Skip to content

[BREAKING] feat: decouple PPO epochs from critic #1358

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
emergenz wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[BREAKING] feat: decouple PPO epochs from critic #1358

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
2 changes: 1 addition & 1 deletion examples/split_placement/config/ppo_trainer_split.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -119,7 +119,7 @@ critic:
ppo_max_token_len_per_gpu: 32768 # (${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}) * 2
forward_max_token_len_per_gpu: ${critic.ppo_max_token_len_per_gpu}
ulysses_sequence_parallel_size: 1 # sp size
ppo_epochs: ${actor_rollout_ref.actor.ppo_epochs}
epochs: 1
shuffle: ${actor_rollout_ref.actor.shuffle}
grad_clip: 1.0
cliprange_value: 0.5
Expand Down
2 changes: 1 addition & 1 deletion verl/trainer/config/ppo_megatron_trainer.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -192,7 +192,7 @@ critic:
ppo_micro_batch_size: null # will be deprecated, use ppo_micro_batch_size_per_gpu
ppo_micro_batch_size_per_gpu: null
use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
ppo_epochs: ${actor_rollout_ref.actor.ppo_epochs}
epochs: 1
data_loader_seed: ${actor_rollout_ref.actor.data_loader_seed}
shuffle: ${actor_rollout_ref.actor.shuffle}
cliprange_value: 0.5
Expand Down
2 changes: 1 addition & 1 deletion verl/trainer/config/ppo_trainer.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -160,7 +160,7 @@ critic:
ppo_max_token_len_per_gpu: 32768 # (${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}) * 2
forward_max_token_len_per_gpu: ${critic.ppo_max_token_len_per_gpu}
ulysses_sequence_parallel_size: 1 # sp size
ppo_epochs: ${actor_rollout_ref.actor.ppo_epochs}
epochs: 1
shuffle: ${actor_rollout_ref.actor.shuffle}
grad_clip: 1.0
cliprange_value: 0.5
Expand Down
2 changes: 1 addition & 1 deletion verl/workers/critic/dp_critic.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -186,7 +186,7 @@ def update_critic(self, data: DataProto):
else:
dataloader = batch.split(self.config.ppo_mini_batch_size)

for epoch in range(self.config.ppo_epochs):
for epoch in range(self.config.epochs):
for batch_idx, data in enumerate(dataloader):
# split batch into micro_batches
mini_batch = data
Expand Down
2 changes: 1 addition & 1 deletion verl/workers/critic/megatron_critic.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -123,7 +123,7 @@ def make_minibatch_iterator(self, data: DataProto) -> Iterable[DataProto]:
data = data.select(batch_keys=select_keys)
return data.make_iterator(
mini_batch_size=self.config.ppo_mini_batch_size,
epochs=self.config.ppo_epochs,
epochs=self.config.epochs,
seed=self.config.data_loader_seed,
dataloader_kwargs={"shuffle": self.config.shuffle},
)
Expand Down
2 changes: 1 addition & 1 deletion verl/workers/fsdp_workers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -921,7 +921,7 @@ def update_critic(self, data: DataProto):

global_num_tokens = data.meta_info["global_token_num"]
estimated_flops, promised_flops = self.flops_counter.estimate_flops(global_num_tokens, delta_time)
metrics["perf/mfu/critic"] = estimated_flops * self.config.ppo_epochs / promised_flops / self.world_size
metrics["perf/mfu/critic"] = estimated_flops * self.config.epochs / promised_flops / self.world_size

self.critic_lr_scheduler.step()
lr = self.critic_lr_scheduler.get_last_lr()[0]
Expand Down
2 changes: 1 addition & 1 deletion verl/workers/megatron_workers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -653,7 +653,7 @@ def update_critic(self, data: DataProto):
delta_time = timer.last
global_num_tokens = data.meta_info["global_token_num"]
estimated_flops, promised_flops = self.flops_counter.estimate_flops(global_num_tokens, delta_time)
metrics["perf/mfu/critic"] = estimated_flops * self.config.ppo_epochs / promised_flops / self.world_size
metrics["perf/mfu/critic"] = estimated_flops * self.config.epochs / promised_flops / self.world_size
output = DataProto(batch=None, meta_info={"metrics": metrics})

if self._is_offload_param:
Expand Down