ti3x-m/laya-multilingual-onnx

multilingual Laya ONNX

Standard ONNX Runtime files are under onnx/. Browser-specific graphs with large gather tensors split for WebGPU are under onnxruntime/webgpu/.

The graph uses fixed batch size 1 with dynamic sequence length and option count. Run multiple questions as separate inference calls. See manifest.json for pinned source revisions, checksums, and validation status. Standard and transformed FP32 graphs passed CPU parity checks; browser WebGPU execution has not been validated. Pure FP16-compute exports are unavailable because conversion produced an invalid graph.

Weight-only INT8 (CPU / WASM target)

onnxruntime/wasm/int8-block-64/model.onnx uses symmetric 8-bit encoder MatMul weights, block size 64, FP32 activations and decision-head computation. Exactly representable embedding/head weights use FP16 storage with FP32 compute; this is not an FP16-compute model. Recipe informed by laya-web.

Quantized encoder MatMuls: 24. First/last encoder layers retained at higher precision: 8 each. Graph plus external weights: 617.68 MB.

CPU conversion fidelity against 34 current PyTorch cases: argmax agreement 100.0%, max calibrated probability error 0.019774, mean KL 0.000105079. These are parity fixtures, not a task accuracy benchmark. Browser WASM execution has not been validated. This variant is not for WebGPU. Use ONNX Runtime 1.30.0 for the tested CPU path; load the external data file beside the graph. See the accompanying reference, quantization, and validation JSON files.

Conversion fidelity by weight variant

These results measure reproduction of PyTorch outputs, not task accuracy.

Weights	Tested backend	Reference cases	Result
Standard FP32	ONNX Runtime CPU	34	Logit parity passed (atol/rtol 0.001); 100.0% argmax agreement; max probability drift 1.90215e-06
WebGPU-transformed FP32	ONNX Runtime CPU only	3	Logit parity passed (atol/rtol 0.001); browser untested
Mixed INT8 weight-only / FP16 storage, FP32 compute	ONNX Runtime CPU	34	100.0% argmax agreement; max probability drift 0.019774; mean KL 0.000105079

FP16-compute weights are not available. The mixed variant's exact retained layers and storage conversions are listed in quantization.json.

Reproduce the conversion-fidelity test

Download this repository, then run from onnxruntime/wasm/int8-block-64/:

python -m pip install numpy==2.5.3 onnxruntime==1.30.0
python evaluate.py --standard ../../../onnx/model.onnx

Evaluator · PyTorch references and inputs · Per-case results · Quantization recipe · All attempts.

Requires Python 3.12. The evaluator exits nonzero if a gate fails. Argmax agreement measures agreement with PyTorch, not correctness against human labels. Probability error is measured after the checkpoint's temperature calibration; mean KL measures distribution drift. These 34 synthetic cases cover question types, cardinalities, multilingual text, empty state and truncation. They have no ground-truth labels. The precision fallback was selected on this same set, so these results do not establish held-out accuracy or calibration quality. To regenerate PyTorch references from the pinned checkpoint, use laya-models int8 --model multilingual from laya-webgpu.