Laya-1.2-speedup-standalone

Self-contained FP32 option-scoring model derived from vtava/Laya-1.2-speedup. Includes the complete original 28-layer ModernBERT encoder, selected 64-token decision head, type embeddings, scorer, tokenizer and configuration. The source model's auxiliary action classifier is not included because V4.6 only supports option-scoring decisions.

Real measured size

• Model parameters: 421,029,889
• Full weight checkpoint: 1,684,141,004 bytes (1684.141 MB)
• Preexisting decision-head-only file: 100,772,088 bytes
• Original upstream base checkpoint: 842,609,220 bytes
• No inference-time download of either ancestor repository.

Usage

First download this one repository (once):

from huggingface_hub import snapshot_download
local_dir = snapshot_download("vtava/Laya-1.2-speedup-standalone")

Then inference, entirely offline:

from modeling_laya_standalone import load_model
model,tokenizer=load_model(local_dir,device="cuda",local_files_only=True)
answers=model.predict(tokenizer,
    {"road_direction":"left","speed":7.0},
    {"steering":{"type":"choice","instructions":"Choose road direction.",
                 "criteria":{"left":"left","straight":"straight","right":"right"}}})
print(answers)

Copy modeling_laya_standalone.py to your Python import path as needed. Requires packages from requirements.txt; no TinyCeNN-LM or base-model download during inference.

Research caveats

The original V4.6 narrow Tesla T4 audit reported 1.203× median GPU speedup, 200/200 teacher agreement and 84.5% gold accuracy relative to native option scoring. Packaging has not repeated that held-out audit or benchmarked new hardware. This export tests FP32 reload parity and FP16 predictions/probabilities on synthetic examples only. Small FP16 rounding differences are allowed only when the independent FP32 comparison also passes. The repository name does not guarantee 1.2× speedup in all contexts. Inspect the upstream model's license before redistributing weights.

Base revision: e929ae5cf69bc34259cd2f95c9e91145b818b1f0. Head revision: ac6305fb66912b25a30b1e4ab39b7034825ae71a.