Skip to content

[megatron] feat: use yaml to manage mbridge args #4584

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
Kite0011 wants to merge 15 commits into
base: main
Choose a base branch
from
Open

[megatron] feat: use yaml to manage mbridge args #4584

Show file tree
Hide file tree
Changes from 10 commits
Commits
Show all changes
15 commits
Select commit Hold shift + click to select a range
6a98545
add mbridge config
Kite0011 Dec 18, 2025
7ca6247
use general inferface for mbridge save_weights
Kite0011 Dec 18, 2025
3e7b555
fint fix
Kite0011 Dec 18, 2025
0611191
add distributed_filesystem arg
Kite0011 Dec 18, 2025
0079597
add default value
Kite0011 Dec 18, 2025
2c6c967
update memory_efficient default value
Kite0011 Dec 18, 2025
0eb6a9d
Update verl/utils/checkpoint/megatron_checkpoint_manager.py
Kite0011 Dec 18, 2025
a704639
Update verl/utils/checkpoint/megatron_checkpoint_manager.py
Kite0011 Dec 18, 2025
59d46f9
Update verl/trainer/config/config.py
Kite0011 Dec 18, 2025
e85bd9a
fix lint
Kite0011 Dec 18, 2025
63c19a3
update mbridge setting
Kite0011 Dec 19, 2025
71bc169
fix bug
Kite0011 Dec 19, 2025
87565d0
fix lint
Kite0011 Dec 19, 2025
ec9aa98
update code
Kite0011 Dec 25, 2025
e2b1e06
fix ci
Kite0011 Dec 31, 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
6 changes: 6 additions & 0 deletions verl/trainer/config/_generated_ppo_megatron_trainer.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -97,6 +97,9 @@ actor_rollout_ref:
- extra
load_contents: ${.save_contents}
async_save: false
mbridge_config:
memory_efficient: true
distributed_filesystem: true
use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
profiler:
_target_: verl.utils.profiler.ProfilerConfig
Expand Down Expand Up @@ -491,6 +494,9 @@ critic:
- extra
load_contents: ${.save_contents}
async_save: false
mbridge_config:
memory_efficient: true
distributed_filesystem: true
profiler:
_target_: verl.utils.profiler.ProfilerConfig
tool: ${oc.select:global_profiler.tool,null}
Expand Down
6 changes: 6 additions & 0 deletions verl/trainer/config/_generated_ppo_trainer.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -84,6 +84,9 @@ actor_rollout_ref:
- extra
load_contents: ${.save_contents}
async_save: false
mbridge_config:
memory_efficient: true
distributed_filesystem: true
use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
profiler:
_target_: verl.utils.profiler.ProfilerConfig
Expand Down Expand Up @@ -425,6 +428,9 @@ critic:
- extra
load_contents: ${.save_contents}
async_save: false
mbridge_config:
memory_efficient: true
distributed_filesystem: true
profiler:
_target_: verl.utils.profiler.ProfilerConfig
tool: ${oc.select:global_profiler.tool,null}
Expand Down
5 changes: 5 additions & 0 deletions verl/trainer/config/actor/actor.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -122,6 +122,11 @@ checkpoint:

# Whether to save checkpoints asynchronously. Only effective for Megatron as of now.
async_save: False

# Mbridge config extension.
mbridge_config:
Copy link
Collaborator

@ISEEKYAN ISEEKYAN Dec 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

please don't add these yaml configs.

memory_efficient: true
distributed_filesystem: true

# optimizer configs
optim:
Expand Down
3 changes: 3 additions & 0 deletions verl/trainer/config/config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,6 +36,9 @@ class CheckpointConfig(BaseConfig):
save_contents: list[str] = field(default_factory=lambda: ["model", "optimizer", "extra"])
load_contents: list[str] = field(default_factory=lambda: ["model", "optimizer", "extra"])
async_save: bool = False
mbridge_config: dict[str, Any] = field(
default_factory=lambda: {"memory_efficient": True, "distributed_filesystem": True}
)


@dataclass
Expand Down
5 changes: 5 additions & 0 deletions verl/trainer/config/critic/critic.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -98,6 +98,11 @@ checkpoint:
# Whether to save checkpoints asynchronously. Only effective for Megatron as of now.
async_save: False

# Mbridge config extension.
mbridge_config:
memory_efficient: true
distributed_filesystem: true

# profile the critic model in `update_critic`
profiler:

Expand Down
4 changes: 4 additions & 0 deletions verl/trainer/config/sft_trainer_engine.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,6 +52,10 @@ checkpoint:

# For more flexibility, you can specify the contents to load from the checkpoint.
load_contents: ${checkpoint.save_contents}
# Mbridge config extension.
mbridge_config:
memory_efficient: true
distributed_filesystem: true

trainer:
default_local_dir: checkpoints/${trainer.project_name}/${trainer.experiment_name}
Expand Down
6 changes: 4 additions & 2 deletions verl/utils/checkpoint/megatron_checkpoint_manager.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -500,7 +500,7 @@ def save_checkpoint(self, local_path: str, hdfs_path: str = None, global_step: i
hf_ckpt_path = get_hf_model_checkpoint_path(local_path)
if self.vanilla_bridge:
self.bridge.save_weights(
Copy link
Collaborator

@ISEEKYAN ISEEKYAN Dec 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

recommend to modify like this:


import inspect
extended_args = {}
if "distributed_filesystem" in inspect.signature(self.bridge.save_weights):
    extended_args["distributed_filesystem"] = True
    extended_args["memory_efficient"] = True
self.bridge.save_weights(self.model, hf_ckpt_path, **extended_args)

self.model, hf_ckpt_path, distributed_filesystem=True, memory_efficient=True
self.model, hf_ckpt_path, **(getattr(self.checkpoint_config, "mbridge_config", None) or {})
)
else:
self.bridge.save_hf_weights(self.model, hf_ckpt_path)
Expand Down Expand Up @@ -572,7 +572,9 @@ def save_checkpoint(self, local_path: str, hdfs_path: str = None, global_step: i
hf_model_ckpt_path = get_hf_model_checkpoint_path(local_path)
if self.vanilla_bridge:
self.bridge.save_weights(
self.model, hf_model_ckpt_path, distributed_filesystem=True, memory_efficient=True
self.model,
Copy link
Collaborator

@ISEEKYAN ISEEKYAN Dec 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

same as above

hf_model_ckpt_path,
**(getattr(self.checkpoint_config, "mbridge_config", None) or {}),
)
else:
self.bridge.save_hf_weights(self.model, hf_model_ckpt_path)
Expand Down
Loading