[QUESTION] Training Mixtral 8x7B on 16 x H100 only achieves low throughput of 130 TFLOPS

As the title says, I wonder if this is normal. If not, how should I optimize it?

Logs

``` using world size: 16, data-parallel size: 4, context-parallel size: 1 tensor-model-parallel size: 4, pipeline-model-parallel size: 1 WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:Llama2Tokenizer WARNING: Setting args.overlap_p2p_comm to False since non-interleaved schedule does not support overlapping p2p communication accumulate and all-reduce gradients in fp32 for bfloat16 data type. using torch.bfloat16 for parameters ... ------------------------ arguments ------------------------ accumulate_allreduce_grads_in_fp32 .............. True adam_beta1 ...................................... 0.9 adam_beta2 ...................................... 0.999 adam_eps ........................................ 1e-08 add_bias_linear ................................. False add_position_embedding .......................... False add_qkv_bias .................................... False adlr_autoresume ................................. False adlr_autoresume_interval ........................ 1000 apply_layernorm_1p .............................. False apply_query_key_layer_scaling ................... False apply_residual_connection_post_layernorm ........ False apply_rope_fusion ............................... True async_tensor_model_parallel_allreduce ........... False attention_dropout ............................... 0.0 attention_softmax_in_fp32 ....................... False auto_detect_ckpt_format ......................... False barrier_with_L1_time ............................ True bert_binary_head ................................ True bert_embedder_type .............................. megatron bert_load ....................................... None bf16 ............................................ True bias_dropout_fusion ............................. True bias_gelu_fusion ................................ False bias_swiglu_fusion .............................. True biencoder_projection_dim ........................ 0 biencoder_shared_query_context_model ............ False block_data_path ................................. None check_for_nan_in_loss_and_grad .................. True classes_fraction ................................ 1.0 clip_grad ....................................... 1.0 clone_scatter_output_in_embedding ............... True consumed_train_samples .......................... 0 consumed_valid_samples .......................... 0 context_parallel_size ........................... 1 create_attention_mask_in_dataloader ............. True data_cache_path ................................. None data_parallel_random_init ....................... False data_parallel_size .............................. 4 data_path ....................................... [] data_per_class_fraction ......................... 1.0 data_sharding ................................... True dataloader_type ................................. single decoder_num_layers .............................. None decoder_seq_length .............................. None delay_grad_reduce ............................... True delay_param_gather .............................. False dino_bottleneck_size ............................ 256 dino_freeze_last_layer .......................... 1 dino_head_hidden_size ........................... 2048 dino_local_crops_number ......................... 10 dino_local_img_size ............................. 96 dino_norm_last_layer ............................ False dino_teacher_temp ............................... 0.07 dino_warmup_teacher_temp ........................ 0.04 dino_warmup_teacher_temp_epochs ................. 30 dist_ckpt_format ................................ torch_dist distribute_saved_activations .................... False distributed_backend ............................. nccl distributed_timeout_minutes ..................... 10 embedding_path .................................. None empty_unused_memory_level ....................... 0 enable_one_logger ............................... False encoder_num_layers .............................. 32 encoder_seq_length .............................. 2048 end_weight_decay ................................ 0.1 eod_mask_loss ................................... False eval_interval ................................... 1000 eval_iters ...................................... 10 evidence_data_path .............................. None exit_duration_in_mins ........................... None exit_interval ................................... None exit_on_missing_checkpoint ...................... False exit_signal_handler ............................. False expert_model_parallel_size ...................... 4 ffn_hidden_size ................................. 14336 finetune ........................................ False fp16 ............................................ False fp16_lm_cross_entropy ........................... False fp32_residual_connection ........................ False fp8 ............................................. None fp8_amax_compute_algo ........................... most_recent fp8_amax_history_len ............................ 1 fp8_interval .................................... 1 fp8_margin ...................................... 0 fp8_wgrad ....................................... True global_batch_size ............................... 128 gradient_accumulation_fusion .................... True group_query_attention ........................... True head_lr_mult .................................... 1.0 hidden_dropout .................................. 0.0 hidden_size ..................................... 4096 hysteresis ...................................... 2 ict_head_size ................................... None ict_load ........................................ None img_h ........................................... 224 img_w ........................................... 224 indexer_batch_size .............................. 128 indexer_log_interval ............................ 1000 inference_batch_times_seqlen_threshold .......... 512 init_method_std ................................. 0.02 init_method_xavier_uniform ...................... False initial_loss_scale .............................. 4294967296 iter_per_epoch .................................. 1250 kv_channels ..................................... 128 lazy_mpu_init ................................... None load ............................................ custom/ckpt/mixtral-8x7b local_rank ...................................... None log_batch_size_to_tensorboard ................... False log_interval .................................... 1 log_learning_rate_to_tensorboard ................ True log_loss_scale_to_tensorboard ................... True log_memory_to_tensorboard ....................... False log_num_zeros_in_grad ........................... False log_params_norm ................................. False log_progress .................................... True log_throughput .................................. True log_timers_to_tensorboard ....................... False log_validation_ppl_to_tensorboard ............... False log_world_size_to_tensorboard ................... False loss_scale ...................................... None loss_scale_window ............................... 1000 lr .............................................. 0.0001 lr_decay_iters .................................. 320000 lr_decay_samples ................................ None lr_decay_style .................................. cosine lr_warmup_fraction .............................. None lr_warmup_init .................................. 0.0 lr_warmup_iters ................................. 500 lr_warmup_samples ............................... 0 make_vocab_size_divisible_by .................... 128 manual_gc ....................................... False manual_gc_eval .................................. True manual_gc_interval .............................. 0 mask_factor ..................................... 1.0 mask_prob ....................................... 0.15 mask_type ....................................... random masked_softmax_fusion ........................... False max_position_embeddings ......................... 32768 max_tokens_to_oom ............................... 12000 merge_file ...................................... None micro_batch_size ................................ 1 min_loss_scale .................................. 1.0 min_lr .......................................... 1e-05 mmap_bin_files .................................. True mock_data ....................................... True moe_aux_loss_coeff .............................. 0.01 moe_grouped_gemm ................................ True moe_input_jitter_eps ............................ None moe_router_load_balancing_type .................. aux_loss moe_router_topk ................................. 2 moe_token_dropping .............................. False moe_z_loss_coeff ................................ None nccl_communicator_config_path ................... None no_load_optim ................................... True no_load_rng ..................................... True no_persist_layer_norm ........................... False no_save_optim ................................... None no_save_rng ..................................... None norm_epsilon .................................... 1e-05 normalization ................................... RMSNorm num_attention_heads ............................. 32 num_channels .................................... 3 num_classes ..................................... 1000 num_experts ..................................... 8 num_layers ...................................... 32 num_layers_per_virtual_pipeline_stage ........... None num_query_groups ................................ 8 num_workers ..................................... 2 one_logger_entity ............................... hwinf_dcm one_logger_project .............................. e2e-tracking one_logger_run_name ............................. None onnx_safe ....................................... None openai_gelu ..................................... False optimizer ....................................... adam output_bert_embeddings .......................... False overlap_grad_reduce ............................. False overlap_p2p_comm ................................ False overlap_param_gather ............................ False override_opt_param_scheduler .................... False params_dtype .................................... torch.bfloat16 patch_dim ....................................... 16 perform_initialization .......................... True pipeline_model_parallel_size .................... 1 pipeline_model_parallel_split_rank .............. None position_embedding_type ......................... rope profile ......................................... True profile_ranks ................................... [0] profile_step_end ................................ 12 profile_step_start .............................. 10 qk_layernorm .................................... False query_in_block_prob ............................. 0.1 rampup_batch_size ............................... None rank ............................................ 0 recompute_granularity ........................... None recompute_method ................................ None recompute_num_layers ............................ None reset_attention_mask ............................ False reset_position_ids .............................. False retriever_report_topk_accuracies ................ [] retriever_score_scaling ......................... False retriever_seq_length ............................ 256 retro_add_retriever ............................. False retro_attention_gate ............................ 1 retro_cyclic_train_iters ........................ None retro_encoder_attention_dropout ................. 0.1 retro_encoder_hidden_dropout .................... 0.1 retro_encoder_layers ............................ 2 retro_num_neighbors ............................. 2 retro_num_retrieved_chunks ...................... 2 retro_project_dir ............................... None retro_verify_neighbor_count ..................... True rotary_interleaved .............................. False rotary_percent .................................. 1.0 rotary_seq_len_interpolation_factor ............. None sample_rate ..................................... 1.0 save ............................................ custom/ckpt/mixtral-8x7b save_interval ................................... 10000 scatter_gather_tensors_in_pipeline .............. True seed ............................................ 1234 seq_length ...................................... 2048 sequence_parallel ............................... True sgd_momentum .................................... 0.9 short_seq_prob .................................. 0.1 skip_train ...................................... False spec ............................................ None split ........................................... 99990,8,2 squared_relu .................................... False standalone_embedding_stage ...................... False start_weight_decay .............................. 0.1 swiglu .......................................... True swin_backbone_type .............................. tiny tensor_model_parallel_size ...................... 4 tensorboard_dir ................................. custom/ckpt/mixtral-8x7b/tensorboard tensorboard_log_interval ........................ 1 tensorboard_queue_size .......................... 1000 test_data_path .................................. None test_mode ....................................... False timing_log_level ................................ 0 timing_log_option ............................... minmax titles_data_path ................................ None tokenizer_model ................................. tokenizer.model tokenizer_type .................................. Llama2Tokenizer tp_comm_bulk_dgrad .............................. True tp_comm_bulk_wgrad .............................. True tp_comm_overlap ................................. False tp_comm_overlap_cfg ............................. None tp_comm_split_ag ................................ True tp_comm_split_rs ................................ True train_data_path ................................. None train_iters ..................................... 500000 train_samples ................................... None transformer_impl ................................ transformer_engine transformer_pipeline_model_parallel_size ........ 1 untie_embeddings_and_output_weights ............. True use_checkpoint_args ............................. False use_checkpoint_opt_param_scheduler .............. False use_cpu_initialization .......................... True use_dist_ckpt ................................... False use_distributed_optimizer ....................... True use_flash_attn .................................. True use_gpu_initialization .......................... None use_mcore_models ................................ True use_one_sent_docs ............................... False use_ring_exchange_p2p ........................... False use_rotary_position_embeddings .................. False valid_data_path ................................. None variable_seq_lengths ............................ False virtual_pipeline_model_parallel_size ............ None vision_backbone_type ............................ vit vision_pretraining .............................. False vision_pretraining_type ......................... classify vocab_extra_ids ................................. 0 vocab_file ...................................... None vocab_size ...................................... None wandb_exp_name .................................. mixtral-8x7b wandb_project ................................... megatron wandb_save_dir .................................. weight_decay .................................... 0.1 weight_decay_incr_style ......................... constant world_size ...................................... 16 yaml_cfg ........................................ None -------------------- end of arguments --------------------- setting number of micro-batches to constant 32 > building Llama2Tokenizer tokenizer ... > padded vocab (size: 32000) with 256 dummy tokens (new size: 32256) > initializing torch distributed ... > initialized tensor model parallel with size 4 > initialized pipeline model parallel with size 1 > setting random seeds to 1234 ... > compiling dataset index builder ... >>> done with dataset index builder. Compilation time: 0.087 seconds WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations. > compiling and loading fused kernels ... >>> done with compiling and loading fused kernels. Compilation time: 7.672 seconds [rank1]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank8]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank2]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank9]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank10]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank3]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank11]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank4]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank12]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank13]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank14]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank5]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank15]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank6]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank7]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank0]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) time to initialize megatron (seconds): 16.718 [after megatron is initialized] datetime: 2024-03-30 19:59:35 building GPT model ... > number of parameters on (tensor, pipeline) model parallel rank (3, 0): 3221491712 > number of parameters on (tensor, pipeline) model parallel rank (1, 0): 3221491712 > number of parameters on (tensor, pipeline) model parallel rank (2, 0): 3221491712 > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 3221491712 INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 1 INFO:megatron.core.distributed.param_and_grad_buffer:Params for bucket 1 (402919424 elements): INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.embedding.word_embeddings.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.final_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.output_layer.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 1 INFO:megatron.core.distributed.param_and_grad_buffer:Params for bucket 1 (2818572288 elements): INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.mlp.experts.weight1 INFO:megatron.core.optimizer:Setting up optimizer with OptimizerConfig(fp16=False, bf16=True, params_dtype=torch.bfloat16, optimizer='adam', lr=0.0001, loss_scale=None, initial_loss_scale=4294967296, min_loss_scale=1.0, loss_scale_window=1000, hysteresis=2, weight_decay=0.1, adam_beta1=0.9, adam_beta2=0.999, adam_eps=1e-08, sgd_momentum=0.9, use_distributed_optimizer=True, overlap_grad_reduce=False, overlap_param_gather=False, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=) > learning rate decay style: cosine WARNING: could not find the metadata file custom/ckpt/mixtral-8x7b/latest_checkpointed_iteration.txt will not load any checkpoints and will start from random > setting tensorboard ... (min, max) time across ranks (ms): load-checkpoint ................................: (0.65, 0.93) [after model, optimizer, and learning rate scheduler are built] datetime: 2024-03-30 20:00:43 > building train, validation, and test datasets ... > datasets target sizes (minimum size): train: 64000000 validation: 641280 test: 1280 INFO:megatron.core.datasets.blended_megatron_dataset_config:mock = True > building train, validation, and test datasets for GPT ... > finished creating GPT datasets ... [after dataloaders are built] datetime: 2024-03-30 20:00:43 done with setup ... training ... (min, max) time across ranks (ms): model-and-optimizer-setup ......................: (67432.26, 67517.36) train/valid/test-data-iterators-setup ..........: (3.74, 338.85) [before the start of training step] datetime: 2024-03-30 20:00:43 [Rank 0] (after 1 iterations) memory (MB) | allocated: 51955.275390625 | max allocated: 51955.291015625 | reserved: 62292.0 | max reserved: 62292.0 [Rank 1] (after 1 iterations) memory (MB) | allocated: 51955.275390625 | max allocated: 51955.291015625 | reserved: 62292.0 | max reserved: 62292.0 [2024-03-30 20:01:13] iteration 1/ 500000 | consumed samples: 128 | elapsed time per iteration (ms): 29428.3 | throughput per GPU (TFLOP/s/GPU): 44.4 | learning rate: 2.000E-07 | global batch size: 128 | lm loss: 1.038043E+01 | loss scale: 1.0 | grad norm: 526.452 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [Rank 3] (after 1 iterations) memory (MB) | allocated: 51955.275390625 | max allocated: 51955.291015625 | reserved: 62294.0 | max reserved: 62294.0 [Rank 2] (after 1 iterations) memory (MB) | allocated: 51955.275390625 | max allocated: 51955.291015625 | reserved: 62294.0 | max reserved: 62294.0 [2024-03-30 20:01:22] iteration 2/ 500000 | consumed samples: 256 | elapsed time per iteration (ms): 9845.6 | throughput per GPU (TFLOP/s/GPU): 132.6 | learning rate: 4.000E-07 | global batch size: 128 | lm loss: 1.047649E+01 | loss scale: 1.0 | grad norm: 506.118 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-03-30 20:01:32] iteration 3/ 500000 | consumed samples: 384 | elapsed time per iteration (ms): 9638.2 | throughput per GPU (TFLOP/s/GPU): 135.5 | learning rate: 6.000E-07 | global batch size: 128 | lm loss: 1.027612E+01 | loss scale: 1.0 | grad norm: 519.891 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-03-30 20:01:42] iteration 4/ 500000 | consumed samples: 512 | elapsed time per iteration (ms): 9702.8 | throughput per GPU (TFLOP/s/GPU): 134.6 | learning rate: 8.000E-07 | global batch size: 128 | lm loss: 9.807467E+00 | loss scale: 1.0 | grad norm: 517.413 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-03-30 20:01:51] iteration 5/ 500000 | consumed samples: 640 | elapsed time per iteration (ms): 9683.9 | throughput per GPU (TFLOP/s/GPU): 134.9 | learning rate: 1.000E-06 | global batch size: 128 | lm loss: 7.764119E+00 | loss scale: 1.0 | grad norm: 492.510 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-03-30 20:02:01] iteration 6/ 500000 | consumed samples: 768 | elapsed time per iteration (ms): 9675.6 | throughput per GPU (TFLOP/s/GPU): 135.0 | learning rate: 1.200E-06 | global batch size: 128 | lm loss: 2.630678E+00 | loss scale: 1.0 | grad norm: 323.002 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-03-30 20:02:11] iteration 7/ 500000 | consumed samples: 896 | elapsed time per iteration (ms): 9454.1 | throughput per GPU (TFLOP/s/GPU): 138.1 | learning rate: 1.400E-06 | global batch size: 128 | lm loss: 1.398795E+00 | loss scale: 1.0 | grad norm: 213.771 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-03-30 20:02:20] iteration 8/ 500000 | consumed samples: 1024 | elapsed time per iteration (ms): 9471.4 | throughput per GPU (TFLOP/s/GPU): 137.9 | learning rate: 1.600E-06 | global batch size: 128 | lm loss: 1.726107E+00 | loss scale: 1.0 | grad norm: 420.698 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-03-30 20:02:30] iteration 9/ 500000 | consumed samples: 1152 | elapsed time per iteration (ms): 10085.2 | throughput per GPU (TFLOP/s/GPU): 129.5 | learning rate: 1.800E-06 | global batch size: 128 | lm loss: 2.890289E-01 | loss scale: 1.0 | grad norm: 83.644 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-03-30 20:02:40] iteration 10/ 500000 | consumed samples: 1280 | elapsed time per iteration (ms): 9496.4 | throughput per GPU (TFLOP/s/GPU): 137.5 | learning rate: 2.000E-06 | global batch size: 128 | lm loss: 2.092005E-01 | loss scale: 1.0 | grad norm: 51.010 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-03-30 20:02:50] iteration 11/ 500000 | consumed samples: 1408 | elapsed time per iteration (ms): 10036.9 | throughput per GPU (TFLOP/s/GPU): 130.1 | learning rate: 2.200E-06 | global batch size: 128 | lm loss: 2.352597E-01 | loss scale: 1.0 | grad norm: 106.730 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-03-30 20:03:00] iteration 12/ 500000 | consumed samples: 1536 | elapsed time per iteration (ms): 10198.4 | throughput per GPU (TFLOP/s/GPU): 128.1 | learning rate: 2.400E-06 | global batch size: 128 | lm loss: 7.243721E-01 | loss scale: 1.0 | grad norm: 163.466 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-03-30 20:03:10] iteration 13/ 500000 | consumed samples: 1664 | elapsed time per iteration (ms): 10269.3 | throughput per GPU (TFLOP/s/GPU): 127.2 | learning rate: 2.600E-06 | global batch size: 128 | lm loss: 1.757669E+00 | loss scale: 1.0 | grad norm: 356.809 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-03-30 20:03:21] iteration 14/ 500000 | consumed samples: 1792 | elapsed time per iteration (ms): 10330.7 | throughput per GPU (TFLOP/s/GPU): 126.4 | learning rate: 2.800E-06 | global batch size: 128 | lm loss: 2.853365E-01 | loss scale: 1.0 | grad norm: 93.354 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-03-30 20:03:31] iteration 15/ 500000 | consumed samples: 1920 | elapsed time per iteration (ms): 10106.2 | throughput per GPU (TFLOP/s/GPU): 129.2 | learning rate: 3.000E-06 | global batch size: 128 | lm loss: 5.018836E-01 | loss scale: 1.0 | grad norm: 165.646 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-03-30 20:03:41] iteration 16/ 500000 | consumed samples: 2048 | elapsed time per iteration (ms): 10102.3 | throughput per GPU (TFLOP/s/GPU): 129.3 | learning rate: 3.200E-06 | global batch size: 128 | lm loss: 9.302688E-01 | loss scale: 1.0 | grad norm: 170.065 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-03-30 20:03:51] iteration 17/ 500000 | consumed samples: 2176 | elapsed time per iteration (ms): 9946.1 | throughput per GPU (TFLOP/s/GPU): 131.3 | learning rate: 3.400E-06 | global batch size: 128 | lm loss: 8.015128E-02 | loss scale: 1.0 | grad norm: 47.780 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | ```

Thank you for reporting this issue. 130 TFLOPS is indeed too low for the H100. I quickly reviewed your script and have some suggestions:

Update the code to the latest main branch and upgrade grouped_gemm to v1.0.
Use alltoall dispathcer: --moe-token-dispatcher-type alltoall.
Use EP8TP2.
Train for a while (at least 400 steps) before checking performance, or load a pretrained checkpoint. This is because router weights in early stage are not sufficiently trained, leading to imbalanced token distribution.

Hi, thanks for the suggestions. I retested the throuput according to your suggestion. To be more specific:

Update Megatron-LM the latest commit (https://github.com/NVIDIA/Megatron-LM/commit/ba773259dbe5735fbd91ca41e7f4ded60b335c52)
Update grouped_gemm to v1.0.0 (https://github.com/fanshiqing/grouped_gemm/commit/7a7f0189797889e926a30b3487512f9539161060)
Set --moe-token-dispatcher-type alltoall
Switch to EP=8 & TP=2
Use the pre-trained weights from Mixtral AI (converted from hf checkpoint)

The throughput has indeed increased significantly, reaching around 230 TFLOP/s. However, for H100, it's still pretty low, isn't it? May I ask, theoretically, what would be a more reasonable throughput?

Here is the logs

``` using world size: 16, data-parallel size: 8, context-parallel size: 1 tensor-model-parallel size: 2, pipeline-model-parallel size: 1 WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:Llama2Tokenizer WARNING: Setting args.overlap_p2p_comm to False since non-interleaved schedule does not support overlapping p2p communication accumulate and all-reduce gradients in fp32 for bfloat16 data type. using torch.bfloat16 for parameters ... ------------------------ arguments ------------------------ accumulate_allreduce_grads_in_fp32 .............. True adam_beta1 ...................................... 0.9 adam_beta2 ...................................... 0.999 adam_eps ........................................ 1e-08 add_bias_linear ................................. False add_position_embedding .......................... True add_qkv_bias .................................... False adlr_autoresume ................................. False adlr_autoresume_interval ........................ 1000 apply_layernorm_1p .............................. False apply_query_key_layer_scaling ................... False apply_residual_connection_post_layernorm ........ False apply_rope_fusion ............................... True async_tensor_model_parallel_allreduce ........... False attention_dropout ............................... 0.0 attention_softmax_in_fp32 ....................... False auto_detect_ckpt_format ......................... False barrier_with_L1_time ............................ True bert_binary_head ................................ True bert_embedder_type .............................. megatron bert_load ....................................... None bf16 ............................................ True bias_dropout_fusion ............................. True bias_gelu_fusion ................................ False bias_swiglu_fusion .............................. True biencoder_projection_dim ........................ 0 biencoder_shared_query_context_model ............ False block_data_path ................................. None check_for_nan_in_loss_and_grad .................. True ckpt_fully_parallel_save ........................ False ckpt_step ....................................... None classes_fraction ................................ 1.0 clip_grad ....................................... 1.0 clone_scatter_output_in_embedding ............... True consumed_train_samples .......................... 0 consumed_valid_samples .......................... 0 context_parallel_size ........................... 1 create_attention_mask_in_dataloader ............. True data_cache_path ................................. None data_parallel_random_init ....................... False data_parallel_size .............................. 8 data_path ....................................... ['custom/data/wudao/wudao_mistralbpe_content_document'] data_per_class_fraction ......................... 1.0 data_sharding ................................... True dataloader_type ................................. single decoder_num_layers .............................. None decoder_seq_length .............................. None decoupled_lr .................................... None decoupled_min_lr ................................ None delay_grad_reduce ............................... True delay_param_gather .............................. False dino_bottleneck_size ............................ 256 dino_freeze_last_layer .......................... 1 dino_head_hidden_size ........................... 2048 dino_local_crops_number ......................... 10 dino_local_img_size ............................. 96 dino_norm_last_layer ............................ False dino_teacher_temp ............................... 0.07 dino_warmup_teacher_temp ........................ 0.04 dino_warmup_teacher_temp_epochs ................. 30 dist_ckpt_format ................................ torch_dist distribute_saved_activations .................... False distributed_backend ............................. nccl distributed_timeout_minutes ..................... 10 embedding_path .................................. None empty_unused_memory_level ....................... 0 enable_one_logger ............................... False encoder_num_layers .............................. 32 encoder_seq_length .............................. 2048 end_weight_decay ................................ 0.1 eod_mask_loss ................................... False eval_interval ................................... 1000 eval_iters ...................................... 1 evidence_data_path .............................. None exit_duration_in_mins ........................... None exit_interval ................................... None exit_on_missing_checkpoint ...................... False exit_signal_handler ............................. False expert_model_parallel_size ...................... 8 ffn_hidden_size ................................. 14336 finetune ........................................ False fp16 ............................................ False fp16_lm_cross_entropy ........................... False fp32_residual_connection ........................ False fp8 ............................................. None fp8_amax_compute_algo ........................... most_recent fp8_amax_history_len ............................ 1 fp8_interval .................................... 1 fp8_margin ...................................... 0 fp8_wgrad ....................................... True global_batch_size ............................... 128 gradient_accumulation_fusion .................... True group_query_attention ........................... True head_lr_mult .................................... 1.0 hidden_dropout .................................. 0.0 hidden_size ..................................... 4096 hysteresis ...................................... 2 ict_head_size ................................... None ict_load ........................................ None img_h ........................................... 224 img_w ........................................... 224 indexer_batch_size .............................. 128 indexer_log_interval ............................ 1000 inference_batch_times_seqlen_threshold .......... 512 init_method_std ................................. 0.02 init_method_xavier_uniform ...................... False initial_loss_scale .............................. 4294967296 iter_per_epoch .................................. 1250 kv_channels ..................................... 128 lazy_mpu_init ................................... None load ............................................ custom/ckpt/mixtral-8x7b-tp2-ep8-mgg local_rank ...................................... None log_batch_size_to_tensorboard ................... False log_interval .................................... 1 log_learning_rate_to_tensorboard ................ True log_loss_scale_to_tensorboard ................... True log_memory_to_tensorboard ....................... False log_num_zeros_in_grad ........................... False log_params_norm ................................. False log_progress .................................... True log_throughput .................................. True log_timers_to_tensorboard ....................... False log_validation_ppl_to_tensorboard ............... False log_world_size_to_tensorboard ................... False loss_scale ...................................... None loss_scale_window ............................... 1000 lr .............................................. 0.0001 lr_decay_iters .................................. 320000 lr_decay_samples ................................ None lr_decay_style .................................. cosine lr_warmup_fraction .............................. None lr_warmup_init .................................. 0.0 lr_warmup_iters ................................. 500 lr_warmup_samples ............................... 0 make_vocab_size_divisible_by .................... 128 manual_gc ....................................... False manual_gc_eval .................................. True manual_gc_interval .............................. 0 mask_factor ..................................... 1.0 mask_prob ....................................... 0.15 mask_type ....................................... random masked_softmax_fusion ........................... False max_position_embeddings ......................... 32768 max_tokens_to_oom ............................... 12000 merge_file ...................................... None micro_batch_size ................................ 1 min_loss_scale .................................. 1.0 min_lr .......................................... 1e-05 mmap_bin_files .................................. True mock_data ....................................... False moe_aux_loss_coeff .............................. 0.01 moe_grouped_gemm ................................ True moe_input_jitter_eps ............................ None moe_per_layer_logging ........................... False moe_router_load_balancing_type .................. aux_loss moe_router_topk ................................. 2 moe_token_dispatcher_type ....................... alltoall moe_token_dropping .............................. False moe_z_loss_coeff ................................ None nccl_communicator_config_path ................... None no_load_optim ................................... True no_load_rng ..................................... True no_persist_layer_norm ........................... False no_save_optim ................................... None no_save_rng ..................................... None norm_epsilon .................................... 1e-05 normalization ................................... RMSNorm num_attention_heads ............................. 32 num_channels .................................... 3 num_classes ..................................... 1000 num_experts ..................................... 8 num_layers ...................................... 32 num_layers_per_virtual_pipeline_stage ........... None num_query_groups ................................ 8 num_workers ..................................... 2 one_logger_entity ............................... hwinf_dcm one_logger_project .............................. e2e-tracking one_logger_run_name ............................. None onnx_safe ....................................... None openai_gelu ..................................... False optimizer ....................................... adam output_bert_embeddings .......................... False overlap_grad_reduce ............................. False overlap_p2p_comm ................................ False overlap_param_gather ............................ False override_opt_param_scheduler .................... False params_dtype .................................... torch.bfloat16 patch_dim ....................................... 16 perform_initialization .......................... True pipeline_model_parallel_size .................... 1 pipeline_model_parallel_split_rank .............. None position_embedding_type ......................... rope pretrained_checkpoint ........................... None profile ......................................... True profile_ranks ................................... [0] profile_step_end ................................ 12 profile_step_start .............................. 10 qk_layernorm .................................... False query_in_block_prob ............................. 0.1 rampup_batch_size ............................... None rank ............................................ 0 recompute_granularity ........................... None recompute_method ................................ None recompute_num_layers ............................ None reset_attention_mask ............................ False reset_position_ids .............................. False retriever_report_topk_accuracies ................ [] retriever_score_scaling ......................... False retriever_seq_length ............................ 256 retro_add_retriever ............................. False retro_attention_gate ............................ 1 retro_cyclic_train_iters ........................ None retro_encoder_attention_dropout ................. 0.1 retro_encoder_hidden_dropout .................... 0.1 retro_encoder_layers ............................ 2 retro_num_neighbors ............................. 2 retro_num_retrieved_chunks ...................... 2 retro_project_dir ............................... None retro_verify_neighbor_count ..................... True rotary_interleaved .............................. False rotary_percent .................................. 1.0 rotary_seq_len_interpolation_factor ............. None sample_rate ..................................... 1.0 save ............................................ custom/ckpt/mixtral-8x7b-tp2-ep8-mgg save_interval ................................... 1000 scatter_gather_tensors_in_pipeline .............. True seed ............................................ 1234 seq_length ...................................... 2048 sequence_parallel ............................... True sgd_momentum .................................... 0.9 short_seq_prob .................................. 0.1 skip_train ...................................... False spec ............................................ None split ........................................... 99990,8,2 squared_relu .................................... False standalone_embedding_stage ...................... False start_weight_decay .............................. 0.1 swiglu .......................................... True swin_backbone_type .............................. tiny tensor_model_parallel_size ...................... 2 tensorboard_dir ................................. custom/ckpt/mixtral-8x7b-tp2-ep8-mgg/tensorboard tensorboard_log_interval ........................ 1 tensorboard_queue_size .......................... 1000 test_data_path .................................. None test_mode ....................................... False timing_log_level ................................ 0 timing_log_option ............................... minmax titles_data_path ................................ None tokenizer_model ................................. custom/ckpt/mixtral-8x7b/tokenizer.model tokenizer_type .................................. Llama2Tokenizer tp_comm_bulk_dgrad .............................. True tp_comm_bulk_wgrad .............................. True tp_comm_overlap ................................. False tp_comm_overlap_ag .............................. True tp_comm_overlap_cfg ............................. None tp_comm_overlap_rs .............................. True tp_comm_split_ag ................................ True tp_comm_split_rs ................................ True train_data_path ................................. None train_iters ..................................... 100 train_samples ................................... None transformer_impl ................................ transformer_engine transformer_pipeline_model_parallel_size ........ 1 untie_embeddings_and_output_weights ............. True use_checkpoint_args ............................. False use_checkpoint_opt_param_scheduler .............. False use_cpu_initialization .......................... None use_dist_ckpt ................................... False use_distributed_optimizer ....................... True use_flash_attn .................................. True use_mcore_models ................................ True use_one_sent_docs ............................... False use_ring_exchange_p2p ........................... False use_rotary_position_embeddings .................. False valid_data_path ................................. None variable_seq_lengths ............................ False virtual_pipeline_model_parallel_size ............ None vision_backbone_type ............................ vit vision_pretraining .............................. False vision_pretraining_type ......................... classify vocab_extra_ids ................................. 0 vocab_file ...................................... None vocab_size ...................................... None wandb_exp_name .................................. wandb_project ................................... wandb_save_dir .................................. weight_decay .................................... 0.1 weight_decay_incr_style ......................... constant world_size ...................................... 16 yaml_cfg ........................................ None -------------------- end of arguments --------------------- setting number of micro-batches to constant 16 > building Llama2Tokenizer tokenizer ... > padded vocab (size: 32000) with 0 dummy tokens (new size: 32000) > initializing torch distributed ... make: Entering directory '.../Megatron-LM/megatron/core/datasets' make: Nothing to be done for 'default'. make: Leaving directory '.../Megatron-LM/megatron/core/datasets' > initialized tensor model parallel with size 2 > initialized pipeline model parallel with size 1 > setting random seeds to 1234 ... > compiling dataset index builder ... >>> done with dataset index builder. Compilation time: 0.104 seconds WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations. > compiling and loading fused kernels ... >>> done with compiling and loading fused kernels. Compilation time: 7.866 seconds [rank1]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank8]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank2]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank9]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank10]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank0]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank3]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank11]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank4]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank12]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank5]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank13]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank6]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank7]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank14]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank15]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) time to initialize megatron (seconds): 14.235 [after megatron is initialized] datetime: 2024-04-06 02:54:57 building GPT model ... > number of parameters on (tensor, pipeline) model parallel rank (1, 0): 3622047744 > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 3622047744 INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 1 INFO:megatron.core.distributed.param_and_grad_buffer:Params for bucket 1 (803475456 elements): INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.embedding.word_embeddings.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.final_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.output_layer.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 1 INFO:megatron.core.distributed.param_and_grad_buffer:Params for bucket 1 (2818572288 elements): INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.mlp.experts.weight1 INFO:megatron.core.optimizer:Setting up optimizer with OptimizerConfig(optimizer='adam', lr=0.0001, min_lr=1e-05, decoupled_lr=None, decoupled_min_lr=None, weight_decay=0.1, fp16=False, bf16=True, params_dtype=torch.bfloat16, loss_scale=None, initial_loss_scale=4294967296, min_loss_scale=1.0, loss_scale_window=1000, hysteresis=2, adam_beta1=0.9, adam_beta2=0.999, adam_eps=1e-08, sgd_momentum=0.9, use_distributed_optimizer=True, overlap_grad_reduce=False, overlap_param_gather=False, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=) > learning rate decay style: cosine loading release checkpoint from custom/ckpt/mixtral-8x7b-tp2-ep8-mgg could not find arguments in the checkpoint ... checkpoint version 0 succesfully fixed query-key-values ordering for checkpoint version 0 successfully loaded checkpoint from custom/ckpt/mixtral-8x7b-tp2-ep8-mgg [ t 0, p 0 ] at iteration 0 > setting tensorboard ... (min, max) time across ranks (ms): load-checkpoint ................................: (8126.15, 8126.65) [after model, optimizer, and learning rate scheduler are built] datetime: 2024-04-06 02:55:06 > building train, validation, and test datasets ... > datasets target sizes (minimum size): train: 12800 validation: 128 test: 128 INFO:megatron.core.datasets.blended_megatron_dataset_config:mock = False INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 0.9999), (0.9999, 0.99998), (0.99998, 1.0)] > building train, validation, and test datasets for GPT ... INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from custom/data/wudao/wudao_mistralbpe_content_document.idx INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers INFO:megatron.core.datasets.indexed_dataset: Extract the document indices INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 59132211 INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 59132211 INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices INFO:megatron.core.datasets.gpt_dataset: Load the document index from cc3235b81bd7fd0fa07cabe05d15043d-GPTDataset-document_index.npy INFO:megatron.core.datasets.gpt_dataset: Load the sample index from cc3235b81bd7fd0fa07cabe05d15043d-GPTDataset-sample_index.npy INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from cc3235b81bd7fd0fa07cabe05d15043d-GPTDataset-shuffle_index.npy INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 40201537 INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset valid indices INFO:megatron.core.datasets.gpt_dataset: Load the document index from a625518736b8143e22f4f34c6682183e-GPTDataset-document_index.npy INFO:megatron.core.datasets.gpt_dataset: Load the sample index from a625518736b8143e22f4f34c6682183e-GPTDataset-sample_index.npy INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from a625518736b8143e22f4f34c6682183e-GPTDataset-shuffle_index.npy INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 6204 INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset test indices INFO:megatron.core.datasets.gpt_dataset: Load the document index from 052434ed70ae721ed70b2219cf2deb88-GPTDataset-document_index.npy INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 052434ed70ae721ed70b2219cf2deb88-GPTDataset-sample_index.npy INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 052434ed70ae721ed70b2219cf2deb88-GPTDataset-shuffle_index.npy INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 2332 > finished creating GPT datasets ... [after dataloaders are built] datetime: 2024-04-06 02:55:07 done with setup ... (min, max) time across ranks (ms): model-and-optimizer-setup ......................: (8592.94, 8605.02) train/valid/test-data-iterators-setup ..........: (569.02, 865.21) training ... [before the start of training step] datetime: 2024-04-06 02:55:07 Number of parameters in transformer layers in billions: 46.44 Number of parameters in embedding layers in billions: 0.26 Total number of parameters in billions: 46.70 Number of parameters in most loaded shard in billions: 23.3510 Theoretical memory footprints: weight and optimizer=167019.40 MB [Rank 0] (after 1 iterations) memory (MB) | allocated: 54250.97802734375 | max allocated: 54250.98583984375 | reserved: 61470.0 | max reserved: 61470.0 [2024-04-06 02:55:39] iteration 1/ 100 | consumed samples: 128 | elapsed time per iteration (ms): 32269.4 | throughput per GPU (TFLOP/s/GPU): 40.5 | learning rate: 2.000000E-07 | global batch size: 128 | lm loss: 1.985617E+00 | load_balancing_loss: 1.089786E+00 | loss scale: 1.0 | grad norm: 6.396 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [Rank 1] (after 1 iterations) memory (MB) | allocated: 54250.97802734375 | max allocated: 54250.98583984375 | reserved: 61480.0 | max reserved: 61480.0 [2024-04-06 02:55:45] iteration 2/ 100 | consumed samples: 256 | elapsed time per iteration (ms): 5630.1 | throughput per GPU (TFLOP/s/GPU): 231.9 | learning rate: 4.000000E-07 | global batch size: 128 | lm loss: 2.021530E+00 | load_balancing_loss: 1.087362E+00 | loss scale: 1.0 | grad norm: 6.895 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:55:50] iteration 3/ 100 | consumed samples: 384 | elapsed time per iteration (ms): 5410.6 | throughput per GPU (TFLOP/s/GPU): 241.4 | learning rate: 6.000000E-07 | global batch size: 128 | lm loss: 2.003316E+00 | load_balancing_loss: 1.085377E+00 | loss scale: 1.0 | grad norm: 6.603 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:55:55] iteration 4/ 100 | consumed samples: 512 | elapsed time per iteration (ms): 5364.1 | throughput per GPU (TFLOP/s/GPU): 243.5 | learning rate: 8.000000E-07 | global batch size: 128 | lm loss: 2.009657E+00 | load_balancing_loss: 1.091695E+00 | loss scale: 1.0 | grad norm: 6.619 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:01] iteration 5/ 100 | consumed samples: 640 | elapsed time per iteration (ms): 5496.7 | throughput per GPU (TFLOP/s/GPU): 237.6 | learning rate: 1.000000E-06 | global batch size: 128 | lm loss: 2.002326E+00 | load_balancing_loss: 1.091539E+00 | loss scale: 1.0 | grad norm: 6.612 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:06] iteration 6/ 100 | consumed samples: 768 | elapsed time per iteration (ms): 5364.8 | throughput per GPU (TFLOP/s/GPU): 243.4 | learning rate: 1.200000E-06 | global batch size: 128 | lm loss: 1.933151E+00 | load_balancing_loss: 1.086472E+00 | loss scale: 1.0 | grad norm: 5.765 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:12] iteration 7/ 100 | consumed samples: 896 | elapsed time per iteration (ms): 5682.7 | throughput per GPU (TFLOP/s/GPU): 229.8 | learning rate: 1.400000E-06 | global batch size: 128 | lm loss: 2.016085E+00 | load_balancing_loss: 1.085193E+00 | loss scale: 1.0 | grad norm: 5.821 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:17] iteration 8/ 100 | consumed samples: 1024 | elapsed time per iteration (ms): 5408.6 | throughput per GPU (TFLOP/s/GPU): 241.4 | learning rate: 1.600000E-06 | global batch size: 128 | lm loss: 1.965713E+00 | load_balancing_loss: 1.080933E+00 | loss scale: 1.0 | grad norm: 4.774 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:23] iteration 9/ 100 | consumed samples: 1152 | elapsed time per iteration (ms): 5590.1 | throughput per GPU (TFLOP/s/GPU): 233.6 | learning rate: 1.800000E-06 | global batch size: 128 | lm loss: 1.919308E+00 | load_balancing_loss: 1.089582E+00 | loss scale: 1.0 | grad norm: 4.267 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:28] iteration 10/ 100 | consumed samples: 1280 | elapsed time per iteration (ms): 5443.7 | throughput per GPU (TFLOP/s/GPU): 239.9 | learning rate: 2.000000E-06 | global batch size: 128 | lm loss: 1.978377E+00 | load_balancing_loss: 1.089948E+00 | loss scale: 1.0 | grad norm: 4.069 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:34] iteration 11/ 100 | consumed samples: 1408 | elapsed time per iteration (ms): 5984.1 | throughput per GPU (TFLOP/s/GPU): 218.2 | learning rate: 2.200000E-06 | global batch size: 128 | lm loss: 1.889895E+00 | load_balancing_loss: 1.083618E+00 | loss scale: 1.0 | grad norm: 3.361 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:40] iteration 12/ 100 | consumed samples: 1536 | elapsed time per iteration (ms): 5821.8 | throughput per GPU (TFLOP/s/GPU): 224.3 | learning rate: 2.400000E-06 | global batch size: 128 | lm loss: 1.932808E+00 | load_balancing_loss: 1.085315E+00 | loss scale: 1.0 | grad norm: 3.336 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:46] iteration 13/ 100 | consumed samples: 1664 | elapsed time per iteration (ms): 5962.2 | throughput per GPU (TFLOP/s/GPU): 219.0 | learning rate: 2.600000E-06 | global batch size: 128 | lm loss: 1.911683E+00 | load_balancing_loss: 1.079515E+00 | loss scale: 1.0 | grad norm: 3.183 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:52] iteration 14/ 100 | consumed samples: 1792 | elapsed time per iteration (ms): 5927.4 | throughput per GPU (TFLOP/s/GPU): 220.3 | learning rate: 2.800000E-06 | global batch size: 128 | lm loss: 1.913695E+00 | load_balancing_loss: 1.076165E+00 | loss scale: 1.0 | grad norm: 2.994 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:58] iteration 15/ 100 | consumed samples: 1920 | elapsed time per iteration (ms): 5926.4 | throughput per GPU (TFLOP/s/GPU): 220.4 | learning rate: 3.000000E-06 | global batch size: 128 | lm loss: 1.957101E+00 | load_balancing_loss: 1.069903E+00 | loss scale: 1.0 | grad norm: 2.853 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:04] iteration 16/ 100 | consumed samples: 2048 | elapsed time per iteration (ms): 5912.7 | throughput per GPU (TFLOP/s/GPU): 220.9 | learning rate: 3.200000E-06 | global batch size: 128 | lm loss: 1.915763E+00 | load_balancing_loss: 1.065748E+00 | loss scale: 1.0 | grad norm: 2.778 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:10] iteration 17/ 100 | consumed samples: 2176 | elapsed time per iteration (ms): 5706.3 | throughput per GPU (TFLOP/s/GPU): 228.9 | learning rate: 3.400000E-06 | global batch size: 128 | lm loss: 1.918353E+00 | load_balancing_loss: 1.064678E+00 | loss scale: 1.0 | grad norm: 2.911 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:15] iteration 18/ 100 | consumed samples: 2304 | elapsed time per iteration (ms): 5732.8 | throughput per GPU (TFLOP/s/GPU): 227.8 | learning rate: 3.600000E-06 | global batch size: 128 | lm loss: 1.861051E+00 | load_balancing_loss: 1.058054E+00 | loss scale: 1.0 | grad norm: 2.449 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:21] iteration 19/ 100 | consumed samples: 2432 | elapsed time per iteration (ms): 5684.9 | throughput per GPU (TFLOP/s/GPU): 229.7 | learning rate: 3.800000E-06 | global batch size: 128 | lm loss: 1.934895E+00 | load_balancing_loss: 1.049081E+00 | loss scale: 1.0 | grad norm: 2.447 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:27] iteration 20/ 100 | consumed samples: 2560 | elapsed time per iteration (ms): 5770.6 | throughput per GPU (TFLOP/s/GPU): 226.3 | learning rate: 4.000000E-06 | global batch size: 128 | lm loss: 1.932632E+00 | load_balancing_loss: 1.052491E+00 | loss scale: 1.0 | grad norm: 2.456 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:32] iteration 21/ 100 | consumed samples: 2688 | elapsed time per iteration (ms): 5541.8 | throughput per GPU (TFLOP/s/GPU): 235.6 | learning rate: 4.200000E-06 | global batch size: 128 | lm loss: 1.904877E+00 | load_balancing_loss: 1.047207E+00 | loss scale: 1.0 | grad norm: 2.213 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:38] iteration 22/ 100 | consumed samples: 2816 | elapsed time per iteration (ms): 5576.7 | throughput per GPU (TFLOP/s/GPU): 234.2 | learning rate: 4.400000E-06 | global batch size: 128 | lm loss: 1.872380E+00 | load_balancing_loss: 1.039512E+00 | loss scale: 1.0 | grad norm: 2.116 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:44] iteration 23/ 100 | consumed samples: 2944 | elapsed time per iteration (ms): 5807.4 | throughput per GPU (TFLOP/s/GPU): 224.9 | learning rate: 4.600000E-06 | global batch size: 128 | lm loss: 1.835408E+00 | load_balancing_loss: 1.042104E+00 | loss scale: 1.0 | grad norm: 2.034 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:50] iteration 24/ 100 | consumed samples: 3072 | elapsed time per iteration (ms): 5727.3 | throughput per GPU (TFLOP/s/GPU): 228.0 | learning rate: 4.800000E-06 | global batch size: 128 | lm loss: 1.898657E+00 | load_balancing_loss: 1.029742E+00 | loss scale: 1.0 | grad norm: 1.982 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:55] iteration 25/ 100 | consumed samples: 3200 | elapsed time per iteration (ms): 5498.4 | throughput per GPU (TFLOP/s/GPU): 237.5 | learning rate: 5.000000E-06 | global batch size: 128 | lm loss: 1.904866E+00 | load_balancing_loss: 1.034888E+00 | loss scale: 1.0 | grad norm: 1.872 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:01] iteration 26/ 100 | consumed samples: 3328 | elapsed time per iteration (ms): 5531.7 | throughput per GPU (TFLOP/s/GPU): 236.1 | learning rate: 5.200000E-06 | global batch size: 128 | lm loss: 1.889752E+00 | load_balancing_loss: 1.028931E+00 | loss scale: 1.0 | grad norm: 1.793 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:06] iteration 27/ 100 | consumed samples: 3456 | elapsed time per iteration (ms): 5678.3 | throughput per GPU (TFLOP/s/GPU): 230.0 | learning rate: 5.400000E-06 | global batch size: 128 | lm loss: 1.866109E+00 | load_balancing_loss: 1.031736E+00 | loss scale: 1.0 | grad norm: 1.773 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:12] iteration 28/ 100 | consumed samples: 3584 | elapsed time per iteration (ms): 5650.6 | throughput per GPU (TFLOP/s/GPU): 231.1 | learning rate: 5.600000E-06 | global batch size: 128 | lm loss: 1.914117E+00 | load_balancing_loss: 1.027364E+00 | loss scale: 1.0 | grad norm: 1.709 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:18] iteration 29/ 100 | consumed samples: 3712 | elapsed time per iteration (ms): 5912.1 | throughput per GPU (TFLOP/s/GPU): 220.9 | learning rate: 5.800000E-06 | global batch size: 128 | lm loss: 1.867856E+00 | load_balancing_loss: 1.023825E+00 | loss scale: 1.0 | grad norm: 1.769 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:23] iteration 30/ 100 | consumed samples: 3840 | elapsed time per iteration (ms): 5571.1 | throughput per GPU (TFLOP/s/GPU): 234.4 | learning rate: 6.000000E-06 | global batch size: 128 | lm loss: 1.924535E+00 | load_balancing_loss: 1.025294E+00 | loss scale: 1.0 | grad norm: 1.572 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:29] iteration 31/ 100 | consumed samples: 3968 | elapsed time per iteration (ms): 5718.9 | throughput per GPU (TFLOP/s/GPU): 228.3 | learning rate: 6.200000E-06 | global batch size: 128 | lm loss: 1.830754E+00 | load_balancing_loss: 1.028048E+00 | loss scale: 1.0 | grad norm: 1.555 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:35] iteration 32/ 100 | consumed samples: 4096 | elapsed time per iteration (ms): 5629.3 | throughput per GPU (TFLOP/s/GPU): 232.0 | learning rate: 6.400000E-06 | global batch size: 128 | lm loss: 1.848776E+00 | load_balancing_loss: 1.021549E+00 | loss scale: 1.0 | grad norm: 1.592 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:40] iteration 33/ 100 | consumed samples: 4224 | elapsed time per iteration (ms): 5600.4 | throughput per GPU (TFLOP/s/GPU): 233.2 | learning rate: 6.600000E-06 | global batch size: 128 | lm loss: 1.917658E+00 | load_balancing_loss: 1.032319E+00 | loss scale: 1.0 | grad norm: 1.519 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:46] iteration 34/ 100 | consumed samples: 4352 | elapsed time per iteration (ms): 5643.8 | throughput per GPU (TFLOP/s/GPU): 231.4 | learning rate: 6.800000E-06 | global batch size: 128 | lm loss: 1.844636E+00 | load_balancing_loss: 1.019185E+00 | loss scale: 1.0 | grad norm: 1.626 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:51] iteration 35/ 100 | consumed samples: 4480 | elapsed time per iteration (ms): 5367.8 | throughput per GPU (TFLOP/s/GPU): 243.3 | learning rate: 7.000000E-06 | global batch size: 128 | lm loss: 1.853418E+00 | load_balancing_loss: 1.020990E+00 | loss scale: 1.0 | grad norm: 1.760 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:57] iteration 36/ 100 | consumed samples: 4608 | elapsed time per iteration (ms): 5399.9 | throughput per GPU (TFLOP/s/GPU): 241.8 | learning rate: 7.200000E-06 | global batch size: 128 | lm loss: 1.842918E+00 | load_balancing_loss: 1.023077E+00 | loss scale: 1.0 | grad norm: 1.409 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:02] iteration 37/ 100 | consumed samples: 4736 | elapsed time per iteration (ms): 5515.8 | throughput per GPU (TFLOP/s/GPU): 236.8 | learning rate: 7.400000E-06 | global batch size: 128 | lm loss: 1.862270E+00 | load_balancing_loss: 1.023782E+00 | loss scale: 1.0 | grad norm: 1.718 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:08] iteration 38/ 100 | consumed samples: 4864 | elapsed time per iteration (ms): 5477.8 | throughput per GPU (TFLOP/s/GPU): 238.4 | learning rate: 7.600000E-06 | global batch size: 128 | lm loss: 1.862543E+00 | load_balancing_loss: 1.019304E+00 | loss scale: 1.0 | grad norm: 1.722 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:13] iteration 39/ 100 | consumed samples: 4992 | elapsed time per iteration (ms): 5649.1 | throughput per GPU (TFLOP/s/GPU): 231.2 | learning rate: 7.800000E-06 | global batch size: 128 | lm loss: 1.863421E+00 | load_balancing_loss: 1.017805E+00 | loss scale: 1.0 | grad norm: 1.469 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:19] iteration 40/ 100 | consumed samples: 5120 | elapsed time per iteration (ms): 5810.4 | throughput per GPU (TFLOP/s/GPU): 224.8 | learning rate: 8.000000E-06 | global batch size: 128 | lm loss: 1.879655E+00 | load_balancing_loss: 1.017568E+00 | loss scale: 1.0 | grad norm: 1.633 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:25] iteration 41/ 100 | consumed samples: 5248 | elapsed time per iteration (ms): 5462.9 | throughput per GPU (TFLOP/s/GPU): 239.1 | learning rate: 8.200000E-06 | global batch size: 128 | lm loss: 1.812076E+00 | load_balancing_loss: 1.020508E+00 | loss scale: 1.0 | grad norm: 1.419 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:30] iteration 42/ 100 | consumed samples: 5376 | elapsed time per iteration (ms): 5452.3 | throughput per GPU (TFLOP/s/GPU): 239.5 | learning rate: 8.400000E-06 | global batch size: 128 | lm loss: 1.824542E+00 | load_balancing_loss: 1.017472E+00 | loss scale: 1.0 | grad norm: 1.400 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:36] iteration 43/ 100 | consumed samples: 5504 | elapsed time per iteration (ms): 5444.9 | throughput per GPU (TFLOP/s/GPU): 239.8 | learning rate: 8.600000E-06 | global batch size: 128 | lm loss: 1.825991E+00 | load_balancing_loss: 1.019746E+00 | loss scale: 1.0 | grad norm: 1.426 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:41] iteration 44/ 100 | consumed samples: 5632 | elapsed time per iteration (ms): 5533.8 | throughput per GPU (TFLOP/s/GPU): 236.0 | learning rate: 8.800000E-06 | global batch size: 128 | lm loss: 1.875063E+00 | load_balancing_loss: 1.020033E+00 | loss scale: 1.0 | grad norm: 1.327 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:47] iteration 45/ 100 | consumed samples: 5760 | elapsed time per iteration (ms): 5718.6 | throughput per GPU (TFLOP/s/GPU): 228.4 | learning rate: 9.000000E-06 | global batch size: 128 | lm loss: 1.834162E+00 | load_balancing_loss: 1.018004E+00 | loss scale: 1.0 | grad norm: 1.611 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:52] iteration 46/ 100 | consumed samples: 5888 | elapsed time per iteration (ms): 5567.2 | throughput per GPU (TFLOP/s/GPU): 234.6 | learning rate: 9.200000E-06 | global batch size: 128 | lm loss: 1.883577E+00 | load_balancing_loss: 1.016062E+00 | loss scale: 1.0 | grad norm: 1.439 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:58] iteration 47/ 100 | consumed samples: 6016 | elapsed time per iteration (ms): 5692.2 | throughput per GPU (TFLOP/s/GPU): 229.4 | learning rate: 9.400000E-06 | global batch size: 128 | lm loss: 1.836727E+00 | load_balancing_loss: 1.019520E+00 | loss scale: 1.0 | grad norm: 1.372 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:04] iteration 48/ 100 | consumed samples: 6144 | elapsed time per iteration (ms): 5872.4 | throughput per GPU (TFLOP/s/GPU): 222.4 | learning rate: 9.600000E-06 | global batch size: 128 | lm loss: 1.855191E+00 | load_balancing_loss: 1.017754E+00 | loss scale: 1.0 | grad norm: 1.508 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:09] iteration 49/ 100 | consumed samples: 6272 | elapsed time per iteration (ms): 5528.7 | throughput per GPU (TFLOP/s/GPU): 236.2 | learning rate: 9.800000E-06 | global batch size: 128 | lm loss: 1.806294E+00 | load_balancing_loss: 1.017504E+00 | loss scale: 1.0 | grad norm: 1.529 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:15] iteration 50/ 100 | consumed samples: 6400 | elapsed time per iteration (ms): 5531.5 | throughput per GPU (TFLOP/s/GPU): 236.1 | learning rate: 1.000000E-05 | global batch size: 128 | lm loss: 1.887587E+00 | load_balancing_loss: 1.016094E+00 | loss scale: 1.0 | grad norm: 1.439 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:21] iteration 51/ 100 | consumed samples: 6528 | elapsed time per iteration (ms): 5501.3 | throughput per GPU (TFLOP/s/GPU): 237.4 | learning rate: 1.020000E-05 | global batch size: 128 | lm loss: 1.834414E+00 | load_balancing_loss: 1.015084E+00 | loss scale: 1.0 | grad norm: 1.599 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:26] iteration 52/ 100 | consumed samples: 6656 | elapsed time per iteration (ms): 5520.9 | throughput per GPU (TFLOP/s/GPU): 236.5 | learning rate: 1.040000E-05 | global batch size: 128 | lm loss: 1.847078E+00 | load_balancing_loss: 1.015950E+00 | loss scale: 1.0 | grad norm: 1.486 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:32] iteration 53/ 100 | consumed samples: 6784 | elapsed time per iteration (ms): 5711.6 | throughput per GPU (TFLOP/s/GPU): 228.6 | learning rate: 1.060000E-05 | global batch size: 128 | lm loss: 1.862840E+00 | load_balancing_loss: 1.016317E+00 | loss scale: 1.0 | grad norm: 1.522 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:37] iteration 54/ 100 | consumed samples: 6912 | elapsed time per iteration (ms): 5689.4 | throughput per GPU (TFLOP/s/GPU): 229.5 | learning rate: 1.080000E-05 | global batch size: 128 | lm loss: 1.897956E+00 | load_balancing_loss: 1.017408E+00 | loss scale: 1.0 | grad norm: 1.383 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:43] iteration 55/ 100 | consumed samples: 7040 | elapsed time per iteration (ms): 5763.8 | throughput per GPU (TFLOP/s/GPU): 226.6 | learning rate: 1.100000E-05 | global batch size: 128 | lm loss: 1.863309E+00 | load_balancing_loss: 1.014457E+00 | loss scale: 1.0 | grad norm: 1.534 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:49] iteration 56/ 100 | consumed samples: 7168 | elapsed time per iteration (ms): 5742.1 | throughput per GPU (TFLOP/s/GPU): 227.4 | learning rate: 1.120000E-05 | global batch size: 128 | lm loss: 1.899538E+00 | load_balancing_loss: 1.018558E+00 | loss scale: 1.0 | grad norm: 1.470 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:54] iteration 57/ 100 | consumed samples: 7296 | elapsed time per iteration (ms): 5450.5 | throughput per GPU (TFLOP/s/GPU): 239.6 | learning rate: 1.140000E-05 | global batch size: 128 | lm loss: 1.864605E+00 | load_balancing_loss: 1.015150E+00 | loss scale: 1.0 | grad norm: 1.244 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:00] iteration 58/ 100 | consumed samples: 7424 | elapsed time per iteration (ms): 5538.9 | throughput per GPU (TFLOP/s/GPU): 235.8 | learning rate: 1.160000E-05 | global batch size: 128 | lm loss: 1.812579E+00 | load_balancing_loss: 1.020851E+00 | loss scale: 1.0 | grad norm: 1.610 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:05] iteration 59/ 100 | consumed samples: 7552 | elapsed time per iteration (ms): 5410.9 | throughput per GPU (TFLOP/s/GPU): 241.3 | learning rate: 1.180000E-05 | global batch size: 128 | lm loss: 1.848337E+00 | load_balancing_loss: 1.013638E+00 | loss scale: 1.0 | grad norm: 1.351 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:11] iteration 60/ 100 | consumed samples: 7680 | elapsed time per iteration (ms): 5603.1 | throughput per GPU (TFLOP/s/GPU): 233.1 | learning rate: 1.200000E-05 | global batch size: 128 | lm loss: 1.801180E+00 | load_balancing_loss: 1.019084E+00 | loss scale: 1.0 | grad norm: 1.549 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:16] iteration 61/ 100 | consumed samples: 7808 | elapsed time per iteration (ms): 5495.5 | throughput per GPU (TFLOP/s/GPU): 237.6 | learning rate: 1.220000E-05 | global batch size: 128 | lm loss: 1.813972E+00 | load_balancing_loss: 1.014779E+00 | loss scale: 1.0 | grad norm: 1.427 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:22] iteration 62/ 100 | consumed samples: 7936 | elapsed time per iteration (ms): 5753.1 | throughput per GPU (TFLOP/s/GPU): 227.0 | learning rate: 1.240000E-05 | global batch size: 128 | lm loss: 1.808689E+00 | load_balancing_loss: 1.022012E+00 | loss scale: 1.0 | grad norm: 1.398 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:28] iteration 63/ 100 | consumed samples: 8064 | elapsed time per iteration (ms): 5650.1 | throughput per GPU (TFLOP/s/GPU): 231.1 | learning rate: 1.260000E-05 | global batch size: 128 | lm loss: 1.781526E+00 | load_balancing_loss: 1.013716E+00 | loss scale: 1.0 | grad norm: 1.494 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:33] iteration 64/ 100 | consumed samples: 8192 | elapsed time per iteration (ms): 5539.7 | throughput per GPU (TFLOP/s/GPU): 235.7 | learning rate: 1.280000E-05 | global batch size: 128 | lm loss: 1.871476E+00 | load_balancing_loss: 1.019044E+00 | loss scale: 1.0 | grad norm: 1.369 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:39] iteration 65/ 100 | consumed samples: 8320 | elapsed time per iteration (ms): 5493.9 | throughput per GPU (TFLOP/s/GPU): 237.7 | learning rate: 1.300000E-05 | global batch size: 128 | lm loss: 1.846450E+00 | load_balancing_loss: 1.017387E+00 | loss scale: 1.0 | grad norm: 1.308 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:44] iteration 66/ 100 | consumed samples: 8448 | elapsed time per iteration (ms): 5590.8 | throughput per GPU (TFLOP/s/GPU): 233.6 | learning rate: 1.320000E-05 | global batch size: 128 | lm loss: 1.873755E+00 | load_balancing_loss: 1.014257E+00 | loss scale: 1.0 | grad norm: 1.411 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:50] iteration 67/ 100 | consumed samples: 8576 | elapsed time per iteration (ms): 5710.3 | throughput per GPU (TFLOP/s/GPU): 228.7 | learning rate: 1.340000E-05 | global batch size: 128 | lm loss: 1.765591E+00 | load_balancing_loss: 1.016482E+00 | loss scale: 1.0 | grad norm: 1.414 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:56] iteration 68/ 100 | consumed samples: 8704 | elapsed time per iteration (ms): 5734.5 | throughput per GPU (TFLOP/s/GPU): 227.7 | learning rate: 1.360000E-05 | global batch size: 128 | lm loss: 1.839895E+00 | load_balancing_loss: 1.012786E+00 | loss scale: 1.0 | grad norm: 1.371 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:01] iteration 69/ 100 | consumed samples: 8832 | elapsed time per iteration (ms): 5478.6 | throughput per GPU (TFLOP/s/GPU): 238.4 | learning rate: 1.380000E-05 | global batch size: 128 | lm loss: 1.912256E+00 | load_balancing_loss: 1.013041E+00 | loss scale: 1.0 | grad norm: 1.485 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:07] iteration 70/ 100 | consumed samples: 8960 | elapsed time per iteration (ms): 5514.8 | throughput per GPU (TFLOP/s/GPU): 236.8 | learning rate: 1.400000E-05 | global batch size: 128 | lm loss: 1.873068E+00 | load_balancing_loss: 1.012509E+00 | loss scale: 1.0 | grad norm: 1.467 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:12] iteration 71/ 100 | consumed samples: 9088 | elapsed time per iteration (ms): 5361.6 | throughput per GPU (TFLOP/s/GPU): 243.6 | learning rate: 1.420000E-05 | global batch size: 128 | lm loss: 1.818812E+00 | load_balancing_loss: 1.013377E+00 | loss scale: 1.0 | grad norm: 1.300 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:18] iteration 72/ 100 | consumed samples: 9216 | elapsed time per iteration (ms): 5470.7 | throughput per GPU (TFLOP/s/GPU): 238.7 | learning rate: 1.440000E-05 | global batch size: 128 | lm loss: 1.820313E+00 | load_balancing_loss: 1.019612E+00 | loss scale: 1.0 | grad norm: 1.305 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:24] iteration 73/ 100 | consumed samples: 9344 | elapsed time per iteration (ms): 5829.9 | throughput per GPU (TFLOP/s/GPU): 224.0 | learning rate: 1.460000E-05 | global batch size: 128 | lm loss: 1.798953E+00 | load_balancing_loss: 1.010977E+00 | loss scale: 1.0 | grad norm: 1.539 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:29] iteration 74/ 100 | consumed samples: 9472 | elapsed time per iteration (ms): 5702.4 | throughput per GPU (TFLOP/s/GPU): 229.0 | learning rate: 1.480000E-05 | global batch size: 128 | lm loss: 1.774078E+00 | load_balancing_loss: 1.012441E+00 | loss scale: 1.0 | grad norm: 1.471 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:35] iteration 75/ 100 | consumed samples: 9600 | elapsed time per iteration (ms): 5599.5 | throughput per GPU (TFLOP/s/GPU): 233.2 | learning rate: 1.500000E-05 | global batch size: 128 | lm loss: 1.838492E+00 | load_balancing_loss: 1.015038E+00 | loss scale: 1.0 | grad norm: 1.445 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:40] iteration 76/ 100 | consumed samples: 9728 | elapsed time per iteration (ms): 5588.2 | throughput per GPU (TFLOP/s/GPU): 233.7 | learning rate: 1.520000E-05 | global batch size: 128 | lm loss: 1.860703E+00 | load_balancing_loss: 1.012689E+00 | loss scale: 1.0 | grad norm: 1.500 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:46] iteration 77/ 100 | consumed samples: 9856 | elapsed time per iteration (ms): 5425.4 | throughput per GPU (TFLOP/s/GPU): 240.7 | learning rate: 1.540000E-05 | global batch size: 128 | lm loss: 1.827507E+00 | load_balancing_loss: 1.012502E+00 | loss scale: 1.0 | grad norm: 1.491 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:52] iteration 78/ 100 | consumed samples: 9984 | elapsed time per iteration (ms): 5652.9 | throughput per GPU (TFLOP/s/GPU): 231.0 | learning rate: 1.560000E-05 | global batch size: 128 | lm loss: 1.784492E+00 | load_balancing_loss: 1.013809E+00 | loss scale: 1.0 | grad norm: 1.407 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:57] iteration 79/ 100 | consumed samples: 10112 | elapsed time per iteration (ms): 5577.0 | throughput per GPU (TFLOP/s/GPU): 234.2 | learning rate: 1.580000E-05 | global batch size: 128 | lm loss: 1.858489E+00 | load_balancing_loss: 1.011662E+00 | loss scale: 1.0 | grad norm: 1.621 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:03] iteration 80/ 100 | consumed samples: 10240 | elapsed time per iteration (ms): 5712.8 | throughput per GPU (TFLOP/s/GPU): 228.6 | learning rate: 1.600000E-05 | global batch size: 128 | lm loss: 1.842588E+00 | load_balancing_loss: 1.011640E+00 | loss scale: 1.0 | grad norm: 1.631 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:09] iteration 81/ 100 | consumed samples: 10368 | elapsed time per iteration (ms): 5684.5 | throughput per GPU (TFLOP/s/GPU): 229.7 | learning rate: 1.620000E-05 | global batch size: 128 | lm loss: 1.818980E+00 | load_balancing_loss: 1.012697E+00 | loss scale: 1.0 | grad norm: 1.564 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:14] iteration 82/ 100 | consumed samples: 10496 | elapsed time per iteration (ms): 5592.0 | throughput per GPU (TFLOP/s/GPU): 233.5 | learning rate: 1.640000E-05 | global batch size: 128 | lm loss: 1.805010E+00 | load_balancing_loss: 1.012805E+00 | loss scale: 1.0 | grad norm: 1.545 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:20] iteration 83/ 100 | consumed samples: 10624 | elapsed time per iteration (ms): 5641.6 | throughput per GPU (TFLOP/s/GPU): 231.5 | learning rate: 1.660000E-05 | global batch size: 128 | lm loss: 1.812314E+00 | load_balancing_loss: 1.011967E+00 | loss scale: 1.0 | grad norm: 1.530 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:25] iteration 84/ 100 | consumed samples: 10752 | elapsed time per iteration (ms): 5563.7 | throughput per GPU (TFLOP/s/GPU): 234.7 | learning rate: 1.680000E-05 | global batch size: 128 | lm loss: 1.822110E+00 | load_balancing_loss: 1.009684E+00 | loss scale: 1.0 | grad norm: 1.799 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:31] iteration 85/ 100 | consumed samples: 10880 | elapsed time per iteration (ms): 5580.9 | throughput per GPU (TFLOP/s/GPU): 234.0 | learning rate: 1.700000E-05 | global batch size: 128 | lm loss: 1.831795E+00 | load_balancing_loss: 1.009344E+00 | loss scale: 1.0 | grad norm: 1.578 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:37] iteration 86/ 100 | consumed samples: 11008 | elapsed time per iteration (ms): 5695.8 | throughput per GPU (TFLOP/s/GPU): 229.3 | learning rate: 1.720000E-05 | global batch size: 128 | lm loss: 1.831625E+00 | load_balancing_loss: 1.011533E+00 | loss scale: 1.0 | grad norm: 1.515 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:42] iteration 87/ 100 | consumed samples: 11136 | elapsed time per iteration (ms): 5444.5 | throughput per GPU (TFLOP/s/GPU): 239.9 | learning rate: 1.740000E-05 | global batch size: 128 | lm loss: 1.814374E+00 | load_balancing_loss: 1.010052E+00 | loss scale: 1.0 | grad norm: 1.365 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:48] iteration 88/ 100 | consumed samples: 11264 | elapsed time per iteration (ms): 5462.7 | throughput per GPU (TFLOP/s/GPU): 239.1 | learning rate: 1.760000E-05 | global batch size: 128 | lm loss: 1.825778E+00 | load_balancing_loss: 1.010838E+00 | loss scale: 1.0 | grad norm: 1.506 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:53] iteration 89/ 100 | consumed samples: 11392 | elapsed time per iteration (ms): 5633.2 | throughput per GPU (TFLOP/s/GPU): 231.8 | learning rate: 1.780000E-05 | global batch size: 128 | lm loss: 1.818898E+00 | load_balancing_loss: 1.011014E+00 | loss scale: 1.0 | grad norm: 1.358 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:59] iteration 90/ 100 | consumed samples: 11520 | elapsed time per iteration (ms): 5567.8 | throughput per GPU (TFLOP/s/GPU): 234.5 | learning rate: 1.800000E-05 | global batch size: 128 | lm loss: 1.813602E+00 | load_balancing_loss: 1.022434E+00 | loss scale: 1.0 | grad norm: 1.590 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:04] iteration 91/ 100 | consumed samples: 11648 | elapsed time per iteration (ms): 5691.9 | throughput per GPU (TFLOP/s/GPU): 229.4 | learning rate: 1.820000E-05 | global batch size: 128 | lm loss: 1.797111E+00 | load_balancing_loss: 1.011964E+00 | loss scale: 1.0 | grad norm: 1.436 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:10] iteration 92/ 100 | consumed samples: 11776 | elapsed time per iteration (ms): 5451.5 | throughput per GPU (TFLOP/s/GPU): 239.6 | learning rate: 1.840000E-05 | global batch size: 128 | lm loss: 1.809117E+00 | load_balancing_loss: 1.012038E+00 | loss scale: 1.0 | grad norm: 1.577 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:15] iteration 93/ 100 | consumed samples: 11904 | elapsed time per iteration (ms): 5599.2 | throughput per GPU (TFLOP/s/GPU): 233.2 | learning rate: 1.860000E-05 | global batch size: 128 | lm loss: 1.797812E+00 | load_balancing_loss: 1.011838E+00 | loss scale: 1.0 | grad norm: 1.553 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:21] iteration 94/ 100 | consumed samples: 12032 | elapsed time per iteration (ms): 5443.7 | throughput per GPU (TFLOP/s/GPU): 239.9 | learning rate: 1.880000E-05 | global batch size: 128 | lm loss: 1.865515E+00 | load_balancing_loss: 1.013109E+00 | loss scale: 1.0 | grad norm: 1.603 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:26] iteration 95/ 100 | consumed samples: 12160 | elapsed time per iteration (ms): 5540.0 | throughput per GPU (TFLOP/s/GPU): 235.7 | learning rate: 1.900000E-05 | global batch size: 128 | lm loss: 1.845348E+00 | load_balancing_loss: 1.012796E+00 | loss scale: 1.0 | grad norm: 1.599 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:32] iteration 96/ 100 | consumed samples: 12288 | elapsed time per iteration (ms): 5702.2 | throughput per GPU (TFLOP/s/GPU): 229.0 | learning rate: 1.920000E-05 | global batch size: 128 | lm loss: 1.843516E+00 | load_balancing_loss: 1.010116E+00 | loss scale: 1.0 | grad norm: 1.851 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:38] iteration 97/ 100 | consumed samples: 12416 | elapsed time per iteration (ms): 5733.2 | throughput per GPU (TFLOP/s/GPU): 227.8 | learning rate: 1.940000E-05 | global batch size: 128 | lm loss: 1.876754E+00 | load_balancing_loss: 1.011542E+00 | loss scale: 1.0 | grad norm: 1.748 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:43] iteration 98/ 100 | consumed samples: 12544 | elapsed time per iteration (ms): 5556.4 | throughput per GPU (TFLOP/s/GPU): 235.0 | learning rate: 1.960000E-05 | global batch size: 128 | lm loss: 1.810738E+00 | load_balancing_loss: 1.010371E+00 | loss scale: 1.0 | grad norm: 1.472 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:49] iteration 99/ 100 | consumed samples: 12672 | elapsed time per iteration (ms): 5523.5 | throughput per GPU (TFLOP/s/GPU): 236.4 | learning rate: 1.980000E-05 | global batch size: 128 | lm loss: 1.872008E+00 | load_balancing_loss: 1.008882E+00 | loss scale: 1.0 | grad norm: 1.681 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:54] iteration 100/ 100 | consumed samples: 12800 | elapsed time per iteration (ms): 5540.0 | throughput per GPU (TFLOP/s/GPU): 235.7 | learning rate: 2.000000E-05 | global batch size: 128 | lm loss: 1.824753E+00 | load_balancing_loss: 1.009905E+00 | loss scale: 1.0 | grad norm: 1.625 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [after training is done] datetime: 2024-04-06 03:04:54 ```

Thank you for reporting this issue. 130 TFLOPS is indeed too low for the H100. I quickly reviewed your script and have some suggestions:

Update the code to the latest main branch and upgrade grouped_gemm to v1.0.

Use alltoall dispathcer: --moe-token-dispatcher-type alltoall.

Use EP8TP2.

Train for a while (at least 400 steps) before checking performance, or load a pretrained checkpoint. This is because router weights in early stage are not sufficiently trained, leading to imbalanced token distribution.

If expert_parallel_size==num_moe_experts, the num_local_experts is 1 and GroupedMLP is same as SequentialMLP, is it right? And as I know, the communication overhead of pp is less than tp and ep if the proportion of bubble time is not too high, is MoE support pp and make it more efficient?

Hi, thanks for the suggestions. I retested the throuput according to your suggestion. To be more specific:

Update Megatron-LM the latest commit (ba77325)

Update grouped_gemm to v1.0.0 (fanshiqing/grouped_gemm@7a7f018)

Set --moe-token-dispatcher-type alltoall

Switch to EP=8 & TP=2

Use the pre-trained weights from Mixtral AI (converted from hf checkpoint)

The throughput has indeed increased significantly, reaching around 230 TFLOP/s. However, for H100, it's still pretty low, isn't it? May I ask, theoretically, what would be a more reasonable throughput?

Here is the logs

Apologies for the delayed reply. 230 TFLOPS falls below our expectations; Currently, we can exceed 330TFLOPS on the H100 and potentially higher by switching to EP8TP1 with re-computation.

Does that mean you can achieve over 330 TFLOPS in the same or similar software environment and settings? Should I then suspect hardware-related issues, such as network speeds between nodes?

Hi @ShinoharaHare , our env is:

DGX H100, 64 GPUs.
pytorch 24.03 image..

I double-checked your scripts and suggest the following modifications:

Seq Len: 2048 -> 4096
enable dp overlap: --overlap-grad-reduce --overlap-param-gather

Let's see how performance changes after these changes ^ ^.

Thank you for reporting this issue. 130 TFLOPS is indeed too low for the H100. I quickly reviewed your script and have some suggestions:

Update the code to the latest main branch and upgrade grouped_gemm to v1.0.

Use alltoall dispathcer: --moe-token-dispatcher-type alltoall.

Use EP8TP2.

Train for a while (at least 400 steps) before checking performance, or load a pretrained checkpoint. This is because router weights in early stage are not sufficiently trained, leading to imbalanced token distribution.

If expert_parallel_size==num_moe_experts, the num_local_experts is 1 and GroupedMLP is same as SequentialMLP, is it right? And as I know, the communication overhead of pp is less than tp and ep if the proportion of bubble time is not too high, is MoE support pp and make it more efficient?

Hi XLZed, MCore MoE does support PP, but for the Mixtral 8x7B model, we prefer EP and TP.

Thank you for reporting this issue. 130 TFLOPS is indeed too low for the H100. I quickly reviewed your script and have some suggestions:

Update the code to the latest main branch and upgrade grouped_gemm to v1.0.

Use alltoall dispathcer: --moe-token-dispatcher-type alltoall.

Use EP8TP2.

Train for a while (at least 400 steps) before checking performance, or load a pretrained checkpoint. This is because router weights in early stage are not sufficiently trained, leading to imbalanced token distribution.

If expert_parallel_size==num_moe_experts, the num_local_experts is 1 and GroupedMLP is same as SequentialMLP, is it right? And as I know, the communication overhead of pp is less than tp and ep if the proportion of bubble time is not too high, is MoE support pp and make it more efficient?

Hi XLZed, MCore MoE does support PP, but for the Mixtral 8x7B model, we prefer EP and TP.

Does grouped_gemm support variable token lengths to local experts on the same rank?

Does grouped_gemm support variable token lengths to local experts on the same rank?

Yes, we support variable lengths for inputs from each local expert.

Hi @ShinoharaHare , our env is:

DGX H100, 64 GPUs.

pytorch 24.03 image..

I double-checked your scripts and suggest the following modifications:

Seq Len: 2048 -> 4096

enable dp overlap: --overlap-grad-reduce --overlap-param-gather

Let's see how performance changes after these changes ^ ^.

@yanring Enabling --overlap-grad-reduce and --overlap-param-gather will result in a CUDA error: uncorrectable ECC error encountered, which seems essentially caused by OOM. I've tried setting the sequence length to 4096 before, but doing so results in a CUDA OOM directly. I've also tried adding --recompute-activations in both scenarios, but still get OOM.

Hi, thanks for the suggestions. I retested the throuput according to your suggestion. To be more specific:

Update Megatron-LM the latest commit (https://github.com/NVIDIA/Megatron-LM/commit/ba773259dbe5735fbd91ca41e7f4ded60b335c52)

Update grouped_gemm to v1.0.0 (https://github.com/fanshiqing/grouped_gemm/commit/7a7f0189797889e926a30b3487512f9539161060)

Set --moe-token-dispatcher-type alltoall

Switch to EP=8 & TP=2

Use the pre-trained weights from Mixtral AI (converted from hf checkpoint)

The throughput has indeed increased significantly, reaching around 230 TFLOP/s. However, for H100, it's still pretty low, isn't it? May I ask, theoretically, what would be a more reasonable throughput?

Here is the logs
``` using world size: 16, data-parallel size: 8, context-parallel size: 1 tensor-model-parallel size: 2, pipeline-model-parallel size: 1 WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:Llama2Tokenizer WARNING: Setting args.overlap_p2p_comm to False since non-interleaved schedule does not support overlapping p2p communication accumulate and all-reduce gradients in fp32 for bfloat16 data type. using torch.bfloat16 for parameters ... ------------------------ arguments ------------------------ accumulate_allreduce_grads_in_fp32 .............. True adam_beta1 ...................................... 0.9 adam_beta2 ...................................... 0.999 adam_eps ........................................ 1e-08 add_bias_linear ................................. False add_position_embedding .......................... True add_qkv_bias .................................... False adlr_autoresume ................................. False adlr_autoresume_interval ........................ 1000 apply_layernorm_1p .............................. False apply_query_key_layer_scaling ................... False apply_residual_connection_post_layernorm ........ False apply_rope_fusion ............................... True async_tensor_model_parallel_allreduce ........... False attention_dropout ............................... 0.0 attention_softmax_in_fp32 ....................... False auto_detect_ckpt_format ......................... False barrier_with_L1_time ............................ True bert_binary_head ................................ True bert_embedder_type .............................. megatron bert_load ....................................... None bf16 ............................................ True bias_dropout_fusion ............................. True bias_gelu_fusion ................................ False bias_swiglu_fusion .............................. True biencoder_projection_dim ........................ 0 biencoder_shared_query_context_model ............ False block_data_path ................................. None check_for_nan_in_loss_and_grad .................. True ckpt_fully_parallel_save ........................ False ckpt_step ....................................... None classes_fraction ................................ 1.0 clip_grad ....................................... 1.0 clone_scatter_output_in_embedding ............... True consumed_train_samples .......................... 0 consumed_valid_samples .......................... 0 context_parallel_size ........................... 1 create_attention_mask_in_dataloader ............. True data_cache_path ................................. None data_parallel_random_init ....................... False data_parallel_size .............................. 8 data_path ....................................... ['custom/data/wudao/wudao_mistralbpe_content_document'] data_per_class_fraction ......................... 1.0 data_sharding ................................... True dataloader_type ................................. single decoder_num_layers .............................. None decoder_seq_length .............................. None decoupled_lr .................................... None decoupled_min_lr ................................ None delay_grad_reduce ............................... True delay_param_gather .............................. False dino_bottleneck_size ............................ 256 dino_freeze_last_layer .......................... 1 dino_head_hidden_size ........................... 2048 dino_local_crops_number ......................... 10 dino_local_img_size ............................. 96 dino_norm_last_layer ............................ False dino_teacher_temp ............................... 0.07 dino_warmup_teacher_temp ........................ 0.04 dino_warmup_teacher_temp_epochs ................. 30 dist_ckpt_format ................................ torch_dist distribute_saved_activations .................... False distributed_backend ............................. nccl distributed_timeout_minutes ..................... 10 embedding_path .................................. None empty_unused_memory_level ....................... 0 enable_one_logger ............................... False encoder_num_layers .............................. 32 encoder_seq_length .............................. 2048 end_weight_decay ................................ 0.1 eod_mask_loss ................................... False eval_interval ................................... 1000 eval_iters ...................................... 1 evidence_data_path .............................. None exit_duration_in_mins ........................... None exit_interval ................................... None exit_on_missing_checkpoint ...................... False exit_signal_handler ............................. False expert_model_parallel_size ...................... 8 ffn_hidden_size ................................. 14336 finetune ........................................ False fp16 ............................................ False fp16_lm_cross_entropy ........................... False fp32_residual_connection ........................ False fp8 ............................................. None fp8_amax_compute_algo ........................... most_recent fp8_amax_history_len ............................ 1 fp8_interval .................................... 1 fp8_margin ...................................... 0 fp8_wgrad ....................................... True global_batch_size ............................... 128 gradient_accumulation_fusion .................... True group_query_attention ........................... True head_lr_mult .................................... 1.0 hidden_dropout .................................. 0.0 hidden_size ..................................... 4096 hysteresis ...................................... 2 ict_head_size ................................... None ict_load ........................................ None img_h ........................................... 224 img_w ........................................... 224 indexer_batch_size .............................. 128 indexer_log_interval ............................ 1000 inference_batch_times_seqlen_threshold .......... 512 init_method_std ................................. 0.02 init_method_xavier_uniform ...................... False initial_loss_scale .............................. 4294967296 iter_per_epoch .................................. 1250 kv_channels ..................................... 128 lazy_mpu_init ................................... None load ............................................ custom/ckpt/mixtral-8x7b-tp2-ep8-mgg local_rank ...................................... None log_batch_size_to_tensorboard ................... False log_interval .................................... 1 log_learning_rate_to_tensorboard ................ True log_loss_scale_to_tensorboard ................... True log_memory_to_tensorboard ....................... False log_num_zeros_in_grad ........................... False log_params_norm ................................. False log_progress .................................... True log_throughput .................................. True log_timers_to_tensorboard ....................... False log_validation_ppl_to_tensorboard ............... False log_world_size_to_tensorboard ................... False loss_scale ...................................... None loss_scale_window ............................... 1000 lr .............................................. 0.0001 lr_decay_iters .................................. 320000 lr_decay_samples ................................ None lr_decay_style .................................. cosine lr_warmup_fraction .............................. None lr_warmup_init .................................. 0.0 lr_warmup_iters ................................. 500 lr_warmup_samples ............................... 0 make_vocab_size_divisible_by .................... 128 manual_gc ....................................... False manual_gc_eval .................................. True manual_gc_interval .............................. 0 mask_factor ..................................... 1.0 mask_prob ....................................... 0.15 mask_type ....................................... random masked_softmax_fusion ........................... False max_position_embeddings ......................... 32768 max_tokens_to_oom ............................... 12000 merge_file ...................................... None micro_batch_size ................................ 1 min_loss_scale .................................. 1.0 min_lr .......................................... 1e-05 mmap_bin_files .................................. True mock_data ....................................... False moe_aux_loss_coeff .............................. 0.01 moe_grouped_gemm ................................ True moe_input_jitter_eps ............................ None moe_per_layer_logging ........................... False moe_router_load_balancing_type .................. aux_loss moe_router_topk ................................. 2 moe_token_dispatcher_type ....................... alltoall moe_token_dropping .............................. False moe_z_loss_coeff ................................ None nccl_communicator_config_path ................... None no_load_optim ................................... True no_load_rng ..................................... True no_persist_layer_norm ........................... False no_save_optim ................................... None no_save_rng ..................................... None norm_epsilon .................................... 1e-05 normalization ................................... RMSNorm num_attention_heads ............................. 32 num_channels .................................... 3 num_classes ..................................... 1000 num_experts ..................................... 8 num_layers ...................................... 32 num_layers_per_virtual_pipeline_stage ........... None num_query_groups ................................ 8 num_workers ..................................... 2 one_logger_entity ............................... hwinf_dcm one_logger_project .............................. e2e-tracking one_logger_run_name ............................. None onnx_safe ....................................... None openai_gelu ..................................... False optimizer ....................................... adam output_bert_embeddings .......................... False overlap_grad_reduce ............................. False overlap_p2p_comm ................................ False overlap_param_gather ............................ False override_opt_param_scheduler .................... False params_dtype .................................... torch.bfloat16 patch_dim ....................................... 16 perform_initialization .......................... True pipeline_model_parallel_size .................... 1 pipeline_model_parallel_split_rank .............. None position_embedding_type ......................... rope pretrained_checkpoint ........................... None profile ......................................... True profile_ranks ................................... [0] profile_step_end ................................ 12 profile_step_start .............................. 10 qk_layernorm .................................... False query_in_block_prob ............................. 0.1 rampup_batch_size ............................... None rank ............................................ 0 recompute_granularity ........................... None recompute_method ................................ None recompute_num_layers ............................ None reset_attention_mask ............................ False reset_position_ids .............................. False retriever_report_topk_accuracies ................ [] retriever_score_scaling ......................... False retriever_seq_length ............................ 256 retro_add_retriever ............................. False retro_attention_gate ............................ 1 retro_cyclic_train_iters ........................ None retro_encoder_attention_dropout ................. 0.1 retro_encoder_hidden_dropout .................... 0.1 retro_encoder_layers ............................ 2 retro_num_neighbors ............................. 2 retro_num_retrieved_chunks ...................... 2 retro_project_dir ............................... None retro_verify_neighbor_count ..................... True rotary_interleaved .............................. False rotary_percent .................................. 1.0 rotary_seq_len_interpolation_factor ............. None sample_rate ..................................... 1.0 save ............................................ custom/ckpt/mixtral-8x7b-tp2-ep8-mgg save_interval ................................... 1000 scatter_gather_tensors_in_pipeline .............. True seed ............................................ 1234 seq_length ...................................... 2048 sequence_parallel ............................... True sgd_momentum .................................... 0.9 short_seq_prob .................................. 0.1 skip_train ...................................... False spec ............................................ None split ........................................... 99990,8,2 squared_relu .................................... False standalone_embedding_stage ...................... False start_weight_decay .............................. 0.1 swiglu .......................................... True swin_backbone_type .............................. tiny tensor_model_parallel_size ...................... 2 tensorboard_dir ................................. custom/ckpt/mixtral-8x7b-tp2-ep8-mgg/tensorboard tensorboard_log_interval ........................ 1 tensorboard_queue_size .......................... 1000 test_data_path .................................. None test_mode ....................................... False timing_log_level ................................ 0 timing_log_option ............................... minmax titles_data_path ................................ None tokenizer_model ................................. custom/ckpt/mixtral-8x7b/tokenizer.model tokenizer_type .................................. Llama2Tokenizer tp_comm_bulk_dgrad .............................. True tp_comm_bulk_wgrad .............................. True tp_comm_overlap ................................. False tp_comm_overlap_ag .............................. True tp_comm_overlap_cfg ............................. None tp_comm_overlap_rs .............................. True tp_comm_split_ag ................................ True tp_comm_split_rs ................................ True train_data_path ................................. None train_iters ..................................... 100 train_samples ................................... None transformer_impl ................................ transformer_engine transformer_pipeline_model_parallel_size ........ 1 untie_embeddings_and_output_weights ............. True use_checkpoint_args ............................. False use_checkpoint_opt_param_scheduler .............. False use_cpu_initialization .......................... None use_dist_ckpt ................................... False use_distributed_optimizer ....................... True use_flash_attn .................................. True use_mcore_models ................................ True use_one_sent_docs ............................... False use_ring_exchange_p2p ........................... False use_rotary_position_embeddings .................. False valid_data_path ................................. None variable_seq_lengths ............................ False virtual_pipeline_model_parallel_size ............ None vision_backbone_type ............................ vit vision_pretraining .............................. False vision_pretraining_type ......................... classify vocab_extra_ids ................................. 0 vocab_file ...................................... None vocab_size ...................................... None wandb_exp_name .................................. wandb_project ................................... wandb_save_dir .................................. weight_decay .................................... 0.1 weight_decay_incr_style ......................... constant world_size ...................................... 16 yaml_cfg ........................................ None -------------------- end of arguments --------------------- setting number of micro-batches to constant 16 > building Llama2Tokenizer tokenizer ... > padded vocab (size: 32000) with 0 dummy tokens (new size: 32000) > initializing torch distributed ... make: Entering directory '.../Megatron-LM/megatron/core/datasets' make: Nothing to be done for 'default'. make: Leaving directory '.../Megatron-LM/megatron/core/datasets' > initialized tensor model parallel with size 2 > initialized pipeline model parallel with size 1 > setting random seeds to 1234 ... > compiling dataset index builder ... >>> done with dataset index builder. Compilation time: 0.104 seconds WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations. > compiling and loading fused kernels ... >>> done with compiling and loading fused kernels. Compilation time: 7.866 seconds [rank1]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank8]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank2]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank9]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank10]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank0]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank3]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank11]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank4]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank12]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank5]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank13]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank6]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank7]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank14]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) [rank15]:[W init.cpp:767] Warning: nvfuser is no longer supported in torch script, use _jit_set_nvfuser_enabled is deprecated and a no-op (function operator()) time to initialize megatron (seconds): 14.235 [after megatron is initialized] datetime: 2024-04-06 02:54:57 building GPT model ... > number of parameters on (tensor, pipeline) model parallel rank (1, 0): 3622047744 > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 3622047744 INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 1 INFO:megatron.core.distributed.param_and_grad_buffer:Params for bucket 1 (803475456 elements): INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.embedding.word_embeddings.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.final_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.output_layer.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.self_attention.linear_proj.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.pre_mlp_layernorm.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.mlp.router.weight INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.self_attention.linear_qkv.weight INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 1 INFO:megatron.core.distributed.param_and_grad_buffer:Params for bucket 1 (2818572288 elements): INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.12.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.26.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.20.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.1.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.30.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.18.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.29.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.23.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.5.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.0.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.15.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.3.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.27.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.21.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.14.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.13.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.8.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.31.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.25.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.19.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.24.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.22.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.17.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.11.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.7.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.4.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.28.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.16.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.6.mlp.experts.weight1 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.2.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.10.mlp.experts.weight2 INFO:megatron.core.distributed.param_and_grad_buffer: module.decoder.layers.9.mlp.experts.weight1 INFO:megatron.core.optimizer:Setting up optimizer with OptimizerConfig(optimizer='adam', lr=0.0001, min_lr=1e-05, decoupled_lr=None, decoupled_min_lr=None, weight_decay=0.1, fp16=False, bf16=True, params_dtype=torch.bfloat16, loss_scale=None, initial_loss_scale=4294967296, min_loss_scale=1.0, loss_scale_window=1000, hysteresis=2, adam_beta1=0.9, adam_beta2=0.999, adam_eps=1e-08, sgd_momentum=0.9, use_distributed_optimizer=True, overlap_grad_reduce=False, overlap_param_gather=False, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=) > learning rate decay style: cosine loading release checkpoint from custom/ckpt/mixtral-8x7b-tp2-ep8-mgg could not find arguments in the checkpoint ... checkpoint version 0 succesfully fixed query-key-values ordering for checkpoint version 0 successfully loaded checkpoint from custom/ckpt/mixtral-8x7b-tp2-ep8-mgg [ t 0, p 0 ] at iteration 0 > setting tensorboard ... (min, max) time across ranks (ms): load-checkpoint ................................: (8126.15, 8126.65) [after model, optimizer, and learning rate scheduler are built] datetime: 2024-04-06 02:55:06 > building train, validation, and test datasets ... > datasets target sizes (minimum size): train: 12800 validation: 128 test: 128 INFO:megatron.core.datasets.blended_megatron_dataset_config:mock = False INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 0.9999), (0.9999, 0.99998), (0.99998, 1.0)] > building train, validation, and test datasets for GPT ... INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from custom/data/wudao/wudao_mistralbpe_content_document.idx INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers INFO:megatron.core.datasets.indexed_dataset: Extract the document indices INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 59132211 INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 59132211 INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices INFO:megatron.core.datasets.gpt_dataset: Load the document index from cc3235b81bd7fd0fa07cabe05d15043d-GPTDataset-document_index.npy INFO:megatron.core.datasets.gpt_dataset: Load the sample index from cc3235b81bd7fd0fa07cabe05d15043d-GPTDataset-sample_index.npy INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from cc3235b81bd7fd0fa07cabe05d15043d-GPTDataset-shuffle_index.npy INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 40201537 INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset valid indices INFO:megatron.core.datasets.gpt_dataset: Load the document index from a625518736b8143e22f4f34c6682183e-GPTDataset-document_index.npy INFO:megatron.core.datasets.gpt_dataset: Load the sample index from a625518736b8143e22f4f34c6682183e-GPTDataset-sample_index.npy INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from a625518736b8143e22f4f34c6682183e-GPTDataset-shuffle_index.npy INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 6204 INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset test indices INFO:megatron.core.datasets.gpt_dataset: Load the document index from 052434ed70ae721ed70b2219cf2deb88-GPTDataset-document_index.npy INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 052434ed70ae721ed70b2219cf2deb88-GPTDataset-sample_index.npy INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 052434ed70ae721ed70b2219cf2deb88-GPTDataset-shuffle_index.npy INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 2332 > finished creating GPT datasets ... [after dataloaders are built] datetime: 2024-04-06 02:55:07 done with setup ... (min, max) time across ranks (ms): model-and-optimizer-setup ......................: (8592.94, 8605.02) train/valid/test-data-iterators-setup ..........: (569.02, 865.21) training ... [before the start of training step] datetime: 2024-04-06 02:55:07 Number of parameters in transformer layers in billions: 46.44 Number of parameters in embedding layers in billions: 0.26 Total number of parameters in billions: 46.70 Number of parameters in most loaded shard in billions: 23.3510 Theoretical memory footprints: weight and optimizer=167019.40 MB [Rank 0] (after 1 iterations) memory (MB) | allocated: 54250.97802734375 | max allocated: 54250.98583984375 | reserved: 61470.0 | max reserved: 61470.0 [2024-04-06 02:55:39] iteration 1/ 100 | consumed samples: 128 | elapsed time per iteration (ms): 32269.4 | throughput per GPU (TFLOP/s/GPU): 40.5 | learning rate: 2.000000E-07 | global batch size: 128 | lm loss: 1.985617E+00 | load_balancing_loss: 1.089786E+00 | loss scale: 1.0 | grad norm: 6.396 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [Rank 1] (after 1 iterations) memory (MB) | allocated: 54250.97802734375 | max allocated: 54250.98583984375 | reserved: 61480.0 | max reserved: 61480.0 [2024-04-06 02:55:45] iteration 2/ 100 | consumed samples: 256 | elapsed time per iteration (ms): 5630.1 | throughput per GPU (TFLOP/s/GPU): 231.9 | learning rate: 4.000000E-07 | global batch size: 128 | lm loss: 2.021530E+00 | load_balancing_loss: 1.087362E+00 | loss scale: 1.0 | grad norm: 6.895 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:55:50] iteration 3/ 100 | consumed samples: 384 | elapsed time per iteration (ms): 5410.6 | throughput per GPU (TFLOP/s/GPU): 241.4 | learning rate: 6.000000E-07 | global batch size: 128 | lm loss: 2.003316E+00 | load_balancing_loss: 1.085377E+00 | loss scale: 1.0 | grad norm: 6.603 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:55:55] iteration 4/ 100 | consumed samples: 512 | elapsed time per iteration (ms): 5364.1 | throughput per GPU (TFLOP/s/GPU): 243.5 | learning rate: 8.000000E-07 | global batch size: 128 | lm loss: 2.009657E+00 | load_balancing_loss: 1.091695E+00 | loss scale: 1.0 | grad norm: 6.619 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:01] iteration 5/ 100 | consumed samples: 640 | elapsed time per iteration (ms): 5496.7 | throughput per GPU (TFLOP/s/GPU): 237.6 | learning rate: 1.000000E-06 | global batch size: 128 | lm loss: 2.002326E+00 | load_balancing_loss: 1.091539E+00 | loss scale: 1.0 | grad norm: 6.612 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:06] iteration 6/ 100 | consumed samples: 768 | elapsed time per iteration (ms): 5364.8 | throughput per GPU (TFLOP/s/GPU): 243.4 | learning rate: 1.200000E-06 | global batch size: 128 | lm loss: 1.933151E+00 | load_balancing_loss: 1.086472E+00 | loss scale: 1.0 | grad norm: 5.765 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:12] iteration 7/ 100 | consumed samples: 896 | elapsed time per iteration (ms): 5682.7 | throughput per GPU (TFLOP/s/GPU): 229.8 | learning rate: 1.400000E-06 | global batch size: 128 | lm loss: 2.016085E+00 | load_balancing_loss: 1.085193E+00 | loss scale: 1.0 | grad norm: 5.821 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:17] iteration 8/ 100 | consumed samples: 1024 | elapsed time per iteration (ms): 5408.6 | throughput per GPU (TFLOP/s/GPU): 241.4 | learning rate: 1.600000E-06 | global batch size: 128 | lm loss: 1.965713E+00 | load_balancing_loss: 1.080933E+00 | loss scale: 1.0 | grad norm: 4.774 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:23] iteration 9/ 100 | consumed samples: 1152 | elapsed time per iteration (ms): 5590.1 | throughput per GPU (TFLOP/s/GPU): 233.6 | learning rate: 1.800000E-06 | global batch size: 128 | lm loss: 1.919308E+00 | load_balancing_loss: 1.089582E+00 | loss scale: 1.0 | grad norm: 4.267 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:28] iteration 10/ 100 | consumed samples: 1280 | elapsed time per iteration (ms): 5443.7 | throughput per GPU (TFLOP/s/GPU): 239.9 | learning rate: 2.000000E-06 | global batch size: 128 | lm loss: 1.978377E+00 | load_balancing_loss: 1.089948E+00 | loss scale: 1.0 | grad norm: 4.069 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:34] iteration 11/ 100 | consumed samples: 1408 | elapsed time per iteration (ms): 5984.1 | throughput per GPU (TFLOP/s/GPU): 218.2 | learning rate: 2.200000E-06 | global batch size: 128 | lm loss: 1.889895E+00 | load_balancing_loss: 1.083618E+00 | loss scale: 1.0 | grad norm: 3.361 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:40] iteration 12/ 100 | consumed samples: 1536 | elapsed time per iteration (ms): 5821.8 | throughput per GPU (TFLOP/s/GPU): 224.3 | learning rate: 2.400000E-06 | global batch size: 128 | lm loss: 1.932808E+00 | load_balancing_loss: 1.085315E+00 | loss scale: 1.0 | grad norm: 3.336 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:46] iteration 13/ 100 | consumed samples: 1664 | elapsed time per iteration (ms): 5962.2 | throughput per GPU (TFLOP/s/GPU): 219.0 | learning rate: 2.600000E-06 | global batch size: 128 | lm loss: 1.911683E+00 | load_balancing_loss: 1.079515E+00 | loss scale: 1.0 | grad norm: 3.183 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:52] iteration 14/ 100 | consumed samples: 1792 | elapsed time per iteration (ms): 5927.4 | throughput per GPU (TFLOP/s/GPU): 220.3 | learning rate: 2.800000E-06 | global batch size: 128 | lm loss: 1.913695E+00 | load_balancing_loss: 1.076165E+00 | loss scale: 1.0 | grad norm: 2.994 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:56:58] iteration 15/ 100 | consumed samples: 1920 | elapsed time per iteration (ms): 5926.4 | throughput per GPU (TFLOP/s/GPU): 220.4 | learning rate: 3.000000E-06 | global batch size: 128 | lm loss: 1.957101E+00 | load_balancing_loss: 1.069903E+00 | loss scale: 1.0 | grad norm: 2.853 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:04] iteration 16/ 100 | consumed samples: 2048 | elapsed time per iteration (ms): 5912.7 | throughput per GPU (TFLOP/s/GPU): 220.9 | learning rate: 3.200000E-06 | global batch size: 128 | lm loss: 1.915763E+00 | load_balancing_loss: 1.065748E+00 | loss scale: 1.0 | grad norm: 2.778 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:10] iteration 17/ 100 | consumed samples: 2176 | elapsed time per iteration (ms): 5706.3 | throughput per GPU (TFLOP/s/GPU): 228.9 | learning rate: 3.400000E-06 | global batch size: 128 | lm loss: 1.918353E+00 | load_balancing_loss: 1.064678E+00 | loss scale: 1.0 | grad norm: 2.911 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:15] iteration 18/ 100 | consumed samples: 2304 | elapsed time per iteration (ms): 5732.8 | throughput per GPU (TFLOP/s/GPU): 227.8 | learning rate: 3.600000E-06 | global batch size: 128 | lm loss: 1.861051E+00 | load_balancing_loss: 1.058054E+00 | loss scale: 1.0 | grad norm: 2.449 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:21] iteration 19/ 100 | consumed samples: 2432 | elapsed time per iteration (ms): 5684.9 | throughput per GPU (TFLOP/s/GPU): 229.7 | learning rate: 3.800000E-06 | global batch size: 128 | lm loss: 1.934895E+00 | load_balancing_loss: 1.049081E+00 | loss scale: 1.0 | grad norm: 2.447 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:27] iteration 20/ 100 | consumed samples: 2560 | elapsed time per iteration (ms): 5770.6 | throughput per GPU (TFLOP/s/GPU): 226.3 | learning rate: 4.000000E-06 | global batch size: 128 | lm loss: 1.932632E+00 | load_balancing_loss: 1.052491E+00 | loss scale: 1.0 | grad norm: 2.456 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:32] iteration 21/ 100 | consumed samples: 2688 | elapsed time per iteration (ms): 5541.8 | throughput per GPU (TFLOP/s/GPU): 235.6 | learning rate: 4.200000E-06 | global batch size: 128 | lm loss: 1.904877E+00 | load_balancing_loss: 1.047207E+00 | loss scale: 1.0 | grad norm: 2.213 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:38] iteration 22/ 100 | consumed samples: 2816 | elapsed time per iteration (ms): 5576.7 | throughput per GPU (TFLOP/s/GPU): 234.2 | learning rate: 4.400000E-06 | global batch size: 128 | lm loss: 1.872380E+00 | load_balancing_loss: 1.039512E+00 | loss scale: 1.0 | grad norm: 2.116 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:44] iteration 23/ 100 | consumed samples: 2944 | elapsed time per iteration (ms): 5807.4 | throughput per GPU (TFLOP/s/GPU): 224.9 | learning rate: 4.600000E-06 | global batch size: 128 | lm loss: 1.835408E+00 | load_balancing_loss: 1.042104E+00 | loss scale: 1.0 | grad norm: 2.034 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:50] iteration 24/ 100 | consumed samples: 3072 | elapsed time per iteration (ms): 5727.3 | throughput per GPU (TFLOP/s/GPU): 228.0 | learning rate: 4.800000E-06 | global batch size: 128 | lm loss: 1.898657E+00 | load_balancing_loss: 1.029742E+00 | loss scale: 1.0 | grad norm: 1.982 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:57:55] iteration 25/ 100 | consumed samples: 3200 | elapsed time per iteration (ms): 5498.4 | throughput per GPU (TFLOP/s/GPU): 237.5 | learning rate: 5.000000E-06 | global batch size: 128 | lm loss: 1.904866E+00 | load_balancing_loss: 1.034888E+00 | loss scale: 1.0 | grad norm: 1.872 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:01] iteration 26/ 100 | consumed samples: 3328 | elapsed time per iteration (ms): 5531.7 | throughput per GPU (TFLOP/s/GPU): 236.1 | learning rate: 5.200000E-06 | global batch size: 128 | lm loss: 1.889752E+00 | load_balancing_loss: 1.028931E+00 | loss scale: 1.0 | grad norm: 1.793 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:06] iteration 27/ 100 | consumed samples: 3456 | elapsed time per iteration (ms): 5678.3 | throughput per GPU (TFLOP/s/GPU): 230.0 | learning rate: 5.400000E-06 | global batch size: 128 | lm loss: 1.866109E+00 | load_balancing_loss: 1.031736E+00 | loss scale: 1.0 | grad norm: 1.773 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:12] iteration 28/ 100 | consumed samples: 3584 | elapsed time per iteration (ms): 5650.6 | throughput per GPU (TFLOP/s/GPU): 231.1 | learning rate: 5.600000E-06 | global batch size: 128 | lm loss: 1.914117E+00 | load_balancing_loss: 1.027364E+00 | loss scale: 1.0 | grad norm: 1.709 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:18] iteration 29/ 100 | consumed samples: 3712 | elapsed time per iteration (ms): 5912.1 | throughput per GPU (TFLOP/s/GPU): 220.9 | learning rate: 5.800000E-06 | global batch size: 128 | lm loss: 1.867856E+00 | load_balancing_loss: 1.023825E+00 | loss scale: 1.0 | grad norm: 1.769 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:23] iteration 30/ 100 | consumed samples: 3840 | elapsed time per iteration (ms): 5571.1 | throughput per GPU (TFLOP/s/GPU): 234.4 | learning rate: 6.000000E-06 | global batch size: 128 | lm loss: 1.924535E+00 | load_balancing_loss: 1.025294E+00 | loss scale: 1.0 | grad norm: 1.572 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:29] iteration 31/ 100 | consumed samples: 3968 | elapsed time per iteration (ms): 5718.9 | throughput per GPU (TFLOP/s/GPU): 228.3 | learning rate: 6.200000E-06 | global batch size: 128 | lm loss: 1.830754E+00 | load_balancing_loss: 1.028048E+00 | loss scale: 1.0 | grad norm: 1.555 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:35] iteration 32/ 100 | consumed samples: 4096 | elapsed time per iteration (ms): 5629.3 | throughput per GPU (TFLOP/s/GPU): 232.0 | learning rate: 6.400000E-06 | global batch size: 128 | lm loss: 1.848776E+00 | load_balancing_loss: 1.021549E+00 | loss scale: 1.0 | grad norm: 1.592 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:40] iteration 33/ 100 | consumed samples: 4224 | elapsed time per iteration (ms): 5600.4 | throughput per GPU (TFLOP/s/GPU): 233.2 | learning rate: 6.600000E-06 | global batch size: 128 | lm loss: 1.917658E+00 | load_balancing_loss: 1.032319E+00 | loss scale: 1.0 | grad norm: 1.519 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:46] iteration 34/ 100 | consumed samples: 4352 | elapsed time per iteration (ms): 5643.8 | throughput per GPU (TFLOP/s/GPU): 231.4 | learning rate: 6.800000E-06 | global batch size: 128 | lm loss: 1.844636E+00 | load_balancing_loss: 1.019185E+00 | loss scale: 1.0 | grad norm: 1.626 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:51] iteration 35/ 100 | consumed samples: 4480 | elapsed time per iteration (ms): 5367.8 | throughput per GPU (TFLOP/s/GPU): 243.3 | learning rate: 7.000000E-06 | global batch size: 128 | lm loss: 1.853418E+00 | load_balancing_loss: 1.020990E+00 | loss scale: 1.0 | grad norm: 1.760 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:58:57] iteration 36/ 100 | consumed samples: 4608 | elapsed time per iteration (ms): 5399.9 | throughput per GPU (TFLOP/s/GPU): 241.8 | learning rate: 7.200000E-06 | global batch size: 128 | lm loss: 1.842918E+00 | load_balancing_loss: 1.023077E+00 | loss scale: 1.0 | grad norm: 1.409 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:02] iteration 37/ 100 | consumed samples: 4736 | elapsed time per iteration (ms): 5515.8 | throughput per GPU (TFLOP/s/GPU): 236.8 | learning rate: 7.400000E-06 | global batch size: 128 | lm loss: 1.862270E+00 | load_balancing_loss: 1.023782E+00 | loss scale: 1.0 | grad norm: 1.718 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:08] iteration 38/ 100 | consumed samples: 4864 | elapsed time per iteration (ms): 5477.8 | throughput per GPU (TFLOP/s/GPU): 238.4 | learning rate: 7.600000E-06 | global batch size: 128 | lm loss: 1.862543E+00 | load_balancing_loss: 1.019304E+00 | loss scale: 1.0 | grad norm: 1.722 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:13] iteration 39/ 100 | consumed samples: 4992 | elapsed time per iteration (ms): 5649.1 | throughput per GPU (TFLOP/s/GPU): 231.2 | learning rate: 7.800000E-06 | global batch size: 128 | lm loss: 1.863421E+00 | load_balancing_loss: 1.017805E+00 | loss scale: 1.0 | grad norm: 1.469 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:19] iteration 40/ 100 | consumed samples: 5120 | elapsed time per iteration (ms): 5810.4 | throughput per GPU (TFLOP/s/GPU): 224.8 | learning rate: 8.000000E-06 | global batch size: 128 | lm loss: 1.879655E+00 | load_balancing_loss: 1.017568E+00 | loss scale: 1.0 | grad norm: 1.633 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:25] iteration 41/ 100 | consumed samples: 5248 | elapsed time per iteration (ms): 5462.9 | throughput per GPU (TFLOP/s/GPU): 239.1 | learning rate: 8.200000E-06 | global batch size: 128 | lm loss: 1.812076E+00 | load_balancing_loss: 1.020508E+00 | loss scale: 1.0 | grad norm: 1.419 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:30] iteration 42/ 100 | consumed samples: 5376 | elapsed time per iteration (ms): 5452.3 | throughput per GPU (TFLOP/s/GPU): 239.5 | learning rate: 8.400000E-06 | global batch size: 128 | lm loss: 1.824542E+00 | load_balancing_loss: 1.017472E+00 | loss scale: 1.0 | grad norm: 1.400 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:36] iteration 43/ 100 | consumed samples: 5504 | elapsed time per iteration (ms): 5444.9 | throughput per GPU (TFLOP/s/GPU): 239.8 | learning rate: 8.600000E-06 | global batch size: 128 | lm loss: 1.825991E+00 | load_balancing_loss: 1.019746E+00 | loss scale: 1.0 | grad norm: 1.426 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:41] iteration 44/ 100 | consumed samples: 5632 | elapsed time per iteration (ms): 5533.8 | throughput per GPU (TFLOP/s/GPU): 236.0 | learning rate: 8.800000E-06 | global batch size: 128 | lm loss: 1.875063E+00 | load_balancing_loss: 1.020033E+00 | loss scale: 1.0 | grad norm: 1.327 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:47] iteration 45/ 100 | consumed samples: 5760 | elapsed time per iteration (ms): 5718.6 | throughput per GPU (TFLOP/s/GPU): 228.4 | learning rate: 9.000000E-06 | global batch size: 128 | lm loss: 1.834162E+00 | load_balancing_loss: 1.018004E+00 | loss scale: 1.0 | grad norm: 1.611 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:52] iteration 46/ 100 | consumed samples: 5888 | elapsed time per iteration (ms): 5567.2 | throughput per GPU (TFLOP/s/GPU): 234.6 | learning rate: 9.200000E-06 | global batch size: 128 | lm loss: 1.883577E+00 | load_balancing_loss: 1.016062E+00 | loss scale: 1.0 | grad norm: 1.439 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 02:59:58] iteration 47/ 100 | consumed samples: 6016 | elapsed time per iteration (ms): 5692.2 | throughput per GPU (TFLOP/s/GPU): 229.4 | learning rate: 9.400000E-06 | global batch size: 128 | lm loss: 1.836727E+00 | load_balancing_loss: 1.019520E+00 | loss scale: 1.0 | grad norm: 1.372 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:04] iteration 48/ 100 | consumed samples: 6144 | elapsed time per iteration (ms): 5872.4 | throughput per GPU (TFLOP/s/GPU): 222.4 | learning rate: 9.600000E-06 | global batch size: 128 | lm loss: 1.855191E+00 | load_balancing_loss: 1.017754E+00 | loss scale: 1.0 | grad norm: 1.508 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:09] iteration 49/ 100 | consumed samples: 6272 | elapsed time per iteration (ms): 5528.7 | throughput per GPU (TFLOP/s/GPU): 236.2 | learning rate: 9.800000E-06 | global batch size: 128 | lm loss: 1.806294E+00 | load_balancing_loss: 1.017504E+00 | loss scale: 1.0 | grad norm: 1.529 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:15] iteration 50/ 100 | consumed samples: 6400 | elapsed time per iteration (ms): 5531.5 | throughput per GPU (TFLOP/s/GPU): 236.1 | learning rate: 1.000000E-05 | global batch size: 128 | lm loss: 1.887587E+00 | load_balancing_loss: 1.016094E+00 | loss scale: 1.0 | grad norm: 1.439 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:21] iteration 51/ 100 | consumed samples: 6528 | elapsed time per iteration (ms): 5501.3 | throughput per GPU (TFLOP/s/GPU): 237.4 | learning rate: 1.020000E-05 | global batch size: 128 | lm loss: 1.834414E+00 | load_balancing_loss: 1.015084E+00 | loss scale: 1.0 | grad norm: 1.599 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:26] iteration 52/ 100 | consumed samples: 6656 | elapsed time per iteration (ms): 5520.9 | throughput per GPU (TFLOP/s/GPU): 236.5 | learning rate: 1.040000E-05 | global batch size: 128 | lm loss: 1.847078E+00 | load_balancing_loss: 1.015950E+00 | loss scale: 1.0 | grad norm: 1.486 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:32] iteration 53/ 100 | consumed samples: 6784 | elapsed time per iteration (ms): 5711.6 | throughput per GPU (TFLOP/s/GPU): 228.6 | learning rate: 1.060000E-05 | global batch size: 128 | lm loss: 1.862840E+00 | load_balancing_loss: 1.016317E+00 | loss scale: 1.0 | grad norm: 1.522 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:37] iteration 54/ 100 | consumed samples: 6912 | elapsed time per iteration (ms): 5689.4 | throughput per GPU (TFLOP/s/GPU): 229.5 | learning rate: 1.080000E-05 | global batch size: 128 | lm loss: 1.897956E+00 | load_balancing_loss: 1.017408E+00 | loss scale: 1.0 | grad norm: 1.383 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:43] iteration 55/ 100 | consumed samples: 7040 | elapsed time per iteration (ms): 5763.8 | throughput per GPU (TFLOP/s/GPU): 226.6 | learning rate: 1.100000E-05 | global batch size: 128 | lm loss: 1.863309E+00 | load_balancing_loss: 1.014457E+00 | loss scale: 1.0 | grad norm: 1.534 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:49] iteration 56/ 100 | consumed samples: 7168 | elapsed time per iteration (ms): 5742.1 | throughput per GPU (TFLOP/s/GPU): 227.4 | learning rate: 1.120000E-05 | global batch size: 128 | lm loss: 1.899538E+00 | load_balancing_loss: 1.018558E+00 | loss scale: 1.0 | grad norm: 1.470 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:00:54] iteration 57/ 100 | consumed samples: 7296 | elapsed time per iteration (ms): 5450.5 | throughput per GPU (TFLOP/s/GPU): 239.6 | learning rate: 1.140000E-05 | global batch size: 128 | lm loss: 1.864605E+00 | load_balancing_loss: 1.015150E+00 | loss scale: 1.0 | grad norm: 1.244 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:00] iteration 58/ 100 | consumed samples: 7424 | elapsed time per iteration (ms): 5538.9 | throughput per GPU (TFLOP/s/GPU): 235.8 | learning rate: 1.160000E-05 | global batch size: 128 | lm loss: 1.812579E+00 | load_balancing_loss: 1.020851E+00 | loss scale: 1.0 | grad norm: 1.610 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:05] iteration 59/ 100 | consumed samples: 7552 | elapsed time per iteration (ms): 5410.9 | throughput per GPU (TFLOP/s/GPU): 241.3 | learning rate: 1.180000E-05 | global batch size: 128 | lm loss: 1.848337E+00 | load_balancing_loss: 1.013638E+00 | loss scale: 1.0 | grad norm: 1.351 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:11] iteration 60/ 100 | consumed samples: 7680 | elapsed time per iteration (ms): 5603.1 | throughput per GPU (TFLOP/s/GPU): 233.1 | learning rate: 1.200000E-05 | global batch size: 128 | lm loss: 1.801180E+00 | load_balancing_loss: 1.019084E+00 | loss scale: 1.0 | grad norm: 1.549 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:16] iteration 61/ 100 | consumed samples: 7808 | elapsed time per iteration (ms): 5495.5 | throughput per GPU (TFLOP/s/GPU): 237.6 | learning rate: 1.220000E-05 | global batch size: 128 | lm loss: 1.813972E+00 | load_balancing_loss: 1.014779E+00 | loss scale: 1.0 | grad norm: 1.427 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:22] iteration 62/ 100 | consumed samples: 7936 | elapsed time per iteration (ms): 5753.1 | throughput per GPU (TFLOP/s/GPU): 227.0 | learning rate: 1.240000E-05 | global batch size: 128 | lm loss: 1.808689E+00 | load_balancing_loss: 1.022012E+00 | loss scale: 1.0 | grad norm: 1.398 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:28] iteration 63/ 100 | consumed samples: 8064 | elapsed time per iteration (ms): 5650.1 | throughput per GPU (TFLOP/s/GPU): 231.1 | learning rate: 1.260000E-05 | global batch size: 128 | lm loss: 1.781526E+00 | load_balancing_loss: 1.013716E+00 | loss scale: 1.0 | grad norm: 1.494 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:33] iteration 64/ 100 | consumed samples: 8192 | elapsed time per iteration (ms): 5539.7 | throughput per GPU (TFLOP/s/GPU): 235.7 | learning rate: 1.280000E-05 | global batch size: 128 | lm loss: 1.871476E+00 | load_balancing_loss: 1.019044E+00 | loss scale: 1.0 | grad norm: 1.369 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:39] iteration 65/ 100 | consumed samples: 8320 | elapsed time per iteration (ms): 5493.9 | throughput per GPU (TFLOP/s/GPU): 237.7 | learning rate: 1.300000E-05 | global batch size: 128 | lm loss: 1.846450E+00 | load_balancing_loss: 1.017387E+00 | loss scale: 1.0 | grad norm: 1.308 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:44] iteration 66/ 100 | consumed samples: 8448 | elapsed time per iteration (ms): 5590.8 | throughput per GPU (TFLOP/s/GPU): 233.6 | learning rate: 1.320000E-05 | global batch size: 128 | lm loss: 1.873755E+00 | load_balancing_loss: 1.014257E+00 | loss scale: 1.0 | grad norm: 1.411 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:50] iteration 67/ 100 | consumed samples: 8576 | elapsed time per iteration (ms): 5710.3 | throughput per GPU (TFLOP/s/GPU): 228.7 | learning rate: 1.340000E-05 | global batch size: 128 | lm loss: 1.765591E+00 | load_balancing_loss: 1.016482E+00 | loss scale: 1.0 | grad norm: 1.414 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:01:56] iteration 68/ 100 | consumed samples: 8704 | elapsed time per iteration (ms): 5734.5 | throughput per GPU (TFLOP/s/GPU): 227.7 | learning rate: 1.360000E-05 | global batch size: 128 | lm loss: 1.839895E+00 | load_balancing_loss: 1.012786E+00 | loss scale: 1.0 | grad norm: 1.371 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:01] iteration 69/ 100 | consumed samples: 8832 | elapsed time per iteration (ms): 5478.6 | throughput per GPU (TFLOP/s/GPU): 238.4 | learning rate: 1.380000E-05 | global batch size: 128 | lm loss: 1.912256E+00 | load_balancing_loss: 1.013041E+00 | loss scale: 1.0 | grad norm: 1.485 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:07] iteration 70/ 100 | consumed samples: 8960 | elapsed time per iteration (ms): 5514.8 | throughput per GPU (TFLOP/s/GPU): 236.8 | learning rate: 1.400000E-05 | global batch size: 128 | lm loss: 1.873068E+00 | load_balancing_loss: 1.012509E+00 | loss scale: 1.0 | grad norm: 1.467 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:12] iteration 71/ 100 | consumed samples: 9088 | elapsed time per iteration (ms): 5361.6 | throughput per GPU (TFLOP/s/GPU): 243.6 | learning rate: 1.420000E-05 | global batch size: 128 | lm loss: 1.818812E+00 | load_balancing_loss: 1.013377E+00 | loss scale: 1.0 | grad norm: 1.300 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:18] iteration 72/ 100 | consumed samples: 9216 | elapsed time per iteration (ms): 5470.7 | throughput per GPU (TFLOP/s/GPU): 238.7 | learning rate: 1.440000E-05 | global batch size: 128 | lm loss: 1.820313E+00 | load_balancing_loss: 1.019612E+00 | loss scale: 1.0 | grad norm: 1.305 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:24] iteration 73/ 100 | consumed samples: 9344 | elapsed time per iteration (ms): 5829.9 | throughput per GPU (TFLOP/s/GPU): 224.0 | learning rate: 1.460000E-05 | global batch size: 128 | lm loss: 1.798953E+00 | load_balancing_loss: 1.010977E+00 | loss scale: 1.0 | grad norm: 1.539 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:29] iteration 74/ 100 | consumed samples: 9472 | elapsed time per iteration (ms): 5702.4 | throughput per GPU (TFLOP/s/GPU): 229.0 | learning rate: 1.480000E-05 | global batch size: 128 | lm loss: 1.774078E+00 | load_balancing_loss: 1.012441E+00 | loss scale: 1.0 | grad norm: 1.471 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:35] iteration 75/ 100 | consumed samples: 9600 | elapsed time per iteration (ms): 5599.5 | throughput per GPU (TFLOP/s/GPU): 233.2 | learning rate: 1.500000E-05 | global batch size: 128 | lm loss: 1.838492E+00 | load_balancing_loss: 1.015038E+00 | loss scale: 1.0 | grad norm: 1.445 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:40] iteration 76/ 100 | consumed samples: 9728 | elapsed time per iteration (ms): 5588.2 | throughput per GPU (TFLOP/s/GPU): 233.7 | learning rate: 1.520000E-05 | global batch size: 128 | lm loss: 1.860703E+00 | load_balancing_loss: 1.012689E+00 | loss scale: 1.0 | grad norm: 1.500 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:46] iteration 77/ 100 | consumed samples: 9856 | elapsed time per iteration (ms): 5425.4 | throughput per GPU (TFLOP/s/GPU): 240.7 | learning rate: 1.540000E-05 | global batch size: 128 | lm loss: 1.827507E+00 | load_balancing_loss: 1.012502E+00 | loss scale: 1.0 | grad norm: 1.491 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:52] iteration 78/ 100 | consumed samples: 9984 | elapsed time per iteration (ms): 5652.9 | throughput per GPU (TFLOP/s/GPU): 231.0 | learning rate: 1.560000E-05 | global batch size: 128 | lm loss: 1.784492E+00 | load_balancing_loss: 1.013809E+00 | loss scale: 1.0 | grad norm: 1.407 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:02:57] iteration 79/ 100 | consumed samples: 10112 | elapsed time per iteration (ms): 5577.0 | throughput per GPU (TFLOP/s/GPU): 234.2 | learning rate: 1.580000E-05 | global batch size: 128 | lm loss: 1.858489E+00 | load_balancing_loss: 1.011662E+00 | loss scale: 1.0 | grad norm: 1.621 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:03] iteration 80/ 100 | consumed samples: 10240 | elapsed time per iteration (ms): 5712.8 | throughput per GPU (TFLOP/s/GPU): 228.6 | learning rate: 1.600000E-05 | global batch size: 128 | lm loss: 1.842588E+00 | load_balancing_loss: 1.011640E+00 | loss scale: 1.0 | grad norm: 1.631 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:09] iteration 81/ 100 | consumed samples: 10368 | elapsed time per iteration (ms): 5684.5 | throughput per GPU (TFLOP/s/GPU): 229.7 | learning rate: 1.620000E-05 | global batch size: 128 | lm loss: 1.818980E+00 | load_balancing_loss: 1.012697E+00 | loss scale: 1.0 | grad norm: 1.564 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:14] iteration 82/ 100 | consumed samples: 10496 | elapsed time per iteration (ms): 5592.0 | throughput per GPU (TFLOP/s/GPU): 233.5 | learning rate: 1.640000E-05 | global batch size: 128 | lm loss: 1.805010E+00 | load_balancing_loss: 1.012805E+00 | loss scale: 1.0 | grad norm: 1.545 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:20] iteration 83/ 100 | consumed samples: 10624 | elapsed time per iteration (ms): 5641.6 | throughput per GPU (TFLOP/s/GPU): 231.5 | learning rate: 1.660000E-05 | global batch size: 128 | lm loss: 1.812314E+00 | load_balancing_loss: 1.011967E+00 | loss scale: 1.0 | grad norm: 1.530 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:25] iteration 84/ 100 | consumed samples: 10752 | elapsed time per iteration (ms): 5563.7 | throughput per GPU (TFLOP/s/GPU): 234.7 | learning rate: 1.680000E-05 | global batch size: 128 | lm loss: 1.822110E+00 | load_balancing_loss: 1.009684E+00 | loss scale: 1.0 | grad norm: 1.799 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:31] iteration 85/ 100 | consumed samples: 10880 | elapsed time per iteration (ms): 5580.9 | throughput per GPU (TFLOP/s/GPU): 234.0 | learning rate: 1.700000E-05 | global batch size: 128 | lm loss: 1.831795E+00 | load_balancing_loss: 1.009344E+00 | loss scale: 1.0 | grad norm: 1.578 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:37] iteration 86/ 100 | consumed samples: 11008 | elapsed time per iteration (ms): 5695.8 | throughput per GPU (TFLOP/s/GPU): 229.3 | learning rate: 1.720000E-05 | global batch size: 128 | lm loss: 1.831625E+00 | load_balancing_loss: 1.011533E+00 | loss scale: 1.0 | grad norm: 1.515 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:42] iteration 87/ 100 | consumed samples: 11136 | elapsed time per iteration (ms): 5444.5 | throughput per GPU (TFLOP/s/GPU): 239.9 | learning rate: 1.740000E-05 | global batch size: 128 | lm loss: 1.814374E+00 | load_balancing_loss: 1.010052E+00 | loss scale: 1.0 | grad norm: 1.365 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:48] iteration 88/ 100 | consumed samples: 11264 | elapsed time per iteration (ms): 5462.7 | throughput per GPU (TFLOP/s/GPU): 239.1 | learning rate: 1.760000E-05 | global batch size: 128 | lm loss: 1.825778E+00 | load_balancing_loss: 1.010838E+00 | loss scale: 1.0 | grad norm: 1.506 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:53] iteration 89/ 100 | consumed samples: 11392 | elapsed time per iteration (ms): 5633.2 | throughput per GPU (TFLOP/s/GPU): 231.8 | learning rate: 1.780000E-05 | global batch size: 128 | lm loss: 1.818898E+00 | load_balancing_loss: 1.011014E+00 | loss scale: 1.0 | grad norm: 1.358 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:03:59] iteration 90/ 100 | consumed samples: 11520 | elapsed time per iteration (ms): 5567.8 | throughput per GPU (TFLOP/s/GPU): 234.5 | learning rate: 1.800000E-05 | global batch size: 128 | lm loss: 1.813602E+00 | load_balancing_loss: 1.022434E+00 | loss scale: 1.0 | grad norm: 1.590 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:04] iteration 91/ 100 | consumed samples: 11648 | elapsed time per iteration (ms): 5691.9 | throughput per GPU (TFLOP/s/GPU): 229.4 | learning rate: 1.820000E-05 | global batch size: 128 | lm loss: 1.797111E+00 | load_balancing_loss: 1.011964E+00 | loss scale: 1.0 | grad norm: 1.436 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:10] iteration 92/ 100 | consumed samples: 11776 | elapsed time per iteration (ms): 5451.5 | throughput per GPU (TFLOP/s/GPU): 239.6 | learning rate: 1.840000E-05 | global batch size: 128 | lm loss: 1.809117E+00 | load_balancing_loss: 1.012038E+00 | loss scale: 1.0 | grad norm: 1.577 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:15] iteration 93/ 100 | consumed samples: 11904 | elapsed time per iteration (ms): 5599.2 | throughput per GPU (TFLOP/s/GPU): 233.2 | learning rate: 1.860000E-05 | global batch size: 128 | lm loss: 1.797812E+00 | load_balancing_loss: 1.011838E+00 | loss scale: 1.0 | grad norm: 1.553 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:21] iteration 94/ 100 | consumed samples: 12032 | elapsed time per iteration (ms): 5443.7 | throughput per GPU (TFLOP/s/GPU): 239.9 | learning rate: 1.880000E-05 | global batch size: 128 | lm loss: 1.865515E+00 | load_balancing_loss: 1.013109E+00 | loss scale: 1.0 | grad norm: 1.603 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:26] iteration 95/ 100 | consumed samples: 12160 | elapsed time per iteration (ms): 5540.0 | throughput per GPU (TFLOP/s/GPU): 235.7 | learning rate: 1.900000E-05 | global batch size: 128 | lm loss: 1.845348E+00 | load_balancing_loss: 1.012796E+00 | loss scale: 1.0 | grad norm: 1.599 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:32] iteration 96/ 100 | consumed samples: 12288 | elapsed time per iteration (ms): 5702.2 | throughput per GPU (TFLOP/s/GPU): 229.0 | learning rate: 1.920000E-05 | global batch size: 128 | lm loss: 1.843516E+00 | load_balancing_loss: 1.010116E+00 | loss scale: 1.0 | grad norm: 1.851 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:38] iteration 97/ 100 | consumed samples: 12416 | elapsed time per iteration (ms): 5733.2 | throughput per GPU (TFLOP/s/GPU): 227.8 | learning rate: 1.940000E-05 | global batch size: 128 | lm loss: 1.876754E+00 | load_balancing_loss: 1.011542E+00 | loss scale: 1.0 | grad norm: 1.748 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:43] iteration 98/ 100 | consumed samples: 12544 | elapsed time per iteration (ms): 5556.4 | throughput per GPU (TFLOP/s/GPU): 235.0 | learning rate: 1.960000E-05 | global batch size: 128 | lm loss: 1.810738E+00 | load_balancing_loss: 1.010371E+00 | loss scale: 1.0 | grad norm: 1.472 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:49] iteration 99/ 100 | consumed samples: 12672 | elapsed time per iteration (ms): 5523.5 | throughput per GPU (TFLOP/s/GPU): 236.4 | learning rate: 1.980000E-05 | global batch size: 128 | lm loss: 1.872008E+00 | load_balancing_loss: 1.008882E+00 | loss scale: 1.0 | grad norm: 1.681 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [2024-04-06 03:04:54] iteration 100/ 100 | consumed samples: 12800 | elapsed time per iteration (ms): 5540.0 | throughput per GPU (TFLOP/s/GPU): 235.7 | learning rate: 2.000000E-05 | global batch size: 128 | lm loss: 1.824753E+00 | load_balancing_loss: 1.009905E+00 | loss scale: 1.0 | grad norm: 1.625 | num zeros: 0.0 | number of skipped iterations: 0 | number of nan iterations: 0 | [after training is done] datetime: 2024-04-06 03:04:54 ```

which modification brings the most speed improvement? btw I encountered some error when converting mixtral from transformers to Megatron when grouped-gemm is set, can you share some converting scripts?

@ShinoharaHare Could you please share your checkpoint conversion script?

which modification brings the most speed improvement? btw I encountered some error when converting mixtral from transformers to Megatron when grouped-gemm is set, can you share some converting scripts?

The most significant performance change is achieved by resuming from a trained checkpoint. If you do not have pretrained weights, you can train from scratch for about 500 steps. We noticed that after several hundred steps, the token distribution will become quite balanced.

@yanring , @ShinoharaHare , can you please share a conversion script for Mixtral from HF weights ?

@yanring , @ShinoharaHare , can you please share a conversion script for Mixtral from HF weights ?

Hi Vlad, we are working on the converter; it is already in the review process.

@yanring @ShinoharaHare

Hi, I'm in a similar situation to this issue. But we also have some differences. For example, we use 8 h800, 64 experts, ep=8, tp=1, pp=1. I also encountered some training efficiency issues, but they were not a top priority.

What bothers me now is that after I used ep8 and grouped-gemm, my model structure changed. when I try to merge the model with ep=8 into the model with ep=1, it can be loaded by the inference program normally, indicating that the merged shape is correct.

But the inference result is incorrect. I want to know if Megatron-LM will develop a model convert tool that can facilitate me to merge the ep=8 model into the ep=1 model.

Or could you provide some information on how to merge a grouped-gemm enabled model?

Hello @hwdef , thank you for the update. Currently, the format for the weights in GroupedGEMM for each expert is [input_size, output_size], which is different from the format used in SequentialMLP's ParallelLinear, [output_size, input_size]. Did you transpose the weight during your conversion? @cb521 can help to take a look if this issue continues.

By the way, we are also working on supporting distributed checkpointing with Grouped GEMM.

By the way, we are also working on supporting distributed checkpointing with Grouped GEMM.

Yes, we have considered the order of output_size and input_size

@yanring Hi, Could you please help me check my convert tool?

@yanring , @ShinoharaHare , can you please share a conversion script for Mixtral from HF weights ?

Hi Vlad, we are working on the converter; it is already in the review process.

I’m excited about this. When do you plan to merge it into the main branch?

@yanring Hi, Could you please help me check my convert tool?

@hwdef 你好，我也遇到同样的问题，请问现在有解决方法了吗？

@yanring Hi, Could you please help me check my convert tool?

@hwdef 你好，我也遇到同样的问题，请问现在有解决方法了吗？

没有

NVIDIA / Megatron-LM

[QUESTION] Training Mixtral 8x7B on 16 x H100 only achieves low throughput of 130 TFLOPS #756