Skip to content

[megatron] fix: set model to eval during compute_log_prob/compute_values #4708

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
wuxibin89 merged 1 commit into from
Dec 29, 2025
Merged

[megatron] fix: set model to eval during compute_log_prob/compute_values #4708

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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: 5 additions & 0 deletions verl/workers/actor/megatron_actor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -196,6 +196,9 @@ def compute_log_prob(self, data: DataProto, calculate_entropy=False) -> torch.Te
Returns:
DataProto: torch.Tensor: the log_prob tensor
"""
prev_modes = [m.training for m in self.actor_module]
for module in self.actor_module:
module.eval()
use_dynamic_bsz = data.meta_info.get("use_dynamic_bsz", False)
micro_batch_size = data.meta_info.get("micro_batch_size", None)
max_token_len = data.meta_info.get("max_token_len", None)
Expand Down Expand Up @@ -306,6 +309,8 @@ def compute_logprobs_fn(output, data, use_dynamic_bsz=False, indices=None):
# add empty cache after each compute
get_torch_device().empty_cache()

for module, mode in zip(self.actor_module, prev_modes, strict=False):
module.train(mode)
return log_probs, entropys, layers_topk_idx

def make_minibatch_iterator(self, data: DataProto) -> Iterable[DataProto]:
Expand Down
5 changes: 5 additions & 0 deletions verl/workers/critic/megatron_critic.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,6 +87,9 @@ def _validate_config(self, config) -> None:

@GPUMemoryLogger("megatron critic", logger=logger)
def compute_values(self, data: DataProto) -> DataProto:
prev_modes = [m.training for m in self.critic_module]
for module in self.critic_module:
module.eval()
responses = data.batch["responses"]
attention_mask = data.batch["attention_mask"]
use_dynamic_bsz = data.meta_info.get("use_dynamic_bsz", False)
Expand Down Expand Up @@ -139,6 +142,8 @@ def compute_values(self, data: DataProto) -> DataProto:
# add empty cache after each compute
get_torch_device().empty_cache()

for module, mode in zip(self.critic_module, prev_modes, strict=False):
module.train(mode)
return values

def make_minibatch_iterator(self, data: DataProto) -> Iterable[DataProto]:
Expand Down
Loading