SillyTavern/src/transformers.mjs

150 lines
4.9 KiB
JavaScript

import { pipeline, env, RawImage, Pipeline } from 'sillytavern-transformers';
import { getConfigValue } from './util.js';
import path from 'path';
import fs from 'fs';
configureTransformers();
function configureTransformers() {
// Limit the number of threads to 1 to avoid issues on Android
env.backends.onnx.wasm.numThreads = 1;
// Use WASM from a local folder to avoid CDN connections
env.backends.onnx.wasm.wasmPaths = path.join(process.cwd(), 'dist') + path.sep;
}
const tasks = {
'text-classification': {
defaultModel: 'Cohee/distilbert-base-uncased-go-emotions-onnx',
pipeline: null,
configField: 'extras.classificationModel',
quantized: true,
},
'image-to-text': {
defaultModel: 'Xenova/vit-gpt2-image-captioning',
pipeline: null,
configField: 'extras.captioningModel',
quantized: true,
},
'feature-extraction': {
defaultModel: 'Xenova/all-mpnet-base-v2',
pipeline: null,
configField: 'extras.embeddingModel',
quantized: true,
},
'text-generation': {
defaultModel: 'Cohee/fooocus_expansion-onnx',
pipeline: null,
configField: 'extras.promptExpansionModel',
quantized: false,
},
'automatic-speech-recognition': {
defaultModel: 'Xenova/whisper-small',
pipeline: null,
configField: 'extras.speechToTextModel',
quantized: true,
},
'text-to-speech': {
defaultModel: 'Xenova/speecht5_tts',
pipeline: null,
configField: 'extras.textToSpeechModel',
quantized: false,
},
};
/**
* Gets a RawImage object from a base64-encoded image.
* @param {string} image Base64-encoded image
* @returns {Promise<RawImage|null>} Object representing the image
*/
async function getRawImage(image) {
try {
const buffer = Buffer.from(image, 'base64');
const byteArray = new Uint8Array(buffer);
const blob = new Blob([byteArray]);
const rawImage = await RawImage.fromBlob(blob);
return rawImage;
} catch {
return null;
}
}
/**
* Gets the model to use for a given transformers.js task.
* @param {string} task The task to get the model for
* @returns {string} The model to use for the given task
*/
function getModelForTask(task) {
const defaultModel = tasks[task].defaultModel;
try {
const model = getConfigValue(tasks[task].configField, null);
return model || defaultModel;
} catch (error) {
console.warn('Failed to read config.yaml, using default classification model.');
return defaultModel;
}
}
async function migrateCacheToDataDir() {
const oldCacheDir = path.join(process.cwd(), 'cache');
const newCacheDir = path.join(global.DATA_ROOT, '_cache');
if (!fs.existsSync(newCacheDir)) {
fs.mkdirSync(newCacheDir, { recursive: true });
}
if (fs.existsSync(oldCacheDir) && fs.statSync(oldCacheDir).isDirectory()) {
const files = fs.readdirSync(oldCacheDir);
if (files.length === 0) {
return;
}
console.log('Migrating model cache files to data directory. Please wait...');
for (const file of files) {
try {
const oldPath = path.join(oldCacheDir, file);
const newPath = path.join(newCacheDir, file);
fs.cpSync(oldPath, newPath, { recursive: true, force: true });
fs.rmSync(oldPath, { recursive: true, force: true });
} catch (error) {
console.warn('Failed to migrate cache file. The model will be re-downloaded.', error);
}
}
}
}
/**
* Gets the transformers.js pipeline for a given task.
* @param {import('sillytavern-transformers').PipelineType} task The task to get the pipeline for
* @param {string} forceModel The model to use for the pipeline, if any
* @returns {Promise<Pipeline>} Pipeline for the task
*/
async function getPipeline(task, forceModel = '') {
await migrateCacheToDataDir();
if (tasks[task].pipeline) {
if (forceModel === '' || tasks[task].currentModel === forceModel) {
return tasks[task].pipeline;
}
console.log('Disposing transformers.js pipeline for for task', task, 'with model', tasks[task].currentModel);
await tasks[task].pipeline.dispose();
}
const cacheDir = path.join(global.DATA_ROOT, '_cache');
const model = forceModel || getModelForTask(task);
const localOnly = getConfigValue('extras.disableAutoDownload', false);
console.log('Initializing transformers.js pipeline for task', task, 'with model', model);
const instance = await pipeline(task, model, { cache_dir: cacheDir, quantized: tasks[task].quantized ?? true, local_files_only: localOnly });
tasks[task].pipeline = instance;
tasks[task].currentModel = model;
return instance;
}
export default {
getPipeline,
getRawImage,
};