This is a
RetNet model, accompanying the paper
Cross-Architecture Transfer Learning for Linear-Cost Inference Transformers,
In this work, we proposed to
not 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 ($W_O$) has been transferred from
pythia-410m. For more detail, please refer to the paper.
Use the code below to get started with the model.
1import torch
2from transformers import AutoModelForCausalLM, AutoTokenizer
3
4torch.set_default_device("cuda")
5
6model = AutoModelForCausalLM.from_pretrained("NucleusAI/RetNet-410m-XATL", torch_dtype="auto", trust_remote_code=True)
7tokenizer = AutoTokenizer.from_pretrained("NucleusAI/RetNet-410m-XATL", trust_remote_code=True) # same as EleutherAI/pythia-1B
8
9inputs = tokenizer("Hi there!", return_tensors="pt", return_attention_mask=False)
10
11outputs = model.generate(**inputs, max_length=200)
12text = tokenizer.batch_decode(outputs)[0]
13print(text)
The model has been trained with
pile_dedup dataset, in favor of comparison with the same sized pythia models.
