Skip to content

[Unified Checkpoint] update async_save_info in develop #9173

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Sep 26, 2024
Merged

[Unified Checkpoint] update async_save_info in develop #9173

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d7ae1d0
update async_save_info
DesmonDay Sep 23, 2024
5e1c691
Merge branch 'develop' of https://github.com/PaddlePaddle/PaddleNLP i…
DesmonDay Sep 23, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
29 changes: 15 additions & 14 deletions paddlenlp/trainer/trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2297,16 +2297,7 @@ def save_model(self, output_dir: Optional[str] = None, merge_tensor_parallel: Op
self.model_wrapped.get_all_parameters(convert2cpu=True)

if self.args.should_save_model_state:
unified_checkpoint_config_backup = self.args.unified_checkpoint_config
# backup and remove unified_checkpoint_config for not trine stage
if not self.is_in_train:
self.args.unified_checkpoint_config = []

self._save(output_dir=output_dir, merge_tensor_parallel=merge_tensor_parallel)

# recover unified_checkpoint_config for not trine stage
if not self.is_in_train:
self.args.unified_checkpoint_config = unified_checkpoint_config_backup
else:
if self.args.unified_checkpoint and "async_save" in self.args.unified_checkpoint_config:
os.makedirs(output_dir, exist_ok=True)
Expand Down Expand Up @@ -2584,10 +2575,9 @@ def _save(self, output_dir: Optional[str] = None, state_dict=None, merge_tensor_
# Save a trained model and configuration using `save_pretrained()`.
# They can then be reloaded using `from_pretrained()`

local_rank = int(os.getenv("PADDLE_RANK_IN_NODE", 0))
if (
strtobool(os.getenv("FLAG_LLM_PDC", "False"))
and local_rank == 0
and paddle.distributed.get_rank() == 0
and self.args.unified_checkpoint
and "async_save" in self.args.unified_checkpoint_config
):
Expand All @@ -2598,9 +2588,10 @@ def _save(self, output_dir: Optional[str] = None, state_dict=None, merge_tensor_
"ignore_save_lr_and_optim": self.args.ignore_save_lr_and_optim,
"skip_save_model_weight": "skip_save_model_weight" in self.args.unified_checkpoint_config,
}
if not os.path.exists(os.path.join(self.args.logging_dir, "async_save_info.json")):
with open(os.path.join(self.args.logging_dir, "async_save_info.json"), "w") as f:
json.dump(save_info, f)
if os.path.exists(os.path.join(self.args.logging_dir, "async_save_info.json")): # afs cannot overwrite
os.remove(os.path.join(self.args.logging_dir, "async_save_info.json"))
with open(os.path.join(self.args.logging_dir, "async_save_info.json"), "w") as f:
json.dump(save_info, f)

if self.args.should_save:
if self.tokenizer is not None:
Expand All @@ -2609,7 +2600,17 @@ def _save(self, output_dir: Optional[str] = None, state_dict=None, merge_tensor_
paddle.save(self.args, os.path.join(output_dir, TRAINING_ARGS_NAME))

if self.args.unified_checkpoint:
unified_checkpoint_config_backup = self.args.unified_checkpoint_config
# backup and remove unified_checkpoint_config for not trine stage
if not self.is_in_train:
self.args.unified_checkpoint_config = []

self.unified_checkpoint_handler.save_unified_checkpoint(self.model, self.optimizer, output_dir)

# recover unified_checkpoint_config for not trine stage
if not self.is_in_train:
self.args.unified_checkpoint_config = unified_checkpoint_config_backup

return

merge_tensor_parallel = merge_tensor_parallel and self.args.use_hybrid_parallel
Expand Down