metadata

license: other

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]))

If you cannot or do not want to use the triton kernels, you can change them to native PyTorch implementations:

xlstm_config = AutoConfig.from_pretrained("NX-AI/xLSTM-7b")
xlstm_config.step_kernel = "native"
xlstm_config.chunkwise_kernel = "chunkwise--native_autograd"
xlstm_config.sequence_kernel = "native_sequence__native"

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

# verify selected kernels
from pprint import pprint
pprint(xlstm.backbone.blocks[0].mlstm_layer.config)

Speed results

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

Performance

Using HuggingFace's lm_eval:

BBH	MMLU-Pro	Math	MUSR	GPQA	IfEval
0.381	0.242	0.036	0.379	0.280	0.244

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

Arc-Challenge (25-shot)	MMLU (5-shot)	Hellaswag (10-shot)	Winogrande (5-shot)	TruthfulQA (0-shot)	GSM8k (5-shot)	OpenbookQA (5-shot)	PiQA (5-shot)
0.584	0.589	0.710	0.742	0.420	0.004	0.443	0.817

License

NXAI Community License (see LICENSE file)