TDoSiMa/sima-laya

Decision models (Laya and CLM), compiled for the SiMa.ai MLSoC Modalix

Laya by Convai Innovations is a "System-1" decision model: an encoder (ModernBERT-large, or mmBERT-base for the multilingual one) with a decision head that answers a typed question about a piece of text in one forward pass. It picks one of several options, scores on a scale, or says yes or no.

These are those checkpoints compiled for the Modalix MLA. Every graph is a single MLA stage, with no layers on the CPU, and a decision takes about 19 ms at 128 tokens (8 ms for the multilingual model).

folder	model	precision	latency per decision, by tokens	same decision as PyTorch	size
general	Laya	BF16	128: 19.5 ms, 256: 35.2 ms, 512: 87.5 ms	100 / 100	3.27 GB
general-int8	Laya INT8	A_BF16_W_INT8	128: 14 ms	98 / 100	0.67 GB
typed-decisions	Laya typed-decisions	BF16	128: 19.4 ms, 512: 87.5 ms, 1024: 282 ms	99 / 100	4.35 GB
multilingual	Laya multilingual	BF16	128: 8.1 ms, 256: 15.3 ms, 512: 33.9 ms, 1024: 116.5 ms	100 / 100	2.61 GB
clm	CLM v0.1 8B	A_BF16_W_INT8	128: 254 ms	9 of 13 reference decisions (int8 weights; an approximation)	9.12 GB
dino	Laya-dino	BF16	64: 18 ms	30 / 30 game states	0.95 GB

clm is a second kind of decision model, CLM v0.1 8B by Contrastive-LM: a frozen Qwen3-8B embeds the state and each option, and two small projection heads compare them. Its encoder is compiled as a chain of 18 graphs with int8 weights (7.25 GB on the MLA), and its latency above is one pass through them, which holds a question's state and all its options: about 0.3 s for a new question, nothing for one seen before. With int8 weights it is an approximation of the PyTorch model (embedding cosine about 0.99; 9 of 13 reference decisions the same), and a one-token option is embedded poorly.

Latency is one decision on a Modalix DevKit, measured inside the runtime. "Same decision" is against the fp32 PyTorch model on 100 decisions with the 128-token graph. Precision BF16 is weights and activations in bfloat16; A_BF16_W_INT8 keeps bfloat16 activations and stores the weights as int8.

What is in a folder

file	what it is
laya_s<N>_stage1_mla.elf	the compiled graph for sequences of up to N tokens
token_embeddings.bf16	the embedding table, looked up on the CPU
act_tail.f32	the last layer of the act head, run on the CPU
tokenizer.json	the checkpoint's tokenizer
laya_config.json	token ids, calibration temperatures, and which graphs there are

and in clm:

file	what it is
qwen_s128_l<NN>n2_stage1_mla.elf	two layers of the encoder, from layer NN: 18 graphs that run one after another
token_embeddings.bf16	Qwen3-8B's embedding table, looked up on the CPU
heads.f32, heads.json	CLM's state and action heads, run on the CPU, and their layout
tokenizer.json	Qwen3-8B's tokenizer
clm_config.json	which graphs there are, in order

models.json lists every folder with file sizes and SHA-256 sums.

Using them

The files run on a Modalix board with the runtime and web app from neat-decision-studio: its Models page downloads a model from this repository onto the board and loads it on the MLA. By hand:

hf download TDoSiMa/sima-laya --include "general/*" --local-dir .
./laya run general --state "We were billed twice for March." \
    --question '{"type": "noul", "instructions": "Is this a billing problem?"}'

They are not usable with PyTorch or ONNX Runtime; for that, use the original checkpoints at convaiinnovations/laya.

License and attribution

Apache-2.0. The models are Convai Innovations' Laya checkpoints (https://github.com/NandhaKishorM/laya, Apache-2.0), compiled without retraining. The exception is dino, whose decision head was fine-tuned on top of the English checkpoint for a browser game. Zero-shot quality is the checkpoints': the port reproduces PyTorch's answers, including its wrong ones.

clm holds the projection heads of Contrastive-LM's CLM-v0.1-8B (https://github.com/Contrastive-LM/CLM, Apache-2.0), unchanged, and Alibaba Cloud's Qwen3-8B (https://huggingface.co/Qwen/Qwen3-8B, Apache-2.0): its embedding table and tokenizer unchanged, and its 36 decoder layers compiled for the MLA with their weights rounded to int8 and no language-model head.