// Web Worker: carga SigLIP 2 (torres de visión y de texto por separado) y calcula embeddings. Nada sale del navegador. import { AutoTokenizer, AutoProcessor, SiglipTextModel, SiglipVisionModel, RawImage, env } from '@huggingface/transformers'; import { MODEL_ID, MODEL_REVISION, VISION_DTYPE, TEXT_DTYPE, TEXT_MAX_TOKENS } from './config.js'; env.allowLocalModels = false; env.allowRemoteModels = true; env.useBrowserCache = true; let tok, proc, textModel, visionModel, device; function configureWasm(wasmBase, wantWebGPU) { const suffix = wantWebGPU ? '.asyncify' : ''; env.backends.onnx.wasm.wasmPaths = { mjs: `${wasmBase}ort-wasm-simd-threaded${suffix}.mjs`, wasm: `${wasmBase}ort-wasm-simd-threaded${suffix}.wasm` }; const cores = self.navigator?.hardwareConcurrency || 2; env.backends.onnx.wasm.numThreads = self.crossOriginIsolated ? Math.min(4, cores) : 1; } async function load({ wantWebGPU, wasmBase }) { configureWasm(wasmBase, wantWebGPU); const progress_callback = (p) => self.postMessage({ type: 'progress', data: { status: p.status, file: p.file, loaded: p.loaded, total: p.total } }); const opts = { revision: MODEL_REVISION, progress_callback }; tok = await AutoTokenizer.from_pretrained(MODEL_ID, opts); proc = await AutoProcessor.from_pretrained(MODEL_ID, opts); const dev = wantWebGPU ? 'webgpu' : 'wasm'; try { visionModel = await SiglipVisionModel.from_pretrained(MODEL_ID, { ...opts, dtype: VISION_DTYPE, device: dev }); textModel = await SiglipTextModel.from_pretrained(MODEL_ID, { ...opts, dtype: TEXT_DTYPE, device: dev }); } catch (e) { if (wantWebGPU) { self.postMessage({ type: 'webgpu-failed', reason: String(e?.message || e).slice(0, 200) }); return; } throw e; } device = dev; self.postMessage({ type: 'ready', device, threads: env.backends.onnx.wasm.numThreads }); } function norm(arr) { let s = 0; for (const x of arr) s += x * x; s = Math.sqrt(s) || 1; return Float32Array.from(arr, (x) => x / s); } async function embedImage({ id, width, height, buffer }) { const image = new RawImage(new Uint8ClampedArray(buffer), width, height, 4).rgb(); const t0 = performance.now(); const inputs = await proc(image); const out = await visionModel(inputs); const vec = norm(out.pooler_output.data); self.postMessage({ type: 'vector', id, vec, ms: performance.now() - t0, device }, [vec.buffer]); } async function embedText({ id, text }) { const t0 = performance.now(); const inputs = tok([text], { padding: 'max_length', max_length: TEXT_MAX_TOKENS, truncation: true }); const out = await textModel(inputs); const vec = norm(out.pooler_output.data); self.postMessage({ type: 'vector', id, vec, ms: performance.now() - t0, device }, [vec.buffer]); } self.addEventListener('message', async (ev) => { const msg = ev.data || {}; try { if (msg.type === 'load') await load(msg); else if (msg.type === 'image') await embedImage(msg); else if (msg.type === 'text') await embedText(msg); } catch (e) { self.postMessage({ type: 'error', id: msg.id, stage: msg.type, message: String(e?.message || e).slice(0, 400) }); } });