import onnxruntime as ort

# Load with automatic provider selection (CUDA > CoreML > CPU)
providers = ort.get_available_providers()
session = ort.InferenceSession("encoder_model.onnx", providers=providers)
outputs = session.run(None, inputs)

With Transformers.js (Browser/WebGPU)

javascript

import { pipeline } from "@xenova/transformers";

// Load the quantized ONNX model
const pipe = await pipeline("image-to-text", "pix2act-weblinx-base-onnx", {
  quantized: true,  // Use INT8 quantized version
});

// Run inference on a screenshot
const result = await pipe(image);
console.log(result);

Browser Requirements:

WebGPU-enabled browser (Chrome 113+, Edge 113+, or Firefox Nightly)
Falls back to WebAssembly (WASM) if WebGPU unavailable
Recommended: 4GB+ system RAM for optimal performance

Export Details

Exported using HotelBench ONNX export utilities.

bash

# Export with INT8 quantization (recommended for browser)
hotelbench export-onnx pix2act-weblinx-base -o models/pix2act-weblinx-base-onnx --quantize

# Push to HuggingFace Hub
hotelbench export-onnx pix2act-weblinx-base -o models/pix2act-weblinx-base-onnx -q \
  --push-to-hub username/pix2act-weblinx-base-onnx

Performance Notes

Browser (WebGPU): ~200-500ms inference time on modern GPU
Browser (WASM): ~1-3s inference time (CPU fallback)
Server (CUDA): ~50-100ms inference time
Server (CPU): ~500ms-1s inference time

License

Apache 2.0 - See original model for full license details.

pierretokns/pix2act-weblinx-base-onnx

Pix2Act-WebLINX-Base ONNX

Model Details

Files

Usage

With ONNX Runtime (Python/Server)

With Transformers.js (Browser/WebGPU)

Export Details

Performance Notes

License

pierretokns/
pix2act-weblinx-base-onnx