mizchi/laya-multilingual-onnx

Laya multilingual — ONNX (float16)

ONNX export of convaiinnovations/laya-multilingual (mmBERT-base encoder, 322M parameters) for onnxruntime, including onnxruntime-web with the WebGPU backend. Produced by laya-mlx with laya-mlx export-onnx --dtype float16 from the published aac6fef/laya-multilingual-mlx checkpoint (snapshot f2b4faf51023039425946074e2cf1361d2db11d5, itself converted from upstream revision 052592a15d198d9ad47da779604259b10b47b7aa).

File	Purpose
model.onnx	Encoder, decision head, scorer and action head in one graph. Opset 18, standard operators only, float16 weights with a float32 decision tail.
rl_agent_config.json	Prompt limits (max_len, head_max_len) and calibration temperatures.
tokenizer/	Hugging Face tokenizer.json and tokenizer_config.json.
onnx_config.json	Export metadata: dtype, opset, input and output names, source checkpoint.

Inputs: input_ids, attention_mask (int64 [batch, sequence]), marker_pos (int64 [batch, markers]), marker_mask (bool [batch, markers]), qtype (int64 [batch]; 0 choice, 1 score, 2 noul). Outputs: logits [batch, markers] and act_logits [batch, actions], both float32. logits still need the calibration temperature and a softmax. Prompt construction, calibration and answer formatting follow upstream Laya; the reference implementation is laya_mlx.common / laya_mlx.agent and the browser port is @laya-mlx/web in the repository.

Validation on the repository's 63-question parity fixtures against the MLX runtime: 63/63 selected answers agree, maximum probability error 5.1e-4 with the onnxruntime CPU provider and 1.2e-2 with onnxruntime-web WebGPU (float16 compute). See docs/ONNX_EXPORT.md in the repository for the method, the browser measurements and the size analysis. onnxruntime-web 1.30.0 needs graphOptimizationLevel: "basic" on WebGPU for this graph.

This is an independent port, not an official Convai Innovations release. Weights are Apache-2.0 as upstream.

Live demo: https://huggingface.co/spaces/mizchi/laya-web-demo (Snake and Chess, both run the model in your browser — direct pages: https://mizchi-laya-web-demo.static.hf.space/snake.html and https://mizchi-laya-web-demo.static.hf.space/chess.html) — also on the repository's GitHub Pages site (for the mizchi/laya-mlx fork: https://mizchi.github.io/laya-mlx/).