Skip to content

feat: separately wakeup vllm to reduce peak memory in refitting #190

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

Closed
zpqiu wants to merge 1 commit into from
Closed

feat: separately wakeup vllm to reduce peak memory in refitting #190

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
3 changes: 2 additions & 1 deletion nemo_reinforcer/algorithms/grpo.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -277,9 +277,10 @@ def refit_policy_generation(
"""Refit the policy generation interface with the latest policy weights."""
policy.offload_before_refit()
ipc_handles = policy.get_weights_ipc_handles()
policy_generation.prepare_for_generation()
policy_generation.prepare_for_generation(tags=["weights"])
policy_generation.update_weights(ipc_handles)
policy.offload_after_refit()
policy_generation.prepare_for_generation(tags=["kv_cache"])


def generate_responses(
Expand Down
10 changes: 7 additions & 3 deletions nemo_reinforcer/models/generation/vllm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -410,8 +410,12 @@ def sleep(self):
gc.collect()
torch.cuda.empty_cache()

def wake_up(self):
self.llm.wake_up()
def wake_up(self, **kwargs):
# tags like ["weights", "kv_cache"]
if "tags" in kwargs:
self.llm.wake_up(tags=kwargs["tags"])
else:
self.llm.wake_up()


class VllmGeneration(GenerationInterface):
Expand Down Expand Up @@ -580,7 +584,7 @@ def prepare_for_generation(self, *args, **kwargs):
try:
# Use run_all_workers_single_data for methods that don't need data
futures = self.worker_group.run_all_workers_single_data(
"wake_up", respect_tied_workers=True
"wake_up", respect_tied_workers=True, **kwargs
)
# Wait for all futures to complete
results = ray.get(futures)
Expand Down