Skip to content

added pad tokens as dummy sequence #2591

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

added pad tokens as dummy sequence #2591

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
17 changes: 11 additions & 6 deletions megatron/rl/rl_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -799,13 +799,18 @@ def get_logprobs(model, tokens, position_ids, attention_mask, no_grad=False, pac
tp_world_size = get_tensor_model_parallel_world_size()
if actual_len % tp_world_size != 0:
actual_len = ((actual_len + tp_world_size - 1) // tp_world_size) * tp_world_size
# Update cu_seqlens to match the padded length.
# The last entry of cu_seqlens must equal the tensor's sequence dimension.
# Without this, TE attention/rotary ops see mismatched dimensions.
# We need to add padding tokens to make the sequence divisible by TP.
Copy link
Contributor

@yobibyte yobibyte Mar 5, 2026

Choose a reason for hiding this comment

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

If we need this to be divisible, we need to have a unit test checking this and preventing this bug in the future.

# if we extend cu_seqlens[-1] to include padding in the
# last real sequence, this would cause that sequence to attend to
# padding tokens (garbage values), leading to numerical instability/NaN.
# Instead, we add a new "padding sequence" entry to cu_seqlens.
if packed_seq_params.cu_seqlens_q[-1].item() != actual_len:
# Clone to avoid modifying cached params
new_cu_seqlens = packed_seq_params.cu_seqlens_q.clone()
new_cu_seqlens[-1] = actual_len
# Add a new entry for the padding pseudo-sequence
# cu_seqlens goes from [0, ..., total_real] to [0, ..., total_real, padded_len]
new_cu_seqlens = torch.cat([
packed_seq_params.cu_seqlens_q,
torch.tensor([actual_len], dtype=torch.int32, device=packed_seq_params.cu_seqlens_q.device)
])
packed_seq_params = PackedSeqParams(
qkv_format=packed_seq_params.qkv_format,
cu_seqlens_q=new_cu_seqlens,
Expand Down
Loading