Skip to content

Fix nemo 1 packed sequence TE version error #11874

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
cuichenx merged 5 commits into from
Jan 21, 2025
Merged

Fix nemo 1 packed sequence TE version error #11874

Changes from 1 commit
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
82f376c
fix nemo 1 packed sequence TE version error
cuichenx Jan 16, 2025
2178ea1
limit condition to only CP=1 case
cuichenx Jan 17, 2025
6609a3b
Merge branch 'main' into chcui/fix-nemo1-packedseq
cuichenx Jan 17, 2025
6e41169
add cudnn version guard
cuichenx Jan 17, 2025
1bd68ad
Apply isort and black reformatting
cuichenx Jan 17, 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
5 changes: 4 additions & 1 deletion nemo/collections/nlp/models/language_modeling/megatron_gpt_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,6 +102,7 @@
drain_embedding_wgrad_compute,
get_model_config,
init_method_normal,
is_te_min_version,
scaled_init_method_normal,
)

Expand Down Expand Up @@ -1366,7 +1367,9 @@ def fwd_output_and_loss_func(dataloader_iter, model, checkpoint_activations_all_
'attention_mask': None if self.get_attention_mask_from_fusion else batch['attention_mask'],
'labels': batch['labels'] if 'labels' in batch else None,
}

if not is_te_min_version("1.13", check_equality=False):
# cu_seqlens_unpadded != cu_seqlens is not supported in 1.13 or earlier
cu_seqlens_unpadded = cu_seqlens
forward_args['packed_seq_params'] = PackedSeqParams(
cu_seqlens_q=cu_seqlens_unpadded,
cu_seqlens_kv=cu_seqlens_unpadded,
Expand Down
Loading