pprp
commented
3 months ago hi, sorry for the late reply.
For lora fine-tuning, we just employ the same method as Wanda, which is not taking the sparsity into consideration.
For sparse lora-finetuning, you can refer to :
- LoRAPrune: Structured Pruning Meets Low-Rank Parameter-Efficient Fine-Tuning
- Sparse Fine-tuning for Inference Acceleration of Large Language Models
- APT: Adaptive Pruning and Tuning Pretrained Language Models for Efficient Training and Inference (ICML24 Oral)
Hello @pprp,
The lora fine-tuned models cannot be merged as the lora branch is dense. This makes the LoRA fine-tuning not very useful to regain performance lost during pruning. Please let me know if there is any sparse lora fine-tuning that you are employing.