Laya-1.2-speedup — V4.6 experimental option-scoring model

This is a derived model for the option-scoring part of convaiinnovations/laya-typed-decisions. The original frozen 28-layer ModernBERT encoder, FFNs, Q/K/V and RoPE are downloaded from upstream. The only redistributed weights are the post-encoder decision Transformer head in `head.safetensors`.

Training steps executed: 120. Selected head weights: pretrained_unmodified (selected checkpoint step: 0). If step is 0, training was attempted but the original pretrained head won the PROBE validation; do not describe the published head as finetuned.

Independent held-out V4.6 AUDIT (200 option decisions)

• Teacher agreement: 100.0000%
• Gold accuracy: 84.5000%
• Teacher gold accuracy: 84.5000%
• Gold accuracy drop: 0.0000%
• Mean JS: 0.00002392
• Paired T4 GPU median speed ratio: 1.1808x
• Paired T4 wall-clock median speed ratio: 1.1806x
• All gates passed: True

The '1.2' in the repository name refers to a previous T4 1.203x result. Speed is not guaranteed on other hardware, batch sizes or sequence lengths. The report covers two narrow benchmark shapes on one T4 and does not prove generally faster or more accurate behavior. The exported model returns option scores only, not the original auxiliary action logits.

Inference

```python from modeling_laya_speedup import load_laya_speedup model, tokenizer = load_laya_speedup( "vtava/Laya-1.2-speedup", device="cuda", compile_head=True )

Prepare typed-decision tensor inputs using TinyCeNN-LM

standalone_decision.build_sequence and collate_items.

with torch.inference_mode(): logits, _ = model(**batch)

```

Install the packages in `requirements.txt` first. The original pretrained model remains the deployment reference for use cases requiring action logits or until broader independent tests validate this variant.