YingLong model is introduced in this paper. This version is pre-trained on 78B time points. More details can be found at our github and on our project page.

Quickstart

bash

pip install xformers transformers
pip install flash-attn --no-build-isolation
git clone https://github.com/Dao-AILab/flash-attention && cd flash-attention
cd csrc/rotary && pip install .
cd ../layer_norm && pip install .

The flash attention is not required. If you use V100 or other GPU doesn't support flash attention, just change the FlashAttention2Available = RequirementCache("flash-attn>=2.0.0.post1") to FlashAttention2Available = False in the model.py file. It should be able to run.

python

import torch
from transformers import AutoModelForCausalLM

# load pretrain model
model = AutoModelForCausalLM.from_pretrained('qcw2333/YingLong_110m', trust_remote_code=True,torch_dtype=torch.bfloat16).cuda()

# prepare input
batch_size, lookback_length = 1, 2880
seqs = torch.randn(batch_size, lookback_length).bfloat16().cuda()

# generate forecast
prediction_length = 96
output = model.generate(seqs, future_token=prediction_length)

print(output.shape)

A notebook example is also provided here. The sample codes for long-term forecasting tasks and gift-eval tasks are provided at link.

Citation

Coming soon...

Contact

If you have any questions or want to use the code, feel free to contact:

Xue Wang (xue.w@alibaba-inc.com)

Tian Zhou (tian.zt@alibaba-inc.com)

License

This model is licensed under the cc-by-4.0 License.