Skip to content

[reward_manager] fix: guard against empty responses and None overlong_buffer_cfg #6484

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

Open
imitater-dou wants to merge 2 commits into
base: main
Choose a base branch
from
Open

[reward_manager] fix: guard against empty responses and None overlong_buffer_cfg #6484

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9dcf37e
[reward_manager] fix: guard against empty responses and None overlong…
imitater-dou May 26, 2026
31d143a
[reward_manager] fix: use .item() to avoid implicit CPU-GPU sync in r…
imitater-dou May 26, 2026
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: 3 additions & 2 deletions verl/workers/reward_manager/dapo.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -118,7 +118,7 @@ def __call__(self, data: DataProto, return_dict: bool = False):

reward = score

if self.overlong_buffer_cfg.enable:
if self.overlong_buffer_cfg is not None and self.overlong_buffer_cfg.enable:
overlong_buffer_len = self.overlong_buffer_cfg.len
expected_len = self.max_resp_len - overlong_buffer_len
exceed_len = valid_response_length - expected_len
Expand All @@ -129,7 +129,8 @@ def __call__(self, data: DataProto, return_dict: bool = False):
reward_extra_info["overlong_reward"].append(overlong_reward)
reward_extra_info["overlong"].append(overlong_reward < 0)

reward_tensor[i, valid_response_length - 1] = reward
if valid_response_length.item() > 0:
reward_tensor[i, valid_response_length.item() - 1] = reward

if data_source not in already_print_data_sources:
already_print_data_sources[data_source] = 0
Expand Down
3 changes: 2 additions & 1 deletion verl/workers/reward_manager/naive.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -97,7 +97,8 @@ def __call__(self, data: DataProto, return_dict: bool = False) -> torch.Tensor |
else:
reward = score

reward_tensor[i, valid_response_length - 1] = reward
if valid_response_length.item() > 0:
reward_tensor[i, valid_response_length.item() - 1] = reward

if data_source not in already_print_data_sources:
already_print_data_sources[data_source] = 0
Expand Down
3 changes: 2 additions & 1 deletion verl/workers/reward_manager/prime.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -174,7 +174,8 @@ def __call__(self, data: DataProto, return_dict: bool = False) -> torch.Tensor |

for i in range(len(data)):
data_source = data_sources[i]
reward_tensor[i, valid_response_length[i].item() - 1] = scores[i]
if valid_response_length[i].item() > 0:
reward_tensor[i, valid_response_length[i].item() - 1] = scores[i]

if data_source not in already_print_data_sources:
already_print_data_sources[data_source] = 0
Expand Down