Update README.md

README.md

@@ -10,10 +10,6 @@ RWKV-3 1.5B on A40 (tf32) = always 0.015 sec/token, tested using simple pytorch
 
 GPT2-XL 1.3B on A40 (tf32) = 0.032 sec/token (for ctxlen 1000), tested using HF, GPU utilization 45% too (interesting), VRAM 9655M
 
-How it works: RWKV gathers information to a number of channels, which are also decaying with different speeds as you move to the next token. It's very simple once you understand it.
-
-**RWKV is parallelizable because the time-decay of each channel is data-independent (and trainable)**. For example, in usual RNN you can adjust the time-decay of a channel from say 0.8 to 0.5 (these are called "gates"), while in RWKV you simply move the information from a W-0.8-channel to a W-0.5-channel to achieve the same effect. Moreover, you can fine-tune RWKV into a non-parallelizable RNN (then you can use outputs of later layers of the previous token) if you want extra performance.
-
 ## Join our Discord: https://discord.gg/bDSBUMeFpc :)
 
 You are welcome to join the RWKV discord https://discord.gg/bDSBUMeFpc to build upon it. We have plenty of potential compute (A100 40Gs) now (thanks to CoreWeave), so if you have interesting ideas I can run them.
@@ -24,6 +20,10 @@ I am training RWKV-3 on the Pile (https://github.com/BlinkDL/RWKV-v2-RNN-Pile):
 
 All of the trained models will be open-source. Inference is very fast (only matrix-vector multiplications, no matrix-matrix multiplications) even on CPUs, so you can even run a LLM on your phone.
 
+How it works: RWKV gathers information to a number of channels, which are also decaying with different speeds as you move to the next token. It's very simple once you understand it.
+
+**RWKV is parallelizable because the time-decay of each channel is data-independent (and trainable)**. For example, in usual RNN you can adjust the time-decay of a channel from say 0.8 to 0.5 (these are called "gates"), while in RWKV you simply move the information from a W-0.8-channel to a W-0.5-channel to achieve the same effect. Moreover, you can fine-tune RWKV into a non-parallelizable RNN (then you can use outputs of later layers of the previous token) if you want extra performance.
+
 Here are some of my TODOs. Let's work together :)
 
 * Now we have RWKV-4 with DeepSpeedStage2 & FP16 & Better CUDA Kernel (100% faster training than tf32): https://github.com/BlinkDL/RWKV-LM/tree/main/RWKV-v4. It will be great if someone can take a look to make it support multinode and Stage3.