423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-08-23 14:22:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix save function

Browse Source 
Former-commit-id: d2f18197e379601a60fa878af975c68d7c8b9648
This commit is contained in:
hiyouga 2023-07-21 14:09:07 +08:00
parent 42c78bf591
commit daf81288a1
2 changed files with 4 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
src/llmtuner/extras/save_and_load.py
View File

@ -1,6 +1,6 @@
import os import os
import torch import torch
from typing import Dict from typing import Dict, Optional
from transformers.trainer import WEIGHTS_NAME, WEIGHTS_INDEX_NAME from transformers.trainer import WEIGHTS_NAME, WEIGHTS_INDEX_NAME
from transformers.modeling_utils import load_sharded_checkpoint from transformers.modeling_utils import load_sharded_checkpoint
@ -12,12 +12,12 @@ from llmtuner.extras.logging import get_logger
logger = get_logger(__name__) logger = get_logger(__name__)
def get_state_dict(model: torch.nn.Module) -> Dict[str, torch.Tensor]: # get state dict containing trainable parameters def get_state_dict(model: torch.nn.Module, trainable_only: Optional[bool] = True) -> Dict[str, torch.Tensor]:
state_dict = model.state_dict() state_dict = model.state_dict()
filtered_state_dict = {} filtered_state_dict = {}
for k, v in model.named_parameters(): for k, v in model.named_parameters():
if v.requires_grad: if (not trainable_only) or v.requires_grad:
filtered_state_dict[k] = state_dict[k].cpu().clone().detach() filtered_state_dict[k] = state_dict[k].cpu().clone().detach()
return filtered_state_dict return filtered_state_dict

2
src/llmtuner/tuner/core/trainer.py
View File

@ -56,7 +56,7 @@ class PeftTrainer(Seq2SeqTrainer):
backbone_model.config.use_cache = True backbone_model.config.use_cache = True
backbone_model.save_pretrained( backbone_model.save_pretrained(
output_dir, output_dir,
state_dict=get_state_dict(backbone_model), state_dict=get_state_dict(backbone_model, trainable_only=(self.finetuning_args.finetuning_type != "full")),
safe_serialization=self.args.save_safetensors safe_serialization=self.args.save_safetensors
) )
backbone_model.config.use_cache = False backbone_model.config.use_cache = False