I use LoRA adapter, but the following message appears: "llama_lora_adapter_set: flash_attn is not compatible with LoRA"

[WangHao311]

When I use llama-server to start the service, I use the "-fa" parameter. When I start the log and call the interface "/v1/chat/completions", the prompt message "llama_lora_adapter_set: flash_attn is not compatible with LoRA appears". Is there a problem? How can I modify it?

[ggerganov]

I am not sure why FA is not allowed with LoRA. It seems to me that they are both compatible. cc @ngxson

[ngxson]

Hmm yeah I'm not sure why I added this check in the first place. I removed it and it still works with llama model. I'll make a PR for it.

[WangHao311]

Thanks for your reply. Is there any way I can try it? I am using Qwen2 model.