LCKV

This is a research-purpose pretrained model described in paper "Layer-Condensed KV Cache for Efficient Inference of Large Language Models".

About

Layer-Condensed KV Cache (LCKV) is a variant of transformer decoders in which queries of all layers are paired with keys and values of just the top layer. It reduces the memory and computation cost, reduces the number of parameters, significantly improves the inference throughput with comparable or better task performance. See more details in our github repo: https://github.com/whyNLP/LCKV

Quick Start

# Use a pipeline as a high-level helper
from transformers import pipeline
pipe = pipeline("text-generation", model="whynlp/tinyllama-lckv-w2-ft-100b", trust_remote_code=True)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("whynlp/tinyllama-lckv-w2-ft-100b", trust_remote_code=True)

Sample text generation script:

# This is consistent with the `run_generation.py` script in the github repo: https://github.com/whyNLP/LCKV
import torch
from accelerate.utils import set_seed

from transformers import pipeline


set_seed(42)

pipe = pipeline(
    "text-generation",
    model="whynlp/tinyllama-lckv-w2-ft-100b",
    torch_dtype=torch.bfloat16,
    device="cuda",
    trust_remote_code=True,
    model_kwargs={"attn_implementation": "flash_attention_2"},
)

response = pipe(
    "the meaning of life is",
    add_special_tokens=False,
    max_new_tokens=50,
    temperature=1.0,
    top_k=0,
    top_p=0.9,
    repetition_penalty=1.0,
    do_sample=True,
)

print(response[0]["generated_text"])
# the meaning of life is the tension that this presence gives rise to each moment of the thought to let live out the moment of my appearance. For Sarkar, sense is what has also forgotten: It is forgets.
# On kiu3/ this is and

The LCKV Collection

The model has 2 warmup layers. i.e. 3/22 KV cache of a standard TinyLlama.

This model was first initialized from the TinyLlama 2.5T checkpoint, then continued pre-training on 100B tokens from SlimPajama.

Since the model structure has been changed, the initialization cannot inherit the performance of the TinyLlama checkpoint, but it effectively boosts the training process compared to pre-training from scratch.

The evaluation follows that of TinyLlama. Refer to our paper for more details.