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

* fix convert megatron model too large * fix convert megatron model too large
huggingface · Sep 5, 2023 · 172f42c · 172f42c
1 parent b8def68
commit 172f42c
Showing 1 changed file with 3 additions and 3 deletions.
diff --git a/src/transformers/models/megatron_gpt2/checkpoint_reshaping_and_interoperability.py b/src/transformers/models/megatron_gpt2/checkpoint_reshaping_and_interoperability.py
@@ -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")
@@ -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:
@@ -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):