Skip to content

[trainer] refactor: move rollout log to inheritable trainer #3576

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 23, 2025
Merged

[trainer] refactor: move rollout log to inheritable trainer #3576

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
ed47967
move log rollout to function
ccclyu Sep 23, 2025
b88a3eb
copy reward info dict
ccclyu Sep 23, 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: 5 additions & 0 deletions recipe/dapo/dapo_ray_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -332,6 +332,11 @@ def fit(self):
actor_output_metrics = reduce_metrics(actor_output.meta_info["metrics"])
metrics.update(actor_output_metrics)

# Log rollout generations if enabled
rollout_data_dir = self.config.trainer.get("rollout_data_dir", None)
if rollout_data_dir:
self._log_rollout_data(batch, reward_extra_infos_dict, timing_raw, rollout_data_dir)

# validate
if (
self.val_reward_fn is not None
Expand Down
17 changes: 1 addition & 16 deletions recipe/one_step_off_policy/ray_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -552,22 +552,7 @@ def fit(self):
# Log rollout generations if enabled
rollout_data_dir = self.config.trainer.get("rollout_data_dir", None)
if rollout_data_dir:
with marked_timer("dump_rollout_generations", timing_raw, color="green"):
inputs = self.tokenizer.batch_decode(batch.batch["prompts"], skip_special_tokens=True)
outputs = self.tokenizer.batch_decode(batch.batch["responses"], skip_special_tokens=True)
scores = batch.batch["token_level_scores"].sum(-1).cpu().tolist()
sample_gts = [
item.non_tensor_batch.get("reward_model", {}).get("ground_truth", None) for item in batch
]

self._dump_generations(
inputs=inputs,
outputs=outputs,
gts=sample_gts,
scores=scores,
reward_extra_infos_dict=reward_extra_infos_dict,
dump_path=rollout_data_dir,
)
self._log_rollout_data(batch, reward_extra_infos_dict, timing_raw, rollout_data_dir)

# validate
if (
Expand Down
17 changes: 1 addition & 16 deletions recipe/sppo/sppo_ray_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -308,22 +308,7 @@ def fit(self):
# Log rollout generations if enabled
rollout_data_dir = self.config.trainer.get("rollout_data_dir", None)
if rollout_data_dir:
with simple_timer("dump_rollout_generations", timing_raw):
print(batch.batch.keys())
inputs = self.tokenizer.batch_decode(batch.batch["prompts"], skip_special_tokens=True)
outputs = self.tokenizer.batch_decode(batch.batch["responses"], skip_special_tokens=True)
scores = batch.batch["token_level_scores"].sum(-1).cpu().tolist()
sample_gts = [
item.non_tensor_batch.get("reward_model", {}).get("ground_truth", None) for item in batch
]
self._dump_generations(
inputs=inputs,
outputs=outputs,
scores=scores,
reward_extra_infos_dict=reward_extra_infos_dict,
gts=sample_gts,
dump_path=rollout_data_dir,
)
self._log_rollout_data(batch, reward_extra_infos_dict, timing_raw, rollout_data_dir)

# validate
if (
Expand Down
56 changes: 33 additions & 23 deletions verl/trainer/ppo/ray_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -441,6 +441,38 @@ def _dump_generations(self, inputs, outputs, gts, scores, reward_extra_infos_dic

print(f"Dumped generations to {filename}")

def _log_rollout_data(
self, batch: DataProto, reward_extra_infos_dict: dict, timing_raw: dict, rollout_data_dir: str
):
"""Log rollout data to disk.
Args:
batch (DataProto): The batch containing rollout data
reward_extra_infos_dict (dict): Additional reward information to log
timing_raw (dict): Timing information for profiling
rollout_data_dir (str): Directory path to save the rollout data
"""
with marked_timer("dump_rollout_generations", timing_raw, color="green"):
inputs = self.tokenizer.batch_decode(batch.batch["prompts"], skip_special_tokens=True)
outputs = self.tokenizer.batch_decode(batch.batch["responses"], skip_special_tokens=True)
scores = batch.batch["token_level_scores"].sum(-1).cpu().tolist()
sample_gts = [item.non_tensor_batch.get("reward_model", {}).get("ground_truth", None) for item in batch]

reward_extra_infos_to_dump = reward_extra_infos_dict.copy()
if "request_id" in batch.non_tensor_batch:
reward_extra_infos_dict.setdefault(
"request_id",
batch.non_tensor_batch["request_id"].tolist(),
)

self._dump_generations(
inputs=inputs,
outputs=outputs,
gts=sample_gts,
scores=scores,
reward_extra_infos_dict=reward_extra_infos_to_dump,
dump_path=rollout_data_dir,
)

def _maybe_log_val_generations(self, inputs, outputs, scores):
"""Log a table of validation samples to the configured logger (wandb or swanlab)"""

Expand Down Expand Up @@ -1111,29 +1143,7 @@ def fit(self):
# Log rollout generations if enabled
rollout_data_dir = self.config.trainer.get("rollout_data_dir", None)
if rollout_data_dir:
with marked_timer("dump_rollout_generations", timing_raw, color="green"):
inputs = self.tokenizer.batch_decode(batch.batch["prompts"], skip_special_tokens=True)
outputs = self.tokenizer.batch_decode(batch.batch["responses"], skip_special_tokens=True)
scores = batch.batch["token_level_scores"].sum(-1).cpu().tolist()
sample_gts = [
item.non_tensor_batch.get("reward_model", {}).get("ground_truth", None)
for item in batch
]

if "request_id" in batch.non_tensor_batch:
reward_extra_infos_dict.setdefault(
"request_id",
batch.non_tensor_batch["request_id"].tolist(),
)

self._dump_generations(
inputs=inputs,
outputs=outputs,
gts=sample_gts,
scores=scores,
reward_extra_infos_dict=reward_extra_infos_dict,
dump_path=rollout_data_dir,
)
self._log_rollout_data(batch, reward_extra_infos_dict, timing_raw, rollout_data_dir)

# validate
if (
Expand Down
Loading