ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

refactor pissa, improve llamaboard

Browse Source 
Former-commit-id: 619556e46c19718f702c97df5d570a2a4c5fb13a
This commit is contained in:
hiyouga
2024-06-28 01:04:24 +08:00
parent edc7498111
commit 46f0189e88
16 changed files with 219 additions and 216 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
src/llamafactory/hparams/finetuning_args.py
View File

@@ -379,10 +379,10 @@ class FinetuningArguments(FreezeArguments, LoraArguments, RLHFArguments, GaloreA
if self.loraplus_lr_ratio is not None and self.finetuning_type != "lora":
raise ValueError("`loraplus_lr_ratio` is only valid for LoRA training.")
if self.pissa_convert and self.finetuning_type != "lora":
raise ValueError("`pissa_convert` is only valid for LoRA training.")
if self.pissa_init and self.finetuning_type != "lora":
raise ValueError("`pissa_init` is only valid for LoRA training.")
if self.pissa_convert and (self.stage in ["rm", "ppo", "kto"] or self.use_ref_model):
if self.pissa_init and (self.stage in ["ppo", "kto"] or self.use_ref_model):
raise ValueError("Cannot use PiSSA for current training stage.")
if self.train_mm_proj_only and self.finetuning_type != "full":