Skip to content

perf: Reduce memory footprint for ChunkedDistribuedLogProb #1895

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
nujoug wants to merge 2 commits into
base: main
Choose a base branch
from
Open

perf: Reduce memory footprint for ChunkedDistribuedLogProb #1895

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
f10e4e4
Reduce memory footprint for ChunkedDistributedLogProb
nujoug Feb 6, 2026
ac29781
Explicity set the dtype for gradient tensor
nujoug Feb 6, 2026
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
19 changes: 13 additions & 6 deletions nemo_rl/distributed/model_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -220,7 +220,9 @@ def backward(
seq_size = int(vocab_parallel_logits.shape[1])
num_chunks = (seq_size + chunk_size - 1) // chunk_size

all_grad_input = []
grad_input: torch.Tensor = torch.empty_like(
vocab_parallel_logits, dtype=torch.float32
)

for chunk_idx in range(num_chunks):
chunk_start = chunk_idx * chunk_size
Expand All @@ -243,13 +245,18 @@ def backward(
num_classes=partition_vocab_size,
)

grad_input = is_chosen.float().sub_(softmax_output)

grad_input.mul_(grad_output[:, chunk_start:chunk_end].unsqueeze(dim=-1))
# Inplace index into the preallocated grad_input tensor
grad_input_chunk = grad_input[:, chunk_start:chunk_end, :]

all_grad_input.append(grad_input)
grad_input_chunk.copy_(
is_chosen.float().sub_(softmax_output)
) # inplace copy
grad_input_chunk.mul_(
grad_output[:, chunk_start:chunk_end].unsqueeze(dim=-1)
)

grad_input = torch.cat(all_grad_input, dim=1)
# Explicitly free before next iteration allocates
del softmax_output, is_chosen, logits

# if you add an argument to the forward method, then you must add a corresponding None here
return grad_input, None, None, None, None, None, None
Expand Down
Loading