Massive Activations 🥳

1️⃣ Introduction

This is a Jittor implementation of the paper: 《Massive Activations in Large Language Models》[COLM 2024]

If you have any questions about this repository, please leave a comment in the Issues section.🥰🥰🥰

2️⃣ Project Structure

Massive/
├── data/
│   ├── corpus/                # corpus for LLM
│   ├── image/                 # image for ViT
│   ├── models/                # Hugging Face model checkpoints
│   └── source/                # Figure from paper
│
├── jittor/
│   ├── jittor_llama/          # LLaMA implementation (Jittor)
│   └── jittor_gpt2/           # GPT2 implementation (Jittor)
│
├── pytorch/
│   ├── gpt-2/                 # GPT-2 related implementation
│   ├── lib/                   # shared libraries
│   ├── monkey_patch/          # tool
│   ├── main_llm.py            # entry for LLM experiments
│   └── main_vit.py            # entry for ViT experiments
│
├── results/                   # Experiment log and results
│
├── README.md
└── requirements.txt

Note

Massive/jittor 内没有任何pytorch函数，就像Massive/pytorch 内没有任何jittor函数

3️⃣ Experiment

✨ Exp1：Massive activations in LLM.

🧾 源论文实验结果
- 源图 Figure1 LLAMA-7b
- 源图 Figure2 LLAMA-13b

🔍 本人实验结果

由于论文未阐明指定激活值的检测层，故本人选用LLAMA-7b的第1、10、20、31层作为指定层。
按照原论文结论，初始层和结尾层激活值相比于中间层更低(参考原论文Figure 4及表述),本人对以上四层的实验结果也验证了该原文表述,即下图中起始层Layer-1,结尾层Layer-31相比中间层Layer-10,Layer-20的超级 激活值更少。

📈 LLAMA2-7b Layer-1
📈 LLAMA2-7b Layer-10
📈 LLAMA2-7b Layer-20
📈 LLAMA2-7b Layer-31

📈 LLAMA2-13b Layer-1
📈 LLAMA2-13b Layer-15
📈 LLAMA2-13b Layer-30
📈 LLAMA2-13b Layer-39

✒️ 本人实验日志
- Click here,Check llama2-7b exp1 logs
- Click here,Check llama2-7b exp1 logs

✨ Exp2：Largest activation magnitudes and the median magnitude at each layer in LLMs.

🧾 源论文实验结果
- 源图 Figure4

🔍 本人实验结果

下图仅展示LLAMA-7b、LLAMA-13b结果，其余LLM实现思路一致。
各模型展示三个种子(42、2026、10096)下的实验结果，保证普遍性。

📈 LLAMA-7b Result under seed of {42,2026,10096}

📈 LLAMA-13b Result under seed of {42,2026,10096}

✒️ 本人实验日志
- Click here,Check llama2-7b exp2 logs
- Click here,Check llama2-13b exp2 logs

✨ Exp3：Intervention analysis of massive activations in LLaMA2-7B and 13B.

🧾 源论文实验结果
- 源图 Table3
🔍 本人实验结果
```
展示结果的参数网格：{seed:42,2026; layer_id:2,20}
呃，种子的选择基本没有影响,因为本身仅推理过程...
```
- 📈 LLAMA2-7b Seed=42 Layer_id=2
  
  Intervention WikiText C4 PG-19
  
  Original 5.116 7.620 8.077
  
  Set to zero inf inf inf
  
  Set to mean 5.120 7.624 8.091
- 📈 LLAMA2-7b Seed=42 Layer_id=20
  
  Intervention WikiText C4 PG-19
  
  Original 5.120 7.619 8.077
  
  Set to zero inf inf inf
  
  Set to mean 5.120 7.624 8.089
- 📈 LLAMA2-7b Seed=2026 Layer_id=2
  
  Intervention WikiText C4 PG-19
  
  Original 5.116 7.619 8.077
  
  Set to zero inf inf inf
  
  Set to mean 5.120 7.624 8.091
- 📈 LLAMA2-7b Seed=2026 Layer_id=20
  
  Intervention WikiText C4 PG-19
  
  Original 5.116 7.619 8.077
  
  Set to zero inf inf inf
  
  Set to mean 5.120 7.624 8.089
- 📈 LLAMA2-13b Seed=42 Layer_id=2
  
  Intervention WikiText C4 PG-19
  
  Original 4.573 7.043 7.0149
  
  Set to zero 655.524 699.837 589.110
  
  Set to mean 238.952 200.041 187.394
- 📈 LLAMA2-13b Seed=42 Layer_id=20
  
  Intervention WikiText C4 PG-19
  
  Original 4.573 7.043 7.015
  
  Set to zero 284.995 258.798 204.537
  
  Set to mean 4.578 7.049 7.021
- 📈 LLAMA2-13b Seed=2026 Layer_id=2
  
  Intervention WikiText C4 PG-19
  
  Original 4.573 7.043 7.014
  
  Set to zero 655.528 699.819 589.093
  
  Set to mean 238.958 200.048 187.398
- 📈 LLAMA2-13b Seed=2026 Layer_id=20
  
  Intervention WikiText C4 PG-19
  
  Original 4.573 7.043 7.014
  
  Set to zero 285.029 258.801 204.546
  
  Set to mean 4.578 7.049 7.021
✒️ 本人实验日志
- Click here,Check llama2-7b exp3 logs
- Click here,Check llama2-13b exp3 logs

✨ Exp4：Attention patterns before and after massive activations appear in LLM.

🧾 源论文实验结果
- 源图 Figure5

🔍 本人实验结果

下图仅展示LLAMA-7b、LLAMA-13b结果，其余LLM实现思路一致。
各模型展示三个种子(42、2026)下的实验结果，层选择与原文一致。

📈 LLAMA2-7b Result seed=42
📈 LLAMA2-7b Result seed=2026

📈 LLAMA2-13b Result seed=42
📈 LLAMA2-7b Result seed=2026

✒️ 本人实验日志

✨ Exp5：About the values and distributions of Activations

🧾 源论文实验结果
- 源图 Table1
- 源图 Table2

🔍 本人实验结果

下图仅展示LLAMA-7b、LLAMA-13b结果，其余LLM实现思路一致。
实验参数网格：{seed: 42, 2026; layer_id: 2, 15}
该实验结果也可以看出，起始层和结尾层的超级激活值特征不明显;下图只标出中间层的超级激活之;

📈 Table1 seed=42 layer_id=2

Model	Top 1	Top 2	Top 3	Top 4	Top 5	Top-10	Top-100	Top 1%	Top 10%	median
LLaMA2-7B	2483.4600	1464.2400	769.9925	461.7525	160.8406	54.4975	3.8934	0.1077	0.0535	0.0188
LLaMA2-13B	91.8762	80.0900	40.9753	34.9484	31.4344	7.0844	1.8285	0.1203	0.0608	0.0212

📉 Table2 seed=42 layer_id=2

Model	Top 1	Top 2	Top 1%	Top 10%	Median
LLaMA2-7B	2483.4600 ± 201.6950	-1464.2400 ± 118.3621	-0.0131 ± 0.1075	0.0021 ± 0.0538	-0.0008 ± 0.0189
LLaMA2-13B	-91.8762 ± 5.3607	-80.0900 ± 12.7943	-0.0025 ± 0.1209	-0.0049 ± 0.0609	0.0023 ± 0.0212

📈 Table1 seed=42 layer_id=15

Model	Top 1	Top 2	Top 3	Top 4	Top 5	Top-10	Top-100	Top 1%	Top 10%	median
LLaMA2-7B	2497.5700	1472.3900	782.7950	468.3200	154.7988	46.8253	9.9369	0.9760	0.5701	0.2266
LLaMA2-13B	1271.2700	787.0900	70.1469	64.9359	53.2294	40.5575	11.7909	1.0746	0.6311	0.2511

📉 Table2 seed=42 layer_id=15

Model	Top 1	Top 2	Top 1%	Top 10%	Median
LLaMA2-7B	2497.5700 ± 201.8074	-1472.3900 ± 118.5022	0.0571 ± 0.9794	0.0459 ± 0.571	-0.0009 ± 0.2278
LLaMA2-13B	-1271.2700 ± 12.3049	-787.0900 ± 44.7056	-0.1714 ± 1.0663	0.0274 ± 0.6338	-0.0097 ± 0.2522

📈 Table1 seed=2026 layer_id=2

Model	Top 1	Top 2	Top 3	Top 4	Top 5	Top-10	Top-100	Top 1%	Top 10%	median
LLaMA2-7B	2529.5000	1491.1800	769.8650	459.7175	159.7562	52.9612	3.9493	0.1078	0.0535	0.0188
LLaMA2-13B	92.2856	79.5378	41.0412	35.2935	31.1086	7.1793	1.8688	0.1207	0.0609	0.0212

📉 Table2 seed=2026 layer_id=2

Model	Top 1	Top 2	Top 1%	Top 10%	Median
LLaMA2-7B	2529.5000 ± 190.8036	-1491.1800 ± 111.8119	0.0062 ± 0.1082	0.0032 ± 0.0537	-0.0004 ± 0.0189
LLaMA2-13B	92.2856 ± 3.5321	-79.5378 ± 11.9228	-0.0105 ± 0.1208	0.0072 ± 0.0608	-0.0049 ± 0.0207

📈 Table1 seed=2026 layer_id=15

Model	Top 1	Top 2	Top 3	Top 4	Top 5	Top-10	Top-100	Top 1%	Top 10%	median
LLaMA2-7B	2543.7400	1499.3700	782.6300	466.2175	154.0350	44.9828	9.8195	0.9751	0.5694	0.2262
LLaMA2-13B	1270.3000	781.8750	58.4928	53.0266	41.4734	33.2675	11.6262	1.0717	0.6295	0.2504

📉 Table2 seed=2026 layer_id=15

Model	Top 1	Top 2	Top 1%	Top 10%	Median
LLaMA2-7B	2543.7400 ± 190.9264	-1499.3700 ± 111.9376	0.1537 ± 0.9680	-0.0349 ± 0.5714	-0.0182 ± 0.2267
LLaMA2-13B	-1270.3000 ± 13.3132	-781.8750 ± 7.7075	-0.0203 ± 1.0770	0.0507 ± 0.6308	-0.0445 ± 0.2477

✒️ 本人实验日志
- Click here,Check llama2-7b exp5 logs
- Click here,Check llama2-13b exp5 logs

✨ Exp6：Activation trajectory starting from input hidden states to query, key and value states.

🧾 源论文实验结果
- 源图 Figure7
🔍 本人实验结果
```
LLAMA2-7b 模型
实验参数网格：{seed: 42, 2026; layer_id: 3, 15}
```
- 📈 LLAMA2-7b Result seed=42 layer_id=3
- 📈 LLAMA2-7b Result seed=42 layer_id=15
- 📈 LLAMA2-7b Result seed=2026 layer_id=3
- 📈 LLAMA2-7b Result seed=2026 layer_id=15
- 📈 LLAMA2-13b Result seed=42 layer_id=3
- 📈 LLAMA2-13b Result seed=42 layer_id=15
- 📈 LLAMA2-13b Result seed=2026 layer_id=3
- 📈 LLAMA2-13b Result seed=2026 layer_id=15
✒️ 本人实验日志
- Click here,Check llama2-7b exp6 logs
- Click here,Check llama2-13b exp6 logs

✨ Exp7：Explicit Attention Biases Eliminate Massive Activations.

🧾 源论文实验结果
- 源图 Figure9
- 源图 Figure10
🔍 本人实验结果
```
模型: GPT-2,基于nano GPT.
实验参数网格:{layer_id: 5, 8; sample_seq: "Summer is warm. Winter is cold.", "I love computer science and computer vision."}
```
- 📈 Sample_seq="Summer is warm. Winter is cold."
- 📈 Sample_seq="I love computer science and computer vision."
- 📈 LLAMA2-7b Result seed=2026
✒️ 本人实验日志
- Click here,Check gpt2_default exp logs
- Click here,Check gpt2_attn_bias exp logs

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Massive Activations 🥳

1️⃣ Introduction

2️⃣ Project Structure

3️⃣ Experiment

✨ Exp1：Massive activations in LLM.

✨ Exp2：Largest activation magnitudes and the median magnitude at each layer in LLMs.

✨ Exp3：Intervention analysis of massive activations in LLaMA2-7B and 13B.

✨ Exp4：Attention patterns before and after massive activations appear in LLM.

✨ Exp5：About the values and distributions of Activations

✨ Exp6：Activation trajectory starting from input hidden states to query, key and value states.

✨ Exp7：Explicit Attention Biases Eliminate Massive Activations.

About

Contributors

Languages

Name		Name	Last commit message	Last commit date
Latest commit History 17 Commits
data		data
jittor		jittor
pytorch		pytorch
results/jittor		results/jittor
.gitignore		.gitignore
README.md		README.md
requirements.txt		requirements.txt

Intervention	WikiText	C4	PG-19
Original	5.116	7.620	8.077
Set to zero	inf	inf	inf
Set to mean	5.120	7.624	8.091

Intervention	WikiText	C4	PG-19
Original	5.120	7.619	8.077
Set to zero	inf	inf	inf
Set to mean	5.120	7.624	8.089

Intervention	WikiText	C4	PG-19
Original	4.573	7.043	7.0149
Set to zero	655.524	699.837	589.110
Set to mean	238.952	200.041	187.394

Folders and files

Latest commit

History

Repository files navigation

Massive Activations 🥳

1️⃣ Introduction

2️⃣ Project Structure

3️⃣ Experiment

✨ Exp1：Massive activations in LLM.

✨ Exp2：Largest activation magnitudes and the median magnitude at each layer in LLMs.

✨ Exp3：Intervention analysis of massive activations in LLaMA2-7B and 13B.

✨ Exp4：Attention patterns before and after massive activations appear in LLM.

✨ Exp5：About the values and distributions of Activations

✨ Exp6：Activation trajectory starting from input hidden states to query, key and value states.

✨ Exp7：Explicit Attention Biases Eliminate Massive Activations.

About

Topics

Resources

Stars

Watchers

Forks

Contributors

Languages