Skip to content

save space when converting hf model to megatron model. #25950

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
LysandreJik merged 4 commits into from
Sep 5, 2023
Merged

save space when converting hf model to megatron model. #25950

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
23d9387
fix convert megatron model too large
flower-with-safe Sep 4, 2023
9b59900
Merge branch 'huggingface:main' into main
flower-with-safe Sep 4, 2023
0a2877d
fix convert megatron model too large
flower-with-safe Sep 4, 2023
2c585ac
Merge branch 'main' of github.com:flower-with-safe/transformers into …
flower-with-safe Sep 5, 2023
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
6 changes: 3 additions & 3 deletions src/transformers/models/megatron_gpt2/checkpoint_reshaping_and_interoperability.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -737,7 +737,7 @@ def convert_checkpoint_from_transformers_to_megatron(args):
word_emb_dict = get_element_from_dict_by_path(
output_state_dict[i], "model.language_model.embedding.word_embeddings"
)
word_emb_dict["weight"] = out_word_embed[i]
word_emb_dict["weight"] = out_word_embed[i].clone()

# Transformer layers
print("converting transformer layers")
Expand Down Expand Up @@ -845,7 +845,7 @@ def convert_checkpoint_from_transformers_to_megatron(args):
for i in range(args.target_tensor_model_parallel_size):
params_dict = get_element_from_dict_by_path(output_state_dict[i], "model.language_model.encoder")
params_dict[layer_name] = (
params[i] if (op_name + "." + weight_or_bias in tensor_parallel_params) else params
params[i].clone() if (op_name + "." + weight_or_bias in tensor_parallel_params) else params
)

if pp_rank == args.target_pipeline_model_parallel_size - 1:
Expand All @@ -860,7 +860,7 @@ def convert_checkpoint_from_transformers_to_megatron(args):
# add the LM head
for i in range(args.target_tensor_model_parallel_size):
params_dict = get_element_from_dict_by_path(output_state_dict[i], "model.word_embeddings_for_head")
params_dict["weight"] = out_word_embed[i]
params_dict["weight"] = out_word_embed[i].clone()

# saving the state dict as per the tp_rank and pp_rank
for tp_rank in range(args.target_tensor_model_parallel_size):
Expand Down