alexdoandev/laya-multilingual-onnx

Laya multilingual — ONNX (fp32, dynamo export)

ONNX export of the Laya multilingual System-1 decision model (© Convai Innovations, Apache-2.0) — a calibrated decision engine that answers typed questions (choice / score / noul) about a state in one forward pass: no text generation, sub-second on CPU.

This is the first publicly available multilingual ONNX bundle of Laya — the published receptron/laya-onnx ships English-only weights.

Why the community needs this

LLM routing and guardrails need thousands of cheap calibrated decisions, not one expensive generation. Run this 322M-parameter encoder locally and ask it typed questions — "which team handles this?", "is this urgent?", "P(destructive)?" — in ~0.15–0.4 s per batch, offline, $0/token. Reference integration: jev-laya-dsh (a System-1 decision layer for DeepSeek Harness agents).

Files

File	Purpose
laya.onnx + laya.onnx.data	the graph (opset 18, dynamo export) + fp32 weights
laya_config.json	max_len, head_max_len, per-cardinality temperature
tokenizer/	mmBERT tokenizer

Inputs: input_ids [B,L] i64, attention_mask [B,L] i64, marker_pos [B,K] i64, marker_mask [B,K] bool, qtype [B] i64 (0=choice, 1=score, 2=noul). Outputs: logits [B,K] f32 (uncalibrated; apply the temperature from laya_config.json then softmax) and act_probs [B,2] f32.

Usage

Node — @receptron/laya ≥ 0.1.2 with the multilingual specials patch (or any onnxruntime binding that builds the sequences below):

import { Laya } from "@receptron/laya";
const laya = await Laya.load({ repo: "alexdoandev/laya-multilingual-onnx" });
// note: multilingual (mmBERT) checkpoints need specials read from
// tokenizer_config.json — the upstream JS hardcodes BERT tokens;
// see jev-laya-dsh onnx-server/patches/apply-multilingual-patch.js

Python — reference laya package loads the original torch checkpoint; use the export script in jev-laya-dsh/onnx-export/ to reproduce this bundle.

Rust — ort crate + Tensor::from_shape (see jev-laya-dsh benchmarks/rust/).

Provenance & parity

• Exported 2026-10-02 from convaiinnovations/laya-multilingual (revision e4e9ddf2) with torch 2.14, opset 18, dynamo exporter.
• Parity vs PyTorch: max |Δlogits| = 7.4e-06.
• Served in production as a prompt router + risk gate for coding agents (fused routing accuracy 86.9% on a 61-prompt Vietnamese benchmark — see the integration repo).

License

Weights and this export: Apache-2.0 (inherited from the Convai checkpoint). "Jev" is a TypeSafe product and is unrelated to this artifact.