The Liger-Kernel library is running slower.

[kadirnar]

trafficstars

I'm training the Orpheus-TTS model using the transformers library. To speed it up, I'm using fsdp + sdpa + compile. However, when I tried liger-kernel for further acceleration, compile doesn't work. It uses much less memory, which is truly incredible. But it trains 40 hours slower.

Train code: https://github.com/canopyai/Orpheus-TTS/tree/main/finetune

I just added this. model = AutoLigerKernelForCausalLM.from_pretrained(model_name, attn_implementation="sdpa")