NX-AI / xLSTM-7b

xLSTM-7B

This xLSTM-7B was pre-trained on the DCLM and selected high-quality data for in a total of approx. 2.3 T tokens using the xlstm-jax framework.

How to use it

First, install xlstm , which now uses the mlstm_kernels package for triton kernels:

pip install xlstm
pip install mlstm_kernels

For now, install the transformers repositiory fork from NX-AI (until it is merged):

pip install 'transformers @ git+ssh://[email protected]/NX-AI/transformers.git@integrate_xlstm'

Use this model as:

from transformers import AutoModelForCausalLM, AutoTokenizer

xlstm = AutoModelForCausalLM.from_pretrained("NX-AI/xLSTM-7b", device_map="auto")

# this is a fork of EleutherAI/gpt-neox-20b
tokenizer = AutoTokenizer.from_pretrained("NX-AI/xLSTM-7b")

tokens = tokenizer("Hello xLSTM, how are you doing?", return_tensors='pt')['input_ids'].to(device="cuda")

out = xlstm.generate(tokens, max_new_tokens=20)

print(tokenizer.decode(out[0]))

Speed results

Generation Speed using torch.cuda.graph and torch.compile optimizations on one NVIDIA H100:

Performance

Using HuggingFace's lm_eval :

Using HuggingFace's lighteval in the Leaderboard-v1 settings:

License

NXAI Community License (see LICENSE file)