---
license: apache-2.0
---

Use rwkv pip package 0.8.6+ for RWKV-5. Might overflow in fp16. Use fp32.

Very interesting. RWKV-5 is great at benchmarks (excellent zeroshot performance), but generates quite worse music (just like GPT models) despite lower loss.

This fits my theory: Dot-product is good for uncreative work, while Channelwise is good for creative work.

Therefore use https://huggingface.co/BlinkDL/rwkv-4-music for better music. Or let's see if someone can find a better sampling method to improve RWKV-5 results.