diff --git a/.gitattributes b/.gitattributes index 3dbdeda7cc429b94c960c8d239bf3b5df7b5e267..060e32cb7bc99a40d2ffd8a862fb9a6edfa276b7 100644 --- a/.gitattributes +++ b/.gitattributes @@ -51,3 +51,4 @@ echo/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__echo__matched__seq8192__gbs8__cap echo/tierbtrace/tierbtrace_gbs8/QWEN2_57B__echo__matched__seq4096__gbs8__cap1p08984375/nsys/QWEN2_57B__echo__matched__seq4096__gbs8__cap1p08984375.nsys-rep filter=lfs diff=lfs merge=lfs -text ultraep/tierbtrace/tierbtrace_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/nsys/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375.nsys-rep filter=lfs diff=lfs merge=lfs -text ultraep/tierbtrace/tierbtrace_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/nsys/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125.nsys-rep filter=lfs diff=lfs merge=lfs -text +ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep filter=lfs diff=lfs merge=lfs -text diff --git a/ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep b/ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep new file mode 100644 index 0000000000000000000000000000000000000000..1a1997bfa7d4f712ddebdda1cd1944c7e8a98db0 --- /dev/null +++ b/ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c2bde598254f4608754087aa5a247757f07b2d1d39cee3af337a603d33aee67 +size 13026056 diff --git a/ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/run.env b/ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/run.env new file mode 100644 index 0000000000000000000000000000000000000000..a55da0e6535746f731724401d046194546844f44 --- /dev/null +++ b/ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/run.env @@ -0,0 +1,61 @@ +backend=ultraep +config=DSV2_LITE_L24 +family=tierB +repo=/workspace/Megatron-UltraEP +branch=ultraep-bracing-regions-through-kernels +commit=03d57ef7 +num_layers=24 +hidden=2048 +ffn_hidden=10944 +moe_ffn_hidden=1408 +num_experts=64 +topk=6 +experts_per_rank=8 +expert_params=8650752 +heads=16 +kv_groups=0 +kv_channels=128 +qk_layernorm=1 +shared_ffn=2816 +shared_expert_overlap=0 +mla=1 +dense_first_layers=1 +moe_layers=23 +rotary_base=10000 +rotary_scaling=40 +qkv_bias=0 +router_balancing=seq_aux_loss +aux_loss_coeff=1e-3 +router_pre_softmax=1 +router_fusion=0 +router_dtype=fp32 +device_initiated_grouped_gemm=1 +norm_epsilon=1e-6 +seq_length=8192 +micro_batch_size=1 +global_batch_size=8 +num_microbatches=1 +assignments_per_rank=49152 +mean_tokens_per_expert=6144 +slot_policy=matched +slots_per_rank=8 +capacity_mode=static +capacity=1.0755208333333333 +cuda_graph=1 +probe_mode=0 +async_grad_reduce=0 +branch_role=timing +memory_dump_mode=none +memory_snapshot_path= +cuda_graph_warmup_steps=3 +recompute=none +force_balance=0 +train_iters=10 +ep_size=8 +echo_dispatch_overlap=1 +echo_dispatch_recompute=0 +ultraep_quota_min_tokens=768 +ultraep_weight_sync_join= +profile_enable=1 +kernel_trace=1 +fp8_recipe=mxfp8 diff --git a/ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/train.log b/ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/train.log new file mode 100644 index 0000000000000000000000000000000000000000..33541fda41e077db9fdf481f2b155222f3ab29ba --- /dev/null +++ b/ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/train.log @@ -0,0 +1,1628 @@ +============================================================== + DSV2_LITE_L24 / ultraep (slots/rank=8 policy=matched) + E=64 K=6 H'=1408 m=8 + L=24 seq=8192 gbs=8 (1 microbatches) + cap_mode=static cap=1.0755208333333333 cg=1 recompute=none + memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/memory_snapshots/snapshot.pickle + repo=/workspace/Megatron-UltraEP @ 03d57ef7 +============================================================== +W0826 20:10:30.209000 748463 torch/distributed/run.py:766] +W0826 20:10:30.209000 748463 torch/distributed/run.py:766] ***************************************** +W0826 20:10:30.209000 748463 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0826 20:10:30.209000 748463 torch/distributed/run.py:766] ***************************************** +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +WARNING: one_logger package is required to enable e2e metrics tracking. please go to https://confluence.nvidia.com/display/MLWFO/Package+Repositories for details to install it +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +using world size: 8, data-parallel size: 8, context-parallel size: 1, hierarchical context-parallel sizes: None, tensor-model-parallel size: 1, pipeline-model-parallel size: 1 +WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer +Number of virtual stages per pipeline stage: None +accumulate and all-reduce gradients in fp32 for bfloat16 data type. +using torch.bfloat16 for parameters ... +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +------------------------ arguments ------------------------ + account_for_embedding_in_pipeline_split ......... False + account_for_loss_in_pipeline_split .............. False + 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 + align_grad_reduce ............................... True + align_param_gather .............................. False + app_tag_run_name ................................ None + app_tag_run_version ............................. 0.0.0 + apply_layernorm_1p .............................. False + apply_query_key_layer_scaling ................... False + apply_residual_connection_post_layernorm ........ False + apply_rope_fusion ............................... False + async_save ...................................... None + async_tensor_model_parallel_allreduce ........... True + attention_backend ............................... AttnBackend.auto + 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 + calc_ft_timeouts ................................ False + calculate_per_token_loss ........................ False + check_for_large_grads ........................... False + check_for_nan_in_loss_and_grad .................. False + check_for_spiky_loss ............................ False + check_weight_hash_across_dp_replicas_interval ... None + ckpt_assume_constant_structure .................. False + ckpt_convert_format ............................. None + ckpt_convert_save ............................... None + ckpt_convert_update_legacy_dist_opt_format ...... False + ckpt_format ..................................... torch_dist + ckpt_fully_parallel_load ........................ False + ckpt_fully_parallel_save ........................ True + ckpt_fully_parallel_save_deprecated ............. False + ckpt_step ....................................... None + classes_fraction ................................ 1.0 + clip_grad ....................................... 1.0 + clone_scatter_output_in_embedding ............... True + config_logger_dir ............................... + consumed_train_samples .......................... 0 + consumed_valid_samples .......................... 0 + context_parallel_size ........................... 1 + cp_comm_type .................................... ['p2p'] + create_attention_mask_in_dataloader ............. True + cross_entropy_fusion_impl ....................... native + cross_entropy_loss_fusion ....................... False + cuda_graph_scope ................................ full_iteration + cuda_graph_warmup_steps ......................... 3 + data_args_path .................................. None + data_cache_path ................................. None + data_parallel_random_init ....................... False + data_parallel_sharding_strategy ................. no_shard + data_parallel_size .............................. 8 + data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document'] + data_per_class_fraction ......................... 1.0 + data_sharding ................................... True + dataloader_type ................................. single + ddp_average_in_collective ....................... False + ddp_bucket_size ................................. None + ddp_num_buckets ................................. None + ddp_pad_buckets_for_high_nccl_busbw ............. False + decoder_first_pipeline_num_layers ............... None + decoder_last_pipeline_num_layers ................ None + decoder_num_layers .............................. None + decoder_seq_length .............................. None + decoupled_lr .................................... None + decoupled_min_lr ................................ None + decrease_batch_size_if_needed ................... False + defer_embedding_wgrad_compute ................... False + delay_wgrad_compute ............................. False + deprecated_use_mcore_models ..................... True + deterministic_mode .............................. 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 + disable_bf16_reduced_precision_matmul ........... False + disable_mamba_mem_eff_path ...................... False + disable_straggler_on_startup .................... False + dist_ckpt_format_deprecated ..................... None + dist_ckpt_strictness ............................ assume_ok_unexpected + distribute_saved_activations .................... False + distributed_backend ............................. nccl + distributed_timeout_minutes ..................... 10 + embedding_init_method_std ....................... None + embedding_path .................................. None + empty_unused_memory_level ....................... 0 + enable_cuda_graph ............................... True + enable_experimental ............................. False + enable_ft_package ............................... False + enable_gloo_process_groups ...................... True + enable_msc ...................................... True + enable_one_logger ............................... True + encoder_num_layers .............................. 24 + encoder_seq_length .............................. 8192 + end_weight_decay ................................ 0.1 + eod_mask_loss ................................... False + error_injection_rate ............................ 0 + error_injection_type ............................ transient_error + eval_interval ................................... 1000000 + eval_iters ...................................... 0 + evidence_data_path .............................. None + exit_duration_in_mins ........................... None + exit_interval ................................... None + exit_on_missing_checkpoint ...................... False + exit_signal_handler ............................. False + exp_avg_dtype ................................... torch.float32 + exp_avg_sq_dtype ................................ torch.float32 + expert_model_parallel_size ...................... 8 + expert_tensor_parallel_size ..................... 1 + export_force_local_attention .................... False + export_kd_cfg ................................... None + export_kd_teacher_ckpt_format ................... None + export_kd_teacher_load .......................... None + export_kv_cache_quant ........................... False + export_legacy_megatron .......................... False + export_model_type ............................... GPTModel + export_moe_apply_probs_on_input ................. False + export_qk_l2_norm ............................... False + export_quant_cfg ................................ None + export_real_quant_cfg ........................... None + export_te_mcore_model ........................... False + external_cuda_graph ............................. False + ffn_hidden_size ................................. 10944 + finetune ........................................ False + finetune_data_split ............................. train + finetune_hf_dataset ............................. None + first_last_layers_bf16 .......................... False + flash_decode .................................... False + fp16 ............................................ False + fp16_lm_cross_entropy ........................... False + fp32_residual_connection ........................ False + fp8 ............................................. e4m3 + fp8_amax_compute_algo ........................... most_recent + fp8_amax_history_len ............................ 1 + fp8_interval .................................... 1 + fp8_margin ...................................... 0 + fp8_param_gather ................................ False + fp8_recipe ...................................... mxfp8 + fp8_wgrad ....................................... True + fsdp_double_buffer .............................. False + global_batch_size ............................... 8 + grad_reduce_in_bf16 ............................. False + gradient_accumulation_fusion .................... True + gradient_reduce_div_fusion ...................... True + group_query_attention ........................... False + head_lr_mult .................................... 1.0 + heterogeneous_layers_config_encoded_json ........ None + heterogeneous_layers_config_path ................ None + hidden_dropout .................................. 0.0 + hidden_size ..................................... 2048 + hierarchical_context_parallel_sizes ............. None + high_priority_stream_groups ..................... [] + hybrid_attention_ratio .......................... 0.0 + hybrid_mlp_ratio ................................ 0.0 + hybrid_override_pattern ......................... None + 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 .......... -1 + inference_dynamic_batching ...................... False + inference_dynamic_batching_buffer_guaranteed_fraction 0.2 + inference_dynamic_batching_buffer_overflow_factor None + inference_dynamic_batching_buffer_size_gb ....... 40.0 + inference_dynamic_batching_chunk_size ........... 256 + inference_dynamic_batching_max_requests_override None + inference_dynamic_batching_max_tokens_override .. None + inference_dynamic_batching_num_cuda_graphs ...... 16 + inference_max_batch_size ........................ 8 + inference_max_seq_length ........................ 2560 + inference_rng_tracker ........................... False + init_method_std ................................. 0.01 + init_method_xavier_uniform ...................... False + init_model_with_meta_device ..................... False + initial_loss_scale .............................. 4294967296 + inprocess_active_world_size ..................... 8 + inprocess_barrier_timeout ....................... 120 + inprocess_completion_timeout .................... 120 + inprocess_empty_cuda_cache ...................... False + inprocess_granularity ........................... node + inprocess_hard_timeout .......................... 90 + inprocess_heartbeat_interval .................... 30 + inprocess_heartbeat_timeout ..................... 60 + inprocess_last_call_wait ........................ 1 + inprocess_max_iterations ........................ None + inprocess_monitor_process_interval .............. 1.0 + inprocess_monitor_thread_interval ............... 1.0 + inprocess_progress_watchdog_interval ............ 1.0 + inprocess_restart ............................... False + inprocess_soft_timeout .......................... 60 + inprocess_termination_grace_time ................ 1 + is_hybrid_model ................................. False + iter_per_epoch .................................. 1250 + iterations_to_skip .............................. [] + keep_fp8_transpose_cache_when_using_custom_fsdp . False + kitchen_config_file ............................. None + kitchen_recipe_number ........................... None + kv_channels ..................................... 128 + kv_lora_rank .................................... 512 + lazy_mpu_init ................................... None + load ............................................ None + load_main_params_from_ckpt ...................... None + load_model_opt_format ........................... False + local_rank ...................................... 0 + log_energy ...................................... False + log_interval .................................... 1 + log_loss_scale_to_tensorboard ................... True + log_memory_to_tensorboard ....................... False + log_num_zeros_in_grad ........................... False + log_params_norm ................................. False + log_progress .................................... False + log_straggler ................................... False + log_throughput .................................. True + log_timers_to_tensorboard ....................... False + log_validation_ppl_to_tensorboard ............... False + log_world_size_to_tensorboard ................... False + logging_level ................................... None + loss_scale ...................................... None + loss_scale_window ............................... 1000 + lr .............................................. 0.0001 + lr_decay_iters .................................. 8 + lr_decay_samples ................................ None + lr_decay_style .................................. cosine + lr_warmup_fraction .............................. None + lr_warmup_init .................................. 0.0 + lr_warmup_iters ................................. 5 + lr_warmup_samples ............................... 0 + lr_wsd_decay_iters .............................. None + lr_wsd_decay_samples ............................ None + lr_wsd_decay_style .............................. exponential + main_grads_dtype ................................ torch.float32 + main_params_dtype ............................... torch.float32 + make_vocab_size_divisible_by .................... 128 + mamba_head_dim .................................. 64 + mamba_num_groups ................................ 8 + mamba_num_heads ................................. None + mamba_state_dim ................................. 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 + memory_snapshot_path ............................ snapshot.pickle + merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt + micro_batch_size ................................ 1 + microbatch_group_size_per_vp_stage .............. None + mid_level_dataset_surplus ....................... 0.005 + min_loss_scale .................................. 1.0 + min_lr .......................................... 1e-05 + mlp_chunks_for_prefill .......................... 1 + mmap_bin_files .................................. True + mock_data ....................................... False + moe_apply_probs_on_input ........................ False + moe_aux_loss_coeff .............................. 0.001 + moe_capacity_monitor ............................ True + moe_deepep_num_sms .............................. 20 + moe_echo_enable_random_offloading ............... False + moe_echo_expert_dispatch_overlap ................ False + moe_echo_expert_dispatcher_type ................. hybridep + moe_echo_recompute_expert_dispatch .............. False + moe_enable_deepep ............................... False + moe_enable_echo ................................. False + moe_enable_hybridep ............................. True + moe_enable_ultraep .............................. True + moe_expert_capacity_factor ...................... None + moe_extended_tp ................................. False + moe_ffn_hidden_size ............................. 1408 + moe_grouped_gemm ................................ True + moe_hybridep_num_sms ............................ 16 + moe_input_jitter_eps ............................ None + moe_layer_freq .................................. [0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] + moe_layer_recompute ............................. False + moe_num_echo_experts ............................ None + moe_num_redundant_experts_per_rank .............. 8 + moe_pad_expert_input_to_capacity ................ False + moe_per_layer_logging ........................... False + moe_permute_fusion .............................. False + moe_received_token_capacity ..................... 1.0755208333333333 + moe_router_bias_update_rate ..................... 0.001 + moe_router_dtype ................................ fp32 + moe_router_enable_expert_bias ................... False + moe_router_force_load_balancing ................. False + moe_router_group_topk ........................... None + moe_router_load_balancing_type .................. seq_aux_loss + moe_router_num_groups ........................... None + moe_router_padding_for_fp8 ...................... False + moe_router_pre_softmax .......................... True + moe_router_score_function ....................... softmax + moe_router_topk ................................. 6 + moe_router_topk_scaling_factor .................. None + moe_shared_expert_intermediate_size ............. 2816 + moe_shared_expert_overlap ....................... False + moe_token_dispatcher_type ....................... flex + moe_token_drop_policy ........................... probs + moe_ultraep_max_microbatches .................... None + moe_ultraep_overlap_weight_sync ................. True + moe_ultraep_weight_sync_join .................... before_dispatch + moe_upcycling_granularity ....................... 1 + moe_use_device_initiated_grouped_gemm ........... True + moe_use_legacy_grouped_gemm ..................... False + moe_use_upcycling ............................... False + moe_z_loss_coeff ................................ None + mrope_section ................................... None + mscale .......................................... 0.707 + mscale_all_dim .................................. 0.707 + mtp_loss_scaling_factor ......................... 0.1 + mtp_num_layers .................................. None + multi_latent_attention .......................... True + nccl_all_reduce_for_prefill ..................... False + nccl_communicator_config_path ................... None + nccl_ub ......................................... False + no_load_optim ................................... None + no_load_rng ..................................... None + no_persist_layer_norm ........................... False + no_rope_freq .................................... None + no_save_optim ................................... None + no_save_rng ..................................... None + non_persistent_ckpt_type ........................ None + non_persistent_global_ckpt_dir .................. None + non_persistent_local_ckpt_algo .................. fully_parallel + non_persistent_local_ckpt_dir ................... None + non_persistent_save_interval .................... None + norm_epsilon .................................... 1e-06 + normalization ................................... RMSNorm + num_attention_heads ............................. 16 + num_channels .................................... 3 + num_classes ..................................... 1000 + num_dataset_builder_threads ..................... 1 + num_distributed_optimizer_instances ............. 1 + num_experts ..................................... 64 + num_layers ...................................... 24 + num_layers_at_end_in_bf16 ....................... 1 + num_layers_at_start_in_bf16 ..................... 1 + num_layers_per_virtual_pipeline_stage ........... None + num_query_groups ................................ 1 + num_virtual_stages_per_pipeline_rank ............ None + num_workers ..................................... 2 + object_storage_cache_path ....................... None + one_logger_async ................................ False + one_logger_project .............................. megatron-lm + one_logger_run_name ............................. None + onnx_safe ....................................... None + openai_gelu ..................................... False + optimizer ....................................... adam + optimizer_cpu_offload ........................... False + optimizer_offload_fraction ...................... 1.0 + output_bert_embeddings .......................... False + overlap_cpu_optimizer_d2h_h2d ................... False + overlap_grad_reduce ............................. True + overlap_moe_expert_parallel_comm ................ False + overlap_p2p_comm ................................ False + overlap_p2p_comm_warmup_flush ................... False + overlap_param_gather ............................ True + overlap_param_gather_with_optimizer_step ........ False + override_opt_param_scheduler .................... False + params_dtype .................................... torch.bfloat16 + patch_dim ....................................... 16 + per_split_data_args_path ........................ None + perform_initialization .......................... True + pin_cpu_grads ................................... True + pin_cpu_params .................................. True + pipeline_model_parallel_comm_backend ............ None + pipeline_model_parallel_layout .................. None + pipeline_model_parallel_size .................... 1 + position_embedding_type ......................... rope + pretrained_checkpoint ........................... None + profile ......................................... True + profile_ranks ................................... [0, 1, 2, 3, 4, 5, 6, 7] + profile_step_end ................................ 8 + profile_step_start .............................. 6 + q_lora_rank ..................................... None + qk_head_dim ..................................... 128 + qk_l2_norm ...................................... False + qk_layernorm .................................... True + qk_pos_emb_head_dim ............................. 64 + query_in_block_prob ............................. 0.1 + rampup_batch_size ............................... None + rank ............................................ 0 + recompute_granularity ........................... None + recompute_method ................................ None + recompute_modules ............................... None + recompute_num_layers ............................ None + record_memory_history ........................... False + relative_attention_max_distance ................. 128 + relative_attention_num_buckets .................. 32 + replication ..................................... False + replication_factor .............................. 2 + replication_jump ................................ None + rerun_mode ...................................... validate_results + reset_attention_mask ............................ False + reset_position_ids .............................. False + result_rejected_tracker_filename ................ None + 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 + reuse_grad_buf_for_mxfp8_param_ag ............... False + rope_scaling_factor ............................. 8.0 + rope_type ....................................... None + rotary_base ..................................... 10000 + rotary_interleaved .............................. False + rotary_percent .................................. 1.0 + rotary_scaling_factor ........................... 40.0 + rotary_seq_len_interpolation_factor ............. None + run_workload_inspector_server ................... False + sample_rate ..................................... 1.0 + save ............................................ None + save_interval ................................... None + save_retain_interval ............................ None + scatter_gather_tensors_in_pipeline .............. True + seed ............................................ 1234 + seq_length ...................................... 8192 + sequence_parallel ............................... False + sft ............................................. False + sft_tokenizer_prompt_format ..................... nemotron-h-aligned + sgd_momentum .................................... 0.9 + short_seq_prob .................................. 0.1 + skip_train ...................................... False + skipped_train_samples ........................... 0 + spec ............................................ None + split ........................................... 100,0,0 + squared_relu .................................... False + start_weight_decay .............................. 0.1 + straggler_ctrlr_port ............................ 65535 + straggler_minmax_count .......................... 1 + suggested_communication_unit_size ............... None + swiglu .......................................... True + swin_backbone_type .............................. tiny + symmetric_ar_type ............................... None + te_rng_tracker .................................. True + tensor_model_parallel_size ...................... 1 + tensorboard_dir ................................. None + tensorboard_log_interval ........................ 1 + tensorboard_queue_size .......................... 1000 + test_data_path .................................. None + test_mode ....................................... False + tiktoken_num_special_tokens ..................... 1000 + tiktoken_pattern ................................ None + tiktoken_special_tokens ......................... None + timing_log_level ................................ 0 + timing_log_option ............................... minmax + titles_data_path ................................ None + tokenizer_model ................................. None + tokenizer_type .................................. GPT2BPETokenizer + torch_fsdp2_reshard_after_forward ............... True + tp_comm_bootstrap_backend ....................... nccl + 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_overlap_rs_dgrad ........................ False + tp_comm_split_ag ................................ True + tp_comm_split_rs ................................ True + train_data_path ................................. None + train_iters ..................................... 10 + train_samples ................................... None + train_sync_interval ............................. 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_custom_fsdp ................................. False + use_dist_ckpt ................................... True + use_dist_ckpt_deprecated ........................ False + use_distributed_optimizer ....................... True + use_flash_attn .................................. False + use_legacy_models ............................... False + use_mp_args_from_checkpoint_args ................ False + use_one_sent_docs ............................... False + use_persistent_ckpt_worker ...................... False + use_precision_aware_optimizer ................... False + use_pytorch_profiler ............................ False + use_ring_exchange_p2p ........................... False + use_rope_scaling ................................ False + use_rotary_position_embeddings .................. False + use_sharp ....................................... False + use_tokenizer_model_from_checkpoint_args ........ True + use_torch_fsdp2 ................................. False + use_torch_optimizer_for_cpu_offload ............. False + use_tp_pp_dp_mapping ............................ False + v_head_dim ...................................... 128 + 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 ...................................... /workspace/data/wikitext/gpt2-vocab.json + vocab_size ...................................... None + wandb_exp_name .................................. + wandb_project ................................... + wandb_save_dir .................................. + weight_decay .................................... 0.1 + weight_decay_incr_style ......................... constant + wgrad_deferral_limit ............................ 0 + world_size ...................................... 8 + yaml_cfg ........................................ None +-------------------- end of arguments --------------------- +INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1 +> building GPT2BPETokenizer tokenizer ... +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") + > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304) +WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS +> initializing torch distributed ... +[Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 7 is connected to 7 peer ranks. [Gloo] Rank Expected number of connected peer ranks is : 7 +2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 7 is connected to [Gloo] Rank 7 peer ranks. Expected number of connected peer ranks is : 67 is connected to 7 + peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +> initialized tensor model parallel with size 1 +> initialized pipeline model parallel with size 1 +> setting random seeds to 1234 ... +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0[Gloo] Rank +0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[rank1]:[W826 20:10:43.929175513 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank5]:[W826 20:10:43.929207131 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 5] using GPU 5 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank3]:[W826 20:10:43.929262116 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank4]:[W826 20:10:43.929864259 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 4] using GPU 4 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank2]:[W826 20:10:43.930152318 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 2] using GPU 2 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank6]:[W826 20:10:43.930180991 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 6] using GPU 6 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank7]:[W826 20:10:43.930258368 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +> compiling dataset index builder ... +make: Entering directory '/workspace/Megatron-UltraEP/megatron/core/datasets' +make: Nothing to be done for 'default'. +make: Leaving directory '/workspace/Megatron-UltraEP/megatron/core/datasets' +>>> done with dataset index builder. Compilation time: 0.424 seconds +WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations. +> compiling and loading fused kernels ... +[rank0]:[W826 20:10:44.515308624 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 0] using GPU 0 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +>>> done with compiling and loading fused kernels. Compilation time: 3.773 seconds +time to initialize megatron (seconds): 9.503 +[after megatron is initialized] datetime: 2026-08-26 20:10:50 +building GPT model ... +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +[INFO] Use MNNVL fabric: False +[WARN] MAX_NUM_NVL_PEERS is not set. Using detected value 8. +NVSHMEM v3.7.2 +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +UltraEP: NVL domain size <= 8; forcing weight_sync plan mode to direct (was adaptive). + > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 4188125184 +INFO:megatron.core.distributed.distributed_data_parallel:Setting up DistributedDataParallel with config DistributedDataParallelConfig(grad_reduce_in_fp32=True, overlap_grad_reduce=True, overlap_param_gather=True, align_param_gather=False, use_distributed_optimizer=True, num_distributed_optimizer_instances=1, check_for_nan_in_grad=False, check_for_large_grads=False, bucket_size=40000000, pad_buckets_for_high_nccl_busbw=False, average_in_collective=False, fp8_param_gather=False, reuse_grad_buf_for_mxfp8_param_ag=False, use_custom_fsdp=False, data_parallel_sharding_strategy='no_shard', gradient_reduce_div_fusion=True, suggested_communication_unit_size=None, preserve_fp32_weights=True, keep_fp8_transpose_cache_when_using_custom_fsdp=False, nccl_ub=False, fsdp_double_buffer=False) +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 20 +Params for bucket 1 (103022592 elements, 103022592 padded size): + module.output_layer.weight +Params for bucket 2 (48503296 elements, 48503296 padded size): + module.decoder.final_layernorm.weight + module.decoder.layers.23.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.23.self_attention.linear_kv_up_proj.weight + module.decoder.layers.22.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.23.mlp.router.weight + module.decoder.layers.23.pre_mlp_layernorm.weight + module.decoder.layers.23.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.23.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.23.self_attention.linear_kv_down_proj.weight + module.decoder.layers.23.self_attention.linear_q_proj.weight + module.decoder.layers.23.self_attention.linear_proj.weight + module.decoder.layers.23.input_layernorm.weight + module.decoder.layers.22.mlp.shared_experts.linear_fc1.weight +Params for bucket 3 (40901632 elements, 40901632 padded size): + module.decoder.layers.22.mlp.router.weight + module.decoder.layers.22.self_attention.linear_kv_down_proj.weight + module.decoder.layers.22.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.22.pre_mlp_layernorm.weight + module.decoder.layers.22.self_attention.linear_q_proj.weight + module.decoder.layers.22.self_attention.linear_proj.weight + module.decoder.layers.22.input_layernorm.weight + module.decoder.layers.21.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.21.pre_mlp_layernorm.weight + module.decoder.layers.21.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.22.self_attention.linear_kv_up_proj.weight + module.decoder.layers.21.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.21.mlp.router.weight + module.decoder.layers.21.self_attention.linear_kv_up_proj.weight + module.decoder.layers.21.self_attention.linear_kv_down_proj.weight + module.decoder.layers.21.self_attention.linear_q_proj.weight +Params for bucket 4 (41163264 elements, 41163264 padded size): + module.decoder.layers.21.self_attention.linear_proj.weight + module.decoder.layers.20.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.21.input_layernorm.weight + module.decoder.layers.20.pre_mlp_layernorm.weight + module.decoder.layers.20.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.20.self_attention.linear_kv_down_proj.weight + module.decoder.layers.20.self_attention.linear_q_proj.weight + module.decoder.layers.20.self_attention.linear_proj.weight + module.decoder.layers.20.input_layernorm.weight + module.decoder.layers.20.mlp.router.weight + module.decoder.layers.20.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.20.self_attention.linear_kv_up_proj.weight + module.decoder.layers.19.mlp.shared_experts.linear_fc2.weight +Params for bucket 5 (42734080 elements, 42734080 padded size): + module.decoder.layers.19.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.19.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.19.pre_mlp_layernorm.weight + module.decoder.layers.19.self_attention.linear_kv_down_proj.weight + module.decoder.layers.19.self_attention.linear_q_proj.weight + module.decoder.layers.19.self_attention.linear_proj.weight + module.decoder.layers.19.input_layernorm.weight + module.decoder.layers.18.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.19.self_attention.linear_kv_up_proj.weight + module.decoder.layers.19.mlp.router.weight + module.decoder.layers.18.mlp.shared_experts.linear_fc2.weight +Params for bucket 6 (40901632 elements, 40901632 padded size): + module.decoder.layers.18.self_attention.linear_kv_down_proj.weight + module.decoder.layers.18.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.18.pre_mlp_layernorm.weight + module.decoder.layers.18.self_attention.linear_q_proj.weight + module.decoder.layers.18.self_attention.linear_proj.weight + module.decoder.layers.18.input_layernorm.weight + module.decoder.layers.17.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.17.pre_mlp_layernorm.weight + module.decoder.layers.17.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.17.self_attention.linear_kv_down_proj.weight + module.decoder.layers.17.self_attention.linear_q_proj.weight + module.decoder.layers.18.mlp.router.weight + module.decoder.layers.18.self_attention.linear_kv_up_proj.weight + module.decoder.layers.17.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.17.mlp.router.weight + module.decoder.layers.17.self_attention.linear_kv_up_proj.weight +Params for bucket 7 (41163264 elements, 41163264 padded size): + module.decoder.layers.17.self_attention.linear_proj.weight + module.decoder.layers.17.input_layernorm.weight + module.decoder.layers.16.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.16.pre_mlp_layernorm.weight + module.decoder.layers.16.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.16.self_attention.linear_kv_down_proj.weight + module.decoder.layers.16.self_attention.linear_q_proj.weight + module.decoder.layers.16.self_attention.linear_proj.weight + module.decoder.layers.16.input_layernorm.weight + module.decoder.layers.16.mlp.router.weight + module.decoder.layers.16.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.16.self_attention.linear_kv_up_proj.weight + module.decoder.layers.15.mlp.shared_experts.linear_fc2.weight +Params for bucket 8 (42734080 elements, 42734080 padded size): + module.decoder.layers.15.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.15.pre_mlp_layernorm.weight + module.decoder.layers.15.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.15.self_attention.linear_kv_down_proj.weight + module.decoder.layers.15.self_attention.linear_q_proj.weight + module.decoder.layers.15.self_attention.linear_proj.weight + module.decoder.layers.15.input_layernorm.weight + module.decoder.layers.14.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.15.self_attention.linear_kv_up_proj.weight + module.decoder.layers.15.mlp.router.weight + module.decoder.layers.14.mlp.shared_experts.linear_fc2.weight +Params for bucket 9 (40901632 elements, 40901632 padded size): + module.decoder.layers.14.mlp.router.weight + module.decoder.layers.14.self_attention.linear_kv_up_proj.weight + module.decoder.layers.13.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.13.mlp.router.weight + module.decoder.layers.13.self_attention.linear_kv_up_proj.weight + module.decoder.layers.14.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.14.pre_mlp_layernorm.weight + module.decoder.layers.14.self_attention.linear_kv_down_proj.weight + module.decoder.layers.14.self_attention.linear_q_proj.weight + module.decoder.layers.14.self_attention.linear_proj.weight + module.decoder.layers.14.input_layernorm.weight + module.decoder.layers.13.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.13.pre_mlp_layernorm.weight + module.decoder.layers.13.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.13.self_attention.linear_kv_down_proj.weight + module.decoder.layers.13.self_attention.linear_q_proj.weight +Params for bucket 10 (41163264 elements, 41163264 padded size): + module.decoder.layers.12.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.12.mlp.router.weight + module.decoder.layers.12.self_attention.linear_kv_up_proj.weight + module.decoder.layers.11.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.13.self_attention.linear_proj.weight + module.decoder.layers.12.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.13.input_layernorm.weight + module.decoder.layers.12.pre_mlp_layernorm.weight + module.decoder.layers.12.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.12.self_attention.linear_kv_down_proj.weight + module.decoder.layers.12.self_attention.linear_q_proj.weight + module.decoder.layers.12.self_attention.linear_proj.weight + module.decoder.layers.12.input_layernorm.weight +Params for bucket 11 (42734080 elements, 42734080 padded size): + module.decoder.layers.11.self_attention.linear_kv_up_proj.weight + module.decoder.layers.11.mlp.router.weight + module.decoder.layers.10.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.11.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.11.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.11.pre_mlp_layernorm.weight + module.decoder.layers.11.self_attention.linear_kv_down_proj.weight + module.decoder.layers.11.self_attention.linear_q_proj.weight + module.decoder.layers.11.self_attention.linear_proj.weight + module.decoder.layers.11.input_layernorm.weight + module.decoder.layers.10.mlp.shared_experts.linear_fc1.weight +Params for bucket 12 (40901632 elements, 40901632 padded size): + module.decoder.layers.10.mlp.router.weight + module.decoder.layers.10.self_attention.linear_kv_up_proj.weight + module.decoder.layers.9.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.9.pre_mlp_layernorm.weight + module.decoder.layers.9.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.9.self_attention.linear_kv_down_proj.weight + module.decoder.layers.10.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.10.pre_mlp_layernorm.weight + module.decoder.layers.10.self_attention.linear_kv_down_proj.weight + module.decoder.layers.10.self_attention.linear_q_proj.weight + module.decoder.layers.10.self_attention.linear_proj.weight + module.decoder.layers.10.input_layernorm.weight + module.decoder.layers.9.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.9.mlp.router.weight + module.decoder.layers.9.self_attention.linear_kv_up_proj.weight + module.decoder.layers.9.self_attention.linear_q_proj.weight +Params for bucket 13 (41163264 elements, 41163264 padded size): + module.decoder.layers.8.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.9.input_layernorm.weight + module.decoder.layers.8.mlp.router.weight + module.decoder.layers.8.self_attention.linear_kv_up_proj.weight + module.decoder.layers.7.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.9.self_attention.linear_proj.weight + module.decoder.layers.8.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.8.pre_mlp_layernorm.weight + module.decoder.layers.8.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.8.self_attention.linear_kv_down_proj.weight + module.decoder.layers.8.self_attention.linear_q_proj.weight + module.decoder.layers.8.self_attention.linear_proj.weight + module.decoder.layers.8.input_layernorm.weight +Params for bucket 14 (42734080 elements, 42734080 padded size): + module.decoder.layers.7.self_attention.linear_kv_up_proj.weight + module.decoder.layers.7.mlp.router.weight + module.decoder.layers.6.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.7.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.7.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.7.pre_mlp_layernorm.weight + module.decoder.layers.7.self_attention.linear_kv_down_proj.weight + module.decoder.layers.7.self_attention.linear_q_proj.weight + module.decoder.layers.7.self_attention.linear_proj.weight + module.decoder.layers.7.input_layernorm.weight + module.decoder.layers.6.mlp.shared_experts.linear_fc1.weight +Params for bucket 15 (40901632 elements, 40901632 padded size): + module.decoder.layers.6.mlp.router.weight + module.decoder.layers.6.self_attention.linear_kv_up_proj.weight + module.decoder.layers.5.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.5.mlp.router.weight + module.decoder.layers.5.self_attention.linear_kv_up_proj.weight + module.decoder.layers.6.self_attention.linear_kv_down_proj.weight + module.decoder.layers.6.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.6.pre_mlp_layernorm.weight + module.decoder.layers.6.self_attention.linear_q_proj.weight + module.decoder.layers.6.self_attention.linear_proj.weight + module.decoder.layers.6.input_layernorm.weight + module.decoder.layers.5.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.5.pre_mlp_layernorm.weight + module.decoder.layers.5.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.5.self_attention.linear_kv_down_proj.weight + module.decoder.layers.5.self_attention.linear_q_proj.weight +Params for bucket 16 (41163264 elements, 41163264 padded size): + module.decoder.layers.4.mlp.router.weight + module.decoder.layers.4.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.4.self_attention.linear_kv_up_proj.weight + module.decoder.layers.4.input_layernorm.weight + module.decoder.layers.5.self_attention.linear_proj.weight + module.decoder.layers.4.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.5.input_layernorm.weight + module.decoder.layers.4.pre_mlp_layernorm.weight + module.decoder.layers.4.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.4.self_attention.linear_kv_down_proj.weight + module.decoder.layers.4.self_attention.linear_q_proj.weight + module.decoder.layers.4.self_attention.linear_proj.weight + module.decoder.layers.3.mlp.shared_experts.linear_fc2.weight +Params for bucket 17 (42734080 elements, 42734080 padded size): + module.decoder.layers.3.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.3.pre_mlp_layernorm.weight + module.decoder.layers.3.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.3.self_attention.linear_kv_down_proj.weight + module.decoder.layers.3.self_attention.linear_q_proj.weight + module.decoder.layers.3.self_attention.linear_proj.weight + module.decoder.layers.3.input_layernorm.weight + module.decoder.layers.2.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.3.mlp.router.weight + module.decoder.layers.3.self_attention.linear_kv_up_proj.weight + module.decoder.layers.2.mlp.shared_experts.linear_fc2.weight +Params for bucket 18 (40901632 elements, 40901632 padded size): + module.decoder.layers.2.mlp.router.weight + module.decoder.layers.2.self_attention.linear_kv_up_proj.weight + module.decoder.layers.2.self_attention.linear_q_proj.weight + module.decoder.layers.2.self_attention.linear_proj.weight + module.decoder.layers.2.input_layernorm.weight + module.decoder.layers.1.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.1.mlp.router.weight + module.decoder.layers.1.self_attention.linear_kv_up_proj.weight + module.decoder.layers.1.self_attention.linear_kv_down_proj.weight + module.decoder.layers.2.self_attention.linear_kv_down_proj.weight + module.decoder.layers.2.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.2.pre_mlp_layernorm.weight + module.decoder.layers.1.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.1.pre_mlp_layernorm.weight + module.decoder.layers.1.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.1.self_attention.linear_q_proj.weight +Params for bucket 19 (71436288 elements, 71436288 padded size): + module.decoder.layers.1.self_attention.linear_proj.weight + module.decoder.layers.0.mlp.linear_fc2.weight + module.decoder.layers.0.mlp.linear_fc1.weight + module.decoder.layers.1.input_layernorm.weight +Params for bucket 20 (116789760 elements, 116789760 padded size): + module.decoder.layers.0.self_attention.linear_q_proj.weight + module.decoder.layers.0.self_attention.linear_kv_up_proj.weight + module.decoder.layers.0.self_attention.linear_proj.weight + module.decoder.layers.0.mlp.linear_fc1.layer_norm_weight + module.decoder.layers.0.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.0.self_attention.linear_kv_down_proj.weight + module.decoder.layers.0.input_layernorm.weight + module.embedding.word_embeddings.weight +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 40 +Params for bucket 1 (40370176 elements, 40370176 padded size): + module.decoder.layers.23.mlp.experts.linear_fc2.weight7 + module.decoder.layers.23.mlp.experts.linear_fc2.weight5 + module.decoder.layers.23.mlp.experts.linear_fc2.weight3 + module.decoder.layers.23.mlp.experts.linear_fc2.weight1 + module.decoder.layers.23.mlp.experts.linear_fc1.weight7 + module.decoder.layers.23.mlp.experts.linear_fc1.weight5 + module.decoder.layers.23.mlp.experts.linear_fc2.weight4 + module.decoder.layers.23.mlp.experts.linear_fc2.weight6 + module.decoder.layers.23.mlp.experts.linear_fc2.weight2 + module.decoder.layers.23.mlp.experts.linear_fc2.weight0 + module.decoder.layers.23.mlp.experts.linear_fc1.weight6 +Params for bucket 2 (40370176 elements, 40370176 padded size): + module.decoder.layers.23.mlp.experts.linear_fc1.weight1 + module.decoder.layers.23.mlp.experts.linear_fc1.weight3 + module.decoder.layers.22.mlp.experts.linear_fc2.weight7 + module.decoder.layers.22.mlp.experts.linear_fc2.weight5 + module.decoder.layers.23.mlp.experts.linear_fc1.weight4 + module.decoder.layers.23.mlp.experts.linear_fc1.weight2 + module.decoder.layers.23.mlp.experts.linear_fc1.weight0 + module.decoder.layers.22.mlp.experts.linear_fc2.weight6 + module.decoder.layers.22.mlp.experts.linear_fc2.weight4 +Params for bucket 3 (40370176 elements, 40370176 padded size): + module.decoder.layers.22.mlp.experts.linear_fc2.weight3 + module.decoder.layers.22.mlp.experts.linear_fc2.weight1 + module.decoder.layers.22.mlp.experts.linear_fc1.weight6 + module.decoder.layers.22.mlp.experts.linear_fc1.weight4 + module.decoder.layers.22.mlp.experts.linear_fc1.weight7 + module.decoder.layers.22.mlp.experts.linear_fc2.weight2 + module.decoder.layers.22.mlp.experts.linear_fc2.weight0 + module.decoder.layers.22.mlp.experts.linear_fc1.weight5 + module.decoder.layers.22.mlp.experts.linear_fc1.weight3 +Params for bucket 4 (40370176 elements, 40370176 padded size): + module.decoder.layers.22.mlp.experts.linear_fc1.weight2 + module.decoder.layers.21.mlp.experts.linear_fc2.weight6 + module.decoder.layers.22.mlp.experts.linear_fc1.weight0 + module.decoder.layers.21.mlp.experts.linear_fc2.weight4 + module.decoder.layers.21.mlp.experts.linear_fc2.weight2 + module.decoder.layers.21.mlp.experts.linear_fc2.weight0 + module.decoder.layers.22.mlp.experts.linear_fc1.weight1 + module.decoder.layers.21.mlp.experts.linear_fc2.weight7 + module.decoder.layers.21.mlp.experts.linear_fc2.weight5 + module.decoder.layers.21.mlp.experts.linear_fc2.weight3 + module.decoder.layers.21.mlp.experts.linear_fc2.weight1 +Params for bucket 5 (40370176 elements, 40370176 padded size): + module.decoder.layers.21.mlp.experts.linear_fc1.weight7 + module.decoder.layers.21.mlp.experts.linear_fc1.weight5 + module.decoder.layers.21.mlp.experts.linear_fc1.weight3 + module.decoder.layers.21.mlp.experts.linear_fc1.weight1 + module.decoder.layers.21.mlp.experts.linear_fc1.weight6 + module.decoder.layers.21.mlp.experts.linear_fc1.weight4 + module.decoder.layers.21.mlp.experts.linear_fc1.weight2 +Params for bucket 6 (40370176 elements, 40370176 padded size): + module.decoder.layers.20.mlp.experts.linear_fc2.weight4 + module.decoder.layers.20.mlp.experts.linear_fc2.weight6 + module.decoder.layers.20.mlp.experts.linear_fc2.weight2 + module.decoder.layers.20.mlp.experts.linear_fc2.weight0 + module.decoder.layers.20.mlp.experts.linear_fc1.weight7 + module.decoder.layers.21.mlp.experts.linear_fc1.weight0 + module.decoder.layers.20.mlp.experts.linear_fc2.weight7 + module.decoder.layers.20.mlp.experts.linear_fc2.weight5 + module.decoder.layers.20.mlp.experts.linear_fc2.weight3 + module.decoder.layers.20.mlp.experts.linear_fc2.weight1 + module.decoder.layers.20.mlp.experts.linear_fc1.weight6 +Params for bucket 7 (40370176 elements, 40370176 padded size): + module.decoder.layers.20.mlp.experts.linear_fc1.weight5 + module.decoder.layers.20.mlp.experts.linear_fc1.weight1 + module.decoder.layers.20.mlp.experts.linear_fc1.weight3 + module.decoder.layers.19.mlp.experts.linear_fc2.weight6 + module.decoder.layers.20.mlp.experts.linear_fc1.weight2 + module.decoder.layers.20.mlp.experts.linear_fc1.weight4 + module.decoder.layers.20.mlp.experts.linear_fc1.weight0 + module.decoder.layers.19.mlp.experts.linear_fc2.weight7 +Params for bucket 8 (40370176 elements, 40370176 padded size): + module.decoder.layers.19.mlp.experts.linear_fc2.weight4 + module.decoder.layers.19.mlp.experts.linear_fc2.weight2 + module.decoder.layers.19.mlp.experts.linear_fc2.weight0 + module.decoder.layers.19.mlp.experts.linear_fc1.weight7 + module.decoder.layers.19.mlp.experts.linear_fc1.weight5 + module.decoder.layers.19.mlp.experts.linear_fc2.weight5 + module.decoder.layers.19.mlp.experts.linear_fc2.weight1 + module.decoder.layers.19.mlp.experts.linear_fc2.weight3 + module.decoder.layers.19.mlp.experts.linear_fc1.weight6 + module.decoder.layers.19.mlp.experts.linear_fc1.weight4 +Params for bucket 9 (40370176 elements, 40370176 padded size): + module.decoder.layers.19.mlp.experts.linear_fc1.weight3 + module.decoder.layers.19.mlp.experts.linear_fc1.weight1 + module.decoder.layers.18.mlp.experts.linear_fc2.weight6 + module.decoder.layers.18.mlp.experts.linear_fc2.weight4 + module.decoder.layers.18.mlp.experts.linear_fc2.weight2 + module.decoder.layers.18.mlp.experts.linear_fc2.weight7 + module.decoder.layers.19.mlp.experts.linear_fc1.weight2 + module.decoder.layers.19.mlp.experts.linear_fc1.weight0 + module.decoder.layers.18.mlp.experts.linear_fc2.weight5 + module.decoder.layers.18.mlp.experts.linear_fc2.weight3 +Params for bucket 10 (40370176 elements, 40370176 padded size): + module.decoder.layers.18.mlp.experts.linear_fc1.weight7 + module.decoder.layers.18.mlp.experts.linear_fc1.weight5 + module.decoder.layers.18.mlp.experts.linear_fc2.weight0 + module.decoder.layers.18.mlp.experts.linear_fc1.weight3 + module.decoder.layers.18.mlp.experts.linear_fc2.weight1 + module.decoder.layers.18.mlp.experts.linear_fc1.weight6 + module.decoder.layers.18.mlp.experts.linear_fc1.weight4 + module.decoder.layers.18.mlp.experts.linear_fc1.weight2 +Params for bucket 11 (40370176 elements, 40370176 padded size): + module.decoder.layers.18.mlp.experts.linear_fc1.weight1 + module.decoder.layers.17.mlp.experts.linear_fc2.weight6 + module.decoder.layers.17.mlp.experts.linear_fc2.weight4 + module.decoder.layers.17.mlp.experts.linear_fc2.weight2 + module.decoder.layers.17.mlp.experts.linear_fc2.weight0 + module.decoder.layers.17.mlp.experts.linear_fc1.weight7 + module.decoder.layers.17.mlp.experts.linear_fc2.weight7 + module.decoder.layers.17.mlp.experts.linear_fc2.weight5 + module.decoder.layers.18.mlp.experts.linear_fc1.weight0 + module.decoder.layers.17.mlp.experts.linear_fc2.weight3 + module.decoder.layers.17.mlp.experts.linear_fc2.weight1 +Params for bucket 12 (40370176 elements, 40370176 padded size): + module.decoder.layers.17.mlp.experts.linear_fc1.weight5 + module.decoder.layers.17.mlp.experts.linear_fc1.weight3 + module.decoder.layers.17.mlp.experts.linear_fc1.weight1 + module.decoder.layers.17.mlp.experts.linear_fc1.weight6 + module.decoder.layers.17.mlp.experts.linear_fc1.weight4 + module.decoder.layers.17.mlp.experts.linear_fc1.weight2 + module.decoder.layers.17.mlp.experts.linear_fc1.weight0 +Params for bucket 13 (40370176 elements, 40370176 padded size): + module.decoder.layers.16.mlp.experts.linear_fc2.weight4 + module.decoder.layers.16.mlp.experts.linear_fc2.weight6 + module.decoder.layers.16.mlp.experts.linear_fc2.weight2 + module.decoder.layers.16.mlp.experts.linear_fc2.weight0 + module.decoder.layers.16.mlp.experts.linear_fc1.weight7 + module.decoder.layers.16.mlp.experts.linear_fc1.weight5 + module.decoder.layers.16.mlp.experts.linear_fc2.weight7 + module.decoder.layers.16.mlp.experts.linear_fc2.weight3 + module.decoder.layers.16.mlp.experts.linear_fc2.weight5 + module.decoder.layers.16.mlp.experts.linear_fc2.weight1 + module.decoder.layers.16.mlp.experts.linear_fc1.weight6 +Params for bucket 14 (40370176 elements, 40370176 padded size): + module.decoder.layers.16.mlp.experts.linear_fc1.weight3 + module.decoder.layers.16.mlp.experts.linear_fc1.weight1 + module.decoder.layers.15.mlp.experts.linear_fc2.weight6 + module.decoder.layers.15.mlp.experts.linear_fc2.weight4 + module.decoder.layers.16.mlp.experts.linear_fc1.weight4 + module.decoder.layers.16.mlp.experts.linear_fc1.weight0 + module.decoder.layers.16.mlp.experts.linear_fc1.weight2 + module.decoder.layers.15.mlp.experts.linear_fc2.weight7 + module.decoder.layers.15.mlp.experts.linear_fc2.weight5 +Params for bucket 15 (40370176 elements, 40370176 padded size): + module.decoder.layers.15.mlp.experts.linear_fc2.weight0 + module.decoder.layers.15.mlp.experts.linear_fc2.weight2 + module.decoder.layers.15.mlp.experts.linear_fc1.weight7 + module.decoder.layers.15.mlp.experts.linear_fc1.weight5 + module.decoder.layers.15.mlp.experts.linear_fc1.weight3 + module.decoder.layers.15.mlp.experts.linear_fc2.weight3 + module.decoder.layers.15.mlp.experts.linear_fc2.weight1 + module.decoder.layers.15.mlp.experts.linear_fc1.weight6 + module.decoder.layers.15.mlp.experts.linear_fc1.weight4 +Params for bucket 16 (40370176 elements, 40370176 padded size): + module.decoder.layers.15.mlp.experts.linear_fc1.weight1 + module.decoder.layers.14.mlp.experts.linear_fc2.weight7 + module.decoder.layers.14.mlp.experts.linear_fc2.weight5 + module.decoder.layers.14.mlp.experts.linear_fc2.weight3 + module.decoder.layers.14.mlp.experts.linear_fc2.weight1 + module.decoder.layers.15.mlp.experts.linear_fc1.weight2 + module.decoder.layers.15.mlp.experts.linear_fc1.weight0 + module.decoder.layers.14.mlp.experts.linear_fc2.weight6 + module.decoder.layers.14.mlp.experts.linear_fc2.weight4 + module.decoder.layers.14.mlp.experts.linear_fc2.weight2 + module.decoder.layers.14.mlp.experts.linear_fc2.weight0 +Params for bucket 17 (40370176 elements, 40370176 padded size): + module.decoder.layers.14.mlp.experts.linear_fc1.weight6 + module.decoder.layers.14.mlp.experts.linear_fc1.weight4 + module.decoder.layers.14.mlp.experts.linear_fc1.weight2 + module.decoder.layers.14.mlp.experts.linear_fc1.weight7 + module.decoder.layers.14.mlp.experts.linear_fc1.weight3 + module.decoder.layers.14.mlp.experts.linear_fc1.weight5 + module.decoder.layers.14.mlp.experts.linear_fc1.weight1 +Params for bucket 18 (40370176 elements, 40370176 padded size): + module.decoder.layers.14.mlp.experts.linear_fc1.weight0 + module.decoder.layers.13.mlp.experts.linear_fc2.weight5 + module.decoder.layers.13.mlp.experts.linear_fc2.weight7 + module.decoder.layers.13.mlp.experts.linear_fc2.weight3 + module.decoder.layers.13.mlp.experts.linear_fc2.weight1 + module.decoder.layers.13.mlp.experts.linear_fc1.weight6 + module.decoder.layers.13.mlp.experts.linear_fc2.weight6 + module.decoder.layers.13.mlp.experts.linear_fc2.weight4 + module.decoder.layers.13.mlp.experts.linear_fc2.weight2 + module.decoder.layers.13.mlp.experts.linear_fc2.weight0 + module.decoder.layers.13.mlp.experts.linear_fc1.weight7 +Params for bucket 19 (40370176 elements, 40370176 padded size): + module.decoder.layers.13.mlp.experts.linear_fc1.weight4 + module.decoder.layers.13.mlp.experts.linear_fc1.weight2 + module.decoder.layers.13.mlp.experts.linear_fc1.weight0 + module.decoder.layers.12.mlp.experts.linear_fc2.weight7 + module.decoder.layers.13.mlp.experts.linear_fc1.weight5 + module.decoder.layers.13.mlp.experts.linear_fc1.weight3 + module.decoder.layers.13.mlp.experts.linear_fc1.weight1 + module.decoder.layers.12.mlp.experts.linear_fc2.weight6 +Params for bucket 20 (40370176 elements, 40370176 padded size): + module.decoder.layers.12.mlp.experts.linear_fc2.weight5 + module.decoder.layers.12.mlp.experts.linear_fc2.weight1 + module.decoder.layers.12.mlp.experts.linear_fc2.weight3 + module.decoder.layers.12.mlp.experts.linear_fc1.weight6 + module.decoder.layers.12.mlp.experts.linear_fc1.weight4 + module.decoder.layers.12.mlp.experts.linear_fc2.weight2 + module.decoder.layers.12.mlp.experts.linear_fc2.weight4 + module.decoder.layers.12.mlp.experts.linear_fc2.weight0 + module.decoder.layers.12.mlp.experts.linear_fc1.weight7 + module.decoder.layers.12.mlp.experts.linear_fc1.weight5 +Params for bucket 21 (40370176 elements, 40370176 padded size): + module.decoder.layers.11.mlp.experts.linear_fc2.weight7 + module.decoder.layers.12.mlp.experts.linear_fc1.weight2 + module.decoder.layers.12.mlp.experts.linear_fc1.weight0 + module.decoder.layers.11.mlp.experts.linear_fc2.weight5 + module.decoder.layers.11.mlp.experts.linear_fc2.weight3 + module.decoder.layers.12.mlp.experts.linear_fc1.weight3 + module.decoder.layers.12.mlp.experts.linear_fc1.weight1 + module.decoder.layers.11.mlp.experts.linear_fc2.weight6 + module.decoder.layers.11.mlp.experts.linear_fc2.weight4 + module.decoder.layers.11.mlp.experts.linear_fc2.weight2 +Params for bucket 22 (40370176 elements, 40370176 padded size): + module.decoder.layers.11.mlp.experts.linear_fc2.weight1 + module.decoder.layers.11.mlp.experts.linear_fc1.weight6 + module.decoder.layers.11.mlp.experts.linear_fc1.weight4 + module.decoder.layers.11.mlp.experts.linear_fc1.weight2 + module.decoder.layers.11.mlp.experts.linear_fc1.weight5 + module.decoder.layers.11.mlp.experts.linear_fc2.weight0 + module.decoder.layers.11.mlp.experts.linear_fc1.weight7 + module.decoder.layers.11.mlp.experts.linear_fc1.weight3 +Params for bucket 23 (40370176 elements, 40370176 padded size): + module.decoder.layers.10.mlp.experts.linear_fc2.weight7 + module.decoder.layers.10.mlp.experts.linear_fc2.weight5 + module.decoder.layers.11.mlp.experts.linear_fc1.weight0 + module.decoder.layers.10.mlp.experts.linear_fc2.weight3 + module.decoder.layers.10.mlp.experts.linear_fc2.weight1 + module.decoder.layers.11.mlp.experts.linear_fc1.weight1 + module.decoder.layers.10.mlp.experts.linear_fc2.weight6 + module.decoder.layers.10.mlp.experts.linear_fc2.weight4 + module.decoder.layers.10.mlp.experts.linear_fc2.weight2 + module.decoder.layers.10.mlp.experts.linear_fc2.weight0 + module.decoder.layers.10.mlp.experts.linear_fc1.weight7 +Params for bucket 24 (40370176 elements, 40370176 padded size): + module.decoder.layers.10.mlp.experts.linear_fc1.weight6 + module.decoder.layers.10.mlp.experts.linear_fc1.weight4 + module.decoder.layers.10.mlp.experts.linear_fc1.weight2 + module.decoder.layers.10.mlp.experts.linear_fc1.weight0 + module.decoder.layers.10.mlp.experts.linear_fc1.weight5 + module.decoder.layers.10.mlp.experts.linear_fc1.weight3 + module.decoder.layers.10.mlp.experts.linear_fc1.weight1 +Params for bucket 25 (40370176 elements, 40370176 padded size): + module.decoder.layers.9.mlp.experts.linear_fc2.weight7 + module.decoder.layers.9.mlp.experts.linear_fc2.weight5 + module.decoder.layers.9.mlp.experts.linear_fc2.weight3 + module.decoder.layers.9.mlp.experts.linear_fc2.weight1 + module.decoder.layers.9.mlp.experts.linear_fc1.weight6 + module.decoder.layers.9.mlp.experts.linear_fc2.weight4 + module.decoder.layers.9.mlp.experts.linear_fc2.weight6 + module.decoder.layers.9.mlp.experts.linear_fc2.weight2 + module.decoder.layers.9.mlp.experts.linear_fc2.weight0 + module.decoder.layers.9.mlp.experts.linear_fc1.weight7 + module.decoder.layers.9.mlp.experts.linear_fc1.weight5 +Params for bucket 26 (40370176 elements, 40370176 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight4 + module.decoder.layers.9.mlp.experts.linear_fc1.weight2 + module.decoder.layers.9.mlp.experts.linear_fc1.weight0 + module.decoder.layers.8.mlp.experts.linear_fc2.weight6 + module.decoder.layers.8.mlp.experts.linear_fc2.weight4 + module.decoder.layers.9.mlp.experts.linear_fc1.weight1 + module.decoder.layers.9.mlp.experts.linear_fc1.weight3 + module.decoder.layers.8.mlp.experts.linear_fc2.weight7 + module.decoder.layers.8.mlp.experts.linear_fc2.weight5 +Params for bucket 27 (40370176 elements, 40370176 padded size): + module.decoder.layers.8.mlp.experts.linear_fc2.weight0 + module.decoder.layers.8.mlp.experts.linear_fc2.weight2 + module.decoder.layers.8.mlp.experts.linear_fc1.weight6 + module.decoder.layers.8.mlp.experts.linear_fc1.weight4 + module.decoder.layers.8.mlp.experts.linear_fc2.weight3 + module.decoder.layers.8.mlp.experts.linear_fc2.weight1 + module.decoder.layers.8.mlp.experts.linear_fc1.weight7 + module.decoder.layers.8.mlp.experts.linear_fc1.weight5 + module.decoder.layers.8.mlp.experts.linear_fc1.weight3 +Params for bucket 28 (40370176 elements, 40370176 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight2 + module.decoder.layers.7.mlp.experts.linear_fc2.weight7 + module.decoder.layers.8.mlp.experts.linear_fc1.weight0 + module.decoder.layers.7.mlp.experts.linear_fc2.weight5 + module.decoder.layers.7.mlp.experts.linear_fc2.weight3 + module.decoder.layers.7.mlp.experts.linear_fc2.weight1 + module.decoder.layers.7.mlp.experts.linear_fc2.weight6 + module.decoder.layers.8.mlp.experts.linear_fc1.weight1 + module.decoder.layers.7.mlp.experts.linear_fc2.weight4 + module.decoder.layers.7.mlp.experts.linear_fc2.weight2 + module.decoder.layers.7.mlp.experts.linear_fc2.weight0 +Params for bucket 29 (40370176 elements, 40370176 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight6 + module.decoder.layers.7.mlp.experts.linear_fc1.weight4 + module.decoder.layers.7.mlp.experts.linear_fc1.weight2 + module.decoder.layers.7.mlp.experts.linear_fc1.weight7 + module.decoder.layers.7.mlp.experts.linear_fc1.weight5 + module.decoder.layers.7.mlp.experts.linear_fc1.weight3 + module.decoder.layers.7.mlp.experts.linear_fc1.weight1 +Params for bucket 30 (40370176 elements, 40370176 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight0 + module.decoder.layers.6.mlp.experts.linear_fc2.weight5 + module.decoder.layers.6.mlp.experts.linear_fc2.weight7 + module.decoder.layers.6.mlp.experts.linear_fc2.weight3 + module.decoder.layers.6.mlp.experts.linear_fc2.weight1 + module.decoder.layers.6.mlp.experts.linear_fc1.weight6 + module.decoder.layers.6.mlp.experts.linear_fc2.weight6 + module.decoder.layers.6.mlp.experts.linear_fc2.weight4 + module.decoder.layers.6.mlp.experts.linear_fc2.weight2 + module.decoder.layers.6.mlp.experts.linear_fc2.weight0 + module.decoder.layers.6.mlp.experts.linear_fc1.weight7 +Params for bucket 31 (40370176 elements, 40370176 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight4 + module.decoder.layers.6.mlp.experts.linear_fc1.weight2 + module.decoder.layers.6.mlp.experts.linear_fc1.weight0 + module.decoder.layers.5.mlp.experts.linear_fc2.weight7 + module.decoder.layers.6.mlp.experts.linear_fc1.weight5 + module.decoder.layers.6.mlp.experts.linear_fc1.weight1 + module.decoder.layers.6.mlp.experts.linear_fc1.weight3 + module.decoder.layers.5.mlp.experts.linear_fc2.weight6 +Params for bucket 32 (40370176 elements, 40370176 padded size): + module.decoder.layers.5.mlp.experts.linear_fc2.weight5 + module.decoder.layers.5.mlp.experts.linear_fc2.weight1 + module.decoder.layers.5.mlp.experts.linear_fc2.weight3 + module.decoder.layers.5.mlp.experts.linear_fc1.weight6 + module.decoder.layers.5.mlp.experts.linear_fc1.weight4 + module.decoder.layers.5.mlp.experts.linear_fc2.weight2 + module.decoder.layers.5.mlp.experts.linear_fc2.weight4 + module.decoder.layers.5.mlp.experts.linear_fc2.weight0 + module.decoder.layers.5.mlp.experts.linear_fc1.weight7 + module.decoder.layers.5.mlp.experts.linear_fc1.weight5 +Params for bucket 33 (40370176 elements, 40370176 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight7 + module.decoder.layers.5.mlp.experts.linear_fc1.weight2 + module.decoder.layers.5.mlp.experts.linear_fc1.weight0 + module.decoder.layers.4.mlp.experts.linear_fc2.weight5 + module.decoder.layers.4.mlp.experts.linear_fc2.weight3 + module.decoder.layers.5.mlp.experts.linear_fc1.weight3 + module.decoder.layers.5.mlp.experts.linear_fc1.weight1 + module.decoder.layers.4.mlp.experts.linear_fc2.weight6 + module.decoder.layers.4.mlp.experts.linear_fc2.weight4 + module.decoder.layers.4.mlp.experts.linear_fc2.weight2 +Params for bucket 34 (40370176 elements, 40370176 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight1 + module.decoder.layers.4.mlp.experts.linear_fc1.weight6 + module.decoder.layers.4.mlp.experts.linear_fc1.weight4 + module.decoder.layers.4.mlp.experts.linear_fc1.weight2 + module.decoder.layers.4.mlp.experts.linear_fc2.weight0 + module.decoder.layers.4.mlp.experts.linear_fc1.weight7 + module.decoder.layers.4.mlp.experts.linear_fc1.weight5 + module.decoder.layers.4.mlp.experts.linear_fc1.weight3 +Params for bucket 35 (40370176 elements, 40370176 padded size): + module.decoder.layers.3.mlp.experts.linear_fc2.weight6 + module.decoder.layers.4.mlp.experts.linear_fc1.weight0 + module.decoder.layers.3.mlp.experts.linear_fc2.weight4 + module.decoder.layers.3.mlp.experts.linear_fc2.weight2 + module.decoder.layers.3.mlp.experts.linear_fc2.weight0 + module.decoder.layers.3.mlp.experts.linear_fc1.weight7 + module.decoder.layers.4.mlp.experts.linear_fc1.weight1 + module.decoder.layers.3.mlp.experts.linear_fc2.weight7 + module.decoder.layers.3.mlp.experts.linear_fc2.weight5 + module.decoder.layers.3.mlp.experts.linear_fc2.weight3 + module.decoder.layers.3.mlp.experts.linear_fc2.weight1 +Params for bucket 36 (40370176 elements, 40370176 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight3 + module.decoder.layers.3.mlp.experts.linear_fc1.weight5 + module.decoder.layers.3.mlp.experts.linear_fc1.weight1 + module.decoder.layers.3.mlp.experts.linear_fc1.weight6 + module.decoder.layers.3.mlp.experts.linear_fc1.weight2 + module.decoder.layers.3.mlp.experts.linear_fc1.weight4 + module.decoder.layers.3.mlp.experts.linear_fc1.weight0 +Params for bucket 37 (40370176 elements, 40370176 padded size): + module.decoder.layers.2.mlp.experts.linear_fc2.weight7 + module.decoder.layers.2.mlp.experts.linear_fc2.weight5 + module.decoder.layers.2.mlp.experts.linear_fc2.weight3 + module.decoder.layers.2.mlp.experts.linear_fc2.weight1 + module.decoder.layers.2.mlp.experts.linear_fc1.weight7 + module.decoder.layers.2.mlp.experts.linear_fc1.weight5 + module.decoder.layers.2.mlp.experts.linear_fc2.weight6 + module.decoder.layers.2.mlp.experts.linear_fc2.weight4 + module.decoder.layers.2.mlp.experts.linear_fc2.weight2 + module.decoder.layers.2.mlp.experts.linear_fc2.weight0 + module.decoder.layers.2.mlp.experts.linear_fc1.weight6 +Params for bucket 38 (40370176 elements, 40370176 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight3 + module.decoder.layers.2.mlp.experts.linear_fc1.weight1 + module.decoder.layers.1.mlp.experts.linear_fc2.weight6 + module.decoder.layers.1.mlp.experts.linear_fc2.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight2 + module.decoder.layers.2.mlp.experts.linear_fc1.weight0 + module.decoder.layers.1.mlp.experts.linear_fc2.weight7 + module.decoder.layers.1.mlp.experts.linear_fc2.weight5 +Params for bucket 39 (40370176 elements, 40370176 padded size): + module.decoder.layers.1.mlp.experts.linear_fc2.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight7 + module.decoder.layers.1.mlp.experts.linear_fc2.weight0 + module.decoder.layers.1.mlp.experts.linear_fc1.weight5 + module.decoder.layers.1.mlp.experts.linear_fc1.weight3 + module.decoder.layers.1.mlp.experts.linear_fc2.weight3 + module.decoder.layers.1.mlp.experts.linear_fc2.weight1 + module.decoder.layers.1.mlp.experts.linear_fc1.weight6 + module.decoder.layers.1.mlp.experts.linear_fc1.weight4 +Params for bucket 40 (17301504 elements, 17301504 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight0 + module.decoder.layers.1.mlp.experts.linear_fc1.weight1 +INFO:megatron.core.optimizer:Setting up optimizer with config OptimizerConfig(optimizer='adam', lr=0.0001, min_lr=1e-05, decoupled_lr=None, decoupled_min_lr=None, weight_decay=0.1, fp8_recipe='mxfp8', fp16=False, bf16=True, reuse_grad_buf_for_mxfp8_param_ag=False, params_dtype=torch.bfloat16, use_precision_aware_optimizer=False, store_param_remainders=True, main_grads_dtype=torch.float32, main_params_dtype=torch.float32, exp_avg_dtype=torch.float32, exp_avg_sq_dtype=torch.float32, 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_param_gather_with_optimizer_step=False, optimizer_cpu_offload=False, optimizer_offload_fraction=1.0, use_torch_optimizer_for_cpu_offload=False, overlap_cpu_optimizer_d2h_h2d=False, pin_cpu_grads=True, pin_cpu_params=True, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=, config_logger_dir='') +INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine +[after model, optimizer, and learning rate scheduler are built] datetime: 2026-08-26 20:10:53 +> building train, validation, and test datasets ... + > datasets target sizes (minimum size): + train: 80 + validation: 0 + test: 0 +INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None] +> building train, validation, and test datasets for GPT ... +INFO:megatron.core.datasets.blended_megatron_dataset_builder:Building GPTDataset splits with sizes=(80, 0, 0) and config=GPTDatasetConfig(random_seed=1234, sequence_length=8192, blend=(['/workspace/data/wikitext/wikitext103_gpt2_text_document'], None), blend_per_split=None, split='100,0,0', split_matrix=[(0, 1.0), None, None], num_dataset_builder_threads=1, path_to_cache=None, mmap_bin_files=True, mock=False, tokenizer=, mid_level_dataset_surplus=0.005, reset_position_ids=False, reset_attention_mask=False, eod_mask_loss=False, create_attention_mask=True, drop_last_partial_validation_sequence=True, add_extra_token_to_sequence=True, object_storage_cache_path=None) +INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_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: 1165029 +INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350 +INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices +INFO:megatron.core.datasets.gpt_dataset: Load the document index from 715d60356d3f5be048391704c2aaf5b2-GPTDataset-train-document_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 715d60356d3f5be048391704c2aaf5b2-GPTDataset-train-sample_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 715d60356d3f5be048391704c2aaf5b2-GPTDataset-train-shuffle_index.npy +INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 14536 +> finished creating GPT datasets ... +[after dataloaders are built] datetime: 2026-08-26 20:10:54 +done with setup ... +(min, max) time across ranks (ms): + model-and-optimizer-setup ......................: (3512.50, 3521.06) + train/valid/test-data-iterators-setup ..........: (96.67, 124.47) +training ... +Overwriting rerun_state_machine.current_iteration from -1 to 0... +[before the start of training step] datetime: 2026-08-26 20:10:54 +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") + [2026-08-26 20:11:25] iteration 1/ 10 | consumed samples: 8 | elapsed time per iteration (ms): 31288.2 | throughput per GPU (TFLOP/s/GPU): 4.1 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.092127E+01 | loss scale: 1.0 | grad norm: 30.623 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | +Number of parameters in transformer block in billions: 13.53 +Number of parameters in embedding layers in billions: 0.21 +Total number of parameters in billions: 13.74 +Number of parameters in most loaded shard in billions: 13.7356 +Theoretical memory footprints: weight and optimizer=98244.96 MB +[Rank 0] (after 1 iterations) memory (MB) | allocated: 37186.783203125 | max allocated: 102778.46142578125 | reserved: 105676.0 | max reserved: 105676.0 +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") + [2026-08-26 20:11:26] iteration 2/ 10 | consumed samples: 16 | elapsed time per iteration (ms): 597.2 | throughput per GPU (TFLOP/s/GPU): 213.6 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.091995E+01 | loss scale: 1.0 | grad norm: 30.567 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 20:11:26] iteration 3/ 10 | consumed samples: 24 | elapsed time per iteration (ms): 597.9 | throughput per GPU (TFLOP/s/GPU): 213.4 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 9.952354E+00 | loss scale: 1.0 | grad norm: 42.232 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | +Capture CUDA graph for training!!! +CUDA graph capture done!!! + [2026-08-26 20:11:28] iteration 4/ 10 | consumed samples: 32 | elapsed time per iteration (ms): 1840.6 | throughput per GPU (TFLOP/s/GPU): 69.3 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 9.477323E+00 | loss scale: 1.0 | grad norm: 7.143 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 20:11:29] iteration 5/ 10 | consumed samples: 40 | elapsed time per iteration (ms): 577.7 | throughput per GPU (TFLOP/s/GPU): 220.9 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.119604E+01 | loss scale: 1.0 | grad norm: 396.977 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 20:11:29] iteration 6/ 10 | consumed samples: 48 | elapsed time per iteration (ms): 583.0 | throughput per GPU (TFLOP/s/GPU): 218.8 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 9.367741E+00 | loss scale: 1.0 | grad norm: 5.090 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 20:11:32] iteration 7/ 10 | consumed samples: 56 | elapsed time per iteration (ms): 3317.3 | throughput per GPU (TFLOP/s/GPU): 38.5 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 8.973585E+00 | loss scale: 1.0 | grad norm: 3.690 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 20:11:33] iteration 8/ 10 | consumed samples: 64 | elapsed time per iteration (ms): 596.9 | throughput per GPU (TFLOP/s/GPU): 213.8 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.787342E+00 | loss scale: 1.0 | grad norm: 5.890 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | +Capture range started in the application. +Capture range ended in the application. +Generating '/tmp/nsys-report-da7b.qdstrm' + [1/1] [0% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-repProcessing events... + [1/1] [0% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [0% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [6% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [8% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [11% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [13% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [=15% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [==18% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [==20% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [===22% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [====25% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [====27% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [=====29% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [=====32% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [======33% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [======34% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [======35% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [=======36% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [=======37% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [=======38% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [=======39% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [========40% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [========41% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [========42% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [=========43% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [=========44% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [=========45% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [=========46% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [==========47% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [==========49% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [===========50% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [===========51% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [===========53% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [============54% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [============55% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [============56% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [============57% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [==============63% ] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [========================100%] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep [1/1] [========================100%] DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep + [2026-08-26 20:12:16] iteration 9/ 10 | consumed samples: 72 | elapsed time per iteration (ms): 43014.9 | throughput per GPU (TFLOP/s/GPU): 3.0 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.612353E+00 | loss scale: 1.0 | grad norm: 3.038 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 20:12:17] iteration 10/ 10 | consumed samples: 80 | elapsed time per iteration (ms): 597.6 | throughput per GPU (TFLOP/s/GPU): 213.5 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.076241E+00 | loss scale: 1.0 | grad norm: 29.812 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | +[after training is done] datetime: 2026-08-26 20:12:17 +[rank0]:[W826 20:12:18.445141812 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank6]:[W826 20:12:19.777471706 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank7]:[W826 20:12:19.798234331 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank2]:[W826 20:12:19.801784905 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank1]:[W826 20:12:19.804290024 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank5]:[W826 20:12:19.807250034 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank3]:[W826 20:12:19.807614604 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank4]:[W826 20:12:20.185246139 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +Collection stop is not allowed in this state. +[rank4]:[W826 20:12:23.163578737 observer.cpp:116] Warning: Leaked callback handle: 1 (function operator()) +pure virtual method called +terminate called without an active exception +[rank0]:[W826 20:12:23.341004661 observer.cpp:116] Warning: Leaked callback handle: 1 (function operator()) +pure virtual method called +terminate called without an active exception +[rank6]:[W826 20:12:23.460738146 observer.cpp:116] Warning: Leaked callback handle: 1 (function operator()) +pure virtual method called +terminate called without an active exception +W0826 20:12:23.710000 748463 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 748572 closing signal SIGTERM +W0826 20:12:23.710000 748463 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 748573 closing signal SIGTERM +W0826 20:12:23.710000 748463 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 748574 closing signal SIGTERM +W0826 20:12:23.712000 748463 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 748575 closing signal SIGTERM +W0826 20:12:23.713000 748463 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 748577 closing signal SIGTERM +W0826 20:12:23.715000 748463 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 748578 closing signal SIGTERM +W0826 20:12:23.716000 748463 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 748579 closing signal SIGTERM +E0826 20:12:25.711000 748463 torch/distributed/elastic/multiprocessing/api.py:874] failed (exitcode: -6) local_rank: 4 (pid: 748576) of binary: /usr/bin/python +Traceback (most recent call last): + File "/usr/local/bin/torchrun", line 8, in + sys.exit(main()) + ^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper + return f(*args, **kwargs) + ^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 892, in main + run(args) + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 883, in run + elastic_launch( + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__ + return launch_agent(self._config, self._entrypoint, list(args)) + ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 270, in launch_agent + raise ChildFailedError( +torch.distributed.elastic.multiprocessing.errors.ChildFailedError: +============================================================ +pretrain_gpt.py FAILED +------------------------------------------------------------ +Failures: + +------------------------------------------------------------ +Root Cause (first observed failure): +[0]: + time : 2026-08-26_20:12:23 + host : sp-aiinfra-interactive-aghosh-b5xzj-8xb200-moe-cg-0 + rank : 4 (local_rank: 4) + exitcode : -6 (pid: 748576) + error_file: + traceback : Signal 6 (SIGABRT) received by PID 748576 +============================================================ +Generated: + /workspace/tierb_study_20260826_144644/ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep diff --git a/ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/verify_trace_ranks.log b/ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/verify_trace_ranks.log new file mode 100644 index 0000000000000000000000000000000000000000..94a339829fe59fa9f4bfadb56c5edae614409a4e --- /dev/null +++ b/ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/verify_trace_ranks.log @@ -0,0 +1,5 @@ +report : /workspace/tierb_study_20260826_144644/ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep +kernels : 146491 +devices : 8 of 8 -> [0, 1, 2, 3, 4, 5, 6, 7] +processes: 1 +PASS: all ranks present diff --git a/ultraep/tierbtrace/tierbtrace_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/nsys/run.env b/ultraep/tierbtrace/tierbtrace_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/nsys/run.env new file mode 100644 index 0000000000000000000000000000000000000000..83438b5901a24078c158226b91bbfafb44196a68 --- /dev/null +++ b/ultraep/tierbtrace/tierbtrace_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/nsys/run.env @@ -0,0 +1,61 @@ +backend=ultraep +config=QWEN2_57B +family=tierB +repo=/workspace/Megatron-UltraEP +branch=ultraep-bracing-regions-through-kernels +commit=03d57ef7 +num_layers=10 +hidden=3584 +ffn_hidden=2560 +moe_ffn_hidden=2560 +num_experts=64 +topk=8 +experts_per_rank=8 +expert_params=27525120 +heads=28 +kv_groups=4 +kv_channels= +qk_layernorm=0 +shared_ffn=20480 +shared_expert_overlap=0 +mla=0 +dense_first_layers=0 +moe_layers=10 +rotary_base=1000000 +rotary_scaling= +qkv_bias=1 +router_balancing=aux_loss +aux_loss_coeff=1e-2 +router_pre_softmax=1 +router_fusion=0 +router_dtype=fp32 +device_initiated_grouped_gemm=1 +norm_epsilon=1e-6 +seq_length=4096 +micro_batch_size=1 +global_batch_size=8 +num_microbatches=1 +assignments_per_rank=32768 +mean_tokens_per_expert=4096 +slot_policy=matched +slots_per_rank=8 +capacity_mode=static +capacity=1.08984375 +cuda_graph=1 +probe_mode=0 +async_grad_reduce=0 +branch_role=timing +memory_dump_mode=none +memory_snapshot_path= +cuda_graph_warmup_steps=3 +recompute=none +force_balance=0 +train_iters=10 +ep_size=8 +echo_dispatch_overlap=1 +echo_dispatch_recompute=0 +ultraep_quota_min_tokens=512 +ultraep_weight_sync_join= +profile_enable=1 +kernel_trace=1 +fp8_recipe=mxfp8 diff --git a/ultraep/tierbtrace/tierbtrace_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/nsys/train.log b/ultraep/tierbtrace/tierbtrace_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/nsys/train.log new file mode 100644 index 0000000000000000000000000000000000000000..bd074432b2ca7a65bbe2e42491dc30e36ff8ff03 --- /dev/null +++ b/ultraep/tierbtrace/tierbtrace_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/nsys/train.log @@ -0,0 +1,1151 @@ +============================================================== + QWEN2_57B / ultraep (slots/rank=8 policy=matched) + E=64 K=8 H'=2560 m=8 + L=10 seq=4096 gbs=8 (1 microbatches) + cap_mode=static cap=1.08984375 cg=1 recompute=none + memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/ultraep/tierbtrace/tierbtrace_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/nsys/memory_snapshots/snapshot.pickle + repo=/workspace/Megatron-UltraEP @ 03d57ef7 +============================================================== +W0826 20:13:02.429000 755703 torch/distributed/run.py:766] +W0826 20:13:02.429000 755703 torch/distributed/run.py:766] ***************************************** +W0826 20:13:02.429000 755703 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0826 20:13:02.429000 755703 torch/distributed/run.py:766] ***************************************** +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +using world size: 8, data-parallel size: 8, context-parallel size: 1, hierarchical context-parallel sizes: None, tensor-model-parallel size: 1, pipeline-model-parallel size: 1 +WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer +Number of virtual stages per pipeline stage: None +accumulate and all-reduce gradients in fp32 for bfloat16 data type. +using torch.bfloat16 for parameters ... +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +------------------------ arguments ------------------------ + account_for_embedding_in_pipeline_split ......... False + account_for_loss_in_pipeline_split .............. False + 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 .................................... True + adlr_autoresume ................................. False + adlr_autoresume_interval ........................ 1000 + align_grad_reduce ............................... True + align_param_gather .............................. False + app_tag_run_name ................................ None + app_tag_run_version ............................. 0.0.0 + apply_layernorm_1p .............................. False + apply_query_key_layer_scaling ................... False + apply_residual_connection_post_layernorm ........ False + apply_rope_fusion ............................... True + async_save ...................................... None + async_tensor_model_parallel_allreduce ........... True + attention_backend ............................... AttnBackend.auto + 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 + calc_ft_timeouts ................................ False + calculate_per_token_loss ........................ False + check_for_large_grads ........................... False + check_for_nan_in_loss_and_grad .................. False + check_for_spiky_loss ............................ False + check_weight_hash_across_dp_replicas_interval ... None + ckpt_assume_constant_structure .................. False + ckpt_convert_format ............................. None + ckpt_convert_save ............................... None + ckpt_convert_update_legacy_dist_opt_format ...... False + ckpt_format ..................................... torch_dist + ckpt_fully_parallel_load ........................ False + ckpt_fully_parallel_save ........................ True + ckpt_fully_parallel_save_deprecated ............. False + ckpt_step ....................................... None + classes_fraction ................................ 1.0 + clip_grad ....................................... 1.0 + clone_scatter_output_in_embedding ............... True + config_logger_dir ............................... + consumed_train_samples .......................... 0 + consumed_valid_samples .......................... 0 + context_parallel_size ........................... 1 + cp_comm_type .................................... ['p2p'] + create_attention_mask_in_dataloader ............. True + cross_entropy_fusion_impl ....................... native + cross_entropy_loss_fusion ....................... False + cuda_graph_scope ................................ full_iteration + cuda_graph_warmup_steps ......................... 3 + data_args_path .................................. None + data_cache_path ................................. None + data_parallel_random_init ....................... False + data_parallel_sharding_strategy ................. no_shard + data_parallel_size .............................. 8 + data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document'] + data_per_class_fraction ......................... 1.0 + data_sharding ................................... True + dataloader_type ................................. single + ddp_average_in_collective ....................... False + ddp_bucket_size ................................. None + ddp_num_buckets ................................. None + ddp_pad_buckets_for_high_nccl_busbw ............. False + decoder_first_pipeline_num_layers ............... None + decoder_last_pipeline_num_layers ................ None + decoder_num_layers .............................. None + decoder_seq_length .............................. None + decoupled_lr .................................... None + decoupled_min_lr ................................ None + decrease_batch_size_if_needed ................... False + defer_embedding_wgrad_compute ................... False + delay_wgrad_compute ............................. False + deprecated_use_mcore_models ..................... True + deterministic_mode .............................. 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 + disable_bf16_reduced_precision_matmul ........... False + disable_mamba_mem_eff_path ...................... False + disable_straggler_on_startup .................... False + dist_ckpt_format_deprecated ..................... None + dist_ckpt_strictness ............................ assume_ok_unexpected + distribute_saved_activations .................... False + distributed_backend ............................. nccl + distributed_timeout_minutes ..................... 10 + embedding_init_method_std ....................... None + embedding_path .................................. None + empty_unused_memory_level ....................... 0 + enable_cuda_graph ............................... True + enable_experimental ............................. False + enable_ft_package ............................... False + enable_gloo_process_groups ...................... True + enable_msc ...................................... True + enable_one_logger ............................... True + encoder_num_layers .............................. 10 + encoder_seq_length .............................. 4096 + end_weight_decay ................................ 0.1 + eod_mask_loss ................................... False + error_injection_rate ............................ 0 + error_injection_type ............................ transient_error + eval_interval ................................... 1000000 + eval_iters ...................................... 0 + evidence_data_path .............................. None + exit_duration_in_mins ........................... None + exit_interval ................................... None + exit_on_missing_checkpoint ...................... False + exit_signal_handler ............................. False + exp_avg_dtype ................................... torch.float32 + exp_avg_sq_dtype ................................ torch.float32 + expert_model_parallel_size ...................... 8 + expert_tensor_parallel_size ..................... 1 + export_force_local_attention .................... False + export_kd_cfg ................................... None + export_kd_teacher_ckpt_format ................... None + export_kd_teacher_load .......................... None + export_kv_cache_quant ........................... False + export_legacy_megatron .......................... False + export_model_type ............................... GPTModel + export_moe_apply_probs_on_input ................. False + export_qk_l2_norm ............................... False + export_quant_cfg ................................ None + export_real_quant_cfg ........................... None + export_te_mcore_model ........................... False + external_cuda_graph ............................. False + ffn_hidden_size ................................. 2560 + finetune ........................................ False + finetune_data_split ............................. train + finetune_hf_dataset ............................. None + first_last_layers_bf16 .......................... False + flash_decode .................................... False + fp16 ............................................ False + fp16_lm_cross_entropy ........................... False + fp32_residual_connection ........................ False + fp8 ............................................. e4m3 + fp8_amax_compute_algo ........................... most_recent + fp8_amax_history_len ............................ 1 + fp8_interval .................................... 1 + fp8_margin ...................................... 0 + fp8_param_gather ................................ False + fp8_recipe ...................................... mxfp8 + fp8_wgrad ....................................... True + fsdp_double_buffer .............................. False + global_batch_size ............................... 8 + grad_reduce_in_bf16 ............................. False + gradient_accumulation_fusion .................... True + gradient_reduce_div_fusion ...................... True + group_query_attention ........................... True + head_lr_mult .................................... 1.0 + heterogeneous_layers_config_encoded_json ........ None + heterogeneous_layers_config_path ................ None + hidden_dropout .................................. 0.0 + hidden_size ..................................... 3584 + hierarchical_context_parallel_sizes ............. None + high_priority_stream_groups ..................... [] + hybrid_attention_ratio .......................... 0.0 + hybrid_mlp_ratio ................................ 0.0 + hybrid_override_pattern ......................... None + 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 .......... -1 + inference_dynamic_batching ...................... False + inference_dynamic_batching_buffer_guaranteed_fraction 0.2 + inference_dynamic_batching_buffer_overflow_factor None + inference_dynamic_batching_buffer_size_gb ....... 40.0 + inference_dynamic_batching_chunk_size ........... 256 + inference_dynamic_batching_max_requests_override None + inference_dynamic_batching_max_tokens_override .. None + inference_dynamic_batching_num_cuda_graphs ...... 16 + inference_max_batch_size ........................ 8 + inference_max_seq_length ........................ 2560 + inference_rng_tracker ........................... False + init_method_std ................................. 0.01 + init_method_xavier_uniform ...................... False + init_model_with_meta_device ..................... False + initial_loss_scale .............................. 4294967296 + inprocess_active_world_size ..................... 8 + inprocess_barrier_timeout ....................... 120 + inprocess_completion_timeout .................... 120 + inprocess_empty_cuda_cache ...................... False + inprocess_granularity ........................... node + inprocess_hard_timeout .......................... 90 + inprocess_heartbeat_interval .................... 30 + inprocess_heartbeat_timeout ..................... 60 + inprocess_last_call_wait ........................ 1 + inprocess_max_iterations ........................ None + inprocess_monitor_process_interval .............. 1.0 + inprocess_monitor_thread_interval ............... 1.0 + inprocess_progress_watchdog_interval ............ 1.0 + inprocess_restart ............................... False + inprocess_soft_timeout .......................... 60 + inprocess_termination_grace_time ................ 1 + is_hybrid_model ................................. False + iter_per_epoch .................................. 1250 + iterations_to_skip .............................. [] + keep_fp8_transpose_cache_when_using_custom_fsdp . False + kitchen_config_file ............................. None + kitchen_recipe_number ........................... None + kv_channels ..................................... 128 + kv_lora_rank .................................... 32 + lazy_mpu_init ................................... None + load ............................................ None + load_main_params_from_ckpt ...................... None + load_model_opt_format ........................... False + local_rank ...................................... 0 + log_energy ...................................... False + log_interval .................................... 1 + log_loss_scale_to_tensorboard ................... True + log_memory_to_tensorboard ....................... False + log_num_zeros_in_grad ........................... False + log_params_norm ................................. False + log_progress .................................... False + log_straggler ................................... False + log_throughput .................................. True + log_timers_to_tensorboard ....................... False + log_validation_ppl_to_tensorboard ............... False + log_world_size_to_tensorboard ................... False + logging_level ................................... None + loss_scale ...................................... None + loss_scale_window ............................... 1000 + lr .............................................. 0.0001 + lr_decay_iters .................................. 8 + lr_decay_samples ................................ None + lr_decay_style .................................. cosine + lr_warmup_fraction .............................. None + lr_warmup_init .................................. 0.0 + lr_warmup_iters ................................. 5 + lr_warmup_samples ............................... 0 + lr_wsd_decay_iters .............................. None + lr_wsd_decay_samples ............................ None + lr_wsd_decay_style .............................. exponential + main_grads_dtype ................................ torch.float32 + main_params_dtype ............................... torch.float32 + make_vocab_size_divisible_by .................... 128 + mamba_head_dim .................................. 64 + mamba_num_groups ................................ 8 + mamba_num_heads ................................. None + mamba_state_dim ................................. 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 + memory_snapshot_path ............................ snapshot.pickle + merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt + micro_batch_size ................................ 1 + microbatch_group_size_per_vp_stage .............. None + mid_level_dataset_surplus ....................... 0.005 + min_loss_scale .................................. 1.0 + min_lr .......................................... 1e-05 + mlp_chunks_for_prefill .......................... 1 + mmap_bin_files .................................. True + mock_data ....................................... False + moe_apply_probs_on_input ........................ False + moe_aux_loss_coeff .............................. 0.01 + moe_capacity_monitor ............................ True + moe_deepep_num_sms .............................. 20 + moe_echo_enable_random_offloading ............... False + moe_echo_expert_dispatch_overlap ................ False + moe_echo_expert_dispatcher_type ................. hybridep + moe_echo_recompute_expert_dispatch .............. False + moe_enable_deepep ............................... False + moe_enable_echo ................................. False + moe_enable_hybridep ............................. True + moe_enable_ultraep .............................. True + moe_expert_capacity_factor ...................... None + moe_extended_tp ................................. False + moe_ffn_hidden_size ............................. 2560 + moe_grouped_gemm ................................ True + moe_hybridep_num_sms ............................ 16 + moe_input_jitter_eps ............................ None + moe_layer_freq .................................. 1 + moe_layer_recompute ............................. False + moe_num_echo_experts ............................ None + moe_num_redundant_experts_per_rank .............. 8 + moe_pad_expert_input_to_capacity ................ False + moe_per_layer_logging ........................... False + moe_permute_fusion .............................. False + moe_received_token_capacity ..................... 1.08984375 + moe_router_bias_update_rate ..................... 0.001 + moe_router_dtype ................................ fp32 + moe_router_enable_expert_bias ................... False + moe_router_force_load_balancing ................. False + moe_router_group_topk ........................... None + moe_router_load_balancing_type .................. aux_loss + moe_router_num_groups ........................... None + moe_router_padding_for_fp8 ...................... False + moe_router_pre_softmax .......................... True + moe_router_score_function ....................... softmax + moe_router_topk ................................. 8 + moe_router_topk_scaling_factor .................. None + moe_shared_expert_intermediate_size ............. 20480 + moe_shared_expert_overlap ....................... False + moe_token_dispatcher_type ....................... flex + moe_token_drop_policy ........................... probs + moe_ultraep_max_microbatches .................... None + moe_ultraep_overlap_weight_sync ................. True + moe_ultraep_weight_sync_join .................... before_dispatch + moe_upcycling_granularity ....................... 1 + moe_use_device_initiated_grouped_gemm ........... True + moe_use_legacy_grouped_gemm ..................... False + moe_use_upcycling ............................... False + moe_z_loss_coeff ................................ None + mrope_section ................................... None + mscale .......................................... 1.0 + mscale_all_dim .................................. 0.0 + mtp_loss_scaling_factor ......................... 0.1 + mtp_num_layers .................................. None + multi_latent_attention .......................... False + nccl_all_reduce_for_prefill ..................... False + nccl_communicator_config_path ................... None + nccl_ub ......................................... False + no_load_optim ................................... None + no_load_rng ..................................... None + no_persist_layer_norm ........................... False + no_rope_freq .................................... None + no_save_optim ................................... None + no_save_rng ..................................... None + non_persistent_ckpt_type ........................ None + non_persistent_global_ckpt_dir .................. None + non_persistent_local_ckpt_algo .................. fully_parallel + non_persistent_local_ckpt_dir ................... None + non_persistent_save_interval .................... None + norm_epsilon .................................... 1e-06 + normalization ................................... RMSNorm + num_attention_heads ............................. 28 + num_channels .................................... 3 + num_classes ..................................... 1000 + num_dataset_builder_threads ..................... 1 + num_distributed_optimizer_instances ............. 1 + num_experts ..................................... 64 + num_layers ...................................... 10 + num_layers_at_end_in_bf16 ....................... 1 + num_layers_at_start_in_bf16 ..................... 1 + num_layers_per_virtual_pipeline_stage ........... None + num_query_groups ................................ 4 + num_virtual_stages_per_pipeline_rank ............ None + num_workers ..................................... 2 + object_storage_cache_path ....................... None + one_logger_async ................................ False + one_logger_project .............................. megatron-lm + one_logger_run_name ............................. None + onnx_safe ....................................... None + openai_gelu ..................................... False + optimizer ....................................... adam + optimizer_cpu_offload ........................... False + optimizer_offload_fraction ...................... 1.0 + output_bert_embeddings .......................... False + overlap_cpu_optimizer_d2h_h2d ................... False + overlap_grad_reduce ............................. True + overlap_moe_expert_parallel_comm ................ False + overlap_p2p_comm ................................ False + overlap_p2p_comm_warmup_flush ................... False + overlap_param_gather ............................ True + overlap_param_gather_with_optimizer_step ........ False + override_opt_param_scheduler .................... False + params_dtype .................................... torch.bfloat16 + patch_dim ....................................... 16 + per_split_data_args_path ........................ None + perform_initialization .......................... True + pin_cpu_grads ................................... True + pin_cpu_params .................................. True + pipeline_model_parallel_comm_backend ............ None + pipeline_model_parallel_layout .................. None + pipeline_model_parallel_size .................... 1 + position_embedding_type ......................... rope + pretrained_checkpoint ........................... None + profile ......................................... True + profile_ranks ................................... [0, 1, 2, 3, 4, 5, 6, 7] + profile_step_end ................................ 8 + profile_step_start .............................. 6 + q_lora_rank ..................................... None + qk_head_dim ..................................... 128 + qk_l2_norm ...................................... False + qk_layernorm .................................... False + qk_pos_emb_head_dim ............................. 64 + query_in_block_prob ............................. 0.1 + rampup_batch_size ............................... None + rank ............................................ 0 + recompute_granularity ........................... None + recompute_method ................................ None + recompute_modules ............................... None + recompute_num_layers ............................ None + record_memory_history ........................... False + relative_attention_max_distance ................. 128 + relative_attention_num_buckets .................. 32 + replication ..................................... False + replication_factor .............................. 2 + replication_jump ................................ None + rerun_mode ...................................... validate_results + reset_attention_mask ............................ False + reset_position_ids .............................. False + result_rejected_tracker_filename ................ None + 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 + reuse_grad_buf_for_mxfp8_param_ag ............... False + rope_scaling_factor ............................. 8.0 + rope_type ....................................... None + rotary_base ..................................... 1000000 + rotary_interleaved .............................. False + rotary_percent .................................. 1.0 + rotary_scaling_factor ........................... 1.0 + rotary_seq_len_interpolation_factor ............. None + run_workload_inspector_server ................... False + sample_rate ..................................... 1.0 + save ............................................ None + save_interval ................................... None + save_retain_interval ............................ None + scatter_gather_tensors_in_pipeline .............. True + seed ............................................ 1234 + seq_length ...................................... 4096 + sequence_parallel ............................... False + sft ............................................. False + sft_tokenizer_prompt_format ..................... nemotron-h-aligned + sgd_momentum .................................... 0.9 + short_seq_prob .................................. 0.1 + skip_train ...................................... False + skipped_train_samples ........................... 0 + spec ............................................ None + split ........................................... 100,0,0 + squared_relu .................................... False + start_weight_decay .............................. 0.1 + straggler_ctrlr_port ............................ 65535 + straggler_minmax_count .......................... 1 + suggested_communication_unit_size ............... None + swiglu .......................................... True + swin_backbone_type .............................. tiny + symmetric_ar_type ............................... None + te_rng_tracker .................................. True + tensor_model_parallel_size ...................... 1 + tensorboard_dir ................................. None + tensorboard_log_interval ........................ 1 + tensorboard_queue_size .......................... 1000 + test_data_path .................................. None + test_mode ....................................... False + tiktoken_num_special_tokens ..................... 1000 + tiktoken_pattern ................................ None + tiktoken_special_tokens ......................... None + timing_log_level ................................ 0 + timing_log_option ............................... minmax + titles_data_path ................................ None + tokenizer_model ................................. None + tokenizer_type .................................. GPT2BPETokenizer + torch_fsdp2_reshard_after_forward ............... True + tp_comm_bootstrap_backend ....................... nccl + 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_overlap_rs_dgrad ........................ False + tp_comm_split_ag ................................ True + tp_comm_split_rs ................................ True + train_data_path ................................. None + train_iters ..................................... 10 + train_samples ................................... None + train_sync_interval ............................. 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_custom_fsdp ................................. False + use_dist_ckpt ................................... True + use_dist_ckpt_deprecated ........................ False + use_distributed_optimizer ....................... True + use_flash_attn .................................. False + use_legacy_models ............................... False + use_mp_args_from_checkpoint_args ................ False + use_one_sent_docs ............................... False + use_persistent_ckpt_worker ...................... False + use_precision_aware_optimizer ................... False + use_pytorch_profiler ............................ False + use_ring_exchange_p2p ........................... False + use_rope_scaling ................................ False + use_rotary_position_embeddings .................. False + use_sharp ....................................... False + use_tokenizer_model_from_checkpoint_args ........ True + use_torch_fsdp2 ................................. False + use_torch_optimizer_for_cpu_offload ............. False + use_tp_pp_dp_mapping ............................ False + v_head_dim ...................................... 128 + 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 ...................................... /workspace/data/wikitext/gpt2-vocab.json + vocab_size ...................................... None + wandb_exp_name .................................. + wandb_project ................................... + wandb_save_dir .................................. + weight_decay .................................... 0.1 + weight_decay_incr_style ......................... constant + wgrad_deferral_limit ............................ 0 + world_size ...................................... 8 + yaml_cfg ........................................ None +-------------------- end of arguments --------------------- +INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1 +> building GPT2BPETokenizer tokenizer ... +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +WARNING: one_logger package is required to enable e2e metrics tracking. please go to https://confluence.nvidia.com/display/MLWFO/Package+Repositories for details to install it + > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304) +WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS +> initializing torch distributed ... +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +[Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to [Gloo] Rank 0 peer ranks. 0Expected number of connected peer ranks is : is connected to 00 + peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +> initialized tensor model parallel with size 1 +> initialized pipeline model parallel with size 1 +> setting random seeds to 1234 ... +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[rank4]:[W826 20:13:16.599732262 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 4] using GPU 4 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank6]:[W826 20:13:16.605387205 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 6] using GPU 6 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank7]:[W826 20:13:16.605496515 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank5]:[W826 20:13:16.605559966 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 5] using GPU 5 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank1]:[W826 20:13:16.605692935 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank3]:[W826 20:13:16.605720898 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank2]:[W826 20:13:16.605757600 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 2] using GPU 2 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +> compiling dataset index builder ... +make: Entering directory '/workspace/Megatron-UltraEP/megatron/core/datasets' +make: Nothing to be done for 'default'. +make: Leaving directory '/workspace/Megatron-UltraEP/megatron/core/datasets' +>>> done with dataset index builder. Compilation time: 0.429 seconds +WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations. +> compiling and loading fused kernels ... +[rank0]:[W826 20:13:17.210198673 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 0] using GPU 0 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +>>> done with compiling and loading fused kernels. Compilation time: 3.956 seconds +time to initialize megatron (seconds): 7.611 +[after megatron is initialized] datetime: 2026-08-26 20:13:23 +building GPT model ... +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +[INFO] Use MNNVL fabric: False +[WARN] MAX_NUM_NVL_PEERS is not set. Using detected value 8. +NVSHMEM v3.7.2 +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +UltraEP: NVL domain size <= 8; forcing weight_sync plan mode to direct (was adaptive). + > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 7262624256 +INFO:megatron.core.distributed.distributed_data_parallel:Setting up DistributedDataParallel with config DistributedDataParallelConfig(grad_reduce_in_fp32=True, overlap_grad_reduce=True, overlap_param_gather=True, align_param_gather=False, use_distributed_optimizer=True, num_distributed_optimizer_instances=1, check_for_nan_in_grad=False, check_for_large_grads=False, bucket_size=40000000, pad_buckets_for_high_nccl_busbw=False, average_in_collective=False, fp8_param_gather=False, reuse_grad_buf_for_mxfp8_param_ag=False, use_custom_fsdp=False, data_parallel_sharding_strategy='no_shard', gradient_reduce_div_fusion=True, suggested_communication_unit_size=None, preserve_fp32_weights=True, keep_fp8_transpose_cache_when_using_custom_fsdp=False, nccl_ub=False, fsdp_double_buffer=False) +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 22 +Params for bucket 1 (180289536 elements, 180289536 padded size): + module.output_layer.weight +Params for bucket 2 (73403904 elements, 73403904 padded size): + module.decoder.final_layernorm.weight + module.decoder.layers.9.mlp.shared_experts.linear_fc2.weight +Params for bucket 3 (146800640 elements, 146800640 padded size): + module.decoder.layers.9.mlp.shared_experts.linear_fc1.weight +Params for bucket 4 (103001600 elements, 103001600 padded size): + module.decoder.layers.9.pre_mlp_layernorm.weight + module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.9.self_attention.linear_qkv.bias + module.decoder.layers.9.mlp.router.weight + module.decoder.layers.9.self_attention.linear_qkv.weight + module.decoder.layers.9.self_attention.linear_proj.weight + module.decoder.layers.8.mlp.shared_experts.linear_fc2.weight +Params for bucket 5 (146800640 elements, 146800640 padded size): + module.decoder.layers.8.mlp.shared_experts.linear_fc1.weight +Params for bucket 6 (103001600 elements, 103001600 padded size): + module.decoder.layers.8.mlp.router.weight + module.decoder.layers.8.self_attention.linear_qkv.weight + module.decoder.layers.8.self_attention.linear_proj.weight + module.decoder.layers.7.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.8.pre_mlp_layernorm.weight + module.decoder.layers.8.self_attention.linear_qkv.bias +Params for bucket 7 (146800640 elements, 146800640 padded size): + module.decoder.layers.7.mlp.shared_experts.linear_fc1.weight +Params for bucket 8 (103001600 elements, 103001600 padded size): + module.decoder.layers.7.pre_mlp_layernorm.weight + module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.7.self_attention.linear_qkv.bias + module.decoder.layers.7.mlp.router.weight + module.decoder.layers.7.self_attention.linear_qkv.weight + module.decoder.layers.7.self_attention.linear_proj.weight + module.decoder.layers.6.mlp.shared_experts.linear_fc2.weight +Params for bucket 9 (146800640 elements, 146800640 padded size): + module.decoder.layers.6.mlp.shared_experts.linear_fc1.weight +Params for bucket 10 (103001600 elements, 103001600 padded size): + module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.6.pre_mlp_layernorm.weight + module.decoder.layers.6.self_attention.linear_qkv.bias + module.decoder.layers.6.mlp.router.weight + module.decoder.layers.6.self_attention.linear_qkv.weight + module.decoder.layers.6.self_attention.linear_proj.weight + module.decoder.layers.5.mlp.shared_experts.linear_fc2.weight +Params for bucket 11 (146800640 elements, 146800640 padded size): + module.decoder.layers.5.mlp.shared_experts.linear_fc1.weight +Params for bucket 12 (103001600 elements, 103001600 padded size): + module.decoder.layers.5.pre_mlp_layernorm.weight + module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.5.self_attention.linear_qkv.bias + module.decoder.layers.5.mlp.router.weight + module.decoder.layers.5.self_attention.linear_qkv.weight + module.decoder.layers.5.self_attention.linear_proj.weight + module.decoder.layers.4.mlp.shared_experts.linear_fc2.weight +Params for bucket 13 (146800640 elements, 146800640 padded size): + module.decoder.layers.4.mlp.shared_experts.linear_fc1.weight +Params for bucket 14 (103001600 elements, 103001600 padded size): + module.decoder.layers.4.mlp.router.weight + module.decoder.layers.4.self_attention.linear_qkv.weight + module.decoder.layers.4.self_attention.linear_proj.weight + module.decoder.layers.3.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.4.self_attention.linear_qkv.bias + module.decoder.layers.4.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.4.pre_mlp_layernorm.weight +Params for bucket 15 (146800640 elements, 146800640 padded size): + module.decoder.layers.3.mlp.shared_experts.linear_fc1.weight +Params for bucket 16 (103001600 elements, 103001600 padded size): + module.decoder.layers.3.mlp.router.weight + module.decoder.layers.3.self_attention.linear_qkv.weight + module.decoder.layers.3.self_attention.linear_proj.weight + module.decoder.layers.2.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.3.self_attention.linear_qkv.bias + module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.3.pre_mlp_layernorm.weight +Params for bucket 17 (146800640 elements, 146800640 padded size): + module.decoder.layers.2.mlp.shared_experts.linear_fc1.weight +Params for bucket 18 (103001600 elements, 103001600 padded size): + module.decoder.layers.2.mlp.router.weight + module.decoder.layers.2.self_attention.linear_qkv.weight + module.decoder.layers.2.self_attention.linear_proj.weight + module.decoder.layers.1.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.2.pre_mlp_layernorm.weight + module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.2.self_attention.linear_qkv.bias +Params for bucket 19 (146800640 elements, 146800640 padded size): + module.decoder.layers.1.mlp.shared_experts.linear_fc1.weight +Params for bucket 20 (103001600 elements, 103001600 padded size): + module.decoder.layers.1.mlp.router.weight + module.decoder.layers.1.self_attention.linear_qkv.weight + module.decoder.layers.1.self_attention.linear_proj.weight + module.decoder.layers.0.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.1.pre_mlp_layernorm.weight + module.decoder.layers.1.self_attention.linear_qkv.bias +Params for bucket 21 (146800640 elements, 146800640 padded size): + module.decoder.layers.0.mlp.shared_experts.linear_fc1.weight +Params for bucket 22 (209890816 elements, 209890816 padded size): + module.decoder.layers.0.mlp.router.weight + module.decoder.layers.0.self_attention.linear_qkv.bias + module.embedding.word_embeddings.weight + module.decoder.layers.0.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.0.self_attention.linear_qkv.weight + module.decoder.layers.0.pre_mlp_layernorm.weight + module.decoder.layers.0.self_attention.linear_proj.weight +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 44 +Params for bucket 1 (45875200 elements, 45875200 padded size): + module.decoder.layers.9.mlp.experts.linear_fc2.weight4 + module.decoder.layers.9.mlp.experts.linear_fc2.weight6 + module.decoder.layers.9.mlp.experts.linear_fc2.weight7 + module.decoder.layers.9.mlp.experts.linear_fc2.weight3 + module.decoder.layers.9.mlp.experts.linear_fc2.weight5 +Params for bucket 2 (45875200 elements, 45875200 padded size): + module.decoder.layers.9.mlp.experts.linear_fc2.weight2 + module.decoder.layers.9.mlp.experts.linear_fc2.weight1 + module.decoder.layers.9.mlp.experts.linear_fc1.weight7 + module.decoder.layers.9.mlp.experts.linear_fc2.weight0 +Params for bucket 3 (55050240 elements, 55050240 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight6 + module.decoder.layers.9.mlp.experts.linear_fc1.weight4 + module.decoder.layers.9.mlp.experts.linear_fc1.weight5 +Params for bucket 4 (55050240 elements, 55050240 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight3 + module.decoder.layers.9.mlp.experts.linear_fc1.weight1 + module.decoder.layers.9.mlp.experts.linear_fc1.weight2 +Params for bucket 5 (45875200 elements, 45875200 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight0 + module.decoder.layers.8.mlp.experts.linear_fc2.weight6 + module.decoder.layers.8.mlp.experts.linear_fc2.weight7 + module.decoder.layers.8.mlp.experts.linear_fc2.weight5 +Params for bucket 6 (45875200 elements, 45875200 padded size): + module.decoder.layers.8.mlp.experts.linear_fc2.weight4 + module.decoder.layers.8.mlp.experts.linear_fc2.weight0 + module.decoder.layers.8.mlp.experts.linear_fc2.weight2 + module.decoder.layers.8.mlp.experts.linear_fc2.weight3 + module.decoder.layers.8.mlp.experts.linear_fc2.weight1 +Params for bucket 7 (55050240 elements, 55050240 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight7 + module.decoder.layers.8.mlp.experts.linear_fc1.weight5 + module.decoder.layers.8.mlp.experts.linear_fc1.weight6 +Params for bucket 8 (55050240 elements, 55050240 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight4 + module.decoder.layers.8.mlp.experts.linear_fc1.weight2 + module.decoder.layers.8.mlp.experts.linear_fc1.weight3 +Params for bucket 9 (45875200 elements, 45875200 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight1 + module.decoder.layers.7.mlp.experts.linear_fc2.weight7 + module.decoder.layers.8.mlp.experts.linear_fc1.weight0 +Params for bucket 10 (45875200 elements, 45875200 padded size): + module.decoder.layers.7.mlp.experts.linear_fc2.weight5 + module.decoder.layers.7.mlp.experts.linear_fc2.weight3 + module.decoder.layers.7.mlp.experts.linear_fc2.weight6 + module.decoder.layers.7.mlp.experts.linear_fc2.weight4 + module.decoder.layers.7.mlp.experts.linear_fc2.weight2 +Params for bucket 11 (55050240 elements, 55050240 padded size): + module.decoder.layers.7.mlp.experts.linear_fc2.weight1 + module.decoder.layers.7.mlp.experts.linear_fc1.weight6 + module.decoder.layers.7.mlp.experts.linear_fc1.weight7 + module.decoder.layers.7.mlp.experts.linear_fc2.weight0 +Params for bucket 12 (55050240 elements, 55050240 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight5 + module.decoder.layers.7.mlp.experts.linear_fc1.weight3 + module.decoder.layers.7.mlp.experts.linear_fc1.weight4 +Params for bucket 13 (55050240 elements, 55050240 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight2 + module.decoder.layers.7.mlp.experts.linear_fc1.weight0 + module.decoder.layers.7.mlp.experts.linear_fc1.weight1 +Params for bucket 14 (45875200 elements, 45875200 padded size): + module.decoder.layers.6.mlp.experts.linear_fc2.weight4 + module.decoder.layers.6.mlp.experts.linear_fc2.weight6 + module.decoder.layers.6.mlp.experts.linear_fc2.weight7 + module.decoder.layers.6.mlp.experts.linear_fc2.weight5 + module.decoder.layers.6.mlp.experts.linear_fc2.weight3 +Params for bucket 15 (45875200 elements, 45875200 padded size): + module.decoder.layers.6.mlp.experts.linear_fc2.weight2 + module.decoder.layers.6.mlp.experts.linear_fc2.weight1 + module.decoder.layers.6.mlp.experts.linear_fc1.weight7 + module.decoder.layers.6.mlp.experts.linear_fc2.weight0 +Params for bucket 16 (55050240 elements, 55050240 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight6 + module.decoder.layers.6.mlp.experts.linear_fc1.weight4 + module.decoder.layers.6.mlp.experts.linear_fc1.weight5 +Params for bucket 17 (55050240 elements, 55050240 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight3 + module.decoder.layers.6.mlp.experts.linear_fc1.weight1 + module.decoder.layers.6.mlp.experts.linear_fc1.weight2 +Params for bucket 18 (45875200 elements, 45875200 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight0 + module.decoder.layers.5.mlp.experts.linear_fc2.weight6 + module.decoder.layers.5.mlp.experts.linear_fc2.weight7 + module.decoder.layers.5.mlp.experts.linear_fc2.weight5 +Params for bucket 19 (45875200 elements, 45875200 padded size): + module.decoder.layers.5.mlp.experts.linear_fc2.weight4 + module.decoder.layers.5.mlp.experts.linear_fc2.weight0 + module.decoder.layers.5.mlp.experts.linear_fc2.weight2 + module.decoder.layers.5.mlp.experts.linear_fc2.weight3 + module.decoder.layers.5.mlp.experts.linear_fc2.weight1 +Params for bucket 20 (55050240 elements, 55050240 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight7 + module.decoder.layers.5.mlp.experts.linear_fc1.weight6 + module.decoder.layers.5.mlp.experts.linear_fc1.weight5 +Params for bucket 21 (55050240 elements, 55050240 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight4 + module.decoder.layers.5.mlp.experts.linear_fc1.weight3 + module.decoder.layers.5.mlp.experts.linear_fc1.weight2 +Params for bucket 22 (45875200 elements, 45875200 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight1 + module.decoder.layers.5.mlp.experts.linear_fc1.weight0 + module.decoder.layers.4.mlp.experts.linear_fc2.weight7 +Params for bucket 23 (45875200 elements, 45875200 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight3 + module.decoder.layers.4.mlp.experts.linear_fc2.weight5 + module.decoder.layers.4.mlp.experts.linear_fc2.weight6 + module.decoder.layers.4.mlp.experts.linear_fc2.weight4 + module.decoder.layers.4.mlp.experts.linear_fc2.weight2 +Params for bucket 24 (55050240 elements, 55050240 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight1 + module.decoder.layers.4.mlp.experts.linear_fc2.weight0 + module.decoder.layers.4.mlp.experts.linear_fc1.weight7 + module.decoder.layers.4.mlp.experts.linear_fc1.weight6 +Params for bucket 25 (55050240 elements, 55050240 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight5 + module.decoder.layers.4.mlp.experts.linear_fc1.weight4 + module.decoder.layers.4.mlp.experts.linear_fc1.weight3 +Params for bucket 26 (55050240 elements, 55050240 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight2 + module.decoder.layers.4.mlp.experts.linear_fc1.weight0 + module.decoder.layers.4.mlp.experts.linear_fc1.weight1 +Params for bucket 27 (45875200 elements, 45875200 padded size): + module.decoder.layers.3.mlp.experts.linear_fc2.weight6 + module.decoder.layers.3.mlp.experts.linear_fc2.weight4 + module.decoder.layers.3.mlp.experts.linear_fc2.weight7 + module.decoder.layers.3.mlp.experts.linear_fc2.weight5 + module.decoder.layers.3.mlp.experts.linear_fc2.weight3 +Params for bucket 28 (45875200 elements, 45875200 padded size): + module.decoder.layers.3.mlp.experts.linear_fc2.weight2 + module.decoder.layers.3.mlp.experts.linear_fc1.weight7 + module.decoder.layers.3.mlp.experts.linear_fc2.weight0 + module.decoder.layers.3.mlp.experts.linear_fc2.weight1 +Params for bucket 29 (55050240 elements, 55050240 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight6 + module.decoder.layers.3.mlp.experts.linear_fc1.weight4 + module.decoder.layers.3.mlp.experts.linear_fc1.weight5 +Params for bucket 30 (55050240 elements, 55050240 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight3 + module.decoder.layers.3.mlp.experts.linear_fc1.weight1 + module.decoder.layers.3.mlp.experts.linear_fc1.weight2 +Params for bucket 31 (45875200 elements, 45875200 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight0 + module.decoder.layers.2.mlp.experts.linear_fc2.weight6 + module.decoder.layers.2.mlp.experts.linear_fc2.weight7 + module.decoder.layers.2.mlp.experts.linear_fc2.weight5 +Params for bucket 32 (45875200 elements, 45875200 padded size): + module.decoder.layers.2.mlp.experts.linear_fc2.weight3 + module.decoder.layers.2.mlp.experts.linear_fc2.weight1 + module.decoder.layers.2.mlp.experts.linear_fc2.weight4 + module.decoder.layers.2.mlp.experts.linear_fc2.weight0 + module.decoder.layers.2.mlp.experts.linear_fc2.weight2 +Params for bucket 33 (55050240 elements, 55050240 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight7 + module.decoder.layers.2.mlp.experts.linear_fc1.weight6 + module.decoder.layers.2.mlp.experts.linear_fc1.weight5 +Params for bucket 34 (55050240 elements, 55050240 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight3 + module.decoder.layers.2.mlp.experts.linear_fc1.weight2 +Params for bucket 35 (45875200 elements, 45875200 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight1 + module.decoder.layers.1.mlp.experts.linear_fc2.weight7 + module.decoder.layers.2.mlp.experts.linear_fc1.weight0 +Params for bucket 36 (45875200 elements, 45875200 padded size): + module.decoder.layers.1.mlp.experts.linear_fc2.weight5 + module.decoder.layers.1.mlp.experts.linear_fc2.weight3 + module.decoder.layers.1.mlp.experts.linear_fc2.weight6 + module.decoder.layers.1.mlp.experts.linear_fc2.weight4 + module.decoder.layers.1.mlp.experts.linear_fc2.weight2 +Params for bucket 37 (55050240 elements, 55050240 padded size): + module.decoder.layers.1.mlp.experts.linear_fc2.weight1 + module.decoder.layers.1.mlp.experts.linear_fc1.weight7 + module.decoder.layers.1.mlp.experts.linear_fc1.weight6 + module.decoder.layers.1.mlp.experts.linear_fc2.weight0 +Params for bucket 38 (55050240 elements, 55050240 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight5 + module.decoder.layers.1.mlp.experts.linear_fc1.weight4 + module.decoder.layers.1.mlp.experts.linear_fc1.weight3 +Params for bucket 39 (55050240 elements, 55050240 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight0 + module.decoder.layers.1.mlp.experts.linear_fc1.weight1 +Params for bucket 40 (45875200 elements, 45875200 padded size): + module.decoder.layers.0.mlp.experts.linear_fc2.weight6 + module.decoder.layers.0.mlp.experts.linear_fc2.weight4 + module.decoder.layers.0.mlp.experts.linear_fc2.weight7 + module.decoder.layers.0.mlp.experts.linear_fc2.weight5 + module.decoder.layers.0.mlp.experts.linear_fc2.weight3 +Params for bucket 41 (45875200 elements, 45875200 padded size): + module.decoder.layers.0.mlp.experts.linear_fc2.weight2 + module.decoder.layers.0.mlp.experts.linear_fc1.weight7 + module.decoder.layers.0.mlp.experts.linear_fc2.weight0 + module.decoder.layers.0.mlp.experts.linear_fc2.weight1 +Params for bucket 42 (55050240 elements, 55050240 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight6 + module.decoder.layers.0.mlp.experts.linear_fc1.weight4 + module.decoder.layers.0.mlp.experts.linear_fc1.weight5 +Params for bucket 43 (55050240 elements, 55050240 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight3 + module.decoder.layers.0.mlp.experts.linear_fc1.weight1 + module.decoder.layers.0.mlp.experts.linear_fc1.weight2 +Params for bucket 44 (18350080 elements, 18350080 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight0 +INFO:megatron.core.optimizer:Setting up optimizer with config OptimizerConfig(optimizer='adam', lr=0.0001, min_lr=1e-05, decoupled_lr=None, decoupled_min_lr=None, weight_decay=0.1, fp8_recipe='mxfp8', fp16=False, bf16=True, reuse_grad_buf_for_mxfp8_param_ag=False, params_dtype=torch.bfloat16, use_precision_aware_optimizer=False, store_param_remainders=True, main_grads_dtype=torch.float32, main_params_dtype=torch.float32, exp_avg_dtype=torch.float32, exp_avg_sq_dtype=torch.float32, 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_param_gather_with_optimizer_step=False, optimizer_cpu_offload=False, optimizer_offload_fraction=1.0, use_torch_optimizer_for_cpu_offload=False, overlap_cpu_optimizer_d2h_h2d=False, pin_cpu_grads=True, pin_cpu_params=True, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=, config_logger_dir='') +INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine +[after model, optimizer, and learning rate scheduler are built] datetime: 2026-08-26 20:13:27 +> building train, validation, and test datasets ... + > datasets target sizes (minimum size): + train: 80 + validation: 0 + test: 0 +INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None] +> building train, validation, and test datasets for GPT ... +INFO:megatron.core.datasets.blended_megatron_dataset_builder:Building GPTDataset splits with sizes=(80, 0, 0) and config=GPTDatasetConfig(random_seed=1234, sequence_length=4096, blend=(['/workspace/data/wikitext/wikitext103_gpt2_text_document'], None), blend_per_split=None, split='100,0,0', split_matrix=[(0, 1.0), None, None], num_dataset_builder_threads=1, path_to_cache=None, mmap_bin_files=True, mock=False, tokenizer=, mid_level_dataset_surplus=0.005, reset_position_ids=False, reset_attention_mask=False, eod_mask_loss=False, create_attention_mask=True, drop_last_partial_validation_sequence=True, add_extra_token_to_sequence=True, object_storage_cache_path=None) +INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_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: 1165029 +INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350 +INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices +INFO:megatron.core.datasets.gpt_dataset: Load the document index from 71e99bc7c46bc3412b339a272e0f9cf3-GPTDataset-train-document_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 71e99bc7c46bc3412b339a272e0f9cf3-GPTDataset-train-sample_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 71e99bc7c46bc3412b339a272e0f9cf3-GPTDataset-train-shuffle_index.npy +INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073 +> finished creating GPT datasets ... +[after dataloaders are built] datetime: 2026-08-26 20:13:27 +done with setup ... +(min, max) time across ranks (ms): + model-and-optimizer-setup ......................: (3676.43, 3687.11) + train/valid/test-data-iterators-setup ..........: (98.18, 123.36) +training ... +Overwriting rerun_state_machine.current_iteration from -1 to 0... +[before the start of training step] datetime: 2026-08-26 20:13:27 +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +Number of parameters in transformer block in billions: 20.11 + [2026-08-26 20:13:56] iteration 1/ 10 | consumed samples: 8 | elapsed time per iteration (ms): 29655.7 | throughput per GPU (TFLOP/s/GPU): 4.2 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.097922E+01 | loss scale: 1.0 | grad norm: 29.875 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33664 | min safe capacity: 1.0273 |Number of parameters in embedding layers in billions: 0.36 + +Total number of parameters in billions: 20.47 +Number of parameters in most loaded shard in billions: 20.4724 +Theoretical memory footprints: weight and optimizer=146430.14 MB +[Rank 0] (after 1 iterations) memory (MB) | allocated: 64278.42138671875 | max allocated: 65496.45947265625 | reserved: 73306.0 | max reserved: 73306.0 +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") + [2026-08-26 20:13:57] iteration 2/ 10 | consumed samples: 16 | elapsed time per iteration (ms): 301.2 | throughput per GPU (TFLOP/s/GPU): 410.0 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.097874E+01 | loss scale: 1.0 | grad norm: 30.625 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33792 | min safe capacity: 1.0312 | + [2026-08-26 20:13:57] iteration 3/ 10 | consumed samples: 24 | elapsed time per iteration (ms): 216.2 | throughput per GPU (TFLOP/s/GPU): 571.2 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 9.539545E+00 | loss scale: 1.0 | grad norm: 32.426 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33792 | min safe capacity: 1.0312 | +Capture CUDA graph for training!!! +CUDA graph capture done!!! + [2026-08-26 20:13:59] iteration 4/ 10 | consumed samples: 32 | elapsed time per iteration (ms): 1714.3 | throughput per GPU (TFLOP/s/GPU): 72.0 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 1.108675E+01 | loss scale: 1.0 | grad norm: 106.907 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33920 | min safe capacity: 1.0352 | + [2026-08-26 20:13:59] iteration 5/ 10 | consumed samples: 40 | elapsed time per iteration (ms): 198.8 | throughput per GPU (TFLOP/s/GPU): 621.3 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.001003E+01 | loss scale: 1.0 | grad norm: 26.914 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34048 | min safe capacity: 1.0391 | + [2026-08-26 20:13:59] iteration 6/ 10 | consumed samples: 48 | elapsed time per iteration (ms): 192.1 | throughput per GPU (TFLOP/s/GPU): 642.8 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 1.050245E+01 | loss scale: 1.0 | grad norm: 8.738 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34048 | min safe capacity: 1.0391 | + [2026-08-26 20:14:02] iteration 7/ 10 | consumed samples: 56 | elapsed time per iteration (ms): 2933.2 | throughput per GPU (TFLOP/s/GPU): 42.1 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 9.451855E+00 | loss scale: 1.0 | grad norm: 5.610 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34048 | min safe capacity: 1.0391 | + [2026-08-26 20:14:02] iteration 8/ 10 | consumed samples: 64 | elapsed time per iteration (ms): 202.4 | throughput per GPU (TFLOP/s/GPU): 610.1 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.983458E+00 | loss scale: 1.0 | grad norm: 4.580 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34048 | min safe capacity: 1.0391 | diff --git a/ultraep/tierbtrace/tierbtrace_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/nsys/train.log b/ultraep/tierbtrace/tierbtrace_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/nsys/train.log index 4970aeda05fbba965e8d3f63c106c93b6dc9bcf9..f789b003c23cc4475f3db98faa39a697f9807ff6 100644 --- a/ultraep/tierbtrace/tierbtrace_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/nsys/train.log +++ b/ultraep/tierbtrace/tierbtrace_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/nsys/train.log @@ -2836,3 +2836,47 @@ HYBRID-EP ALLGATHER TIMEOUT:SM 9 [0]:expecting 3448 got 3445 HYBRID-EP ALLGATHER TIMEOUT:SM 18 [0]:expecting 3448 got 3445 HYBRID-EP ALLGATHER TIMEOUT:SM 25 [0]:expecting 3448 got 3445 HYBRID-EP ALLGATHER TIMEOUT:SM 22 [0]:expecting 3448 got 3445 +Collection stop is not allowed in this state. +W0826 20:10:03.249000 739270 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers +W0826 20:10:03.250000 739270 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 739385 closing signal SIGTERM +Generated: + /workspace/tierb_study_20260826_144644/ultraep/tierbtrace/tierbtrace_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/nsys/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125.nsys-rep +W0826 20:10:03.258000 739270 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 739386 closing signal SIGTERM +W0826 20:10:03.274000 739270 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 739387 closing signal SIGTERM +W0826 20:10:03.279000 739270 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 739388 closing signal SIGTERM +W0826 20:10:03.287000 739270 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 739389 closing signal SIGTERM +W0826 20:10:03.300000 739270 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 739390 closing signal SIGTERM +W0826 20:10:03.319000 739270 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 739391 closing signal SIGTERM +W0826 20:10:03.351000 739270 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 739392 closing signal SIGTERM +Traceback (most recent call last): + File "/usr/local/bin/torchrun", line 8, in + sys.exit(main()) + ^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper + return f(*args, **kwargs) + ^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 892, in main + run(args) + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 883, in run + elastic_launch( + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__ + return launch_agent(self._config, self._entrypoint, list(args)) + ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 261, in launch_agent + result = agent.run() + ^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper + result = f(*args, **kwargs) + ^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 711, in run + result = self._invoke_run(role) + ^^^^^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 870, in _invoke_run + time.sleep(monitor_interval) + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 84, in _terminate_process_handler + raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval) +torch.distributed.elastic.multiprocessing.api.SignalException: Process 739270 got signal: 15 +The target application terminated. One or more process it created re-parented. +Waiting for termination of re-parented processes. +Use the `--wait` option to modify this behavior. +There are no active sessions. diff --git a/ultraep/tierbtrace/tierbtrace_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/nsys/verify_trace_ranks.log b/ultraep/tierbtrace/tierbtrace_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/nsys/verify_trace_ranks.log new file mode 100644 index 0000000000000000000000000000000000000000..bb1d389a384e9d0ff6aca8facf2be2063fc9c87c --- /dev/null +++ b/ultraep/tierbtrace/tierbtrace_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/nsys/verify_trace_ranks.log @@ -0,0 +1,5 @@ +report : /workspace/tierb_study_20260826_144644/ultraep/tierbtrace/tierbtrace_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/nsys/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125.nsys-rep +kernels : 253335 +devices : 8 of 8 -> [0, 1, 2, 3, 4, 5, 6, 7] +processes: 1 +PASS: all ranks present diff --git a/ultraep/tierbtrace/tierbtrace_gbs8/trace_results.tsv b/ultraep/tierbtrace/tierbtrace_gbs8/trace_results.tsv index a6ea12f8a28471420b9d84b135b62b69839a5c76..5a1e135fc6b92345a7a44a2b55cb245e39f09c7a 100644 --- a/ultraep/tierbtrace/tierbtrace_gbs8/trace_results.tsv +++ b/ultraep/tierbtrace/tierbtrace_gbs8/trace_results.tsv @@ -1,2 +1,4 @@ model backend capacity status report QWEN3_235B ultraep 1.083984375 OK /workspace/tierb_study_20260826_144644/ultraep/tierbtrace/tierbtrace_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/nsys/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375.nsys-rep +QWEN3_30B ultraep 1.11328125 OK /workspace/tierb_study_20260826_144644/ultraep/tierbtrace/tierbtrace_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/nsys/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125.nsys-rep +DSV2_LITE_L24 ultraep 1.0755208333333333 OK /workspace/tierb_study_20260826_144644/ultraep/tierbtrace/tierbtrace_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/nsys/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333.nsys-rep