Upload serve_gemma.js with huggingface_hub

serve_gemma.js

@@ -0,0 +1,71 @@
+const http = require('http'), fs = require('fs'), path = require('path');
+const PORT = 8150;
+const ROOT = __dirname;
+const SPLITS = path.join(ROOT, 'model_splits');
+
+const MIME = {
+  '.html': 'text/html', '.js': 'text/javascript', '.mjs': 'text/javascript',
+  '.json': 'application/json', '.gguf': 'application/octet-stream',
+  '.wasm': 'application/wasm', '.ts': 'text/javascript',
+};
+
+http.createServer((req, res) => {
+  if (req.method === 'OPTIONS') {
+    res.writeHead(200, {
+      'Access-Control-Allow-Origin': '*',
+      'Access-Control-Allow-Methods': 'GET,HEAD',
+      'Access-Control-Allow-Headers': 'Content-Type,Range',
+      'Cross-Origin-Embedder-Policy': 'require-corp',
+      'Cross-Origin-Opener-Policy': 'same-origin',
+    });
+    return res.end();
+  }
+
+  let p = decodeURIComponent(req.url.split('?')[0]);
+  if (p === '/') p = '/index.html';
+
+  let fp;
+  if (p.startsWith('/model/')) {
+    fp = path.join(SPLITS, p.slice('/model/'.length));
+  } else {
+    fp = path.join(ROOT, p);
+  }
+
+  fs.stat(fp, (e, st) => {
+    if (e) { res.writeHead(404); return res.end('not found: ' + fp); }
+    const total = st.size;
+    const range = req.headers.range;
+    const ct = MIME[path.extname(fp)] || 'application/octet-stream';
+
+    const headers = {
+      'Content-Type': ct,
+      'Accept-Ranges': 'bytes',
+      'Access-Control-Allow-Origin': '*',
+      'Cross-Origin-Embedder-Policy': 'require-corp',
+      'Cross-Origin-Opener-Policy': 'same-origin',
+    };
+
+    if (range) {
+      const m = range.match(/bytes=(\d+)-(\d*)/);
+      if (m) {
+        const start = parseInt(m[1]);
+        const end = m[2] ? parseInt(m[2]) : total - 1;
+        headers['Content-Range'] = 'bytes ' + start + '-' + end + '/' + total;
+        headers['Content-Length'] = end - start + 1;
+        res.writeHead(206, headers);
+        fs.createReadStream(fp, { start, end }).pipe(res);
+        return;
+      }
+    }
+
+    headers['Content-Length'] = total;
+    res.writeHead(200, headers);
+    fs.createReadStream(fp).pipe(res);
+  });
+}).listen(PORT, () => {
+  console.log('Gemma WebGPU on :' + PORT);
+  console.log('Model splits: ' + SPLITS);
+  const files = fs.readdirSync(SPLITS).filter(f => f.endsWith('.gguf'));
+  console.log('Split files: ' + files.length);
+  console.log('CORS + COEP/COOP headers enabled for wllama multi-threading');
+});