Massive Activations in Large Language Models

Official PyTorch implementation of our paper:

Massive Activations in Large Language Models
Mingjie Sun, Xinlei Chen, J. Zico Kolter, Zhuang Liu
Carnegie Mellon University, Meta AI Research and Bosch Center for AI
Paper - Project page

Most of the experiments in this paper were done on one A6000 GPU.

---

This paper studies the existence of massive activations in Large Language Models (LLMs). These activations have significantly larger magnitudes than other activations while on the other hand are extremely few in quantity.

This repository

Setup

Installation instructions can be found in INSTALL.md.

Outline

The contents of this repository are as follows:

• lib contains the util function for loading models, plotting figures and evaluation.
• monkey_patch contains the code for monkey patching LLMs with custom forward function, with a goal of collecting internal activation and attention statistics.
• gpt-2 contains the code for training GPT-2 with explicit attention biases.
• main_llm.py contains the code for reproducing our experiments on LLMs.
• main_vit.py contains the code for reproducing our experiments on ViTs.

Large Language Models (LLMs)

• We provide an example command to visualize a hidden state feature on the residual stream:

CUDA_VISIBLE_DEVICES=0 python main_llm.py --model llama2_7b \
    --exp1 --layer_id 2 \
    --savedir results/llm/3d_feat_vis/

Running this command will visualize the output feature of layer 2 in LLaMA-2-7B, on the input prompt "Summer is warm. Winter is cold.\n". The resulting visualizations are saved in results/llm/3d_feat_vis/.

For some LLMs, e.g., LLaMA2-7B, you need to set the argument --access-token in order to access the weights.

• We provide an example command to visualize the layerwise top 3 largest activation magnitudes:

CUDA_VISIBLE_DEVICES=0 python main_llm.py --model llama2_7b \
    --exp2 \
    --savedir results/llm/layerwise/

Running this command will visualize the per layer top activation magnitudes. The resulting visualizations are saved in results/llm/layerwise.

• We provide an example command to run the intervention analysis:

CUDA_VISIBLE_DEVICES=0 python main_llm.py --model llama2_7b \
    --exp3 \
    --reset_type set_zero \
    --layer_id 2 \
    --savedir results/llm/intervention_analysis/

Here the argument --reset_type can be either set_zero or set_mean. This command will zero the massive activations in the output feature of layer 2 in LLaMA-2-7B. The evaluation results are saved in results/llm/intervention_analysis.

• We provide an example command for attention visualization:

CUDA_VISIBLE_DEVICES=0 python main_llm.py --model llama2_7b \
    --exp4 \
    --layer_id 3 \
    --savedir results/llm/attn_vis/

Running this command will visualize the attention logits (average over attention heads) in layer 3 of LLaMA-2-7B. The visualizations are saved in results/llm/attn_vis/.

Vision Transformers (ViTs)

• We provide an example command for visualizing the activation magnitudes of the output feature of an intermediate layer:

CUDA_VISIBLE_DEVICES=0 python main_vit.py --model_family dinov2_reg --model_size giant \
    --exp1 \
    --layer_id 40 \
    --savedir results/vit/3d_feat_vis/

• We provide an example command for visualizing the layer-wise largest activation magnitudes:

CUDA_VISIBLE_DEVICES=0 python main_vit.py --model_family dinov2_reg --model_size giant \
    --exp2 \
    --savedir results/vit/layerwise/

• For reproducing the results of Fix-Reg-Mean on DINOv2-reg, run the following commands:

for model_size in small base large giant 
do 
CUDA_VISIBLE_DEVICES=0 python main_vit.py \
    --model_family dinov2_reg --model_size ${model_size} --exp3 \
    --reg_feat_mean assets/reg_feat_mean \
    --imagenet_dir [Path to ImageNet validation set] \
    --savedir results/vit/exp4/dinov2_reg_${model_size}
done

The argument --reg_feat_mean corresponds to the directory containing the mean of the register features at all layers collected over 10k ImageNet training images with data augmentations.

Results

DINOv2-reg	ViT-S	ViT-B	ViT-L	ViT-G
Original	81.9	84.8	86.3	87.0
Fix-Reg-Mean	81.7	85.0	86.2	87.0

License

This project is released under the MIT license. Please see the LICENSE file for more information.

Reference

@article{sun2024massive,
  title={Massive Activations in Large Language Models}, 
  author={Sun, Mingjie and Chen, Xinlei and Kolter, J. Zico and Liu, Zhuang},
  year={2024},
  journal={arXiv preprint arXiv:2402.17762}
}