Laya — ONNX export
ONNX export of convaiinnovations/laya (ModernBERT-large
encoder + Laya's decision head) for use with @receptron/laya
from Node.js / TypeScript, or with ONNX Runtime directly.
| File | Contents |
|---|---|
laya.onnx / laya.onnx.data |
graph + fp32 weights (English checkpoint, 421M parameters) |
laya_config.json |
max_len, head_max_len and the per-cardinality temperatures from rl_agent_config.json |
tokenizer/ |
the checkpoint's tokenizer |
Inputs: input_ids [B,L] int64, attention_mask [B,L] int64, marker_pos [B,K] int64, marker_mask [B,K] bool, qtype [B] int64.
Outputs: logits [B,K] float32 (uncalibrated; masked slots = -1e4), act_probs [B,2] float32.
Built with export/export_onnx.py;
max logit difference vs. the PyTorch reference ≈ 1e-5.
import { Laya } from "@receptron/laya";
const laya = await Laya.load(); // downloads this bundle on first use
Weights are Convai Innovations' and remain under Apache 2.0. Export code: MIT, https://github.com/receptron/laya