Skip to content

[data] switch pretrain from broadcast to replicated load #491

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
ananthsub merged 1 commit into from
Aug 26, 2025
Merged

[data] switch pretrain from broadcast to replicated load #491

Show file tree
Hide file tree
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
3 changes: 2 additions & 1 deletion src/megatron/bridge/data/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,8 +69,9 @@ def pretrain_train_valid_test_datasets_provider(

print_rank_0("> building train, validation, and test datasets for GPT ...")

# Build the dataset on all ranks for TP-replicated loading
train_ds, valid_ds, test_ds = BlendedMegatronDatasetBuilder(
dataset_type, train_val_test_num_samples, is_dataset_built_on_rank, dataset_config
dataset_type, train_val_test_num_samples, lambda: True, dataset_config
).build()

print_rank_0("> finished creating GPT datasets ...")
Expand Down
9 changes: 1 addition & 8 deletions src/megatron/bridge/training/gpt_step.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -269,14 +269,7 @@ def get_batch(
if (not parallel_state.is_pipeline_first_stage()) and (not parallel_state.is_pipeline_last_stage()):
return None, None, None, None, None, None, None, None

if isinstance(cfg.dataset, FinetuningDatasetConfig):
batch = get_batch_from_iterator(data_iterator, use_mtp)
else:
# get batches based on the TP rank you are on
batch = get_batch_on_this_tp_rank(data_iterator, cfg, use_mtp)
batch["cu_seqlens"] = None
batch["cu_seqlens_argmin"] = None
batch["max_seqlen"] = None
batch = get_batch_from_iterator(data_iterator, use_mtp)

# slice batch along sequence dimension for context parallelism
batch = get_batch_on_this_cp_rank(batch)
Expand Down
Loading