verdict-models/laya-multilingual

Laya Multilingual (Verdict-ready)

The Laya Multilingual System One decision model (mmBERT-base, 100+ languages), exported to ONNX fp32 and packaged for Verdict -- a high-performance serving engine for System One decision models. Serve it with one command:

# CPU host
docker run --rm -p 8080:8080 \
    -v ~/.cache/verdict:/root/.cache/verdict \
    sgyaqing/verdict:latest-cpu serve verdict-models/laya-multilingual --host 0.0.0.0

# GPU host
docker run --rm --gpus all -p 8080:8080 \
    -v ~/.cache/verdict:/root/.cache/verdict \
    sgyaqing/verdict:latest-cuda serve verdict-models/laya-multilingual --host 0.0.0.0

# ModelScope:
docker run --rm -p 8080:8080 \
    -e VERDICT_USE_MODELSCOPE=1 \
    -v ~/.cache/verdict:/root/.cache/verdict \
    sgyaqing/verdict:latest-cpu serve sgyaqing/laya-multilingual --host 0.0.0.0

For English-only workloads, use verdict-models/laya.

Endpoints: the standard decision protocols POST /v1/systemone and POST /v1/decisions, native POST /v1/choice | /v1/score | /v1/noul, plus an OpenAI-compatible layer (/v1/chat/completions, /v1/models).

Performance

Measured on one RTX 4090, end to end over HTTP, against the model author's own server (laya-serve). A ratio above 1x means Verdict is ahead.

concurrency	Verdict rps	laya-serve rps	rps ratio	Verdict P50 / P99 (ms)	laya-serve P50 / P99 (ms)	P50 / P99 ratio
1	123.2	49.1	2.51x	8.03 / 10.29	18.54 / 28.92	2.31 / 2.81x
8	553.4	46.3	11.9x	13.99 / 19.79	184.36 / 204.24	13.2 / 10.3x
64	2534.4	refused requests	—	24.67 / 32.23	refused requests	—

Files

File	Purpose
model.json	manifest: graph file name, budgets, calibration, special-token names
laya-multilingual.onnx	the ONNX graph (fp32)
tokenizer/tokenizer.json (+ tokenizer_config.json)	the tokenizer
SHA256SUMS	checksums for every file above

Provenance & fidelity

Weights are unchanged from convaiinnovations/laya-multilingual; this repo adds only the ONNX export and the serving manifest. Conversion and serving are verified against the golden standard (the official Laya SDK, CPU fp32, 43-case corpus): identical answers on CPU; on CUDA (RTX 4090) within 1e-4.

Accuracy PK — against laya-serve 0.4.7; both servers on one RTX 4090 (CUDA), graded against the same golden standard:

	Verdict	laya-serve	result
worst deviation	1e-4	1.74	Verdict ahead

License

Apache-2.0, inherited from the base model. Credit for the model itself belongs to the original authors; this repository only repackages it for serving.