Skip to content

Allow uneven shards for multi-GPU inference in vllm backend #494

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
SahilJain314 merged 2 commits into from
Jun 27, 2025
Merged

Allow uneven shards for multi-GPU inference in vllm backend #494

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
19a83ff
allow uneven shards in generate_text
KiddoZhu Jun 9, 2025
98d682a
Merge branch 'main' into zhaochengz/multi-gpu_eval
SahilJain314 Jun 26, 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
7 changes: 3 additions & 4 deletions nemo_rl/models/generation/vllm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1381,12 +1381,11 @@ def generate_text(
f"data must be a BatchedDataDict, got type: {type(data)}"
)

# Get total batch size
batch_size = len(data["prompts"])

# Shard the data across the tied worker groups
dp_size = self.sharding_annotations.get_axis_size("data_parallel")
sharded_data = data.shard_by_batch_size(dp_size, batch_size=batch_size)
sharded_data: list[SlicedDataDict] = data.shard_by_batch_size(
dp_size, allow_uneven_shards=True
)
future_bundle = self.worker_group.run_all_workers_sharded_data(
"generate_text",
sharded_data,
Expand Down
Loading