File size: 1,984 Bytes
ac5f02d bb32640 ac5f02d 0f48bd8 ac5f02d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 |
---
license: apache-2.0
datasets:
- EleutherAI/the_pile_deduplicated
language:
- en
---
# Hybrid RetNet
This is a hybrid architecture between self-attention based Transformer and [RetNet](https://arxiv.org/abs/2307.08621), where only the 2nd and middle layer is multi-head attention, and otherwise RetNet.
This is the model weight accompanying the paper [Cross-Architecture Transfer Learning for Linear-Cost Inference Transformers](https://arxiv.org/abs/2404.02684v1),
in which new Linear-Cost Inference models (e.g. RetNet) are not trained from scratch but transfer shared weight components from other PTLMs.
The model's input/output embeddings, MLP weights, & Layer Norms has been transferred from [pythia-1B](https://huggingface.co/EleutherAI/pythia-1b). For more detail, please refer to the paper.
## Model Details
### Model Description
- **Developed by:** NucleusAI, Sehyun Choi
- **Model type:** RetNet & Transformer Hybrid
### Model Sources
- **Repository:** [RetNet-XATL](https://github.com/syncdoth/RetNet-XATL)
- **Paper:** [Cross-Architecture Transfer Learning for Linear-Cost Inference Transformers](https://arxiv.org/abs/2404.02684v1)
## How to Get Started with the Model
Use the code below to get started with the model.
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
torch.set_default_device("cuda")
model = AutoModelForCausalLM.from_pretrained("NucleusAI/RetNet-1B-Hybrid-XATL", torch_dtype="auto", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("NucleusAI/RetNet-1B-Hybrid-XATL", trust_remote_code=True) # same as EleutherAI/pythia-1B
inputs = tokenizer("Hi there!", return_tensors="pt", return_attention_mask=False)
outputs = model.generate(**inputs, max_length=200)
text = tokenizer.batch_decode(outputs)[0]
print(text)
```
## Training Data
The model has been trained with [pile_dedup](EleutherAI/the_pile_deduplicated) dataset, in favor of comparison with the same sized pythia models.
|