laya-executorch / README.md
kloodia's picture
Laya multilingual ExecuTorch gate (MLX INT4 iOS + XNNPACK 8DA4W Android, max_len=2048)
34021fe verified
|
Raw History Blame Contribute Delete
914 Bytes
metadata
license: apache-2.0
language:
  - multilingual
tags:
  - executorch
  - decision-model
  - laya
  - localia
base_model: convaiinnovations/laya

Laya multilingual ExecuTorch (Localia gate)

Exported from convaiinnovations/laya (subfolder=multilingual, mmBERT).

Path Backend Quant Size (approx)
ios/laya_gate_mlx_int4.pte MLX INT4 linear + INT8 embedding ~280 MB
android/laya_gate_xnnpack_8da4w.pte XNNPACK 8DA4W linears (embeddings fp32) ~830 MB
*/tokenizer.json — mmBERT multilingual tokenizer ~33 MB

I/O

  • Inputs: input_ids[1,2048], attention_mask[1,2048], marker_pos[1,4], marker_mask[1,4], qtype[1] (0=choice, 1=score, 2=noul)
  • Output: logits[1,4] (unused options masked ≈ -1e4)

License

Apache-2.0 (same as upstream Convai Innovations Laya).