Maximum sample-based training for Megatron NMT and Text Memmap based Seq2seq Pre-training

Jenkinsfile

@@ -2814,7 +2814,7 @@ pipeline {
         trainer.devices=2 \
         trainer.accelerator=gpu \
         trainer.log_every_n_steps=1 \
-        trainer.val_check_interval=10 \
+        trainer.val_check_interval=2 \
         trainer.limit_val_batches=2 \
         trainer.accumulate_grad_batches=1 \
         trainer.max_steps=3 \
@@ -2839,36 +2839,36 @@ pipeline {
         model.activations_checkpoint_num_layers=1 \
         model.data.data_prefix=[.5,/home/TestData/nlp/megatron_gpt/data/gpt/simple_wiki_gpt_preproc_text_document,.5,/home/TestData/nlp/megatron_gpt/data/gpt/simple_wiki_gpt_preproc_text_document] \
         model.data.index_mapping_dir=examples/nlp/language_modeling/gpt_index_mappings"
-//        sh "python examples/nlp/language_modeling/megatron_gpt_pretraining.py \
-//        trainer.devices=2 \
-//        trainer.accelerator=gpu \
-//        trainer.log_every_n_steps=1 \
-//        trainer.val_check_interval=10 \
-//        trainer.limit_val_batches=1 \
-//        trainer.accumulate_grad_batches=1 \
-//        trainer.max_steps=20 \
-//        trainer.precision=16 \
-//        trainer.gradient_clip_val=1.0 \
-//        exp_manager.exp_dir=examples/nlp/language_modeling/gpt_pretrain_results \
-//        exp_manager.resume_if_exists=True \
-//        model.tensor_model_parallel_size=2 \
-//        model.optim.name=fused_adam \
-//        model.optim.lr=2e-4 \
-//        model.optim.sched.warmup_steps=2 \
-//        model.optim.sched.constant_steps=2 \
-//        model.optim.sched.min_lr=8e-5 \
-//        model.max_position_embeddings=128 \
-//        model.encoder_seq_length=128 \
-//        model.data.seq_length=128 \
-//        model.tokenizer.vocab_file=/home/TestData/nlp/megatron_gpt/data/gpt/vocab.json \
-//        model.tokenizer.merge_file=/home/TestData/nlp/megatron_gpt/data/gpt/merges.txt \
-//        model.num_layers=8 \
-//        model.hidden_size=256 \
-//        model.num_attention_heads=8 \
-//        model.activations_checkpoint_method='block' \
-//        model.activations_checkpoint_num_layers=1 \
-//        model.data.data_prefix=[.5,/home/TestData/nlp/megatron_gpt/data/gpt/simple_wiki_gpt_preproc_text_document,.5,/home/TestData/nlp/megatron_gpt/data/gpt/simple_wiki_gpt_preproc_text_document] \
-//        model.data.index_mapping_dir=examples/nlp/language_modeling/gpt_index_mappings"
+        sh "python examples/nlp/language_modeling/megatron_gpt_pretraining.py \
+        trainer.devices=2 \
+        trainer.accelerator=gpu \
+        trainer.log_every_n_steps=1 \
+        trainer.val_check_interval=2 \
+        trainer.limit_val_batches=1 \
+        trainer.accumulate_grad_batches=1 \
+        trainer.max_steps=6 \
+        trainer.precision=16 \
+        trainer.gradient_clip_val=1.0 \
+        exp_manager.exp_dir=examples/nlp/language_modeling/gpt_pretrain_results \
+        exp_manager.resume_if_exists=True \
+        model.tensor_model_parallel_size=2 \
+        model.optim.name=fused_adam \
+        model.optim.lr=2e-4 \
+        model.optim.sched.warmup_steps=2 \
+        model.optim.sched.constant_steps=2 \
+        model.optim.sched.min_lr=8e-5 \
+        model.max_position_embeddings=128 \
+        model.encoder_seq_length=128 \
+        model.data.seq_length=128 \
+        model.tokenizer.vocab_file=/home/TestData/nlp/megatron_gpt/data/gpt/vocab.json \
+        model.tokenizer.merge_file=/home/TestData/nlp/megatron_gpt/data/gpt/merges.txt \
+        model.num_layers=8 \
+        model.hidden_size=256 \
+        model.num_attention_heads=8 \
+        model.activations_checkpoint_method='block' \
+        model.activations_checkpoint_num_layers=1 \
+        model.data.data_prefix=[.5,/home/TestData/nlp/megatron_gpt/data/gpt/simple_wiki_gpt_preproc_text_document,.5,/home/TestData/nlp/megatron_gpt/data/gpt/simple_wiki_gpt_preproc_text_document] \
+        model.data.index_mapping_dir=examples/nlp/language_modeling/gpt_index_mappings"
         sh "rm -rf examples/nlp/language_modeling/gpt_pretrain_results"
         sh "rm -rf examples/nlp/language_modeling/gpt_index_mappings"
       }
@@ -2885,7 +2885,7 @@ pipeline {
         sh "python examples/nlp/language_modeling/megatron_gpt_pretraining.py \
         trainer.devices=2 \
         trainer.log_every_n_steps=1 \
-        trainer.val_check_interval=10 \
+        trainer.val_check_interval=2 \
         trainer.limit_val_batches=2 \
         trainer.accumulate_grad_batches=1 \
         trainer.max_steps=3 \
@@ -2911,36 +2911,36 @@ pipeline {
         model.activations_checkpoint_num_layers=1 \
         model.data.data_prefix=[.5,/home/TestData/nlp/megatron_gpt/data/gpt/simple_wiki_gpt_preproc_text_document,.5,/home/TestData/nlp/megatron_gpt/data/gpt/simple_wiki_gpt_preproc_text_document] \
         model.data.index_mapping_dir=examples/nlp/language_modeling/gpt_index_mappings"
-        // sh "python examples/nlp/language_modeling/megatron_gpt_pretraining.py \
-        // trainer.devices=2 \
-        // trainer.log_every_n_steps=1 \
-        // trainer.val_check_interval=10 \
-        // trainer.limit_val_batches=2 \
-        // trainer.accumulate_grad_batches=1 \
-        // trainer.max_steps=20 \
-        // trainer.precision=16 \
-        // trainer.gradient_clip_val=1.0 \
-        // exp_manager.exp_dir=examples/nlp/language_modeling/gpt_pretrain_results \
-        // exp_manager.resume_if_exists=True \
-        // model.pipeline_model_parallel_size=2 \
-        // model.tensor_model_parallel_size=1 \
-        // model.optim.name=fused_adam \
-        // model.optim.lr=2e-4 \
-        // model.optim.sched.warmup_steps=2 \
-        // model.optim.sched.constant_steps=2 \
-        // model.optim.sched.min_lr=8e-5 \
-        // model.max_position_embeddings=128 \
-        // model.encoder_seq_length=128 \
-        // model.data.seq_length=128 \
-        // model.tokenizer.vocab_file=/home/TestData/nlp/megatron_gpt/data/gpt/vocab.json \
-        // model.tokenizer.merge_file=/home/TestData/nlp/megatron_gpt/data/gpt/merges.txt \
-        // model.num_layers=8 \
-        // model.hidden_size=256 \
-        // model.num_attention_heads=8 \
-        // model.activations_checkpoint_method='block' \
-        // model.activations_checkpoint_num_layers=1 \
-        // model.data.data_prefix=[.5,/home/TestData/nlp/megatron_gpt/data/gpt/simple_wiki_gpt_preproc_text_document,.5,/home/TestData/nlp/megatron_gpt/data/gpt/simple_wiki_gpt_preproc_text_document] \
-        // model.data.index_mapping_dir=examples/nlp/language_modeling/gpt_index_mappings"
+        sh "python examples/nlp/language_modeling/megatron_gpt_pretraining.py \
+        trainer.devices=2 \
+        trainer.log_every_n_steps=1 \
+        trainer.val_check_interval=2 \
+        trainer.limit_val_batches=2 \
+        trainer.accumulate_grad_batches=1 \
+        trainer.max_steps=6 \
+        trainer.precision=16 \
+        trainer.gradient_clip_val=1.0 \
+        exp_manager.exp_dir=examples/nlp/language_modeling/gpt_pretrain_results \
+        exp_manager.resume_if_exists=True \
+        model.pipeline_model_parallel_size=2 \
+        model.tensor_model_parallel_size=1 \
+        model.optim.name=fused_adam \
+        model.optim.lr=2e-4 \
+        model.optim.sched.warmup_steps=2 \
+        model.optim.sched.constant_steps=2 \
+        model.optim.sched.min_lr=8e-5 \
+        model.max_position_embeddings=128 \
+        model.encoder_seq_length=128 \
+        model.data.seq_length=128 \
+        model.tokenizer.vocab_file=/home/TestData/nlp/megatron_gpt/data/gpt/vocab.json \
+        model.tokenizer.merge_file=/home/TestData/nlp/megatron_gpt/data/gpt/merges.txt \
+        model.num_layers=8 \
+        model.hidden_size=256 \
+        model.num_attention_heads=8 \
+        model.activations_checkpoint_method='block' \
+        model.activations_checkpoint_num_layers=1 \
+        model.data.data_prefix=[.5,/home/TestData/nlp/megatron_gpt/data/gpt/simple_wiki_gpt_preproc_text_document,.5,/home/TestData/nlp/megatron_gpt/data/gpt/simple_wiki_gpt_preproc_text_document] \
+        model.data.index_mapping_dir=examples/nlp/language_modeling/gpt_index_mappings"
         sh "rm -rf examples/nlp/language_modeling/gpt_pretrain_results"
         sh "rm -rf examples/nlp/language_modeling/gpt_index_mappings"
       }
@@ -3144,10 +3144,14 @@ pipeline {
         model.activations_checkpoint_method='block' \
         model.activations_checkpoint_num_layers=1 \
         model.transformer_block_type='pre_ln' \
-        model.data.data_prefix=[.5,/home/TestData/nlp/megatron_t5/data/pile_val_small_bert_tokenizer_text_document,.5,/home/TestData/nlp/megatron_t5/data/pile_val_small_bert_tokenizer_text_document] \
+        model.data.data_prefix=[.5,/home/TestData/nlp/nmt/toy_data/wmt14-de-en.src,.5,/home/TestData/nlp/nmt/toy_data/wmt14-de-en.ref] \
         model.position_embedding_type=relative \
         model.data.index_mapping_dir=examples/nlp/language_modeling/t5_index_mappings \
-        model.data.respect_document_boundaries=False \
+        model.data.data_impl=text_mmap \
+        +model.data.data_impl_kwargs.newline_int=10 \
+        +model.data.data_impl_kwargs.header_lines=0 \
+        +model.data.data_impl_kwargs.workers=null \
+        +model.data.data_impl_kwargs.sort_dataset_paths=False \
         model.share_token_embeddings=False \
         model.share_decoder_tokens_head_embeddings=False"
         sh "python examples/nlp/language_modeling/megatron_t5_pretraining.py \
@@ -3171,11 +3175,14 @@ pipeline {
         model.bias_activation_fusion=False \
         model.activations_checkpoint_method='block' \
         model.activations_checkpoint_num_layers=1 \
-        model.transformer_block_type='pre_ln' \
-        model.data.data_prefix=[.5,/home/TestData/nlp/megatron_t5/data/pile_val_small_bert_tokenizer_text_document,.5,/home/TestData/nlp/megatron_t5/data/pile_val_small_bert_tokenizer_text_document] \
+        model.data.data_prefix=[.5,/home/TestData/nlp/nmt/toy_data/wmt14-de-en.src,.5,/home/TestData/nlp/nmt/toy_data/wmt14-de-en.ref] \
         model.position_embedding_type=relative \
         model.data.index_mapping_dir=examples/nlp/language_modeling/t5_index_mappings \
-        model.data.respect_document_boundaries=False \
+        model.data.data_impl=text_mmap \
+        +model.data.data_impl_kwargs.newline_int=10 \
+        +model.data.data_impl_kwargs.header_lines=0 \
+        +model.data.data_impl_kwargs.workers=null \
+        +model.data.data_impl_kwargs.sort_dataset_paths=False \
         model.share_token_embeddings=False \
         model.share_decoder_tokens_head_embeddings=False"
         sh "rm -rf examples/nlp/language_modeling/t5_pretrain_results"

examples/nlp/language_modeling/conf/megatron_bart_config.yaml

@@ -122,6 +122,19 @@ model:
     data_prefix: ???
     index_mapping_dir: null # path to save index mapping .npy files, by default will save in the same location as data_prefix
     data_impl: mmap
+    # data_impl_kwargs: # currently used only for text_mmap, csv_mmap (should be data_impl dependant)
+    #     # defaults for text_memmap
+    #     newline_int: 10 # byte-value of newline (Use ord('\n') to get value)
+    #     header_lines: 0 # skip first N header lines
+    #     workers: null # number of workers when creating missing index files (null defaults to cpu_num // 2)
+    #     sort_dataset_paths: False # if True datasets will be sorted by name
+    #     # defaults for csv_memmap
+    #     newline_int: 10 # byte-value of newline
+    #     header_lines: 1 # skip first N header lines
+    #     workers: null # number of workers when creating missing index files (null defaults to cpu_num // 2)
+    #     sort_dataset_paths: False # if True datasets will be sorted by name
+    #     data_col: 1 # column to use for data
+    #     data_sep: ',' # string to split text into columns
     splits_string: 949,45,5
     seq_length: ${model.seq_length}
     skip_warmup: True

examples/nlp/language_modeling/conf/megatron_t5_config.yaml

@@ -124,7 +124,20 @@ model:
     #   - /raid/data/pile/my-t5_01_text_document
     data_prefix: ???
     index_mapping_dir: null # path to save index mapping .npy files, by default will save in the same location as data_prefix
-    data_impl: mmap
+    data_impl: mmap # mmap, retmmap, text_mmap, csv_mmap
+    # data_impl_kwargs: # currently used only for text_mmap, csv_mmap (should be data_impl dependant)
+    #     # defaults for text_memmap
+    #     newline_int: 10 # byte-value of newline (Use ord('\n') to get value)
+    #     header_lines: 0 # skip first N header lines
+    #     workers: null # number of workers when creating missing index files (null defaults to cpu_num // 2)
+    #     sort_dataset_paths: False # if True datasets will be sorted by name
+    #     # defaults for csv_memmap
+    #     newline_int: 10 # byte-value of newline
+    #     header_lines: 1 # skip first N header lines
+    #     workers: null # number of workers when creating missing index files (null defaults to cpu_num // 2)
+    #     sort_dataset_paths: False # if True datasets will be sorted by name
+    #     data_col: 1 # column to use for data
+    #     data_sep: ',' # string to split text into columns
     splits_string: 949,45,5
     seq_length: ${model.seq_length}
     seq_length_dec: 128

examples/nlp/language_modeling/conf/megatron_ul2_config.yaml

@@ -121,6 +121,19 @@ model:
     data_prefix: ???
     index_mapping_dir: null # path to save index mapping .npy files, by default will save in the same location as data_prefix
     data_impl: mmap
+    # data_impl_kwargs: # currently used only for text_mmap, csv_mmap (should be data_impl dependant)
+    #     # defaults for text_memmap
+    #     newline_int: 10 # byte-value of newline (Use ord('\n') to get value)
+    #     header_lines: 0 # skip first N header lines
+    #     workers: null # number of workers when creating missing index files (null defaults to cpu_num // 2)
+    #     sort_dataset_paths: False # if True datasets will be sorted by name
+    #     # defaults for csv_memmap
+    #     newline_int: 10 # byte-value of newline
+    #     header_lines: 1 # skip first N header lines
+    #     workers: null # number of workers when creating missing index files (null defaults to cpu_num // 2)
+    #     sort_dataset_paths: False # if True datasets will be sorted by name
+    #     data_col: 1 # column to use for data
+    #     data_sep: ',' # string to split text into columns
     splits_string: 949,45,5
     seq_length: ${model.seq_length}
     seq_length_dec: ${model.seq_length}

examples/nlp/machine_translation/megatron_nmt_training.py

@@ -145,8 +145,12 @@ def main(cfg) -> None:
             pretrained_cfg.train_ds = cfg.model.train_ds
             pretrained_cfg.train_ds.micro_batch_size = cfg.model.micro_batch_size
             pretrained_cfg.train_ds.global_batch_size = cfg.model.global_batch_size
-            pretrained_cfg.validation_ds = cfg.model.validation_ds
-            pretrained_cfg.test_ds = cfg.model.test_ds
+            if hasattr(cfg.model, 'validation_ds'):
+                pretrained_cfg.validation_ds = cfg.model.validation_ds
+            else:
+                raise AttributeError(f"No validation dataset found in config.")
+            if hasattr(cfg.model, 'test_ds'):
+                pretrained_cfg.test_ds = cfg.model.test_ds
 
             # Class target for the new class being restored.
             pretrained_cfg.target = (