emozilla/landmark-llama-7b

This model is an out-of-the-box ready version of the LLaMA-7B variant of Landmark Attention. The original code is modified from the Landmark GitHub and the weights from here.

As a LLaMA variant, this model may be subject to the LLaMA license.

To use

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
import torch

tokenizer = AutoTokenizer.from_pretrained("emozilla/landmark-llama-7b", use_fast=False)
model = AutoModelForCausalLM.from_pretrained("emozilla/landmark-llama-7b", \
  torch_dtype=torch.bfloat16, trust_remote_code=True, device_map="auto")
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)

print(pipe("Somebody once told me the world is gonna roll me", \
           max_new_tokens=256, temperature=0.8, do_sample=True))

You can configure the Landmark parameters by editing mem_freq, mem_top_k, mem_max_seq_len, and mem_max_cache_size.

config = AutoConfig.from_pretrained("emozilla/landmark-llama-7b", trust_remote_code=True)
config.mem_top_k = 6
model = AutoModelForCausalLM.from_pretrained("emozilla/landmark-llama-7b", \
  torch_dtype=torch.bfloat16, trust_remote_code=True, device_map="auto", config=config)