[optimization] support 8bit optims from bistandbytes (#163)

* support 8bit optims from bnb.

* fix exist_ok

* use_8bit_bnb.

* fix.

* note in readme.
This commit is contained in:
Sayak Paul
2024-12-30 14:03:56 +05:30
committed by GitHub
parent e9fba4ab7f
commit 191ea8b0cc
4 changed files with 12 additions and 2 deletions
+2 -1
View File
@@ -506,6 +506,7 @@ class Trainer:
beta3=self.args.beta3,
epsilon=self.args.epsilon,
weight_decay=self.args.weight_decay,
use_8bit=self.args.use_8bit_bnb,
use_deepspeed=use_deepspeed_opt,
)
@@ -1003,7 +1004,7 @@ class Trainer:
if self.args.push_to_hub:
repo_id = self.args.hub_model_id or Path(self.args.output_dir).name
self.state.repo_id = create_repo(token=self.args.hub_token, repo_id=repo_id).repo_id
self.state.repo_id = create_repo(token=self.args.hub_token, repo_id=repo_id, exist_ok=True).repo_id
def _move_components_to_device(self):
if self.text_encoder is not None: