AdhyanshVerma/browser-laya-onnx

Laya Browser Decision Head (Ultra-Edge INT8 ONNX)

This repository provides the official Ultra-Edge INT8 Quantized ONNX version of the Laya System-1 Decision Head for browser agents, created for the Pooki-Pra anti-detect browser engine.

Model Highlights

• Model Architecture: ModernBERT unpadded sequence classification fine-tuned for DOM action spaces (CLICK, TYPE_TEXT, SELECT, SCROLL_DOWN, SCROLL_UP, WAIT, DONE).
• Quantization: Dynamic INT8 symmetric quantization (onnxruntime.quantization.QuantType.QInt8).
• Ultra-Lightweight: 310.1 MB (over 4x size reduction from the 1.28 GB FP32 model).
• Zero GPU Required: Sub-45ms inference on standard x86_64 CPUs, laptops, and ARM edge devices.

Files in this Repository

File	Size	Description
model.onnx / model_int8.onnx	310 MB	Self-contained INT8 quantized ONNX decision model
config.json	1.8 KB	ModernBERT base configuration
rl_agent_config.json	68 B	Reinforcement learning action head mapping
tokenizer/	Directory	Tokenizer JSON vocabulary (256,000 tokens)

Quick Inference Example

import onnxruntime as ort
import numpy as np

# Load the 310 MB INT8 model
session = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])

# Forward pass with dynamic sequence length and action markers
outputs = session.run(None, {
    "input_ids": input_ids,
    "attention_mask": attention_mask,
    "marker_pos": marker_pos,
    "marker_mask": marker_mask,
    "qtype": qtype
})
action_logits = outputs[0]

Part of Pooki-Pra

• Developed by Adhyansh Verma.
• Designed for Pooki-Pra & Camoufox.