8000 [fix] Make checkpint directory only on rank0 to avoid race condition by BlankCheng · Pull Request #88 · LLM360/Reasoning360 · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

[fix] Make checkpint directory only on rank0 to avoid race condition #88

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 8000 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 5 commits into from
Jun 17, 2025
Merged

[fix] Make checkpint directory only on rank0 to avoid race condition #88

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
1132a79
Making checkpint directory only on rank0 to avoid race condition caus…
BlankCheng Jun 17, 2025
6906538
add megatron patch
ZYHowell Jun 17, 2025
eaa8cbc
Merge branch 'main' into zhoujun/fix-fsdp-saveckpt
ZYHowell Jun 17, 2025
7bd35ce
Update fsdp_checkpoint_manager.py
BlankCheng Jun 17, 2025
eb2e7da
Update megatron_checkpoint_manager.py
ZYHowell Jun 17, 2025
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
5 changes: 4 additions & 1 deletion verl/utils/checkpoint/fsdp_checkpoint_manager.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,7 @@
# limitations under the License.

import os
import time
import warnings
from typing import Optional, Union

Expand Down Expand Up @@ -207,7 +208,9 @@ def save_checkpoint(self, local_path: str, hdfs_path: str = None, global_step: i

if "hf_model" in self.checkpoint_contents:
hf_local_path = os.path.join(local_path, "huggingface")
os.makedirs(hf_local_path, exist_ok=True)
if self.rank == 0:
os.makedirs(hf_local_path, exist_ok=True)
torch.distributed.barrier()

# Only rank 0 will save hf model and,
# offload to cpu to save LLMs which may be too large to fit in one GPU
Expand Down
18 changes: 14 additions & 4 deletions verl/utils/checkpoint/megatron_checkpoint_manager.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -167,7 +167,8 @@ def get_checkpoint_name(
if expert_parallel:
common_path = common_path + f"_{expert_rank:03d}"

os.makedirs(common_path, exist_ok=True)
# replace os.makedirs by local_mkdir
self.local_mkdir(common_path)

if return_base_dir:
return common_path
Expand Down Expand Up @@ -230,10 +231,18 @@ def save_checkpoint(self, local_path: str, hdfs_path: str = None, global_step: i
self.remove_previous_save_local_path(self.previous_saved_paths[:keep_start])
self.previous_saved_paths = self.previous_saved_paths[keep_start:]

local_path = self.local_mkdir(local_path)
if self.rank == 0:
# NOTE: bug fix by Reasoning360: avoid multiple nodes creating the same directory
local_path = self.local_mkdir(local_path)
model_ckpt_path = get_model_checkpoint_path(local_path)
model_ckpt_path = self.local_mkdir(model_ckpt_path)
torch.distributed.barrier()

# Save Model
if "model" in self.checkpoint_contents and mpu.get_data_parallel_rank() == 0:
# NOTE: bug fix by Reasoning360: only save one copy for the CP group
if ("model" in self.checkpoint_contents and mpu.get_data_parallel_rank() == 0 and
(mpu.get_context_parallel_world_size() <= 1 or mpu.get_context_parallel_rank() == 0)
):
state_dicts = []

for vpp_rank, model in enumerate(self.model):
Expand Down 5971 Expand Up @@ -320,7 +329,8 @@ def save_checkpoint(self, local_path: str, hdfs_path: str = None, global_step: i
torch.distributed.barrier()
# NOTE: bug saving by Reasoning360: here we implicitly create a local path and multiple nodes
# may try to access it together.
rng_state_parent_path = self.local_mkdir(os.path.join(local_path, "rng_states"))
if self.rank == 0:
rng_state_parent_path = self.local_mkdir(os.path.join(local_path, "rng_states"))
torch.distributed.barrier()

rng_state_path = get_rng_states_checkpoint_path(local_path, only_rank0_save=False)
Expand Down
0