Model parameters: d_model 768 ffw_size 3072 kv_size 64 n_heads 12 n_layers 15 Megatron-DeepSpeed/pretrain_gpt.py --tensor-model-parallel-size 1 --pipeline-model-parallel-size 1 --num-layers 15 --hidden-size 768 --num-attention-heads 12 --kv-channels 64 --ffn-hidden-size 3072 --seq-length 2048 --max-position-embeddings 2048 --micro-batch-size 4 --global-batch-size 256 --train-samples 84_762_549 --vocab-file gpt2/vocab.json --merge-file gpt2/merges.txt --loss-scale 12 --clip-grad 1.0 --kill-switch-path kill-switch-146m174b100m --bf16 --checkpoint-activations --optimizer adam --adam-beta1 0.9 --adam-beta2 0.999 --adam-eps 1e-8 --lr 2e-4 --min-lr 2e-5 --lr-decay-style cosine --lr-decay-samples 84_762_549 --lr-warmup-samples 847_625 --clip-grad 1.0 --weight-decay 1e-1 --log-interval 100 --save-interval 10000 --eval-interval 10000 --eval-iters 1 --tensorboard-dir tensorboard_146m174b100m --tensorboard-queue-size 5 --log-timers-to-tensorboard --log-batch-size-to-tensorboard --log-validation-ppl-to-tensorboard --save checkpoints_146m174b100m --load checkpoints_146m174b100m --train-weighted-split-paths-path train100m.txt --valid-weighted-split-paths-path val.txt --data-impl mmap --deepspeed --deepspeed_config ds_configs/3319491.json --zero-stage 0 START 3319491: Fri 17 Mar 2023 01:50:53 PM EET 0: 0: 0: ======================= ROCm System Management Interface ======================= 0: ================================= Concise Info ================================= 0: GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% 0: 0 46.0c 96.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 0: 1 44.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 0: 2 40.0c 92.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 0: 3 43.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 0: 4 45.0c 87.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 0: 5 42.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 0: 6 38.0c 89.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 0: 7 47.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 0: ================================================================================ 0: ============================= End of ROCm SMI Log ============================== 7: 7: 7: ======================= ROCm System Management Interface ======================= 7: ================================= Concise Info ================================= 7: GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% 7: 0 45.0c 88.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 7: 1 49.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 7: 2 38.0c 83.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 7: 3 43.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 7: 4 43.0c 86.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 7: 5 43.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 7: 6 42.0c 91.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 7: 7 43.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 7: ================================================================================ 7: ============================= End of ROCm SMI Log ============================== 1: 1: 1: ======================= ROCm System Management Interface ======================= 1: ================================= Concise Info ================================= 1: GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% 1: 0 45.0c 95.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 1: 1 43.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 1: 2 42.0c 86.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 1: 3 47.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 1: 4 49.0c 95.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 1: 5 52.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 1: 6 42.0c 87.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 1: 7 46.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 1: ================================================================================ 1: ============================= End of ROCm SMI Log ============================== 4: 4: 4: ======================= ROCm System Management Interface ======================= 4: ================================= Concise Info ================================= 4: GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% 4: 0 49.0c 88.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 4: 1 46.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 4: 2 41.0c 90.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 4: 3 46.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 4: 4 42.0c 90.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 4: 5 45.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 4: 6 43.0c 92.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 4: 7 39.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 4: ================================================================================ 4: ============================= End of ROCm SMI Log ============================== 5: 5: 5: ======================= ROCm System Management Interface ======================= 5: ================================= Concise Info ================================= 5: GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% 5: 0 46.0c 90.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 5: 1 45.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 5: 2 42.0c 87.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 5: 3 44.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 5: 4 44.0c 84.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 5: 5 48.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 5: 6 35.0c 85.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 5: 7 40.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 5: ================================================================================ 5: ============================= End of ROCm SMI Log ============================== 3: 3: 3: ======================= ROCm System Management Interface ======================= 3: ================================= Concise Info ================================= 3: GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% 3: 0 46.0c 88.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 3: 1 52.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 3: 2 46.0c 91.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 3: 3 43.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 3: 4 42.0c 94.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 3: 5 44.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 3: 6 47.0c 86.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 3: 7 44.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 3: ================================================================================ 3: ============================= End of ROCm SMI Log ============================== 2: 2: 2: ======================= ROCm System Management Interface ======================= 2: ================================= Concise Info ================================= 2: GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% 2: 0 45.0c 91.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 2: 1 48.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 2: 2 40.0c 89.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 2: 3 47.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 2: 4 45.0c 85.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 2: 5 47.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 2: 6 40.0c 84.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 2: 7 42.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 2: ================================================================================ 2: ============================= End of ROCm SMI Log ============================== 6: 6: 6: ======================= ROCm System Management Interface ======================= 6: ================================= Concise Info ================================= 6: GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% 6: 0 48.0c 83.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 6: 1 46.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 6: 2 41.0c 96.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 6: 3 45.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 6: 4 42.0c 91.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 6: 5 46.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 6: 6 39.0c 87.0W 800Mhz 1600Mhz 0% auto 560.0W 0% 0% 6: 7 44.0c N/A 800Mhz 1600Mhz 0% auto 0.0W 0% 0% 6: ================================================================================ 6: ============================= End of ROCm SMI Log ============================== 7: Launching on nid006946 (7/8), master nid006939 port 9999, GPUs 8, CUDA: True 4: Launching on nid006943 (4/8), master nid006939 port 9999, GPUs 8, CUDA: True 6: Launching on nid006945 (6/8), master nid006939 port 9999, GPUs 8, CUDA: True 3: Launching on nid006942 (3/8), master nid006939 port 9999, GPUs 8, CUDA: True 0: Launching on nid006939 (0/8), master nid006939 port 9999, GPUs 8, CUDA: True 5: Launching on nid006944 (5/8), master nid006939 port 9999, GPUs 8, CUDA: True 1: Launching on nid006940 (1/8), master nid006939 port 9999, GPUs 8, CUDA: True 2: Launching on nid006941 (2/8), master nid006939 port 9999, GPUs 8, CUDA: True 7: > setting tensorboard ... 0: using world size: 64, data-parallel-size: 64, tensor-model-parallel size: 1, pipeline-model-parallel size: 1 0: accumulate and all-reduce gradients in fp32 for bfloat16 data type. 0: using torch.bfloat16 for parameters ... 0: ------------------------ arguments ------------------------ 0: abort_on_unmet_fused_kernel_constraints ......... False 0: accumulate_allreduce_grads_in_fp32 .............. True 0: adam_beta1 ...................................... 0.9 0: adam_beta2 ...................................... 0.999 0: adam_eps ........................................ 1e-08 0: adlr_autoresume ................................. False 0: adlr_autoresume_interval ........................ 1000 0: apply_query_key_layer_scaling ................... True 0: apply_residual_connection_post_layernorm ........ False 0: attention_dropout ............................... 0.1 0: attention_softmax_in_fp32 ....................... False 0: bert_binary_head ................................ True 0: bert_load ....................................... None 0: bf16 ............................................ True 0: bias_dropout_fusion ............................. True 0: bias_gelu_fusion ................................ True 0: biencoder_projection_dim ........................ 0 0: biencoder_shared_query_context_model ............ False 0: block_data_path ................................. None 0: checkpoint_activations .......................... True 0: checkpoint_in_cpu ............................... False 0: checkpoint_num_layers ........................... 1 0: clip_grad ....................................... 1.0 0: codecarbon_dir .................................. None 0: consumed_train_samples .......................... 0 0: consumed_train_tokens ........................... 0 0: consumed_valid_samples .......................... 0 0: contigious_checkpointing ........................ False 0: cpu_optimizer ................................... False 0: cpu_torch_adam .................................. False 0: curriculum_learning ............................. False 0: data_impl ....................................... mmap 0: data_parallel_size .............................. 64 0: data_path ....................................... None 0: dataloader_type ................................. single 0: DDP_impl ........................................ local 0: decoder_seq_length .............................. None 0: deepscale ....................................... False 0: deepscale_config ................................ None 0: deepspeed ....................................... True 0: deepspeed_activation_checkpointing .............. False 0: deepspeed_config ................................ ds_configs/3319491.json 0: deepspeed_mpi ................................... False 0: distribute_checkpointed_activations ............. False 0: distributed_backend ............................. nccl 0: embed_layernorm ................................. False 0: embedding_path .................................. None 0: encoder_seq_length .............................. 2048 0: eod_mask_loss ................................... False 0: eval_interval ................................... 10000 0: eval_iters ...................................... 1 0: eval_only ....................................... None 0: evidence_data_path .............................. None 0: exit_duration_in_mins ........................... None 0: exit_interval ................................... None 0: ffn_hidden_size ................................. 3072 0: finetune ........................................ False 0: fp16 ............................................ False 0: fp16_lm_cross_entropy ........................... False 0: fp32_residual_connection ........................ False 0: gigaflos_no_embeds .............................. 0 0: global_batch_size ............................... 256 0: glu_activation .................................. None 0: hidden_dropout .................................. 0.1 0: hidden_size ..................................... 768 0: hysteresis ...................................... 2 0: ict_head_size ................................... None 0: ict_load ........................................ None 0: img_dim ......................................... 224 0: indexer_batch_size .............................. 128 0: indexer_log_interval ............................ 1000 0: inference ....................................... False 0: init_method_std ................................. 0.02 0: init_method_xavier_uniform ...................... False 0: initial_loss_scale .............................. 4294967296 0: kill_switch_path ................................ kill-switch-146m174b100m 0: kv_channels ..................................... 64 0: layer_norm_fusion ............................... True 0: layernorm_epsilon ............................... 1e-05 0: lazy_mpu_init ................................... None 0: load ............................................ checkpoints_146m174b100m 0: local_rank ...................................... None 0: log_batch_size_to_tensorboard ................... True 0: log_interval .................................... 100 0: log_learning_rate_to_tensorboard ................ True 0: log_level ....................................... None 0: log_level_replica ............................... None 0: log_loss_scale_to_tensorboard ................... True 0: log_num_zeros_in_grad ........................... False 0: log_params_norm ................................. False 0: log_path ........................................ None 0: log_timers_to_tensorboard ....................... True 0: log_validation_ppl_to_tensorboard ............... True 0: loss_on_targets_only ............................ False 0: loss_scale ...................................... 12.0 0: loss_scale_window ............................... 1000 0: lr .............................................. 0.0002 0: lr_decay_iters .................................. None 0: lr_decay_samples ................................ 84762549 0: lr_decay_style .................................. cosine 0: lr_decay_tokens ................................. None 0: lr_warmup_fraction .............................. None 0: lr_warmup_iters ................................. 0 0: lr_warmup_samples ............................... 847625 0: make_vocab_size_divisible_by .................... 128 0: mask_prob ....................................... 0.15 0: masked_softmax_fusion ........................... True 0: max_position_embeddings ......................... 2048 0: mean_noise_span_length .......................... None 0: memory_centric_tiled_linear ..................... False 0: merge_file ...................................... gpt2/merges.txt 0: micro_batch_size ................................ 4 0: min_loss_scale .................................. 1.0 0: min_lr .......................................... 2e-05 0: mmap_warmup ..................................... False 0: no_load_optim ................................... None 0: no_load_rng ..................................... None 0: no_save_optim ................................... None 0: no_save_rng ..................................... None 0: noise_density ................................... None 0: num_attention_heads ............................. 12 0: num_channels .................................... 3 0: num_classes ..................................... 1000 0: num_layers ...................................... 15 0: num_layers_per_virtual_pipeline_stage ........... None 0: num_workers ..................................... 2 0: onnx_safe ....................................... None 0: openai_gelu ..................................... False 0: optimizer ....................................... adam 0: optimizer_fusion ................................ True 0: override_lr_scheduler ........................... False 0: pad_vocab_size_to ............................... None 0: params_dtype .................................... torch.bfloat16 0: partition_activations ........................... False 0: patch_dim ....................................... 16 0: pipeline_model_parallel_size .................... 1 0: position_embedding_type ......................... PositionEmbeddingType.absolute 0: pp_partition_method ............................. None 0: profile_backward ................................ False 0: query_in_block_prob ............................. 0.1 0: rampup_batch_size ............................... None 0: rank ............................................ 0 0: remote_device ................................... none 0: reset_attention_mask ............................ False 0: reset_position_ids .............................. False 0: reset_progress .................................. None 0: retriever_report_topk_accuracies ................ [] 0: retriever_score_scaling ......................... False 0: retriever_seq_length ............................ 256 0: reweight_loss_based_on_position_frequency ....... False 0: sample_rate ..................................... 1.0 0: save ............................................ checkpoints_146m174b100m 0: save_interval ................................... 10000 0: scatter_gather_tensors_in_pipeline .............. True 0: scattered_embeddings ............................ False 0: seed ............................................ 1234 0: seq_length ...................................... 2048 0: sgd_momentum .................................... 0.9 0: short_seq_prob .................................. 0.1 0: skip_train_iteration_range ...................... None 0: split ........................................... None 0: split_transformers .............................. False 0: sync_tp_duplicated_parameters ................... False 0: synchronize_each_layer .......................... False 0: tensor_model_parallel_size ...................... 1 0: tensorboard_dir ................................. tensorboard_146m174b100m 0: tensorboard_log_interval ........................ 1 0: tensorboard_queue_size .......................... 5 0: test_weighted_split_paths ....................... None 0: test_weighted_split_paths_path .................. None 0: tile_factor ..................................... 1 0: titles_data_path ................................ None 0: tokenizer_name_or_path .......................... None 0: tokenizer_type .................................. GPT2BPETokenizer 0: train_iters ..................................... None 0: train_samples ................................... 84762549 0: train_tokens .................................... None 0: train_weighted_split_names ...................... ['train'] 0: train_weighted_split_paths ...................... [['/scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_100M_text_document']] 0: train_weighted_split_paths_path ................. None 0: train_weighted_split_splits ..................... [['0:1']] 0: train_weighted_split_weights .................... [['1.0']] 0: universal_checkpoint ............................ False 0: use_bnb_optimizer ............................... False 0: use_checkpoint_lr_scheduler ..................... False 0: use_contiguous_buffers_in_ddp ................... True 0: use_cpu_initialization .......................... None 0: use_one_sent_docs ............................... False 0: use_pin_memory .................................. False 0: valid_num_workers ............................... 2 0: valid_weighted_split_names ...................... ['validation'] 0: valid_weighted_split_paths ...................... [['/scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document']] 0: valid_weighted_split_paths_path ................. None 0: valid_weighted_split_splits ..................... [['0:1']] 0: valid_weighted_split_weights .................... [['1.0']] 0: virtual_pipeline_model_parallel_size ............ None 0: vocab_extra_ids ................................. 0 0: vocab_file ...................................... gpt2/vocab.json 0: weight_decay .................................... 0.1 0: world_size ...................................... 64 0: zero_allgather_bucket_size ...................... 0.0 0: zero_contigious_gradients ....................... False 0: zero_reduce_bucket_size ......................... 0.0 0: zero_reduce_scatter ............................. False 0: zero_stage ...................................... 0 0: -------------------- end of arguments --------------------- 0: setting number of micro-batches to constant 1 0: > building GPT2BPETokenizer tokenizer ... 0: > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304) 0: DeepSpeed general environment info: 0: torch install path ............... ['/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/venv/lib/python3.9/site-packages/torch'] 0: torch version .................... 1.13.0+rocm5.2 0: torch cuda version ............... None 0: torch hip version ................ 5.2.21151-afdc89f8 0: nvcc version ..................... None 0: deepspeed install path ........... ['/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/venv/lib/python3.9/site-packages/deepspeed'] 0: deepspeed info ................... 0.7.5, unknown, unknown 0: deepspeed wheel compiled w. ...... torch 1.13, hip 5.1 0: **** Git info for Megatron: git_hash=unknown git_branch=unknown **** 0: > initializing torch distributed ... 0: [2023-03-17 13:53:41,482] [INFO] [comm.py:633:init_distributed] Initializing TorchBackend in DeepSpeed with backend nccl 0: > initializing tensor model parallel with size 1 0: > initializing pipeline model parallel with size 1 0: > setting random seeds to 1234 ... 0: > 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 0: > compiling dataset index builder ... 0: make: Entering directory '/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/Megatron-DeepSpeed/megatron/data' 0: make: Nothing to be done for 'default'. 0: make: Leaving directory '/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/Megatron-DeepSpeed/megatron/data' 0: >>> done with dataset index builder. Compilation time: 0.065 seconds 0: > compiling and loading fused kernels ...