laya-executorch / meta.json
kloodia's picture
Laya multilingual ExecuTorch gate (MLX INT4 iOS + XNNPACK 8DA4W Android, max_len=2048)
34021fe verified
Raw History Blame Contribute Delete
672 Bytes
{
"source": "convaiinnovations/laya/multilingual",
"role": "localia-command-gate",
"modalities": [
"text"
],
"languages": "100+",
"quant": {
"ios": "mlx weight-only int4 group32 + embedding int8",
"android": "xnnpack 8da4w (int8 dynamic act + int4 weights group32)"
},
"max_len": 2048,
"max_opts": 4,
"android": "android/laya_gate_xnnpack_8da4w.pte",
"ios": "ios/laya_gate_mlx_int4.pte",
"inputs": [
"input_ids[1,2048] long",
"attention_mask[1,2048] long",
"marker_pos[1,4] long",
"marker_mask[1,4] bool",
"qtype[1] long (0=choice, 1=score, 2=noul)"
],
"output": "logits[1,4] float",
"license": "Apache-2.0"
}