ttkciar@alien.topB to LocalLLaMA@poweruser.forumEnglish · 10 months agoPractical Tips for Finetuning LLMs Using LoRA (Low-Rank Adaptation)magazine.sebastianraschka.comexternal-linkmessage-square1fedilinkarrow-up11arrow-down10cross-posted to: machinelearning@academy.garden
arrow-up11arrow-down1external-linkPractical Tips for Finetuning LLMs Using LoRA (Low-Rank Adaptation)magazine.sebastianraschka.comttkciar@alien.topB to LocalLLaMA@poweruser.forumEnglish · 10 months agomessage-square1fedilinkcross-posted to: machinelearning@academy.garden
minus-squareRelevant_Outcome_726@alien.topBlinkfedilinkEnglisharrow-up1·10 months agoFrom my experience, Here are some other things related to Lora: + FSDP doesn’t work for Lora because FSDP requires all parameters to be trainable or frozen. + For Qlora, currently we can only use deepspeed zero2 (deepspeed zero 3 is not supported)
From my experience, Here are some other things related to Lora:
+ FSDP doesn’t work for Lora because FSDP requires all parameters to be trainable or frozen.
+ For Qlora, currently we can only use deepspeed zero2 (deepspeed zero 3 is not supported)