Skip to content

Merge IFlow_XLUC20251119310498000001598005_1 to main #1

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
blackcode1 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Merge IFlow_XLUC20251119310498000001598005_1 to main #1

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
10 changes: 8 additions & 2 deletions roll/configs/base_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -309,6 +309,11 @@ class PPOConfig(BaseConfig):
default_factory=WorkerConfig, metadata={"help": "Configuration for the reference role."}
)

use_reference: bool = field(
default=True,
metadata={"help": "Whether to use reference model for KL divergence computation. If False, reference model will not be initialized."}
)

async_generation_ratio: float = field(
default=0,
metadata={
Expand Down Expand Up @@ -389,11 +394,12 @@ def __post_init__(self):
if (
self.actor_train.model_args.model_name_or_path is None
or self.actor_infer.model_args.model_name_or_path is None
or self.reference.model_args.model_name_or_path is None
or (self.use_reference and self.reference.model_args.model_name_or_path is None)
):
self.actor_train.model_args.model_name_or_path = self.pretrain
self.actor_infer.model_args.model_name_or_path = self.pretrain
self.reference.model_args.model_name_or_path = self.pretrain
if self.use_reference:
self.reference.model_args.model_name_or_path = self.pretrain

if self.critic.model_args.model_name_or_path is None:
self.critic.model_args.model_name_or_path = self.reward_pretrain
Expand Down
16 changes: 10 additions & 6 deletions roll/pipeline/rlvr/rlvr_math_vlm_pipeline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -222,7 +222,7 @@ def __init__(self, pipeline_config: RLVRConfig):
worker_config=self.pipeline_config.actor_infer,
)
# use unwrapped model as reference for lora training
if not self.is_lora:
if not self.is_lora and self.pipeline_config.use_reference:
self.reference: Any = Cluster(
name=self.pipeline_config.reference.name,
worker_cls=self.pipeline_config.reference.worker_cls,
Expand Down Expand Up @@ -361,15 +361,19 @@ def run(self):

with Timer(name="cal_ref_log_probs_reward", logger=None) as cal_timer:
if self.is_lora:
batch.meta_info["disable_adapter"] = True
batch.meta_info["is_offload_states"] = False
ref_log_probs_refs: List[ray.ObjectRef] = self.actor_train.compute_log_probs(
batch, blocking=False
)
else:
ref_log_probs_refs: List[ray.ObjectRef] = self.reference.compute_log_probs(
batch, blocking=False
)
if self.pipeline_config.use_reference:
ref_log_probs_refs: List[ray.ObjectRef] = self.reference.compute_log_probs(
batch, blocking=False
)
else:
# When reference model is disabled, use actor's own log probabilities as reference
ref_log_probs_refs: List[ray.ObjectRef] = self.actor_infer.compute_log_probs(
batch, blocking=False
)
rewards_refs: List[ray.ObjectRef] = self.reward.compute_rewards(batch, blocking=False)

ref_log_probs = DataProto.materialize_concat(data_refs=ref_log_probs_refs)
Expand Down
26 changes: 15 additions & 11 deletions roll/pipeline/rlvr/rlvr_pipeline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -216,7 +216,7 @@ def __init__(self, pipeline_config: RLVRConfig):
)
download_clusters = [self.actor_train, self.actor_infer]
# use unwrapped model as reference for lora training
if not self.is_lora:
if not self.is_lora and self.pipeline_config.use_reference:
self.reference: Any = Cluster(
name=self.pipeline_config.reference.name,
worker_cls=self.pipeline_config.reference.worker_cls,
Expand Down Expand Up @@ -544,16 +544,20 @@ def run(self):
batch.meta_info["is_offload_states"] = False
ref_log_probs = self.actor_train.compute_log_probs(batch, blocking=True)
else:
if self.pipeline_config.reference.use_dynamic_batching_in_infer:
batch, dynamic_batching_metrics = dynamic_batching_shard(
batch,
self.reference.dp_size,
self.pipeline_config.reference.max_tokens_per_microbatch_in_infer,
self.pipeline_config.reference.sequence_length_round_in_infer,
"reference/compute_log_probs",
)
metrics_mgr.add_metrics(dynamic_batching_metrics)
ref_log_probs = self.reference.compute_log_probs(batch, blocking=True)
if self.pipeline_config.use_reference:
if self.pipeline_config.reference.use_dynamic_batching_in_infer:
batch, dynamic_batching_metrics = dynamic_batching_shard(
batch,
self.reference.dp_size,
self.pipeline_config.reference.max_tokens_per_microbatch_in_infer,
self.pipeline_config.reference.sequence_length_round_in_infer,
"reference/compute_log_probs",
)
metrics_mgr.add_metrics(dynamic_batching_metrics)
ref_log_probs = self.reference.compute_log_probs(batch, blocking=True)
else:
# When reference model is disabled, use actor's own log probabilities as reference
ref_log_probs = self.actor_infer.compute_log_probs(batch, blocking=True)
metrics_mgr.add_reduced_metrics(ref_log_probs.meta_info.pop("metrics", {}))
ref_log_probs.rename(old_keys="log_probs", new_keys="ref_log_probs")
batch = batch.union(ref_log_probs)
Expand Down
23 changes: 15 additions & 8 deletions roll/pipeline/rlvr/rlvr_vlm_pipeline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -313,13 +313,16 @@ def __init__(self, pipeline_config: RLVRConfig):
resource_manager=self.resource_manager,
worker_config=self.pipeline_config.actor_infer,
)
self.reference: Any = Cluster(
name=self.pipeline_config.reference.name,
worker_cls=self.pipeline_config.reference.worker_cls,
resource_manager=self.resource_manager,
worker_config=self.pipeline_config.reference,
)
download_clusters = [self.actor_train, self.actor_infer, self.reference]
if self.pipeline_config.use_reference:
self.reference: Any = Cluster(
name=self.pipeline_config.reference.name,
worker_cls=self.pipeline_config.reference.worker_cls,
resource_manager=self.resource_manager,
worker_config=self.pipeline_config.reference,
)
download_clusters = [self.actor_train, self.actor_infer, self.reference]
else:
download_clusters = [self.actor_train, self.actor_infer]
if self.pipeline_config.adv_estimator == "gae":
self.critic: Any = Cluster(
name=self.pipeline_config.critic.name,
Expand Down Expand Up @@ -542,7 +545,11 @@ def run(self):
batch.meta_info["_broadcast_non_tensor_batch"]= True

with Timer(name="cal_ref_log_probs", logger=None) as cal_ref_log_probs_timer:
ref_log_probs = self.reference.compute_log_probs(batch, blocking=True)
if self.pipeline_config.use_reference:
ref_log_probs = self.reference.compute_log_probs(batch, blocking=True)
else:
# When reference model is disabled, use actor's own log probabilities as reference
ref_log_probs = self.actor_infer.compute_log_probs(batch, blocking=True)
metrics_mgr.add_reduced_metrics(ref_log_probs.meta_info.pop("metrics", {}))
ref_log_probs.rename(old_keys="log_probs", new_keys="ref_log_probs")
batch = batch.union(ref_log_probs)
Expand Down