rixty/SillyTavern
1
0
Fork 0
mirror of https://github.com/SillyTavern/SillyTavern.git synced 2025-03-02 19:07:40 +01:00
Code Issues Packages Projects Releases Wiki Activity

Fix Mistral's Chat Completion moment

Browse Source
This commit is contained in:
Cohee 2024-05-03 20:22:03 +03:00
parent 6b2b849a26
commit 3c2017c7b8
2 changed files with 68 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
src/endpoints/backends/chat-completions.js
View File

@ -5,7 +5,7 @@ const Readable = require('stream').Readable;
const { jsonParser } = require('../../express-common');
const { CHAT_COMPLETION_SOURCES, GEMINI_SAFETY, BISON_SAFETY, OPENROUTER_HEADERS } = require('../../constants');
const { forwardFetchResponse, getConfigValue, tryParse, uuidv4, mergeObjectWithYaml, excludeKeysByYaml, color } = require('../../util');
const { convertClaudeMessages, convertGooglePrompt, convertTextCompletionPrompt, convertCohereMessages } = require('../../prompt-converters');
const { convertClaudeMessages, convertGooglePrompt, convertTextCompletionPrompt, convertCohereMessages, convertMistralMessages } = require('../../prompt-converters');
const { readSecret, SECRET_KEYS } = require('../secrets');
const { getTokenizerModel, getSentencepiceTokenizer, getTiktokenTokenizer, sentencepieceTokenizers, TEXT_COMPLETION_MODELS } = require('../tokenizers');
@ -465,35 +465,7 @@ async function sendMistralAIRequest(request, response) {
}
try {
//must send a user role as last message
const messages = Array.isArray(request.body.messages) ? request.body.messages : [];
//large seems to be throwing a 500 error if we don't make the first message a user role, most likely a bug since the other models won't do this
if (request.body.model.includes('large'))
messages[0].role = 'user';
const lastMsg = messages[messages.length - 1];
if (messages.length > 0 && lastMsg && (lastMsg.role === 'system' || lastMsg.role === 'assistant')) {
if (lastMsg.role === 'assistant' && lastMsg.name) {
lastMsg.content = lastMsg.name + ': ' + lastMsg.content;
} else if (lastMsg.role === 'system') {
lastMsg.content = '[INST] ' + lastMsg.content + ' [/INST]';
}
lastMsg.role = 'user';
}
//system prompts can be stacked at the start, but any futher sys prompts after the first user/assistant message will break the model
let encounteredNonSystemMessage = false;
messages.forEach(msg => {
if ((msg.role === 'user' || msg.role === 'assistant') && !encounteredNonSystemMessage) {
encounteredNonSystemMessage = true;
}
if (encounteredNonSystemMessage && msg.role === 'system') {
msg.role = 'user';
//unsure if the instruct version is what they've deployed on their endpoints and if this will make a difference or not.
//it should be better than just sending the message as a user role without context though
msg.content = '[INST] ' + msg.content + ' [/INST]';
}
});
const messages = convertMistralMessages(request.body.messages, request.body.model, request.body.char_name, request.body.user_name);
const controller = new AbortController();
request.socket.removeAllListeners('close');
request.socket.on('close', function () {

66
src/prompt-converters.js
View File

@ -354,6 +354,71 @@ function convertGooglePrompt(messages, model, useSysPrompt = false, charName = '
return { contents: contents, system_instruction: system_instruction };
}
/**
* Convert a prompt from the ChatML objects to the format used by MistralAI.
* @param {object[]} messages Array of messages
* @param {string} model Model name
* @param {string} charName Character name
* @param {string} userName User name
*/
function convertMistralMessages(messages, model, charName = '', userName = '') {
if (!Array.isArray(messages)) {
return [];
}
//large seems to be throwing a 500 error if we don't make the first message a user role, most likely a bug since the other models won't do this
if (model.includes('large')) {
messages[0].role = 'user';
}
//must send a user role as last message
const lastMsg = messages[messages.length - 1];
if (messages.length > 0 && lastMsg && (lastMsg.role === 'system' || lastMsg.role === 'assistant')) {
if (lastMsg.role === 'assistant' && lastMsg.name) {
lastMsg.content = lastMsg.name + ': ' + lastMsg.content;
} else if (lastMsg.role === 'system') {
lastMsg.content = '[INST] ' + lastMsg.content + ' [/INST]';
}
lastMsg.role = 'user';
}
//system prompts can be stacked at the start, but any futher sys prompts after the first user/assistant message will break the model
let encounteredNonSystemMessage = false;
messages.forEach(msg => {
if (msg.role === 'system' && msg.name === 'example_assistant') {
if (charName) {
msg.content = `${charName}: ${msg.content}`;
}
delete msg.name;
}
if (msg.role === 'system' && msg.name === 'example_user') {
if (userName) {
msg.content = `${userName}: ${msg.content}`;
}
delete msg.name;
}
if (msg.name) {
msg.content = `${msg.name}: ${msg.content}`;
delete msg.name;
}
if ((msg.role === 'user' || msg.role === 'assistant') && !encounteredNonSystemMessage) {
encounteredNonSystemMessage = true;
}
if (encounteredNonSystemMessage && msg.role === 'system') {
msg.role = 'user';
//unsure if the instruct version is what they've deployed on their endpoints and if this will make a difference or not.
//it should be better than just sending the message as a user role without context though
msg.content = '[INST] ' + msg.content + ' [/INST]';
}
});
return messages;
}
/**
* Convert a prompt from the ChatML objects to the format used by Text Completion API.
* @param {object[]} messages Array of messages
@ -385,4 +450,5 @@ module.exports = {
convertGooglePrompt,
convertTextCompletionPrompt,
convertCohereMessages,
convertMistralMessages,
};