vtava/Laya-Integrated-Memory-V23-Decision

Integrated Memory V2.3 — Standalone Decision Model

This repository contains the complete exported model, not an adapter. It does not require the laya Python package for inference.

Architecture

• Source teacher: convaiinnovations/laya-typed-decisions
• Replacement: TinyCeNN Integrated Memory V2.3
• All source full_attention layers replaced: [0, 3, 6, 9, 12, 15, 18, 21, 24, 27]
• Sliding-attention layers remain unchanged
• Export format: tinycenn-standalone-decision-v23

Fast evaluation

• Teacher/student decision agreement: 53.12%
• Student speedup vs teacher: 0.96x

See standalone_config.json for the complete validation and fast-evaluation metadata.

Install

pip install torch transformers huggingface_hub safetensors

Load directly from Hugging Face

import sys
import torch
from huggingface_hub import snapshot_download

model_dir = snapshot_download("vtava/Laya-Integrated-Memory-V23-Decision")
sys.path.insert(0, model_dir)
from standalone_decision import load_standalone

runtime = load_standalone(
    model_dir,
    device="cuda" if torch.cuda.is_available() else "cpu",
)

result = runtime.decide(
    state={"speed_kmh": 61, "track_error": -0.16, "curve": "left"},
    actions={
        "left": "steer left",
        "straight": "hold steering",
        "right": "steer right",
        "brake": "reduce speed",
    },
    instruction="Choose the safest next control action.",
)
print(result)

You can also run example_usage.py from this model repository.