Files
Qwen2.5-3B-trit-uniform-d3/README.md
ModelHub XC 4e132ccf87 初始化项目,由ModelHub XC社区提供模型
Model: Entrit/Qwen2.5-3B-trit-uniform-d3
Source: Original Platform
2026-06-15 12:42:17 +08:00

61 lines
2.0 KiB
Markdown

---
license: apache-2.0
base_model: Qwen/Qwen2.5-3B
tags:
- quantization
- ternary
- balanced-ternary
- tritllm
library_name: transformers
---
# Qwen2.5-3B-trit-uniform-d3
Balanced ternary quantization of [`Qwen/Qwen2.5-3B`](https://huggingface.co/Qwen/Qwen2.5-3B) at depth **d=3** (27 levels per weight, **5.05 bits per weight**).
Produced with the codec from **"Balanced Ternary Post-Training Quantization for Large Language Models"** (Stentzel, 2026). See [Entrit/tritllm-codec](https://huggingface.co/Entrit/tritllm-codec) for the codec source.
## Quick load
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("Entrit/Qwen2.5-3B-trit-uniform-d3")
tokenizer = AutoTokenizer.from_pretrained("Entrit/Qwen2.5-3B-trit-uniform-d3")
```
The weights are dequantized to FP16 for stock-`transformers` compatibility. The on-disk size is therefore the same as the FP16 source. The 5.05-bpw figure refers to the *information content* of the quantized matrices and is what matters for inference on hardware that consumes the packed trit format directly (see [Entrit/tritllm-kernel](https://huggingface.co/Entrit/tritllm-kernel)).
## Quantization details
| Field | Value |
|---|---|
| Source model | [`Qwen/Qwen2.5-3B`](https://huggingface.co/Qwen/Qwen2.5-3B) |
| Depth | d=3 (27 levels) |
| Bits per weight | 5.05 |
| Group size | 16 |
| Scale codebook | 27-entry log-spaced (scale_depth=3) |
| Method | Uniform PTQ |
| Quantized layers | all 2D linear matrices |
| Kept FP16 | `lm_head`, token embeddings, all `*_norm` layers |
| Codec | tritllm v2 |
## Citation
```
@article{stentzel2026ternaryptq,
title = {Balanced Ternary Post-Training Quantization for Large Language Models},
author = {Stentzel, Eric},
year = 2026,
note = {Entrit Systems}
}
```
## Reproducibility
```bash
git clone https://huggingface.co/Entrit/tritllm-codec
cd tritllm-codec
python quantize_model_v2.py --model Qwen/Qwen2.5-3B --configs uniform-d3 --out ./out
```