search

BlinkDL / RWKV-LM

RWKV is an RNN with transformer-level LLM performance. It can be directly trained like a GPT (parallelizable). So it's combining the best of RNN and transformer - great performance, fast inference, saves VRAM, fast training, "infinite" ctx_len, and free sentence embedding.
Apache License 2.0
11.99k stars 825 forks source link

Flash Attention #239

Closed fakerybakery closed 2 months ago

fakerybakery commented 2 months ago

Hi, Thanks for releasing RWKV! I got an error saying RWKV doesn't support Flash Attention. Is Flash Attention support planned? Thank you!

BlinkDL commented 2 months ago

Hi RWKV is attention-free so there is no need to use flash attention

fakerybakery commented 2 months ago

Sorry about that, Thank you for clarifying!