https://github.com/BlinkDL/LinearAttentionArena

Here we will test various linear attention designs.
https://github.com/BlinkDL/LinearAttentionArena

Last synced: about 1 month ago
JSON representation

Here we will test various linear attention designs.

• Host: GitHub
• URL: https://github.com/BlinkDL/LinearAttentionArena
• Owner: BlinkDL
• License: apache-2.0
• Created: 2024-04-05T10:55:43.000Z (5 months ago)
• Default Branch: main
• Last Pushed: 2024-04-25T09:35:47.000Z (5 months ago)
• Last Synced: 2024-05-30T05:18:20.099Z (4 months ago)
• Language: Python
• Homepage:
• Size: 229 KB
• Stars: 50
• Watchers: 8
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# LinearAttentionArena

Here we will test various linear attention designs.

```

pip install pytorch-lightning==1.9.5 torch deepspeed wandb ninja --upgrade

```

RWKV-6.0b differences (vs RWKV-6.0): GroupNorm => LayerNorm, and remove "gate" in TimeMix, so the params count is lower.

```

# Example: RWKV-6.0b L12-D768 (189M params) on 4x4090, minipile 1.5B tokens loss 2.812

./prepare.sh --model_type "x060b" --layer 12 --emb 768 --ctx_len 512 --suffix "-0"

./train.sh --model_type "x060b" --layer 12 --emb 768 --lr_init "6e-4" --lr_final "2e-4" --ctx_len 512 --n_gpu 4 --m_bsz 32 --grad_cp 0 --save_period 1000 --suffix "-0"

```

```

# Example: Mamba L12-D768 (191M params) on 4x4090, minipile 1.5B tokens loss 2.885

./prepare.sh --model_type "mamba" --layer 12 --emb 768 --ctx_len 512 --suffix "-0"

./train.sh --model_type "mamba" --layer 12 --emb 768 --lr_init "6e-4" --lr_final "2e-4" --ctx_len 512 --n_gpu 4 --m_bsz 32 --grad_cp 0 --save_period 1000 --suffix "-0"

```

![rwkv-x060b-mamba](rwkv-x060b-mamba.png)