sg-nta/llm-attention

Attention & Routing Matrices of some MoE-LLMs

Dataset structure

...
./
├─ nhap.py
├─ usage.py
├─ usage_example.ipynb
├─ outputs/
│  ├─ mixtral-8x7b/
│  │  ├─ routing_matrices.npz
│  │  ├─ attention_matrices_multihead.npz
│  │  ├─ tokenized_input.npz
│  │  └─ metadata.txt
│  ├─ olmoe-7b/
│  │  └─ ...
│  ├─ qwen-moe-a2.7b/
│  │  └─ ...
│  └─ qwen-moe-a2.7b-chat/
│     └─ ...

Top directory contain:

• outputs.zip: contain attention and routing matrices
• nhap.py: code to extract attention and routing matrices
• usage.py and usage_example.ipynb: example loading of attention and routing matrices

Each subfolder outputs/<model> contain

• attention_matrices_multihead.npz - dict {layer_id -> attn_matrix} with shape [num_heads, seq_len, seq_len]
• routing_matrices.npz - dict {layer_id -> routing_logits} with shape [seq_len, num_experts]
• tokenized_input.npz - dict with input_ids store token ids of input.
• metadata.txt - model/context information

Requirements

• python 3.10
• numpy 1.24.0
• pytorch 2.4.0
• transformers 4.57.0
• datasets 4.1.1

To use

• Download this dataset

hf download sg-nta/llm-attention --repo-type dataset

• unzip outputs.zip

unzip outputs.zip

• See usage.py and usage_example.ipynb for example of loading matrices from outputs/