Skip to content

cp: fix(deepseek-v4): restore batch axis for packed-sequence (THD) forward (2651) into r0.5.0 #2745

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
akoumpa merged 1 commit into from
Jun 24, 2026
Merged

cp: fix(deepseek-v4): restore batch axis for packed-sequence (THD) forward (2651) into r0.5.0 #2745

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
7 changes: 7 additions & 0 deletions nemo_automodel/components/models/deepseek_v4/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -436,6 +436,13 @@ def forward(
raise ValueError("First PP stage requires input_ids or inputs_embeds")
if inputs_embeds is None:
inputs_embeds = self.embed_tokens(input_ids)
# Packed-sequence (THD) inputs arrive with the batch axis collapsed
# (``process_input_for_thd`` flattens ids to ``[T]`` -> embeds ``[T, dim]``).
# Restore the leading batch dim so the hc_mult expand below sees the
# expected ``[B, S, dim]`` rank, mirroring the output-side THD
# ``unsqueeze(0)`` in ``compute_lm_head_logits(is_thd=True)``.
if inputs_embeds.dim() == 2:
inputs_embeds = inputs_embeds.unsqueeze(0)
# Expand embeddings to hc_mult copies: [B,S,dim] -> [B,S,hc_mult,dim]
h = inputs_embeds.unsqueeze(2).expand(-1, -1, self.config.hc_mult, -1).contiguous()
shape_ref = inputs_embeds # 3D ref for rotary / mask sizing
Expand Down
Loading