Gemma3n / index.js
tylermullen's picture
Update index.js
81a9a45 verified
// Copyright 2025 The MediaPipe Authors.
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
// http://www.apache.org/licenses/LICENSE-2.0
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
// -------------------------------------------------------------------------- //
import {oauthLoginUrl, oauthHandleRedirectIfPresent} from "@huggingface/hub";
import {FilesetResolver, LlmInference} from '@mediapipe/tasks-genai';
// --- DOM Element References ---
const webcamElement = document.getElementById('webcam');
const statusMessageElement = document.getElementById(
'status-message',
);
const responseContainer = document.getElementById(
'response-container',
);
const promptInputElement = document.getElementById(
'prompt-input',
);
const recordButton = document.getElementById(
'record-button',
);
const sendButton = document.getElementById('send-button');
const clearCacheButton = document.getElementById('clear-cache-button');
const recordButtonIcon = recordButton.querySelector('i');
const loaderOverlay = document.getElementById('loader-overlay');
const progressBarFill = document.getElementById('progress-bar-fill');
const signInMessage = document.getElementById('sign-in-message');
const loaderMessage = document.getElementById('loader-message');
const versionText = document.getElementById('version-text');
const toggleVersionButton = document.getElementById('toggle-version-button');
// --- State Management ---
let isRecording = false;
let isLoading = false;
let mediaRecorder = null;
let audioChunks = [];
// --- Model-specific constants ---
// If the user wants to try running on a more limited device, they can switch
// the demo from default E4B to E2B by appending '?e2b' to the URL.
const thisUrl = new URL(window.location.href);
const use_e4b = !thisUrl.searchParams.has('e2b');
const cacheFileName = use_e4b ? "3n_e4b" : "3n_e2b";
const remoteFileUrl = use_e4b ? 'https://huggingface.co/google/gemma-3n-E4B-it-litert-lm/resolve/main/gemma-3n-E4B-it-int4-Web.litertlm'
: 'https://huggingface.co/google/gemma-3n-E2B-it-litert-lm/resolve/main/gemma-3n-E2B-it-int4-Web.litertlm';
// Model size in bytes for reliable progress indication; just hard-coded for now
const modelSize = use_e4b ? 4275044352 : 3038117888;
// --- Core Functions ---
/**
* Updates the progress bar's width.
* @param {number} percentage The progress percentage (0-100).
*/
function updateProgressBar(percentage) {
if (progressBarFill) {
progressBarFill.style.width = `${percentage}%`;
}
}
/**
* Initializes our local LLM from a StreamReader.
*/
let llmInference;
async function initLlm(modelReader) {
console.log('Initializing LLM');
loaderMessage.textContent = "Initializing model...";
// We have no actual progress updates for this last initialization step, but
// it's relatively short (<10s on a decent laptop). So we just set to 90%.
// TODO: It'd look nicer to have this go from 0 to 100% instead.
updateProgressBar(90);
const genaiFileset = await FilesetResolver.forGenAiTasks(
'https://cdn.jsdelivr.net/npm/@mediapipe/tasks-genai/wasm');
try {
llmInference = await LlmInference.createFromOptions(genaiFileset, {
baseOptions: {modelAssetBuffer: modelReader},
maxTokens: 2048,
maxNumImages: 1,
supportAudio: true,
});
// Enable demo now that loading has fully finished.
loaderOverlay.style.opacity = '0';
setTimeout(() => {
loaderOverlay.style.display = 'none';
promptInputElement.disabled = false;
sendButton.disabled = false;
recordButton.disabled = false;
}, 300);
} catch (error) {
console.error('Failed to initialize the LLM', error);
loaderOverlay.style.display = 'none'; // Hide loader on error
}
}
/**
* Replaces our demo with a sign-in button for HuggingFace.
*/
function requireSignIn() {
document.getElementById('loader-overlay').style = "display:none";
document.getElementById('main-container').style = "display:none";
document.body.parentElement.prepend(document.getElementById('title-container'));
document.body.appendChild(document.getElementById('version-info'));
document.getElementById("signin").style.removeProperty("display");
document.getElementById('sign-in-message').style.removeProperty("display");
document.getElementById("signin").onclick = async function() {
// prompt=consent to re-trigger the consent screen instead of silently redirecting
window.location.href = (await oauthLoginUrl({scopes: window.huggingface.variables.OAUTH_SCOPES})) + "&prompt=consent";
};
// clear old oauth, if any
localStorage.removeItem("oauth");
}
/**
* Utility function to show progress while we load from remote file into local
* cache.
*/
async function pipeStreamAndReportProgress(readableStream, writableStream) {
// Alert the user if browser expects caching to run out of memory.
const cacheEstimate = await navigator.storage.estimate();
if (modelSize > (cacheEstimate.quota - cacheEstimate.usage)) {
alert(`The browser reports it does not have enough space in cache for this model. Ensure you are not running in incognito mode, or else try to free up some space. Model size: ${modelSize}. Cache quota: ${cacheEstimate.quota}. Cache usage: ${cacheEstimate.usage}.`);
}
// Effectively "await responseStream.pipeTo(writeStream)", but with progress
// reporting.
const reader = readableStream.getReader();
const writer = writableStream.getWriter();
let bytesCount = 0;
let progressBarPercent = 0;
let wasAborted = false;
try {
while (true) {
const {done, value} = await reader.read();
if (done) {
break;
}
if (value) {
bytesCount += value.length;
const percentage = Math.round(bytesCount / modelSize * 90);
if (percentage > progressBarPercent) {
progressBarPercent = percentage;
updateProgressBar(progressBarPercent);
const downloadedMB = (bytesCount / 1e6).toFixed(2);
const totalMB = (modelSize / 1e6).toFixed(2);
loaderMessage.textContent =
`Downloading model: ${downloadedMB}MB / ${totalMB}MB`;
}
await writer.write(value);
}
}
} catch (error) {
console.error('Error while piping stream:', error);
// Abort the writer if there's an error
wasAborted = true;
await writer.abort(error);
throw error;
} finally {
// Release the reader lock
reader.releaseLock();
// Close the writer only if the stream wasn't aborted
if (!wasAborted) {
console.log('Closing the writer, and hence the stream');
await writer.close();
}
}
}
/**
* Loads the LLM file from either cache or OAuth-guarded remote download.
*/
async function loadLlm() {
// For browsers other than Chrome and Edge, we just show an alert and error out.
const isChrome = navigator.userAgent.includes('Chrome');
const isEdge = navigator.userAgent.includes('Edg');
if (!isChrome && !isEdge) {
alert('Non-Chromium browsers are not supported yet. Please run demo on Chrome for the best experience.');
loaderMessage.textContent = "Non-Chromium browsers are not supported yet. Please run on Chrome for best experience.";
return;
}
let opfs = await navigator.storage.getDirectory();
// If we can load the model from cache, then do so.
try {
const fileHandle = await opfs.getFileHandle(cacheFileName);
// Check to make sure size is as expected, and not a partially-downloaded
// or corrupt file.
console.log('Model found in cache; checking size.');
const file = await fileHandle.getFile();
console.log('File size is: ', file.size);
if (file.size !== modelSize) {
console.error('Cached model had unexpected size. Redownloading.');
throw new Error('Unexpected cached model size');
}
console.log('Model found in cache of expected size, reusing.');
const fileReader = file.stream().getReader();
await initLlm(fileReader);
} catch {
// Otherwise, we need to be download remotely, which requires oauth.
console.log('Model not found in cache: oauth and download required.');
// We first remove from cache, in case model file is corrupted/partial.
try {
await opfs.removeEntry(cacheFileName);
} catch {}
let oauthResult = localStorage.getItem("oauth");
if (oauthResult) {
try {
oauthResult = JSON.parse(oauthResult);
} catch {
oauthResult = null;
}
}
oauthResult ||= await oauthHandleRedirectIfPresent();
// If we have successful oauth from one of the methods above, download from
// remote.
if (oauthResult?.accessToken) {
localStorage.setItem("oauth", JSON.stringify(oauthResult));
const modelUrl = remoteFileUrl;
const oauthHeaders = {
"Authorization": `Bearer ${oauthResult.accessToken}`
};
const response = await fetch(modelUrl, {headers: oauthHeaders});
if (response.ok) {
const responseStream = await response.body;
// Cache locally, so we can avoid this next time.
const fileHandle =
await opfs.getFileHandle(cacheFileName, {create: true});
const writeStream = await fileHandle.createWritable();
await pipeStreamAndReportProgress(responseStream, writeStream);
console.log('Model written to cache!');
const file = await fileHandle.getFile();
const fileReader = file.stream().getReader();
await initLlm(fileReader);
} else {
console.error('Model fetch encountered error. Likely requires sign-in or Gemma license acknowledgement.');
requireSignIn();
}
} else {
// No successful oauth, so replace our demo with a HuggingFace sign-in button.
console.log('No oauth detected. Requiring sign-in.');
requireSignIn();
}
}
}
/**
* Initializes the webcam and microphone.
*/
let audioUrl = undefined;
async function initMedia() {
versionText.textContent = use_e4b ? 'E4B' : 'E2B';
toggleVersionButton.textContent = use_e4b ? 'Switch to E2B' : 'Switch to E4B';
// Disable controls on startup
promptInputElement.disabled = true;
sendButton.disabled = true;
recordButton.disabled = true;
try {
const videoStream = await navigator.mediaDevices.getUserMedia({video: true});
const audioStream = await navigator.mediaDevices.getUserMedia({audio: true});
webcamElement.srcObject = videoStream;
statusMessageElement.style.display = 'none';
webcamElement.style.display = 'block';
await loadLlm();
// Set up MediaRecorder for audio
mediaRecorder = new MediaRecorder(audioStream);
mediaRecorder.ondataavailable = (event) => {
console.log('ondataavailable event: ', event);
audioChunks.push(event.data);
};
mediaRecorder.onstop = () => {
// We process the audio here, and free the previous one, if any
const blob = new Blob(audioChunks, {type: 'audio/webm'});
if (audioUrl) window.URL.revokeObjectURL(audioUrl);
audioUrl = window.URL.createObjectURL(blob);
audioChunks = [];
sendQuery({audioSource: audioUrl});
};
} catch (error) {
console.error('Error accessing media devices.', error);
audioUrl = undefined;
statusMessageElement.textContent =
'Error: Could not access camera or microphone. Please check permissions.';
loaderOverlay.style.display = 'none'; // Hide loader on error
}
}
/**
* Toggles the audio recording state.
*/
function toggleRecording() {
isRecording = !isRecording;
if (isRecording) {
if (mediaRecorder && mediaRecorder.state === 'inactive') {
mediaRecorder.start();
}
recordButton.classList.add('recording');
if (recordButtonIcon) {
recordButtonIcon.className = 'fa-solid fa-stop';
}
promptInputElement.placeholder = 'Recording... Press stop when done.';
} else {
if (mediaRecorder && mediaRecorder.state === 'recording') {
mediaRecorder.stop();
}
recordButton.classList.remove('recording');
if (recordButtonIcon) {
recordButtonIcon.className = 'fa-solid fa-microphone';
}
promptInputElement.placeholder = 'Ask a question about what you see...';
}
}
/**
* Sends a text prompt with webcam frame to the Gemma 3n model.
*/
async function sendTextQuery() {
const prompt = promptInputElement.value.trim();
sendQuery(prompt);
}
/**
* Sends the user's prompt (text or audio) with webcam frame to the Gemma 3n model.
*/
async function sendQuery(prompt) {
if (!prompt || isLoading) {
return;
}
setLoading(true);
try {
const query = [
'<start_of_turn>user\n',
prompt, // audio or text
{imageSource: webcam},
'<end_of_turn>\n<start_of_turn>model\n',
];
let resultSoFar = '';
await llmInference.generateResponse(query, (newText, isDone) => {
resultSoFar += newText;
updateResponse(resultSoFar);
});
promptInputElement.value = '';
} catch (error) {
console.error('Error running Gemma 3n on query.', error);
updateResponse(
`Error: Could not get a response. ${error instanceof Error ? error.message : String(error)}`,
);
} finally {
setLoading(false);
}
}
/**
* Updates the response container with new content.
* @param {string} text The text to display.
*/
function updateResponse(text) {
responseContainer.classList.remove('thinking');
responseContainer.innerHTML = '';
const p = document.createElement('p');
p.textContent = text;
responseContainer.appendChild(p);
}
/**
* Sets the loading state of the UI.
* @param {boolean} loading - True if loading, false otherwise.
*/
function setLoading(loading) {
isLoading = loading;
promptInputElement.disabled = loading;
sendButton.disabled = loading;
recordButton.disabled = loading;
if (loading) {
responseContainer.classList.add('thinking');
responseContainer.innerHTML = '<p>Processing...</p>';
}
}
// --- Event Listeners ---
recordButton.addEventListener('click', toggleRecording);
sendButton.addEventListener('click', sendTextQuery);
promptInputElement.addEventListener('keydown', (event) => {
if (event.key === 'Enter') {
sendTextQuery();
}
});
clearCacheButton.addEventListener('click', async () => {
const userConfirmed = confirm(
'Are you sure you want to clear the cached model? ' +
'This will require re-downloading the model on the next visit.'
);
if (userConfirmed) {
try {
const opfs = await navigator.storage.getDirectory();
await opfs.removeEntry(cacheFileName);
console.log('Cache cleared successfully.');
clearCacheButton.style.display = 'none';
} catch (error) {
console.error('Error clearing cache:', error);
}
}
});
toggleVersionButton.addEventListener('click', () => {
const url = new URL(window.location.href);
if (use_e4b) {
url.searchParams.set('e2b', 'true');
} else {
url.searchParams.delete('e2b');
}
window.location.href = url.href;
});
// --- Initialization ---
document.addEventListener('DOMContentLoaded', initMedia);