Laya multilingual¶

Adapter: laya — fused encoder + decision head, marker_pos / qtype, calibrated temperatures
License: Apache-2.0
Parameters: 322M · Architecture: mmBERT-base · Languages: 100+ languages
Context: 1,024 tokens in this export (the upstream checkpoint supports up to 8,192)

Default variant¶


Repo	`onnx-community/laya-multilingual-ONNX`
Files	`onnx/model_fp16.onnx` + `.onnx_data`, `tokenizer.json`, `config.json`
Quantization	fp16 weights and activations, fp32 inputs/outputs
Download	682 MB
Role	default

Variants¶

Variant	Quantization	Size	Role	Install
`fp16` *	fp16	681.8 MB	default	`dtk models download laya-multilingual --variant fp16`
`fp32`	fp32	1.33 GB	quality	`dtk models download laya-multilingual --variant fp32`

dtk models download laya-multilingual gets fp16; --variant fp32 downloads the full-precision build. Select one with dtk models use laya-multilingual --variant <name>. See Variants and quantization.

Reproducibility¶

Revision: 46b77bbf5642fec5f14e540570228a8cbe8ab81f
Default file hashes (SHA256):
onnx/model_fp16.onnx — 81f641e21289ddb2cd8ac774c1cddebfdb0fa7e4b183e49ad476d2faee46d98a
onnx/model_fp16.onnx_data — 3c9eb52f5f795f76d7a681197228131fac7e92c3e5a4c90b68ec4065699ee4cf
tokenizer.json — 609d8f4c067cd3950f88594c5a802616cea245823836ef5848ee4fc40aab5b6f
config.json — cc32109b1fded91ec6734127ffb1827f80cdcf8ffdd4eeda75a1b0f4d3687dcc
Verify: dtk models verify laya-multilingual --variant fp16
Bench: uv run python bench/run.py --suite bench/suites/smoke.jsonl --models laya-multilingual
The fp32 variant hashes live in catalog.py.

Evidence (2026-10-21)¶

CPU p50: 37.4 ms fp16, 34.1 ms fp32 on the audit machine after warmup.
Local example (Help! My payouts have been failing for 3 days!): team billing 1.00, urgency 0.756, frustration 1.26.
The upstream conversion reports max |Δp| 1e-4 fp32 and 0.035 fp16 against the PyTorch checkpoint on typed-decision sequences.
fp16 and fp32 agree on the answers in the local smoke/audit cases.

Contract notes¶

Sequence: [CLS] <type> question: <instructions> [SEP] ([MASK] option) … [SEP] state [SEP].
One fused graph: input_ids, attention_mask, marker_pos, marker_mask, qtype → logits per option + act_logits.
Temperatures come from the laya section of config.json and are clamped to [0.5, 5], matching the reference runtime.
The adapter also supports the split encoder + head Laya layout; the multilingual export is the fused layout.

Known limitations¶

Low-resource languages remain weak in the upstream checkpoint; validate your language and label set before relying on calibrated confidence.
This export's max_len is 1,024; the upstream checkpoint supports 8,192 with a larger setting, but the pinned export config uses 1,024.
The adapter uses the fused graph path; the English split checkpoint remains a separate catalog entry (laya).

Loading each variant¶

REST selects the quantization through the standard model field using the alias laya-multilingual@variant; there is no variant field in the request body.

`fp16` (fp16, default)¶

Library

from decision_toolkit import DecisionModel

model = DecisionModel("laya-multilingual", variant="fp16")

CLI

dtk models download laya-multilingual --variant fp16
dtk models use laya-multilingual --variant fp16
dtk decide --model laya-multilingual --variant fp16 \
  --state "..." --q 'urgent=noul:Is this urgent?'

REST

curl -s http://127.0.0.1:8000/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{"model": "laya-multilingual@fp16", "state": "...",
       "questions": {"urgent": {"type": "noul",
                     "instructions": "Is this urgent?"}}}}'

`fp32` (fp32, quality)¶