Merge pull request #996 from kingbased/scale

Alternative method for scale generations
This commit is contained in:
Cohee 2023-08-22 16:41:07 +03:00 committed by GitHub
commit 0ffb3a8dda
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
6 changed files with 183 additions and 20 deletions

View File

@ -635,7 +635,7 @@
Max prompt cost: <span id="openrouter_max_prompt_cost">Unknown</span>
</div>
<hr>
<div class="range-block" data-source="openai,claude,windowai,openrouter,ai21">
<div class="range-block" data-source="openai,claude,windowai,openrouter,ai21,scale">
<div class="range-block-title" data-i18n="Temperature">
Temperature
</div>
@ -710,7 +710,7 @@
</div>
</div>
</div>
<div class="range-block" data-source="openai,claude,openrouter,ai21">
<div class="range-block" data-source="openai,claude,openrouter,ai21,scale">
<div class="range-block-title" data-i18n="Top-p">
Top P
</div>
@ -1578,7 +1578,7 @@
</div>
</div>
</div>
<div class="range-block m-t-1" data-source="openai,openrouter">
<div class="range-block m-t-1" data-source="openai,openrouter,scale">
<div class="range-block-title openai_restorable" data-i18n="Logit Bias">
Logit Bias
</div>
@ -1976,18 +1976,34 @@
</form>
<form id="scale_form" data-source="scale" action="javascript:void(null);" method="post" enctype="multipart/form-data">
<h4>Scale API Key</h4>
<div class="flex-container">
<input id="api_key_scale" name="api_key_scale" class="text_pole flex1" maxlength="500" value="" autocomplete="off">
<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_scale"></div>
<div id="normal_scale_form">
<h4>Scale API Key</h4>
<div class="flex-container">
<input id="api_key_scale" name="api_key_scale" class="text_pole flex1" maxlength="500" value="" autocomplete="off">
<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_scale"></div>
</div>
<div data-for="api_key_scale" class="neutral_warning">
For privacy reasons, your API key will be hidden after you reload the page.
</div>
<h4>Scale API URL</h4>
<input id="api_url_scale" name="api_url_scale" class="text_pole" maxlength="500" value="" autocomplete="off" placeholder="https://dashboard.scale.com/spellbook/api/v2/deploy/xxxxxxx">
</div>
<div data-for="api_key_scale" class="neutral_warning">
For privacy reasons, your API key will be hidden after you reload the page.
<div id="alt_scale_form">
<h4>Scale Cookie (_jwt)</h4>
<div class="flex-container">
<input id="scale_cookie" name="scale_cookie" class="text_pole flex1" maxlength="500" value="" autocomplete="off">
<div title="Clear your cookie" data-i18n="[title]Clear your cookie" class="menu_button fa-solid fa-circle-xmark clear-api-key" data-key="scale_cookie"></div>
</div>
<div data-for="scale_cookie" class="neutral_warning">
For privacy reasons, your cookie will be hidden after you reload the page.
</div>
</div>
<h4>Scale API URL</h4>
<input id="api_url_scale" name="api_url_scale" class="text_pole" maxlength="500" value="" autocomplete="off" placeholder="https://dashboard.scale.com/spellbook/api/v2/deploy/xxxxxxx">
<!-- Its only purpose is to trigger max context size check -->
<select id="model_scale_select" class="displayNone"></select>
<label for="scale-alt" class="checkbox_label">
<input id="scale-alt" type="checkbox" checked>
<span data-i18n="Alt Method">Alt Method</span>
</label>
</form>
<form id="ai21_form" data-source="ai21" action="javascript:void(null);" method="post" enctype="multipart/form-data">

View File

@ -365,7 +365,7 @@ function RA_autoconnect(PrevApi) {
case 'openai':
if (((secret_state[SECRET_KEYS.OPENAI] || oai_settings.reverse_proxy) && oai_settings.chat_completion_source == chat_completion_sources.OPENAI)
|| ((secret_state[SECRET_KEYS.CLAUDE] || oai_settings.reverse_proxy) && oai_settings.chat_completion_source == chat_completion_sources.CLAUDE)
|| (secret_state[SECRET_KEYS.SCALE] && oai_settings.chat_completion_source == chat_completion_sources.SCALE)
|| ((secret_state[SECRET_KEYS.SCALE] || secret_state[SECRET_KEYS.SCALE_COOKIE]) && oai_settings.chat_completion_source == chat_completion_sources.SCALE)
|| (oai_settings.chat_completion_source == chat_completion_sources.WINDOWAI)
|| (secret_state[SECRET_KEYS.OPENROUTER] && oai_settings.chat_completion_source == chat_completion_sources.OPENROUTER)
|| (secret_state[SECRET_KEYS.AI21] && oai_settings.chat_completion_source == chat_completion_sources.AI21)

View File

@ -1386,7 +1386,7 @@ export async function createNewGroupChat(groupId) {
group.chat_metadata = {};
updateChatMetadata(group.chat_metadata, true);
await editGroup(group.id, true);
await editGroup(group.id, true, false);
await getGroupChat(group.id);
}

View File

@ -109,8 +109,8 @@ const max_4k = 4095;
const max_8k = 8191;
const max_16k = 16383;
const max_32k = 32767;
const scale_max = 7900; // Probably more. Save some for the system prompt defined on Scale site.
const claude_max = 8000; // We have a proper tokenizer, so theoretically could be larger (up to 9k)
const scale_max = 8191;
const claude_max = 9000; // We have a proper tokenizer, so theoretically could be larger (up to 9k)
const palm2_max = 7500; // The real context window is 8192, spare some for padding due to using turbo tokenizer
const claude_100k_max = 99000;
let ai21_max = 9200; //can easily fit 9k gpt tokens because j2's tokenizer is efficient af
@ -219,6 +219,7 @@ const default_settings = {
assistant_prefill: '',
use_ai21_tokenizer: false,
exclude_assistant: false,
use_alt_scale: false,
};
const oai_settings = {
@ -261,6 +262,7 @@ const oai_settings = {
assistant_prefill: '',
use_ai21_tokenizer: false,
exclude_assistant: false,
use_alt_scale: false,
};
let openai_setting_names;
@ -1082,6 +1084,47 @@ function saveModelList(data) {
}
}
async function sendAltScaleRequest(openai_msgs_tosend, logit_bias, signal) {
const generate_url = '/generate_altscale';
let firstSysMsgs = []
for(let msg of openai_msgs_tosend){
if(msg.role === 'system') {
firstSysMsgs.push(substituteParams(msg.name ? msg.name + ": " + msg.content : msg.content));
} else {
break;
}
}
let subsequentMsgs = openai_msgs_tosend.slice(firstSysMsgs.length);
const joinedSysMsgs = substituteParams(firstSysMsgs.join("\n"));
const joinedSubsequentMsgs = subsequentMsgs.reduce((acc, obj) => {
return acc + obj.role + ": " + obj.content + "\n";
}, "");
openai_msgs_tosend = substituteParams(joinedSubsequentMsgs);
const generate_data = {
sysprompt: joinedSysMsgs,
prompt: openai_msgs_tosend,
temp: parseFloat(oai_settings.temp_openai),
top_p: parseFloat(oai_settings.top_p_openai),
max_tokens: parseFloat(oai_settings.openai_max_tokens),
logit_bias: logit_bias,
}
const response = await fetch(generate_url, {
method: 'POST',
body: JSON.stringify(generate_data),
headers: getRequestHeaders(),
signal: signal
});
const data = await response.json();
return data.output;
}
async function sendOpenAIRequest(type, openai_msgs_tosend, signal) {
// Provide default abort signal
if (!signal) {
@ -1118,7 +1161,7 @@ async function sendOpenAIRequest(type, openai_msgs_tosend, signal) {
return sendWindowAIRequest(openai_msgs_tosend, signal, stream);
}
const logitBiasSources = [chat_completion_sources.OPENAI, chat_completion_sources.OPENROUTER];
const logitBiasSources = [chat_completion_sources.OPENAI, chat_completion_sources.OPENROUTER, chat_completion_sources.SCALE];
if (oai_settings.bias_preset_selected
&& logitBiasSources.includes(oai_settings.chat_completion_source)
&& Array.isArray(oai_settings.bias_presets[oai_settings.bias_preset_selected])
@ -1127,6 +1170,10 @@ async function sendOpenAIRequest(type, openai_msgs_tosend, signal) {
biasCache = logit_bias;
}
if (isScale && oai_settings.use_alt_scale) {
return sendAltScaleRequest(openai_msgs_tosend, logit_bias, signal)
}
const model = getChatCompletionModel();
const generate_data = {
"messages": openai_msgs_tosend,
@ -1971,6 +2018,7 @@ function loadOpenAISettings(data, settings) {
if (settings.openai_model !== undefined) oai_settings.openai_model = settings.openai_model;
if (settings.use_ai21_tokenizer !== undefined) { oai_settings.use_ai21_tokenizer = !!settings.use_ai21_tokenizer; oai_settings.use_ai21_tokenizer ? ai21_max = 8191 : ai21_max = 9200; }
if (settings.exclude_assistant !== undefined) oai_settings.exclude_assistant = !!settings.exclude_assistant;
if (settings.use_alt_scale !== undefined) { oai_settings.use_alt_scale = !!settings.use_alt_scale; updateScaleForm(); }
$('#stream_toggle').prop('checked', oai_settings.stream_openai);
$('#api_url_scale').val(oai_settings.api_url_scale);
$('#openai_proxy_password').val(oai_settings.proxy_password);
@ -2001,6 +2049,7 @@ function loadOpenAISettings(data, settings) {
$('#openai_external_category').toggle(oai_settings.show_external_models);
$('#use_ai21_tokenizer').prop('checked', oai_settings.use_ai21_tokenizer);
$('#exclude_assistant').prop('checked', oai_settings.exclude_assistant);
$('#scale-alt').prop('checked', oai_settings.use_alt_scale);
if (settings.impersonation_prompt !== undefined) oai_settings.impersonation_prompt = settings.impersonation_prompt;
$('#impersonation_prompt_textarea').val(oai_settings.impersonation_prompt);
@ -2199,6 +2248,7 @@ async function saveOpenAIPreset(name, settings, triggerUi = true) {
assistant_prefill: settings.assistant_prefill,
use_ai21_tokenizer: settings.use_ai21_tokenizer,
exclude_assistant: settings.exclude_assistant,
use_alt_scale: settings.use_alt_scale,
};
const savePresetSettings = await fetch(`/savepreset_openai?name=${name}`, {
@ -2536,7 +2586,8 @@ function onSettingsPresetChange() {
assistant_prefill: ['#claude_assistant_prefill', 'assistant_prefill', false],
use_ai21_tokenizer: ['#use_ai21_tokenizer', 'use_ai21_tokenizer', false],
exclude_assistant: ['#exclude_assistant', 'exclude_assistant', false],
};
use_alt_scale: ['#use_alt_scale', 'use_alt_scale', false],
};
const presetName = $('#settings_perset_openai').find(":selected").text();
oai_settings.preset_settings_openai = presetName;
@ -2831,20 +2882,31 @@ async function onConnectButtonClick(e) {
if (oai_settings.chat_completion_source == chat_completion_sources.SCALE) {
const api_key_scale = $('#api_key_scale').val().trim();
const scale_cookie = $('#scale_cookie').val().trim();
if (api_key_scale.length) {
await writeSecret(SECRET_KEYS.SCALE, api_key_scale);
}
if (!oai_settings.api_url_scale) {
if (scale_cookie.length) {
await writeSecret(SECRET_KEYS.SCALE_COOKIE, scale_cookie);
}
if (!oai_settings.api_url_scale && !oai_settings.use_alt_scale) {
console.log('No API URL saved for Scale');
return;
}
if (!secret_state[SECRET_KEYS.SCALE]) {
if (!secret_state[SECRET_KEYS.SCALE] && !oai_settings.use_alt_scale) {
console.log('No secret key saved for Scale');
return;
}
if (!secret_state[SECRET_KEYS.SCALE_COOKIE] && oai_settings.use_alt_scale) {
console.log("No cookie set for Scale");
return;
}
}
if (oai_settings.chat_completion_source == chat_completion_sources.CLAUDE) {
@ -2958,11 +3020,27 @@ function onProxyPasswordShowClick() {
$(this).toggleClass('fa-eye-slash fa-eye');
}
function updateScaleForm() {
if (oai_settings.use_alt_scale) {
$('#normal_scale_form').css('display', 'none');
$('#alt_scale_form').css('display', '');
} else {
$('#normal_scale_form').css('display', '');
$('#alt_scale_form').css('display', 'none');
}
}
$(document).ready(async function () {
await loadTokenCache();
$('#test_api_button').on('click', testApiConnection);
$('#scale-alt').on('change', function () {
oai_settings.use_alt_scale = !!$('#scale-alt').prop('checked');
saveSettingsDebounced();
updateScaleForm();
});
$(document).on('input', '#temp_openai', function () {
oai_settings.temp_openai = Number($(this).val());
$('#temp_counter_openai').text(Number($(this).val()).toFixed(2));

View File

@ -9,6 +9,7 @@ export const SECRET_KEYS = {
OPENROUTER: 'api_key_openrouter',
SCALE: 'api_key_scale',
AI21: 'api_key_ai21',
SCALE_COOKIE: 'scale_cookie',
}
const INPUT_MAP = {
@ -20,6 +21,7 @@ const INPUT_MAP = {
[SECRET_KEYS.OPENROUTER]: '#api_key_openrouter',
[SECRET_KEYS.SCALE]: '#api_key_scale',
[SECRET_KEYS.AI21]: '#api_key_ai21',
[SECRET_KEYS.SCALE_COOKIE]: '#scale_cookie',
}
async function clearSecret() {

View File

@ -3300,6 +3300,72 @@ async function sendScaleRequest(request, response) {
}
}
app.post("/generate_altscale", jsonParser, function (request, response_generate_scale) {
if(!request.body) return response_generate_scale.sendStatus(400);
fetch('https://dashboard.scale.com/spellbook/api/trpc/v2.variant.run', {
method: 'POST',
headers: {
'Content-Type': 'application/json',
'cookie': `_jwt=${readSecret(SECRET_KEYS.SCALE_COOKIE)}`,
},
body: JSON.stringify({
json: {
variant: {
name: 'New Variant',
appId: '',
taxonomy: null
},
prompt: {
id: '',
template: '{{input}}\n',
exampleVariables: {},
variablesSourceDataId: null,
systemMessage: request.body.sysprompt
},
modelParameters: {
id: '',
modelId: 'GPT4',
modelType: 'OpenAi',
maxTokens: request.body.max_tokens,
temperature: request.body.temp,
stop: "user:",
suffix: null,
topP: request.body.top_p,
logprobs: null,
logitBias: request.body.logit_bias
},
inputs: [
{
index: '-1',
valueByName: {
input: request.body.prompt
}
}
]
},
meta: {
values: {
'variant.taxonomy': ['undefined'],
'prompt.variablesSourceDataId': ['undefined'],
'modelParameters.suffix': ['undefined'],
'modelParameters.logprobs': ['undefined'],
}
}
})
})
.then(response => response.json())
.then(data => {
console.log(data.result.data.json.outputs[0])
return response_generate_scale.send({output: data.result.data.json.outputs[0]});
})
.catch((error) => {
console.error('Error:', error)
return response_generate_scale.send({error: true})
});
});
async function sendClaudeRequest(request, response) {
const fetch = require('node-fetch').default;
@ -3988,7 +4054,8 @@ const SECRET_KEYS = {
DEEPL: 'deepl',
OPENROUTER: 'api_key_openrouter',
SCALE: 'api_key_scale',
AI21: 'api_key_ai21'
AI21: 'api_key_ai21',
SCALE_COOKIE: 'scale_cookie',
}
function migrateSecrets() {