort_types.auto) {
return undefined;
}
// These sources require effort as a string
if (oai_settings.reasoning_effort === reasoning_effort_types.min) {
switch (oai_settings.chat_completion_source) {
case chat_completion_sources.OPENAI:
case chat_completion_sources.CUSTOM:
case chat_completion_sources.XAI:
return reasoning_effort_types.low;
}
}
// Same here, but max effort
if (oai_settings.reasoning_effort === reasoning_effort_types.max) {
switch (oai_settings.chat_completion_source) {
case chat_completion_sources.OPENAI:
case chat_completion_sources.CUSTOM:
case chat_completion_sources.XAI:
return reasoning_effort_types.high;
}
}
return oai_settings.reasoning_effort;
}
/**
* Send a chat completion request to backend
* @param {string} type (impersonate, quiet, continue, etc)
@ -2034,17 +2061,12 @@ async function sendOpenAIRequest(type, messages, signal) {
'char_name': name2,
'group_names': getGroupNames(),
'include_reasoning': Boolean(oai_settings.show_thoughts),
'reasoning_effort': String(oai_settings.reasoning_effort),
'reasoning_effort': getReasoningEffort(),
'enable_web_search': Boolean(oai_settings.enable_web_search),
'request_images': Boolean(oai_settings.request_images),
'custom_prompt_post_processing': oai_settings.custom_prompt_post_processing,
};
if (thinkingBudgetModels.includes(model)) {
generate_data['enable_thinking'] = oai_settings.enable_thinking;
generate_data['thinking_budget'] = oai_settings.thinking_budget;
}
if (!canMultiSwipe && ToolManager.canPerformToolCalls(type)) {
await ToolManager.registerFunctionToolsOpenAI(generate_data);
}
@ -3300,8 +3322,6 @@ function loadOpenAISettings(data, settings) {
oai_settings.show_thoughts = settings.show_thoughts ?? default_settings.show_thoughts;
oai_settings.reasoning_effort = settings.reasoning_effort ?? default_settings.reasoning_effort;
oai_settings.enable_web_search = settings.enable_web_search ?? default_settings.enable_web_search;
oai_settings.enable_thinking = settings.enable_thinking ?? default_settings.enable_thinking;
oai_settings.thinking_budget = settings.thinking_budget ?? default_settings.thinking_budget;
oai_settings.request_images = settings.request_images ?? default_settings.request_images;
oai_settings.seed = settings.seed ?? default_settings.seed;
oai_settings.n = settings.n ?? default_settings.n;
@ -3437,10 +3457,6 @@ function loadOpenAISettings(data, settings) {
$('#openai_reasoning_effort').val(oai_settings.reasoning_effort);
$(`#openai_reasoning_effort option[value="${oai_settings.reasoning_effort}"]`).prop('selected', true);
$('#enable_thinking').prop('checked', oai_settings.enable_thinking);
$('#thinking_budget').val(oai_settings.thinking_budget);
$('#thinking_budget_counter').val(oai_settings.thinking_budget);
if (settings.reverse_proxy !== undefined) oai_settings.reverse_proxy = settings.reverse_proxy;
$('#openai_reverse_proxy').val(oai_settings.reverse_proxy);
@ -3471,7 +3487,6 @@ function loadOpenAISettings(data, settings) {
setNamesBehaviorControls();
setContinuePostfixControls();
updateThinkingBudgetUI();
if (oai_settings.custom_prompt_post_processing === custom_prompt_post_processing_types.CLAUDE) {
oai_settings.custom_prompt_post_processing = custom_prompt_post_processing_types.MERGE;
@ -3529,14 +3544,6 @@ function setContinuePostfixControls() {
$('#continue_postfix_display').text(checkedItemText);
}
/**
* Updates the visibility and state of the Thinking Budget controls.
*/
function updateThinkingBudgetUI() {
const modelSupportsControl = thinkingBudgetModels.includes(getChatCompletionModel());
$('#thinking_budget_controls').toggle(modelSupportsControl);
}
async function getStatusOpen() {
if (oai_settings.chat_completion_source == chat_completion_sources.WINDOWAI) {
let status;
@ -3717,8 +3724,6 @@ async function saveOpenAIPreset(name, settings, triggerUi = true) {
reasoning_effort: settings.reasoning_effort,
enable_web_search: settings.enable_web_search,
request_images: settings.request_images,
enable_thinking: settings.enable_thinking,
thinking_budget: settings.thinking_budget,
seed: settings.seed,
n: settings.n,
};
@ -4749,7 +4754,6 @@ async function onModelChange() {
$('#openai_max_context_counter').attr('max', Number($('#openai_max_context').attr('max')));
updateThinkingBudgetUI();
saveSettingsDebounced();
eventSource.emit(event_types.CHATCOMPLETION_MODEL_CHANGED, value);
}
@ -5568,7 +5572,6 @@ export function initOpenAI() {
$('#chat_completion_source').on('change', function () {
oai_settings.chat_completion_source = String($(this).find(':selected').val());
toggleChatCompletionForms();
updateThinkingBudgetUI();
saveSettingsDebounced();
reconnectOpenAi();
forceCharacterEditorTokenize();
@ -5754,29 +5757,6 @@ export function initOpenAI() {
saveSettingsDebounced();
});
$('#enable_thinking').on('input', function () {
oai_settings.enable_thinking = !!$(this).prop('checked');
updateThinkingBudgetUI();
saveSettingsDebounced();
});
$('#thinking_budget').on('input', function () {
oai_settings.thinking_budget = Number($(this).val());
$('#thinking_budget_counter').val(oai_settings.thinking_budget);
saveSettingsDebounced();
});
$('#thinking_budget_counter').on('input', function () {
let value = Number($(this).val());
const min = Number($('#thinking_budget').attr('min'));
const max = Number($('#thinking_budget').attr('max'));
value = Math.max(min, Math.min(max, value));
$(this).val(value);
oai_settings.thinking_budget = value;
$('#thinking_budget').val(value);
saveSettingsDebounced();
});
$('#openai_enable_web_search').on('input', function () {
oai_settings.enable_web_search = !!$(this).prop('checked');
calculateOpenRouterCost();

18
src/endpoints/backends/chat-completions.js
View File

@ -28,7 +28,8 @@ import {
cachingAtDepthForOpenRouterClaude,
cachingAtDepthForClaude,
getPromptNames,
calculateBudgetTokens,
calculateClaudeBudgetTokens,
calculateGoogleBudgetTokens,
} from '../../prompt-converters.js';
import { readSecret, SECRET_KEYS } from '../secrets.js';
@ -202,7 +203,7 @@ async function sendClaudeRequest(request, response) {
// No prefill when thinking
voidPrefill = true;
const reasoningEffort = request.body.reasoning_effort;
const budgetTokens = calculateBudgetTokens(requestBody.max_tokens, reasoningEffort, requestBody.stream);
const budgetTokens = calculateClaudeBudgetTokens(requestBody.max_tokens, reasoningEffort, requestBody.stream);
const minThinkTokens = 1024;
if (requestBody.max_tokens <= minThinkTokens) {
const newValue = requestBody.max_tokens + minThinkTokens;
@ -340,6 +341,7 @@ async function sendMakerSuiteRequest(request, response) {
const stream = Boolean(request.body.stream);
const enableWebSearch = Boolean(request.body.enable_web_search);
const requestImages = Boolean(request.body.request_images);
const reasoningEffort = String(request.body.reasoning_effort);
const isThinking = model.includes('thinking');
const isGemma = model.includes('gemma');
@ -412,11 +414,15 @@ async function sendMakerSuiteRequest(request, response) {
tools.push({ function_declarations: functionDeclarations });
}
if ('enable_thinking' in request.body && 'thinking_budget' in request.body) {
const thinkingEnabled = Boolean(request.body.enable_thinking);
const thinkingBudget = Number(request.body.thinking_budget);
// One more models list to maintain, yay
const thinkingBudgetModels = [
'gemini-2.5-flash-preview-04-17',
];
if (thinkingEnabled) {
if (thinkingBudgetModels.includes(model)) {
const thinkingBudget = calculateGoogleBudgetTokens(generationConfig.maxOutputTokens, reasoningEffort);
if (Number.isInteger(thinkingBudget)) {
generationConfig.thinkingConfig = { thinkingBudget: thinkingBudget };
}
}

63
src/prompt-converters.js
View File

@ -3,6 +3,15 @@ import { getConfigValue, tryParse } from './util.js';
const PROMPT_PLACEHOLDER = getConfigValue('promptPlaceholder', 'Let\'s get started.');
const REASONING_EFFORT = {
auto: 'auto',
low: 'low',
medium: 'medium',
high: 'high',
min: 'min',
max: 'max',
};
/**
* @typedef {object} PromptNames
* @property {string} charName Character name
@ -944,25 +953,35 @@ export function cachingAtDepthForOpenRouterClaude(messages, cachingAtDepth) {
}
/**
* Calculate the budget tokens for a given reasoning effort.
* Calculate the Claude budget tokens for a given reasoning effort.
* @param {number} maxTokens Maximum tokens
* @param {string} reasoningEffort Reasoning effort
* @param {boolean} stream If streaming is enabled
* @returns {number} Budget tokens
*/
export function calculateBudgetTokens(maxTokens, reasoningEffort, stream) {
export function calculateClaudeBudgetTokens(maxTokens, reasoningEffort, stream) {
let budgetTokens = 0;
switch (reasoningEffort) {
case 'low':
// Claude doesn't have a default budget value. Use same as min.
case REASONING_EFFORT.auto:
budgetTokens = 1024;
break;
case REASONING_EFFORT.min:
budgetTokens = 1024;
break;
case REASONING_EFFORT.low:
budgetTokens = Math.floor(maxTokens * 0.1);
break;
case 'medium':
case REASONING_EFFORT.medium:
budgetTokens = Math.floor(maxTokens * 0.25);
break;
case 'high':
case REASONING_EFFORT.high:
budgetTokens = Math.floor(maxTokens * 0.5);
break;
case REASONING_EFFORT.max:
budgetTokens = maxTokens;
break;
}
budgetTokens = Math.max(budgetTokens, 1024);
@ -973,3 +992,37 @@ export function calculateBudgetTokens(maxTokens, reasoningEffort, stream) {
return budgetTokens;
}
/**
* Calculate the Google budget tokens for a given reasoning effort.
* @param {number} maxTokens Maximum tokens
* @param {string} reasoningEffort Reasoning effort
* @returns {number?} Budget tokens
*/
export function calculateGoogleBudgetTokens(maxTokens, reasoningEffort) {
let budgetTokens = 0;
switch (reasoningEffort) {
case REASONING_EFFORT.auto:
return null;
case REASONING_EFFORT.min:
budgetTokens = 0;
break;
case REASONING_EFFORT.low:
budgetTokens = Math.floor(maxTokens * 0.1);
break;
case REASONING_EFFORT.medium:
budgetTokens = Math.floor(maxTokens * 0.25);
break;
case REASONING_EFFORT.high:
budgetTokens = Math.floor(maxTokens * 0.5);
break;
case REASONING_EFFORT.max:
budgetTokens = maxTokens;
break;
}
budgetTokens = Math.min(budgetTokens, 24576);
return budgetTokens;
}