diff --git a/.gitattributes b/.gitattributes index f69f1f446b67de79d4f1243822fa94f15bfc2626..02f2d7bca451c8723b290668ccebba44ee7be098 100644 --- a/.gitattributes +++ b/.gitattributes @@ -91,3 +91,6 @@ granularity_next_b200/tiera_topk_trace/echo/tierbtrace/tierbtrace_gbs16/G4_K6__e granularity_next_b200/tiera_topk_trace/echo/tierbtrace/tierbtrace_gbs16/G4_K8__echo__matched__seq4096__gbs16__cap1p109375/nsys/G4_K8__echo__matched__seq4096__gbs16__cap1p109375.nsys-rep filter=lfs diff=lfs merge=lfs -text granularity_next_b200/tiera_topk_trace/echo/tierbtrace/tierbtrace_gbs16/G4_K10__echo__matched__seq4096__gbs16__cap1p1/nsys/G4_K10__echo__matched__seq4096__gbs16__cap1p1.nsys-rep filter=lfs diff=lfs merge=lfs -text granularity_next_b200/tiera_topk_trace/echo/tierbtrace/tierbtrace_gbs16/G4_K16__echo__matched__seq4096__gbs16__cap1p08203125/nsys/G4_K16__echo__matched__seq4096__gbs16__cap1p08203125.nsys-rep filter=lfs diff=lfs merge=lfs -text +granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep filter=lfs diff=lfs merge=lfs -text +granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep filter=lfs diff=lfs merge=lfs -text +granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep filter=lfs diff=lfs merge=lfs -text diff --git a/granularity_next_b200/tiera_topk_trace/hf_watcher.log b/granularity_next_b200/tiera_topk_trace/hf_watcher.log index 3fe9e933766d7a80271655b3260b67c46248b9bb..7cee7cfb3d380e534ad2ed8c248810a7baf8c384 100644 --- a/granularity_next_b200/tiera_topk_trace/hf_watcher.log +++ b/granularity_next_b200/tiera_topk_trace/hf_watcher.log @@ -28,3 +28,6 @@ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main uploaded: /workspace/granularity_next_b200/tiera_topk_trace -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tiera_topk_trace https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main [2026-09-17T03:25:01+00:00] backup pass + uploaded: /workspace/granularity_next_b200/tiera_topk_trace -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tiera_topk_trace +https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main +[2026-09-17T03:35:04+00:00] backup pass diff --git a/granularity_next_b200/tiera_topk_trace/study.log b/granularity_next_b200/tiera_topk_trace/study.log index c6cde9f3817d3feea21b86fac7554187f8a93330..176645b53bd69e36cf90ab0929e37aa1078f7bc1 100644 --- a/granularity_next_b200/tiera_topk_trace/study.log +++ b/granularity_next_b200/tiera_topk_trace/study.log @@ -104,3 +104,19 @@ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main results -> /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace ============================================================== [03:22:42] building ultraep marker extension (log: /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/trace_extension_build.log) +[03:26:51] marker extension ready: _ultraep_trace_C.cpython-312-x86_64-linux-gnu.so +[03:26:51] TRACE-ONLY G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625 (ranks: 0 1 2 3 4 5 6 7) + [guard] iter 8/10 (70s) + [guard] no progress for 60s at iter 8/10 (past profile window); treating as done + [guard] nsys stop (flushing report) + [guard] tearing down process group (outcome=STALLED) +[03:29:49] trace verified -> G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep (11M, all 8 GPUs + ultraep markers) +[03:29:49] TRACE-ONLY G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875 (ranks: 0 1 2 3 4 5 6 7) + [guard] iter 7/10 (70s) + [guard] iter 8/10 (80s) + [guard] no progress for 60s at iter 8/10 (past profile window); treating as done + [guard] nsys stop (flushing report) + [guard] tearing down process group (outcome=STALLED) +[03:32:56] trace verified -> G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep (12M, all 8 GPUs + ultraep markers) +[03:32:56] TRACE-ONLY G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375 (ranks: 0 1 2 3 4 5 6 7) + [guard] iter 8/10 (70s) diff --git a/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep new file mode 100644 index 0000000000000000000000000000000000000000..e03d5d6bd01bc9e606541ceb9b7ed989f0daf53f --- /dev/null +++ b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16d8b184e95b4cc82e363b6b774e38586cdc69ef72bb35e09082e92a2190d7f6 +size 10833175 diff --git a/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/run.env b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/run.env new file mode 100644 index 0000000000000000000000000000000000000000..436b2c3807196bd7f66912af65b23bdc58c753d1 --- /dev/null +++ b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/run.env @@ -0,0 +1,102 @@ +backend=ultraep +config=G2_K2 +family=tierA +source_model= +standard_num_layers= +base_rung=G2 +repo=/workspace/Megatron-UltraEP +branch=ultraep-bracing-regions-through-kernels +commit=03d57ef7 +num_layers=16 +hidden=4096 +ffn_hidden=14336 +moe_ffn_hidden=3584 +num_experts=32 +topk=2 +experts_per_rank=4 +expert_params=44040192 +heads=32 +kv_groups=8 +kv_channels= +qk_layernorm=0 +shared_ffn=0 +shared_expert_overlap=0 +moe_token_dispatcher=flex +mla=0 +q_lora_rank= +kv_lora_rank=512 +qk_head_dim=128 +qk_pos_emb_head_dim=64 +v_head_dim=128 +dense_first_layers=0 +moe_layers=16 +rotary_base=1000000 +rotary_percent=1.0 +rotary_scaling= +mscale= +mscale_all_dim= +qkv_bias=0 +router_balancing=aux_loss +aux_loss_coeff=1e-2 +router_pre_softmax=0 +router_fusion=0 +router_dtype=fp32 +router_groups= +router_group_topk= +router_score_function=softmax +router_expert_bias=0 +router_bias_update_rate= +router_topk_scaling_factor= +shared_expert_gate=0 +apply_layernorm_1p=0 +attention_output_gate=0 +linear_attention_type= +linear_attention_freq= +linear_conv_kernel_dim= +linear_key_head_dim= +linear_value_head_dim= +linear_num_key_heads= +linear_num_value_heads= +mtp_num_layers=0 +mtp_loss_scaling_factor=0.1 +no_weight_decay_cond_type= +max_position_embeddings=32768 +init_method_std=0.01 +device_initiated_grouped_gemm=1 +norm_epsilon=1e-5 +seq_length=4096 +micro_batch_size=1 +global_batch_size=16 +num_microbatches=2 +assignments_per_rank=8192 +mean_tokens_per_expert=2048 +slot_policy=matched +slots_per_rank=4 +capacity_mode=static +capacity=1.15625 +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=256 +ultraep_weight_sync_join= +profile_enable=1 +kernel_trace=1 +fp8_mode=mxfp8 +fp8_recipe=mxfp8 +training_seed=1234 +learning_rate=1e-4 +min_learning_rate=1.0e-5 +lr_warmup_iters=5 +state_dump_enable=0 +state_dump_dir= +moonep_assert_no_pending_reduce=0 diff --git a/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/train.log b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/train.log new file mode 100644 index 0000000000000000000000000000000000000000..0fa822119d28849c35d1f9d7f338653beec69b57 --- /dev/null +++ b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/train.log @@ -0,0 +1,1176 @@ +============================================================== + G2_K2 / ultraep (slots/rank=4 policy=matched) + E=32 K=2 H'=3584 m=4 + L=16 seq=4096 gbs=16 (2 microbatches) + cap_mode=static cap=1.15625 cg=1 recompute=none + memory_dump=none snapshots=/workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/memory_snapshots/snapshot.pickle + repo=/workspace/Megatron-UltraEP @ 03d57ef7 +============================================================== +W0917 03:26:59.698000 948326 torch/distributed/run.py:766] +W0917 03:26:59.698000 948326 torch/distributed/run.py:766] ***************************************** +W0917 03:26:59.698000 948326 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. +W0917 03:26:59.698000 948326 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 .................................... 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 ............................... 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 .............................. 16 + 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 ................................. 14336 + 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 ............................... 16 + 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 ..................................... 4096 + 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 ............................. 3584 + 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 .............. 4 + moe_pad_expert_input_to_capacity ................ False + moe_per_layer_logging ........................... False + moe_permute_fusion .............................. False + moe_received_token_capacity ..................... 1.15625 + 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 .......................... False + moe_router_score_function ....................... softmax + moe_router_topk ................................. 2 + moe_router_topk_scaling_factor .................. None + moe_shared_expert_intermediate_size ............. None + 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-05 + normalization ................................... RMSNorm + num_attention_heads ............................. 32 + num_channels .................................... 3 + num_classes ..................................... 1000 + num_dataset_builder_threads ..................... 1 + num_distributed_optimizer_instances ............. 1 + num_experts ..................................... 32 + num_layers ...................................... 16 + num_layers_at_end_in_bf16 ....................... 1 + num_layers_at_start_in_bf16 ..................... 1 + num_layers_per_virtual_pipeline_stage ........... None + num_query_groups ................................ 8 + 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 2 +> building GPT2BPETokenizer tokenizer ... + > 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") +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") +[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 7[Gloo] Rank is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +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 5 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 [Gloo] Rank 45 is connected to is connected to 77 peer ranks. peer ranks. Expected number of connected peer ranks is : 7Expected 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 [Gloo] Rank 76 is connected to is connected to 77 peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : 77 + +[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 +> 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 +[rank3]:[W917 03:27:12.759074705 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. +[rank1]:[W917 03:27:12.759291933 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]:[W917 03:27:12.759731800 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. +[rank6]:[W917 03:27:12.760213509 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. +[rank2]:[W917 03:27:12.763662209 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. +[rank4]:[W917 03:27:12.763770125 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. +[rank7]:[W917 03:27:12.763810340 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: 0.437 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]:[W917 03:27:13.386048103 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: 4.013 seconds +time to initialize megatron (seconds): 9.439 +[after megatron is initialized] datetime: 2026-09-17 03:27:19 +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): 6722555904 +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: 18 +Params for bucket 1 (206045184 elements, 206045184 padded size): + module.output_layer.weight +Params for bucket 2 (42086400 elements, 42086400 padded size): + module.decoder.layers.15.self_attention.linear_qkv.weight + module.decoder.layers.15.pre_mlp_layernorm.weight + module.decoder.final_layernorm.weight + module.decoder.layers.15.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.15.mlp.router.weight + module.decoder.layers.15.self_attention.linear_proj.weight +Params for bucket 3 (42082304 elements, 42082304 padded size): + module.decoder.layers.14.self_attention.linear_qkv.weight + module.decoder.layers.14.pre_mlp_layernorm.weight + module.decoder.layers.14.mlp.router.weight + module.decoder.layers.14.self_attention.linear_proj.weight + module.decoder.layers.14.self_attention.linear_qkv.layer_norm_weight +Params for bucket 4 (42082304 elements, 42082304 padded size): + module.decoder.layers.13.pre_mlp_layernorm.weight + module.decoder.layers.13.self_attention.linear_qkv.weight + module.decoder.layers.13.mlp.router.weight + module.decoder.layers.13.self_attention.linear_proj.weight + module.decoder.layers.13.self_attention.linear_qkv.layer_norm_weight +Params for bucket 5 (42082304 elements, 42082304 padded size): + module.decoder.layers.12.self_attention.linear_qkv.weight + module.decoder.layers.12.pre_mlp_layernorm.weight + module.decoder.layers.12.mlp.router.weight + module.decoder.layers.12.self_attention.linear_proj.weight + module.decoder.layers.12.self_attention.linear_qkv.layer_norm_weight +Params for bucket 6 (42082304 elements, 42082304 padded size): + module.decoder.layers.11.self_attention.linear_qkv.weight + module.decoder.layers.11.pre_mlp_layernorm.weight + module.decoder.layers.11.mlp.router.weight + module.decoder.layers.11.self_attention.linear_proj.weight + module.decoder.layers.11.self_attention.linear_qkv.layer_norm_weight +Params for bucket 7 (42082304 elements, 42082304 padded size): + module.decoder.layers.10.self_attention.linear_qkv.weight + module.decoder.layers.10.pre_mlp_layernorm.weight + module.decoder.layers.10.mlp.router.weight + module.decoder.layers.10.self_attention.linear_proj.weight + module.decoder.layers.10.self_attention.linear_qkv.layer_norm_weight +Params for bucket 8 (42082304 elements, 42082304 padded size): + module.decoder.layers.9.self_attention.linear_qkv.weight + module.decoder.layers.9.pre_mlp_layernorm.weight + module.decoder.layers.9.mlp.router.weight + module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.9.self_attention.linear_proj.weight +Params for bucket 9 (42082304 elements, 42082304 padded size): + module.decoder.layers.8.self_attention.linear_qkv.weight + module.decoder.layers.8.pre_mlp_layernorm.weight + module.decoder.layers.8.mlp.router.weight + module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.8.self_attention.linear_proj.weight +Params for bucket 10 (42082304 elements, 42082304 padded size): + module.decoder.layers.7.mlp.router.weight + module.decoder.layers.7.self_attention.linear_proj.weight + module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.7.self_attention.linear_qkv.weight + module.decoder.layers.7.pre_mlp_layernorm.weight +Params for bucket 11 (42082304 elements, 42082304 padded size): + module.decoder.layers.6.mlp.router.weight + module.decoder.layers.6.self_attention.linear_proj.weight + module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.6.self_attention.linear_qkv.weight + module.decoder.layers.6.pre_mlp_layernorm.weight +Params for bucket 12 (42082304 elements, 42082304 padded size): + module.decoder.layers.5.self_attention.linear_qkv.weight + module.decoder.layers.5.self_attention.linear_proj.weight + module.decoder.layers.5.pre_mlp_layernorm.weight + module.decoder.layers.5.mlp.router.weight + module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight +Params for bucket 13 (42082304 elements, 42082304 padded size): + module.decoder.layers.4.mlp.router.weight + module.decoder.layers.4.self_attention.linear_proj.weight + module.decoder.layers.4.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.4.self_attention.linear_qkv.weight + module.decoder.layers.4.pre_mlp_layernorm.weight +Params for bucket 14 (42082304 elements, 42082304 padded size): + module.decoder.layers.3.mlp.router.weight + module.decoder.layers.3.self_attention.linear_proj.weight + module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.3.self_attention.linear_qkv.weight + module.decoder.layers.3.pre_mlp_layernorm.weight +Params for bucket 15 (42082304 elements, 42082304 padded size): + module.decoder.layers.2.mlp.router.weight + module.decoder.layers.2.self_attention.linear_proj.weight + module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.2.pre_mlp_layernorm.weight + module.decoder.layers.2.self_attention.linear_qkv.weight +Params for bucket 16 (42082304 elements, 42082304 padded size): + module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.1.self_attention.linear_proj.weight + module.decoder.layers.1.mlp.router.weight + module.decoder.layers.1.self_attention.linear_qkv.weight + module.decoder.layers.1.pre_mlp_layernorm.weight +Params for bucket 17 (42082304 elements, 42082304 padded size): + module.decoder.layers.0.pre_mlp_layernorm.weight + module.decoder.layers.0.self_attention.linear_proj.weight + module.decoder.layers.0.mlp.router.weight + module.decoder.layers.0.self_attention.linear_qkv.weight + module.decoder.layers.0.self_attention.linear_qkv.layer_norm_weight +Params for bucket 18 (206045184 elements, 206045184 padded size): + module.embedding.word_embeddings.weight +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 56 +Params for bucket 1 (44040192 elements, 44040192 padded size): + 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_fc2.weight2 +Params for bucket 2 (44040192 elements, 44040192 padded size): + module.decoder.layers.15.mlp.experts.linear_fc2.weight0 + module.decoder.layers.15.mlp.experts.linear_fc1.weight3 +Params for bucket 3 (58720256 elements, 58720256 padded size): + module.decoder.layers.15.mlp.experts.linear_fc1.weight2 + module.decoder.layers.15.mlp.experts.linear_fc1.weight1 +Params for bucket 4 (44040192 elements, 44040192 padded size): + module.decoder.layers.15.mlp.experts.linear_fc1.weight0 + module.decoder.layers.14.mlp.experts.linear_fc2.weight3 +Params for bucket 5 (44040192 elements, 44040192 padded size): + module.decoder.layers.14.mlp.experts.linear_fc2.weight2 + module.decoder.layers.14.mlp.experts.linear_fc2.weight0 + module.decoder.layers.14.mlp.experts.linear_fc2.weight1 +Params for bucket 6 (58720256 elements, 58720256 padded size): + module.decoder.layers.14.mlp.experts.linear_fc1.weight3 + module.decoder.layers.14.mlp.experts.linear_fc1.weight2 +Params for bucket 7 (58720256 elements, 58720256 padded size): + module.decoder.layers.14.mlp.experts.linear_fc1.weight1 + module.decoder.layers.14.mlp.experts.linear_fc1.weight0 +Params for bucket 8 (44040192 elements, 44040192 padded size): + 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_fc2.weight2 +Params for bucket 9 (44040192 elements, 44040192 padded size): + module.decoder.layers.13.mlp.experts.linear_fc2.weight0 + module.decoder.layers.13.mlp.experts.linear_fc1.weight3 +Params for bucket 10 (58720256 elements, 58720256 padded size): + module.decoder.layers.13.mlp.experts.linear_fc1.weight2 + module.decoder.layers.13.mlp.experts.linear_fc1.weight1 +Params for bucket 11 (44040192 elements, 44040192 padded size): + module.decoder.layers.13.mlp.experts.linear_fc1.weight0 + module.decoder.layers.12.mlp.experts.linear_fc2.weight3 +Params for bucket 12 (44040192 elements, 44040192 padded size): + module.decoder.layers.12.mlp.experts.linear_fc2.weight2 + module.decoder.layers.12.mlp.experts.linear_fc2.weight0 + module.decoder.layers.12.mlp.experts.linear_fc2.weight1 +Params for bucket 13 (58720256 elements, 58720256 padded size): + module.decoder.layers.12.mlp.experts.linear_fc1.weight3 + module.decoder.layers.12.mlp.experts.linear_fc1.weight2 +Params for bucket 14 (58720256 elements, 58720256 padded size): + module.decoder.layers.12.mlp.experts.linear_fc1.weight1 + module.decoder.layers.12.mlp.experts.linear_fc1.weight0 +Params for bucket 15 (44040192 elements, 44040192 padded size): + module.decoder.layers.11.mlp.experts.linear_fc2.weight3 + module.decoder.layers.11.mlp.experts.linear_fc2.weight1 + module.decoder.layers.11.mlp.experts.linear_fc2.weight2 +Params for bucket 16 (44040192 elements, 44040192 padded size): + module.decoder.layers.11.mlp.experts.linear_fc2.weight0 + module.decoder.layers.11.mlp.experts.linear_fc1.weight3 +Params for bucket 17 (58720256 elements, 58720256 padded size): + module.decoder.layers.11.mlp.experts.linear_fc1.weight2 + module.decoder.layers.11.mlp.experts.linear_fc1.weight1 +Params for bucket 18 (44040192 elements, 44040192 padded size): + module.decoder.layers.11.mlp.experts.linear_fc1.weight0 + module.decoder.layers.10.mlp.experts.linear_fc2.weight3 +Params for bucket 19 (44040192 elements, 44040192 padded size): + module.decoder.layers.10.mlp.experts.linear_fc2.weight2 + module.decoder.layers.10.mlp.experts.linear_fc2.weight1 + module.decoder.layers.10.mlp.experts.linear_fc2.weight0 +Params for bucket 20 (58720256 elements, 58720256 padded size): + module.decoder.layers.10.mlp.experts.linear_fc1.weight3 + module.decoder.layers.10.mlp.experts.linear_fc1.weight2 +Params for bucket 21 (58720256 elements, 58720256 padded size): + module.decoder.layers.10.mlp.experts.linear_fc1.weight1 + module.decoder.layers.10.mlp.experts.linear_fc1.weight0 +Params for bucket 22 (44040192 elements, 44040192 padded size): + 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_fc2.weight2 +Params for bucket 23 (44040192 elements, 44040192 padded size): + module.decoder.layers.9.mlp.experts.linear_fc2.weight0 + module.decoder.layers.9.mlp.experts.linear_fc1.weight3 +Params for bucket 24 (58720256 elements, 58720256 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight2 + module.decoder.layers.9.mlp.experts.linear_fc1.weight1 +Params for bucket 25 (44040192 elements, 44040192 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight0 + module.decoder.layers.8.mlp.experts.linear_fc2.weight3 +Params for bucket 26 (44040192 elements, 44040192 padded size): + module.decoder.layers.8.mlp.experts.linear_fc2.weight2 + module.decoder.layers.8.mlp.experts.linear_fc2.weight1 + module.decoder.layers.8.mlp.experts.linear_fc2.weight0 +Params for bucket 27 (58720256 elements, 58720256 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight3 + module.decoder.layers.8.mlp.experts.linear_fc1.weight2 +Params for bucket 28 (58720256 elements, 58720256 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight1 + module.decoder.layers.8.mlp.experts.linear_fc1.weight0 +Params for bucket 29 (44040192 elements, 44040192 padded size): + 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.weight2 +Params for bucket 30 (44040192 elements, 44040192 padded size): + module.decoder.layers.7.mlp.experts.linear_fc2.weight0 + module.decoder.layers.7.mlp.experts.linear_fc1.weight3 +Params for bucket 31 (58720256 elements, 58720256 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight2 + module.decoder.layers.7.mlp.experts.linear_fc1.weight1 +Params for bucket 32 (44040192 elements, 44040192 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight0 + module.decoder.layers.6.mlp.experts.linear_fc2.weight3 +Params for bucket 33 (44040192 elements, 44040192 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_fc2.weight0 +Params for bucket 34 (58720256 elements, 58720256 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight3 + module.decoder.layers.6.mlp.experts.linear_fc1.weight2 +Params for bucket 35 (58720256 elements, 58720256 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight1 + module.decoder.layers.6.mlp.experts.linear_fc1.weight0 +Params for bucket 36 (44040192 elements, 44040192 padded size): + module.decoder.layers.5.mlp.experts.linear_fc2.weight3 + module.decoder.layers.5.mlp.experts.linear_fc2.weight1 + module.decoder.layers.5.mlp.experts.linear_fc2.weight2 +Params for bucket 37 (44040192 elements, 44040192 padded size): + module.decoder.layers.5.mlp.experts.linear_fc2.weight0 + module.decoder.layers.5.mlp.experts.linear_fc1.weight3 +Params for bucket 38 (58720256 elements, 58720256 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight2 + module.decoder.layers.5.mlp.experts.linear_fc1.weight1 +Params for bucket 39 (44040192 elements, 44040192 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight0 + module.decoder.layers.4.mlp.experts.linear_fc2.weight3 +Params for bucket 40 (44040192 elements, 44040192 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight2 + module.decoder.layers.4.mlp.experts.linear_fc2.weight1 + module.decoder.layers.4.mlp.experts.linear_fc2.weight0 +Params for bucket 41 (58720256 elements, 58720256 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight3 + module.decoder.layers.4.mlp.experts.linear_fc1.weight2 +Params for bucket 42 (58720256 elements, 58720256 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight1 + module.decoder.layers.4.mlp.experts.linear_fc1.weight0 +Params for bucket 43 (44040192 elements, 44040192 padded size): + module.decoder.layers.3.mlp.experts.linear_fc2.weight3 + module.decoder.layers.3.mlp.experts.linear_fc2.weight1 + module.decoder.layers.3.mlp.experts.linear_fc2.weight2 +Params for bucket 44 (44040192 elements, 44040192 padded size): + module.decoder.layers.3.mlp.experts.linear_fc2.weight0 + module.decoder.layers.3.mlp.experts.linear_fc1.weight3 +Params for bucket 45 (58720256 elements, 58720256 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight2 + module.decoder.layers.3.mlp.experts.linear_fc1.weight1 +Params for bucket 46 (44040192 elements, 44040192 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight0 + module.decoder.layers.2.mlp.experts.linear_fc2.weight3 +Params for bucket 47 (44040192 elements, 44040192 padded size): + 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_fc2.weight1 +Params for bucket 48 (58720256 elements, 58720256 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight3 + module.decoder.layers.2.mlp.experts.linear_fc1.weight2 +Params for bucket 49 (58720256 elements, 58720256 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight1 + module.decoder.layers.2.mlp.experts.linear_fc1.weight0 +Params for bucket 50 (44040192 elements, 44040192 padded size): + 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_fc2.weight2 +Params for bucket 51 (44040192 elements, 44040192 padded size): + module.decoder.layers.1.mlp.experts.linear_fc2.weight0 + module.decoder.layers.1.mlp.experts.linear_fc1.weight3 +Params for bucket 52 (58720256 elements, 58720256 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight1 +Params for bucket 53 (44040192 elements, 44040192 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight0 + module.decoder.layers.0.mlp.experts.linear_fc2.weight3 +Params for bucket 54 (44040192 elements, 44040192 padded size): + module.decoder.layers.0.mlp.experts.linear_fc2.weight2 + module.decoder.layers.0.mlp.experts.linear_fc2.weight0 + module.decoder.layers.0.mlp.experts.linear_fc2.weight1 +Params for bucket 55 (58720256 elements, 58720256 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight3 + module.decoder.layers.0.mlp.experts.linear_fc1.weight2 +Params for bucket 56 (58720256 elements, 58720256 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight1 + 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: 2026-09-17 03:27:22 +> building train, validation, and test datasets ... + > datasets target sizes (minimum size): + train: 160 + 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=(160, 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 68684a515e118253a8572c4517cf2ffe-GPTDataset-train-document_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 68684a515e118253a8572c4517cf2ffe-GPTDataset-train-sample_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 68684a515e118253a8572c4517cf2ffe-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: 2026-09-17 03:27:22 +done with setup ... +(min, max) time across ranks (ms): + model-and-optimizer-setup ......................: (3347.95, 3351.05) + train/valid/test-data-iterators-setup ..........: (113.53, 139.33) +training ... +Overwriting rerun_state_machine.current_iteration from -1 to 0... +[before the start of training step] datetime: 2026-09-17 03:27:23 +/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: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: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.") +/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.") + [2026-09-17 03:27:52] iteration 1/ 10 | consumed samples: 16 | elapsed time per iteration (ms): 29813.9 | throughput per GPU (TFLOP/s/GPU): 4.2 | learning rate: 2.000000E-05 | global batch size: 16 | lm loss: 1.102902E+01 | loss scale: 1.0 | grad norm: 44.186 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 8832 | min safe capacity: 1.0781 | +Number of parameters in transformer block in billions: 23.22 +Number of parameters in embedding layers in billions: 0.41 +Total number of parameters in billions: 23.63 +Number of parameters in most loaded shard in billions: 23.6320 +Theoretical memory footprints: weight and optimizer=169029.43 MB +[Rank 0] (after 1 iterations) memory (MB) | allocated: 58124.0458984375 | max allocated: 65718.16796875 | reserved: 79458.0 | max reserved: 79458.0 + [2026-09-17 03:27:53] iteration 2/ 10 | consumed samples: 32 | elapsed time per iteration (ms): 586.2 | throughput per GPU (TFLOP/s/GPU): 214.2 | learning rate: 4.000000E-05 | global batch size: 16 | lm loss: 1.101823E+01 | loss scale: 1.0 | grad norm: 45.729 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 8832 | min safe capacity: 1.0781 | + [2026-09-17 03:27:53] iteration 3/ 10 | consumed samples: 48 | elapsed time per iteration (ms): 434.7 | throughput per GPU (TFLOP/s/GPU): 288.9 | learning rate: 6.000000E-05 | global batch size: 16 | lm loss: 9.972256E+00 | loss scale: 1.0 | grad norm: 157.296 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 8832 | min safe capacity: 1.0781 | +Capture CUDA graph for training!!! +CUDA graph capture done!!! + [2026-09-17 03:27:57] iteration 4/ 10 | consumed samples: 64 | elapsed time per iteration (ms): 3115.4 | throughput per GPU (TFLOP/s/GPU): 40.3 | learning rate: 8.000000E-05 | global batch size: 16 | lm loss: 1.311298E+01 | loss scale: 1.0 | grad norm: 41.251 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 8832 | min safe capacity: 1.0781 | + [2026-09-17 03:27:57] iteration 5/ 10 | consumed samples: 80 | elapsed time per iteration (ms): 323.9 | throughput per GPU (TFLOP/s/GPU): 387.7 | learning rate: 1.000000E-04 | global batch size: 16 | lm loss: 1.040733E+01 | loss scale: 1.0 | grad norm: 25.625 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 8832 | min safe capacity: 1.0781 | + [2026-09-17 03:27:57] iteration 6/ 10 | consumed samples: 96 | elapsed time per iteration (ms): 333.2 | throughput per GPU (TFLOP/s/GPU): 376.9 | learning rate: 7.750000E-05 | global batch size: 16 | lm loss: 1.025381E+01 | loss scale: 1.0 | grad norm: 19.172 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 8832 | min safe capacity: 1.0781 | + [2026-09-17 03:28:00] iteration 7/ 10 | consumed samples: 112 | elapsed time per iteration (ms): 3161.7 | throughput per GPU (TFLOP/s/GPU): 39.7 | learning rate: 3.250000E-05 | global batch size: 16 | lm loss: 1.002502E+01 | loss scale: 1.0 | grad norm: 8.664 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 8832 | min safe capacity: 1.0781 | + [2026-09-17 03:28:01] iteration 8/ 10 | consumed samples: 128 | elapsed time per iteration (ms): 342.7 | throughput per GPU (TFLOP/s/GPU): 366.4 | learning rate: 1.000000E-05 | global batch size: 16 | lm loss: 9.464363E+00 | loss scale: 1.0 | grad norm: 8.108 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 8832 | min safe capacity: 1.0781 | +Capture range started in the application. +Capture range ended in the application. +Generating '/tmp/nsys-report-2929.qdstrm' + [1/1] [0% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-repProcessing events... + [1/1] [0% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [0% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [5% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [8% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [11% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [13% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [=16% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [==19% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [==21% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [===24% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [====27% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [=====29% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [=====32% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [======35% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [=======37% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [=======38% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [=======39% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [========40% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [========41% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [========42% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [=========44% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [=========45% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [=========46% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [==========48% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [==========49% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [===========51% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [===========52% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [===========53% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [============55% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [============56% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [=============58% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [=============60% ] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [========================100%] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep [1/1] [========================100%] G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep +Collection stop is not allowed in this state. +W0917 03:29:32.760000 948326 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers +W0917 03:29:32.761000 948326 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 948649 closing signal SIGTERM +Generated: + /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep +W0917 03:29:32.765000 948326 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 948650 closing signal SIGTERM +W0917 03:29:32.772000 948326 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 948651 closing signal SIGTERM +W0917 03:29:32.773000 948326 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 948652 closing signal SIGTERM +W0917 03:29:32.777000 948326 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 948653 closing signal SIGTERM +W0917 03:29:32.783000 948326 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 948654 closing signal SIGTERM +W0917 03:29:32.794000 948326 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 948655 closing signal SIGTERM +W0917 03:29:32.804000 948326 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 948656 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 948326 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/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/verify_trace_ranks.log b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/verify_trace_ranks.log new file mode 100644 index 0000000000000000000000000000000000000000..3512a874fb7a5d849cbce38fa54d6568b0eb2126 --- /dev/null +++ b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/verify_trace_ranks.log @@ -0,0 +1,6 @@ +report : /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep +kernels : 143382 +devices : 8 of 8 -> [0, 1, 2, 3, 4, 5, 6, 7] +processes: 1 +markers : 26624 ultraep_trace__ kernels on devices [0, 1, 2, 3, 4, 5, 6, 7] +PASS: all ranks present and backend markers emitted diff --git a/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep new file mode 100644 index 0000000000000000000000000000000000000000..b2b8a2df3cc0197c0f58276dddb1d0c2b0866357 --- /dev/null +++ b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b72936061abb85f8abaf8716a77beb00af3bf28399bb4c5824e9a1ef123b6169 +size 12226799 diff --git a/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/run.env b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/run.env new file mode 100644 index 0000000000000000000000000000000000000000..b1b7bbbb397ac16d0e1231fbdec458dfd500c53b --- /dev/null +++ b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/run.env @@ -0,0 +1,102 @@ +backend=ultraep +config=G3_K2 +family=tierA +source_model= +standard_num_layers= +base_rung=G3 +repo=/workspace/Megatron-UltraEP +branch=ultraep-bracing-regions-through-kernels +commit=03d57ef7 +num_layers=16 +hidden=4096 +ffn_hidden=14336 +moe_ffn_hidden=1792 +num_experts=64 +topk=2 +experts_per_rank=8 +expert_params=22020096 +heads=32 +kv_groups=8 +kv_channels= +qk_layernorm=0 +shared_ffn=0 +shared_expert_overlap=0 +moe_token_dispatcher=flex +mla=0 +q_lora_rank= +kv_lora_rank=512 +qk_head_dim=128 +qk_pos_emb_head_dim=64 +v_head_dim=128 +dense_first_layers=0 +moe_layers=16 +rotary_base=1000000 +rotary_percent=1.0 +rotary_scaling= +mscale= +mscale_all_dim= +qkv_bias=0 +router_balancing=aux_loss +aux_loss_coeff=1e-2 +router_pre_softmax=0 +router_fusion=0 +router_dtype=fp32 +router_groups= +router_group_topk= +router_score_function=softmax +router_expert_bias=0 +router_bias_update_rate= +router_topk_scaling_factor= +shared_expert_gate=0 +apply_layernorm_1p=0 +attention_output_gate=0 +linear_attention_type= +linear_attention_freq= +linear_conv_kernel_dim= +linear_key_head_dim= +linear_value_head_dim= +linear_num_key_heads= +linear_num_value_heads= +mtp_num_layers=0 +mtp_loss_scaling_factor=0.1 +no_weight_decay_cond_type= +max_position_embeddings=32768 +init_method_std=0.01 +device_initiated_grouped_gemm=1 +norm_epsilon=1e-5 +seq_length=4096 +micro_batch_size=1 +global_batch_size=16 +num_microbatches=2 +assignments_per_rank=8192 +mean_tokens_per_expert=1024 +slot_policy=matched +slots_per_rank=8 +capacity_mode=static +capacity=1.1875 +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=128 +ultraep_weight_sync_join= +profile_enable=1 +kernel_trace=1 +fp8_mode=mxfp8 +fp8_recipe=mxfp8 +training_seed=1234 +learning_rate=1e-4 +min_learning_rate=1.0e-5 +lr_warmup_iters=5 +state_dump_enable=0 +state_dump_dir= +moonep_assert_no_pending_reduce=0 diff --git a/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/train.log b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/train.log new file mode 100644 index 0000000000000000000000000000000000000000..163437bbed5fa33d1d425be78219fdfd917d2d74 --- /dev/null +++ b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/train.log @@ -0,0 +1,1312 @@ +============================================================== + G3_K2 / ultraep (slots/rank=8 policy=matched) + E=64 K=2 H'=1792 m=8 + L=16 seq=4096 gbs=16 (2 microbatches) + cap_mode=static cap=1.1875 cg=1 recompute=none + memory_dump=none snapshots=/workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/memory_snapshots/snapshot.pickle + repo=/workspace/Megatron-UltraEP @ 03d57ef7 +============================================================== +W0917 03:29:57.172000 960711 torch/distributed/run.py:766] +W0917 03:29:57.172000 960711 torch/distributed/run.py:766] ***************************************** +W0917 03:29:57.172000 960711 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. +W0917 03:29:57.172000 960711 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") +/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") +/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 ............................... 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 .............................. 16 + 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 ................................. 14336 + 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 ............................... 16 + 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 ..................................... 4096 + 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 ............................. 1792 + 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 ..................... 1.1875 + 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 .......................... False + moe_router_score_function ....................... softmax + moe_router_topk ................................. 2 + moe_router_topk_scaling_factor .................. None + moe_shared_expert_intermediate_size ............. None + 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-05 + normalization ................................... RMSNorm + num_attention_heads ............................. 32 + num_channels .................................... 3 + num_classes ..................................... 1000 + num_dataset_builder_threads ..................... 1 + num_distributed_optimizer_instances ............. 1 + num_experts ..................................... 64 + num_layers ...................................... 16 + num_layers_at_end_in_bf16 ....................... 1 + num_layers_at_start_in_bf16 ..................... 1 + num_layers_per_virtual_pipeline_stage ........... None + num_query_groups ................................ 8 + 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 2 +> 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 ... +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 +[Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank [Gloo] Rank 0 is connected to 37 is connected to peer ranks. Expected number of connected peer ranks is : 77 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 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank [Gloo] Rank 7 is connected to 76 peer ranks. Expected number of connected peer ranks is : is connected to 77 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 [Gloo] Rank 5 is connected to 37 is connected to peer ranks. 7Expected number of connected peer ranks is : peer ranks. 7Expected 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 7 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 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[Gloo] Rank peer ranks. Expected number of connected peer ranks is : 0 +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 +[rank3]:[W917 03:30:09.927160051 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. +[rank1]:[W917 03:30:09.927160874 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]:[W917 03:30:09.927340509 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. +[rank6]:[W917 03:30:09.927344996 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. +[rank4]:[W917 03:30:09.927925286 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. +[rank7]:[W917 03:30:09.927971518 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. +[rank2]:[W917 03:30:09.928149080 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: 0.437 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]:[W917 03:30:10.553623237 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: 3.679 seconds +time to initialize megatron (seconds): 9.143 +[after megatron is initialized] datetime: 2026-09-17 03:30:16 +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): 6724653056 +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: 18 +Params for bucket 1 (206045184 elements, 206045184 padded size): + module.output_layer.weight +Params for bucket 2 (42217472 elements, 42217472 padded size): + module.decoder.layers.15.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.15.mlp.router.weight + module.decoder.layers.15.self_attention.linear_proj.weight + module.decoder.final_layernorm.weight + module.decoder.layers.15.self_attention.linear_qkv.weight + module.decoder.layers.15.pre_mlp_layernorm.weight +Params for bucket 3 (42213376 elements, 42213376 padded size): + module.decoder.layers.14.mlp.router.weight + module.decoder.layers.14.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.14.self_attention.linear_proj.weight + module.decoder.layers.14.self_attention.linear_qkv.weight + module.decoder.layers.14.pre_mlp_layernorm.weight +Params for bucket 4 (42213376 elements, 42213376 padded size): + module.decoder.layers.13.mlp.router.weight + module.decoder.layers.13.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.13.self_attention.linear_proj.weight + module.decoder.layers.13.self_attention.linear_qkv.weight + module.decoder.layers.13.pre_mlp_layernorm.weight +Params for bucket 5 (42213376 elements, 42213376 padded size): + module.decoder.layers.12.mlp.router.weight + module.decoder.layers.12.self_attention.linear_proj.weight + module.decoder.layers.12.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.12.pre_mlp_layernorm.weight + module.decoder.layers.12.self_attention.linear_qkv.weight +Params for bucket 6 (42213376 elements, 42213376 padded size): + module.decoder.layers.11.mlp.router.weight + module.decoder.layers.11.self_attention.linear_proj.weight + module.decoder.layers.11.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.11.self_attention.linear_qkv.weight + module.decoder.layers.11.pre_mlp_layernorm.weight +Params for bucket 7 (42213376 elements, 42213376 padded size): + module.decoder.layers.10.mlp.router.weight + module.decoder.layers.10.self_attention.linear_proj.weight + module.decoder.layers.10.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.10.pre_mlp_layernorm.weight + module.decoder.layers.10.self_attention.linear_qkv.weight +Params for bucket 8 (42213376 elements, 42213376 padded size): + module.decoder.layers.9.mlp.router.weight + module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.9.self_attention.linear_proj.weight + module.decoder.layers.9.self_attention.linear_qkv.weight + module.decoder.layers.9.pre_mlp_layernorm.weight +Params for bucket 9 (42213376 elements, 42213376 padded size): + module.decoder.layers.8.self_attention.linear_qkv.weight + module.decoder.layers.8.pre_mlp_layernorm.weight + module.decoder.layers.8.mlp.router.weight + module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.8.self_attention.linear_proj.weight +Params for bucket 10 (42213376 elements, 42213376 padded size): + module.decoder.layers.7.pre_mlp_layernorm.weight + module.decoder.layers.7.self_attention.linear_qkv.weight + module.decoder.layers.7.mlp.router.weight + module.decoder.layers.7.self_attention.linear_proj.weight + module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight +Params for bucket 11 (42213376 elements, 42213376 padded size): + module.decoder.layers.6.self_attention.linear_qkv.weight + module.decoder.layers.6.pre_mlp_layernorm.weight + module.decoder.layers.6.mlp.router.weight + module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.6.self_attention.linear_proj.weight +Params for bucket 12 (42213376 elements, 42213376 padded size): + module.decoder.layers.5.self_attention.linear_proj.weight + module.decoder.layers.5.pre_mlp_layernorm.weight + module.decoder.layers.5.self_attention.linear_qkv.weight + module.decoder.layers.5.mlp.router.weight + module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight +Params for bucket 13 (42213376 elements, 42213376 padded size): + module.decoder.layers.4.mlp.router.weight + module.decoder.layers.4.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.4.self_attention.linear_proj.weight + module.decoder.layers.4.self_attention.linear_qkv.weight + module.decoder.layers.4.pre_mlp_layernorm.weight +Params for bucket 14 (42213376 elements, 42213376 padded size): + module.decoder.layers.3.mlp.router.weight + module.decoder.layers.3.self_attention.linear_proj.weight + module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.3.pre_mlp_layernorm.weight + module.decoder.layers.3.self_attention.linear_qkv.weight +Params for bucket 15 (42213376 elements, 42213376 padded size): + module.decoder.layers.2.mlp.router.weight + module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.2.self_attention.linear_proj.weight + module.decoder.layers.2.self_attention.linear_qkv.weight + module.decoder.layers.2.pre_mlp_layernorm.weight +Params for bucket 16 (42213376 elements, 42213376 padded size): + module.decoder.layers.1.mlp.router.weight + module.decoder.layers.1.self_attention.linear_proj.weight + module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.1.self_attention.linear_qkv.weight + module.decoder.layers.1.pre_mlp_layernorm.weight +Params for bucket 17 (42213376 elements, 42213376 padded size): + module.decoder.layers.0.pre_mlp_layernorm.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.mlp.router.weight + module.decoder.layers.0.self_attention.linear_proj.weight +Params for bucket 18 (206045184 elements, 206045184 padded size): + module.embedding.word_embeddings.weight +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 64 +Params for bucket 1 (44040192 elements, 44040192 padded size): + module.decoder.layers.15.mlp.experts.linear_fc2.weight4 + module.decoder.layers.15.mlp.experts.linear_fc2.weight6 + module.decoder.layers.15.mlp.experts.linear_fc2.weight2 + module.decoder.layers.15.mlp.experts.linear_fc2.weight7 + module.decoder.layers.15.mlp.experts.linear_fc2.weight3 + module.decoder.layers.15.mlp.experts.linear_fc2.weight5 +Params for bucket 2 (44040192 elements, 44040192 padded size): + module.decoder.layers.15.mlp.experts.linear_fc2.weight1 + module.decoder.layers.15.mlp.experts.linear_fc2.weight0 + module.decoder.layers.15.mlp.experts.linear_fc1.weight6 + module.decoder.layers.15.mlp.experts.linear_fc1.weight7 +Params for bucket 3 (44040192 elements, 44040192 padded size): + 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_fc1.weight4 +Params for bucket 4 (44040192 elements, 44040192 padded size): + module.decoder.layers.15.mlp.experts.linear_fc1.weight2 + module.decoder.layers.15.mlp.experts.linear_fc1.weight0 + module.decoder.layers.15.mlp.experts.linear_fc1.weight1 +Params for bucket 5 (44040192 elements, 44040192 padded size): + module.decoder.layers.14.mlp.experts.linear_fc2.weight7 + module.decoder.layers.14.mlp.experts.linear_fc2.weight3 + module.decoder.layers.14.mlp.experts.linear_fc2.weight5 + module.decoder.layers.14.mlp.experts.linear_fc2.weight4 + module.decoder.layers.14.mlp.experts.linear_fc2.weight6 + module.decoder.layers.14.mlp.experts.linear_fc2.weight2 +Params for bucket 6 (44040192 elements, 44040192 padded size): + module.decoder.layers.14.mlp.experts.linear_fc2.weight1 + module.decoder.layers.14.mlp.experts.linear_fc2.weight0 + module.decoder.layers.14.mlp.experts.linear_fc1.weight7 + module.decoder.layers.14.mlp.experts.linear_fc1.weight6 +Params for bucket 7 (44040192 elements, 44040192 padded size): + module.decoder.layers.14.mlp.experts.linear_fc1.weight5 + module.decoder.layers.14.mlp.experts.linear_fc1.weight4 + module.decoder.layers.14.mlp.experts.linear_fc1.weight3 +Params for bucket 8 (44040192 elements, 44040192 padded size): + module.decoder.layers.14.mlp.experts.linear_fc1.weight2 + module.decoder.layers.14.mlp.experts.linear_fc1.weight1 + module.decoder.layers.14.mlp.experts.linear_fc1.weight0 +Params for bucket 9 (44040192 elements, 44040192 padded size): + module.decoder.layers.13.mlp.experts.linear_fc2.weight7 + module.decoder.layers.13.mlp.experts.linear_fc2.weight5 + module.decoder.layers.13.mlp.experts.linear_fc2.weight3 + 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 +Params for bucket 10 (44040192 elements, 44040192 padded size): + 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.weight0 + module.decoder.layers.13.mlp.experts.linear_fc1.weight7 +Params for bucket 11 (44040192 elements, 44040192 padded size): + 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.weight4 +Params for bucket 12 (44040192 elements, 44040192 padded size): + module.decoder.layers.13.mlp.experts.linear_fc1.weight2 + module.decoder.layers.13.mlp.experts.linear_fc1.weight0 + module.decoder.layers.13.mlp.experts.linear_fc1.weight1 +Params for bucket 13 (44040192 elements, 44040192 padded size): + module.decoder.layers.12.mlp.experts.linear_fc2.weight7 + module.decoder.layers.12.mlp.experts.linear_fc2.weight5 + module.decoder.layers.12.mlp.experts.linear_fc2.weight3 + module.decoder.layers.12.mlp.experts.linear_fc2.weight6 + module.decoder.layers.12.mlp.experts.linear_fc2.weight4 + module.decoder.layers.12.mlp.experts.linear_fc2.weight2 +Params for bucket 14 (44040192 elements, 44040192 padded size): + module.decoder.layers.12.mlp.experts.linear_fc2.weight1 + module.decoder.layers.12.mlp.experts.linear_fc1.weight6 + module.decoder.layers.12.mlp.experts.linear_fc2.weight0 + module.decoder.layers.12.mlp.experts.linear_fc1.weight7 +Params for bucket 15 (44040192 elements, 44040192 padded size): + module.decoder.layers.12.mlp.experts.linear_fc1.weight5 + module.decoder.layers.12.mlp.experts.linear_fc1.weight3 + module.decoder.layers.12.mlp.experts.linear_fc1.weight4 +Params for bucket 16 (44040192 elements, 44040192 padded size): + module.decoder.layers.12.mlp.experts.linear_fc1.weight2 + module.decoder.layers.12.mlp.experts.linear_fc1.weight0 + module.decoder.layers.12.mlp.experts.linear_fc1.weight1 +Params for bucket 17 (44040192 elements, 44040192 padded size): + module.decoder.layers.11.mlp.experts.linear_fc2.weight7 + module.decoder.layers.11.mlp.experts.linear_fc2.weight5 + module.decoder.layers.11.mlp.experts.linear_fc2.weight3 + module.decoder.layers.11.mlp.experts.linear_fc2.weight4 + module.decoder.layers.11.mlp.experts.linear_fc2.weight6 + module.decoder.layers.11.mlp.experts.linear_fc2.weight2 +Params for bucket 18 (44040192 elements, 44040192 padded size): + module.decoder.layers.11.mlp.experts.linear_fc2.weight1 + module.decoder.layers.11.mlp.experts.linear_fc1.weight7 + module.decoder.layers.11.mlp.experts.linear_fc1.weight6 + module.decoder.layers.11.mlp.experts.linear_fc2.weight0 +Params for bucket 19 (44040192 elements, 44040192 padded size): + module.decoder.layers.11.mlp.experts.linear_fc1.weight5 + module.decoder.layers.11.mlp.experts.linear_fc1.weight4 + module.decoder.layers.11.mlp.experts.linear_fc1.weight3 +Params for bucket 20 (44040192 elements, 44040192 padded size): + module.decoder.layers.11.mlp.experts.linear_fc1.weight2 + module.decoder.layers.11.mlp.experts.linear_fc1.weight1 + module.decoder.layers.11.mlp.experts.linear_fc1.weight0 +Params for bucket 21 (44040192 elements, 44040192 padded size): + module.decoder.layers.10.mlp.experts.linear_fc2.weight7 + module.decoder.layers.10.mlp.experts.linear_fc2.weight3 + module.decoder.layers.10.mlp.experts.linear_fc2.weight5 + 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 +Params for bucket 22 (44040192 elements, 44040192 padded size): + module.decoder.layers.10.mlp.experts.linear_fc2.weight1 + module.decoder.layers.10.mlp.experts.linear_fc2.weight0 + module.decoder.layers.10.mlp.experts.linear_fc1.weight7 + module.decoder.layers.10.mlp.experts.linear_fc1.weight6 +Params for bucket 23 (44040192 elements, 44040192 padded size): + 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.weight4 +Params for bucket 24 (44040192 elements, 44040192 padded size): + 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.weight1 +Params for bucket 25 (44040192 elements, 44040192 padded size): + 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 + 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 +Params for bucket 26 (44040192 elements, 44040192 padded size): + module.decoder.layers.9.mlp.experts.linear_fc2.weight1 + 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.weight6 +Params for bucket 27 (44040192 elements, 44040192 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight5 + module.decoder.layers.9.mlp.experts.linear_fc1.weight3 + module.decoder.layers.9.mlp.experts.linear_fc1.weight4 +Params for bucket 28 (44040192 elements, 44040192 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight2 + module.decoder.layers.9.mlp.experts.linear_fc1.weight0 + module.decoder.layers.9.mlp.experts.linear_fc1.weight1 +Params for bucket 29 (44040192 elements, 44040192 padded size): + module.decoder.layers.8.mlp.experts.linear_fc2.weight7 + module.decoder.layers.8.mlp.experts.linear_fc2.weight5 + module.decoder.layers.8.mlp.experts.linear_fc2.weight3 + module.decoder.layers.8.mlp.experts.linear_fc2.weight6 + module.decoder.layers.8.mlp.experts.linear_fc2.weight4 + module.decoder.layers.8.mlp.experts.linear_fc2.weight2 +Params for bucket 30 (44040192 elements, 44040192 padded size): + 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_fc2.weight0 + module.decoder.layers.8.mlp.experts.linear_fc1.weight6 +Params for bucket 31 (44040192 elements, 44040192 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight5 + module.decoder.layers.8.mlp.experts.linear_fc1.weight4 + module.decoder.layers.8.mlp.experts.linear_fc1.weight3 +Params for bucket 32 (44040192 elements, 44040192 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight2 + module.decoder.layers.8.mlp.experts.linear_fc1.weight0 + module.decoder.layers.8.mlp.experts.linear_fc1.weight1 +Params for bucket 33 (44040192 elements, 44040192 padded size): + module.decoder.layers.7.mlp.experts.linear_fc2.weight7 + module.decoder.layers.7.mlp.experts.linear_fc2.weight3 + module.decoder.layers.7.mlp.experts.linear_fc2.weight5 + module.decoder.layers.7.mlp.experts.linear_fc2.weight4 + module.decoder.layers.7.mlp.experts.linear_fc2.weight6 + module.decoder.layers.7.mlp.experts.linear_fc2.weight2 +Params for bucket 34 (44040192 elements, 44040192 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_fc2.weight0 + module.decoder.layers.7.mlp.experts.linear_fc1.weight7 +Params for bucket 35 (44040192 elements, 44040192 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 36 (44040192 elements, 44040192 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 37 (44040192 elements, 44040192 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.weight2 + 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 38 (44040192 elements, 44040192 padded size): + 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_fc1.weight6 + module.decoder.layers.6.mlp.experts.linear_fc2.weight0 +Params for bucket 39 (44040192 elements, 44040192 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight5 + module.decoder.layers.6.mlp.experts.linear_fc1.weight4 + module.decoder.layers.6.mlp.experts.linear_fc1.weight3 +Params for bucket 40 (44040192 elements, 44040192 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight2 + module.decoder.layers.6.mlp.experts.linear_fc1.weight1 + module.decoder.layers.6.mlp.experts.linear_fc1.weight0 +Params for bucket 41 (44040192 elements, 44040192 padded size): + module.decoder.layers.5.mlp.experts.linear_fc2.weight6 + module.decoder.layers.5.mlp.experts.linear_fc2.weight4 + module.decoder.layers.5.mlp.experts.linear_fc2.weight2 + module.decoder.layers.5.mlp.experts.linear_fc2.weight7 + module.decoder.layers.5.mlp.experts.linear_fc2.weight5 + module.decoder.layers.5.mlp.experts.linear_fc2.weight3 +Params for bucket 42 (44040192 elements, 44040192 padded size): + module.decoder.layers.5.mlp.experts.linear_fc2.weight1 + module.decoder.layers.5.mlp.experts.linear_fc1.weight6 + module.decoder.layers.5.mlp.experts.linear_fc2.weight0 + module.decoder.layers.5.mlp.experts.linear_fc1.weight7 +Params for bucket 43 (44040192 elements, 44040192 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight5 + module.decoder.layers.5.mlp.experts.linear_fc1.weight3 + module.decoder.layers.5.mlp.experts.linear_fc1.weight4 +Params for bucket 44 (44040192 elements, 44040192 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight2 + module.decoder.layers.5.mlp.experts.linear_fc1.weight0 + module.decoder.layers.5.mlp.experts.linear_fc1.weight1 +Params for bucket 45 (44040192 elements, 44040192 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight7 + module.decoder.layers.4.mlp.experts.linear_fc2.weight5 + module.decoder.layers.4.mlp.experts.linear_fc2.weight3 + module.decoder.layers.4.mlp.experts.linear_fc2.weight4 + module.decoder.layers.4.mlp.experts.linear_fc2.weight6 + module.decoder.layers.4.mlp.experts.linear_fc2.weight2 +Params for bucket 46 (44040192 elements, 44040192 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight1 + module.decoder.layers.4.mlp.experts.linear_fc1.weight7 + module.decoder.layers.4.mlp.experts.linear_fc1.weight6 + module.decoder.layers.4.mlp.experts.linear_fc2.weight0 +Params for bucket 47 (44040192 elements, 44040192 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 48 (44040192 elements, 44040192 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight2 + module.decoder.layers.4.mlp.experts.linear_fc1.weight1 + module.decoder.layers.4.mlp.experts.linear_fc1.weight0 +Params for bucket 49 (44040192 elements, 44040192 padded size): + module.decoder.layers.3.mlp.experts.linear_fc2.weight7 + module.decoder.layers.3.mlp.experts.linear_fc2.weight3 + module.decoder.layers.3.mlp.experts.linear_fc2.weight5 + 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.weight2 +Params for bucket 50 (44040192 elements, 44040192 padded size): + module.decoder.layers.3.mlp.experts.linear_fc2.weight1 + module.decoder.layers.3.mlp.experts.linear_fc2.weight0 + module.decoder.layers.3.mlp.experts.linear_fc1.weight7 + module.decoder.layers.3.mlp.experts.linear_fc1.weight6 +Params for bucket 51 (44040192 elements, 44040192 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight5 + module.decoder.layers.3.mlp.experts.linear_fc1.weight3 + module.decoder.layers.3.mlp.experts.linear_fc1.weight4 +Params for bucket 52 (44040192 elements, 44040192 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight2 + module.decoder.layers.3.mlp.experts.linear_fc1.weight0 + module.decoder.layers.3.mlp.experts.linear_fc1.weight1 +Params for bucket 53 (44040192 elements, 44040192 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.weight4 + module.decoder.layers.2.mlp.experts.linear_fc2.weight6 + module.decoder.layers.2.mlp.experts.linear_fc2.weight2 +Params for bucket 54 (44040192 elements, 44040192 padded size): + 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.weight6 + module.decoder.layers.2.mlp.experts.linear_fc2.weight0 +Params for bucket 55 (44040192 elements, 44040192 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight5 + module.decoder.layers.2.mlp.experts.linear_fc1.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight3 +Params for bucket 56 (44040192 elements, 44040192 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight2 + module.decoder.layers.2.mlp.experts.linear_fc1.weight0 + module.decoder.layers.2.mlp.experts.linear_fc1.weight1 +Params for bucket 57 (44040192 elements, 44040192 padded size): + module.decoder.layers.1.mlp.experts.linear_fc2.weight7 + 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 58 (44040192 elements, 44040192 padded size): + 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_fc2.weight0 + module.decoder.layers.1.mlp.experts.linear_fc1.weight7 +Params for bucket 59 (44040192 elements, 44040192 padded size): + 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_fc1.weight4 +Params for bucket 60 (44040192 elements, 44040192 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 61 (44040192 elements, 44040192 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.weight2 + 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 62 (44040192 elements, 44040192 padded size): + module.decoder.layers.0.mlp.experts.linear_fc2.weight1 + module.decoder.layers.0.mlp.experts.linear_fc1.weight6 + module.decoder.layers.0.mlp.experts.linear_fc2.weight0 + module.decoder.layers.0.mlp.experts.linear_fc1.weight7 +Params for bucket 63 (44040192 elements, 44040192 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight5 + module.decoder.layers.0.mlp.experts.linear_fc1.weight3 + module.decoder.layers.0.mlp.experts.linear_fc1.weight4 +Params for bucket 64 (44040192 elements, 44040192 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight2 + module.decoder.layers.0.mlp.experts.linear_fc1.weight1 + 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: 2026-09-17 03:30:19 +> building train, validation, and test datasets ... + > datasets target sizes (minimum size): + train: 160 + 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=(160, 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 68684a515e118253a8572c4517cf2ffe-GPTDataset-train-document_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 68684a515e118253a8572c4517cf2ffe-GPTDataset-train-sample_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 68684a515e118253a8572c4517cf2ffe-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: 2026-09-17 03:30:19 +done with setup ... +(min, max) time across ranks (ms): + model-and-optimizer-setup ......................: (3225.24, 3230.28) + train/valid/test-data-iterators-setup ..........: (108.05, 137.65) +training ... +Overwriting rerun_state_machine.current_iteration from -1 to 0... +[before the start of training step] datetime: 2026-09-17 03:30:19 +/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: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: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.") +/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.") + [2026-09-17 03:30:50] iteration 1/ 10 | consumed samples: 16 | elapsed time per iteration (ms): 31014.6 | throughput per GPU (TFLOP/s/GPU): 2.9 | learning rate: 2.000000E-05 | global batch size: 16 | lm loss: 1.102587E+01 | loss scale: 1.0 | grad norm: 46.862 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 9216 | min safe capacity: 1.1250 | +Number of parameters in transformer block in billions: 23.22 +Number of parameters in embedding layers in billions: 0.41 +Total number of parameters in billions: 23.63 +Number of parameters in most loaded shard in billions: 23.6320 +Theoretical memory footprints: weight and optimizer=169029.43 MB +[Rank 0] (after 1 iterations) memory (MB) | allocated: 58124.50390625 | max allocated: 64453.58203125 | reserved: 80694.0 | max reserved: 80694.0 + [2026-09-17 03:30:51] iteration 2/ 10 | consumed samples: 32 | elapsed time per iteration (ms): 948.0 | throughput per GPU (TFLOP/s/GPU): 95.9 | learning rate: 4.000000E-05 | global batch size: 16 | lm loss: 1.101135E+01 | loss scale: 1.0 | grad norm: 47.659 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 9216 | min safe capacity: 1.1250 | + [2026-09-17 03:30:52] iteration 3/ 10 | consumed samples: 48 | elapsed time per iteration (ms): 389.4 | throughput per GPU (TFLOP/s/GPU): 233.5 | learning rate: 6.000000E-05 | global batch size: 16 | lm loss: 1.024894E+01 | loss scale: 1.0 | grad norm: 187.969 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 9216 | min safe capacity: 1.1250 | +Capture CUDA graph for training!!! +CUDA graph capture done!!! + [2026-09-17 03:30:56] iteration 4/ 10 | consumed samples: 64 | elapsed time per iteration (ms): 3909.6 | throughput per GPU (TFLOP/s/GPU): 23.3 | learning rate: 8.000000E-05 | global batch size: 16 | lm loss: 1.308593E+01 | loss scale: 1.0 | grad norm: 44.378 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 9216 | min safe capacity: 1.1250 | + [2026-09-17 03:30:56] iteration 5/ 10 | consumed samples: 80 | elapsed time per iteration (ms): 250.6 | throughput per GPU (TFLOP/s/GPU): 362.9 | learning rate: 1.000000E-04 | global batch size: 16 | lm loss: 1.040534E+01 | loss scale: 1.0 | grad norm: 20.456 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 9216 | min safe capacity: 1.1250 | + [2026-09-17 03:30:56] iteration 6/ 10 | consumed samples: 96 | elapsed time per iteration (ms): 251.5 | throughput per GPU (TFLOP/s/GPU): 361.6 | learning rate: 7.750000E-05 | global batch size: 16 | lm loss: 1.022187E+01 | loss scale: 1.0 | grad norm: 31.356 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 9216 | min safe capacity: 1.1250 | + [2026-09-17 03:30:59] iteration 7/ 10 | consumed samples: 112 | elapsed time per iteration (ms): 3070.8 | throughput per GPU (TFLOP/s/GPU): 29.6 | learning rate: 3.250000E-05 | global batch size: 16 | lm loss: 9.913305E+00 | loss scale: 1.0 | grad norm: 9.749 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 9216 | min safe capacity: 1.1250 | + [2026-09-17 03:30:59] iteration 8/ 10 | consumed samples: 128 | elapsed time per iteration (ms): 270.2 | throughput per GPU (TFLOP/s/GPU): 336.6 | learning rate: 1.000000E-05 | global batch size: 16 | lm loss: 9.461720E+00 | loss scale: 1.0 | grad norm: 8.259 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 9216 | min safe capacity: 1.1250 | +Capture range started in the application. +Capture range ended in the application. +Generating '/tmp/nsys-report-5702.qdstrm' + [1/1] [0% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-repProcessing events... + [1/1] [0% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [0% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [5% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [7% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [10% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [12% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [14% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [=17% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [==19% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [===22% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [===24% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [====27% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [=====29% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [=====31% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [======34% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [=======36% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [=======37% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [=======38% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [=======39% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [========40% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [========41% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [========42% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [=========43% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [=========44% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [=========45% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [=========46% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [==========47% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [==========48% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [===========50% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [===========51% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [===========53% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [============54% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [============55% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [============56% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [=============58% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [=============59% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [==============63% ] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [========================100%] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep [1/1] [========================100%] G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep +Collection stop is not allowed in this state. +W0917 03:32:40.178000 960711 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers +W0917 03:32:40.179000 960711 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 960827 closing signal SIGTERM +Generated: + /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep +W0917 03:32:40.180000 960711 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 960828 closing signal SIGTERM +W0917 03:32:40.187000 960711 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 960829 closing signal SIGTERM +W0917 03:32:40.189000 960711 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 960830 closing signal SIGTERM +W0917 03:32:40.196000 960711 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 960831 closing signal SIGTERM +W0917 03:32:40.201000 960711 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 960832 closing signal SIGTERM +W0917 03:32:40.214000 960711 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 960833 closing signal SIGTERM +W0917 03:32:40.216000 960711 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 960834 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 960711 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/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/verify_trace_ranks.log b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/verify_trace_ranks.log new file mode 100644 index 0000000000000000000000000000000000000000..68c71a7744d30a806be696cb2613e42b150f4e2e --- /dev/null +++ b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/verify_trace_ranks.log @@ -0,0 +1,6 @@ +report : /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep +kernels : 157846 +devices : 8 of 8 -> [0, 1, 2, 3, 4, 5, 6, 7] +processes: 1 +markers : 26624 ultraep_trace__ kernels on devices [0, 1, 2, 3, 4, 5, 6, 7] +PASS: all ranks present and backend markers emitted diff --git a/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep new file mode 100644 index 0000000000000000000000000000000000000000..b4f60f9bd58f7b9727f86f236f87993c8b13bce8 --- /dev/null +++ b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ccb36de5c1df5199ea5ef7350c20642dfcc3518f1cf7ce829c2b7a31d96b14b2 +size 12191295 diff --git a/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/run.env b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/run.env new file mode 100644 index 0000000000000000000000000000000000000000..24922be79f214e5c3e73442cddf47a7af23a63d2 --- /dev/null +++ b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/run.env @@ -0,0 +1,102 @@ +backend=ultraep +config=G3_K8 +family=tierA +source_model= +standard_num_layers= +base_rung=G3 +repo=/workspace/Megatron-UltraEP +branch=ultraep-bracing-regions-through-kernels +commit=03d57ef7 +num_layers=16 +hidden=4096 +ffn_hidden=14336 +moe_ffn_hidden=1792 +num_experts=64 +topk=8 +experts_per_rank=8 +expert_params=22020096 +heads=32 +kv_groups=8 +kv_channels= +qk_layernorm=0 +shared_ffn=0 +shared_expert_overlap=0 +moe_token_dispatcher=flex +mla=0 +q_lora_rank= +kv_lora_rank=512 +qk_head_dim=128 +qk_pos_emb_head_dim=64 +v_head_dim=128 +dense_first_layers=0 +moe_layers=16 +rotary_base=1000000 +rotary_percent=1.0 +rotary_scaling= +mscale= +mscale_all_dim= +qkv_bias=0 +router_balancing=aux_loss +aux_loss_coeff=1e-2 +router_pre_softmax=0 +router_fusion=0 +router_dtype=fp32 +router_groups= +router_group_topk= +router_score_function=softmax +router_expert_bias=0 +router_bias_update_rate= +router_topk_scaling_factor= +shared_expert_gate=0 +apply_layernorm_1p=0 +attention_output_gate=0 +linear_attention_type= +linear_attention_freq= +linear_conv_kernel_dim= +linear_key_head_dim= +linear_value_head_dim= +linear_num_key_heads= +linear_num_value_heads= +mtp_num_layers=0 +mtp_loss_scaling_factor=0.1 +no_weight_decay_cond_type= +max_position_embeddings=32768 +init_method_std=0.01 +device_initiated_grouped_gemm=1 +norm_epsilon=1e-5 +seq_length=4096 +micro_batch_size=1 +global_batch_size=16 +num_microbatches=2 +assignments_per_rank=32768 +mean_tokens_per_expert=4096 +slot_policy=matched +slots_per_rank=8 +capacity_mode=static +capacity=1.09375 +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_mode=mxfp8 +fp8_recipe=mxfp8 +training_seed=1234 +learning_rate=1e-4 +min_learning_rate=1.0e-5 +lr_warmup_iters=5 +state_dump_enable=0 +state_dump_dir= +moonep_assert_no_pending_reduce=0 diff --git a/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/train.log b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/train.log new file mode 100644 index 0000000000000000000000000000000000000000..8e5057b532affa9632803a1b5bf54c7b806db957 --- /dev/null +++ b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/train.log @@ -0,0 +1,1268 @@ +============================================================== + G3_K8 / ultraep (slots/rank=8 policy=matched) + E=64 K=8 H'=1792 m=8 + L=16 seq=4096 gbs=16 (2 microbatches) + cap_mode=static cap=1.09375 cg=1 recompute=none + memory_dump=none snapshots=/workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/memory_snapshots/snapshot.pickle + repo=/workspace/Megatron-UltraEP @ 03d57ef7 +============================================================== +W0917 03:33:05.052000 969559 torch/distributed/run.py:766] +W0917 03:33:05.052000 969559 torch/distributed/run.py:766] ***************************************** +W0917 03:33:05.052000 969559 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. +W0917 03:33:05.052000 969559 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 .................................... 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 ............................... 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 .............................. 16 + 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 ................................. 14336 + 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 ............................... 16 + 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 ..................................... 4096 + 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 ............................. 1792 + 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 ..................... 1.09375 + 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 .......................... False + moe_router_score_function ....................... softmax + moe_router_topk ................................. 8 + moe_router_topk_scaling_factor .................. None + moe_shared_expert_intermediate_size ............. None + 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-05 + normalization ................................... RMSNorm + num_attention_heads ............................. 32 + num_channels .................................... 3 + num_classes ..................................... 1000 + num_dataset_builder_threads ..................... 1 + num_distributed_optimizer_instances ............. 1 + num_experts ..................................... 64 + num_layers ...................................... 16 + num_layers_at_end_in_bf16 ....................... 1 + num_layers_at_start_in_bf16 ..................... 1 + num_layers_per_virtual_pipeline_stage ........... None + num_query_groups ................................ 8 + 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 2 +> 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") +/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 ... +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") +[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 6 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 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 7 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 [Gloo] Rank 5 is connected to 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 +[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 +> initialized tensor model parallel with size 1 +> initialized pipeline model parallel with size 1 +> setting random seeds to 1234 ... +[rank5]:[W917 03:33:17.749694598 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]:[W917 03:33:17.749736865 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]:[W917 03:33:17.749771189 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]:[W917 03:33:17.749792831 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. +[rank7]:[W917 03:33:17.753345276 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. +[rank4]:[W917 03:33:17.754063808 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]:[W917 03:33:17.754496723 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. +> 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]:[W917 03:33:18.359796766 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: 3.438 seconds +time to initialize megatron (seconds): 8.825 +[after megatron is initialized] datetime: 2026-09-17 03:33:23 +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): 6724653056 +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: 18 +Params for bucket 1 (206045184 elements, 206045184 padded size): + module.output_layer.weight +Params for bucket 2 (42217472 elements, 42217472 padded size): + module.decoder.final_layernorm.weight + module.decoder.layers.15.self_attention.linear_qkv.weight + module.decoder.layers.15.pre_mlp_layernorm.weight + module.decoder.layers.15.mlp.router.weight + module.decoder.layers.15.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.15.self_attention.linear_proj.weight +Params for bucket 3 (42213376 elements, 42213376 padded size): + module.decoder.layers.14.self_attention.linear_qkv.weight + module.decoder.layers.14.pre_mlp_layernorm.weight + module.decoder.layers.14.mlp.router.weight + module.decoder.layers.14.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.14.self_attention.linear_proj.weight +Params for bucket 4 (42213376 elements, 42213376 padded size): + module.decoder.layers.13.pre_mlp_layernorm.weight + module.decoder.layers.13.self_attention.linear_qkv.weight + module.decoder.layers.13.mlp.router.weight + module.decoder.layers.13.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.13.self_attention.linear_proj.weight +Params for bucket 5 (42213376 elements, 42213376 padded size): + module.decoder.layers.12.mlp.router.weight + module.decoder.layers.12.self_attention.linear_proj.weight + module.decoder.layers.12.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.12.pre_mlp_layernorm.weight + module.decoder.layers.12.self_attention.linear_qkv.weight +Params for bucket 6 (42213376 elements, 42213376 padded size): + module.decoder.layers.11.mlp.router.weight + module.decoder.layers.11.self_attention.linear_proj.weight + module.decoder.layers.11.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.11.self_attention.linear_qkv.weight + module.decoder.layers.11.pre_mlp_layernorm.weight +Params for bucket 7 (42213376 elements, 42213376 padded size): + module.decoder.layers.10.mlp.router.weight + module.decoder.layers.10.self_attention.linear_proj.weight + module.decoder.layers.10.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.10.pre_mlp_layernorm.weight + module.decoder.layers.10.self_attention.linear_qkv.weight +Params for bucket 8 (42213376 elements, 42213376 padded size): + module.decoder.layers.9.pre_mlp_layernorm.weight + module.decoder.layers.9.self_attention.linear_qkv.weight + module.decoder.layers.9.mlp.router.weight + module.decoder.layers.9.self_attention.linear_proj.weight + module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight +Params for bucket 9 (42213376 elements, 42213376 padded size): + module.decoder.layers.8.self_attention.linear_qkv.weight + module.decoder.layers.8.pre_mlp_layernorm.weight + module.decoder.layers.8.mlp.router.weight + module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.8.self_attention.linear_proj.weight +Params for bucket 10 (42213376 elements, 42213376 padded size): + module.decoder.layers.7.mlp.router.weight + module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.7.self_attention.linear_proj.weight + module.decoder.layers.7.self_attention.linear_qkv.weight + module.decoder.layers.7.pre_mlp_layernorm.weight +Params for bucket 11 (42213376 elements, 42213376 padded size): + module.decoder.layers.6.mlp.router.weight + module.decoder.layers.6.self_attention.linear_proj.weight + module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.6.self_attention.linear_qkv.weight + module.decoder.layers.6.pre_mlp_layernorm.weight +Params for bucket 12 (42213376 elements, 42213376 padded size): + module.decoder.layers.5.self_attention.linear_qkv.weight + module.decoder.layers.5.self_attention.linear_proj.weight + module.decoder.layers.5.pre_mlp_layernorm.weight + module.decoder.layers.5.mlp.router.weight + module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight +Params for bucket 13 (42213376 elements, 42213376 padded size): + module.decoder.layers.4.self_attention.linear_qkv.weight + module.decoder.layers.4.pre_mlp_layernorm.weight + module.decoder.layers.4.mlp.router.weight + module.decoder.layers.4.self_attention.linear_proj.weight + module.decoder.layers.4.self_attention.linear_qkv.layer_norm_weight +Params for bucket 14 (42213376 elements, 42213376 padded size): + module.decoder.layers.3.mlp.router.weight + module.decoder.layers.3.self_attention.linear_proj.weight + module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.3.pre_mlp_layernorm.weight + module.decoder.layers.3.self_attention.linear_qkv.weight +Params for bucket 15 (42213376 elements, 42213376 padded size): + module.decoder.layers.2.mlp.router.weight + module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.2.self_attention.linear_proj.weight + module.decoder.layers.2.self_attention.linear_qkv.weight + module.decoder.layers.2.pre_mlp_layernorm.weight +Params for bucket 16 (42213376 elements, 42213376 padded size): + module.decoder.layers.1.mlp.router.weight + module.decoder.layers.1.self_attention.linear_proj.weight + module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.1.self_attention.linear_qkv.weight + module.decoder.layers.1.pre_mlp_layernorm.weight +Params for bucket 17 (42213376 elements, 42213376 padded size): + module.decoder.layers.0.pre_mlp_layernorm.weight + module.decoder.layers.0.self_attention.linear_proj.weight + module.decoder.layers.0.self_attention.linear_qkv.weight + module.decoder.layers.0.mlp.router.weight + module.decoder.layers.0.self_attention.linear_qkv.layer_norm_weight +Params for bucket 18 (206045184 elements, 206045184 padded size): + module.embedding.word_embeddings.weight +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 64 +Params for bucket 1 (44040192 elements, 44040192 padded size): + module.decoder.layers.15.mlp.experts.linear_fc2.weight6 + module.decoder.layers.15.mlp.experts.linear_fc2.weight4 + module.decoder.layers.15.mlp.experts.linear_fc2.weight2 + module.decoder.layers.15.mlp.experts.linear_fc2.weight7 + module.decoder.layers.15.mlp.experts.linear_fc2.weight5 + module.decoder.layers.15.mlp.experts.linear_fc2.weight3 +Params for bucket 2 (44040192 elements, 44040192 padded size): + 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_fc2.weight0 + module.decoder.layers.15.mlp.experts.linear_fc1.weight7 +Params for bucket 3 (44040192 elements, 44040192 padded size): + 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_fc1.weight4 +Params for bucket 4 (44040192 elements, 44040192 padded size): + module.decoder.layers.15.mlp.experts.linear_fc1.weight2 + module.decoder.layers.15.mlp.experts.linear_fc1.weight0 + module.decoder.layers.15.mlp.experts.linear_fc1.weight1 +Params for bucket 5 (44040192 elements, 44040192 padded size): + 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.weight7 + module.decoder.layers.14.mlp.experts.linear_fc2.weight3 + module.decoder.layers.14.mlp.experts.linear_fc2.weight5 +Params for bucket 6 (44040192 elements, 44040192 padded size): + module.decoder.layers.14.mlp.experts.linear_fc2.weight1 + module.decoder.layers.14.mlp.experts.linear_fc2.weight0 + module.decoder.layers.14.mlp.experts.linear_fc1.weight6 + module.decoder.layers.14.mlp.experts.linear_fc1.weight7 +Params for bucket 7 (44040192 elements, 44040192 padded size): + module.decoder.layers.14.mlp.experts.linear_fc1.weight5 + module.decoder.layers.14.mlp.experts.linear_fc1.weight4 + module.decoder.layers.14.mlp.experts.linear_fc1.weight3 +Params for bucket 8 (44040192 elements, 44040192 padded size): + module.decoder.layers.14.mlp.experts.linear_fc1.weight2 + module.decoder.layers.14.mlp.experts.linear_fc1.weight1 + module.decoder.layers.14.mlp.experts.linear_fc1.weight0 +Params for bucket 9 (44040192 elements, 44040192 padded size): + module.decoder.layers.13.mlp.experts.linear_fc2.weight7 + module.decoder.layers.13.mlp.experts.linear_fc2.weight5 + module.decoder.layers.13.mlp.experts.linear_fc2.weight3 + module.decoder.layers.13.mlp.experts.linear_fc2.weight4 + module.decoder.layers.13.mlp.experts.linear_fc2.weight6 + module.decoder.layers.13.mlp.experts.linear_fc2.weight2 +Params for bucket 10 (44040192 elements, 44040192 padded size): + module.decoder.layers.13.mlp.experts.linear_fc2.weight1 + module.decoder.layers.13.mlp.experts.linear_fc2.weight0 + module.decoder.layers.13.mlp.experts.linear_fc1.weight6 + module.decoder.layers.13.mlp.experts.linear_fc1.weight7 +Params for bucket 11 (44040192 elements, 44040192 padded size): + 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.weight4 +Params for bucket 12 (44040192 elements, 44040192 padded size): + module.decoder.layers.13.mlp.experts.linear_fc1.weight2 + module.decoder.layers.13.mlp.experts.linear_fc1.weight0 + module.decoder.layers.13.mlp.experts.linear_fc1.weight1 +Params for bucket 13 (44040192 elements, 44040192 padded size): + module.decoder.layers.12.mlp.experts.linear_fc2.weight6 + module.decoder.layers.12.mlp.experts.linear_fc2.weight4 + module.decoder.layers.12.mlp.experts.linear_fc2.weight2 + module.decoder.layers.12.mlp.experts.linear_fc2.weight7 + module.decoder.layers.12.mlp.experts.linear_fc2.weight5 + module.decoder.layers.12.mlp.experts.linear_fc2.weight3 +Params for bucket 14 (44040192 elements, 44040192 padded size): + module.decoder.layers.12.mlp.experts.linear_fc2.weight1 + module.decoder.layers.12.mlp.experts.linear_fc1.weight7 + module.decoder.layers.12.mlp.experts.linear_fc1.weight6 + module.decoder.layers.12.mlp.experts.linear_fc2.weight0 +Params for bucket 15 (44040192 elements, 44040192 padded size): + module.decoder.layers.12.mlp.experts.linear_fc1.weight5 + module.decoder.layers.12.mlp.experts.linear_fc1.weight4 + module.decoder.layers.12.mlp.experts.linear_fc1.weight3 +Params for bucket 16 (44040192 elements, 44040192 padded size): + module.decoder.layers.12.mlp.experts.linear_fc1.weight2 + module.decoder.layers.12.mlp.experts.linear_fc1.weight1 + module.decoder.layers.12.mlp.experts.linear_fc1.weight0 +Params for bucket 17 (44040192 elements, 44040192 padded size): + module.decoder.layers.11.mlp.experts.linear_fc2.weight7 + module.decoder.layers.11.mlp.experts.linear_fc2.weight5 + module.decoder.layers.11.mlp.experts.linear_fc2.weight3 + module.decoder.layers.11.mlp.experts.linear_fc2.weight4 + module.decoder.layers.11.mlp.experts.linear_fc2.weight6 + module.decoder.layers.11.mlp.experts.linear_fc2.weight2 +Params for bucket 18 (44040192 elements, 44040192 padded size): + module.decoder.layers.11.mlp.experts.linear_fc2.weight1 + module.decoder.layers.11.mlp.experts.linear_fc1.weight7 + module.decoder.layers.11.mlp.experts.linear_fc1.weight6 + module.decoder.layers.11.mlp.experts.linear_fc2.weight0 +Params for bucket 19 (44040192 elements, 44040192 padded size): + module.decoder.layers.11.mlp.experts.linear_fc1.weight5 + module.decoder.layers.11.mlp.experts.linear_fc1.weight4 + module.decoder.layers.11.mlp.experts.linear_fc1.weight3 +Params for bucket 20 (44040192 elements, 44040192 padded size): + module.decoder.layers.11.mlp.experts.linear_fc1.weight2 + module.decoder.layers.11.mlp.experts.linear_fc1.weight1 + module.decoder.layers.11.mlp.experts.linear_fc1.weight0 +Params for bucket 21 (44040192 elements, 44040192 padded size): + module.decoder.layers.10.mlp.experts.linear_fc2.weight7 + module.decoder.layers.10.mlp.experts.linear_fc2.weight3 + module.decoder.layers.10.mlp.experts.linear_fc2.weight5 + 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 +Params for bucket 22 (44040192 elements, 44040192 padded size): + module.decoder.layers.10.mlp.experts.linear_fc2.weight1 + module.decoder.layers.10.mlp.experts.linear_fc2.weight0 + module.decoder.layers.10.mlp.experts.linear_fc1.weight7 + module.decoder.layers.10.mlp.experts.linear_fc1.weight6 +Params for bucket 23 (44040192 elements, 44040192 padded size): + 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.weight4 +Params for bucket 24 (44040192 elements, 44040192 padded size): + 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.weight1 +Params for bucket 25 (44040192 elements, 44040192 padded size): + 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 + 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 +Params for bucket 26 (44040192 elements, 44040192 padded size): + module.decoder.layers.9.mlp.experts.linear_fc2.weight1 + 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.weight6 +Params for bucket 27 (44040192 elements, 44040192 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight5 + module.decoder.layers.9.mlp.experts.linear_fc1.weight3 + module.decoder.layers.9.mlp.experts.linear_fc1.weight4 +Params for bucket 28 (44040192 elements, 44040192 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight2 + module.decoder.layers.9.mlp.experts.linear_fc1.weight0 + module.decoder.layers.9.mlp.experts.linear_fc1.weight1 +Params for bucket 29 (44040192 elements, 44040192 padded size): + module.decoder.layers.8.mlp.experts.linear_fc2.weight4 + module.decoder.layers.8.mlp.experts.linear_fc2.weight6 + module.decoder.layers.8.mlp.experts.linear_fc2.weight2 + module.decoder.layers.8.mlp.experts.linear_fc2.weight7 + module.decoder.layers.8.mlp.experts.linear_fc2.weight5 + module.decoder.layers.8.mlp.experts.linear_fc2.weight3 +Params for bucket 30 (44040192 elements, 44040192 padded size): + 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_fc2.weight0 + module.decoder.layers.8.mlp.experts.linear_fc1.weight6 +Params for bucket 31 (44040192 elements, 44040192 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight5 + module.decoder.layers.8.mlp.experts.linear_fc1.weight4 + module.decoder.layers.8.mlp.experts.linear_fc1.weight3 +Params for bucket 32 (44040192 elements, 44040192 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight2 + module.decoder.layers.8.mlp.experts.linear_fc1.weight0 + module.decoder.layers.8.mlp.experts.linear_fc1.weight1 +Params for bucket 33 (44040192 elements, 44040192 padded size): + module.decoder.layers.7.mlp.experts.linear_fc2.weight4 + module.decoder.layers.7.mlp.experts.linear_fc2.weight6 + module.decoder.layers.7.mlp.experts.linear_fc2.weight2 + module.decoder.layers.7.mlp.experts.linear_fc2.weight7 + module.decoder.layers.7.mlp.experts.linear_fc2.weight5 + module.decoder.layers.7.mlp.experts.linear_fc2.weight3 +Params for bucket 34 (44040192 elements, 44040192 padded size): + module.decoder.layers.7.mlp.experts.linear_fc2.weight1 + module.decoder.layers.7.mlp.experts.linear_fc1.weight7 + module.decoder.layers.7.mlp.experts.linear_fc1.weight6 + module.decoder.layers.7.mlp.experts.linear_fc2.weight0 +Params for bucket 35 (44040192 elements, 44040192 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight5 + module.decoder.layers.7.mlp.experts.linear_fc1.weight4 + module.decoder.layers.7.mlp.experts.linear_fc1.weight3 +Params for bucket 36 (44040192 elements, 44040192 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 37 (44040192 elements, 44040192 padded size): + 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 + 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 +Params for bucket 38 (44040192 elements, 44040192 padded size): + 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.weight0 + module.decoder.layers.6.mlp.experts.linear_fc1.weight7 +Params for bucket 39 (44040192 elements, 44040192 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight5 + module.decoder.layers.6.mlp.experts.linear_fc1.weight3 + module.decoder.layers.6.mlp.experts.linear_fc1.weight4 +Params for bucket 40 (44040192 elements, 44040192 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight2 + module.decoder.layers.6.mlp.experts.linear_fc1.weight0 + module.decoder.layers.6.mlp.experts.linear_fc1.weight1 +Params for bucket 41 (44040192 elements, 44040192 padded size): + module.decoder.layers.5.mlp.experts.linear_fc2.weight7 + module.decoder.layers.5.mlp.experts.linear_fc2.weight3 + module.decoder.layers.5.mlp.experts.linear_fc2.weight5 + module.decoder.layers.5.mlp.experts.linear_fc2.weight6 + module.decoder.layers.5.mlp.experts.linear_fc2.weight4 + module.decoder.layers.5.mlp.experts.linear_fc2.weight2 +Params for bucket 42 (44040192 elements, 44040192 padded size): + module.decoder.layers.5.mlp.experts.linear_fc2.weight1 + module.decoder.layers.5.mlp.experts.linear_fc1.weight7 + module.decoder.layers.5.mlp.experts.linear_fc2.weight0 + module.decoder.layers.5.mlp.experts.linear_fc1.weight6 +Params for bucket 43 (44040192 elements, 44040192 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight5 + module.decoder.layers.5.mlp.experts.linear_fc1.weight4 + module.decoder.layers.5.mlp.experts.linear_fc1.weight3 +Params for bucket 44 (44040192 elements, 44040192 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight2 + module.decoder.layers.5.mlp.experts.linear_fc1.weight0 + module.decoder.layers.5.mlp.experts.linear_fc1.weight1 +Params for bucket 45 (44040192 elements, 44040192 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight7 + 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.weight4 + module.decoder.layers.4.mlp.experts.linear_fc2.weight6 + module.decoder.layers.4.mlp.experts.linear_fc2.weight2 +Params for bucket 46 (44040192 elements, 44040192 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 47 (44040192 elements, 44040192 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight5 + module.decoder.layers.4.mlp.experts.linear_fc1.weight3 + module.decoder.layers.4.mlp.experts.linear_fc1.weight4 +Params for bucket 48 (44040192 elements, 44040192 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 49 (44040192 elements, 44040192 padded size): + module.decoder.layers.3.mlp.experts.linear_fc2.weight7 + module.decoder.layers.3.mlp.experts.linear_fc2.weight3 + module.decoder.layers.3.mlp.experts.linear_fc2.weight5 + 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.weight2 +Params for bucket 50 (44040192 elements, 44040192 padded size): + module.decoder.layers.3.mlp.experts.linear_fc2.weight1 + module.decoder.layers.3.mlp.experts.linear_fc2.weight0 + module.decoder.layers.3.mlp.experts.linear_fc1.weight7 + module.decoder.layers.3.mlp.experts.linear_fc1.weight6 +Params for bucket 51 (44040192 elements, 44040192 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight5 + module.decoder.layers.3.mlp.experts.linear_fc1.weight3 + module.decoder.layers.3.mlp.experts.linear_fc1.weight4 +Params for bucket 52 (44040192 elements, 44040192 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight2 + module.decoder.layers.3.mlp.experts.linear_fc1.weight0 + module.decoder.layers.3.mlp.experts.linear_fc1.weight1 +Params for bucket 53 (44040192 elements, 44040192 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.weight4 + module.decoder.layers.2.mlp.experts.linear_fc2.weight6 + module.decoder.layers.2.mlp.experts.linear_fc2.weight2 +Params for bucket 54 (44040192 elements, 44040192 padded size): + 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.weight6 + module.decoder.layers.2.mlp.experts.linear_fc2.weight0 +Params for bucket 55 (44040192 elements, 44040192 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight5 + module.decoder.layers.2.mlp.experts.linear_fc1.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight3 +Params for bucket 56 (44040192 elements, 44040192 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight2 + module.decoder.layers.2.mlp.experts.linear_fc1.weight0 + module.decoder.layers.2.mlp.experts.linear_fc1.weight1 +Params for bucket 57 (44040192 elements, 44040192 padded size): + module.decoder.layers.1.mlp.experts.linear_fc2.weight7 + 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 58 (44040192 elements, 44040192 padded size): + module.decoder.layers.1.mlp.experts.linear_fc2.weight1 + module.decoder.layers.1.mlp.experts.linear_fc2.weight0 + module.decoder.layers.1.mlp.experts.linear_fc1.weight7 + module.decoder.layers.1.mlp.experts.linear_fc1.weight6 +Params for bucket 59 (44040192 elements, 44040192 padded size): + 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_fc1.weight4 +Params for bucket 60 (44040192 elements, 44040192 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 61 (44040192 elements, 44040192 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.weight2 + 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 62 (44040192 elements, 44040192 padded size): + module.decoder.layers.0.mlp.experts.linear_fc2.weight1 + module.decoder.layers.0.mlp.experts.linear_fc1.weight6 + module.decoder.layers.0.mlp.experts.linear_fc2.weight0 + module.decoder.layers.0.mlp.experts.linear_fc1.weight7 +Params for bucket 63 (44040192 elements, 44040192 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight5 + module.decoder.layers.0.mlp.experts.linear_fc1.weight4 + module.decoder.layers.0.mlp.experts.linear_fc1.weight3 +Params for bucket 64 (44040192 elements, 44040192 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight2 + module.decoder.layers.0.mlp.experts.linear_fc1.weight0 + module.decoder.layers.0.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: 2026-09-17 03:33:27 +> building train, validation, and test datasets ... + > datasets target sizes (minimum size): + train: 160 + 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=(160, 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 68684a515e118253a8572c4517cf2ffe-GPTDataset-train-document_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 68684a515e118253a8572c4517cf2ffe-GPTDataset-train-sample_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 68684a515e118253a8572c4517cf2ffe-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: 2026-09-17 03:33:27 +done with setup ... +(min, max) time across ranks (ms): + model-and-optimizer-setup ......................: (3188.55, 3194.23) + train/valid/test-data-iterators-setup ..........: (107.22, 135.77) +training ... +Overwriting rerun_state_machine.current_iteration from -1 to 0... +[before the start of training step] datetime: 2026-09-17 03:33:27 +/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: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: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: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: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.") +/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.") + [2026-09-17 03:33:57] iteration 1/ 10 | consumed samples: 16 | elapsed time per iteration (ms): 30515.0 | throughput per GPU (TFLOP/s/GPU): 6.4 | learning rate: 2.000000E-05 | global batch size: 16 | lm loss: 1.102374E+01 | loss scale: 1.0 | grad norm: 44.521 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33920 | min safe capacity: 1.0352 | +Number of parameters in transformer block in billions: 23.22 +Number of parameters in embedding layers in billions: 0.41 +Total number of parameters in billions: 23.63 +Number of parameters in most loaded shard in billions: 23.6320 +Theoretical memory footprints: weight and optimizer=169029.43 MB +[Rank 0] (after 1 iterations) memory (MB) | allocated: 58125.50390625 | max allocated: 69789.0546875 | reserved: 76118.0 | max reserved: 76118.0 + [2026-09-17 03:33:58] iteration 2/ 10 | consumed samples: 32 | elapsed time per iteration (ms): 1083.3 | throughput per GPU (TFLOP/s/GPU): 179.9 | learning rate: 4.000000E-05 | global batch size: 16 | lm loss: 1.101671E+01 | loss scale: 1.0 | grad norm: 46.188 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33920 | min safe capacity: 1.0352 | + [2026-09-17 03:33:59] iteration 3/ 10 | consumed samples: 48 | elapsed time per iteration (ms): 437.1 | throughput per GPU (TFLOP/s/GPU): 445.7 | learning rate: 6.000000E-05 | global batch size: 16 | lm loss: 1.046371E+01 | loss scale: 1.0 | grad norm: 211.813 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33920 | min safe capacity: 1.0352 | +Capture CUDA graph for training!!! +CUDA graph capture done!!! + [2026-09-17 03:34:02] iteration 4/ 10 | consumed samples: 64 | elapsed time per iteration (ms): 3593.9 | throughput per GPU (TFLOP/s/GPU): 54.2 | learning rate: 8.000000E-05 | global batch size: 16 | lm loss: 1.314799E+01 | loss scale: 1.0 | grad norm: 39.078 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33920 | min safe capacity: 1.0352 | + [2026-09-17 03:34:03] iteration 5/ 10 | consumed samples: 80 | elapsed time per iteration (ms): 360.8 | throughput per GPU (TFLOP/s/GPU): 540.0 | learning rate: 1.000000E-04 | global batch size: 16 | lm loss: 1.038756E+01 | loss scale: 1.0 | grad norm: 17.511 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33920 | min safe capacity: 1.0352 | + [2026-09-17 03:34:03] iteration 6/ 10 | consumed samples: 96 | elapsed time per iteration (ms): 383.9 | throughput per GPU (TFLOP/s/GPU): 507.6 | learning rate: 7.750000E-05 | global batch size: 16 | lm loss: 1.033255E+01 | loss scale: 1.0 | grad norm: 29.428 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34176 | min safe capacity: 1.0430 | + [2026-09-17 03:34:06] iteration 7/ 10 | consumed samples: 112 | elapsed time per iteration (ms): 3303.0 | throughput per GPU (TFLOP/s/GPU): 59.0 | learning rate: 3.250000E-05 | global batch size: 16 | lm loss: 1.003726E+01 | loss scale: 1.0 | grad norm: 7.175 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34176 | min safe capacity: 1.0430 | + [2026-09-17 03:34:07] iteration 8/ 10 | consumed samples: 128 | elapsed time per iteration (ms): 371.2 | throughput per GPU (TFLOP/s/GPU): 524.8 | learning rate: 1.000000E-05 | global batch size: 16 | lm loss: 9.514786E+00 | loss scale: 1.0 | grad norm: 5.913 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34176 | min safe capacity: 1.0430 | +Capture range started in the application. +Capture range ended in the application. +Generating '/tmp/nsys-report-2835.qdstrm' + [1/1] [0% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-repProcessing events... + [1/1] [0% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [0% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [5% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [7% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [9% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [12% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [14% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [=16% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [==19% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [==21% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [===23% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [====26% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [====28% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [=====30% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [======33% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [======34% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [======35% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [=======36% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [=======37% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [=======38% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [=======39% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [========40% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [========41% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [========42% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [=========43% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [=========45% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [=========46% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [==========47% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [==========48% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [===========50% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [===========51% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [===========52% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [===========53% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [============54% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [============56% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [============57% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [==============61% ] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [========================100%] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep [1/1] [========================100%] G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep diff --git a/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/results.tsv b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/results.tsv new file mode 100644 index 0000000000000000000000000000000000000000..d9690c679d3d751b83465f63e6318c6df1d2db06 --- /dev/null +++ b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/results.tsv @@ -0,0 +1 @@ +config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir diff --git a/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/trace_extension_build.log b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/trace_extension_build.log index df8efc98865c41730de31b776796cf5b52e63287..da4f04142c786e1726c2ba802c5d74ba16b2465d 100644 --- a/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/trace_extension_build.log +++ b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/trace_extension_build.log @@ -66,3 +66,9 @@ Emitting ninja build file /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cp Compiling objects... Allowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N) [1/2] c++ -MMD -MF /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cpython-312/megatron/core/transformer/moe/ultraep_trace_ops.o.d -fno-strict-overflow -Wsign-compare -DNDEBUG -g -O2 -Wall -fPIC -I/usr/local/lib/python3.12/dist-packages/torch/include -I/usr/local/lib/python3.12/dist-packages/torch/include/torch/csrc/api/include -I/usr/local/cuda/include -I/usr/include/python3.12 -c -c /workspace/Megatron-UltraEP/megatron/core/transformer/moe/ultraep_trace_ops.cpp -o /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cpython-312/megatron/core/transformer/moe/ultraep_trace_ops.o -O3 -std=c++17 -DTORCH_API_INCLUDE_EXTENSION_H '-DPYBIND11_COMPILER_TYPE="_gcc"' '-DPYBIND11_STDLIB="_libstdcpp"' '-DPYBIND11_BUILD_ABI="_cxxabi1016"' -DTORCH_EXTENSION_NAME=_ultraep_trace_C +[2/2] /usr/local/cuda/bin/nvcc --generate-dependencies-with-compile --dependency-output /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cpython-312/megatron/core/transformer/moe/ultraep_trace_kernels.o.d -I/usr/local/lib/python3.12/dist-packages/torch/include -I/usr/local/lib/python3.12/dist-packages/torch/include/torch/csrc/api/include -I/usr/local/cuda/include -I/usr/include/python3.12 -c -c /workspace/Megatron-UltraEP/megatron/core/transformer/moe/ultraep_trace_kernels.cu -o /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cpython-312/megatron/core/transformer/moe/ultraep_trace_kernels.o -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr --compiler-options ''"'"'-fPIC'"'"'' -O3 -std=c++17 -lineinfo -DTORCH_API_INCLUDE_EXTENSION_H '-DPYBIND11_COMPILER_TYPE="_gcc"' '-DPYBIND11_STDLIB="_libstdcpp"' '-DPYBIND11_BUILD_ABI="_cxxabi1016"' -DTORCH_EXTENSION_NAME=_ultraep_trace_C -gencode=arch=compute_100,code=sm_100 -gencode=arch=compute_120,code=compute_120 -gencode=arch=compute_120,code=sm_120 -gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_80,code=sm_80 -gencode=arch=compute_86,code=sm_86 -gencode=arch=compute_90,code=sm_90 +creating build/lib.linux-x86_64-cpython-312/megatron/core/transformer/moe +x86_64-linux-gnu-g++ -fno-strict-overflow -Wsign-compare -DNDEBUG -g -O2 -Wall -shared -Wl,-O1 -Wl,-Bsymbolic-functions /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cpython-312/megatron/core/transformer/moe/ultraep_trace_kernels.o /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cpython-312/megatron/core/transformer/moe/ultraep_trace_ops.o -L/usr/local/lib/python3.12/dist-packages/torch/lib -L/usr/local/cuda/lib64 -L/usr/lib/x86_64-linux-gnu -lc10 -ltorch -ltorch_cpu -ltorch_python -lcudart -lc10_cuda -ltorch_cuda -o build/lib.linux-x86_64-cpython-312/megatron/core/transformer/moe/_ultraep_trace_C.cpython-312-x86_64-linux-gnu.so +copying build/lib.linux-x86_64-cpython-312/megatron/core/datasets/helpers_cpp.cpython-312-x86_64-linux-gnu.so -> megatron/core/datasets +copying build/lib.linux-x86_64-cpython-312/megatron/core/transformer/moe/_ultraep_trace_C.cpython-312-x86_64-linux-gnu.so -> megatron/core/transformer/moe +load check passed: /workspace/Megatron-UltraEP/megatron/core/transformer/moe/_ultraep_trace_C.cpython-312-x86_64-linux-gnu.so -> torch.ops.megatron_ultraep_trace.fwd_ultraep_forward_begin diff --git a/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/trace_results.tsv b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/trace_results.tsv new file mode 100644 index 0000000000000000000000000000000000000000..fb0db3b164187208f74a23f3b63487dd92fad1cf --- /dev/null +++ b/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/trace_results.tsv @@ -0,0 +1,3 @@ +model backend capacity status report async_overlap_json +G2_K2 ultraep 1.15625 OK /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep +G3_K2 ultraep 1.1875 OK /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep