Retnet 410m XATL

🚀 Hybrid RetNet

This is a RetNet model that accompanies the paper "Cross-Architecture Transfer Learning for Linear-Cost Inference Transformers". Instead of training new Linear-Cost Inference models from scratch, this work transfers shared weight components from other PTLMs.

🚀 Quick Start

Use the following code to start using the model:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

torch.set_default_device("cuda")

model = AutoModelForCausalLM.from_pretrained("NucleusAI/RetNet-410m-XATL", torch_dtype="auto", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("NucleusAI/RetNet-410m-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)

✨ Features

In this work, we proposed not to train new Linear-Cost Inference models (e.g., RetNet) from scratch but to transfer shared weight components from other PTLMs. The model's input/output embeddings, MLP weights, Layer Norms, Attention Output Projections ($W_O$) have been transferred from pythia-410m. For more details, please refer to the paper.

📚 Documentation

Model Details

Model Description

Model Sources

🔧 Technical Details

The model has been trained with the pile_dedup dataset to facilitate comparison with the same-sized pythia models.

📄 License

This model is released under the Apache-2.0 license.