How to understand `no` variable in cuda code?

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

How to understand `no` variable in cuda code? #234

Open yxchng opened 3 months ago

yxchng commented 3 months ago

e.g.

F no = max(o, u + k[ii]);
F A = exp(o - no);
F B = exp(u + k[ii] - no);

There are many parts of the code that uses the no variable like this. I couldn't understand it as it is not in the equations mentioned in the paper. Why minus no? Can you kindly explain?

BlinkDL commented 3 months ago

see "RWKV-4 improvements" in https://github.com/BlinkDL/RWKV-LM readme