***************************************** Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. ***************************************** ***************************************** Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. ***************************************** ***************************************** Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. ***************************************** ***************************************** Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. ***************************************** ***************************************** Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. ***************************************** ***************************************** Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. ***************************************** ***************************************** Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. ***************************************** ***************************************** Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. ***************************************** ***************************************** Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. ***************************************** ***************************************** Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. ***************************************** ***************************************** Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. ***************************************** ***************************************** Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. ***************************************** ***************************************** Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. ***************************************** ***************************************** Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. ***************************************** ***************************************** Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. ***************************************** ***************************************** Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. ***************************************** using world size: 64, data-parallel-size: 1, tensor-model-parallel size: 4, pipeline-model-parallel size: 16 using torch.float16 for parameters ... ------------------------ arguments ------------------------ accumulate_allreduce_grads_in_fp32 .............. False adam_beta1 ...................................... 0.9 adam_beta2 ...................................... 0.999 adam_eps ........................................ 1e-08 adlr_autoresume ................................. False adlr_autoresume_interval ........................ 1000 apply_query_key_layer_scaling ................... True apply_residual_connection_post_layernorm ........ False attention_dropout ............................... 0.1 attention_softmax_in_fp32 ....................... False bert_binary_head ................................ True bert_load ....................................... None bf16 ............................................ False bias_dropout_fusion ............................. True bias_gelu_fusion ................................ True biencoder_projection_dim ........................ 0 biencoder_shared_query_context_model ............ False block_data_path ................................. None checkpoint_activations .......................... True checkpoint_num_layers ........................... 1 clip_grad ....................................... 1.0 consumed_train_samples .......................... 0 consumed_valid_samples .......................... 0 data_impl ....................................... mmap data_parallel_size .............................. 1 data_path ....................................... ['/gpfsscratch/rech/eha/commun/datasets-custom/openwebtext-10k/meg-gpt2_text_document'] dataloader_type ................................. single DDP_impl ........................................ local decoder_seq_length .............................. None distribute_checkpointed_activations ............. False distributed_backend ............................. nccl embedding_path .................................. None encoder_seq_length .............................. 1024 eod_mask_loss ................................... False eval_interval ................................... 100 eval_iters ...................................... 10 evidence_data_path .............................. None exit_duration_in_mins ........................... None exit_interval ................................... None ffn_hidden_size ................................. 32768 finetune ........................................ False fp16 ............................................ True fp16_lm_cross_entropy ........................... False fp32_residual_connection ........................ False global_batch_size ............................... 1024 hidden_dropout .................................. 0.1 hidden_size ..................................... 8192 hysteresis ...................................... 2 ict_head_size ................................... None ict_load ........................................ None img_dim ......................................... 224 indexer_batch_size .............................. 128 indexer_log_interval ............................ 1000 init_method_std ................................. 0.02 init_method_xavier_uniform ...................... False initial_loss_scale .............................. 4294967296 kv_channels ..................................... 256 layernorm_epsilon ............................... 1e-05 lazy_mpu_init ................................... None load ............................................ /gpfsscratch/rech/eha/commun/checkpoints/gpt2-1-node local_rank ...................................... 0 log_batch_size_to_tensorboard ................... False log_interval .................................... 1 log_learning_rate_to_tensorboard ................ True log_loss_scale_to_tensorboard ................... True log_num_zeros_in_grad ........................... False log_params_norm ................................. False log_timers_to_tensorboard ....................... False log_validation_ppl_to_tensorboard ............... False loss_scale ...................................... None loss_scale_window ............................... 1000 lr .............................................. 0.00015 lr_decay_iters .................................. 800 lr_decay_samples ................................ None lr_decay_style .................................. cosine lr_warmup_fraction .............................. 0.01 lr_warmup_iters ................................. 0 lr_warmup_samples ............................... 0 make_vocab_size_divisible_by .................... 128 mask_prob ....................................... 0.15 masked_softmax_fusion ........................... True max_position_embeddings ......................... 1024 merge_file ...................................... /gpfsscratch/rech/eha/commun/models-custom/megatron-gpt2/megatron_lm_345m_v0.0/release/gpt2-merges.txt micro_batch_size ................................ 4 min_loss_scale .................................. 1.0 min_lr .......................................... 1e-05 mmap_warmup ..................................... False no_load_optim ................................... None no_load_rng ..................................... None no_save_optim ................................... None no_save_rng ..................................... None num_attention_heads ............................. 32 num_channels .................................... 3 num_classes ..................................... 1000 num_layers ...................................... 64 num_layers_per_virtual_pipeline_stage ........... None num_workers ..................................... 2 onnx_safe ....................................... None openai_gelu ..................................... False optimizer ....................................... adam override_lr_scheduler ........................... False params_dtype .................................... torch.float16 patch_dim ....................................... 16 pipeline_model_parallel_size .................... 16 query_in_block_prob ............................. 0.1 rampup_batch_size ............................... None rank ............................................ 0 reset_attention_mask ............................ False reset_position_ids .............................. False retriever_report_topk_accuracies ................ [] retriever_score_scaling ......................... False retriever_seq_length ............................ 256 sample_rate ..................................... 1.0 save ............................................ /gpfsscratch/rech/eha/commun/checkpoints/gpt2-1-node save_interval ................................... 500 scatter_gather_tensors_in_pipeline .............. True seed ............................................ 1234 seq_length ...................................... 1024 sgd_momentum .................................... 0.9 short_seq_prob .................................. 0.1 split ........................................... 949,50,1 tensor_model_parallel_size ...................... 4 tensorboard_dir ................................. None tensorboard_log_interval ........................ 1 tensorboard_queue_size .......................... 1000 titles_data_path ................................ None tokenizer_type .................................. GPT2BPETokenizer train_iters ..................................... 1000 train_samples ................................... None use_checkpoint_lr_scheduler ..................... False use_contiguous_buffers_in_ddp ................... False use_cpu_initialization .......................... None use_one_sent_docs ............................... False virtual_pipeline_model_parallel_size ............ None vocab_extra_ids ................................. 0 vocab_file ...................................... /gpfsscratch/rech/eha/commun/models-custom/megatron-gpt2/megatron_lm_345m_v0.0/release/gpt2-vocab.json weight_decay .................................... 0.01 world_size ...................................... 64 -------------------- end of arguments --------------------- setting number of micro-batches to constant 256 > building GPT2BPETokenizer tokenizer ... > padded vocab (size: 50257) with 431 dummy tokens (new size: 50688) > initializing torch distributed ... > initializing tensor model parallel with size 4 > initializing pipeline model parallel with size 16 > setting random seeds to 1234 ... > initializing model parallel cuda seeds on global rank 0, model parallel rank 0, and data parallel rank 0 with model parallel seed: 3952 and data parallel seed: 1234 > compiling dataset index builder ... make: Entering directory '/gpfsdswork/projects/rech/eha/commun/code/megatron-lm/megatron/data' make: Nothing to be done for 'default'. make: Leaving directory '/gpfsdswork/projects/rech/eha/commun/code/megatron-lm/megatron/data' >>> done with dataset index builder. Compilation time: 0.109 seconds > compiling and loading fused kernels ... Detected CUDA files, patching ldflags Emitting ninja build file /gpfsdswork/projects/rech/eha/commun/code/megatron-lm/megatron/fused_kernels/build/build.ninja... Building extension module scaled_upper_triang_masked_softmax_cuda... Allowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N) ninja: no work to do. Loading extension module scaled_upper_triang_masked_softmax_cuda... Detected CUDA files, patching ldflags Emitting ninja build file /gpfsdswork/projects/rech/eha/commun/code/megatron-lm/megatron/fused_kernels/build/build.ninja... Building extension module scaled_masked_softmax_cuda... Allowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N) ninja: no work to do. Loading extension module scaled_masked_softmax_cuda... Detected CUDA files, patching ldflags Emitting ninja build file /gpfsdswork/projects/rech/eha/commun/code/megatron-lm/megatron/fused_kernels/build/build.ninja... Building extension module fused_mix_prec_layer_norm_cuda... Allowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N) ninja: no work to do. Loading extension module fused_mix_prec_layer_norm_cuda... >>> done with compiling and loading fused kernels. Compilation time: 12.395 seconds time to initialize megatron (seconds): -27.152 [after megatron is initialized] datetime: 2021-05-26 03:20:36 building GPT model ... > number of parameters on (tensor, pipeline) model parallel rank (2, 1): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (0, 13): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (1, 2): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (0, 11): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (1, 6): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (3, 1): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (3, 2): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (2, 13): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (3, 6): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (0, 6): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (0, 14): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (3, 10): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (1, 14): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (0, 7): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (1, 5): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (2, 5): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (2, 7): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (3, 4): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (0, 5): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (0, 4): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (1, 4): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (0, 1): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (1, 13): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (3, 13): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (1, 1): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (2, 2): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (0, 2): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (2, 6): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (3, 5): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (0, 12): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (3, 3): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (2, 12): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (1, 3): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (1, 10): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (0, 10): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (2, 4): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (1, 7): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (2, 14): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (3, 14): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (2, 10): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (3, 7): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (1, 11): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (3, 11): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (2, 11): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (0, 3): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (0, 9): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (2, 3): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (3, 12): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (1, 12): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (3, 9): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (1, 9): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (2, 8): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (3, 8): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (1, 8): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (0, 8): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (2, 9): 805560320 > number of parameters on (tensor, pipeline) model parallel rank (3, 15): 909385728 > number of parameters on (tensor, pipeline) model parallel rank (3, 0): 917757952 > number of parameters on (tensor, pipeline) model parallel rank (1, 15): 909385728 > number of parameters on (tensor, pipeline) model parallel rank (0, 15): 909385728 > number of parameters on (tensor, pipeline) model parallel rank (1, 0): 917757952 > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 917757952 > number of parameters on (tensor, pipeline) model parallel rank (2, 0): 917757952 > number of parameters on (tensor, pipeline) model parallel rank (2, 15): 909385728 > learning rate decay style: cosine WARNING: could not find the metadata file /gpfsscratch/rech/eha/commun/checkpoints/gpt2-1-node/latest_checkpointed_iteration.txt will not load any checkpoints and will start from random time (ms) | load-checkpoint: 0.47 [after model, optimizer, and learning rate scheduler are built] datetime: 2021-05-26 03:20:37 > building train, validation, and test datasets ... > datasets target sizes (minimum size): train: 1024000 validation: 112640 test: 10240 > building train, validation, and test datasets for GPT ... > building dataset index ... reading sizes... reading pointers... reading document index... creating numpy buffer of mmap... creating memory view of numpy buffer... > finished creating indexed dataset in 0.001563 seconds number of documents: 10000 > dataset split: train: document indices in [0, 9490) total of 9490 documents validation: document indices in [9490, 9990) total of 500 documents test: document indices in [9990, 10000) total of 10 documents > loading doc-idx mapping from /gpfsscratch/rech/eha/commun/datasets-custom/openwebtext-10k/meg-gpt2_text_document_train_indexmap_1024000ns_1024sl_1234s_doc_idx.npy > loading sample-idx mapping from /gpfsscratch/rech/eha/commun/datasets-custom/openwebtext-10k/meg-gpt2_text_document_train_indexmap_1024000ns_1024sl_1234s_sample_idx.npy > loading shuffle-idx mapping from /gpfsscratch/rech/eha/commun/datasets-custom/openwebtext-10k/meg-gpt2_text_document_train_indexmap_1024000ns_1024sl_1234s_shuffle_idx.npy loaded indexed file in 0.013 seconds total number of samples: 1024856 total number of epochs: 99 > loading doc-idx mapping from /gpfsscratch/rech/eha/commun/datasets-custom/openwebtext-10k/meg-gpt2_text_document_valid_indexmap_112640ns_1024sl_1234s_doc_idx.npy > loading sample-idx mapping from /gpfsscratch/rech/eha/commun/datasets-custom/openwebtext-10k/meg-gpt2_text_document_valid_indexmap_112640ns_1024sl_1234s_sample_idx.npy > loading shuffle-idx mapping from /gpfsscratch/rech/eha/commun/datasets-custom/openwebtext-10k/meg-gpt2_text_document_valid_indexmap_112640ns_1024sl_1234s_shuffle_idx.npy loaded indexed file in 0.002 seconds total number of samples: 113200 total number of epochs: 182 > loading doc-idx mapping from /gpfsscratch/rech/eha/commun/datasets-custom/openwebtext-10k/meg-gpt2_text_document_test_indexmap_10240ns_1024sl_1234s_doc_idx.npy > loading sample-idx mapping from /gpfsscratch/rech/eha/commun/datasets-custom/openwebtext-10k/meg-gpt2_text_document_test_indexmap_10240ns_1024sl_1234s_sample_idx.npy > loading shuffle-idx mapping from /gpfsscratch/rech/eha/commun/datasets-custom/openwebtext-10k/meg-gpt2_text_document_test_indexmap_10240ns_1024sl_1234s_shuffle_idx.npy loaded indexed file in 0.001 seconds total number of samples: 10255 total number of epochs: 672 > finished creating GPT datasets ... [after dataloaders are built] datetime: 2021-05-26 03:20:38 done with setup ... training ... time (ms) | model-and-optimizer-setup: 336.31 | train/valid/test-data-iterators-setup: 494.63 [before the start of training step] datetime: 2021-05-26 03:20:38 iteration 1/ 1000 | consumed samples: 1024 | elapsed time per iteration (ms): 155483.8 | learning rate: 0.000E+00 | global batch size: 1024 | loss scale: 4294967296.0 | number of skipped iterations: 1 | number of nan iterations: 0 | time (ms) | forward-compute: 32054.16 | forward-recv: 19581.85 | backward-compute: 87380.53 | backward-send: 2.90 | backward-send-forward-recv: 11878.08 | backward-params-all-reduce: 26.06 | backward-embedding-all-reduce: 4181.82 | optimizer-copy-to-main-grad: 8.18 | optimizer-unscale-and-check-inf: 333.72 | optimizer: 342.01 | batch-generator: 182.63 iteration 2/ 1000 | consumed samples: 2048 | elapsed time per iteration (ms): 129028.8 | learning rate: 0.000E+00 | global batch size: 1024 | loss scale: 2147483648.0 | number of skipped iterations: 1 | number of nan iterations: 0 | time (ms) | forward-compute: 31749.74 | forward-recv: 1493.80 | backward-compute: 88911.73 | backward-send: 2.63 | backward-send-forward-recv: 2267.37 | backward-params-all-reduce: 26.04 | backward-embedding-all-reduce: 4492.62 | optimizer-copy-to-main-grad: 8.16 | optimizer-unscale-and-check-inf: 42.71 | optimizer: 51.07 | batch-generator: 184.46 iteration 3/ 1000 | consumed samples: 3072 | elapsed time per iteration (ms): 128891.5 | learning rate: 0.000E+00 | global batch size: 1024 | loss scale: 1073741824.0 | number of skipped iterations: 1 | number of nan iterations: 0 | time (ms) | forward-compute: 31888.01 | forward-recv: 1498.23 | backward-compute: 89444.82 | backward-send: 2.80 | backward-send-forward-recv: 1766.94 | backward-params-all-reduce: 26.45 | backward-embedding-all-reduce: 4179.82 | optimizer-copy-to-main-grad: 8.17 | optimizer-unscale-and-check-inf: 42.35 | optimizer: 50.66 | batch-generator: 187.52