diff --git a/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep1/mem_samples.csv b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep1/mem_samples.csv new file mode 100644 index 0000000000000000000000000000000000000000..3627601769063c769b683b73dd473fcd4b55647d --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep1/mem_samples.csv @@ -0,0 +1,232 @@ +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 1 +1, 1 +2, 1 +3, 1 +4, 1 +5, 1 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 1 +5, 1 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 4 +5, 4 +6, 4 +7, 4 +0, 4 +1, 4 +2, 4 +3, 4 +4, 4 +5, 4 +6, 4 +7, 4 +0, 821 +1, 821 +2, 821 +3, 821 +4, 821 +5, 821 +6, 821 +7, 821 +0, 1955 +1, 1955 +2, 1955 +3, 1955 +4, 1955 +5, 1955 +6, 1955 +7, 1955 +0, 2353 +1, 2373 +2, 2357 +3, 2349 +4, 2337 +5, 2359 +6, 2353 +7, 2329 +0, 25283 +1, 25349 +2, 25327 +3, 25327 +4, 25349 +5, 25371 +6, 25393 +7, 25437 +0, 28875 +1, 28993 +2, 28799 +3, 28767 +4, 28823 +5, 28787 +6, 28907 +7, 28937 +0, 39085 +1, 39167 +2, 39019 +3, 39015 +4, 38981 +5, 39009 +6, 39115 +7, 39135 +0, 39773 +1, 39857 +2, 39707 +3, 39705 +4, 39705 +5, 39703 +6, 39839 +7, 39857 +0, 39773 +1, 39857 +2, 39707 +3, 39705 +4, 39705 +5, 39703 +6, 39839 +7, 39857 +0, 39783 +1, 39867 +2, 39717 +3, 39717 +4, 39717 +5, 39713 +6, 39849 +7, 39867 +0, 39783 +1, 39867 +2, 39717 +3, 39717 +4, 39717 +5, 39713 +6, 39849 +7, 39867 +0, 109705 +1, 109749 +2, 110123 +3, 110121 +4, 110057 +5, 110065 +6, 109739 +7, 109747 +0, 110101 +1, 110187 +2, 110513 +3, 110511 +4, 110511 +5, 110509 +6, 110167 +7, 110187 +0, 110105 +1, 110191 +2, 110517 +3, 110515 +4, 110515 +5, 110513 +6, 110171 +7, 110191 +0, 112457 +1, 112501 +2, 110755 +3, 110775 +4, 110731 +5, 110741 +6, 112459 +7, 112477 +0, 113191 +1, 113277 +2, 111555 +3, 111553 +4, 111553 +5, 111551 +6, 113257 +7, 113277 +0, 113921 +1, 113969 +2, 112277 +3, 112239 +4, 112193 +5, 112225 +6, 113921 +7, 113943 +0, 124037 +1, 126189 +2, 130157 +3, 128095 +4, 134269 +5, 128093 +6, 128229 +7, 134419 +0, 142137 +1, 142231 +2, 142081 +3, 142079 +4, 142075 +5, 142077 +6, 142211 +7, 142231 +0, 142137 +1, 142231 +2, 142081 +3, 142079 +4, 142075 +5, 142077 +6, 142211 +7, 142231 +0, 142137 +1, 142231 +2, 142081 +3, 142079 +4, 142075 +5, 142077 +6, 142211 +7, 142231 +0, 138363 +1, 138521 +2, 138953 +3, 138951 +4, 138947 +5, 138687 +6, 139083 +7, 138777 +0, 136941 +1, 137099 +2, 137531 +3, 137267 +4, 137525 +5, 137265 +6, 137661 +7, 137355 +0, 1 +1, 3 +2, 4 +3, 4 +4, 4 +5, 68 +6, 137013 +7, 324 diff --git a/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep1/run.env b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep1/run.env new file mode 100644 index 0000000000000000000000000000000000000000..8907152eacecee56da306170fa30edf972163c8e --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep1/run.env @@ -0,0 +1,61 @@ +backend=ultraep +config=DSV2_LITE_L24 +family=tierB +repo=/workspace/Megatron-UltraEP +branch=ultraep-bracing-regions-through-kernels +commit=03d57ef7 +num_layers=24 +hidden=2048 +ffn_hidden=10944 +moe_ffn_hidden=1408 +num_experts=64 +topk=6 +experts_per_rank=8 +expert_params=8650752 +heads=16 +kv_groups=0 +kv_channels=128 +qk_layernorm=1 +shared_ffn=2816 +shared_expert_overlap=0 +mla=1 +dense_first_layers=1 +moe_layers=23 +rotary_base=10000 +rotary_scaling=40 +qkv_bias=0 +router_balancing=seq_aux_loss +aux_loss_coeff=1e-3 +router_pre_softmax=1 +router_fusion=0 +router_dtype=fp32 +device_initiated_grouped_gemm=1 +norm_epsilon=1e-6 +seq_length=8192 +micro_batch_size=1 +global_batch_size=8 +num_microbatches=1 +assignments_per_rank=49152 +mean_tokens_per_expert=6144 +slot_policy=matched +slots_per_rank=8 +capacity_mode=static +capacity=1.0755208333333333 +cuda_graph=1 +probe_mode=0 +async_grad_reduce=0 +branch_role=timing +memory_dump_mode=none +memory_snapshot_path= +cuda_graph_warmup_steps=3 +recompute=none +force_balance=0 +train_iters=12 +ep_size=8 +echo_dispatch_overlap=1 +echo_dispatch_recompute=0 +ultraep_quota_min_tokens=768 +ultraep_weight_sync_join= +profile_enable=0 +kernel_trace=0 +fp8_recipe=mxfp8 diff --git a/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep1/train.log b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep1/train.log new file mode 100644 index 0000000000000000000000000000000000000000..91c7fe4264fd53c17e82388c5a598818474e989f --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep1/train.log @@ -0,0 +1,1610 @@ +============================================================== + DSV2_LITE_L24 / ultraep (slots/rank=8 policy=matched) + E=64 K=6 H'=1408 m=8 + L=24 seq=8192 gbs=8 (1 microbatches) + cap_mode=static cap=1.0755208333333333 cg=1 recompute=none + memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep1/memory_snapshots/snapshot.pickle + repo=/workspace/Megatron-UltraEP @ 03d57ef7 +============================================================== +W0826 18:35:01.965000 528750 torch/distributed/run.py:766] +W0826 18:35:01.965000 528750 torch/distributed/run.py:766] ***************************************** +W0826 18:35:01.965000 528750 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0826 18:35:01.965000 528750 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") +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") +using world size: 8, data-parallel size: 8, context-parallel size: 1, hierarchical context-parallel sizes: None, tensor-model-parallel size: 1, pipeline-model-parallel size: 1 +WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer +Number of virtual stages per pipeline stage: None +accumulate and all-reduce gradients in fp32 for bfloat16 data type. +using torch.bfloat16 for parameters ... +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +------------------------ arguments ------------------------ + account_for_embedding_in_pipeline_split ......... False + account_for_loss_in_pipeline_split .............. False + accumulate_allreduce_grads_in_fp32 .............. True + adam_beta1 ...................................... 0.9 + adam_beta2 ...................................... 0.999 + adam_eps ........................................ 1e-08 + add_bias_linear ................................. False + add_position_embedding .......................... False + add_qkv_bias .................................... False + adlr_autoresume ................................. False + adlr_autoresume_interval ........................ 1000 + align_grad_reduce ............................... True + align_param_gather .............................. False + app_tag_run_name ................................ None + app_tag_run_version ............................. 0.0.0 + apply_layernorm_1p .............................. False + apply_query_key_layer_scaling ................... False + apply_residual_connection_post_layernorm ........ False + apply_rope_fusion ............................... False + async_save ...................................... None + async_tensor_model_parallel_allreduce ........... True + attention_backend ............................... AttnBackend.auto + attention_dropout ............................... 0.0 + attention_softmax_in_fp32 ....................... False + auto_detect_ckpt_format ......................... False + barrier_with_L1_time ............................ True + bert_binary_head ................................ True + bert_embedder_type .............................. megatron + bert_load ....................................... None + bf16 ............................................ True + bias_dropout_fusion ............................. True + bias_gelu_fusion ................................ False + bias_swiglu_fusion .............................. True + biencoder_projection_dim ........................ 0 + biencoder_shared_query_context_model ............ False + block_data_path ................................. None + calc_ft_timeouts ................................ False + calculate_per_token_loss ........................ False + check_for_large_grads ........................... False + check_for_nan_in_loss_and_grad .................. False + check_for_spiky_loss ............................ False + check_weight_hash_across_dp_replicas_interval ... None + ckpt_assume_constant_structure .................. False + ckpt_convert_format ............................. None + ckpt_convert_save ............................... None + ckpt_convert_update_legacy_dist_opt_format ...... False + ckpt_format ..................................... torch_dist + ckpt_fully_parallel_load ........................ False + ckpt_fully_parallel_save ........................ True + ckpt_fully_parallel_save_deprecated ............. False + ckpt_step ....................................... None + classes_fraction ................................ 1.0 + clip_grad ....................................... 1.0 + clone_scatter_output_in_embedding ............... True + config_logger_dir ............................... + consumed_train_samples .......................... 0 + consumed_valid_samples .......................... 0 + context_parallel_size ........................... 1 + cp_comm_type .................................... ['p2p'] + create_attention_mask_in_dataloader ............. True + cross_entropy_fusion_impl ....................... native + cross_entropy_loss_fusion ....................... False + cuda_graph_scope ................................ full_iteration + cuda_graph_warmup_steps ......................... 3 + data_args_path .................................. None + data_cache_path ................................. None + data_parallel_random_init ....................... False + data_parallel_sharding_strategy ................. no_shard + data_parallel_size .............................. 8 + data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document'] + data_per_class_fraction ......................... 1.0 + data_sharding ................................... True + dataloader_type ................................. single + ddp_average_in_collective ....................... False + ddp_bucket_size ................................. None + ddp_num_buckets ................................. None + ddp_pad_buckets_for_high_nccl_busbw ............. False + decoder_first_pipeline_num_layers ............... None + decoder_last_pipeline_num_layers ................ None + decoder_num_layers .............................. None + decoder_seq_length .............................. None + decoupled_lr .................................... None + decoupled_min_lr ................................ None + decrease_batch_size_if_needed ................... False + defer_embedding_wgrad_compute ................... False + delay_wgrad_compute ............................. False + deprecated_use_mcore_models ..................... True + deterministic_mode .............................. False + dino_bottleneck_size ............................ 256 + dino_freeze_last_layer .......................... 1 + dino_head_hidden_size ........................... 2048 + dino_local_crops_number ......................... 10 + dino_local_img_size ............................. 96 + dino_norm_last_layer ............................ False + dino_teacher_temp ............................... 0.07 + dino_warmup_teacher_temp ........................ 0.04 + dino_warmup_teacher_temp_epochs ................. 30 + disable_bf16_reduced_precision_matmul ........... False + disable_mamba_mem_eff_path ...................... False + disable_straggler_on_startup .................... False + dist_ckpt_format_deprecated ..................... None + dist_ckpt_strictness ............................ assume_ok_unexpected + distribute_saved_activations .................... False + distributed_backend ............................. nccl + distributed_timeout_minutes ..................... 10 + embedding_init_method_std ....................... None + embedding_path .................................. None + empty_unused_memory_level ....................... 0 + enable_cuda_graph ............................... True + enable_experimental ............................. False + enable_ft_package ............................... False + enable_gloo_process_groups ...................... True + enable_msc ...................................... True + enable_one_logger ............................... True + encoder_num_layers .............................. 24 + encoder_seq_length .............................. 8192 + end_weight_decay ................................ 0.1 + eod_mask_loss ................................... False + error_injection_rate ............................ 0 + error_injection_type ............................ transient_error + eval_interval ................................... 1000000 + eval_iters ...................................... 0 + evidence_data_path .............................. None + exit_duration_in_mins ........................... None + exit_interval ................................... None + exit_on_missing_checkpoint ...................... False + exit_signal_handler ............................. False + exp_avg_dtype ................................... torch.float32 + exp_avg_sq_dtype ................................ torch.float32 + expert_model_parallel_size ...................... 8 + expert_tensor_parallel_size ..................... 1 + export_force_local_attention .................... False + export_kd_cfg ................................... None + export_kd_teacher_ckpt_format ................... None + export_kd_teacher_load .......................... None + export_kv_cache_quant ........................... False + export_legacy_megatron .......................... False + export_model_type ............................... GPTModel + export_moe_apply_probs_on_input ................. False + export_qk_l2_norm ............................... False + export_quant_cfg ................................ None + export_real_quant_cfg ........................... None + export_te_mcore_model ........................... False + external_cuda_graph ............................. False + ffn_hidden_size ................................. 10944 + finetune ........................................ False + finetune_data_split ............................. train + finetune_hf_dataset ............................. None + first_last_layers_bf16 .......................... False + flash_decode .................................... False + fp16 ............................................ False + fp16_lm_cross_entropy ........................... False + fp32_residual_connection ........................ False + fp8 ............................................. e4m3 + fp8_amax_compute_algo ........................... most_recent + fp8_amax_history_len ............................ 1 + fp8_interval .................................... 1 + fp8_margin ...................................... 0 + fp8_param_gather ................................ False + fp8_recipe ...................................... mxfp8 + fp8_wgrad ....................................... True + fsdp_double_buffer .............................. False + global_batch_size ............................... 8 + grad_reduce_in_bf16 ............................. False + gradient_accumulation_fusion .................... True + gradient_reduce_div_fusion ...................... True + group_query_attention ........................... False + head_lr_mult .................................... 1.0 + heterogeneous_layers_config_encoded_json ........ None + heterogeneous_layers_config_path ................ None + hidden_dropout .................................. 0.0 + hidden_size ..................................... 2048 + hierarchical_context_parallel_sizes ............. None + high_priority_stream_groups ..................... [] + hybrid_attention_ratio .......................... 0.0 + hybrid_mlp_ratio ................................ 0.0 + hybrid_override_pattern ......................... None + hysteresis ...................................... 2 + ict_head_size ................................... None + ict_load ........................................ None + img_h ........................................... 224 + img_w ........................................... 224 + indexer_batch_size .............................. 128 + indexer_log_interval ............................ 1000 + inference_batch_times_seqlen_threshold .......... -1 + inference_dynamic_batching ...................... False + inference_dynamic_batching_buffer_guaranteed_fraction 0.2 + inference_dynamic_batching_buffer_overflow_factor None + inference_dynamic_batching_buffer_size_gb ....... 40.0 + inference_dynamic_batching_chunk_size ........... 256 + inference_dynamic_batching_max_requests_override None + inference_dynamic_batching_max_tokens_override .. None + inference_dynamic_batching_num_cuda_graphs ...... 16 + inference_max_batch_size ........................ 8 + inference_max_seq_length ........................ 2560 + inference_rng_tracker ........................... False + init_method_std ................................. 0.01 + init_method_xavier_uniform ...................... False + init_model_with_meta_device ..................... False + initial_loss_scale .............................. 4294967296 + inprocess_active_world_size ..................... 8 + inprocess_barrier_timeout ....................... 120 + inprocess_completion_timeout .................... 120 + inprocess_empty_cuda_cache ...................... False + inprocess_granularity ........................... node + inprocess_hard_timeout .......................... 90 + inprocess_heartbeat_interval .................... 30 + inprocess_heartbeat_timeout ..................... 60 + inprocess_last_call_wait ........................ 1 + inprocess_max_iterations ........................ None + inprocess_monitor_process_interval .............. 1.0 + inprocess_monitor_thread_interval ............... 1.0 + inprocess_progress_watchdog_interval ............ 1.0 + inprocess_restart ............................... False + inprocess_soft_timeout .......................... 60 + inprocess_termination_grace_time ................ 1 + is_hybrid_model ................................. False + iter_per_epoch .................................. 1250 + iterations_to_skip .............................. [] + keep_fp8_transpose_cache_when_using_custom_fsdp . False + kitchen_config_file ............................. None + kitchen_recipe_number ........................... None + kv_channels ..................................... 128 + kv_lora_rank .................................... 512 + lazy_mpu_init ................................... None + load ............................................ None + load_main_params_from_ckpt ...................... None + load_model_opt_format ........................... False + local_rank ...................................... 0 + log_energy ...................................... False + log_interval .................................... 1 + log_loss_scale_to_tensorboard ................... True + log_memory_to_tensorboard ....................... False + log_num_zeros_in_grad ........................... False + log_params_norm ................................. False + log_progress .................................... False + log_straggler ................................... False + log_throughput .................................. True + log_timers_to_tensorboard ....................... False + log_validation_ppl_to_tensorboard ............... False + log_world_size_to_tensorboard ................... False + logging_level ................................... None + loss_scale ...................................... None + loss_scale_window ............................... 1000 + lr .............................................. 0.0001 + lr_decay_iters .................................. 8 + lr_decay_samples ................................ None + lr_decay_style .................................. cosine + lr_warmup_fraction .............................. None + lr_warmup_init .................................. 0.0 + lr_warmup_iters ................................. 5 + lr_warmup_samples ............................... 0 + lr_wsd_decay_iters .............................. None + lr_wsd_decay_samples ............................ None + lr_wsd_decay_style .............................. exponential + main_grads_dtype ................................ torch.float32 + main_params_dtype ............................... torch.float32 + make_vocab_size_divisible_by .................... 128 + mamba_head_dim .................................. 64 + mamba_num_groups ................................ 8 + mamba_num_heads ................................. None + mamba_state_dim ................................. 128 + manual_gc ....................................... False + manual_gc_eval .................................. True + manual_gc_interval .............................. 0 + mask_factor ..................................... 1.0 + mask_prob ....................................... 0.15 + mask_type ....................................... random + masked_softmax_fusion ........................... False + max_position_embeddings ......................... 32768 + max_tokens_to_oom ............................... 12000 + memory_snapshot_path ............................ snapshot.pickle + merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt + micro_batch_size ................................ 1 + microbatch_group_size_per_vp_stage .............. None + mid_level_dataset_surplus ....................... 0.005 + min_loss_scale .................................. 1.0 + min_lr .......................................... 1e-05 + mlp_chunks_for_prefill .......................... 1 + mmap_bin_files .................................. True + mock_data ....................................... False + moe_apply_probs_on_input ........................ False + moe_aux_loss_coeff .............................. 0.001 + moe_capacity_monitor ............................ True + moe_deepep_num_sms .............................. 20 + moe_echo_enable_random_offloading ............... False + moe_echo_expert_dispatch_overlap ................ False + moe_echo_expert_dispatcher_type ................. hybridep + moe_echo_recompute_expert_dispatch .............. False + moe_enable_deepep ............................... False + moe_enable_echo ................................. False + moe_enable_hybridep ............................. True + moe_enable_ultraep .............................. True + moe_expert_capacity_factor ...................... None + moe_extended_tp ................................. False + moe_ffn_hidden_size ............................. 1408 + moe_grouped_gemm ................................ True + moe_hybridep_num_sms ............................ 16 + moe_input_jitter_eps ............................ None + moe_layer_freq .................................. [0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] + moe_layer_recompute ............................. False + moe_num_echo_experts ............................ None + moe_num_redundant_experts_per_rank .............. 8 + moe_pad_expert_input_to_capacity ................ False + moe_per_layer_logging ........................... False + moe_permute_fusion .............................. False + moe_received_token_capacity ..................... 1.0755208333333333 + moe_router_bias_update_rate ..................... 0.001 + moe_router_dtype ................................ fp32 + moe_router_enable_expert_bias ................... False + moe_router_force_load_balancing ................. False + moe_router_group_topk ........................... None + moe_router_load_balancing_type .................. seq_aux_loss + moe_router_num_groups ........................... None + moe_router_padding_for_fp8 ...................... False + moe_router_pre_softmax .......................... True + moe_router_score_function ....................... softmax + moe_router_topk ................................. 6 + moe_router_topk_scaling_factor .................. None + moe_shared_expert_intermediate_size ............. 2816 + moe_shared_expert_overlap ....................... False + moe_token_dispatcher_type ....................... flex + moe_token_drop_policy ........................... probs + moe_ultraep_max_microbatches .................... None + moe_ultraep_overlap_weight_sync ................. True + moe_ultraep_weight_sync_join .................... before_dispatch + moe_upcycling_granularity ....................... 1 + moe_use_device_initiated_grouped_gemm ........... True + moe_use_legacy_grouped_gemm ..................... False + moe_use_upcycling ............................... False + moe_z_loss_coeff ................................ None + mrope_section ................................... None + mscale .......................................... 0.707 + mscale_all_dim .................................. 0.707 + mtp_loss_scaling_factor ......................... 0.1 + mtp_num_layers .................................. None + multi_latent_attention .......................... True + nccl_all_reduce_for_prefill ..................... False + nccl_communicator_config_path ................... None + nccl_ub ......................................... False + no_load_optim ................................... None + no_load_rng ..................................... None + no_persist_layer_norm ........................... False + no_rope_freq .................................... None + no_save_optim ................................... None + no_save_rng ..................................... None + non_persistent_ckpt_type ........................ None + non_persistent_global_ckpt_dir .................. None + non_persistent_local_ckpt_algo .................. fully_parallel + non_persistent_local_ckpt_dir ................... None + non_persistent_save_interval .................... None + norm_epsilon .................................... 1e-06 + normalization ................................... RMSNorm + num_attention_heads ............................. 16 + num_channels .................................... 3 + num_classes ..................................... 1000 + num_dataset_builder_threads ..................... 1 + num_distributed_optimizer_instances ............. 1 + num_experts ..................................... 64 + num_layers ...................................... 24 + num_layers_at_end_in_bf16 ....................... 1 + num_layers_at_start_in_bf16 ..................... 1 + num_layers_per_virtual_pipeline_stage ........... None + num_query_groups ................................ 1 + num_virtual_stages_per_pipeline_rank ............ None + num_workers ..................................... 2 + object_storage_cache_path ....................... None + one_logger_async ................................ False + one_logger_project .............................. megatron-lm + one_logger_run_name ............................. None + onnx_safe ....................................... None + openai_gelu ..................................... False + optimizer ....................................... adam + optimizer_cpu_offload ........................... False + optimizer_offload_fraction ...................... 1.0 + output_bert_embeddings .......................... False + overlap_cpu_optimizer_d2h_h2d ................... False + overlap_grad_reduce ............................. True + overlap_moe_expert_parallel_comm ................ False + overlap_p2p_comm ................................ False + overlap_p2p_comm_warmup_flush ................... False + overlap_param_gather ............................ True + overlap_param_gather_with_optimizer_step ........ False + override_opt_param_scheduler .................... False + params_dtype .................................... torch.bfloat16 + patch_dim ....................................... 16 + per_split_data_args_path ........................ None + perform_initialization .......................... True + pin_cpu_grads ................................... True + pin_cpu_params .................................. True + pipeline_model_parallel_comm_backend ............ None + pipeline_model_parallel_layout .................. None + pipeline_model_parallel_size .................... 1 + position_embedding_type ......................... rope + pretrained_checkpoint ........................... None + profile ......................................... False + profile_ranks ................................... [0] + profile_step_end ................................ 12 + profile_step_start .............................. 10 + q_lora_rank ..................................... None + qk_head_dim ..................................... 128 + qk_l2_norm ...................................... False + qk_layernorm .................................... True + qk_pos_emb_head_dim ............................. 64 + query_in_block_prob ............................. 0.1 + rampup_batch_size ............................... None + rank ............................................ 0 + recompute_granularity ........................... None + recompute_method ................................ None + recompute_modules ............................... None + recompute_num_layers ............................ None + record_memory_history ........................... False + relative_attention_max_distance ................. 128 + relative_attention_num_buckets .................. 32 + replication ..................................... False + replication_factor .............................. 2 + replication_jump ................................ None + rerun_mode ...................................... validate_results + reset_attention_mask ............................ False + reset_position_ids .............................. False + result_rejected_tracker_filename ................ None + retriever_report_topk_accuracies ................ [] + retriever_score_scaling ......................... False + retriever_seq_length ............................ 256 + retro_add_retriever ............................. False + retro_attention_gate ............................ 1 + retro_cyclic_train_iters ........................ None + retro_encoder_attention_dropout ................. 0.1 + retro_encoder_hidden_dropout .................... 0.1 + retro_encoder_layers ............................ 2 + retro_num_neighbors ............................. 2 + retro_num_retrieved_chunks ...................... 2 + retro_project_dir ............................... None + retro_verify_neighbor_count ..................... True + reuse_grad_buf_for_mxfp8_param_ag ............... False + rope_scaling_factor ............................. 8.0 + rope_type ....................................... None + rotary_base ..................................... 10000 + rotary_interleaved .............................. False + rotary_percent .................................. 1.0 + rotary_scaling_factor ........................... 40.0 + rotary_seq_len_interpolation_factor ............. None + run_workload_inspector_server ................... False + sample_rate ..................................... 1.0 + save ............................................ None + save_interval ................................... None + save_retain_interval ............................ None + scatter_gather_tensors_in_pipeline .............. True + seed ............................................ 1234 + seq_length ...................................... 8192 + sequence_parallel ............................... False + sft ............................................. False + sft_tokenizer_prompt_format ..................... nemotron-h-aligned + sgd_momentum .................................... 0.9 + short_seq_prob .................................. 0.1 + skip_train ...................................... False + skipped_train_samples ........................... 0 + spec ............................................ None + split ........................................... 100,0,0 + squared_relu .................................... False + start_weight_decay .............................. 0.1 + straggler_ctrlr_port ............................ 65535 + straggler_minmax_count .......................... 1 + suggested_communication_unit_size ............... None + swiglu .......................................... True + swin_backbone_type .............................. tiny + symmetric_ar_type ............................... None + te_rng_tracker .................................. True + tensor_model_parallel_size ...................... 1 + tensorboard_dir ................................. None + tensorboard_log_interval ........................ 1 + tensorboard_queue_size .......................... 1000 + test_data_path .................................. None + test_mode ....................................... False + tiktoken_num_special_tokens ..................... 1000 + tiktoken_pattern ................................ None + tiktoken_special_tokens ......................... None + timing_log_level ................................ 0 + timing_log_option ............................... minmax + titles_data_path ................................ None + tokenizer_model ................................. None + tokenizer_type .................................. GPT2BPETokenizer + torch_fsdp2_reshard_after_forward ............... True + tp_comm_bootstrap_backend ....................... nccl + tp_comm_bulk_dgrad .............................. True + tp_comm_bulk_wgrad .............................. True + tp_comm_overlap ................................. False + tp_comm_overlap_ag .............................. True + tp_comm_overlap_cfg ............................. None + tp_comm_overlap_rs .............................. True + tp_comm_overlap_rs_dgrad ........................ False + tp_comm_split_ag ................................ True + tp_comm_split_rs ................................ True + train_data_path ................................. None + train_iters ..................................... 12 + train_samples ................................... None + train_sync_interval ............................. None + transformer_impl ................................ transformer_engine + transformer_pipeline_model_parallel_size ........ 1 + untie_embeddings_and_output_weights ............. True + use_checkpoint_args ............................. False + use_checkpoint_opt_param_scheduler .............. False + use_cpu_initialization .......................... None + use_custom_fsdp ................................. False + use_dist_ckpt ................................... True + use_dist_ckpt_deprecated ........................ False + use_distributed_optimizer ....................... True + use_flash_attn .................................. False + use_legacy_models ............................... False + use_mp_args_from_checkpoint_args ................ False + use_one_sent_docs ............................... False + use_persistent_ckpt_worker ...................... False + use_precision_aware_optimizer ................... False + use_pytorch_profiler ............................ False + use_ring_exchange_p2p ........................... False + use_rope_scaling ................................ False + use_rotary_position_embeddings .................. False + use_sharp ....................................... False + use_tokenizer_model_from_checkpoint_args ........ True + use_torch_fsdp2 ................................. False + use_torch_optimizer_for_cpu_offload ............. False + use_tp_pp_dp_mapping ............................ False + v_head_dim ...................................... 128 + valid_data_path ................................. None + variable_seq_lengths ............................ False + virtual_pipeline_model_parallel_size ............ None + vision_backbone_type ............................ vit + vision_pretraining .............................. False + vision_pretraining_type ......................... classify + vocab_extra_ids ................................. 0 + vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json + vocab_size ...................................... None + wandb_exp_name .................................. + wandb_project ................................... + wandb_save_dir .................................. + weight_decay .................................... 0.1 + weight_decay_incr_style ......................... constant + wgrad_deferral_limit ............................ 0 + world_size ...................................... 8 + yaml_cfg ........................................ None +-------------------- end of arguments --------------------- +INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1 +> building GPT2BPETokenizer tokenizer ... + > 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") +[Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank [Gloo] Rank [Gloo] Rank [Gloo] Rank 1246 is connected to is connected to is connected to is connected to 7777 peer ranks. peer ranks. peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : Expected number of connected peer ranks is : Expected number of connected peer ranks is : 7777 + + + +[Gloo] Rank 5 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 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank [Gloo] Rank 3 is connected to 17 is connected to peer ranks. 7Expected number of connected peer ranks is : peer ranks. 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 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 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 +> 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 [Gloo] Rank 00 is connected to is connected to 00 peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : 00 + +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[rank1]:[W826 18:35:20.108488939 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. +[rank2]:[W826 18:35:20.116920911 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. +[rank3]:[W826 18:35:20.117022262 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank4]:[W826 18:35:20.117041468 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]:[W826 18:35:20.117043252 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank5]:[W826 18:35:20.117045243 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]:[W826 18:35:20.117059906 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: 0.034 seconds +WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations. +> compiling and loading fused kernels ... +[rank0]:[W826 18:35:20.273067510 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: 3.915 seconds +time to initialize megatron (seconds): 7.106 +[after megatron is initialized] datetime: 2026-08-26 18:35:26 +building GPT model ... +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +[INFO] Use MNNVL fabric: False +[WARN] MAX_NUM_NVL_PEERS is not set. Using detected value 8. +NVSHMEM v3.7.2 +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +UltraEP: NVL domain size <= 8; forcing weight_sync plan mode to direct (was adaptive). + > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 4188125184 +INFO:megatron.core.distributed.distributed_data_parallel:Setting up DistributedDataParallel with config DistributedDataParallelConfig(grad_reduce_in_fp32=True, overlap_grad_reduce=True, overlap_param_gather=True, align_param_gather=False, use_distributed_optimizer=True, num_distributed_optimizer_instances=1, check_for_nan_in_grad=False, check_for_large_grads=False, bucket_size=40000000, pad_buckets_for_high_nccl_busbw=False, average_in_collective=False, fp8_param_gather=False, reuse_grad_buf_for_mxfp8_param_ag=False, use_custom_fsdp=False, data_parallel_sharding_strategy='no_shard', gradient_reduce_div_fusion=True, suggested_communication_unit_size=None, preserve_fp32_weights=True, keep_fp8_transpose_cache_when_using_custom_fsdp=False, nccl_ub=False, fsdp_double_buffer=False) +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 20 +Params for bucket 1 (103022592 elements, 103022592 padded size): + module.output_layer.weight +Params for bucket 2 (48503296 elements, 48503296 padded size): + module.decoder.layers.23.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.23.self_attention.linear_kv_up_proj.weight + module.decoder.layers.22.mlp.shared_experts.linear_fc2.weight + module.decoder.final_layernorm.weight + module.decoder.layers.23.pre_mlp_layernorm.weight + module.decoder.layers.23.mlp.router.weight + module.decoder.layers.23.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.23.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.23.self_attention.linear_kv_down_proj.weight + module.decoder.layers.23.self_attention.linear_q_proj.weight + module.decoder.layers.23.self_attention.linear_proj.weight + module.decoder.layers.23.input_layernorm.weight + module.decoder.layers.22.mlp.shared_experts.linear_fc1.weight +Params for bucket 3 (40901632 elements, 40901632 padded size): + module.decoder.layers.22.mlp.router.weight + module.decoder.layers.22.self_attention.linear_kv_down_proj.weight + module.decoder.layers.22.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.22.pre_mlp_layernorm.weight + module.decoder.layers.22.self_attention.linear_q_proj.weight + module.decoder.layers.22.self_attention.linear_proj.weight + module.decoder.layers.22.input_layernorm.weight + module.decoder.layers.21.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.21.pre_mlp_layernorm.weight + module.decoder.layers.21.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.22.self_attention.linear_kv_up_proj.weight + module.decoder.layers.21.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.21.mlp.router.weight + module.decoder.layers.21.self_attention.linear_kv_up_proj.weight + module.decoder.layers.21.self_attention.linear_kv_down_proj.weight + module.decoder.layers.21.self_attention.linear_q_proj.weight +Params for bucket 4 (41163264 elements, 41163264 padded size): + module.decoder.layers.21.self_attention.linear_proj.weight + module.decoder.layers.20.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.20.mlp.router.weight + module.decoder.layers.21.input_layernorm.weight + module.decoder.layers.20.self_attention.linear_kv_up_proj.weight + module.decoder.layers.19.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.20.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.20.pre_mlp_layernorm.weight + module.decoder.layers.20.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.20.self_attention.linear_kv_down_proj.weight + module.decoder.layers.20.self_attention.linear_q_proj.weight + module.decoder.layers.20.self_attention.linear_proj.weight + module.decoder.layers.20.input_layernorm.weight +Params for bucket 5 (42734080 elements, 42734080 padded size): + module.decoder.layers.19.mlp.router.weight + module.decoder.layers.19.self_attention.linear_kv_up_proj.weight + module.decoder.layers.18.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.19.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.19.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.19.pre_mlp_layernorm.weight + module.decoder.layers.19.self_attention.linear_kv_down_proj.weight + module.decoder.layers.19.self_attention.linear_q_proj.weight + module.decoder.layers.19.self_attention.linear_proj.weight + module.decoder.layers.19.input_layernorm.weight + module.decoder.layers.18.mlp.shared_experts.linear_fc1.weight +Params for bucket 6 (40901632 elements, 40901632 padded size): + module.decoder.layers.18.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.18.pre_mlp_layernorm.weight + module.decoder.layers.18.self_attention.linear_kv_down_proj.weight + module.decoder.layers.18.self_attention.linear_q_proj.weight + module.decoder.layers.18.self_attention.linear_proj.weight + module.decoder.layers.18.input_layernorm.weight + module.decoder.layers.17.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.17.mlp.router.weight + module.decoder.layers.17.self_attention.linear_kv_up_proj.weight + module.decoder.layers.18.mlp.router.weight + module.decoder.layers.18.self_attention.linear_kv_up_proj.weight + module.decoder.layers.17.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.17.pre_mlp_layernorm.weight + module.decoder.layers.17.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.17.self_attention.linear_kv_down_proj.weight + module.decoder.layers.17.self_attention.linear_q_proj.weight +Params for bucket 7 (41163264 elements, 41163264 padded size): + module.decoder.layers.16.mlp.router.weight + module.decoder.layers.16.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.16.self_attention.linear_kv_up_proj.weight + module.decoder.layers.15.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.17.self_attention.linear_proj.weight + module.decoder.layers.16.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.17.input_layernorm.weight + module.decoder.layers.16.pre_mlp_layernorm.weight + module.decoder.layers.16.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.16.self_attention.linear_kv_down_proj.weight + module.decoder.layers.16.self_attention.linear_q_proj.weight + module.decoder.layers.16.self_attention.linear_proj.weight + module.decoder.layers.16.input_layernorm.weight +Params for bucket 8 (42734080 elements, 42734080 padded size): + module.decoder.layers.15.self_attention.linear_kv_up_proj.weight + module.decoder.layers.15.mlp.router.weight + module.decoder.layers.15.self_attention.linear_proj.weight + module.decoder.layers.14.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.15.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.15.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.15.pre_mlp_layernorm.weight + module.decoder.layers.15.self_attention.linear_kv_down_proj.weight + module.decoder.layers.15.self_attention.linear_q_proj.weight + module.decoder.layers.15.input_layernorm.weight + module.decoder.layers.14.mlp.shared_experts.linear_fc1.weight +Params for bucket 9 (40901632 elements, 40901632 padded size): + module.decoder.layers.14.mlp.router.weight + module.decoder.layers.14.self_attention.linear_kv_up_proj.weight + module.decoder.layers.13.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.13.mlp.router.weight + module.decoder.layers.13.self_attention.linear_kv_up_proj.weight + module.decoder.layers.14.self_attention.linear_kv_down_proj.weight + module.decoder.layers.14.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.14.pre_mlp_layernorm.weight + module.decoder.layers.14.self_attention.linear_q_proj.weight + module.decoder.layers.14.self_attention.linear_proj.weight + module.decoder.layers.14.input_layernorm.weight + module.decoder.layers.13.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.13.pre_mlp_layernorm.weight + module.decoder.layers.13.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.13.self_attention.linear_kv_down_proj.weight + module.decoder.layers.13.self_attention.linear_q_proj.weight +Params for bucket 10 (41163264 elements, 41163264 padded size): + module.decoder.layers.12.mlp.router.weight + module.decoder.layers.12.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.12.self_attention.linear_kv_up_proj.weight + module.decoder.layers.11.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.13.self_attention.linear_proj.weight + module.decoder.layers.12.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.13.input_layernorm.weight + module.decoder.layers.12.pre_mlp_layernorm.weight + module.decoder.layers.12.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.12.self_attention.linear_kv_down_proj.weight + module.decoder.layers.12.self_attention.linear_q_proj.weight + module.decoder.layers.12.self_attention.linear_proj.weight + module.decoder.layers.12.input_layernorm.weight +Params for bucket 11 (42734080 elements, 42734080 padded size): + module.decoder.layers.11.mlp.router.weight + module.decoder.layers.11.self_attention.linear_kv_up_proj.weight + module.decoder.layers.10.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.11.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.11.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.11.pre_mlp_layernorm.weight + module.decoder.layers.11.self_attention.linear_kv_down_proj.weight + module.decoder.layers.11.self_attention.linear_q_proj.weight + module.decoder.layers.11.self_attention.linear_proj.weight + module.decoder.layers.11.input_layernorm.weight + module.decoder.layers.10.mlp.shared_experts.linear_fc1.weight +Params for bucket 12 (40901632 elements, 40901632 padded size): + module.decoder.layers.10.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.10.self_attention.linear_kv_down_proj.weight + module.decoder.layers.10.pre_mlp_layernorm.weight + module.decoder.layers.10.self_attention.linear_q_proj.weight + module.decoder.layers.10.self_attention.linear_proj.weight + module.decoder.layers.9.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.9.mlp.router.weight + module.decoder.layers.9.self_attention.linear_kv_up_proj.weight + module.decoder.layers.10.mlp.router.weight + module.decoder.layers.10.self_attention.linear_kv_up_proj.weight + module.decoder.layers.10.input_layernorm.weight + module.decoder.layers.9.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.9.pre_mlp_layernorm.weight + module.decoder.layers.9.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.9.self_attention.linear_kv_down_proj.weight + module.decoder.layers.9.self_attention.linear_q_proj.weight +Params for bucket 13 (41163264 elements, 41163264 padded size): + module.decoder.layers.8.mlp.router.weight + module.decoder.layers.8.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.8.self_attention.linear_kv_up_proj.weight + module.decoder.layers.7.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.9.self_attention.linear_proj.weight + module.decoder.layers.8.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.9.input_layernorm.weight + module.decoder.layers.8.pre_mlp_layernorm.weight + module.decoder.layers.8.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.8.self_attention.linear_kv_down_proj.weight + module.decoder.layers.8.self_attention.linear_q_proj.weight + module.decoder.layers.8.self_attention.linear_proj.weight + module.decoder.layers.8.input_layernorm.weight +Params for bucket 14 (42734080 elements, 42734080 padded size): + module.decoder.layers.7.mlp.router.weight + module.decoder.layers.7.self_attention.linear_kv_up_proj.weight + module.decoder.layers.6.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.7.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.7.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.7.pre_mlp_layernorm.weight + module.decoder.layers.7.self_attention.linear_kv_down_proj.weight + module.decoder.layers.7.self_attention.linear_q_proj.weight + module.decoder.layers.7.self_attention.linear_proj.weight + module.decoder.layers.7.input_layernorm.weight + module.decoder.layers.6.mlp.shared_experts.linear_fc1.weight +Params for bucket 15 (40901632 elements, 40901632 padded size): + module.decoder.layers.6.mlp.router.weight + module.decoder.layers.6.self_attention.linear_kv_up_proj.weight + module.decoder.layers.5.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.5.mlp.router.weight + module.decoder.layers.5.self_attention.linear_kv_up_proj.weight + module.decoder.layers.5.self_attention.linear_q_proj.weight + module.decoder.layers.6.self_attention.linear_kv_down_proj.weight + module.decoder.layers.6.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.6.pre_mlp_layernorm.weight + module.decoder.layers.6.self_attention.linear_q_proj.weight + module.decoder.layers.6.self_attention.linear_proj.weight + module.decoder.layers.6.input_layernorm.weight + module.decoder.layers.5.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.5.pre_mlp_layernorm.weight + module.decoder.layers.5.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.5.self_attention.linear_kv_down_proj.weight +Params for bucket 16 (41163264 elements, 41163264 padded size): + module.decoder.layers.5.self_attention.linear_proj.weight + module.decoder.layers.5.input_layernorm.weight + module.decoder.layers.4.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.4.pre_mlp_layernorm.weight + module.decoder.layers.4.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.4.self_attention.linear_kv_down_proj.weight + module.decoder.layers.4.self_attention.linear_q_proj.weight + module.decoder.layers.4.self_attention.linear_proj.weight + module.decoder.layers.4.input_layernorm.weight + module.decoder.layers.4.mlp.router.weight + module.decoder.layers.4.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.4.self_attention.linear_kv_up_proj.weight + module.decoder.layers.3.mlp.shared_experts.linear_fc2.weight +Params for bucket 17 (42734080 elements, 42734080 padded size): + module.decoder.layers.3.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.3.mlp.router.weight + module.decoder.layers.3.self_attention.linear_kv_up_proj.weight + module.decoder.layers.2.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.3.pre_mlp_layernorm.weight + module.decoder.layers.3.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.3.self_attention.linear_kv_down_proj.weight + module.decoder.layers.3.self_attention.linear_q_proj.weight + module.decoder.layers.3.self_attention.linear_proj.weight + module.decoder.layers.3.input_layernorm.weight + module.decoder.layers.2.mlp.shared_experts.linear_fc1.weight +Params for bucket 18 (40901632 elements, 40901632 padded size): + module.decoder.layers.2.mlp.router.weight + module.decoder.layers.2.self_attention.linear_kv_up_proj.weight + module.decoder.layers.1.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.1.mlp.router.weight + module.decoder.layers.1.self_attention.linear_kv_up_proj.weight + module.decoder.layers.1.self_attention.linear_kv_down_proj.weight + module.decoder.layers.1.self_attention.linear_q_proj.weight + module.decoder.layers.2.self_attention.linear_kv_down_proj.weight + module.decoder.layers.2.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.2.pre_mlp_layernorm.weight + module.decoder.layers.2.self_attention.linear_q_proj.weight + module.decoder.layers.2.self_attention.linear_proj.weight + module.decoder.layers.2.input_layernorm.weight + module.decoder.layers.1.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.1.pre_mlp_layernorm.weight + module.decoder.layers.1.self_attention.linear_kv_up_proj.layer_norm_weight +Params for bucket 19 (71436288 elements, 71436288 padded size): + module.decoder.layers.1.self_attention.linear_proj.weight + module.decoder.layers.0.mlp.linear_fc2.weight + module.decoder.layers.0.mlp.linear_fc1.weight + module.decoder.layers.1.input_layernorm.weight +Params for bucket 20 (116789760 elements, 116789760 padded size): + module.decoder.layers.0.mlp.linear_fc1.layer_norm_weight + module.decoder.layers.0.self_attention.linear_kv_up_proj.weight + module.decoder.layers.0.self_attention.linear_q_proj.weight + module.embedding.word_embeddings.weight + module.decoder.layers.0.self_attention.linear_kv_down_proj.weight + module.decoder.layers.0.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.0.self_attention.linear_proj.weight + module.decoder.layers.0.input_layernorm.weight +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 40 +Params for bucket 1 (40370176 elements, 40370176 padded size): + module.decoder.layers.23.mlp.experts.linear_fc2.weight6 + module.decoder.layers.23.mlp.experts.linear_fc2.weight4 + module.decoder.layers.23.mlp.experts.linear_fc2.weight2 + module.decoder.layers.23.mlp.experts.linear_fc2.weight0 + module.decoder.layers.23.mlp.experts.linear_fc1.weight7 + module.decoder.layers.23.mlp.experts.linear_fc1.weight5 + module.decoder.layers.23.mlp.experts.linear_fc2.weight7 + module.decoder.layers.23.mlp.experts.linear_fc2.weight3 + module.decoder.layers.23.mlp.experts.linear_fc2.weight5 + module.decoder.layers.23.mlp.experts.linear_fc2.weight1 + module.decoder.layers.23.mlp.experts.linear_fc1.weight6 +Params for bucket 2 (40370176 elements, 40370176 padded size): + module.decoder.layers.23.mlp.experts.linear_fc1.weight3 + module.decoder.layers.23.mlp.experts.linear_fc1.weight1 + module.decoder.layers.22.mlp.experts.linear_fc2.weight7 + module.decoder.layers.22.mlp.experts.linear_fc2.weight5 + module.decoder.layers.23.mlp.experts.linear_fc1.weight4 + module.decoder.layers.23.mlp.experts.linear_fc1.weight0 + module.decoder.layers.23.mlp.experts.linear_fc1.weight2 + module.decoder.layers.22.mlp.experts.linear_fc2.weight6 + module.decoder.layers.22.mlp.experts.linear_fc2.weight4 +Params for bucket 3 (40370176 elements, 40370176 padded size): + module.decoder.layers.22.mlp.experts.linear_fc2.weight3 + module.decoder.layers.22.mlp.experts.linear_fc2.weight1 + module.decoder.layers.22.mlp.experts.linear_fc1.weight6 + module.decoder.layers.22.mlp.experts.linear_fc1.weight4 + module.decoder.layers.22.mlp.experts.linear_fc1.weight7 + module.decoder.layers.22.mlp.experts.linear_fc2.weight2 + module.decoder.layers.22.mlp.experts.linear_fc2.weight0 + module.decoder.layers.22.mlp.experts.linear_fc1.weight5 + module.decoder.layers.22.mlp.experts.linear_fc1.weight3 +Params for bucket 4 (40370176 elements, 40370176 padded size): + module.decoder.layers.22.mlp.experts.linear_fc1.weight2 + module.decoder.layers.21.mlp.experts.linear_fc2.weight6 + module.decoder.layers.22.mlp.experts.linear_fc1.weight0 + module.decoder.layers.21.mlp.experts.linear_fc2.weight4 + module.decoder.layers.21.mlp.experts.linear_fc2.weight2 + module.decoder.layers.21.mlp.experts.linear_fc2.weight0 + module.decoder.layers.22.mlp.experts.linear_fc1.weight1 + module.decoder.layers.21.mlp.experts.linear_fc2.weight7 + module.decoder.layers.21.mlp.experts.linear_fc2.weight5 + module.decoder.layers.21.mlp.experts.linear_fc2.weight3 + module.decoder.layers.21.mlp.experts.linear_fc2.weight1 +Params for bucket 5 (40370176 elements, 40370176 padded size): + module.decoder.layers.21.mlp.experts.linear_fc1.weight7 + module.decoder.layers.21.mlp.experts.linear_fc1.weight5 + module.decoder.layers.21.mlp.experts.linear_fc1.weight3 + module.decoder.layers.21.mlp.experts.linear_fc1.weight1 + module.decoder.layers.21.mlp.experts.linear_fc1.weight6 + module.decoder.layers.21.mlp.experts.linear_fc1.weight4 + module.decoder.layers.21.mlp.experts.linear_fc1.weight2 +Params for bucket 6 (40370176 elements, 40370176 padded size): + module.decoder.layers.20.mlp.experts.linear_fc2.weight4 + module.decoder.layers.20.mlp.experts.linear_fc2.weight6 + module.decoder.layers.20.mlp.experts.linear_fc2.weight2 + module.decoder.layers.20.mlp.experts.linear_fc2.weight0 + module.decoder.layers.20.mlp.experts.linear_fc1.weight6 + module.decoder.layers.21.mlp.experts.linear_fc1.weight0 + module.decoder.layers.20.mlp.experts.linear_fc2.weight7 + module.decoder.layers.20.mlp.experts.linear_fc2.weight5 + module.decoder.layers.20.mlp.experts.linear_fc2.weight3 + module.decoder.layers.20.mlp.experts.linear_fc2.weight1 + module.decoder.layers.20.mlp.experts.linear_fc1.weight7 +Params for bucket 7 (40370176 elements, 40370176 padded size): + module.decoder.layers.20.mlp.experts.linear_fc1.weight4 + module.decoder.layers.20.mlp.experts.linear_fc1.weight2 + module.decoder.layers.20.mlp.experts.linear_fc1.weight0 + module.decoder.layers.19.mlp.experts.linear_fc2.weight7 + module.decoder.layers.20.mlp.experts.linear_fc1.weight5 + module.decoder.layers.20.mlp.experts.linear_fc1.weight1 + module.decoder.layers.20.mlp.experts.linear_fc1.weight3 + module.decoder.layers.19.mlp.experts.linear_fc2.weight6 +Params for bucket 8 (40370176 elements, 40370176 padded size): + module.decoder.layers.19.mlp.experts.linear_fc2.weight5 + module.decoder.layers.19.mlp.experts.linear_fc2.weight3 + module.decoder.layers.19.mlp.experts.linear_fc2.weight1 + module.decoder.layers.19.mlp.experts.linear_fc1.weight6 + module.decoder.layers.19.mlp.experts.linear_fc1.weight4 + module.decoder.layers.19.mlp.experts.linear_fc2.weight4 + module.decoder.layers.19.mlp.experts.linear_fc2.weight2 + module.decoder.layers.19.mlp.experts.linear_fc2.weight0 + module.decoder.layers.19.mlp.experts.linear_fc1.weight7 + module.decoder.layers.19.mlp.experts.linear_fc1.weight5 +Params for bucket 9 (40370176 elements, 40370176 padded size): + module.decoder.layers.19.mlp.experts.linear_fc1.weight0 + module.decoder.layers.18.mlp.experts.linear_fc2.weight7 + module.decoder.layers.19.mlp.experts.linear_fc1.weight2 + module.decoder.layers.18.mlp.experts.linear_fc2.weight5 + module.decoder.layers.18.mlp.experts.linear_fc2.weight3 + module.decoder.layers.19.mlp.experts.linear_fc1.weight3 + module.decoder.layers.19.mlp.experts.linear_fc1.weight1 + module.decoder.layers.18.mlp.experts.linear_fc2.weight6 + module.decoder.layers.18.mlp.experts.linear_fc2.weight4 + module.decoder.layers.18.mlp.experts.linear_fc2.weight2 +Params for bucket 10 (40370176 elements, 40370176 padded size): + module.decoder.layers.18.mlp.experts.linear_fc2.weight1 + module.decoder.layers.18.mlp.experts.linear_fc1.weight6 + module.decoder.layers.18.mlp.experts.linear_fc1.weight4 + module.decoder.layers.18.mlp.experts.linear_fc1.weight2 + module.decoder.layers.18.mlp.experts.linear_fc2.weight0 + module.decoder.layers.18.mlp.experts.linear_fc1.weight7 + module.decoder.layers.18.mlp.experts.linear_fc1.weight5 + module.decoder.layers.18.mlp.experts.linear_fc1.weight3 +Params for bucket 11 (40370176 elements, 40370176 padded size): + module.decoder.layers.18.mlp.experts.linear_fc1.weight0 + module.decoder.layers.17.mlp.experts.linear_fc2.weight6 + module.decoder.layers.17.mlp.experts.linear_fc2.weight4 + module.decoder.layers.17.mlp.experts.linear_fc2.weight2 + module.decoder.layers.17.mlp.experts.linear_fc2.weight0 + module.decoder.layers.18.mlp.experts.linear_fc1.weight1 + module.decoder.layers.17.mlp.experts.linear_fc2.weight5 + module.decoder.layers.17.mlp.experts.linear_fc2.weight7 + module.decoder.layers.17.mlp.experts.linear_fc2.weight3 + module.decoder.layers.17.mlp.experts.linear_fc2.weight1 + module.decoder.layers.17.mlp.experts.linear_fc1.weight7 +Params for bucket 12 (40370176 elements, 40370176 padded size): + module.decoder.layers.17.mlp.experts.linear_fc1.weight6 + module.decoder.layers.17.mlp.experts.linear_fc1.weight4 + module.decoder.layers.17.mlp.experts.linear_fc1.weight2 + module.decoder.layers.17.mlp.experts.linear_fc1.weight0 + module.decoder.layers.17.mlp.experts.linear_fc1.weight5 + module.decoder.layers.17.mlp.experts.linear_fc1.weight3 + module.decoder.layers.17.mlp.experts.linear_fc1.weight1 +Params for bucket 13 (40370176 elements, 40370176 padded size): + module.decoder.layers.16.mlp.experts.linear_fc2.weight7 + module.decoder.layers.16.mlp.experts.linear_fc2.weight5 + module.decoder.layers.16.mlp.experts.linear_fc2.weight3 + module.decoder.layers.16.mlp.experts.linear_fc2.weight1 + module.decoder.layers.16.mlp.experts.linear_fc1.weight6 + module.decoder.layers.16.mlp.experts.linear_fc2.weight4 + module.decoder.layers.16.mlp.experts.linear_fc2.weight6 + module.decoder.layers.16.mlp.experts.linear_fc2.weight2 + module.decoder.layers.16.mlp.experts.linear_fc2.weight0 + module.decoder.layers.16.mlp.experts.linear_fc1.weight7 + module.decoder.layers.16.mlp.experts.linear_fc1.weight5 +Params for bucket 14 (40370176 elements, 40370176 padded size): + module.decoder.layers.16.mlp.experts.linear_fc1.weight4 + module.decoder.layers.16.mlp.experts.linear_fc1.weight0 + module.decoder.layers.16.mlp.experts.linear_fc1.weight2 + module.decoder.layers.15.mlp.experts.linear_fc2.weight7 + module.decoder.layers.15.mlp.experts.linear_fc2.weight5 + module.decoder.layers.16.mlp.experts.linear_fc1.weight1 + module.decoder.layers.16.mlp.experts.linear_fc1.weight3 + module.decoder.layers.15.mlp.experts.linear_fc2.weight6 + module.decoder.layers.15.mlp.experts.linear_fc2.weight4 +Params for bucket 15 (40370176 elements, 40370176 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_fc1.weight6 + module.decoder.layers.15.mlp.experts.linear_fc1.weight4 + module.decoder.layers.15.mlp.experts.linear_fc1.weight7 + module.decoder.layers.15.mlp.experts.linear_fc2.weight2 + module.decoder.layers.15.mlp.experts.linear_fc2.weight0 + module.decoder.layers.15.mlp.experts.linear_fc1.weight5 + module.decoder.layers.15.mlp.experts.linear_fc1.weight3 +Params for bucket 16 (40370176 elements, 40370176 padded size): + module.decoder.layers.15.mlp.experts.linear_fc1.weight2 + module.decoder.layers.15.mlp.experts.linear_fc1.weight0 + module.decoder.layers.14.mlp.experts.linear_fc2.weight7 + module.decoder.layers.14.mlp.experts.linear_fc2.weight5 + module.decoder.layers.14.mlp.experts.linear_fc2.weight3 + module.decoder.layers.14.mlp.experts.linear_fc2.weight1 + module.decoder.layers.15.mlp.experts.linear_fc1.weight1 + module.decoder.layers.14.mlp.experts.linear_fc2.weight6 + module.decoder.layers.14.mlp.experts.linear_fc2.weight4 + module.decoder.layers.14.mlp.experts.linear_fc2.weight2 + module.decoder.layers.14.mlp.experts.linear_fc2.weight0 +Params for bucket 17 (40370176 elements, 40370176 padded size): + module.decoder.layers.14.mlp.experts.linear_fc1.weight4 + module.decoder.layers.14.mlp.experts.linear_fc1.weight6 + module.decoder.layers.14.mlp.experts.linear_fc1.weight2 + module.decoder.layers.14.mlp.experts.linear_fc1.weight7 + module.decoder.layers.14.mlp.experts.linear_fc1.weight5 + module.decoder.layers.14.mlp.experts.linear_fc1.weight3 + module.decoder.layers.14.mlp.experts.linear_fc1.weight1 +Params for bucket 18 (40370176 elements, 40370176 padded size): + module.decoder.layers.14.mlp.experts.linear_fc1.weight0 + module.decoder.layers.13.mlp.experts.linear_fc2.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.weight1 + module.decoder.layers.13.mlp.experts.linear_fc1.weight6 + 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 + module.decoder.layers.13.mlp.experts.linear_fc2.weight0 + module.decoder.layers.13.mlp.experts.linear_fc1.weight7 +Params for bucket 19 (40370176 elements, 40370176 padded size): + module.decoder.layers.13.mlp.experts.linear_fc1.weight4 + module.decoder.layers.13.mlp.experts.linear_fc1.weight2 + module.decoder.layers.13.mlp.experts.linear_fc1.weight0 + module.decoder.layers.12.mlp.experts.linear_fc2.weight6 + module.decoder.layers.13.mlp.experts.linear_fc1.weight5 + module.decoder.layers.13.mlp.experts.linear_fc1.weight1 + module.decoder.layers.13.mlp.experts.linear_fc1.weight3 + module.decoder.layers.12.mlp.experts.linear_fc2.weight7 +Params for bucket 20 (40370176 elements, 40370176 padded size): + 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.weight0 + module.decoder.layers.12.mlp.experts.linear_fc1.weight7 + module.decoder.layers.12.mlp.experts.linear_fc1.weight5 + 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.weight1 + module.decoder.layers.12.mlp.experts.linear_fc1.weight6 + module.decoder.layers.12.mlp.experts.linear_fc1.weight4 +Params for bucket 21 (40370176 elements, 40370176 padded size): + module.decoder.layers.12.mlp.experts.linear_fc1.weight3 + module.decoder.layers.12.mlp.experts.linear_fc1.weight1 + module.decoder.layers.11.mlp.experts.linear_fc2.weight7 + module.decoder.layers.11.mlp.experts.linear_fc2.weight5 + module.decoder.layers.11.mlp.experts.linear_fc2.weight3 + module.decoder.layers.12.mlp.experts.linear_fc1.weight0 + module.decoder.layers.12.mlp.experts.linear_fc1.weight2 + module.decoder.layers.11.mlp.experts.linear_fc2.weight6 + module.decoder.layers.11.mlp.experts.linear_fc2.weight4 + module.decoder.layers.11.mlp.experts.linear_fc2.weight2 +Params for bucket 22 (40370176 elements, 40370176 padded size): + module.decoder.layers.11.mlp.experts.linear_fc2.weight1 + module.decoder.layers.11.mlp.experts.linear_fc1.weight6 + module.decoder.layers.11.mlp.experts.linear_fc1.weight4 + module.decoder.layers.11.mlp.experts.linear_fc1.weight2 + module.decoder.layers.11.mlp.experts.linear_fc1.weight5 + module.decoder.layers.11.mlp.experts.linear_fc2.weight0 + module.decoder.layers.11.mlp.experts.linear_fc1.weight7 + module.decoder.layers.11.mlp.experts.linear_fc1.weight3 +Params for bucket 23 (40370176 elements, 40370176 padded size): + module.decoder.layers.10.mlp.experts.linear_fc2.weight7 + module.decoder.layers.10.mlp.experts.linear_fc2.weight5 + module.decoder.layers.11.mlp.experts.linear_fc1.weight0 + module.decoder.layers.10.mlp.experts.linear_fc2.weight3 + module.decoder.layers.10.mlp.experts.linear_fc2.weight1 + module.decoder.layers.10.mlp.experts.linear_fc1.weight7 + module.decoder.layers.11.mlp.experts.linear_fc1.weight1 + module.decoder.layers.10.mlp.experts.linear_fc2.weight6 + module.decoder.layers.10.mlp.experts.linear_fc2.weight4 + module.decoder.layers.10.mlp.experts.linear_fc2.weight2 + module.decoder.layers.10.mlp.experts.linear_fc2.weight0 +Params for bucket 24 (40370176 elements, 40370176 padded size): + module.decoder.layers.10.mlp.experts.linear_fc1.weight3 + module.decoder.layers.10.mlp.experts.linear_fc1.weight5 + module.decoder.layers.10.mlp.experts.linear_fc1.weight1 + module.decoder.layers.10.mlp.experts.linear_fc1.weight6 + module.decoder.layers.10.mlp.experts.linear_fc1.weight2 + module.decoder.layers.10.mlp.experts.linear_fc1.weight4 + module.decoder.layers.10.mlp.experts.linear_fc1.weight0 +Params for bucket 25 (40370176 elements, 40370176 padded size): + module.decoder.layers.9.mlp.experts.linear_fc2.weight7 + module.decoder.layers.9.mlp.experts.linear_fc2.weight3 + module.decoder.layers.9.mlp.experts.linear_fc2.weight5 + module.decoder.layers.9.mlp.experts.linear_fc2.weight1 + module.decoder.layers.9.mlp.experts.linear_fc1.weight6 + module.decoder.layers.9.mlp.experts.linear_fc2.weight6 + module.decoder.layers.9.mlp.experts.linear_fc2.weight4 + module.decoder.layers.9.mlp.experts.linear_fc2.weight2 + module.decoder.layers.9.mlp.experts.linear_fc2.weight0 + module.decoder.layers.9.mlp.experts.linear_fc1.weight7 + module.decoder.layers.9.mlp.experts.linear_fc1.weight5 +Params for bucket 26 (40370176 elements, 40370176 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight4 + module.decoder.layers.9.mlp.experts.linear_fc1.weight2 + module.decoder.layers.9.mlp.experts.linear_fc1.weight0 + module.decoder.layers.8.mlp.experts.linear_fc2.weight7 + module.decoder.layers.8.mlp.experts.linear_fc2.weight5 + module.decoder.layers.9.mlp.experts.linear_fc1.weight3 + module.decoder.layers.9.mlp.experts.linear_fc1.weight1 + module.decoder.layers.8.mlp.experts.linear_fc2.weight6 + module.decoder.layers.8.mlp.experts.linear_fc2.weight4 +Params for bucket 27 (40370176 elements, 40370176 padded size): + module.decoder.layers.8.mlp.experts.linear_fc2.weight3 + module.decoder.layers.8.mlp.experts.linear_fc2.weight1 + module.decoder.layers.8.mlp.experts.linear_fc1.weight6 + module.decoder.layers.8.mlp.experts.linear_fc1.weight4 + module.decoder.layers.8.mlp.experts.linear_fc2.weight0 + module.decoder.layers.8.mlp.experts.linear_fc1.weight7 + module.decoder.layers.8.mlp.experts.linear_fc2.weight2 + module.decoder.layers.8.mlp.experts.linear_fc1.weight5 + module.decoder.layers.8.mlp.experts.linear_fc1.weight3 +Params for bucket 28 (40370176 elements, 40370176 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight2 + module.decoder.layers.8.mlp.experts.linear_fc1.weight0 + 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 + module.decoder.layers.7.mlp.experts.linear_fc2.weight1 + module.decoder.layers.7.mlp.experts.linear_fc2.weight6 + module.decoder.layers.8.mlp.experts.linear_fc1.weight1 + module.decoder.layers.7.mlp.experts.linear_fc2.weight4 + module.decoder.layers.7.mlp.experts.linear_fc2.weight2 + module.decoder.layers.7.mlp.experts.linear_fc2.weight0 +Params for bucket 29 (40370176 elements, 40370176 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight4 + module.decoder.layers.7.mlp.experts.linear_fc1.weight6 + module.decoder.layers.7.mlp.experts.linear_fc1.weight2 + module.decoder.layers.7.mlp.experts.linear_fc1.weight7 + module.decoder.layers.7.mlp.experts.linear_fc1.weight5 + module.decoder.layers.7.mlp.experts.linear_fc1.weight3 + module.decoder.layers.7.mlp.experts.linear_fc1.weight1 +Params for bucket 30 (40370176 elements, 40370176 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight0 + module.decoder.layers.6.mlp.experts.linear_fc2.weight5 + module.decoder.layers.6.mlp.experts.linear_fc2.weight7 + module.decoder.layers.6.mlp.experts.linear_fc2.weight3 + module.decoder.layers.6.mlp.experts.linear_fc2.weight1 + module.decoder.layers.6.mlp.experts.linear_fc1.weight6 + module.decoder.layers.6.mlp.experts.linear_fc2.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.weight0 + module.decoder.layers.6.mlp.experts.linear_fc1.weight7 +Params for bucket 31 (40370176 elements, 40370176 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight2 + module.decoder.layers.6.mlp.experts.linear_fc1.weight4 + module.decoder.layers.6.mlp.experts.linear_fc1.weight0 + module.decoder.layers.5.mlp.experts.linear_fc2.weight7 + module.decoder.layers.6.mlp.experts.linear_fc1.weight5 + module.decoder.layers.6.mlp.experts.linear_fc1.weight3 + module.decoder.layers.6.mlp.experts.linear_fc1.weight1 + module.decoder.layers.5.mlp.experts.linear_fc2.weight6 +Params for bucket 32 (40370176 elements, 40370176 padded size): + module.decoder.layers.5.mlp.experts.linear_fc2.weight5 + module.decoder.layers.5.mlp.experts.linear_fc2.weight1 + module.decoder.layers.5.mlp.experts.linear_fc2.weight3 + module.decoder.layers.5.mlp.experts.linear_fc1.weight6 + module.decoder.layers.5.mlp.experts.linear_fc1.weight4 + module.decoder.layers.5.mlp.experts.linear_fc2.weight2 + module.decoder.layers.5.mlp.experts.linear_fc2.weight4 + module.decoder.layers.5.mlp.experts.linear_fc2.weight0 + module.decoder.layers.5.mlp.experts.linear_fc1.weight7 + module.decoder.layers.5.mlp.experts.linear_fc1.weight5 +Params for bucket 33 (40370176 elements, 40370176 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight0 + module.decoder.layers.5.mlp.experts.linear_fc1.weight2 + module.decoder.layers.4.mlp.experts.linear_fc2.weight6 + module.decoder.layers.4.mlp.experts.linear_fc2.weight4 + module.decoder.layers.4.mlp.experts.linear_fc2.weight2 + module.decoder.layers.5.mlp.experts.linear_fc1.weight3 + module.decoder.layers.4.mlp.experts.linear_fc2.weight7 + module.decoder.layers.5.mlp.experts.linear_fc1.weight1 + module.decoder.layers.4.mlp.experts.linear_fc2.weight5 + module.decoder.layers.4.mlp.experts.linear_fc2.weight3 +Params for bucket 34 (40370176 elements, 40370176 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight0 + module.decoder.layers.4.mlp.experts.linear_fc1.weight5 + module.decoder.layers.4.mlp.experts.linear_fc1.weight7 + module.decoder.layers.4.mlp.experts.linear_fc1.weight3 + module.decoder.layers.4.mlp.experts.linear_fc2.weight1 + module.decoder.layers.4.mlp.experts.linear_fc1.weight6 + module.decoder.layers.4.mlp.experts.linear_fc1.weight4 + module.decoder.layers.4.mlp.experts.linear_fc1.weight2 +Params for bucket 35 (40370176 elements, 40370176 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight1 + module.decoder.layers.3.mlp.experts.linear_fc2.weight5 + 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.weight1 + module.decoder.layers.3.mlp.experts.linear_fc2.weight6 + module.decoder.layers.4.mlp.experts.linear_fc1.weight0 + module.decoder.layers.3.mlp.experts.linear_fc2.weight4 + module.decoder.layers.3.mlp.experts.linear_fc2.weight2 + module.decoder.layers.3.mlp.experts.linear_fc2.weight0 + module.decoder.layers.3.mlp.experts.linear_fc1.weight7 +Params for bucket 36 (40370176 elements, 40370176 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight6 + module.decoder.layers.3.mlp.experts.linear_fc1.weight2 + module.decoder.layers.3.mlp.experts.linear_fc1.weight4 + module.decoder.layers.3.mlp.experts.linear_fc1.weight0 + 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.weight1 +Params for bucket 37 (40370176 elements, 40370176 padded size): + module.decoder.layers.2.mlp.experts.linear_fc2.weight7 + module.decoder.layers.2.mlp.experts.linear_fc2.weight5 + module.decoder.layers.2.mlp.experts.linear_fc2.weight3 + module.decoder.layers.2.mlp.experts.linear_fc2.weight1 + module.decoder.layers.2.mlp.experts.linear_fc1.weight7 + module.decoder.layers.2.mlp.experts.linear_fc1.weight5 + module.decoder.layers.2.mlp.experts.linear_fc2.weight6 + module.decoder.layers.2.mlp.experts.linear_fc2.weight4 + module.decoder.layers.2.mlp.experts.linear_fc2.weight2 + module.decoder.layers.2.mlp.experts.linear_fc2.weight0 + module.decoder.layers.2.mlp.experts.linear_fc1.weight6 +Params for bucket 38 (40370176 elements, 40370176 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight3 + module.decoder.layers.2.mlp.experts.linear_fc1.weight1 + module.decoder.layers.1.mlp.experts.linear_fc2.weight7 + module.decoder.layers.1.mlp.experts.linear_fc2.weight5 + module.decoder.layers.2.mlp.experts.linear_fc1.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight0 + module.decoder.layers.2.mlp.experts.linear_fc1.weight2 + module.decoder.layers.1.mlp.experts.linear_fc2.weight6 + module.decoder.layers.1.mlp.experts.linear_fc2.weight4 +Params for bucket 39 (40370176 elements, 40370176 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_fc1.weight7 + module.decoder.layers.1.mlp.experts.linear_fc1.weight5 + module.decoder.layers.1.mlp.experts.linear_fc1.weight3 + module.decoder.layers.1.mlp.experts.linear_fc2.weight2 + module.decoder.layers.1.mlp.experts.linear_fc2.weight0 + module.decoder.layers.1.mlp.experts.linear_fc1.weight6 + module.decoder.layers.1.mlp.experts.linear_fc1.weight4 +Params for bucket 40 (17301504 elements, 17301504 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight1 + module.decoder.layers.1.mlp.experts.linear_fc1.weight0 +INFO:megatron.core.optimizer:Setting up optimizer with config OptimizerConfig(optimizer='adam', lr=0.0001, min_lr=1e-05, decoupled_lr=None, decoupled_min_lr=None, weight_decay=0.1, fp8_recipe='mxfp8', fp16=False, bf16=True, reuse_grad_buf_for_mxfp8_param_ag=False, params_dtype=torch.bfloat16, use_precision_aware_optimizer=False, store_param_remainders=True, main_grads_dtype=torch.float32, main_params_dtype=torch.float32, exp_avg_dtype=torch.float32, exp_avg_sq_dtype=torch.float32, loss_scale=None, initial_loss_scale=4294967296, min_loss_scale=1.0, loss_scale_window=1000, hysteresis=2, adam_beta1=0.9, adam_beta2=0.999, adam_eps=1e-08, sgd_momentum=0.9, use_distributed_optimizer=True, overlap_param_gather_with_optimizer_step=False, optimizer_cpu_offload=False, optimizer_offload_fraction=1.0, use_torch_optimizer_for_cpu_offload=False, overlap_cpu_optimizer_d2h_h2d=False, pin_cpu_grads=True, pin_cpu_params=True, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=, config_logger_dir='') +INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine +[after model, optimizer, and learning rate scheduler are built] datetime: 2026-08-26 18:35:29 +> building train, validation, and test datasets ... + > datasets target sizes (minimum size): + train: 96 + 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=(96, 0, 0) and config=GPTDatasetConfig(random_seed=1234, sequence_length=8192, blend=(['/workspace/data/wikitext/wikitext103_gpt2_text_document'], None), blend_per_split=None, split='100,0,0', split_matrix=[(0, 1.0), None, None], num_dataset_builder_threads=1, path_to_cache=None, mmap_bin_files=True, mock=False, tokenizer=, mid_level_dataset_surplus=0.005, reset_position_ids=False, reset_attention_mask=False, eod_mask_loss=False, create_attention_mask=True, drop_last_partial_validation_sequence=True, add_extra_token_to_sequence=True, object_storage_cache_path=None) +INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx +INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths +INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers +INFO:megatron.core.datasets.indexed_dataset: Extract the document indices +INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029 +INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350 +INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices +INFO:megatron.core.datasets.gpt_dataset: Load the document index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-document_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the sample index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-sample_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-shuffle_index.npy +INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 14536 +> finished creating GPT datasets ... +[after dataloaders are built] datetime: 2026-08-26 18:35:29 +done with setup ... +(min, max) time across ranks (ms): + model-and-optimizer-setup ......................: (3584.76, 3590.98) + train/valid/test-data-iterators-setup ..........: (68.85, 90.15) +training ... +Overwriting rerun_state_machine.current_iteration from -1 to 0... +[before the start of training step] datetime: 2026-08-26 18:35:29 +/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: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:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") + [2026-08-26 18:35:57] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 27749.6 | throughput per GPU (TFLOP/s/GPU): 4.6 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.092127E+01 | loss scale: 1.0 | grad norm: 30.623 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | +Number of parameters in transformer block in billions: 13.53 +Number of parameters in embedding layers in billions: 0.21 +Total number of parameters in billions: 13.74 +Number of parameters in most loaded shard in billions: 13.7356 +Theoretical memory footprints: weight and optimizer=98244.96 MB +[Rank 0] (after 1 iterations) memory (MB) | allocated: 37181.40869140625 | max allocated: 102778.78173828125 | reserved: 105676.0 | max reserved: 105676.0 +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py: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: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.") + [2026-08-26 18:35:58] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 594.7 | throughput per GPU (TFLOP/s/GPU): 214.5 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.091995E+01 | loss scale: 1.0 | grad norm: 30.567 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:35:58] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 596.3 | throughput per GPU (TFLOP/s/GPU): 214.0 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 9.952354E+00 | loss scale: 1.0 | grad norm: 42.232 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | +Capture CUDA graph for training!!! +CUDA graph capture done!!! + [2026-08-26 18:36:00] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 1769.0 | throughput per GPU (TFLOP/s/GPU): 72.1 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 9.477323E+00 | loss scale: 1.0 | grad norm: 7.143 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:36:01] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 577.4 | throughput per GPU (TFLOP/s/GPU): 221.0 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.119604E+01 | loss scale: 1.0 | grad norm: 396.977 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:36:01] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 582.8 | throughput per GPU (TFLOP/s/GPU): 218.9 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 9.367741E+00 | loss scale: 1.0 | grad norm: 5.090 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:36:02] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 581.5 | throughput per GPU (TFLOP/s/GPU): 219.4 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 8.973585E+00 | loss scale: 1.0 | grad norm: 3.690 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:36:02] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 579.6 | throughput per GPU (TFLOP/s/GPU): 220.1 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.787342E+00 | loss scale: 1.0 | grad norm: 5.890 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:36:03] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 579.0 | throughput per GPU (TFLOP/s/GPU): 220.4 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.612353E+00 | loss scale: 1.0 | grad norm: 3.038 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:36:04] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 578.6 | throughput per GPU (TFLOP/s/GPU): 220.5 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.076241E+00 | loss scale: 1.0 | grad norm: 29.812 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:36:04] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 578.7 | throughput per GPU (TFLOP/s/GPU): 220.4 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.639501E+00 | loss scale: 1.0 | grad norm: 64.594 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:36:05] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 578.9 | throughput per GPU (TFLOP/s/GPU): 220.4 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.624613E+00 | loss scale: 1.0 | grad norm: 71.371 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | +[after training is done] datetime: 2026-08-26 18:36:05 +[rank0]:[W826 18:36:06.392519388 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank3]:[W826 18:36:07.786593783 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank1]:[W826 18:36:07.788536180 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank2]:[W826 18:36:07.790702675 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank7]:[W826 18:36:07.793393604 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank4]:[W826 18:36:07.797344936 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank5]:[W826 18:36:08.084865811 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank6]:[W826 18:36:08.514117245 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +W0826 18:36:10.665000 528750 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers +W0826 18:36:10.665000 528750 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 528836 closing signal SIGTERM +W0826 18:36:10.666000 528750 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 528837 closing signal SIGTERM +W0826 18:36:10.668000 528750 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 528838 closing signal SIGTERM +W0826 18:36:10.670000 528750 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 528839 closing signal SIGTERM +W0826 18:36:10.671000 528750 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 528840 closing signal SIGTERM +W0826 18:36:10.672000 528750 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 528841 closing signal SIGTERM +W0826 18:36:10.674000 528750 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 528842 closing signal SIGTERM +W0826 18:36:10.676000 528750 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 528843 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 528750 got signal: 15 diff --git a/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep2/mem_samples.csv b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep2/mem_samples.csv new file mode 100644 index 0000000000000000000000000000000000000000..ee7b431371a3ce6feecc888d5baf0f860aebdeb2 --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep2/mem_samples.csv @@ -0,0 +1,232 @@ +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 1 +1, 1 +2, 1 +3, 1 +4, 1 +5, 1 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 1 +5, 1 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 4 +5, 4 +6, 4 +7, 4 +0, 4 +1, 629 +2, 629 +3, 629 +4, 629 +5, 629 +6, 629 +7, 629 +0, 927 +1, 929 +2, 929 +3, 939 +4, 937 +5, 949 +6, 915 +7, 941 +0, 1955 +1, 1955 +2, 1955 +3, 1955 +4, 1955 +5, 1955 +6, 1955 +7, 1955 +0, 2343 +1, 2339 +2, 2337 +3, 2331 +4, 2317 +5, 2337 +6, 2333 +7, 2341 +0, 9687 +1, 9819 +2, 9687 +3, 9665 +4, 9585 +5, 9687 +6, 9687 +7, 9433 +0, 28859 +1, 28975 +2, 28819 +3, 28801 +4, 28783 +5, 28807 +6, 28945 +7, 29001 +0, 38967 +1, 39075 +2, 38919 +3, 38903 +4, 38879 +5, 38913 +6, 39029 +7, 39091 +0, 39773 +1, 39857 +2, 39707 +3, 39705 +4, 39705 +5, 39703 +6, 39839 +7, 39857 +0, 39773 +1, 39857 +2, 39707 +3, 39705 +4, 39705 +5, 39703 +6, 39839 +7, 39857 +0, 39783 +1, 39867 +2, 39717 +3, 39717 +4, 39717 +5, 39713 +6, 39849 +7, 39867 +0, 39783 +1, 39867 +2, 39717 +3, 39717 +4, 39717 +5, 39713 +6, 39849 +7, 39867 +0, 109633 +1, 109671 +2, 110011 +3, 110037 +4, 110009 +5, 110033 +6, 109693 +7, 109679 +0, 110101 +1, 110187 +2, 110513 +3, 110511 +4, 110511 +5, 110509 +6, 110167 +7, 110187 +0, 110105 +1, 110191 +2, 110517 +3, 110515 +4, 110515 +5, 110513 +6, 110171 +7, 110191 +0, 112345 +1, 112489 +2, 110743 +3, 110719 +4, 110733 +5, 110739 +6, 112441 +7, 112475 +0, 113191 +1, 113277 +2, 111555 +3, 111553 +4, 111553 +5, 111551 +6, 113257 +7, 113277 +0, 113823 +1, 113907 +2, 112159 +3, 112209 +4, 112193 +5, 112139 +6, 113879 +7, 113913 +0, 121981 +1, 120015 +2, 123973 +3, 128085 +4, 123979 +5, 130155 +6, 128221 +7, 124131 +0, 142137 +1, 142231 +2, 142081 +3, 142079 +4, 142075 +5, 142077 +6, 142211 +7, 142231 +0, 142137 +1, 142231 +2, 142081 +3, 142079 +4, 142075 +5, 142077 +6, 142211 +7, 142231 +0, 142137 +1, 142231 +2, 142081 +3, 142079 +4, 142075 +5, 142077 +6, 142211 +7, 142231 +0, 138747 +1, 139103 +2, 138627 +3, 138369 +4, 138947 +5, 138687 +6, 139083 +7, 139103 +0, 137587 +1, 137355 +2, 136885 +3, 136947 +4, 137525 +5, 137201 +6, 137015 +7, 137099 +0, 1 +1, 4 +2, 4 +3, 4 +4, 68 +5, 136495 +6, 324 +7, 4 diff --git a/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep2/run.env b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep2/run.env new file mode 100644 index 0000000000000000000000000000000000000000..8907152eacecee56da306170fa30edf972163c8e --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep2/run.env @@ -0,0 +1,61 @@ +backend=ultraep +config=DSV2_LITE_L24 +family=tierB +repo=/workspace/Megatron-UltraEP +branch=ultraep-bracing-regions-through-kernels +commit=03d57ef7 +num_layers=24 +hidden=2048 +ffn_hidden=10944 +moe_ffn_hidden=1408 +num_experts=64 +topk=6 +experts_per_rank=8 +expert_params=8650752 +heads=16 +kv_groups=0 +kv_channels=128 +qk_layernorm=1 +shared_ffn=2816 +shared_expert_overlap=0 +mla=1 +dense_first_layers=1 +moe_layers=23 +rotary_base=10000 +rotary_scaling=40 +qkv_bias=0 +router_balancing=seq_aux_loss +aux_loss_coeff=1e-3 +router_pre_softmax=1 +router_fusion=0 +router_dtype=fp32 +device_initiated_grouped_gemm=1 +norm_epsilon=1e-6 +seq_length=8192 +micro_batch_size=1 +global_batch_size=8 +num_microbatches=1 +assignments_per_rank=49152 +mean_tokens_per_expert=6144 +slot_policy=matched +slots_per_rank=8 +capacity_mode=static +capacity=1.0755208333333333 +cuda_graph=1 +probe_mode=0 +async_grad_reduce=0 +branch_role=timing +memory_dump_mode=none +memory_snapshot_path= +cuda_graph_warmup_steps=3 +recompute=none +force_balance=0 +train_iters=12 +ep_size=8 +echo_dispatch_overlap=1 +echo_dispatch_recompute=0 +ultraep_quota_min_tokens=768 +ultraep_weight_sync_join= +profile_enable=0 +kernel_trace=0 +fp8_recipe=mxfp8 diff --git a/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep2/train.log b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep2/train.log new file mode 100644 index 0000000000000000000000000000000000000000..18b229ecbe1c75593738dd49d1c89c5b696fc507 --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep2/train.log @@ -0,0 +1,1610 @@ +============================================================== + DSV2_LITE_L24 / ultraep (slots/rank=8 policy=matched) + E=64 K=6 H'=1408 m=8 + L=24 seq=8192 gbs=8 (1 microbatches) + cap_mode=static cap=1.0755208333333333 cg=1 recompute=none + memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep2/memory_snapshots/snapshot.pickle + repo=/workspace/Megatron-UltraEP @ 03d57ef7 +============================================================== +W0826 18:36:16.170000 531896 torch/distributed/run.py:766] +W0826 18:36:16.170000 531896 torch/distributed/run.py:766] ***************************************** +W0826 18:36:16.170000 531896 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0826 18:36:16.170000 531896 torch/distributed/run.py:766] ***************************************** +using world size: 8, data-parallel size: 8, context-parallel size: 1, hierarchical context-parallel sizes: None, tensor-model-parallel size: 1, pipeline-model-parallel size: 1 +WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer +Number of virtual stages per pipeline stage: None +accumulate and all-reduce gradients in fp32 for bfloat16 data type. +using torch.bfloat16 for parameters ... +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +------------------------ arguments ------------------------ + account_for_embedding_in_pipeline_split ......... False + account_for_loss_in_pipeline_split .............. False + accumulate_allreduce_grads_in_fp32 .............. True + adam_beta1 ...................................... 0.9 + adam_beta2 ...................................... 0.999 + adam_eps ........................................ 1e-08 + add_bias_linear ................................. False + add_position_embedding .......................... False + add_qkv_bias .................................... False + adlr_autoresume ................................. False + adlr_autoresume_interval ........................ 1000 + align_grad_reduce ............................... True + align_param_gather .............................. False + app_tag_run_name ................................ None + app_tag_run_version ............................. 0.0.0 + apply_layernorm_1p .............................. False + apply_query_key_layer_scaling ................... False + apply_residual_connection_post_layernorm ........ False + apply_rope_fusion ............................... False + async_save ...................................... None + async_tensor_model_parallel_allreduce ........... True + attention_backend ............................... AttnBackend.auto + attention_dropout ............................... 0.0 + attention_softmax_in_fp32 ....................... False + auto_detect_ckpt_format ......................... False + barrier_with_L1_time ............................ True + bert_binary_head ................................ True + bert_embedder_type .............................. megatron + bert_load ....................................... None + bf16 ............................................ True + bias_dropout_fusion ............................. True + bias_gelu_fusion ................................ False + bias_swiglu_fusion .............................. True + biencoder_projection_dim ........................ 0 + biencoder_shared_query_context_model ............ False + block_data_path ................................. None + calc_ft_timeouts ................................ False + calculate_per_token_loss ........................ False + check_for_large_grads ........................... False + check_for_nan_in_loss_and_grad .................. False + check_for_spiky_loss ............................ False + check_weight_hash_across_dp_replicas_interval ... None + ckpt_assume_constant_structure .................. False + ckpt_convert_format ............................. None + ckpt_convert_save ............................... None + ckpt_convert_update_legacy_dist_opt_format ...... False + ckpt_format ..................................... torch_dist + ckpt_fully_parallel_load ........................ False + ckpt_fully_parallel_save ........................ True + ckpt_fully_parallel_save_deprecated ............. False + ckpt_step ....................................... None + classes_fraction ................................ 1.0 + clip_grad ....................................... 1.0 + clone_scatter_output_in_embedding ............... True + config_logger_dir ............................... + consumed_train_samples .......................... 0 + consumed_valid_samples .......................... 0 + context_parallel_size ........................... 1 + cp_comm_type .................................... ['p2p'] + create_attention_mask_in_dataloader ............. True + cross_entropy_fusion_impl ....................... native + cross_entropy_loss_fusion ....................... False + cuda_graph_scope ................................ full_iteration + cuda_graph_warmup_steps ......................... 3 + data_args_path .................................. None + data_cache_path ................................. None + data_parallel_random_init ....................... False + data_parallel_sharding_strategy ................. no_shard + data_parallel_size .............................. 8 + data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document'] + data_per_class_fraction ......................... 1.0 + data_sharding ................................... True + dataloader_type ................................. single + ddp_average_in_collective ....................... False + ddp_bucket_size ................................. None + ddp_num_buckets ................................. None + ddp_pad_buckets_for_high_nccl_busbw ............. False + decoder_first_pipeline_num_layers ............... None + decoder_last_pipeline_num_layers ................ None + decoder_num_layers .............................. None + decoder_seq_length .............................. None + decoupled_lr .................................... None + decoupled_min_lr ................................ None + decrease_batch_size_if_needed ................... False + defer_embedding_wgrad_compute ................... False + delay_wgrad_compute ............................. False + deprecated_use_mcore_models ..................... True + deterministic_mode .............................. False + dino_bottleneck_size ............................ 256 + dino_freeze_last_layer .......................... 1 + dino_head_hidden_size ........................... 2048 + dino_local_crops_number ......................... 10 + dino_local_img_size ............................. 96 + dino_norm_last_layer ............................ False + dino_teacher_temp ............................... 0.07 + dino_warmup_teacher_temp ........................ 0.04 + dino_warmup_teacher_temp_epochs ................. 30 + disable_bf16_reduced_precision_matmul ........... False + disable_mamba_mem_eff_path ...................... False + disable_straggler_on_startup .................... False + dist_ckpt_format_deprecated ..................... None + dist_ckpt_strictness ............................ assume_ok_unexpected + distribute_saved_activations .................... False + distributed_backend ............................. nccl + distributed_timeout_minutes ..................... 10 + embedding_init_method_std ....................... None + embedding_path .................................. None + empty_unused_memory_level ....................... 0 + enable_cuda_graph ............................... True + enable_experimental ............................. False + enable_ft_package ............................... False + enable_gloo_process_groups ...................... True + enable_msc ...................................... True + enable_one_logger ............................... True + encoder_num_layers .............................. 24 + encoder_seq_length .............................. 8192 + end_weight_decay ................................ 0.1 + eod_mask_loss ................................... False + error_injection_rate ............................ 0 + error_injection_type ............................ transient_error + eval_interval ................................... 1000000 + eval_iters ...................................... 0 + evidence_data_path .............................. None + exit_duration_in_mins ........................... None + exit_interval ................................... None + exit_on_missing_checkpoint ...................... False + exit_signal_handler ............................. False + exp_avg_dtype ................................... torch.float32 + exp_avg_sq_dtype ................................ torch.float32 + expert_model_parallel_size ...................... 8 + expert_tensor_parallel_size ..................... 1 + export_force_local_attention .................... False + export_kd_cfg ................................... None + export_kd_teacher_ckpt_format ................... None + export_kd_teacher_load .......................... None + export_kv_cache_quant ........................... False + export_legacy_megatron .......................... False + export_model_type ............................... GPTModel + export_moe_apply_probs_on_input ................. False + export_qk_l2_norm ............................... False + export_quant_cfg ................................ None + export_real_quant_cfg ........................... None + export_te_mcore_model ........................... False + external_cuda_graph ............................. False + ffn_hidden_size ................................. 10944 + finetune ........................................ False + finetune_data_split ............................. train + finetune_hf_dataset ............................. None + first_last_layers_bf16 .......................... False + flash_decode .................................... False + fp16 ............................................ False + fp16_lm_cross_entropy ........................... False + fp32_residual_connection ........................ False + fp8 ............................................. e4m3 + fp8_amax_compute_algo ........................... most_recent + fp8_amax_history_len ............................ 1 + fp8_interval .................................... 1 + fp8_margin ...................................... 0 + fp8_param_gather ................................ False + fp8_recipe ...................................... mxfp8 + fp8_wgrad ....................................... True + fsdp_double_buffer .............................. False + global_batch_size ............................... 8 + grad_reduce_in_bf16 ............................. False + gradient_accumulation_fusion .................... True + gradient_reduce_div_fusion ...................... True + group_query_attention ........................... False + head_lr_mult .................................... 1.0 + heterogeneous_layers_config_encoded_json ........ None + heterogeneous_layers_config_path ................ None + hidden_dropout .................................. 0.0 + hidden_size ..................................... 2048 + hierarchical_context_parallel_sizes ............. None + high_priority_stream_groups ..................... [] + hybrid_attention_ratio .......................... 0.0 + hybrid_mlp_ratio ................................ 0.0 + hybrid_override_pattern ......................... None + hysteresis ...................................... 2 + ict_head_size ................................... None + ict_load ........................................ None + img_h ........................................... 224 + img_w ........................................... 224 + indexer_batch_size .............................. 128 + indexer_log_interval ............................ 1000 + inference_batch_times_seqlen_threshold .......... -1 + inference_dynamic_batching ...................... False + inference_dynamic_batching_buffer_guaranteed_fraction 0.2 + inference_dynamic_batching_buffer_overflow_factor None + inference_dynamic_batching_buffer_size_gb ....... 40.0 + inference_dynamic_batching_chunk_size ........... 256 + inference_dynamic_batching_max_requests_override None + inference_dynamic_batching_max_tokens_override .. None + inference_dynamic_batching_num_cuda_graphs ...... 16 + inference_max_batch_size ........................ 8 + inference_max_seq_length ........................ 2560 + inference_rng_tracker ........................... False + init_method_std ................................. 0.01 + init_method_xavier_uniform ...................... False + init_model_with_meta_device ..................... False + initial_loss_scale .............................. 4294967296 + inprocess_active_world_size ..................... 8 + inprocess_barrier_timeout ....................... 120 + inprocess_completion_timeout .................... 120 + inprocess_empty_cuda_cache ...................... False + inprocess_granularity ........................... node + inprocess_hard_timeout .......................... 90 + inprocess_heartbeat_interval .................... 30 + inprocess_heartbeat_timeout ..................... 60 + inprocess_last_call_wait ........................ 1 + inprocess_max_iterations ........................ None + inprocess_monitor_process_interval .............. 1.0 + inprocess_monitor_thread_interval ............... 1.0 + inprocess_progress_watchdog_interval ............ 1.0 + inprocess_restart ............................... False + inprocess_soft_timeout .......................... 60 + inprocess_termination_grace_time ................ 1 + is_hybrid_model ................................. False + iter_per_epoch .................................. 1250 + iterations_to_skip .............................. [] + keep_fp8_transpose_cache_when_using_custom_fsdp . False + kitchen_config_file ............................. None + kitchen_recipe_number ........................... None + kv_channels ..................................... 128 + kv_lora_rank .................................... 512 + lazy_mpu_init ................................... None + load ............................................ None + load_main_params_from_ckpt ...................... None + load_model_opt_format ........................... False + local_rank ...................................... 0 + log_energy ...................................... False + log_interval .................................... 1 + log_loss_scale_to_tensorboard ................... True + log_memory_to_tensorboard ....................... False + log_num_zeros_in_grad ........................... False + log_params_norm ................................. False + log_progress .................................... False + log_straggler ................................... False + log_throughput .................................. True + log_timers_to_tensorboard ....................... False + log_validation_ppl_to_tensorboard ............... False + log_world_size_to_tensorboard ................... False + logging_level ................................... None + loss_scale ...................................... None + loss_scale_window ............................... 1000 + lr .............................................. 0.0001 + lr_decay_iters .................................. 8 + lr_decay_samples ................................ None + lr_decay_style .................................. cosine + lr_warmup_fraction .............................. None + lr_warmup_init .................................. 0.0 + lr_warmup_iters ................................. 5 + lr_warmup_samples ............................... 0 + lr_wsd_decay_iters .............................. None + lr_wsd_decay_samples ............................ None + lr_wsd_decay_style .............................. exponential + main_grads_dtype ................................ torch.float32 + main_params_dtype ............................... torch.float32 + make_vocab_size_divisible_by .................... 128 + mamba_head_dim .................................. 64 + mamba_num_groups ................................ 8 + mamba_num_heads ................................. None + mamba_state_dim ................................. 128 + manual_gc ....................................... False + manual_gc_eval .................................. True + manual_gc_interval .............................. 0 + mask_factor ..................................... 1.0 + mask_prob ....................................... 0.15 + mask_type ....................................... random + masked_softmax_fusion ........................... False + max_position_embeddings ......................... 32768 + max_tokens_to_oom ............................... 12000 + memory_snapshot_path ............................ snapshot.pickle + merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt + micro_batch_size ................................ 1 + microbatch_group_size_per_vp_stage .............. None + mid_level_dataset_surplus ....................... 0.005 + min_loss_scale .................................. 1.0 + min_lr .......................................... 1e-05 + mlp_chunks_for_prefill .......................... 1 + mmap_bin_files .................................. True + mock_data ....................................... False + moe_apply_probs_on_input ........................ False + moe_aux_loss_coeff .............................. 0.001 + moe_capacity_monitor ............................ True + moe_deepep_num_sms .............................. 20 + moe_echo_enable_random_offloading ............... False + moe_echo_expert_dispatch_overlap ................ False + moe_echo_expert_dispatcher_type ................. hybridep + moe_echo_recompute_expert_dispatch .............. False + moe_enable_deepep ............................... False + moe_enable_echo ................................. False + moe_enable_hybridep ............................. True + moe_enable_ultraep .............................. True + moe_expert_capacity_factor ...................... None + moe_extended_tp ................................. False + moe_ffn_hidden_size ............................. 1408 + moe_grouped_gemm ................................ True + moe_hybridep_num_sms ............................ 16 + moe_input_jitter_eps ............................ None + moe_layer_freq .................................. [0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] + moe_layer_recompute ............................. False + moe_num_echo_experts ............................ None + moe_num_redundant_experts_per_rank .............. 8 + moe_pad_expert_input_to_capacity ................ False + moe_per_layer_logging ........................... False + moe_permute_fusion .............................. False + moe_received_token_capacity ..................... 1.0755208333333333 + moe_router_bias_update_rate ..................... 0.001 + moe_router_dtype ................................ fp32 + moe_router_enable_expert_bias ................... False + moe_router_force_load_balancing ................. False + moe_router_group_topk ........................... None + moe_router_load_balancing_type .................. seq_aux_loss + moe_router_num_groups ........................... None + moe_router_padding_for_fp8 ...................... False + moe_router_pre_softmax .......................... True + moe_router_score_function ....................... softmax + moe_router_topk ................................. 6 + moe_router_topk_scaling_factor .................. None + moe_shared_expert_intermediate_size ............. 2816 + moe_shared_expert_overlap ....................... False + moe_token_dispatcher_type ....................... flex + moe_token_drop_policy ........................... probs + moe_ultraep_max_microbatches .................... None + moe_ultraep_overlap_weight_sync ................. True + moe_ultraep_weight_sync_join .................... before_dispatch + moe_upcycling_granularity ....................... 1 + moe_use_device_initiated_grouped_gemm ........... True + moe_use_legacy_grouped_gemm ..................... False + moe_use_upcycling ............................... False + moe_z_loss_coeff ................................ None + mrope_section ................................... None + mscale .......................................... 0.707 + mscale_all_dim .................................. 0.707 + mtp_loss_scaling_factor ......................... 0.1 + mtp_num_layers .................................. None + multi_latent_attention .......................... True + nccl_all_reduce_for_prefill ..................... False + nccl_communicator_config_path ................... None + nccl_ub ......................................... False + no_load_optim ................................... None + no_load_rng ..................................... None + no_persist_layer_norm ........................... False + no_rope_freq .................................... None + no_save_optim ................................... None + no_save_rng ..................................... None + non_persistent_ckpt_type ........................ None + non_persistent_global_ckpt_dir .................. None + non_persistent_local_ckpt_algo .................. fully_parallel + non_persistent_local_ckpt_dir ................... None + non_persistent_save_interval .................... None + norm_epsilon .................................... 1e-06 + normalization ................................... RMSNorm + num_attention_heads ............................. 16 + num_channels .................................... 3 + num_classes ..................................... 1000 + num_dataset_builder_threads ..................... 1 + num_distributed_optimizer_instances ............. 1 + num_experts ..................................... 64 + num_layers ...................................... 24 + num_layers_at_end_in_bf16 ....................... 1 + num_layers_at_start_in_bf16 ..................... 1 + num_layers_per_virtual_pipeline_stage ........... None + num_query_groups ................................ 1 + num_virtual_stages_per_pipeline_rank ............ None + num_workers ..................................... 2 + object_storage_cache_path ....................... None + one_logger_async ................................ False + one_logger_project .............................. megatron-lm + one_logger_run_name ............................. None + onnx_safe ....................................... None + openai_gelu ..................................... False + optimizer ....................................... adam + optimizer_cpu_offload ........................... False + optimizer_offload_fraction ...................... 1.0 + output_bert_embeddings .......................... False + overlap_cpu_optimizer_d2h_h2d ................... False + overlap_grad_reduce ............................. True + overlap_moe_expert_parallel_comm ................ False + overlap_p2p_comm ................................ False + overlap_p2p_comm_warmup_flush ................... False + overlap_param_gather ............................ True + overlap_param_gather_with_optimizer_step ........ False + override_opt_param_scheduler .................... False + params_dtype .................................... torch.bfloat16 + patch_dim ....................................... 16 + per_split_data_args_path ........................ None + perform_initialization .......................... True + pin_cpu_grads ................................... True + pin_cpu_params .................................. True + pipeline_model_parallel_comm_backend ............ None + pipeline_model_parallel_layout .................. None + pipeline_model_parallel_size .................... 1 + position_embedding_type ......................... rope + pretrained_checkpoint ........................... None + profile ......................................... False + profile_ranks ................................... [0] + profile_step_end ................................ 12 + profile_step_start .............................. 10 + q_lora_rank ..................................... None + qk_head_dim ..................................... 128 + qk_l2_norm ...................................... False + qk_layernorm .................................... True + qk_pos_emb_head_dim ............................. 64 + query_in_block_prob ............................. 0.1 + rampup_batch_size ............................... None + rank ............................................ 0 + recompute_granularity ........................... None + recompute_method ................................ None + recompute_modules ............................... None + recompute_num_layers ............................ None + record_memory_history ........................... False + relative_attention_max_distance ................. 128 + relative_attention_num_buckets .................. 32 + replication ..................................... False + replication_factor .............................. 2 + replication_jump ................................ None + rerun_mode ...................................... validate_results + reset_attention_mask ............................ False + reset_position_ids .............................. False + result_rejected_tracker_filename ................ None + retriever_report_topk_accuracies ................ [] + retriever_score_scaling ......................... False + retriever_seq_length ............................ 256 + retro_add_retriever ............................. False + retro_attention_gate ............................ 1 + retro_cyclic_train_iters ........................ None + retro_encoder_attention_dropout ................. 0.1 + retro_encoder_hidden_dropout .................... 0.1 + retro_encoder_layers ............................ 2 + retro_num_neighbors ............................. 2 + retro_num_retrieved_chunks ...................... 2 + retro_project_dir ............................... None + retro_verify_neighbor_count ..................... True + reuse_grad_buf_for_mxfp8_param_ag ............... False + rope_scaling_factor ............................. 8.0 + rope_type ....................................... None + rotary_base ..................................... 10000 + rotary_interleaved .............................. False + rotary_percent .................................. 1.0 + rotary_scaling_factor ........................... 40.0 + rotary_seq_len_interpolation_factor ............. None + run_workload_inspector_server ................... False + sample_rate ..................................... 1.0 + save ............................................ None + save_interval ................................... None + save_retain_interval ............................ None + scatter_gather_tensors_in_pipeline .............. True + seed ............................................ 1234 + seq_length ...................................... 8192 + sequence_parallel ............................... False + sft ............................................. False + sft_tokenizer_prompt_format ..................... nemotron-h-aligned + sgd_momentum .................................... 0.9 + short_seq_prob .................................. 0.1 + skip_train ...................................... False + skipped_train_samples ........................... 0 + spec ............................................ None + split ........................................... 100,0,0 + squared_relu .................................... False + start_weight_decay .............................. 0.1 + straggler_ctrlr_port ............................ 65535 + straggler_minmax_count .......................... 1 + suggested_communication_unit_size ............... None + swiglu .......................................... True + swin_backbone_type .............................. tiny + symmetric_ar_type ............................... None + te_rng_tracker .................................. True + tensor_model_parallel_size ...................... 1 + tensorboard_dir ................................. None + tensorboard_log_interval ........................ 1 + tensorboard_queue_size .......................... 1000 + test_data_path .................................. None + test_mode ....................................... False + tiktoken_num_special_tokens ..................... 1000 + tiktoken_pattern ................................ None + tiktoken_special_tokens ......................... None + timing_log_level ................................ 0 + timing_log_option ............................... minmax + titles_data_path ................................ None + tokenizer_model ................................. None + tokenizer_type .................................. GPT2BPETokenizer + torch_fsdp2_reshard_after_forward ............... True + tp_comm_bootstrap_backend ....................... nccl + tp_comm_bulk_dgrad .............................. True + tp_comm_bulk_wgrad .............................. True + tp_comm_overlap ................................. False + tp_comm_overlap_ag .............................. True + tp_comm_overlap_cfg ............................. None + tp_comm_overlap_rs .............................. True + tp_comm_overlap_rs_dgrad ........................ False + tp_comm_split_ag ................................ True + tp_comm_split_rs ................................ True + train_data_path ................................. None + train_iters ..................................... 12 + train_samples ................................... None + train_sync_interval ............................. None + transformer_impl ................................ transformer_engine + transformer_pipeline_model_parallel_size ........ 1 + untie_embeddings_and_output_weights ............. True + use_checkpoint_args ............................. False + use_checkpoint_opt_param_scheduler .............. False + use_cpu_initialization .......................... None + use_custom_fsdp ................................. False + use_dist_ckpt ................................... True + use_dist_ckpt_deprecated ........................ False + use_distributed_optimizer ....................... True + use_flash_attn .................................. False + use_legacy_models ............................... False + use_mp_args_from_checkpoint_args ................ False + use_one_sent_docs ............................... False + use_persistent_ckpt_worker ...................... False + use_precision_aware_optimizer ................... False + use_pytorch_profiler ............................ False + use_ring_exchange_p2p ........................... False + use_rope_scaling ................................ False + use_rotary_position_embeddings .................. False + use_sharp ....................................... False + use_tokenizer_model_from_checkpoint_args ........ True + use_torch_fsdp2 ................................. False + use_torch_optimizer_for_cpu_offload ............. False + use_tp_pp_dp_mapping ............................ False + v_head_dim ...................................... 128 + valid_data_path ................................. None + variable_seq_lengths ............................ False + virtual_pipeline_model_parallel_size ............ None + vision_backbone_type ............................ vit + vision_pretraining .............................. False + vision_pretraining_type ......................... classify + vocab_extra_ids ................................. 0 + vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json + vocab_size ...................................... None + wandb_exp_name .................................. + wandb_project ................................... + wandb_save_dir .................................. + weight_decay .................................... 0.1 + weight_decay_incr_style ......................... constant + wgrad_deferral_limit ............................ 0 + world_size ...................................... 8 + yaml_cfg ........................................ None +-------------------- end of arguments --------------------- +INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1 +> building GPT2BPETokenizer tokenizer ... +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") + > 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") +/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 +[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 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +2 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 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : [Gloo] Rank 71 + is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 3 is connected to 7 peer ranks. [Gloo] Rank Expected number of connected peer ranks is : 5[Gloo] Rank 74 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 +> initialized tensor model parallel with size 1 +> initialized pipeline model parallel with size 1 +> setting random seeds to 1234 ... +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank [Gloo] Rank 0[Gloo] Rank is connected to 000 is connected to is connected to peer ranks. 00Expected number of connected peer ranks is : peer ranks. peer ranks. 0Expected number of connected peer ranks is : Expected number of connected peer ranks is : +00 + +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[rank6]:[W826 18:36:34.322600135 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. +[rank3]:[W826 18:36:34.322627105 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank4]:[W826 18:36:34.322635313 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. +[rank5]:[W826 18:36:34.322650960 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. +[rank2]:[W826 18:36:34.322721413 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. +[rank1]:[W826 18:36:34.322729711 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. +[rank7]:[W826 18:36:34.324456168 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: 0.034 seconds +WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations. +> compiling and loading fused kernels ... +[rank0]:[W826 18:36:34.492157291 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: 3.649 seconds +time to initialize megatron (seconds): 6.986 +[after megatron is initialized] datetime: 2026-08-26 18:36:40 +building GPT model ... +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +[INFO] Use MNNVL fabric: False +[WARN] MAX_NUM_NVL_PEERS is not set. Using detected value 8. +NVSHMEM v3.7.2 +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +UltraEP: NVL domain size <= 8; forcing weight_sync plan mode to direct (was adaptive). + > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 4188125184 +INFO:megatron.core.distributed.distributed_data_parallel:Setting up DistributedDataParallel with config DistributedDataParallelConfig(grad_reduce_in_fp32=True, overlap_grad_reduce=True, overlap_param_gather=True, align_param_gather=False, use_distributed_optimizer=True, num_distributed_optimizer_instances=1, check_for_nan_in_grad=False, check_for_large_grads=False, bucket_size=40000000, pad_buckets_for_high_nccl_busbw=False, average_in_collective=False, fp8_param_gather=False, reuse_grad_buf_for_mxfp8_param_ag=False, use_custom_fsdp=False, data_parallel_sharding_strategy='no_shard', gradient_reduce_div_fusion=True, suggested_communication_unit_size=None, preserve_fp32_weights=True, keep_fp8_transpose_cache_when_using_custom_fsdp=False, nccl_ub=False, fsdp_double_buffer=False) +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 20 +Params for bucket 1 (103022592 elements, 103022592 padded size): + module.output_layer.weight +Params for bucket 2 (48503296 elements, 48503296 padded size): + module.decoder.final_layernorm.weight + module.decoder.layers.23.pre_mlp_layernorm.weight + module.decoder.layers.23.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.23.mlp.router.weight + module.decoder.layers.23.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.23.self_attention.linear_kv_down_proj.weight + module.decoder.layers.23.self_attention.linear_q_proj.weight + module.decoder.layers.23.self_attention.linear_proj.weight + module.decoder.layers.23.input_layernorm.weight + module.decoder.layers.22.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.23.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.23.self_attention.linear_kv_up_proj.weight + module.decoder.layers.22.mlp.shared_experts.linear_fc2.weight +Params for bucket 3 (40901632 elements, 40901632 padded size): + module.decoder.layers.22.self_attention.linear_kv_up_proj.weight + module.decoder.layers.21.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.21.mlp.router.weight + module.decoder.layers.21.self_attention.linear_kv_up_proj.weight + module.decoder.layers.21.self_attention.linear_q_proj.weight + module.decoder.layers.21.self_attention.linear_kv_down_proj.weight + module.decoder.layers.22.mlp.router.weight + module.decoder.layers.22.self_attention.linear_kv_down_proj.weight + module.decoder.layers.22.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.22.pre_mlp_layernorm.weight + module.decoder.layers.22.self_attention.linear_q_proj.weight + module.decoder.layers.22.self_attention.linear_proj.weight + module.decoder.layers.22.input_layernorm.weight + module.decoder.layers.21.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.21.pre_mlp_layernorm.weight + module.decoder.layers.21.self_attention.linear_kv_up_proj.layer_norm_weight +Params for bucket 4 (41163264 elements, 41163264 padded size): + module.decoder.layers.20.mlp.router.weight + module.decoder.layers.20.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.20.self_attention.linear_kv_up_proj.weight + module.decoder.layers.19.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.21.self_attention.linear_proj.weight + module.decoder.layers.20.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.21.input_layernorm.weight + module.decoder.layers.20.pre_mlp_layernorm.weight + module.decoder.layers.20.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.20.self_attention.linear_kv_down_proj.weight + module.decoder.layers.20.self_attention.linear_q_proj.weight + module.decoder.layers.20.self_attention.linear_proj.weight + module.decoder.layers.20.input_layernorm.weight +Params for bucket 5 (42734080 elements, 42734080 padded size): + module.decoder.layers.19.self_attention.linear_kv_up_proj.weight + module.decoder.layers.19.mlp.router.weight + module.decoder.layers.18.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.19.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.19.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.19.pre_mlp_layernorm.weight + module.decoder.layers.19.self_attention.linear_kv_down_proj.weight + module.decoder.layers.19.self_attention.linear_q_proj.weight + module.decoder.layers.19.self_attention.linear_proj.weight + module.decoder.layers.19.input_layernorm.weight + module.decoder.layers.18.mlp.shared_experts.linear_fc1.weight +Params for bucket 6 (40901632 elements, 40901632 padded size): + module.decoder.layers.18.mlp.router.weight + module.decoder.layers.18.self_attention.linear_kv_up_proj.weight + module.decoder.layers.17.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.17.mlp.router.weight + module.decoder.layers.17.self_attention.linear_kv_up_proj.weight + module.decoder.layers.17.self_attention.linear_q_proj.weight + module.decoder.layers.18.self_attention.linear_kv_down_proj.weight + module.decoder.layers.18.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.18.pre_mlp_layernorm.weight + module.decoder.layers.18.self_attention.linear_q_proj.weight + module.decoder.layers.18.self_attention.linear_proj.weight + module.decoder.layers.18.input_layernorm.weight + module.decoder.layers.17.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.17.pre_mlp_layernorm.weight + module.decoder.layers.17.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.17.self_attention.linear_kv_down_proj.weight +Params for bucket 7 (41163264 elements, 41163264 padded size): + module.decoder.layers.16.mlp.router.weight + module.decoder.layers.16.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.16.self_attention.linear_kv_up_proj.weight + module.decoder.layers.15.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.17.self_attention.linear_proj.weight + module.decoder.layers.17.input_layernorm.weight + module.decoder.layers.16.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.16.pre_mlp_layernorm.weight + module.decoder.layers.16.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.16.self_attention.linear_kv_down_proj.weight + module.decoder.layers.16.self_attention.linear_q_proj.weight + module.decoder.layers.16.self_attention.linear_proj.weight + module.decoder.layers.16.input_layernorm.weight +Params for bucket 8 (42734080 elements, 42734080 padded size): + module.decoder.layers.15.self_attention.linear_kv_up_proj.weight + module.decoder.layers.15.mlp.router.weight + module.decoder.layers.14.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.15.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.15.pre_mlp_layernorm.weight + module.decoder.layers.15.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.15.self_attention.linear_kv_down_proj.weight + module.decoder.layers.15.self_attention.linear_q_proj.weight + module.decoder.layers.15.self_attention.linear_proj.weight + module.decoder.layers.15.input_layernorm.weight + module.decoder.layers.14.mlp.shared_experts.linear_fc1.weight +Params for bucket 9 (40901632 elements, 40901632 padded size): + module.decoder.layers.14.mlp.router.weight + module.decoder.layers.14.self_attention.linear_kv_up_proj.weight + module.decoder.layers.13.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.13.mlp.router.weight + module.decoder.layers.13.self_attention.linear_kv_up_proj.weight + module.decoder.layers.13.self_attention.linear_q_proj.weight + module.decoder.layers.14.pre_mlp_layernorm.weight + module.decoder.layers.14.self_attention.linear_kv_down_proj.weight + module.decoder.layers.14.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.14.self_attention.linear_q_proj.weight + module.decoder.layers.14.self_attention.linear_proj.weight + module.decoder.layers.14.input_layernorm.weight + module.decoder.layers.13.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.13.pre_mlp_layernorm.weight + module.decoder.layers.13.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.13.self_attention.linear_kv_down_proj.weight +Params for bucket 10 (41163264 elements, 41163264 padded size): + module.decoder.layers.12.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.12.pre_mlp_layernorm.weight + module.decoder.layers.12.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.12.self_attention.linear_kv_down_proj.weight + module.decoder.layers.11.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.13.self_attention.linear_proj.weight + module.decoder.layers.13.input_layernorm.weight + module.decoder.layers.12.mlp.router.weight + module.decoder.layers.12.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.12.self_attention.linear_kv_up_proj.weight + module.decoder.layers.12.self_attention.linear_q_proj.weight + module.decoder.layers.12.self_attention.linear_proj.weight + module.decoder.layers.12.input_layernorm.weight +Params for bucket 11 (42734080 elements, 42734080 padded size): + module.decoder.layers.11.self_attention.linear_kv_up_proj.weight + module.decoder.layers.11.mlp.router.weight + module.decoder.layers.10.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.11.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.11.pre_mlp_layernorm.weight + module.decoder.layers.11.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.11.self_attention.linear_kv_down_proj.weight + module.decoder.layers.11.self_attention.linear_q_proj.weight + module.decoder.layers.11.self_attention.linear_proj.weight + module.decoder.layers.11.input_layernorm.weight + module.decoder.layers.10.mlp.shared_experts.linear_fc1.weight +Params for bucket 12 (40901632 elements, 40901632 padded size): + module.decoder.layers.10.mlp.router.weight + module.decoder.layers.10.self_attention.linear_kv_up_proj.weight + module.decoder.layers.9.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.9.pre_mlp_layernorm.weight + module.decoder.layers.9.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.9.self_attention.linear_kv_down_proj.weight + module.decoder.layers.10.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.10.pre_mlp_layernorm.weight + module.decoder.layers.10.self_attention.linear_kv_down_proj.weight + module.decoder.layers.10.self_attention.linear_q_proj.weight + module.decoder.layers.10.self_attention.linear_proj.weight + module.decoder.layers.10.input_layernorm.weight + module.decoder.layers.9.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.9.mlp.router.weight + module.decoder.layers.9.self_attention.linear_kv_up_proj.weight + module.decoder.layers.9.self_attention.linear_q_proj.weight +Params for bucket 13 (41163264 elements, 41163264 padded size): + module.decoder.layers.9.input_layernorm.weight + module.decoder.layers.8.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.8.mlp.router.weight + module.decoder.layers.8.self_attention.linear_kv_up_proj.weight + module.decoder.layers.7.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.9.self_attention.linear_proj.weight + module.decoder.layers.8.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.8.pre_mlp_layernorm.weight + module.decoder.layers.8.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.8.self_attention.linear_kv_down_proj.weight + module.decoder.layers.8.self_attention.linear_q_proj.weight + module.decoder.layers.8.self_attention.linear_proj.weight + module.decoder.layers.8.input_layernorm.weight +Params for bucket 14 (42734080 elements, 42734080 padded size): + module.decoder.layers.7.self_attention.linear_kv_up_proj.weight + module.decoder.layers.7.mlp.router.weight + module.decoder.layers.6.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.7.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.7.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.7.pre_mlp_layernorm.weight + module.decoder.layers.7.self_attention.linear_kv_down_proj.weight + module.decoder.layers.7.self_attention.linear_q_proj.weight + module.decoder.layers.7.self_attention.linear_proj.weight + module.decoder.layers.7.input_layernorm.weight + module.decoder.layers.6.mlp.shared_experts.linear_fc1.weight +Params for bucket 15 (40901632 elements, 40901632 padded size): + module.decoder.layers.6.mlp.router.weight + module.decoder.layers.6.self_attention.linear_kv_up_proj.weight + module.decoder.layers.5.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.5.mlp.router.weight + module.decoder.layers.5.self_attention.linear_kv_up_proj.weight + module.decoder.layers.5.self_attention.linear_q_proj.weight + module.decoder.layers.6.self_attention.linear_kv_down_proj.weight + module.decoder.layers.6.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.6.pre_mlp_layernorm.weight + module.decoder.layers.6.self_attention.linear_q_proj.weight + module.decoder.layers.6.self_attention.linear_proj.weight + module.decoder.layers.6.input_layernorm.weight + module.decoder.layers.5.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.5.pre_mlp_layernorm.weight + module.decoder.layers.5.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.5.self_attention.linear_kv_down_proj.weight +Params for bucket 16 (41163264 elements, 41163264 padded size): + module.decoder.layers.4.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.5.input_layernorm.weight + module.decoder.layers.4.pre_mlp_layernorm.weight + module.decoder.layers.4.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.4.self_attention.linear_kv_down_proj.weight + module.decoder.layers.4.self_attention.linear_q_proj.weight + module.decoder.layers.4.self_attention.linear_proj.weight + module.decoder.layers.4.input_layernorm.weight + module.decoder.layers.5.self_attention.linear_proj.weight + module.decoder.layers.4.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.4.mlp.router.weight + module.decoder.layers.4.self_attention.linear_kv_up_proj.weight + module.decoder.layers.3.mlp.shared_experts.linear_fc2.weight +Params for bucket 17 (42734080 elements, 42734080 padded size): + module.decoder.layers.3.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.3.self_attention.linear_kv_up_proj.weight + module.decoder.layers.3.mlp.router.weight + module.decoder.layers.2.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.3.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.3.pre_mlp_layernorm.weight + module.decoder.layers.3.self_attention.linear_kv_down_proj.weight + module.decoder.layers.3.self_attention.linear_q_proj.weight + module.decoder.layers.3.self_attention.linear_proj.weight + module.decoder.layers.3.input_layernorm.weight + module.decoder.layers.2.mlp.shared_experts.linear_fc1.weight +Params for bucket 18 (40901632 elements, 40901632 padded size): + module.decoder.layers.2.pre_mlp_layernorm.weight + module.decoder.layers.2.self_attention.linear_kv_down_proj.weight + module.decoder.layers.2.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.2.self_attention.linear_q_proj.weight + module.decoder.layers.2.self_attention.linear_proj.weight + module.decoder.layers.2.input_layernorm.weight + module.decoder.layers.1.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.1.pre_mlp_layernorm.weight + module.decoder.layers.1.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.1.self_attention.linear_kv_down_proj.weight + module.decoder.layers.2.mlp.router.weight + module.decoder.layers.2.self_attention.linear_kv_up_proj.weight + module.decoder.layers.1.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.1.mlp.router.weight + module.decoder.layers.1.self_attention.linear_kv_up_proj.weight + module.decoder.layers.1.self_attention.linear_q_proj.weight +Params for bucket 19 (71436288 elements, 71436288 padded size): + module.decoder.layers.1.self_attention.linear_proj.weight + module.decoder.layers.1.input_layernorm.weight + module.decoder.layers.0.mlp.linear_fc1.weight + module.decoder.layers.0.mlp.linear_fc2.weight +Params for bucket 20 (116789760 elements, 116789760 padded size): + module.decoder.layers.0.mlp.linear_fc1.layer_norm_weight + module.decoder.layers.0.self_attention.linear_kv_up_proj.weight + module.decoder.layers.0.self_attention.linear_q_proj.weight + module.decoder.layers.0.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.0.self_attention.linear_kv_down_proj.weight + module.decoder.layers.0.self_attention.linear_proj.weight + module.decoder.layers.0.input_layernorm.weight + module.embedding.word_embeddings.weight +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 40 +Params for bucket 1 (40370176 elements, 40370176 padded size): + module.decoder.layers.23.mlp.experts.linear_fc2.weight7 + module.decoder.layers.23.mlp.experts.linear_fc2.weight3 + module.decoder.layers.23.mlp.experts.linear_fc2.weight5 + module.decoder.layers.23.mlp.experts.linear_fc2.weight1 + module.decoder.layers.23.mlp.experts.linear_fc1.weight6 + module.decoder.layers.23.mlp.experts.linear_fc2.weight6 + module.decoder.layers.23.mlp.experts.linear_fc2.weight4 + module.decoder.layers.23.mlp.experts.linear_fc2.weight2 + module.decoder.layers.23.mlp.experts.linear_fc2.weight0 + module.decoder.layers.23.mlp.experts.linear_fc1.weight7 + module.decoder.layers.23.mlp.experts.linear_fc1.weight5 +Params for bucket 2 (40370176 elements, 40370176 padded size): + module.decoder.layers.23.mlp.experts.linear_fc1.weight4 + module.decoder.layers.23.mlp.experts.linear_fc1.weight0 + module.decoder.layers.23.mlp.experts.linear_fc1.weight2 + module.decoder.layers.22.mlp.experts.linear_fc2.weight6 + module.decoder.layers.22.mlp.experts.linear_fc2.weight4 + module.decoder.layers.23.mlp.experts.linear_fc1.weight1 + module.decoder.layers.23.mlp.experts.linear_fc1.weight3 + module.decoder.layers.22.mlp.experts.linear_fc2.weight7 + module.decoder.layers.22.mlp.experts.linear_fc2.weight5 +Params for bucket 3 (40370176 elements, 40370176 padded size): + module.decoder.layers.22.mlp.experts.linear_fc1.weight7 + module.decoder.layers.22.mlp.experts.linear_fc2.weight2 + module.decoder.layers.22.mlp.experts.linear_fc2.weight0 + module.decoder.layers.22.mlp.experts.linear_fc1.weight5 + module.decoder.layers.22.mlp.experts.linear_fc1.weight3 + module.decoder.layers.22.mlp.experts.linear_fc2.weight3 + module.decoder.layers.22.mlp.experts.linear_fc2.weight1 + module.decoder.layers.22.mlp.experts.linear_fc1.weight6 + module.decoder.layers.22.mlp.experts.linear_fc1.weight4 +Params for bucket 4 (40370176 elements, 40370176 padded size): + module.decoder.layers.22.mlp.experts.linear_fc1.weight1 + module.decoder.layers.21.mlp.experts.linear_fc2.weight7 + module.decoder.layers.21.mlp.experts.linear_fc2.weight5 + module.decoder.layers.21.mlp.experts.linear_fc2.weight3 + module.decoder.layers.21.mlp.experts.linear_fc2.weight1 + module.decoder.layers.22.mlp.experts.linear_fc1.weight2 + module.decoder.layers.21.mlp.experts.linear_fc2.weight6 + module.decoder.layers.22.mlp.experts.linear_fc1.weight0 + module.decoder.layers.21.mlp.experts.linear_fc2.weight4 + module.decoder.layers.21.mlp.experts.linear_fc2.weight2 + module.decoder.layers.21.mlp.experts.linear_fc2.weight0 +Params for bucket 5 (40370176 elements, 40370176 padded size): + module.decoder.layers.21.mlp.experts.linear_fc1.weight6 + module.decoder.layers.21.mlp.experts.linear_fc1.weight4 + module.decoder.layers.21.mlp.experts.linear_fc1.weight2 + module.decoder.layers.21.mlp.experts.linear_fc1.weight7 + module.decoder.layers.21.mlp.experts.linear_fc1.weight5 + module.decoder.layers.21.mlp.experts.linear_fc1.weight3 + module.decoder.layers.21.mlp.experts.linear_fc1.weight1 +Params for bucket 6 (40370176 elements, 40370176 padded size): + module.decoder.layers.21.mlp.experts.linear_fc1.weight0 + module.decoder.layers.20.mlp.experts.linear_fc2.weight7 + module.decoder.layers.20.mlp.experts.linear_fc2.weight5 + module.decoder.layers.20.mlp.experts.linear_fc2.weight3 + module.decoder.layers.20.mlp.experts.linear_fc2.weight1 + module.decoder.layers.20.mlp.experts.linear_fc1.weight6 + module.decoder.layers.20.mlp.experts.linear_fc2.weight6 + module.decoder.layers.20.mlp.experts.linear_fc2.weight4 + module.decoder.layers.20.mlp.experts.linear_fc2.weight2 + module.decoder.layers.20.mlp.experts.linear_fc2.weight0 + module.decoder.layers.20.mlp.experts.linear_fc1.weight7 +Params for bucket 7 (40370176 elements, 40370176 padded size): + module.decoder.layers.20.mlp.experts.linear_fc1.weight2 + module.decoder.layers.20.mlp.experts.linear_fc1.weight4 + module.decoder.layers.20.mlp.experts.linear_fc1.weight0 + module.decoder.layers.19.mlp.experts.linear_fc2.weight7 + module.decoder.layers.20.mlp.experts.linear_fc1.weight5 + module.decoder.layers.20.mlp.experts.linear_fc1.weight3 + module.decoder.layers.20.mlp.experts.linear_fc1.weight1 + module.decoder.layers.19.mlp.experts.linear_fc2.weight6 +Params for bucket 8 (40370176 elements, 40370176 padded size): + module.decoder.layers.19.mlp.experts.linear_fc2.weight5 + module.decoder.layers.19.mlp.experts.linear_fc2.weight1 + module.decoder.layers.19.mlp.experts.linear_fc2.weight3 + module.decoder.layers.19.mlp.experts.linear_fc1.weight6 + module.decoder.layers.19.mlp.experts.linear_fc1.weight4 + module.decoder.layers.19.mlp.experts.linear_fc2.weight4 + module.decoder.layers.19.mlp.experts.linear_fc2.weight2 + module.decoder.layers.19.mlp.experts.linear_fc2.weight0 + module.decoder.layers.19.mlp.experts.linear_fc1.weight7 + module.decoder.layers.19.mlp.experts.linear_fc1.weight5 +Params for bucket 9 (40370176 elements, 40370176 padded size): + module.decoder.layers.18.mlp.experts.linear_fc2.weight7 + module.decoder.layers.19.mlp.experts.linear_fc1.weight0 + module.decoder.layers.19.mlp.experts.linear_fc1.weight2 + module.decoder.layers.18.mlp.experts.linear_fc2.weight5 + module.decoder.layers.18.mlp.experts.linear_fc2.weight3 + module.decoder.layers.19.mlp.experts.linear_fc1.weight3 + module.decoder.layers.19.mlp.experts.linear_fc1.weight1 + module.decoder.layers.18.mlp.experts.linear_fc2.weight6 + module.decoder.layers.18.mlp.experts.linear_fc2.weight4 + module.decoder.layers.18.mlp.experts.linear_fc2.weight2 +Params for bucket 10 (40370176 elements, 40370176 padded size): + module.decoder.layers.18.mlp.experts.linear_fc2.weight1 + module.decoder.layers.18.mlp.experts.linear_fc1.weight6 + module.decoder.layers.18.mlp.experts.linear_fc1.weight4 + module.decoder.layers.18.mlp.experts.linear_fc1.weight2 + module.decoder.layers.18.mlp.experts.linear_fc1.weight7 + module.decoder.layers.18.mlp.experts.linear_fc1.weight5 + module.decoder.layers.18.mlp.experts.linear_fc2.weight0 + module.decoder.layers.18.mlp.experts.linear_fc1.weight3 +Params for bucket 11 (40370176 elements, 40370176 padded size): + module.decoder.layers.18.mlp.experts.linear_fc1.weight0 + module.decoder.layers.17.mlp.experts.linear_fc2.weight7 + module.decoder.layers.17.mlp.experts.linear_fc2.weight5 + module.decoder.layers.17.mlp.experts.linear_fc2.weight3 + module.decoder.layers.17.mlp.experts.linear_fc2.weight1 + module.decoder.layers.18.mlp.experts.linear_fc1.weight1 + module.decoder.layers.17.mlp.experts.linear_fc2.weight6 + module.decoder.layers.17.mlp.experts.linear_fc2.weight4 + module.decoder.layers.17.mlp.experts.linear_fc2.weight2 + module.decoder.layers.17.mlp.experts.linear_fc2.weight0 + module.decoder.layers.17.mlp.experts.linear_fc1.weight7 +Params for bucket 12 (40370176 elements, 40370176 padded size): + module.decoder.layers.17.mlp.experts.linear_fc1.weight6 + module.decoder.layers.17.mlp.experts.linear_fc1.weight4 + module.decoder.layers.17.mlp.experts.linear_fc1.weight2 + module.decoder.layers.17.mlp.experts.linear_fc1.weight0 + module.decoder.layers.17.mlp.experts.linear_fc1.weight5 + module.decoder.layers.17.mlp.experts.linear_fc1.weight3 + module.decoder.layers.17.mlp.experts.linear_fc1.weight1 +Params for bucket 13 (40370176 elements, 40370176 padded size): + module.decoder.layers.16.mlp.experts.linear_fc2.weight7 + module.decoder.layers.16.mlp.experts.linear_fc2.weight3 + module.decoder.layers.16.mlp.experts.linear_fc2.weight5 + module.decoder.layers.16.mlp.experts.linear_fc2.weight1 + module.decoder.layers.16.mlp.experts.linear_fc1.weight6 + module.decoder.layers.16.mlp.experts.linear_fc2.weight4 + module.decoder.layers.16.mlp.experts.linear_fc2.weight6 + module.decoder.layers.16.mlp.experts.linear_fc2.weight2 + module.decoder.layers.16.mlp.experts.linear_fc2.weight0 + module.decoder.layers.16.mlp.experts.linear_fc1.weight7 + module.decoder.layers.16.mlp.experts.linear_fc1.weight5 +Params for bucket 14 (40370176 elements, 40370176 padded size): + module.decoder.layers.16.mlp.experts.linear_fc1.weight4 + module.decoder.layers.16.mlp.experts.linear_fc1.weight0 + module.decoder.layers.16.mlp.experts.linear_fc1.weight2 + module.decoder.layers.15.mlp.experts.linear_fc2.weight7 + module.decoder.layers.15.mlp.experts.linear_fc2.weight5 + module.decoder.layers.16.mlp.experts.linear_fc1.weight3 + module.decoder.layers.16.mlp.experts.linear_fc1.weight1 + module.decoder.layers.15.mlp.experts.linear_fc2.weight6 + module.decoder.layers.15.mlp.experts.linear_fc2.weight4 +Params for bucket 15 (40370176 elements, 40370176 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_fc1.weight6 + module.decoder.layers.15.mlp.experts.linear_fc1.weight4 + module.decoder.layers.15.mlp.experts.linear_fc2.weight0 + module.decoder.layers.15.mlp.experts.linear_fc1.weight7 + module.decoder.layers.15.mlp.experts.linear_fc2.weight2 + module.decoder.layers.15.mlp.experts.linear_fc1.weight5 + module.decoder.layers.15.mlp.experts.linear_fc1.weight3 +Params for bucket 16 (40370176 elements, 40370176 padded size): + module.decoder.layers.15.mlp.experts.linear_fc1.weight2 + module.decoder.layers.15.mlp.experts.linear_fc1.weight0 + module.decoder.layers.14.mlp.experts.linear_fc2.weight7 + module.decoder.layers.14.mlp.experts.linear_fc2.weight5 + module.decoder.layers.14.mlp.experts.linear_fc2.weight3 + module.decoder.layers.14.mlp.experts.linear_fc2.weight1 + module.decoder.layers.15.mlp.experts.linear_fc1.weight1 + module.decoder.layers.14.mlp.experts.linear_fc2.weight6 + module.decoder.layers.14.mlp.experts.linear_fc2.weight4 + module.decoder.layers.14.mlp.experts.linear_fc2.weight2 + module.decoder.layers.14.mlp.experts.linear_fc2.weight0 +Params for bucket 17 (40370176 elements, 40370176 padded size): + module.decoder.layers.14.mlp.experts.linear_fc1.weight6 + module.decoder.layers.14.mlp.experts.linear_fc1.weight4 + module.decoder.layers.14.mlp.experts.linear_fc1.weight2 + module.decoder.layers.14.mlp.experts.linear_fc1.weight7 + module.decoder.layers.14.mlp.experts.linear_fc1.weight3 + module.decoder.layers.14.mlp.experts.linear_fc1.weight5 + module.decoder.layers.14.mlp.experts.linear_fc1.weight1 +Params for bucket 18 (40370176 elements, 40370176 padded size): + module.decoder.layers.14.mlp.experts.linear_fc1.weight0 + module.decoder.layers.13.mlp.experts.linear_fc2.weight5 + module.decoder.layers.13.mlp.experts.linear_fc2.weight7 + module.decoder.layers.13.mlp.experts.linear_fc2.weight3 + module.decoder.layers.13.mlp.experts.linear_fc2.weight1 + module.decoder.layers.13.mlp.experts.linear_fc1.weight6 + module.decoder.layers.13.mlp.experts.linear_fc2.weight6 + module.decoder.layers.13.mlp.experts.linear_fc2.weight4 + module.decoder.layers.13.mlp.experts.linear_fc2.weight2 + module.decoder.layers.13.mlp.experts.linear_fc2.weight0 + module.decoder.layers.13.mlp.experts.linear_fc1.weight7 +Params for bucket 19 (40370176 elements, 40370176 padded size): + module.decoder.layers.13.mlp.experts.linear_fc1.weight2 + module.decoder.layers.13.mlp.experts.linear_fc1.weight4 + module.decoder.layers.13.mlp.experts.linear_fc1.weight0 + module.decoder.layers.12.mlp.experts.linear_fc2.weight7 + module.decoder.layers.13.mlp.experts.linear_fc1.weight5 + module.decoder.layers.13.mlp.experts.linear_fc1.weight1 + module.decoder.layers.13.mlp.experts.linear_fc1.weight3 + module.decoder.layers.12.mlp.experts.linear_fc2.weight6 +Params for bucket 20 (40370176 elements, 40370176 padded size): + module.decoder.layers.12.mlp.experts.linear_fc2.weight5 + module.decoder.layers.12.mlp.experts.linear_fc2.weight3 + 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.weight5 + module.decoder.layers.12.mlp.experts.linear_fc2.weight2 + module.decoder.layers.12.mlp.experts.linear_fc2.weight4 + module.decoder.layers.12.mlp.experts.linear_fc2.weight0 + module.decoder.layers.12.mlp.experts.linear_fc1.weight6 + module.decoder.layers.12.mlp.experts.linear_fc1.weight4 +Params for bucket 21 (40370176 elements, 40370176 padded size): + module.decoder.layers.12.mlp.experts.linear_fc1.weight3 + module.decoder.layers.12.mlp.experts.linear_fc1.weight1 + module.decoder.layers.11.mlp.experts.linear_fc2.weight7 + module.decoder.layers.11.mlp.experts.linear_fc2.weight5 + module.decoder.layers.11.mlp.experts.linear_fc2.weight3 + module.decoder.layers.12.mlp.experts.linear_fc1.weight2 + module.decoder.layers.12.mlp.experts.linear_fc1.weight0 + module.decoder.layers.11.mlp.experts.linear_fc2.weight6 + module.decoder.layers.11.mlp.experts.linear_fc2.weight4 + module.decoder.layers.11.mlp.experts.linear_fc2.weight2 +Params for bucket 22 (40370176 elements, 40370176 padded size): + module.decoder.layers.11.mlp.experts.linear_fc2.weight1 + module.decoder.layers.11.mlp.experts.linear_fc1.weight6 + module.decoder.layers.11.mlp.experts.linear_fc1.weight4 + module.decoder.layers.11.mlp.experts.linear_fc1.weight2 + module.decoder.layers.11.mlp.experts.linear_fc1.weight7 + module.decoder.layers.11.mlp.experts.linear_fc1.weight5 + module.decoder.layers.11.mlp.experts.linear_fc2.weight0 + module.decoder.layers.11.mlp.experts.linear_fc1.weight3 +Params for bucket 23 (40370176 elements, 40370176 padded size): + module.decoder.layers.11.mlp.experts.linear_fc1.weight0 + module.decoder.layers.10.mlp.experts.linear_fc2.weight6 + module.decoder.layers.10.mlp.experts.linear_fc2.weight4 + module.decoder.layers.10.mlp.experts.linear_fc2.weight2 + module.decoder.layers.10.mlp.experts.linear_fc2.weight0 + module.decoder.layers.11.mlp.experts.linear_fc1.weight1 + module.decoder.layers.10.mlp.experts.linear_fc2.weight7 + module.decoder.layers.10.mlp.experts.linear_fc2.weight5 + module.decoder.layers.10.mlp.experts.linear_fc2.weight3 + module.decoder.layers.10.mlp.experts.linear_fc2.weight1 + module.decoder.layers.10.mlp.experts.linear_fc1.weight7 +Params for bucket 24 (40370176 elements, 40370176 padded size): + module.decoder.layers.10.mlp.experts.linear_fc1.weight6 + module.decoder.layers.10.mlp.experts.linear_fc1.weight4 + module.decoder.layers.10.mlp.experts.linear_fc1.weight2 + module.decoder.layers.10.mlp.experts.linear_fc1.weight0 + module.decoder.layers.10.mlp.experts.linear_fc1.weight5 + module.decoder.layers.10.mlp.experts.linear_fc1.weight3 + module.decoder.layers.10.mlp.experts.linear_fc1.weight1 +Params for bucket 25 (40370176 elements, 40370176 padded size): + module.decoder.layers.9.mlp.experts.linear_fc2.weight6 + module.decoder.layers.9.mlp.experts.linear_fc2.weight4 + module.decoder.layers.9.mlp.experts.linear_fc2.weight2 + module.decoder.layers.9.mlp.experts.linear_fc2.weight0 + module.decoder.layers.9.mlp.experts.linear_fc1.weight7 + module.decoder.layers.9.mlp.experts.linear_fc1.weight5 + module.decoder.layers.9.mlp.experts.linear_fc2.weight7 + module.decoder.layers.9.mlp.experts.linear_fc2.weight5 + module.decoder.layers.9.mlp.experts.linear_fc2.weight3 + module.decoder.layers.9.mlp.experts.linear_fc2.weight1 + module.decoder.layers.9.mlp.experts.linear_fc1.weight6 +Params for bucket 26 (40370176 elements, 40370176 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight3 + module.decoder.layers.9.mlp.experts.linear_fc1.weight1 + module.decoder.layers.8.mlp.experts.linear_fc2.weight6 + module.decoder.layers.8.mlp.experts.linear_fc2.weight4 + module.decoder.layers.9.mlp.experts.linear_fc1.weight4 + module.decoder.layers.9.mlp.experts.linear_fc1.weight0 + module.decoder.layers.9.mlp.experts.linear_fc1.weight2 + module.decoder.layers.8.mlp.experts.linear_fc2.weight7 + module.decoder.layers.8.mlp.experts.linear_fc2.weight5 +Params for bucket 27 (40370176 elements, 40370176 padded size): + module.decoder.layers.8.mlp.experts.linear_fc2.weight0 + module.decoder.layers.8.mlp.experts.linear_fc2.weight2 + module.decoder.layers.8.mlp.experts.linear_fc1.weight6 + module.decoder.layers.8.mlp.experts.linear_fc1.weight4 + module.decoder.layers.8.mlp.experts.linear_fc2.weight3 + module.decoder.layers.8.mlp.experts.linear_fc2.weight1 + module.decoder.layers.8.mlp.experts.linear_fc1.weight7 + module.decoder.layers.8.mlp.experts.linear_fc1.weight5 + module.decoder.layers.8.mlp.experts.linear_fc1.weight3 +Params for bucket 28 (40370176 elements, 40370176 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight2 + module.decoder.layers.7.mlp.experts.linear_fc2.weight7 + module.decoder.layers.8.mlp.experts.linear_fc1.weight0 + module.decoder.layers.7.mlp.experts.linear_fc2.weight5 + module.decoder.layers.7.mlp.experts.linear_fc2.weight3 + module.decoder.layers.7.mlp.experts.linear_fc2.weight1 + module.decoder.layers.7.mlp.experts.linear_fc2.weight6 + module.decoder.layers.8.mlp.experts.linear_fc1.weight1 + module.decoder.layers.7.mlp.experts.linear_fc2.weight4 + module.decoder.layers.7.mlp.experts.linear_fc2.weight2 + module.decoder.layers.7.mlp.experts.linear_fc2.weight0 +Params for bucket 29 (40370176 elements, 40370176 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight6 + module.decoder.layers.7.mlp.experts.linear_fc1.weight4 + module.decoder.layers.7.mlp.experts.linear_fc1.weight2 + module.decoder.layers.7.mlp.experts.linear_fc1.weight7 + module.decoder.layers.7.mlp.experts.linear_fc1.weight5 + module.decoder.layers.7.mlp.experts.linear_fc1.weight3 + module.decoder.layers.7.mlp.experts.linear_fc1.weight1 +Params for bucket 30 (40370176 elements, 40370176 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight0 + module.decoder.layers.6.mlp.experts.linear_fc2.weight5 + module.decoder.layers.6.mlp.experts.linear_fc2.weight7 + module.decoder.layers.6.mlp.experts.linear_fc2.weight3 + module.decoder.layers.6.mlp.experts.linear_fc2.weight1 + module.decoder.layers.6.mlp.experts.linear_fc1.weight6 + module.decoder.layers.6.mlp.experts.linear_fc2.weight6 + module.decoder.layers.6.mlp.experts.linear_fc2.weight4 + module.decoder.layers.6.mlp.experts.linear_fc2.weight2 + module.decoder.layers.6.mlp.experts.linear_fc2.weight0 + module.decoder.layers.6.mlp.experts.linear_fc1.weight7 +Params for bucket 31 (40370176 elements, 40370176 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight4 + module.decoder.layers.6.mlp.experts.linear_fc1.weight2 + module.decoder.layers.6.mlp.experts.linear_fc1.weight0 + module.decoder.layers.5.mlp.experts.linear_fc2.weight7 + module.decoder.layers.6.mlp.experts.linear_fc1.weight5 + module.decoder.layers.6.mlp.experts.linear_fc1.weight1 + module.decoder.layers.6.mlp.experts.linear_fc1.weight3 + module.decoder.layers.5.mlp.experts.linear_fc2.weight6 +Params for bucket 32 (40370176 elements, 40370176 padded size): + module.decoder.layers.5.mlp.experts.linear_fc2.weight5 + module.decoder.layers.5.mlp.experts.linear_fc2.weight1 + module.decoder.layers.5.mlp.experts.linear_fc2.weight3 + module.decoder.layers.5.mlp.experts.linear_fc1.weight6 + module.decoder.layers.5.mlp.experts.linear_fc1.weight4 + module.decoder.layers.5.mlp.experts.linear_fc2.weight2 + module.decoder.layers.5.mlp.experts.linear_fc2.weight4 + module.decoder.layers.5.mlp.experts.linear_fc2.weight0 + module.decoder.layers.5.mlp.experts.linear_fc1.weight7 + module.decoder.layers.5.mlp.experts.linear_fc1.weight5 +Params for bucket 33 (40370176 elements, 40370176 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight2 + module.decoder.layers.5.mlp.experts.linear_fc1.weight0 + module.decoder.layers.4.mlp.experts.linear_fc2.weight6 + module.decoder.layers.4.mlp.experts.linear_fc2.weight4 + module.decoder.layers.4.mlp.experts.linear_fc2.weight2 + module.decoder.layers.5.mlp.experts.linear_fc1.weight3 + module.decoder.layers.5.mlp.experts.linear_fc1.weight1 + module.decoder.layers.4.mlp.experts.linear_fc2.weight7 + module.decoder.layers.4.mlp.experts.linear_fc2.weight5 + module.decoder.layers.4.mlp.experts.linear_fc2.weight3 +Params for bucket 34 (40370176 elements, 40370176 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight0 + module.decoder.layers.4.mlp.experts.linear_fc1.weight5 + module.decoder.layers.4.mlp.experts.linear_fc1.weight7 + module.decoder.layers.4.mlp.experts.linear_fc1.weight3 + module.decoder.layers.4.mlp.experts.linear_fc2.weight1 + module.decoder.layers.4.mlp.experts.linear_fc1.weight6 + module.decoder.layers.4.mlp.experts.linear_fc1.weight4 + module.decoder.layers.4.mlp.experts.linear_fc1.weight2 +Params for bucket 35 (40370176 elements, 40370176 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight1 + 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 + module.decoder.layers.3.mlp.experts.linear_fc2.weight0 + module.decoder.layers.4.mlp.experts.linear_fc1.weight0 + module.decoder.layers.3.mlp.experts.linear_fc2.weight7 + module.decoder.layers.3.mlp.experts.linear_fc2.weight5 + module.decoder.layers.3.mlp.experts.linear_fc2.weight3 + module.decoder.layers.3.mlp.experts.linear_fc2.weight1 + module.decoder.layers.3.mlp.experts.linear_fc1.weight7 +Params for bucket 36 (40370176 elements, 40370176 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight6 + module.decoder.layers.3.mlp.experts.linear_fc1.weight2 + module.decoder.layers.3.mlp.experts.linear_fc1.weight4 + module.decoder.layers.3.mlp.experts.linear_fc1.weight0 + 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.weight1 +Params for bucket 37 (40370176 elements, 40370176 padded size): + 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 + module.decoder.layers.2.mlp.experts.linear_fc2.weight0 + module.decoder.layers.2.mlp.experts.linear_fc1.weight6 + module.decoder.layers.2.mlp.experts.linear_fc2.weight7 + module.decoder.layers.2.mlp.experts.linear_fc2.weight3 + module.decoder.layers.2.mlp.experts.linear_fc2.weight5 + module.decoder.layers.2.mlp.experts.linear_fc2.weight1 + module.decoder.layers.2.mlp.experts.linear_fc1.weight7 + module.decoder.layers.2.mlp.experts.linear_fc1.weight5 +Params for bucket 38 (40370176 elements, 40370176 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight2 + module.decoder.layers.2.mlp.experts.linear_fc1.weight0 + module.decoder.layers.1.mlp.experts.linear_fc2.weight6 + module.decoder.layers.1.mlp.experts.linear_fc2.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight1 + module.decoder.layers.2.mlp.experts.linear_fc1.weight3 + module.decoder.layers.1.mlp.experts.linear_fc2.weight7 + module.decoder.layers.1.mlp.experts.linear_fc2.weight5 +Params for bucket 39 (40370176 elements, 40370176 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight7 + module.decoder.layers.1.mlp.experts.linear_fc2.weight2 + module.decoder.layers.1.mlp.experts.linear_fc2.weight0 + module.decoder.layers.1.mlp.experts.linear_fc1.weight5 + module.decoder.layers.1.mlp.experts.linear_fc1.weight3 + module.decoder.layers.1.mlp.experts.linear_fc2.weight3 + module.decoder.layers.1.mlp.experts.linear_fc2.weight1 + module.decoder.layers.1.mlp.experts.linear_fc1.weight6 + module.decoder.layers.1.mlp.experts.linear_fc1.weight4 +Params for bucket 40 (17301504 elements, 17301504 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight0 + module.decoder.layers.1.mlp.experts.linear_fc1.weight1 +INFO:megatron.core.optimizer:Setting up optimizer with config OptimizerConfig(optimizer='adam', lr=0.0001, min_lr=1e-05, decoupled_lr=None, decoupled_min_lr=None, weight_decay=0.1, fp8_recipe='mxfp8', fp16=False, bf16=True, reuse_grad_buf_for_mxfp8_param_ag=False, params_dtype=torch.bfloat16, use_precision_aware_optimizer=False, store_param_remainders=True, main_grads_dtype=torch.float32, main_params_dtype=torch.float32, exp_avg_dtype=torch.float32, exp_avg_sq_dtype=torch.float32, loss_scale=None, initial_loss_scale=4294967296, min_loss_scale=1.0, loss_scale_window=1000, hysteresis=2, adam_beta1=0.9, adam_beta2=0.999, adam_eps=1e-08, sgd_momentum=0.9, use_distributed_optimizer=True, overlap_param_gather_with_optimizer_step=False, optimizer_cpu_offload=False, optimizer_offload_fraction=1.0, use_torch_optimizer_for_cpu_offload=False, overlap_cpu_optimizer_d2h_h2d=False, pin_cpu_grads=True, pin_cpu_params=True, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=, config_logger_dir='') +INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine +[after model, optimizer, and learning rate scheduler are built] datetime: 2026-08-26 18:36:43 +> building train, validation, and test datasets ... + > datasets target sizes (minimum size): + train: 96 + 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=(96, 0, 0) and config=GPTDatasetConfig(random_seed=1234, sequence_length=8192, blend=(['/workspace/data/wikitext/wikitext103_gpt2_text_document'], None), blend_per_split=None, split='100,0,0', split_matrix=[(0, 1.0), None, None], num_dataset_builder_threads=1, path_to_cache=None, mmap_bin_files=True, mock=False, tokenizer=, mid_level_dataset_surplus=0.005, reset_position_ids=False, reset_attention_mask=False, eod_mask_loss=False, create_attention_mask=True, drop_last_partial_validation_sequence=True, add_extra_token_to_sequence=True, object_storage_cache_path=None) +INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx +INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths +INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers +INFO:megatron.core.datasets.indexed_dataset: Extract the document indices +INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029 +INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350 +INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices +INFO:megatron.core.datasets.gpt_dataset: Load the document index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-document_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the sample index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-sample_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-shuffle_index.npy +INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 14536 +> finished creating GPT datasets ... +[after dataloaders are built] datetime: 2026-08-26 18:36:43 +done with setup ... +(min, max) time across ranks (ms): + model-and-optimizer-setup ......................: (3835.78, 3839.13) + train/valid/test-data-iterators-setup ..........: (72.32, 89.61) +training ... +Overwriting rerun_state_machine.current_iteration from -1 to 0... +[before the start of training step] datetime: 2026-08-26 18:36:44 +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py: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:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") + [2026-08-26 18:37:11] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 27838.1 | throughput per GPU (TFLOP/s/GPU): 4.6 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.092127E+01 | loss scale: 1.0 | grad norm: 30.623 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 |Number of parameters in transformer block in billions: 13.53 + +Number of parameters in embedding layers in billions: 0.21 +Total number of parameters in billions: 13.74 +Number of parameters in most loaded shard in billions: 13.7356 +Theoretical memory footprints: weight and optimizer=98244.96 MB +[Rank 0] (after 1 iterations) memory (MB) | allocated: 37181.40869140625 | max allocated: 102778.78173828125 | reserved: 105676.0 | max reserved: 105676.0 +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py: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.") + [2026-08-26 18:37:12] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 596.0 | throughput per GPU (TFLOP/s/GPU): 214.1 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.091995E+01 | loss scale: 1.0 | grad norm: 30.567 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:37:13] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 597.5 | throughput per GPU (TFLOP/s/GPU): 213.5 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 9.952354E+00 | loss scale: 1.0 | grad norm: 42.232 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | +Capture CUDA graph for training!!! +CUDA graph capture done!!! + [2026-08-26 18:37:14] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 1799.8 | throughput per GPU (TFLOP/s/GPU): 70.9 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 9.477323E+00 | loss scale: 1.0 | grad norm: 7.143 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:37:15] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 576.5 | throughput per GPU (TFLOP/s/GPU): 221.3 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.119604E+01 | loss scale: 1.0 | grad norm: 396.977 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:37:16] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 582.9 | throughput per GPU (TFLOP/s/GPU): 218.9 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 9.367741E+00 | loss scale: 1.0 | grad norm: 5.090 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:37:16] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 583.1 | throughput per GPU (TFLOP/s/GPU): 218.8 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 8.973585E+00 | loss scale: 1.0 | grad norm: 3.690 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:37:17] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 579.7 | throughput per GPU (TFLOP/s/GPU): 220.1 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.787342E+00 | loss scale: 1.0 | grad norm: 5.890 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:37:17] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 579.0 | throughput per GPU (TFLOP/s/GPU): 220.4 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.612353E+00 | loss scale: 1.0 | grad norm: 3.038 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:37:18] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 578.5 | throughput per GPU (TFLOP/s/GPU): 220.5 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.076241E+00 | loss scale: 1.0 | grad norm: 29.812 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:37:18] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 578.9 | throughput per GPU (TFLOP/s/GPU): 220.4 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.639501E+00 | loss scale: 1.0 | grad norm: 64.594 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:37:19] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 579.5 | throughput per GPU (TFLOP/s/GPU): 220.1 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.624613E+00 | loss scale: 1.0 | grad norm: 71.371 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | +[after training is done] datetime: 2026-08-26 18:37:19 +[rank0]:[W826 18:37:20.634576390 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank2]:[W826 18:37:22.048266061 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank3]:[W826 18:37:22.049850848 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank6]:[W826 18:37:22.050227136 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank5]:[W826 18:37:22.329439269 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank4]:[W826 18:37:22.343423427 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank7]:[W826 18:37:22.344655150 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank1]:[W826 18:37:22.601189451 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +W0826 18:37:24.869000 531896 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers +W0826 18:37:24.870000 531896 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 531982 closing signal SIGTERM +W0826 18:37:24.870000 531896 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 531983 closing signal SIGTERM +W0826 18:37:24.873000 531896 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 531984 closing signal SIGTERM +W0826 18:37:24.874000 531896 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 531985 closing signal SIGTERM +W0826 18:37:24.875000 531896 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 531986 closing signal SIGTERM +W0826 18:37:24.876000 531896 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 531987 closing signal SIGTERM +W0826 18:37:24.878000 531896 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 531988 closing signal SIGTERM +W0826 18:37:24.879000 531896 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 531989 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 531896 got signal: 15 diff --git a/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep3/mem_samples.csv b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep3/mem_samples.csv new file mode 100644 index 0000000000000000000000000000000000000000..e32c7d960ec35d4b2ba1352a88e6d660e9a8f750 --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep3/mem_samples.csv @@ -0,0 +1,232 @@ +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 1 +1, 1 +2, 1 +3, 1 +4, 1 +5, 1 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 1 +5, 1 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 4 +5, 4 +6, 4 +7, 4 +0, 4 +1, 4 +2, 4 +3, 629 +4, 4 +5, 4 +6, 4 +7, 4 +0, 821 +1, 821 +2, 821 +3, 821 +4, 821 +5, 821 +6, 821 +7, 821 +0, 1955 +1, 1955 +2, 1955 +3, 1955 +4, 1955 +5, 1955 +6, 1955 +7, 1955 +0, 2343 +1, 2345 +2, 2351 +3, 2347 +4, 2339 +5, 2367 +6, 2349 +7, 2347 +0, 9263 +1, 9117 +2, 9311 +3, 9077 +4, 9759 +5, 9311 +6, 9605 +7, 9371 +0, 28851 +1, 28945 +2, 28801 +3, 28795 +4, 28775 +5, 28835 +6, 28977 +7, 28907 +0, 39067 +1, 39151 +2, 39017 +3, 39015 +4, 39001 +5, 38995 +6, 39123 +7, 39167 +0, 39773 +1, 39857 +2, 39707 +3, 39705 +4, 39705 +5, 39703 +6, 39839 +7, 39857 +0, 39773 +1, 39857 +2, 39707 +3, 39705 +4, 39705 +5, 39703 +6, 39839 +7, 39857 +0, 39783 +1, 39867 +2, 39717 +3, 39717 +4, 39717 +5, 39713 +6, 39849 +7, 39867 +0, 39783 +1, 39867 +2, 39717 +3, 39717 +4, 39717 +5, 39713 +6, 39849 +7, 39867 +0, 109693 +1, 109797 +2, 110115 +3, 110121 +4, 110111 +5, 110101 +6, 109757 +7, 109787 +0, 110101 +1, 110187 +2, 110513 +3, 110511 +4, 110511 +5, 110509 +6, 110167 +7, 110187 +0, 110105 +1, 110191 +2, 110517 +3, 110515 +4, 110515 +5, 110513 +6, 110171 +7, 110191 +0, 112269 +1, 112391 +2, 110639 +3, 110677 +4, 110667 +5, 110903 +6, 112363 +7, 112481 +0, 113191 +1, 113277 +2, 111555 +3, 111553 +4, 111553 +5, 111551 +6, 113257 +7, 113277 +0, 113901 +1, 113967 +2, 112249 +3, 112257 +4, 112217 +5, 112283 +6, 113967 +7, 113919 +0, 134333 +1, 134427 +2, 134287 +3, 134271 +4, 136331 +5, 136333 +6, 134417 +7, 134425 +0, 142137 +1, 142231 +2, 142081 +3, 142079 +4, 142075 +5, 142077 +6, 142211 +7, 142231 +0, 142137 +1, 142231 +2, 142081 +3, 142079 +4, 142075 +5, 142077 +6, 142211 +7, 142231 +0, 142137 +1, 142231 +2, 142081 +3, 142079 +4, 142075 +5, 142077 +6, 142211 +7, 142231 +0, 139009 +1, 139103 +2, 138953 +3, 138689 +4, 138947 +5, 138949 +6, 139083 +7, 139103 +0, 136999 +1, 137099 +2, 137531 +3, 137267 +4, 137913 +5, 137265 +6, 137661 +7, 138069 +0, 321 +1, 4 +2, 4 +3, 4 +4, 4 +5, 4 +6, 68 +7, 137033 diff --git a/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep3/run.env b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep3/run.env new file mode 100644 index 0000000000000000000000000000000000000000..8907152eacecee56da306170fa30edf972163c8e --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep3/run.env @@ -0,0 +1,61 @@ +backend=ultraep +config=DSV2_LITE_L24 +family=tierB +repo=/workspace/Megatron-UltraEP +branch=ultraep-bracing-regions-through-kernels +commit=03d57ef7 +num_layers=24 +hidden=2048 +ffn_hidden=10944 +moe_ffn_hidden=1408 +num_experts=64 +topk=6 +experts_per_rank=8 +expert_params=8650752 +heads=16 +kv_groups=0 +kv_channels=128 +qk_layernorm=1 +shared_ffn=2816 +shared_expert_overlap=0 +mla=1 +dense_first_layers=1 +moe_layers=23 +rotary_base=10000 +rotary_scaling=40 +qkv_bias=0 +router_balancing=seq_aux_loss +aux_loss_coeff=1e-3 +router_pre_softmax=1 +router_fusion=0 +router_dtype=fp32 +device_initiated_grouped_gemm=1 +norm_epsilon=1e-6 +seq_length=8192 +micro_batch_size=1 +global_batch_size=8 +num_microbatches=1 +assignments_per_rank=49152 +mean_tokens_per_expert=6144 +slot_policy=matched +slots_per_rank=8 +capacity_mode=static +capacity=1.0755208333333333 +cuda_graph=1 +probe_mode=0 +async_grad_reduce=0 +branch_role=timing +memory_dump_mode=none +memory_snapshot_path= +cuda_graph_warmup_steps=3 +recompute=none +force_balance=0 +train_iters=12 +ep_size=8 +echo_dispatch_overlap=1 +echo_dispatch_recompute=0 +ultraep_quota_min_tokens=768 +ultraep_weight_sync_join= +profile_enable=0 +kernel_trace=0 +fp8_recipe=mxfp8 diff --git a/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep3/train.log b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep3/train.log new file mode 100644 index 0000000000000000000000000000000000000000..66d2f5d203407d108f127b2d3757691e070b1aca --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep3/train.log @@ -0,0 +1,1610 @@ +============================================================== + DSV2_LITE_L24 / ultraep (slots/rank=8 policy=matched) + E=64 K=6 H'=1408 m=8 + L=24 seq=8192 gbs=8 (1 microbatches) + cap_mode=static cap=1.0755208333333333 cg=1 recompute=none + memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep3/memory_snapshots/snapshot.pickle + repo=/workspace/Megatron-UltraEP @ 03d57ef7 +============================================================== +W0826 18:37:30.374000 535038 torch/distributed/run.py:766] +W0826 18:37:30.374000 535038 torch/distributed/run.py:766] ***************************************** +W0826 18:37:30.374000 535038 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0826 18:37:30.374000 535038 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") +using world size: 8, data-parallel size: 8, context-parallel size: 1, hierarchical context-parallel sizes: None, tensor-model-parallel size: 1, pipeline-model-parallel size: 1 +WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer +Number of virtual stages per pipeline stage: None +accumulate and all-reduce gradients in fp32 for bfloat16 data type. +using torch.bfloat16 for parameters ... +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +------------------------ arguments ------------------------ + account_for_embedding_in_pipeline_split ......... False + account_for_loss_in_pipeline_split .............. False + accumulate_allreduce_grads_in_fp32 .............. True + adam_beta1 ...................................... 0.9 + adam_beta2 ...................................... 0.999 + adam_eps ........................................ 1e-08 + add_bias_linear ................................. False + add_position_embedding .......................... False + add_qkv_bias .................................... False + adlr_autoresume ................................. False + adlr_autoresume_interval ........................ 1000 + align_grad_reduce ............................... True + align_param_gather .............................. False + app_tag_run_name ................................ None + app_tag_run_version ............................. 0.0.0 + apply_layernorm_1p .............................. False + apply_query_key_layer_scaling ................... False + apply_residual_connection_post_layernorm ........ False + apply_rope_fusion ............................... False + async_save ...................................... None + async_tensor_model_parallel_allreduce ........... True + attention_backend ............................... AttnBackend.auto + attention_dropout ............................... 0.0 + attention_softmax_in_fp32 ....................... False + auto_detect_ckpt_format ......................... False + barrier_with_L1_time ............................ True + bert_binary_head ................................ True + bert_embedder_type .............................. megatron + bert_load ....................................... None + bf16 ............................................ True + bias_dropout_fusion ............................. True + bias_gelu_fusion ................................ False + bias_swiglu_fusion .............................. True + biencoder_projection_dim ........................ 0 + biencoder_shared_query_context_model ............ False + block_data_path ................................. None + calc_ft_timeouts ................................ False + calculate_per_token_loss ........................ False + check_for_large_grads ........................... False + check_for_nan_in_loss_and_grad .................. False + check_for_spiky_loss ............................ False + check_weight_hash_across_dp_replicas_interval ... None + ckpt_assume_constant_structure .................. False + ckpt_convert_format ............................. None + ckpt_convert_save ............................... None + ckpt_convert_update_legacy_dist_opt_format ...... False + ckpt_format ..................................... torch_dist + ckpt_fully_parallel_load ........................ False + ckpt_fully_parallel_save ........................ True + ckpt_fully_parallel_save_deprecated ............. False + ckpt_step ....................................... None + classes_fraction ................................ 1.0 + clip_grad ....................................... 1.0 + clone_scatter_output_in_embedding ............... True + config_logger_dir ............................... + consumed_train_samples .......................... 0 + consumed_valid_samples .......................... 0 + context_parallel_size ........................... 1 + cp_comm_type .................................... ['p2p'] + create_attention_mask_in_dataloader ............. True + cross_entropy_fusion_impl ....................... native + cross_entropy_loss_fusion ....................... False + cuda_graph_scope ................................ full_iteration + cuda_graph_warmup_steps ......................... 3 + data_args_path .................................. None + data_cache_path ................................. None + data_parallel_random_init ....................... False + data_parallel_sharding_strategy ................. no_shard + data_parallel_size .............................. 8 + data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document'] + data_per_class_fraction ......................... 1.0 + data_sharding ................................... True + dataloader_type ................................. single + ddp_average_in_collective ....................... False + ddp_bucket_size ................................. None + ddp_num_buckets ................................. None + ddp_pad_buckets_for_high_nccl_busbw ............. False + decoder_first_pipeline_num_layers ............... None + decoder_last_pipeline_num_layers ................ None + decoder_num_layers .............................. None + decoder_seq_length .............................. None + decoupled_lr .................................... None + decoupled_min_lr ................................ None + decrease_batch_size_if_needed ................... False + defer_embedding_wgrad_compute ................... False + delay_wgrad_compute ............................. False + deprecated_use_mcore_models ..................... True + deterministic_mode .............................. False + dino_bottleneck_size ............................ 256 + dino_freeze_last_layer .......................... 1 + dino_head_hidden_size ........................... 2048 + dino_local_crops_number ......................... 10 + dino_local_img_size ............................. 96 + dino_norm_last_layer ............................ False + dino_teacher_temp ............................... 0.07 + dino_warmup_teacher_temp ........................ 0.04 + dino_warmup_teacher_temp_epochs ................. 30 + disable_bf16_reduced_precision_matmul ........... False + disable_mamba_mem_eff_path ...................... False + disable_straggler_on_startup .................... False + dist_ckpt_format_deprecated ..................... None + dist_ckpt_strictness ............................ assume_ok_unexpected + distribute_saved_activations .................... False + distributed_backend ............................. nccl + distributed_timeout_minutes ..................... 10 + embedding_init_method_std ....................... None + embedding_path .................................. None + empty_unused_memory_level ....................... 0 + enable_cuda_graph ............................... True + enable_experimental ............................. False + enable_ft_package ............................... False + enable_gloo_process_groups ...................... True + enable_msc ...................................... True + enable_one_logger ............................... True + encoder_num_layers .............................. 24 + encoder_seq_length .............................. 8192 + end_weight_decay ................................ 0.1 + eod_mask_loss ................................... False + error_injection_rate ............................ 0 + error_injection_type ............................ transient_error + eval_interval ................................... 1000000 + eval_iters ...................................... 0 + evidence_data_path .............................. None + exit_duration_in_mins ........................... None + exit_interval ................................... None + exit_on_missing_checkpoint ...................... False + exit_signal_handler ............................. False + exp_avg_dtype ................................... torch.float32 + exp_avg_sq_dtype ................................ torch.float32 + expert_model_parallel_size ...................... 8 + expert_tensor_parallel_size ..................... 1 + export_force_local_attention .................... False + export_kd_cfg ................................... None + export_kd_teacher_ckpt_format ................... None + export_kd_teacher_load .......................... None + export_kv_cache_quant ........................... False + export_legacy_megatron .......................... False + export_model_type ............................... GPTModel + export_moe_apply_probs_on_input ................. False + export_qk_l2_norm ............................... False + export_quant_cfg ................................ None + export_real_quant_cfg ........................... None + export_te_mcore_model ........................... False + external_cuda_graph ............................. False + ffn_hidden_size ................................. 10944 + finetune ........................................ False + finetune_data_split ............................. train + finetune_hf_dataset ............................. None + first_last_layers_bf16 .......................... False + flash_decode .................................... False + fp16 ............................................ False + fp16_lm_cross_entropy ........................... False + fp32_residual_connection ........................ False + fp8 ............................................. e4m3 + fp8_amax_compute_algo ........................... most_recent + fp8_amax_history_len ............................ 1 + fp8_interval .................................... 1 + fp8_margin ...................................... 0 + fp8_param_gather ................................ False + fp8_recipe ...................................... mxfp8 + fp8_wgrad ....................................... True + fsdp_double_buffer .............................. False + global_batch_size ............................... 8 + grad_reduce_in_bf16 ............................. False + gradient_accumulation_fusion .................... True + gradient_reduce_div_fusion ...................... True + group_query_attention ........................... False + head_lr_mult .................................... 1.0 + heterogeneous_layers_config_encoded_json ........ None + heterogeneous_layers_config_path ................ None + hidden_dropout .................................. 0.0 + hidden_size ..................................... 2048 + hierarchical_context_parallel_sizes ............. None + high_priority_stream_groups ..................... [] + hybrid_attention_ratio .......................... 0.0 + hybrid_mlp_ratio ................................ 0.0 + hybrid_override_pattern ......................... None + hysteresis ...................................... 2 + ict_head_size ................................... None + ict_load ........................................ None + img_h ........................................... 224 + img_w ........................................... 224 + indexer_batch_size .............................. 128 + indexer_log_interval ............................ 1000 + inference_batch_times_seqlen_threshold .......... -1 + inference_dynamic_batching ...................... False + inference_dynamic_batching_buffer_guaranteed_fraction 0.2 + inference_dynamic_batching_buffer_overflow_factor None + inference_dynamic_batching_buffer_size_gb ....... 40.0 + inference_dynamic_batching_chunk_size ........... 256 + inference_dynamic_batching_max_requests_override None + inference_dynamic_batching_max_tokens_override .. None + inference_dynamic_batching_num_cuda_graphs ...... 16 + inference_max_batch_size ........................ 8 + inference_max_seq_length ........................ 2560 + inference_rng_tracker ........................... False + init_method_std ................................. 0.01 + init_method_xavier_uniform ...................... False + init_model_with_meta_device ..................... False + initial_loss_scale .............................. 4294967296 + inprocess_active_world_size ..................... 8 + inprocess_barrier_timeout ....................... 120 + inprocess_completion_timeout .................... 120 + inprocess_empty_cuda_cache ...................... False + inprocess_granularity ........................... node + inprocess_hard_timeout .......................... 90 + inprocess_heartbeat_interval .................... 30 + inprocess_heartbeat_timeout ..................... 60 + inprocess_last_call_wait ........................ 1 + inprocess_max_iterations ........................ None + inprocess_monitor_process_interval .............. 1.0 + inprocess_monitor_thread_interval ............... 1.0 + inprocess_progress_watchdog_interval ............ 1.0 + inprocess_restart ............................... False + inprocess_soft_timeout .......................... 60 + inprocess_termination_grace_time ................ 1 + is_hybrid_model ................................. False + iter_per_epoch .................................. 1250 + iterations_to_skip .............................. [] + keep_fp8_transpose_cache_when_using_custom_fsdp . False + kitchen_config_file ............................. None + kitchen_recipe_number ........................... None + kv_channels ..................................... 128 + kv_lora_rank .................................... 512 + lazy_mpu_init ................................... None + load ............................................ None + load_main_params_from_ckpt ...................... None + load_model_opt_format ........................... False + local_rank ...................................... 0 + log_energy ...................................... False + log_interval .................................... 1 + log_loss_scale_to_tensorboard ................... True + log_memory_to_tensorboard ....................... False + log_num_zeros_in_grad ........................... False + log_params_norm ................................. False + log_progress .................................... False + log_straggler ................................... False + log_throughput .................................. True + log_timers_to_tensorboard ....................... False + log_validation_ppl_to_tensorboard ............... False + log_world_size_to_tensorboard ................... False + logging_level ................................... None + loss_scale ...................................... None + loss_scale_window ............................... 1000 + lr .............................................. 0.0001 + lr_decay_iters .................................. 8 + lr_decay_samples ................................ None + lr_decay_style .................................. cosine + lr_warmup_fraction .............................. None + lr_warmup_init .................................. 0.0 + lr_warmup_iters ................................. 5 + lr_warmup_samples ............................... 0 + lr_wsd_decay_iters .............................. None + lr_wsd_decay_samples ............................ None + lr_wsd_decay_style .............................. exponential + main_grads_dtype ................................ torch.float32 + main_params_dtype ............................... torch.float32 + make_vocab_size_divisible_by .................... 128 + mamba_head_dim .................................. 64 + mamba_num_groups ................................ 8 + mamba_num_heads ................................. None + mamba_state_dim ................................. 128 + manual_gc ....................................... False + manual_gc_eval .................................. True + manual_gc_interval .............................. 0 + mask_factor ..................................... 1.0 + mask_prob ....................................... 0.15 + mask_type ....................................... random + masked_softmax_fusion ........................... False + max_position_embeddings ......................... 32768 + max_tokens_to_oom ............................... 12000 + memory_snapshot_path ............................ snapshot.pickle + merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt + micro_batch_size ................................ 1 + microbatch_group_size_per_vp_stage .............. None + mid_level_dataset_surplus ....................... 0.005 + min_loss_scale .................................. 1.0 + min_lr .......................................... 1e-05 + mlp_chunks_for_prefill .......................... 1 + mmap_bin_files .................................. True + mock_data ....................................... False + moe_apply_probs_on_input ........................ False + moe_aux_loss_coeff .............................. 0.001 + moe_capacity_monitor ............................ True + moe_deepep_num_sms .............................. 20 + moe_echo_enable_random_offloading ............... False + moe_echo_expert_dispatch_overlap ................ False + moe_echo_expert_dispatcher_type ................. hybridep + moe_echo_recompute_expert_dispatch .............. False + moe_enable_deepep ............................... False + moe_enable_echo ................................. False + moe_enable_hybridep ............................. True + moe_enable_ultraep .............................. True + moe_expert_capacity_factor ...................... None + moe_extended_tp ................................. False + moe_ffn_hidden_size ............................. 1408 + moe_grouped_gemm ................................ True + moe_hybridep_num_sms ............................ 16 + moe_input_jitter_eps ............................ None + moe_layer_freq .................................. [0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] + moe_layer_recompute ............................. False + moe_num_echo_experts ............................ None + moe_num_redundant_experts_per_rank .............. 8 + moe_pad_expert_input_to_capacity ................ False + moe_per_layer_logging ........................... False + moe_permute_fusion .............................. False + moe_received_token_capacity ..................... 1.0755208333333333 + moe_router_bias_update_rate ..................... 0.001 + moe_router_dtype ................................ fp32 + moe_router_enable_expert_bias ................... False + moe_router_force_load_balancing ................. False + moe_router_group_topk ........................... None + moe_router_load_balancing_type .................. seq_aux_loss + moe_router_num_groups ........................... None + moe_router_padding_for_fp8 ...................... False + moe_router_pre_softmax .......................... True + moe_router_score_function ....................... softmax + moe_router_topk ................................. 6 + moe_router_topk_scaling_factor .................. None + moe_shared_expert_intermediate_size ............. 2816 + moe_shared_expert_overlap ....................... False + moe_token_dispatcher_type ....................... flex + moe_token_drop_policy ........................... probs + moe_ultraep_max_microbatches .................... None + moe_ultraep_overlap_weight_sync ................. True + moe_ultraep_weight_sync_join .................... before_dispatch + moe_upcycling_granularity ....................... 1 + moe_use_device_initiated_grouped_gemm ........... True + moe_use_legacy_grouped_gemm ..................... False + moe_use_upcycling ............................... False + moe_z_loss_coeff ................................ None + mrope_section ................................... None + mscale .......................................... 0.707 + mscale_all_dim .................................. 0.707 + mtp_loss_scaling_factor ......................... 0.1 + mtp_num_layers .................................. None + multi_latent_attention .......................... True + nccl_all_reduce_for_prefill ..................... False + nccl_communicator_config_path ................... None + nccl_ub ......................................... False + no_load_optim ................................... None + no_load_rng ..................................... None + no_persist_layer_norm ........................... False + no_rope_freq .................................... None + no_save_optim ................................... None + no_save_rng ..................................... None + non_persistent_ckpt_type ........................ None + non_persistent_global_ckpt_dir .................. None + non_persistent_local_ckpt_algo .................. fully_parallel + non_persistent_local_ckpt_dir ................... None + non_persistent_save_interval .................... None + norm_epsilon .................................... 1e-06 + normalization ................................... RMSNorm + num_attention_heads ............................. 16 + num_channels .................................... 3 + num_classes ..................................... 1000 + num_dataset_builder_threads ..................... 1 + num_distributed_optimizer_instances ............. 1 + num_experts ..................................... 64 + num_layers ...................................... 24 + num_layers_at_end_in_bf16 ....................... 1 + num_layers_at_start_in_bf16 ..................... 1 + num_layers_per_virtual_pipeline_stage ........... None + num_query_groups ................................ 1 + num_virtual_stages_per_pipeline_rank ............ None + num_workers ..................................... 2 + object_storage_cache_path ....................... None + one_logger_async ................................ False + one_logger_project .............................. megatron-lm + one_logger_run_name ............................. None + onnx_safe ....................................... None + openai_gelu ..................................... False + optimizer ....................................... adam + optimizer_cpu_offload ........................... False + optimizer_offload_fraction ...................... 1.0 + output_bert_embeddings .......................... False + overlap_cpu_optimizer_d2h_h2d ................... False + overlap_grad_reduce ............................. True + overlap_moe_expert_parallel_comm ................ False + overlap_p2p_comm ................................ False + overlap_p2p_comm_warmup_flush ................... False + overlap_param_gather ............................ True + overlap_param_gather_with_optimizer_step ........ False + override_opt_param_scheduler .................... False + params_dtype .................................... torch.bfloat16 + patch_dim ....................................... 16 + per_split_data_args_path ........................ None + perform_initialization .......................... True + pin_cpu_grads ................................... True + pin_cpu_params .................................. True + pipeline_model_parallel_comm_backend ............ None + pipeline_model_parallel_layout .................. None + pipeline_model_parallel_size .................... 1 + position_embedding_type ......................... rope + pretrained_checkpoint ........................... None + profile ......................................... False + profile_ranks ................................... [0] + profile_step_end ................................ 12 + profile_step_start .............................. 10 + q_lora_rank ..................................... None + qk_head_dim ..................................... 128 + qk_l2_norm ...................................... False + qk_layernorm .................................... True + qk_pos_emb_head_dim ............................. 64 + query_in_block_prob ............................. 0.1 + rampup_batch_size ............................... None + rank ............................................ 0 + recompute_granularity ........................... None + recompute_method ................................ None + recompute_modules ............................... None + recompute_num_layers ............................ None + record_memory_history ........................... False + relative_attention_max_distance ................. 128 + relative_attention_num_buckets .................. 32 + replication ..................................... False + replication_factor .............................. 2 + replication_jump ................................ None + rerun_mode ...................................... validate_results + reset_attention_mask ............................ False + reset_position_ids .............................. False + result_rejected_tracker_filename ................ None + retriever_report_topk_accuracies ................ [] + retriever_score_scaling ......................... False + retriever_seq_length ............................ 256 + retro_add_retriever ............................. False + retro_attention_gate ............................ 1 + retro_cyclic_train_iters ........................ None + retro_encoder_attention_dropout ................. 0.1 + retro_encoder_hidden_dropout .................... 0.1 + retro_encoder_layers ............................ 2 + retro_num_neighbors ............................. 2 + retro_num_retrieved_chunks ...................... 2 + retro_project_dir ............................... None + retro_verify_neighbor_count ..................... True + reuse_grad_buf_for_mxfp8_param_ag ............... False + rope_scaling_factor ............................. 8.0 + rope_type ....................................... None + rotary_base ..................................... 10000 + rotary_interleaved .............................. False + rotary_percent .................................. 1.0 + rotary_scaling_factor ........................... 40.0 + rotary_seq_len_interpolation_factor ............. None + run_workload_inspector_server ................... False + sample_rate ..................................... 1.0 + save ............................................ None + save_interval ................................... None + save_retain_interval ............................ None + scatter_gather_tensors_in_pipeline .............. True + seed ............................................ 1234 + seq_length ...................................... 8192 + sequence_parallel ............................... False + sft ............................................. False + sft_tokenizer_prompt_format ..................... nemotron-h-aligned + sgd_momentum .................................... 0.9 + short_seq_prob .................................. 0.1 + skip_train ...................................... False + skipped_train_samples ........................... 0 + spec ............................................ None + split ........................................... 100,0,0 + squared_relu .................................... False + start_weight_decay .............................. 0.1 + straggler_ctrlr_port ............................ 65535 + straggler_minmax_count .......................... 1 + suggested_communication_unit_size ............... None + swiglu .......................................... True + swin_backbone_type .............................. tiny + symmetric_ar_type ............................... None + te_rng_tracker .................................. True + tensor_model_parallel_size ...................... 1 + tensorboard_dir ................................. None + tensorboard_log_interval ........................ 1 + tensorboard_queue_size .......................... 1000 + test_data_path .................................. None + test_mode ....................................... False + tiktoken_num_special_tokens ..................... 1000 + tiktoken_pattern ................................ None + tiktoken_special_tokens ......................... None + timing_log_level ................................ 0 + timing_log_option ............................... minmax + titles_data_path ................................ None + tokenizer_model ................................. None + tokenizer_type .................................. GPT2BPETokenizer + torch_fsdp2_reshard_after_forward ............... True + tp_comm_bootstrap_backend ....................... nccl + tp_comm_bulk_dgrad .............................. True + tp_comm_bulk_wgrad .............................. True + tp_comm_overlap ................................. False + tp_comm_overlap_ag .............................. True + tp_comm_overlap_cfg ............................. None + tp_comm_overlap_rs .............................. True + tp_comm_overlap_rs_dgrad ........................ False + tp_comm_split_ag ................................ True + tp_comm_split_rs ................................ True + train_data_path ................................. None + train_iters ..................................... 12 + train_samples ................................... None + train_sync_interval ............................. None + transformer_impl ................................ transformer_engine + transformer_pipeline_model_parallel_size ........ 1 + untie_embeddings_and_output_weights ............. True + use_checkpoint_args ............................. False + use_checkpoint_opt_param_scheduler .............. False + use_cpu_initialization .......................... None + use_custom_fsdp ................................. False + use_dist_ckpt ................................... True + use_dist_ckpt_deprecated ........................ False + use_distributed_optimizer ....................... True + use_flash_attn .................................. False + use_legacy_models ............................... False + use_mp_args_from_checkpoint_args ................ False + use_one_sent_docs ............................... False + use_persistent_ckpt_worker ...................... False + use_precision_aware_optimizer ................... False + use_pytorch_profiler ............................ False + use_ring_exchange_p2p ........................... False + use_rope_scaling ................................ False + use_rotary_position_embeddings .................. False + use_sharp ....................................... False + use_tokenizer_model_from_checkpoint_args ........ True + use_torch_fsdp2 ................................. False + use_torch_optimizer_for_cpu_offload ............. False + use_tp_pp_dp_mapping ............................ False + v_head_dim ...................................... 128 + valid_data_path ................................. None + variable_seq_lengths ............................ False + virtual_pipeline_model_parallel_size ............ None + vision_backbone_type ............................ vit + vision_pretraining .............................. False + vision_pretraining_type ......................... classify + vocab_extra_ids ................................. 0 + vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json + vocab_size ...................................... None + wandb_exp_name .................................. + wandb_project ................................... + wandb_save_dir .................................. + weight_decay .................................... 0.1 + weight_decay_incr_style ......................... constant + wgrad_deferral_limit ............................ 0 + world_size ...................................... 8 + yaml_cfg ........................................ None +-------------------- end of arguments --------------------- +INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1 +> building GPT2BPETokenizer tokenizer ... +WARNING: one_logger package is required to enable e2e metrics tracking. please go to https://confluence.nvidia.com/display/MLWFO/Package+Repositories for details to install it + > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304) +WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS +> initializing torch distributed ... +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +[Gloo] Rank 1 is connected to [Gloo] Rank 7 peer ranks. 0Expected number of connected peer ranks is : is connected to 77 peer ranks. +Expected number of connected peer ranks is : 7 +[Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 3 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 6 is connected to 77 peer ranks. is connected to Expected number of connected peer ranks is : 77 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 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank [Gloo] Rank 6 is connected to 77 is connected to peer ranks. 7Expected number of connected peer ranks is : peer ranks. 7Expected number of connected peer ranks is : 7 + +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +> initialized tensor model parallel with size 1 +> initialized pipeline model parallel with size 1 +> setting random seeds to 1234 ... +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank [Gloo] Rank 00 is connected to is connected to 00 peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : 00 + +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[rank3]:[W826 18:37:48.700966722 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank4]:[W826 18:37:48.701090734 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 4] using GPU 4 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank6]:[W826 18:37:48.701093654 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]:[W826 18:37:48.701096323 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. +[rank1]:[W826 18:37:48.701121516 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. +[rank7]:[W826 18:37:48.701125902 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank5]:[W826 18:37:48.702620111 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. +> 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: 0.036 seconds +WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations. +> compiling and loading fused kernels ... +[rank0]:[W826 18:37:48.880746081 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: 3.793 seconds +time to initialize megatron (seconds): 7.270 +[after megatron is initialized] datetime: 2026-08-26 18:37:54 +building GPT model ... +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/workspace/Megatron-UltraEP/megatron/core/transformer/transformer_config.py:1526: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +[INFO] Use MNNVL fabric: False +[WARN] MAX_NUM_NVL_PEERS is not set. Using detected value 8. +NVSHMEM v3.7.2 +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +UltraEP: NVL domain size <= 8; forcing weight_sync plan mode to direct (was adaptive). + > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 4188125184 +INFO:megatron.core.distributed.distributed_data_parallel:Setting up DistributedDataParallel with config DistributedDataParallelConfig(grad_reduce_in_fp32=True, overlap_grad_reduce=True, overlap_param_gather=True, align_param_gather=False, use_distributed_optimizer=True, num_distributed_optimizer_instances=1, check_for_nan_in_grad=False, check_for_large_grads=False, bucket_size=40000000, pad_buckets_for_high_nccl_busbw=False, average_in_collective=False, fp8_param_gather=False, reuse_grad_buf_for_mxfp8_param_ag=False, use_custom_fsdp=False, data_parallel_sharding_strategy='no_shard', gradient_reduce_div_fusion=True, suggested_communication_unit_size=None, preserve_fp32_weights=True, keep_fp8_transpose_cache_when_using_custom_fsdp=False, nccl_ub=False, fsdp_double_buffer=False) +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 20 +Params for bucket 1 (103022592 elements, 103022592 padded size): + module.output_layer.weight +Params for bucket 2 (48503296 elements, 48503296 padded size): + module.decoder.final_layernorm.weight + module.decoder.layers.23.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.23.self_attention.linear_kv_up_proj.weight + module.decoder.layers.22.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.23.pre_mlp_layernorm.weight + module.decoder.layers.23.mlp.router.weight + module.decoder.layers.23.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.23.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.23.self_attention.linear_kv_down_proj.weight + module.decoder.layers.23.self_attention.linear_q_proj.weight + module.decoder.layers.23.self_attention.linear_proj.weight + module.decoder.layers.23.input_layernorm.weight + module.decoder.layers.22.mlp.shared_experts.linear_fc1.weight +Params for bucket 3 (40901632 elements, 40901632 padded size): + module.decoder.layers.22.mlp.router.weight + module.decoder.layers.22.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.22.pre_mlp_layernorm.weight + module.decoder.layers.22.self_attention.linear_kv_down_proj.weight + module.decoder.layers.22.self_attention.linear_q_proj.weight + module.decoder.layers.22.self_attention.linear_proj.weight + module.decoder.layers.22.input_layernorm.weight + module.decoder.layers.21.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.21.pre_mlp_layernorm.weight + module.decoder.layers.21.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.21.self_attention.linear_kv_down_proj.weight + module.decoder.layers.22.self_attention.linear_kv_up_proj.weight + module.decoder.layers.21.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.21.mlp.router.weight + module.decoder.layers.21.self_attention.linear_kv_up_proj.weight + module.decoder.layers.21.self_attention.linear_q_proj.weight +Params for bucket 4 (41163264 elements, 41163264 padded size): + module.decoder.layers.21.self_attention.linear_proj.weight + module.decoder.layers.20.mlp.router.weight + module.decoder.layers.20.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.20.self_attention.linear_kv_up_proj.weight + module.decoder.layers.19.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.20.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.21.input_layernorm.weight + module.decoder.layers.20.pre_mlp_layernorm.weight + module.decoder.layers.20.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.20.self_attention.linear_kv_down_proj.weight + module.decoder.layers.20.self_attention.linear_q_proj.weight + module.decoder.layers.20.self_attention.linear_proj.weight + module.decoder.layers.20.input_layernorm.weight +Params for bucket 5 (42734080 elements, 42734080 padded size): + module.decoder.layers.19.self_attention.linear_kv_up_proj.weight + module.decoder.layers.19.mlp.router.weight + module.decoder.layers.18.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.19.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.19.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.19.pre_mlp_layernorm.weight + module.decoder.layers.19.self_attention.linear_kv_down_proj.weight + module.decoder.layers.19.self_attention.linear_q_proj.weight + module.decoder.layers.19.self_attention.linear_proj.weight + module.decoder.layers.19.input_layernorm.weight + module.decoder.layers.18.mlp.shared_experts.linear_fc1.weight +Params for bucket 6 (40901632 elements, 40901632 padded size): + module.decoder.layers.18.mlp.router.weight + module.decoder.layers.18.self_attention.linear_kv_up_proj.weight + module.decoder.layers.17.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.17.mlp.router.weight + module.decoder.layers.17.self_attention.linear_kv_up_proj.weight + module.decoder.layers.17.self_attention.linear_q_proj.weight + module.decoder.layers.18.self_attention.linear_kv_down_proj.weight + module.decoder.layers.18.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.18.pre_mlp_layernorm.weight + module.decoder.layers.18.self_attention.linear_q_proj.weight + module.decoder.layers.18.self_attention.linear_proj.weight + module.decoder.layers.18.input_layernorm.weight + module.decoder.layers.17.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.17.pre_mlp_layernorm.weight + module.decoder.layers.17.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.17.self_attention.linear_kv_down_proj.weight +Params for bucket 7 (41163264 elements, 41163264 padded size): + module.decoder.layers.16.mlp.router.weight + module.decoder.layers.16.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.16.self_attention.linear_kv_up_proj.weight + module.decoder.layers.15.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.17.self_attention.linear_proj.weight + module.decoder.layers.17.input_layernorm.weight + module.decoder.layers.16.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.16.pre_mlp_layernorm.weight + module.decoder.layers.16.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.16.self_attention.linear_kv_down_proj.weight + module.decoder.layers.16.self_attention.linear_q_proj.weight + module.decoder.layers.16.self_attention.linear_proj.weight + module.decoder.layers.16.input_layernorm.weight +Params for bucket 8 (42734080 elements, 42734080 padded size): + module.decoder.layers.15.pre_mlp_layernorm.weight + module.decoder.layers.15.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.15.self_attention.linear_kv_down_proj.weight + module.decoder.layers.15.self_attention.linear_q_proj.weight + module.decoder.layers.15.self_attention.linear_proj.weight + module.decoder.layers.15.input_layernorm.weight + module.decoder.layers.14.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.15.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.15.self_attention.linear_kv_up_proj.weight + module.decoder.layers.15.mlp.router.weight + module.decoder.layers.14.mlp.shared_experts.linear_fc2.weight +Params for bucket 9 (40901632 elements, 40901632 padded size): + module.decoder.layers.14.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.14.pre_mlp_layernorm.weight + module.decoder.layers.14.self_attention.linear_kv_down_proj.weight + module.decoder.layers.14.self_attention.linear_q_proj.weight + module.decoder.layers.14.self_attention.linear_proj.weight + module.decoder.layers.14.input_layernorm.weight + module.decoder.layers.13.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.13.mlp.router.weight + module.decoder.layers.13.self_attention.linear_kv_up_proj.weight + module.decoder.layers.14.mlp.router.weight + module.decoder.layers.14.self_attention.linear_kv_up_proj.weight + module.decoder.layers.13.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.13.pre_mlp_layernorm.weight + module.decoder.layers.13.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.13.self_attention.linear_kv_down_proj.weight + module.decoder.layers.13.self_attention.linear_q_proj.weight +Params for bucket 10 (41163264 elements, 41163264 padded size): + module.decoder.layers.12.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.12.mlp.router.weight + module.decoder.layers.12.self_attention.linear_kv_up_proj.weight + module.decoder.layers.12.self_attention.linear_q_proj.weight + module.decoder.layers.12.self_attention.linear_proj.weight + module.decoder.layers.12.input_layernorm.weight + module.decoder.layers.13.self_attention.linear_proj.weight + module.decoder.layers.13.input_layernorm.weight + module.decoder.layers.12.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.12.pre_mlp_layernorm.weight + module.decoder.layers.12.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.12.self_attention.linear_kv_down_proj.weight + module.decoder.layers.11.mlp.shared_experts.linear_fc2.weight +Params for bucket 11 (42734080 elements, 42734080 padded size): + module.decoder.layers.11.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.11.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.11.pre_mlp_layernorm.weight + module.decoder.layers.11.self_attention.linear_kv_down_proj.weight + module.decoder.layers.11.self_attention.linear_q_proj.weight + module.decoder.layers.11.self_attention.linear_proj.weight + module.decoder.layers.11.input_layernorm.weight + module.decoder.layers.10.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.11.self_attention.linear_kv_up_proj.weight + module.decoder.layers.11.mlp.router.weight + module.decoder.layers.10.mlp.shared_experts.linear_fc2.weight +Params for bucket 12 (40901632 elements, 40901632 padded size): + module.decoder.layers.10.mlp.router.weight + module.decoder.layers.10.self_attention.linear_kv_up_proj.weight + module.decoder.layers.9.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.9.mlp.router.weight + module.decoder.layers.9.self_attention.linear_kv_up_proj.weight + module.decoder.layers.10.pre_mlp_layernorm.weight + module.decoder.layers.10.self_attention.linear_kv_down_proj.weight + module.decoder.layers.10.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.10.self_attention.linear_q_proj.weight + module.decoder.layers.10.self_attention.linear_proj.weight + module.decoder.layers.10.input_layernorm.weight + module.decoder.layers.9.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.9.pre_mlp_layernorm.weight + module.decoder.layers.9.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.9.self_attention.linear_kv_down_proj.weight + module.decoder.layers.9.self_attention.linear_q_proj.weight +Params for bucket 13 (41163264 elements, 41163264 padded size): + module.decoder.layers.8.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.8.mlp.router.weight + module.decoder.layers.8.self_attention.linear_kv_up_proj.weight + module.decoder.layers.7.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.9.self_attention.linear_proj.weight + module.decoder.layers.8.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.9.input_layernorm.weight + module.decoder.layers.8.pre_mlp_layernorm.weight + module.decoder.layers.8.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.8.self_attention.linear_kv_down_proj.weight + module.decoder.layers.8.self_attention.linear_q_proj.weight + module.decoder.layers.8.self_attention.linear_proj.weight + module.decoder.layers.8.input_layernorm.weight +Params for bucket 14 (42734080 elements, 42734080 padded size): + module.decoder.layers.7.self_attention.linear_kv_up_proj.weight + module.decoder.layers.7.mlp.router.weight + module.decoder.layers.7.self_attention.linear_q_proj.weight + module.decoder.layers.7.self_attention.linear_proj.weight + module.decoder.layers.7.input_layernorm.weight + module.decoder.layers.6.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.7.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.7.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.7.pre_mlp_layernorm.weight + module.decoder.layers.7.self_attention.linear_kv_down_proj.weight + module.decoder.layers.6.mlp.shared_experts.linear_fc2.weight +Params for bucket 15 (40901632 elements, 40901632 padded size): + module.decoder.layers.6.mlp.router.weight + module.decoder.layers.6.self_attention.linear_kv_up_proj.weight + module.decoder.layers.5.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.5.pre_mlp_layernorm.weight + module.decoder.layers.5.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.5.self_attention.linear_kv_down_proj.weight + module.decoder.layers.5.self_attention.linear_q_proj.weight + module.decoder.layers.6.pre_mlp_layernorm.weight + module.decoder.layers.6.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.6.self_attention.linear_kv_down_proj.weight + module.decoder.layers.6.self_attention.linear_q_proj.weight + module.decoder.layers.6.self_attention.linear_proj.weight + module.decoder.layers.6.input_layernorm.weight + module.decoder.layers.5.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.5.mlp.router.weight + module.decoder.layers.5.self_attention.linear_kv_up_proj.weight +Params for bucket 16 (41163264 elements, 41163264 padded size): + module.decoder.layers.5.self_attention.linear_proj.weight + module.decoder.layers.4.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.4.mlp.router.weight + module.decoder.layers.4.self_attention.linear_kv_up_proj.weight + module.decoder.layers.3.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.5.input_layernorm.weight + module.decoder.layers.4.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.4.pre_mlp_layernorm.weight + module.decoder.layers.4.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.4.self_attention.linear_kv_down_proj.weight + module.decoder.layers.4.self_attention.linear_q_proj.weight + module.decoder.layers.4.self_attention.linear_proj.weight + module.decoder.layers.4.input_layernorm.weight +Params for bucket 17 (42734080 elements, 42734080 padded size): + module.decoder.layers.3.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.3.pre_mlp_layernorm.weight + module.decoder.layers.3.self_attention.linear_kv_down_proj.weight + module.decoder.layers.3.self_attention.linear_q_proj.weight + module.decoder.layers.3.self_attention.linear_proj.weight + module.decoder.layers.3.input_layernorm.weight + module.decoder.layers.2.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.3.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.3.mlp.router.weight + module.decoder.layers.3.self_attention.linear_kv_up_proj.weight + module.decoder.layers.2.mlp.shared_experts.linear_fc2.weight +Params for bucket 18 (40901632 elements, 40901632 padded size): + module.decoder.layers.2.self_attention.linear_kv_down_proj.weight + module.decoder.layers.2.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.2.pre_mlp_layernorm.weight + module.decoder.layers.2.self_attention.linear_q_proj.weight + module.decoder.layers.2.self_attention.linear_proj.weight + module.decoder.layers.2.input_layernorm.weight + module.decoder.layers.1.mlp.shared_experts.linear_fc1.weight + module.decoder.layers.1.mlp.router.weight + module.decoder.layers.1.pre_mlp_layernorm.weight + module.decoder.layers.1.self_attention.linear_kv_up_proj.weight + module.decoder.layers.1.self_attention.linear_kv_down_proj.weight + module.decoder.layers.2.mlp.router.weight + module.decoder.layers.2.self_attention.linear_kv_up_proj.weight + module.decoder.layers.1.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.1.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.1.self_attention.linear_q_proj.weight +Params for bucket 19 (71436288 elements, 71436288 padded size): + module.decoder.layers.1.self_attention.linear_proj.weight + module.decoder.layers.1.input_layernorm.weight + module.decoder.layers.0.mlp.linear_fc2.weight + module.decoder.layers.0.mlp.linear_fc1.weight +Params for bucket 20 (116789760 elements, 116789760 padded size): + module.decoder.layers.0.mlp.linear_fc1.layer_norm_weight + module.decoder.layers.0.self_attention.linear_kv_up_proj.layer_norm_weight + module.decoder.layers.0.self_attention.linear_kv_down_proj.weight + module.decoder.layers.0.self_attention.linear_proj.weight + module.decoder.layers.0.input_layernorm.weight + module.embedding.word_embeddings.weight + module.decoder.layers.0.self_attention.linear_kv_up_proj.weight + module.decoder.layers.0.self_attention.linear_q_proj.weight +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 40 +Params for bucket 1 (40370176 elements, 40370176 padded size): + module.decoder.layers.23.mlp.experts.linear_fc2.weight7 + module.decoder.layers.23.mlp.experts.linear_fc2.weight5 + module.decoder.layers.23.mlp.experts.linear_fc2.weight3 + module.decoder.layers.23.mlp.experts.linear_fc2.weight1 + module.decoder.layers.23.mlp.experts.linear_fc1.weight6 + module.decoder.layers.23.mlp.experts.linear_fc2.weight6 + module.decoder.layers.23.mlp.experts.linear_fc2.weight4 + module.decoder.layers.23.mlp.experts.linear_fc2.weight2 + module.decoder.layers.23.mlp.experts.linear_fc2.weight0 + module.decoder.layers.23.mlp.experts.linear_fc1.weight7 + module.decoder.layers.23.mlp.experts.linear_fc1.weight5 +Params for bucket 2 (40370176 elements, 40370176 padded size): + module.decoder.layers.23.mlp.experts.linear_fc1.weight4 + module.decoder.layers.23.mlp.experts.linear_fc1.weight2 + module.decoder.layers.23.mlp.experts.linear_fc1.weight0 + module.decoder.layers.22.mlp.experts.linear_fc2.weight7 + module.decoder.layers.22.mlp.experts.linear_fc2.weight5 + module.decoder.layers.23.mlp.experts.linear_fc1.weight3 + module.decoder.layers.23.mlp.experts.linear_fc1.weight1 + module.decoder.layers.22.mlp.experts.linear_fc2.weight6 + module.decoder.layers.22.mlp.experts.linear_fc2.weight4 +Params for bucket 3 (40370176 elements, 40370176 padded size): + module.decoder.layers.22.mlp.experts.linear_fc2.weight3 + module.decoder.layers.22.mlp.experts.linear_fc2.weight1 + module.decoder.layers.22.mlp.experts.linear_fc1.weight6 + module.decoder.layers.22.mlp.experts.linear_fc1.weight4 + module.decoder.layers.22.mlp.experts.linear_fc1.weight7 + module.decoder.layers.22.mlp.experts.linear_fc2.weight2 + module.decoder.layers.22.mlp.experts.linear_fc2.weight0 + module.decoder.layers.22.mlp.experts.linear_fc1.weight5 + module.decoder.layers.22.mlp.experts.linear_fc1.weight3 +Params for bucket 4 (40370176 elements, 40370176 padded size): + module.decoder.layers.22.mlp.experts.linear_fc1.weight2 + module.decoder.layers.21.mlp.experts.linear_fc2.weight6 + module.decoder.layers.22.mlp.experts.linear_fc1.weight0 + module.decoder.layers.21.mlp.experts.linear_fc2.weight4 + module.decoder.layers.21.mlp.experts.linear_fc2.weight2 + module.decoder.layers.21.mlp.experts.linear_fc2.weight0 + module.decoder.layers.21.mlp.experts.linear_fc2.weight7 + module.decoder.layers.22.mlp.experts.linear_fc1.weight1 + module.decoder.layers.21.mlp.experts.linear_fc2.weight5 + module.decoder.layers.21.mlp.experts.linear_fc2.weight3 + module.decoder.layers.21.mlp.experts.linear_fc2.weight1 +Params for bucket 5 (40370176 elements, 40370176 padded size): + module.decoder.layers.21.mlp.experts.linear_fc1.weight7 + module.decoder.layers.21.mlp.experts.linear_fc1.weight3 + module.decoder.layers.21.mlp.experts.linear_fc1.weight5 + module.decoder.layers.21.mlp.experts.linear_fc1.weight1 + module.decoder.layers.21.mlp.experts.linear_fc1.weight6 + module.decoder.layers.21.mlp.experts.linear_fc1.weight4 + module.decoder.layers.21.mlp.experts.linear_fc1.weight2 +Params for bucket 6 (40370176 elements, 40370176 padded size): + module.decoder.layers.20.mlp.experts.linear_fc2.weight7 + module.decoder.layers.20.mlp.experts.linear_fc2.weight5 + module.decoder.layers.20.mlp.experts.linear_fc2.weight3 + module.decoder.layers.20.mlp.experts.linear_fc2.weight1 + module.decoder.layers.20.mlp.experts.linear_fc1.weight6 + module.decoder.layers.21.mlp.experts.linear_fc1.weight0 + module.decoder.layers.20.mlp.experts.linear_fc2.weight6 + module.decoder.layers.20.mlp.experts.linear_fc2.weight4 + module.decoder.layers.20.mlp.experts.linear_fc2.weight2 + module.decoder.layers.20.mlp.experts.linear_fc2.weight0 + module.decoder.layers.20.mlp.experts.linear_fc1.weight7 +Params for bucket 7 (40370176 elements, 40370176 padded size): + module.decoder.layers.20.mlp.experts.linear_fc1.weight2 + module.decoder.layers.20.mlp.experts.linear_fc1.weight4 + module.decoder.layers.20.mlp.experts.linear_fc1.weight0 + module.decoder.layers.19.mlp.experts.linear_fc2.weight7 + module.decoder.layers.20.mlp.experts.linear_fc1.weight5 + module.decoder.layers.20.mlp.experts.linear_fc1.weight3 + module.decoder.layers.20.mlp.experts.linear_fc1.weight1 + module.decoder.layers.19.mlp.experts.linear_fc2.weight6 +Params for bucket 8 (40370176 elements, 40370176 padded size): + module.decoder.layers.19.mlp.experts.linear_fc2.weight5 + module.decoder.layers.19.mlp.experts.linear_fc2.weight1 + module.decoder.layers.19.mlp.experts.linear_fc2.weight3 + module.decoder.layers.19.mlp.experts.linear_fc1.weight6 + module.decoder.layers.19.mlp.experts.linear_fc1.weight4 + module.decoder.layers.19.mlp.experts.linear_fc2.weight4 + module.decoder.layers.19.mlp.experts.linear_fc2.weight2 + module.decoder.layers.19.mlp.experts.linear_fc2.weight0 + module.decoder.layers.19.mlp.experts.linear_fc1.weight7 + module.decoder.layers.19.mlp.experts.linear_fc1.weight5 +Params for bucket 9 (40370176 elements, 40370176 padded size): + module.decoder.layers.18.mlp.experts.linear_fc2.weight7 + module.decoder.layers.19.mlp.experts.linear_fc1.weight0 + module.decoder.layers.19.mlp.experts.linear_fc1.weight2 + module.decoder.layers.18.mlp.experts.linear_fc2.weight5 + module.decoder.layers.18.mlp.experts.linear_fc2.weight3 + module.decoder.layers.19.mlp.experts.linear_fc1.weight3 + module.decoder.layers.19.mlp.experts.linear_fc1.weight1 + module.decoder.layers.18.mlp.experts.linear_fc2.weight6 + module.decoder.layers.18.mlp.experts.linear_fc2.weight4 + module.decoder.layers.18.mlp.experts.linear_fc2.weight2 +Params for bucket 10 (40370176 elements, 40370176 padded size): + module.decoder.layers.18.mlp.experts.linear_fc2.weight1 + module.decoder.layers.18.mlp.experts.linear_fc1.weight6 + module.decoder.layers.18.mlp.experts.linear_fc1.weight4 + module.decoder.layers.18.mlp.experts.linear_fc1.weight2 + module.decoder.layers.18.mlp.experts.linear_fc1.weight7 + module.decoder.layers.18.mlp.experts.linear_fc1.weight5 + module.decoder.layers.18.mlp.experts.linear_fc2.weight0 + module.decoder.layers.18.mlp.experts.linear_fc1.weight3 +Params for bucket 11 (40370176 elements, 40370176 padded size): + module.decoder.layers.17.mlp.experts.linear_fc2.weight7 + module.decoder.layers.18.mlp.experts.linear_fc1.weight0 + module.decoder.layers.17.mlp.experts.linear_fc2.weight5 + module.decoder.layers.17.mlp.experts.linear_fc2.weight3 + module.decoder.layers.17.mlp.experts.linear_fc2.weight1 + module.decoder.layers.18.mlp.experts.linear_fc1.weight1 + module.decoder.layers.17.mlp.experts.linear_fc2.weight6 + module.decoder.layers.17.mlp.experts.linear_fc2.weight4 + module.decoder.layers.17.mlp.experts.linear_fc2.weight2 + module.decoder.layers.17.mlp.experts.linear_fc2.weight0 + module.decoder.layers.17.mlp.experts.linear_fc1.weight7 +Params for bucket 12 (40370176 elements, 40370176 padded size): + module.decoder.layers.17.mlp.experts.linear_fc1.weight6 + module.decoder.layers.17.mlp.experts.linear_fc1.weight4 + module.decoder.layers.17.mlp.experts.linear_fc1.weight2 + module.decoder.layers.17.mlp.experts.linear_fc1.weight0 + module.decoder.layers.17.mlp.experts.linear_fc1.weight5 + module.decoder.layers.17.mlp.experts.linear_fc1.weight3 + module.decoder.layers.17.mlp.experts.linear_fc1.weight1 +Params for bucket 13 (40370176 elements, 40370176 padded size): + module.decoder.layers.16.mlp.experts.linear_fc2.weight7 + module.decoder.layers.16.mlp.experts.linear_fc2.weight3 + module.decoder.layers.16.mlp.experts.linear_fc2.weight5 + module.decoder.layers.16.mlp.experts.linear_fc2.weight1 + module.decoder.layers.16.mlp.experts.linear_fc1.weight6 + module.decoder.layers.16.mlp.experts.linear_fc2.weight4 + module.decoder.layers.16.mlp.experts.linear_fc2.weight6 + module.decoder.layers.16.mlp.experts.linear_fc2.weight2 + module.decoder.layers.16.mlp.experts.linear_fc2.weight0 + module.decoder.layers.16.mlp.experts.linear_fc1.weight7 + module.decoder.layers.16.mlp.experts.linear_fc1.weight5 +Params for bucket 14 (40370176 elements, 40370176 padded size): + module.decoder.layers.16.mlp.experts.linear_fc1.weight4 + module.decoder.layers.16.mlp.experts.linear_fc1.weight0 + module.decoder.layers.16.mlp.experts.linear_fc1.weight2 + module.decoder.layers.15.mlp.experts.linear_fc2.weight6 + module.decoder.layers.15.mlp.experts.linear_fc2.weight4 + module.decoder.layers.16.mlp.experts.linear_fc1.weight3 + module.decoder.layers.16.mlp.experts.linear_fc1.weight1 + module.decoder.layers.15.mlp.experts.linear_fc2.weight7 + module.decoder.layers.15.mlp.experts.linear_fc2.weight5 +Params for bucket 15 (40370176 elements, 40370176 padded size): + module.decoder.layers.15.mlp.experts.linear_fc2.weight2 + module.decoder.layers.15.mlp.experts.linear_fc1.weight7 + module.decoder.layers.15.mlp.experts.linear_fc2.weight0 + module.decoder.layers.15.mlp.experts.linear_fc1.weight5 + module.decoder.layers.15.mlp.experts.linear_fc1.weight3 + module.decoder.layers.15.mlp.experts.linear_fc2.weight3 + module.decoder.layers.15.mlp.experts.linear_fc2.weight1 + module.decoder.layers.15.mlp.experts.linear_fc1.weight6 + module.decoder.layers.15.mlp.experts.linear_fc1.weight4 +Params for bucket 16 (40370176 elements, 40370176 padded size): + module.decoder.layers.14.mlp.experts.linear_fc2.weight6 + module.decoder.layers.15.mlp.experts.linear_fc1.weight1 + module.decoder.layers.14.mlp.experts.linear_fc2.weight4 + module.decoder.layers.14.mlp.experts.linear_fc2.weight2 + module.decoder.layers.14.mlp.experts.linear_fc2.weight0 + module.decoder.layers.15.mlp.experts.linear_fc1.weight2 + module.decoder.layers.15.mlp.experts.linear_fc1.weight0 + module.decoder.layers.14.mlp.experts.linear_fc2.weight7 + module.decoder.layers.14.mlp.experts.linear_fc2.weight5 + module.decoder.layers.14.mlp.experts.linear_fc2.weight3 + module.decoder.layers.14.mlp.experts.linear_fc2.weight1 +Params for bucket 17 (40370176 elements, 40370176 padded size): + module.decoder.layers.14.mlp.experts.linear_fc1.weight7 + module.decoder.layers.14.mlp.experts.linear_fc1.weight5 + module.decoder.layers.14.mlp.experts.linear_fc1.weight3 + module.decoder.layers.14.mlp.experts.linear_fc1.weight1 + module.decoder.layers.14.mlp.experts.linear_fc1.weight6 + module.decoder.layers.14.mlp.experts.linear_fc1.weight4 + module.decoder.layers.14.mlp.experts.linear_fc1.weight2 +Params for bucket 18 (40370176 elements, 40370176 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.weight1 + module.decoder.layers.13.mlp.experts.linear_fc1.weight6 + module.decoder.layers.14.mlp.experts.linear_fc1.weight0 + 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 + module.decoder.layers.13.mlp.experts.linear_fc2.weight0 + module.decoder.layers.13.mlp.experts.linear_fc1.weight7 +Params for bucket 19 (40370176 elements, 40370176 padded size): + module.decoder.layers.13.mlp.experts.linear_fc1.weight4 + module.decoder.layers.13.mlp.experts.linear_fc1.weight2 + module.decoder.layers.13.mlp.experts.linear_fc1.weight0 + module.decoder.layers.12.mlp.experts.linear_fc2.weight7 + module.decoder.layers.13.mlp.experts.linear_fc1.weight5 + module.decoder.layers.13.mlp.experts.linear_fc1.weight1 + module.decoder.layers.13.mlp.experts.linear_fc1.weight3 + module.decoder.layers.12.mlp.experts.linear_fc2.weight6 +Params for bucket 20 (40370176 elements, 40370176 padded size): + module.decoder.layers.12.mlp.experts.linear_fc2.weight5 + module.decoder.layers.12.mlp.experts.linear_fc2.weight1 + module.decoder.layers.12.mlp.experts.linear_fc2.weight3 + module.decoder.layers.12.mlp.experts.linear_fc1.weight6 + module.decoder.layers.12.mlp.experts.linear_fc1.weight4 + module.decoder.layers.12.mlp.experts.linear_fc2.weight2 + module.decoder.layers.12.mlp.experts.linear_fc2.weight4 + module.decoder.layers.12.mlp.experts.linear_fc2.weight0 + module.decoder.layers.12.mlp.experts.linear_fc1.weight7 + module.decoder.layers.12.mlp.experts.linear_fc1.weight5 +Params for bucket 21 (40370176 elements, 40370176 padded size): + module.decoder.layers.12.mlp.experts.linear_fc1.weight2 + module.decoder.layers.12.mlp.experts.linear_fc1.weight0 + module.decoder.layers.11.mlp.experts.linear_fc2.weight6 + module.decoder.layers.11.mlp.experts.linear_fc2.weight4 + module.decoder.layers.11.mlp.experts.linear_fc2.weight2 + module.decoder.layers.12.mlp.experts.linear_fc1.weight3 + module.decoder.layers.11.mlp.experts.linear_fc2.weight7 + module.decoder.layers.12.mlp.experts.linear_fc1.weight1 + module.decoder.layers.11.mlp.experts.linear_fc2.weight5 + module.decoder.layers.11.mlp.experts.linear_fc2.weight3 +Params for bucket 22 (40370176 elements, 40370176 padded size): + module.decoder.layers.11.mlp.experts.linear_fc1.weight7 + module.decoder.layers.11.mlp.experts.linear_fc1.weight5 + module.decoder.layers.11.mlp.experts.linear_fc2.weight0 + module.decoder.layers.11.mlp.experts.linear_fc1.weight3 + module.decoder.layers.11.mlp.experts.linear_fc2.weight1 + module.decoder.layers.11.mlp.experts.linear_fc1.weight6 + module.decoder.layers.11.mlp.experts.linear_fc1.weight4 + module.decoder.layers.11.mlp.experts.linear_fc1.weight2 +Params for bucket 23 (40370176 elements, 40370176 padded size): + module.decoder.layers.11.mlp.experts.linear_fc1.weight1 + module.decoder.layers.10.mlp.experts.linear_fc2.weight6 + module.decoder.layers.10.mlp.experts.linear_fc2.weight4 + module.decoder.layers.10.mlp.experts.linear_fc2.weight2 + module.decoder.layers.10.mlp.experts.linear_fc2.weight0 + module.decoder.layers.10.mlp.experts.linear_fc2.weight7 + module.decoder.layers.11.mlp.experts.linear_fc1.weight0 + module.decoder.layers.10.mlp.experts.linear_fc2.weight5 + module.decoder.layers.10.mlp.experts.linear_fc2.weight3 + module.decoder.layers.10.mlp.experts.linear_fc2.weight1 + module.decoder.layers.10.mlp.experts.linear_fc1.weight7 +Params for bucket 24 (40370176 elements, 40370176 padded size): + module.decoder.layers.10.mlp.experts.linear_fc1.weight6 + module.decoder.layers.10.mlp.experts.linear_fc1.weight2 + module.decoder.layers.10.mlp.experts.linear_fc1.weight4 + module.decoder.layers.10.mlp.experts.linear_fc1.weight0 + module.decoder.layers.10.mlp.experts.linear_fc1.weight5 + module.decoder.layers.10.mlp.experts.linear_fc1.weight3 + module.decoder.layers.10.mlp.experts.linear_fc1.weight1 +Params for bucket 25 (40370176 elements, 40370176 padded size): + module.decoder.layers.9.mlp.experts.linear_fc2.weight7 + module.decoder.layers.9.mlp.experts.linear_fc2.weight5 + module.decoder.layers.9.mlp.experts.linear_fc2.weight3 + module.decoder.layers.9.mlp.experts.linear_fc2.weight1 + module.decoder.layers.9.mlp.experts.linear_fc1.weight6 + module.decoder.layers.9.mlp.experts.linear_fc2.weight6 + module.decoder.layers.9.mlp.experts.linear_fc2.weight4 + module.decoder.layers.9.mlp.experts.linear_fc2.weight2 + module.decoder.layers.9.mlp.experts.linear_fc2.weight0 + module.decoder.layers.9.mlp.experts.linear_fc1.weight7 + module.decoder.layers.9.mlp.experts.linear_fc1.weight5 +Params for bucket 26 (40370176 elements, 40370176 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight4 + module.decoder.layers.9.mlp.experts.linear_fc1.weight2 + module.decoder.layers.9.mlp.experts.linear_fc1.weight0 + module.decoder.layers.8.mlp.experts.linear_fc2.weight7 + module.decoder.layers.8.mlp.experts.linear_fc2.weight5 + module.decoder.layers.9.mlp.experts.linear_fc1.weight3 + module.decoder.layers.9.mlp.experts.linear_fc1.weight1 + module.decoder.layers.8.mlp.experts.linear_fc2.weight6 + module.decoder.layers.8.mlp.experts.linear_fc2.weight4 +Params for bucket 27 (40370176 elements, 40370176 padded size): + module.decoder.layers.8.mlp.experts.linear_fc2.weight3 + module.decoder.layers.8.mlp.experts.linear_fc2.weight1 + module.decoder.layers.8.mlp.experts.linear_fc1.weight6 + module.decoder.layers.8.mlp.experts.linear_fc1.weight4 + module.decoder.layers.8.mlp.experts.linear_fc2.weight2 + module.decoder.layers.8.mlp.experts.linear_fc2.weight0 + module.decoder.layers.8.mlp.experts.linear_fc1.weight7 + module.decoder.layers.8.mlp.experts.linear_fc1.weight5 + module.decoder.layers.8.mlp.experts.linear_fc1.weight3 +Params for bucket 28 (40370176 elements, 40370176 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight2 + module.decoder.layers.7.mlp.experts.linear_fc2.weight7 + module.decoder.layers.8.mlp.experts.linear_fc1.weight0 + module.decoder.layers.7.mlp.experts.linear_fc2.weight5 + module.decoder.layers.7.mlp.experts.linear_fc2.weight3 + module.decoder.layers.7.mlp.experts.linear_fc2.weight1 + module.decoder.layers.7.mlp.experts.linear_fc2.weight6 + module.decoder.layers.8.mlp.experts.linear_fc1.weight1 + module.decoder.layers.7.mlp.experts.linear_fc2.weight4 + module.decoder.layers.7.mlp.experts.linear_fc2.weight2 + module.decoder.layers.7.mlp.experts.linear_fc2.weight0 +Params for bucket 29 (40370176 elements, 40370176 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight6 + module.decoder.layers.7.mlp.experts.linear_fc1.weight4 + module.decoder.layers.7.mlp.experts.linear_fc1.weight2 + module.decoder.layers.7.mlp.experts.linear_fc1.weight7 + module.decoder.layers.7.mlp.experts.linear_fc1.weight5 + module.decoder.layers.7.mlp.experts.linear_fc1.weight3 + module.decoder.layers.7.mlp.experts.linear_fc1.weight1 +Params for bucket 30 (40370176 elements, 40370176 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight0 + module.decoder.layers.6.mlp.experts.linear_fc2.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.weight1 + module.decoder.layers.6.mlp.experts.linear_fc1.weight6 + 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.weight0 + module.decoder.layers.6.mlp.experts.linear_fc1.weight7 +Params for bucket 31 (40370176 elements, 40370176 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight2 + module.decoder.layers.6.mlp.experts.linear_fc1.weight4 + module.decoder.layers.6.mlp.experts.linear_fc1.weight0 + module.decoder.layers.5.mlp.experts.linear_fc2.weight7 + module.decoder.layers.6.mlp.experts.linear_fc1.weight5 + module.decoder.layers.6.mlp.experts.linear_fc1.weight1 + module.decoder.layers.6.mlp.experts.linear_fc1.weight3 + module.decoder.layers.5.mlp.experts.linear_fc2.weight6 +Params for bucket 32 (40370176 elements, 40370176 padded size): + module.decoder.layers.5.mlp.experts.linear_fc2.weight5 + module.decoder.layers.5.mlp.experts.linear_fc2.weight3 + 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_fc1.weight4 + 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.weight0 + module.decoder.layers.5.mlp.experts.linear_fc1.weight7 + module.decoder.layers.5.mlp.experts.linear_fc1.weight5 +Params for bucket 33 (40370176 elements, 40370176 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight2 + module.decoder.layers.5.mlp.experts.linear_fc1.weight0 + 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.5.mlp.experts.linear_fc1.weight3 + module.decoder.layers.5.mlp.experts.linear_fc1.weight1 + module.decoder.layers.4.mlp.experts.linear_fc2.weight6 + module.decoder.layers.4.mlp.experts.linear_fc2.weight4 + module.decoder.layers.4.mlp.experts.linear_fc2.weight2 +Params for bucket 34 (40370176 elements, 40370176 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight1 + module.decoder.layers.4.mlp.experts.linear_fc1.weight6 + module.decoder.layers.4.mlp.experts.linear_fc1.weight4 + module.decoder.layers.4.mlp.experts.linear_fc1.weight2 + module.decoder.layers.4.mlp.experts.linear_fc2.weight0 + module.decoder.layers.4.mlp.experts.linear_fc1.weight7 + module.decoder.layers.4.mlp.experts.linear_fc1.weight5 + module.decoder.layers.4.mlp.experts.linear_fc1.weight3 +Params for bucket 35 (40370176 elements, 40370176 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight0 + module.decoder.layers.3.mlp.experts.linear_fc2.weight7 + module.decoder.layers.3.mlp.experts.linear_fc2.weight5 + module.decoder.layers.3.mlp.experts.linear_fc2.weight3 + module.decoder.layers.3.mlp.experts.linear_fc2.weight1 + module.decoder.layers.4.mlp.experts.linear_fc1.weight1 + 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 + module.decoder.layers.3.mlp.experts.linear_fc2.weight0 + module.decoder.layers.3.mlp.experts.linear_fc1.weight7 +Params for bucket 36 (40370176 elements, 40370176 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight6 + module.decoder.layers.3.mlp.experts.linear_fc1.weight2 + module.decoder.layers.3.mlp.experts.linear_fc1.weight4 + module.decoder.layers.3.mlp.experts.linear_fc1.weight0 + module.decoder.layers.3.mlp.experts.linear_fc1.weight3 + module.decoder.layers.3.mlp.experts.linear_fc1.weight5 + module.decoder.layers.3.mlp.experts.linear_fc1.weight1 +Params for bucket 37 (40370176 elements, 40370176 padded size): + module.decoder.layers.2.mlp.experts.linear_fc2.weight6 + module.decoder.layers.2.mlp.experts.linear_fc2.weight4 + module.decoder.layers.2.mlp.experts.linear_fc2.weight2 + module.decoder.layers.2.mlp.experts.linear_fc2.weight0 + module.decoder.layers.2.mlp.experts.linear_fc1.weight6 + module.decoder.layers.2.mlp.experts.linear_fc2.weight7 + module.decoder.layers.2.mlp.experts.linear_fc2.weight5 + module.decoder.layers.2.mlp.experts.linear_fc2.weight3 + module.decoder.layers.2.mlp.experts.linear_fc2.weight1 + module.decoder.layers.2.mlp.experts.linear_fc1.weight7 + module.decoder.layers.2.mlp.experts.linear_fc1.weight5 +Params for bucket 38 (40370176 elements, 40370176 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight0 + module.decoder.layers.2.mlp.experts.linear_fc1.weight2 + module.decoder.layers.1.mlp.experts.linear_fc2.weight7 + module.decoder.layers.1.mlp.experts.linear_fc2.weight5 + module.decoder.layers.2.mlp.experts.linear_fc1.weight3 + module.decoder.layers.2.mlp.experts.linear_fc1.weight1 + module.decoder.layers.1.mlp.experts.linear_fc2.weight6 + module.decoder.layers.1.mlp.experts.linear_fc2.weight4 +Params for bucket 39 (40370176 elements, 40370176 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_fc1.weight6 + module.decoder.layers.1.mlp.experts.linear_fc1.weight4 + module.decoder.layers.1.mlp.experts.linear_fc2.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight7 + module.decoder.layers.1.mlp.experts.linear_fc2.weight0 + module.decoder.layers.1.mlp.experts.linear_fc1.weight5 + module.decoder.layers.1.mlp.experts.linear_fc1.weight3 +Params for bucket 40 (17301504 elements, 17301504 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight0 + module.decoder.layers.1.mlp.experts.linear_fc1.weight1 +INFO:megatron.core.optimizer:Setting up optimizer with config OptimizerConfig(optimizer='adam', lr=0.0001, min_lr=1e-05, decoupled_lr=None, decoupled_min_lr=None, weight_decay=0.1, fp8_recipe='mxfp8', fp16=False, bf16=True, reuse_grad_buf_for_mxfp8_param_ag=False, params_dtype=torch.bfloat16, use_precision_aware_optimizer=False, store_param_remainders=True, main_grads_dtype=torch.float32, main_params_dtype=torch.float32, exp_avg_dtype=torch.float32, exp_avg_sq_dtype=torch.float32, loss_scale=None, initial_loss_scale=4294967296, min_loss_scale=1.0, loss_scale_window=1000, hysteresis=2, adam_beta1=0.9, adam_beta2=0.999, adam_eps=1e-08, sgd_momentum=0.9, use_distributed_optimizer=True, overlap_param_gather_with_optimizer_step=False, optimizer_cpu_offload=False, optimizer_offload_fraction=1.0, use_torch_optimizer_for_cpu_offload=False, overlap_cpu_optimizer_d2h_h2d=False, pin_cpu_grads=True, pin_cpu_params=True, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=, config_logger_dir='') +INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine +[after model, optimizer, and learning rate scheduler are built] datetime: 2026-08-26 18:37:58 +> building train, validation, and test datasets ... + > datasets target sizes (minimum size): + train: 96 + 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=(96, 0, 0) and config=GPTDatasetConfig(random_seed=1234, sequence_length=8192, blend=(['/workspace/data/wikitext/wikitext103_gpt2_text_document'], None), blend_per_split=None, split='100,0,0', split_matrix=[(0, 1.0), None, None], num_dataset_builder_threads=1, path_to_cache=None, mmap_bin_files=True, mock=False, tokenizer=, mid_level_dataset_surplus=0.005, reset_position_ids=False, reset_attention_mask=False, eod_mask_loss=False, create_attention_mask=True, drop_last_partial_validation_sequence=True, add_extra_token_to_sequence=True, object_storage_cache_path=None) +INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx +INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths +INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers +INFO:megatron.core.datasets.indexed_dataset: Extract the document indices +INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029 +INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350 +INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices +INFO:megatron.core.datasets.gpt_dataset: Load the document index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-document_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the sample index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-sample_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-shuffle_index.npy +INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 14536 +> finished creating GPT datasets ... +[after dataloaders are built] datetime: 2026-08-26 18:37:58 +done with setup ... +training ... +Overwriting rerun_state_machine.current_iteration from -1 to 0... +(min, max) time across ranks (ms): + model-and-optimizer-setup ......................: (3732.27, 3739.39) + train/valid/test-data-iterators-setup ..........: (70.43, 90.52) +[before the start of training step] datetime: 2026-08-26 18:37:58 +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py: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: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:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") + [2026-08-26 18:38:25] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 27097.5 | throughput per GPU (TFLOP/s/GPU): 4.7 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.092127E+01 | loss scale: 1.0 | grad norm: 30.623 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | +Number of parameters in transformer block in billions: 13.53 +Number of parameters in embedding layers in billions: 0.21 +Total number of parameters in billions: 13.74 +Number of parameters in most loaded shard in billions: 13.7356 +Theoretical memory footprints: weight and optimizer=98244.96 MB +[Rank 0] (after 1 iterations) memory (MB) | allocated: 37181.40869140625 | max allocated: 102778.78173828125 | reserved: 105676.0 | max reserved: 105676.0 +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py: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: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.") + [2026-08-26 18:38:26] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 594.9 | throughput per GPU (TFLOP/s/GPU): 214.4 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.091995E+01 | loss scale: 1.0 | grad norm: 30.567 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:38:26] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 598.0 | throughput per GPU (TFLOP/s/GPU): 213.3 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 9.952354E+00 | loss scale: 1.0 | grad norm: 42.232 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | +Capture CUDA graph for training!!! +CUDA graph capture done!!! + [2026-08-26 18:38:28] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 1742.7 | throughput per GPU (TFLOP/s/GPU): 73.2 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 9.477323E+00 | loss scale: 1.0 | grad norm: 7.143 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:38:29] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 576.9 | throughput per GPU (TFLOP/s/GPU): 221.1 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.119604E+01 | loss scale: 1.0 | grad norm: 396.977 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:38:29] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 583.0 | throughput per GPU (TFLOP/s/GPU): 218.9 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 9.367741E+00 | loss scale: 1.0 | grad norm: 5.090 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:38:30] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 581.8 | throughput per GPU (TFLOP/s/GPU): 219.3 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 8.973585E+00 | loss scale: 1.0 | grad norm: 3.690 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:38:30] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 579.9 | throughput per GPU (TFLOP/s/GPU): 220.0 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.787342E+00 | loss scale: 1.0 | grad norm: 5.890 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:38:31] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 579.1 | throughput per GPU (TFLOP/s/GPU): 220.3 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.612353E+00 | loss scale: 1.0 | grad norm: 3.038 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:38:32] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 578.9 | throughput per GPU (TFLOP/s/GPU): 220.4 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.076241E+00 | loss scale: 1.0 | grad norm: 29.812 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:38:32] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 579.0 | throughput per GPU (TFLOP/s/GPU): 220.4 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.639501E+00 | loss scale: 1.0 | grad norm: 64.594 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | + [2026-08-26 18:38:33] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 578.6 | throughput per GPU (TFLOP/s/GPU): 220.5 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.624613E+00 | loss scale: 1.0 | grad norm: 71.371 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 50560 | min safe capacity: 1.0286 | +[after training is done] datetime: 2026-08-26 18:38:33 +[rank0]:[W826 18:38:34.387310058 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank1]:[W826 18:38:35.934452871 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank6]:[W826 18:38:35.936101233 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank4]:[W826 18:38:35.938073110 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank3]:[W826 18:38:35.939802454 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank2]:[W826 18:38:35.942204194 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank5]:[W826 18:38:36.046498396 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank7]:[W826 18:38:36.049853215 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +W0826 18:38:39.071000 535038 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers +W0826 18:38:39.071000 535038 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 535124 closing signal SIGTERM +W0826 18:38:39.072000 535038 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 535125 closing signal SIGTERM +W0826 18:38:39.074000 535038 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 535126 closing signal SIGTERM +W0826 18:38:39.076000 535038 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 535127 closing signal SIGTERM +W0826 18:38:39.077000 535038 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 535128 closing signal SIGTERM +W0826 18:38:39.077000 535038 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 535129 closing signal SIGTERM +W0826 18:38:39.079000 535038 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 535130 closing signal SIGTERM +W0826 18:38:39.081000 535038 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 535131 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 535038 got signal: 15 diff --git a/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep1/mem_samples.csv b/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep1/mem_samples.csv new file mode 100644 index 0000000000000000000000000000000000000000..4340eb944ef589cc8953e6062829d27c97bf4be6 --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep1/mem_samples.csv @@ -0,0 +1,192 @@ +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 2 +1, 1 +2, 1 +3, 1 +4, 1 +5, 1 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 1 +5, 1 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 4 +5, 4 +6, 4 +7, 4 +0, 4 +1, 27 +2, 4 +3, 4 +4, 37 +5, 4 +6, 4 +7, 4 +0, 821 +1, 821 +2, 821 +3, 821 +4, 821 +5, 821 +6, 821 +7, 821 +0, 1803 +1, 1803 +2, 1803 +3, 1803 +4, 1803 +5, 1803 +6, 1803 +7, 1803 +0, 2403 +1, 2375 +2, 2373 +3, 2367 +4, 2347 +5, 2361 +6, 2371 +7, 2369 +0, 46703 +1, 46137 +2, 46137 +3, 46837 +4, 46837 +5, 46837 +6, 46837 +7, 46837 +0, 47125 +1, 46555 +2, 46557 +3, 47255 +4, 47257 +5, 47263 +6, 47255 +7, 47257 +0, 48413 +1, 47843 +2, 47843 +3, 48543 +4, 48543 +5, 48543 +6, 48543 +7, 48543 +0, 48659 +1, 48089 +2, 48089 +3, 48789 +4, 48789 +5, 48789 +6, 48789 +7, 48789 +0, 48659 +1, 48089 +2, 48089 +3, 48789 +4, 48789 +5, 48789 +6, 48789 +7, 48789 +0, 50727 +1, 50157 +2, 50157 +3, 50857 +4, 50857 +5, 50857 +6, 50857 +7, 50857 +0, 73613 +1, 73627 +2, 73609 +3, 73629 +4, 73635 +5, 73633 +6, 73639 +7, 73641 +0, 74163 +1, 74141 +2, 74123 +3, 74167 +4, 74167 +5, 74167 +6, 74167 +7, 74167 +0, 74163 +1, 74143 +2, 74125 +3, 74169 +4, 74169 +5, 74169 +6, 74169 +7, 74169 +0, 74505 +1, 74377 +2, 74373 +3, 74571 +4, 74407 +5, 74355 +6, 74431 +7, 74439 +0, 82097 +1, 81703 +2, 81947 +3, 82057 +4, 82035 +5, 82043 +6, 81979 +7, 82077 +0, 96727 +1, 96777 +2, 96809 +3, 96709 +4, 96709 +5, 96709 +6, 96709 +7, 96709 +0, 98599 +1, 98233 +2, 98423 +3, 98637 +4, 98637 +5, 98637 +6, 98637 +7, 98637 +0, 98599 +1, 98233 +2, 98423 +3, 98637 +4, 98637 +5, 98637 +6, 98637 +7, 98637 +0, 95137 +1, 1284 +2, 4 +3, 4 +4, 4 +5, 4 +6, 4 +7, 260 diff --git a/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep1/run.env b/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep1/run.env new file mode 100644 index 0000000000000000000000000000000000000000..f42d1791470a05fc183c10bb34677f755bbb4e2c --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep1/run.env @@ -0,0 +1,61 @@ +backend=ultraep +config=QWEN2_57B +family=tierB +repo=/workspace/Megatron-UltraEP +branch=ultraep-bracing-regions-through-kernels +commit=03d57ef7 +num_layers=10 +hidden=3584 +ffn_hidden=2560 +moe_ffn_hidden=2560 +num_experts=64 +topk=8 +experts_per_rank=8 +expert_params=27525120 +heads=28 +kv_groups=4 +kv_channels= +qk_layernorm=0 +shared_ffn=20480 +shared_expert_overlap=0 +mla=0 +dense_first_layers=0 +moe_layers=10 +rotary_base=1000000 +rotary_scaling= +qkv_bias=1 +router_balancing=aux_loss +aux_loss_coeff=1e-2 +router_pre_softmax=1 +router_fusion=0 +router_dtype=fp32 +device_initiated_grouped_gemm=1 +norm_epsilon=1e-6 +seq_length=4096 +micro_batch_size=1 +global_batch_size=8 +num_microbatches=1 +assignments_per_rank=32768 +mean_tokens_per_expert=4096 +slot_policy=matched +slots_per_rank=8 +capacity_mode=static +capacity=1.08984375 +cuda_graph=1 +probe_mode=0 +async_grad_reduce=0 +branch_role=timing +memory_dump_mode=none +memory_snapshot_path= +cuda_graph_warmup_steps=3 +recompute=none +force_balance=0 +train_iters=12 +ep_size=8 +echo_dispatch_overlap=1 +echo_dispatch_recompute=0 +ultraep_quota_min_tokens=512 +ultraep_weight_sync_join= +profile_enable=0 +kernel_trace=0 +fp8_recipe=mxfp8 diff --git a/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep1/train.log b/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep1/train.log new file mode 100644 index 0000000000000000000000000000000000000000..d24714382f98d861496dc0973cae23080cae8ee7 --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep1/train.log @@ -0,0 +1,1194 @@ +============================================================== + QWEN2_57B / ultraep (slots/rank=8 policy=matched) + E=64 K=8 H'=2560 m=8 + L=10 seq=4096 gbs=8 (1 microbatches) + cap_mode=static cap=1.08984375 cg=1 recompute=none + memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep1/memory_snapshots/snapshot.pickle + repo=/workspace/Megatron-UltraEP @ 03d57ef7 +============================================================== +W0826 18:38:44.640000 538192 torch/distributed/run.py:766] +W0826 18:38:44.640000 538192 torch/distributed/run.py:766] ***************************************** +W0826 18:38:44.640000 538192 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0826 18:38:44.640000 538192 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") +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") +using world size: 8, data-parallel size: 8, context-parallel size: 1, hierarchical context-parallel sizes: None, tensor-model-parallel size: 1, pipeline-model-parallel size: 1 +WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer +Number of virtual stages per pipeline stage: None +accumulate and all-reduce gradients in fp32 for bfloat16 data type. +using torch.bfloat16 for parameters ... +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +------------------------ arguments ------------------------ + account_for_embedding_in_pipeline_split ......... False + account_for_loss_in_pipeline_split .............. False + accumulate_allreduce_grads_in_fp32 .............. True + adam_beta1 ...................................... 0.9 + adam_beta2 ...................................... 0.999 + adam_eps ........................................ 1e-08 + add_bias_linear ................................. False + add_position_embedding .......................... False + add_qkv_bias .................................... True + adlr_autoresume ................................. False + adlr_autoresume_interval ........................ 1000 + align_grad_reduce ............................... True + align_param_gather .............................. False + app_tag_run_name ................................ None + app_tag_run_version ............................. 0.0.0 + apply_layernorm_1p .............................. False + apply_query_key_layer_scaling ................... False + apply_residual_connection_post_layernorm ........ False + apply_rope_fusion ............................... True + async_save ...................................... None + async_tensor_model_parallel_allreduce ........... True + attention_backend ............................... AttnBackend.auto + attention_dropout ............................... 0.0 + attention_softmax_in_fp32 ....................... False + auto_detect_ckpt_format ......................... False + barrier_with_L1_time ............................ True + bert_binary_head ................................ True + bert_embedder_type .............................. megatron + bert_load ....................................... None + bf16 ............................................ True + bias_dropout_fusion ............................. True + bias_gelu_fusion ................................ False + bias_swiglu_fusion .............................. True + biencoder_projection_dim ........................ 0 + biencoder_shared_query_context_model ............ False + block_data_path ................................. None + calc_ft_timeouts ................................ False + calculate_per_token_loss ........................ False + check_for_large_grads ........................... False + check_for_nan_in_loss_and_grad .................. False + check_for_spiky_loss ............................ False + check_weight_hash_across_dp_replicas_interval ... None + ckpt_assume_constant_structure .................. False + ckpt_convert_format ............................. None + ckpt_convert_save ............................... None + ckpt_convert_update_legacy_dist_opt_format ...... False + ckpt_format ..................................... torch_dist + ckpt_fully_parallel_load ........................ False + ckpt_fully_parallel_save ........................ True + ckpt_fully_parallel_save_deprecated ............. False + ckpt_step ....................................... None + classes_fraction ................................ 1.0 + clip_grad ....................................... 1.0 + clone_scatter_output_in_embedding ............... True + config_logger_dir ............................... + consumed_train_samples .......................... 0 + consumed_valid_samples .......................... 0 + context_parallel_size ........................... 1 + cp_comm_type .................................... ['p2p'] + create_attention_mask_in_dataloader ............. True + cross_entropy_fusion_impl ....................... native + cross_entropy_loss_fusion ....................... False + cuda_graph_scope ................................ full_iteration + cuda_graph_warmup_steps ......................... 3 + data_args_path .................................. None + data_cache_path ................................. None + data_parallel_random_init ....................... False + data_parallel_sharding_strategy ................. no_shard + data_parallel_size .............................. 8 + data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document'] + data_per_class_fraction ......................... 1.0 + data_sharding ................................... True + dataloader_type ................................. single + ddp_average_in_collective ....................... False + ddp_bucket_size ................................. None + ddp_num_buckets ................................. None + ddp_pad_buckets_for_high_nccl_busbw ............. False + decoder_first_pipeline_num_layers ............... None + decoder_last_pipeline_num_layers ................ None + decoder_num_layers .............................. None + decoder_seq_length .............................. None + decoupled_lr .................................... None + decoupled_min_lr ................................ None + decrease_batch_size_if_needed ................... False + defer_embedding_wgrad_compute ................... False + delay_wgrad_compute ............................. False + deprecated_use_mcore_models ..................... True + deterministic_mode .............................. False + dino_bottleneck_size ............................ 256 + dino_freeze_last_layer .......................... 1 + dino_head_hidden_size ........................... 2048 + dino_local_crops_number ......................... 10 + dino_local_img_size ............................. 96 + dino_norm_last_layer ............................ False + dino_teacher_temp ............................... 0.07 + dino_warmup_teacher_temp ........................ 0.04 + dino_warmup_teacher_temp_epochs ................. 30 + disable_bf16_reduced_precision_matmul ........... False + disable_mamba_mem_eff_path ...................... False + disable_straggler_on_startup .................... False + dist_ckpt_format_deprecated ..................... None + dist_ckpt_strictness ............................ assume_ok_unexpected + distribute_saved_activations .................... False + distributed_backend ............................. nccl + distributed_timeout_minutes ..................... 10 + embedding_init_method_std ....................... None + embedding_path .................................. None + empty_unused_memory_level ....................... 0 + enable_cuda_graph ............................... True + enable_experimental ............................. False + enable_ft_package ............................... False + enable_gloo_process_groups ...................... True + enable_msc ...................................... True + enable_one_logger ............................... True + encoder_num_layers .............................. 10 + encoder_seq_length .............................. 4096 + end_weight_decay ................................ 0.1 + eod_mask_loss ................................... False + error_injection_rate ............................ 0 + error_injection_type ............................ transient_error + eval_interval ................................... 1000000 + eval_iters ...................................... 0 + evidence_data_path .............................. None + exit_duration_in_mins ........................... None + exit_interval ................................... None + exit_on_missing_checkpoint ...................... False + exit_signal_handler ............................. False + exp_avg_dtype ................................... torch.float32 + exp_avg_sq_dtype ................................ torch.float32 + expert_model_parallel_size ...................... 8 + expert_tensor_parallel_size ..................... 1 + export_force_local_attention .................... False + export_kd_cfg ................................... None + export_kd_teacher_ckpt_format ................... None + export_kd_teacher_load .......................... None + export_kv_cache_quant ........................... False + export_legacy_megatron .......................... False + export_model_type ............................... GPTModel + export_moe_apply_probs_on_input ................. False + export_qk_l2_norm ............................... False + export_quant_cfg ................................ None + export_real_quant_cfg ........................... None + export_te_mcore_model ........................... False + external_cuda_graph ............................. False + ffn_hidden_size ................................. 2560 + finetune ........................................ False + finetune_data_split ............................. train + finetune_hf_dataset ............................. None + first_last_layers_bf16 .......................... False + flash_decode .................................... False + fp16 ............................................ False + fp16_lm_cross_entropy ........................... False + fp32_residual_connection ........................ False + fp8 ............................................. e4m3 + fp8_amax_compute_algo ........................... most_recent + fp8_amax_history_len ............................ 1 + fp8_interval .................................... 1 + fp8_margin ...................................... 0 + fp8_param_gather ................................ False + fp8_recipe ...................................... mxfp8 + fp8_wgrad ....................................... True + fsdp_double_buffer .............................. False + global_batch_size ............................... 8 + grad_reduce_in_bf16 ............................. False + gradient_accumulation_fusion .................... True + gradient_reduce_div_fusion ...................... True + group_query_attention ........................... True + head_lr_mult .................................... 1.0 + heterogeneous_layers_config_encoded_json ........ None + heterogeneous_layers_config_path ................ None + hidden_dropout .................................. 0.0 + hidden_size ..................................... 3584 + hierarchical_context_parallel_sizes ............. None + high_priority_stream_groups ..................... [] + hybrid_attention_ratio .......................... 0.0 + hybrid_mlp_ratio ................................ 0.0 + hybrid_override_pattern ......................... None + hysteresis ...................................... 2 + ict_head_size ................................... None + ict_load ........................................ None + img_h ........................................... 224 + img_w ........................................... 224 + indexer_batch_size .............................. 128 + indexer_log_interval ............................ 1000 + inference_batch_times_seqlen_threshold .......... -1 + inference_dynamic_batching ...................... False + inference_dynamic_batching_buffer_guaranteed_fraction 0.2 + inference_dynamic_batching_buffer_overflow_factor None + inference_dynamic_batching_buffer_size_gb ....... 40.0 + inference_dynamic_batching_chunk_size ........... 256 + inference_dynamic_batching_max_requests_override None + inference_dynamic_batching_max_tokens_override .. None + inference_dynamic_batching_num_cuda_graphs ...... 16 + inference_max_batch_size ........................ 8 + inference_max_seq_length ........................ 2560 + inference_rng_tracker ........................... False + init_method_std ................................. 0.01 + init_method_xavier_uniform ...................... False + init_model_with_meta_device ..................... False + initial_loss_scale .............................. 4294967296 + inprocess_active_world_size ..................... 8 + inprocess_barrier_timeout ....................... 120 + inprocess_completion_timeout .................... 120 + inprocess_empty_cuda_cache ...................... False + inprocess_granularity ........................... node + inprocess_hard_timeout .......................... 90 + inprocess_heartbeat_interval .................... 30 + inprocess_heartbeat_timeout ..................... 60 + inprocess_last_call_wait ........................ 1 + inprocess_max_iterations ........................ None + inprocess_monitor_process_interval .............. 1.0 + inprocess_monitor_thread_interval ............... 1.0 + inprocess_progress_watchdog_interval ............ 1.0 + inprocess_restart ............................... False + inprocess_soft_timeout .......................... 60 + inprocess_termination_grace_time ................ 1 + is_hybrid_model ................................. False + iter_per_epoch .................................. 1250 + iterations_to_skip .............................. [] + keep_fp8_transpose_cache_when_using_custom_fsdp . False + kitchen_config_file ............................. None + kitchen_recipe_number ........................... None + kv_channels ..................................... 128 + kv_lora_rank .................................... 32 + lazy_mpu_init ................................... None + load ............................................ None + load_main_params_from_ckpt ...................... None + load_model_opt_format ........................... False + local_rank ...................................... 0 + log_energy ...................................... False + log_interval .................................... 1 + log_loss_scale_to_tensorboard ................... True + log_memory_to_tensorboard ....................... False + log_num_zeros_in_grad ........................... False + log_params_norm ................................. False + log_progress .................................... False + log_straggler ................................... False + log_throughput .................................. True + log_timers_to_tensorboard ....................... False + log_validation_ppl_to_tensorboard ............... False + log_world_size_to_tensorboard ................... False + logging_level ................................... None + loss_scale ...................................... None + loss_scale_window ............................... 1000 + lr .............................................. 0.0001 + lr_decay_iters .................................. 8 + lr_decay_samples ................................ None + lr_decay_style .................................. cosine + lr_warmup_fraction .............................. None + lr_warmup_init .................................. 0.0 + lr_warmup_iters ................................. 5 + lr_warmup_samples ............................... 0 + lr_wsd_decay_iters .............................. None + lr_wsd_decay_samples ............................ None + lr_wsd_decay_style .............................. exponential + main_grads_dtype ................................ torch.float32 + main_params_dtype ............................... torch.float32 + make_vocab_size_divisible_by .................... 128 + mamba_head_dim .................................. 64 + mamba_num_groups ................................ 8 + mamba_num_heads ................................. None + mamba_state_dim ................................. 128 + manual_gc ....................................... False + manual_gc_eval .................................. True + manual_gc_interval .............................. 0 + mask_factor ..................................... 1.0 + mask_prob ....................................... 0.15 + mask_type ....................................... random + masked_softmax_fusion ........................... False + max_position_embeddings ......................... 32768 + max_tokens_to_oom ............................... 12000 + memory_snapshot_path ............................ snapshot.pickle + merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt + micro_batch_size ................................ 1 + microbatch_group_size_per_vp_stage .............. None + mid_level_dataset_surplus ....................... 0.005 + min_loss_scale .................................. 1.0 + min_lr .......................................... 1e-05 + mlp_chunks_for_prefill .......................... 1 + mmap_bin_files .................................. True + mock_data ....................................... False + moe_apply_probs_on_input ........................ False + moe_aux_loss_coeff .............................. 0.01 + moe_capacity_monitor ............................ True + moe_deepep_num_sms .............................. 20 + moe_echo_enable_random_offloading ............... False + moe_echo_expert_dispatch_overlap ................ False + moe_echo_expert_dispatcher_type ................. hybridep + moe_echo_recompute_expert_dispatch .............. False + moe_enable_deepep ............................... False + moe_enable_echo ................................. False + moe_enable_hybridep ............................. True + moe_enable_ultraep .............................. True + moe_expert_capacity_factor ...................... None + moe_extended_tp ................................. False + moe_ffn_hidden_size ............................. 2560 + moe_grouped_gemm ................................ True + moe_hybridep_num_sms ............................ 16 + moe_input_jitter_eps ............................ None + moe_layer_freq .................................. 1 + moe_layer_recompute ............................. False + moe_num_echo_experts ............................ None + moe_num_redundant_experts_per_rank .............. 8 + moe_pad_expert_input_to_capacity ................ False + moe_per_layer_logging ........................... False + moe_permute_fusion .............................. False + moe_received_token_capacity ..................... 1.08984375 + moe_router_bias_update_rate ..................... 0.001 + moe_router_dtype ................................ fp32 + moe_router_enable_expert_bias ................... False + moe_router_force_load_balancing ................. False + moe_router_group_topk ........................... None + moe_router_load_balancing_type .................. aux_loss + moe_router_num_groups ........................... None + moe_router_padding_for_fp8 ...................... False + moe_router_pre_softmax .......................... True + moe_router_score_function ....................... softmax + moe_router_topk ................................. 8 + moe_router_topk_scaling_factor .................. None + moe_shared_expert_intermediate_size ............. 20480 + moe_shared_expert_overlap ....................... False + moe_token_dispatcher_type ....................... flex + moe_token_drop_policy ........................... probs + moe_ultraep_max_microbatches .................... None + moe_ultraep_overlap_weight_sync ................. True + moe_ultraep_weight_sync_join .................... before_dispatch + moe_upcycling_granularity ....................... 1 + moe_use_device_initiated_grouped_gemm ........... True + moe_use_legacy_grouped_gemm ..................... False + moe_use_upcycling ............................... False + moe_z_loss_coeff ................................ None + mrope_section ................................... None + mscale .......................................... 1.0 + mscale_all_dim .................................. 0.0 + mtp_loss_scaling_factor ......................... 0.1 + mtp_num_layers .................................. None + multi_latent_attention .......................... False + nccl_all_reduce_for_prefill ..................... False + nccl_communicator_config_path ................... None + nccl_ub ......................................... False + no_load_optim ................................... None + no_load_rng ..................................... None + no_persist_layer_norm ........................... False + no_rope_freq .................................... None + no_save_optim ................................... None + no_save_rng ..................................... None + non_persistent_ckpt_type ........................ None + non_persistent_global_ckpt_dir .................. None + non_persistent_local_ckpt_algo .................. fully_parallel + non_persistent_local_ckpt_dir ................... None + non_persistent_save_interval .................... None + norm_epsilon .................................... 1e-06 + normalization ................................... RMSNorm + num_attention_heads ............................. 28 + num_channels .................................... 3 + num_classes ..................................... 1000 + num_dataset_builder_threads ..................... 1 + num_distributed_optimizer_instances ............. 1 + num_experts ..................................... 64 + num_layers ...................................... 10 + num_layers_at_end_in_bf16 ....................... 1 + num_layers_at_start_in_bf16 ..................... 1 + num_layers_per_virtual_pipeline_stage ........... None + num_query_groups ................................ 4 + num_virtual_stages_per_pipeline_rank ............ None + num_workers ..................................... 2 + object_storage_cache_path ....................... None + one_logger_async ................................ False + one_logger_project .............................. megatron-lm + one_logger_run_name ............................. None + onnx_safe ....................................... None + openai_gelu ..................................... False + optimizer ....................................... adam + optimizer_cpu_offload ........................... False + optimizer_offload_fraction ...................... 1.0 + output_bert_embeddings .......................... False + overlap_cpu_optimizer_d2h_h2d ................... False + overlap_grad_reduce ............................. True + overlap_moe_expert_parallel_comm ................ False + overlap_p2p_comm ................................ False + overlap_p2p_comm_warmup_flush ................... False + overlap_param_gather ............................ True + overlap_param_gather_with_optimizer_step ........ False + override_opt_param_scheduler .................... False + params_dtype .................................... torch.bfloat16 + patch_dim ....................................... 16 + per_split_data_args_path ........................ None + perform_initialization .......................... True + pin_cpu_grads ................................... True + pin_cpu_params .................................. True + pipeline_model_parallel_comm_backend ............ None + pipeline_model_parallel_layout .................. None + pipeline_model_parallel_size .................... 1 + position_embedding_type ......................... rope + pretrained_checkpoint ........................... None + profile ......................................... False + profile_ranks ................................... [0] + profile_step_end ................................ 12 + profile_step_start .............................. 10 + 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 ..................................... 12 + train_samples ................................... None + train_sync_interval ............................. None + transformer_impl ................................ transformer_engine + transformer_pipeline_model_parallel_size ........ 1 + untie_embeddings_and_output_weights ............. True + use_checkpoint_args ............................. False + use_checkpoint_opt_param_scheduler .............. False + use_cpu_initialization .......................... None + use_custom_fsdp ................................. False + use_dist_ckpt ................................... True + use_dist_ckpt_deprecated ........................ False + use_distributed_optimizer ....................... True + use_flash_attn .................................. False + use_legacy_models ............................... False + use_mp_args_from_checkpoint_args ................ False + use_one_sent_docs ............................... False + use_persistent_ckpt_worker ...................... False + use_precision_aware_optimizer ................... False + use_pytorch_profiler ............................ False + use_ring_exchange_p2p ........................... False + use_rope_scaling ................................ False + use_rotary_position_embeddings .................. False + use_sharp ....................................... False + use_tokenizer_model_from_checkpoint_args ........ True + use_torch_fsdp2 ................................. False + use_torch_optimizer_for_cpu_offload ............. False + use_tp_pp_dp_mapping ............................ False + v_head_dim ...................................... 128 + valid_data_path ................................. None + variable_seq_lengths ............................ False + virtual_pipeline_model_parallel_size ............ None + vision_backbone_type ............................ vit + vision_pretraining .............................. False + vision_pretraining_type ......................... classify + vocab_extra_ids ................................. 0 + vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json + vocab_size ...................................... None + wandb_exp_name .................................. + wandb_project ................................... + wandb_save_dir .................................. + weight_decay .................................... 0.1 + weight_decay_incr_style ......................... constant + wgrad_deferral_limit ............................ 0 + world_size ...................................... 8 + yaml_cfg ........................................ None +-------------------- end of arguments --------------------- +INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1 +> building GPT2BPETokenizer tokenizer ... + > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304) +WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS +> initializing torch distributed ... +[Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank [Gloo] Rank 7 is connected to 76 peer ranks. is connected to Expected number of connected peer ranks is : 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 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 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 [Gloo] Rank [Gloo] Rank 567 is connected to is connected to is connected to 777 peer ranks. peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : Expected number of connected peer ranks is : 77 +7 + +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +> initialized tensor model parallel with size 1 +> initialized pipeline model parallel with size 1 +> setting random seeds to 1234 ... +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank [Gloo] Rank 00 is connected to is connected to 00 peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : 00 + +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[rank1]:[W826 18:39:02.758192730 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. +[rank2]:[W826 18:39:02.758218559 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]:[W826 18:39:02.758224235 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]:[W826 18:39:02.758264494 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. +[rank3]:[W826 18:39:02.758263509 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. +[rank5]:[W826 18:39:02.758284080 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]:[W826 18:39:02.758367472 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: 0.036 seconds +WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations. +> compiling and loading fused kernels ... +[rank0]:[W826 18:39:02.928009101 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: 3.911 seconds +time to initialize megatron (seconds): 7.130 +[after megatron is initialized] datetime: 2026-08-26 18:39:08 +building GPT model ... +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect. + warnings.warn( +[INFO] Use MNNVL fabric: False +[WARN] MAX_NUM_NVL_PEERS is not set. Using detected value 8. +NVSHMEM v3.7.2 +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +/dvs/p4/build/sw/rel/gpgpu/toolkit/r12.9/main_nvshmem/src/modules/transport/ibrc/ibrc.cpp:1754: error status: 7 (NVSHMEMX_ERROR_INTERNAL) Device enumeration failed. + +UltraEP: NVL domain size <= 8; forcing weight_sync plan mode to direct (was adaptive). + > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 7262624256 +INFO:megatron.core.distributed.distributed_data_parallel:Setting up DistributedDataParallel with config DistributedDataParallelConfig(grad_reduce_in_fp32=True, overlap_grad_reduce=True, overlap_param_gather=True, align_param_gather=False, use_distributed_optimizer=True, num_distributed_optimizer_instances=1, check_for_nan_in_grad=False, check_for_large_grads=False, bucket_size=40000000, pad_buckets_for_high_nccl_busbw=False, average_in_collective=False, fp8_param_gather=False, reuse_grad_buf_for_mxfp8_param_ag=False, use_custom_fsdp=False, data_parallel_sharding_strategy='no_shard', gradient_reduce_div_fusion=True, suggested_communication_unit_size=None, preserve_fp32_weights=True, keep_fp8_transpose_cache_when_using_custom_fsdp=False, nccl_ub=False, fsdp_double_buffer=False) +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 22 +Params for bucket 1 (180289536 elements, 180289536 padded size): + module.output_layer.weight +Params for bucket 2 (73403904 elements, 73403904 padded size): + module.decoder.final_layernorm.weight + module.decoder.layers.9.mlp.shared_experts.linear_fc2.weight +Params for bucket 3 (146800640 elements, 146800640 padded size): + module.decoder.layers.9.mlp.shared_experts.linear_fc1.weight +Params for bucket 4 (103001600 elements, 103001600 padded size): + module.decoder.layers.9.mlp.router.weight + module.decoder.layers.9.self_attention.linear_qkv.weight + module.decoder.layers.9.self_attention.linear_proj.weight + module.decoder.layers.8.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.9.pre_mlp_layernorm.weight + module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.9.self_attention.linear_qkv.bias +Params for bucket 5 (146800640 elements, 146800640 padded size): + module.decoder.layers.8.mlp.shared_experts.linear_fc1.weight +Params for bucket 6 (103001600 elements, 103001600 padded size): + module.decoder.layers.8.mlp.router.weight + module.decoder.layers.8.self_attention.linear_qkv.weight + module.decoder.layers.8.self_attention.linear_proj.weight + module.decoder.layers.7.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.8.pre_mlp_layernorm.weight + module.decoder.layers.8.self_attention.linear_qkv.bias + module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight +Params for bucket 7 (146800640 elements, 146800640 padded size): + module.decoder.layers.7.mlp.shared_experts.linear_fc1.weight +Params for bucket 8 (103001600 elements, 103001600 padded size): + module.decoder.layers.7.mlp.router.weight + module.decoder.layers.7.self_attention.linear_qkv.weight + module.decoder.layers.7.self_attention.linear_proj.weight + module.decoder.layers.6.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.7.pre_mlp_layernorm.weight + module.decoder.layers.7.self_attention.linear_qkv.bias +Params for bucket 9 (146800640 elements, 146800640 padded size): + module.decoder.layers.6.mlp.shared_experts.linear_fc1.weight +Params for bucket 10 (103001600 elements, 103001600 padded size): + module.decoder.layers.6.mlp.router.weight + module.decoder.layers.6.self_attention.linear_qkv.weight + module.decoder.layers.6.self_attention.linear_proj.weight + module.decoder.layers.5.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.6.self_attention.linear_qkv.bias + module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.6.pre_mlp_layernorm.weight +Params for bucket 11 (146800640 elements, 146800640 padded size): + module.decoder.layers.5.mlp.shared_experts.linear_fc1.weight +Params for bucket 12 (103001600 elements, 103001600 padded size): + module.decoder.layers.5.mlp.router.weight + module.decoder.layers.5.self_attention.linear_qkv.weight + module.decoder.layers.5.self_attention.linear_proj.weight + module.decoder.layers.4.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.5.pre_mlp_layernorm.weight + module.decoder.layers.5.self_attention.linear_qkv.bias +Params for bucket 13 (146800640 elements, 146800640 padded size): + module.decoder.layers.4.mlp.shared_experts.linear_fc1.weight +Params for bucket 14 (103001600 elements, 103001600 padded size): + module.decoder.layers.4.mlp.router.weight + module.decoder.layers.4.self_attention.linear_qkv.weight + module.decoder.layers.4.self_attention.linear_proj.weight + module.decoder.layers.3.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.4.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.4.self_attention.linear_qkv.bias + module.decoder.layers.4.pre_mlp_layernorm.weight +Params for bucket 15 (146800640 elements, 146800640 padded size): + module.decoder.layers.3.mlp.shared_experts.linear_fc1.weight +Params for bucket 16 (103001600 elements, 103001600 padded size): + module.decoder.layers.3.mlp.router.weight + module.decoder.layers.3.self_attention.linear_qkv.weight + module.decoder.layers.3.self_attention.linear_proj.weight + module.decoder.layers.2.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.3.self_attention.linear_qkv.bias + module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.3.pre_mlp_layernorm.weight +Params for bucket 17 (146800640 elements, 146800640 padded size): + module.decoder.layers.2.mlp.shared_experts.linear_fc1.weight +Params for bucket 18 (103001600 elements, 103001600 padded size): + module.decoder.layers.2.mlp.router.weight + module.decoder.layers.2.self_attention.linear_qkv.weight + module.decoder.layers.2.self_attention.linear_proj.weight + module.decoder.layers.1.mlp.shared_experts.linear_fc2.weight + module.decoder.layers.2.pre_mlp_layernorm.weight + module.decoder.layers.2.self_attention.linear_qkv.bias + module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight +Params for bucket 19 (146800640 elements, 146800640 padded size): + module.decoder.layers.1.mlp.shared_experts.linear_fc1.weight +Params for bucket 20 (103001600 elements, 103001600 padded size): + module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.1.self_attention.linear_qkv.bias + module.decoder.layers.1.pre_mlp_layernorm.weight + module.decoder.layers.1.mlp.router.weight + module.decoder.layers.1.self_attention.linear_qkv.weight + module.decoder.layers.1.self_attention.linear_proj.weight + module.decoder.layers.0.mlp.shared_experts.linear_fc2.weight +Params for bucket 21 (146800640 elements, 146800640 padded size): + module.decoder.layers.0.mlp.shared_experts.linear_fc1.weight +Params for bucket 22 (209890816 elements, 209890816 padded size): + module.decoder.layers.0.self_attention.linear_qkv.bias + module.decoder.layers.0.mlp.router.weight + 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.self_attention.linear_proj.weight + module.embedding.word_embeddings.weight +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 44 +Params for bucket 1 (45875200 elements, 45875200 padded size): + module.decoder.layers.9.mlp.experts.linear_fc2.weight7 + module.decoder.layers.9.mlp.experts.linear_fc2.weight5 + module.decoder.layers.9.mlp.experts.linear_fc2.weight3 + module.decoder.layers.9.mlp.experts.linear_fc2.weight4 + module.decoder.layers.9.mlp.experts.linear_fc2.weight6 +Params for bucket 2 (45875200 elements, 45875200 padded size): + module.decoder.layers.9.mlp.experts.linear_fc2.weight2 + module.decoder.layers.9.mlp.experts.linear_fc1.weight7 + module.decoder.layers.9.mlp.experts.linear_fc2.weight1 + module.decoder.layers.9.mlp.experts.linear_fc2.weight0 +Params for bucket 3 (55050240 elements, 55050240 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight6 + module.decoder.layers.9.mlp.experts.linear_fc1.weight4 + module.decoder.layers.9.mlp.experts.linear_fc1.weight5 +Params for bucket 4 (55050240 elements, 55050240 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight3 + module.decoder.layers.9.mlp.experts.linear_fc1.weight1 + module.decoder.layers.9.mlp.experts.linear_fc1.weight2 +Params for bucket 5 (45875200 elements, 45875200 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight0 + module.decoder.layers.8.mlp.experts.linear_fc2.weight6 + module.decoder.layers.8.mlp.experts.linear_fc2.weight7 + module.decoder.layers.8.mlp.experts.linear_fc2.weight5 +Params for bucket 6 (45875200 elements, 45875200 padded size): + module.decoder.layers.8.mlp.experts.linear_fc2.weight3 + module.decoder.layers.8.mlp.experts.linear_fc2.weight1 + module.decoder.layers.8.mlp.experts.linear_fc2.weight4 + module.decoder.layers.8.mlp.experts.linear_fc2.weight0 + module.decoder.layers.8.mlp.experts.linear_fc2.weight2 +Params for bucket 7 (55050240 elements, 55050240 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight7 + module.decoder.layers.8.mlp.experts.linear_fc1.weight6 + module.decoder.layers.8.mlp.experts.linear_fc1.weight5 +Params for bucket 8 (55050240 elements, 55050240 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight4 + module.decoder.layers.8.mlp.experts.linear_fc1.weight3 + module.decoder.layers.8.mlp.experts.linear_fc1.weight2 +Params for bucket 9 (45875200 elements, 45875200 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight1 + module.decoder.layers.8.mlp.experts.linear_fc1.weight0 + module.decoder.layers.7.mlp.experts.linear_fc2.weight7 +Params for bucket 10 (45875200 elements, 45875200 padded size): + module.decoder.layers.7.mlp.experts.linear_fc2.weight5 + module.decoder.layers.7.mlp.experts.linear_fc2.weight3 + module.decoder.layers.7.mlp.experts.linear_fc2.weight6 + module.decoder.layers.7.mlp.experts.linear_fc2.weight2 + module.decoder.layers.7.mlp.experts.linear_fc2.weight4 +Params for bucket 11 (55050240 elements, 55050240 padded size): + module.decoder.layers.7.mlp.experts.linear_fc2.weight1 + module.decoder.layers.7.mlp.experts.linear_fc1.weight6 + module.decoder.layers.7.mlp.experts.linear_fc2.weight0 + module.decoder.layers.7.mlp.experts.linear_fc1.weight7 +Params for bucket 12 (55050240 elements, 55050240 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight5 + module.decoder.layers.7.mlp.experts.linear_fc1.weight4 + module.decoder.layers.7.mlp.experts.linear_fc1.weight3 +Params for bucket 13 (55050240 elements, 55050240 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight2 + module.decoder.layers.7.mlp.experts.linear_fc1.weight1 + module.decoder.layers.7.mlp.experts.linear_fc1.weight0 +Params for bucket 14 (45875200 elements, 45875200 padded size): + module.decoder.layers.6.mlp.experts.linear_fc2.weight7 + module.decoder.layers.6.mlp.experts.linear_fc2.weight3 + module.decoder.layers.6.mlp.experts.linear_fc2.weight5 + module.decoder.layers.6.mlp.experts.linear_fc2.weight6 + module.decoder.layers.6.mlp.experts.linear_fc2.weight4 +Params for bucket 15 (45875200 elements, 45875200 padded size): + module.decoder.layers.6.mlp.experts.linear_fc2.weight2 + module.decoder.layers.6.mlp.experts.linear_fc2.weight0 + module.decoder.layers.6.mlp.experts.linear_fc1.weight7 + module.decoder.layers.6.mlp.experts.linear_fc2.weight1 +Params for bucket 16 (55050240 elements, 55050240 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight6 + module.decoder.layers.6.mlp.experts.linear_fc1.weight4 + module.decoder.layers.6.mlp.experts.linear_fc1.weight5 +Params for bucket 17 (55050240 elements, 55050240 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight3 + module.decoder.layers.6.mlp.experts.linear_fc1.weight2 + module.decoder.layers.6.mlp.experts.linear_fc1.weight1 +Params for bucket 18 (45875200 elements, 45875200 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight0 + module.decoder.layers.5.mlp.experts.linear_fc2.weight5 + module.decoder.layers.5.mlp.experts.linear_fc2.weight6 + module.decoder.layers.5.mlp.experts.linear_fc2.weight7 +Params for bucket 19 (45875200 elements, 45875200 padded size): + module.decoder.layers.5.mlp.experts.linear_fc2.weight1 + module.decoder.layers.5.mlp.experts.linear_fc2.weight3 + module.decoder.layers.5.mlp.experts.linear_fc2.weight4 + module.decoder.layers.5.mlp.experts.linear_fc2.weight2 + module.decoder.layers.5.mlp.experts.linear_fc2.weight0 +Params for bucket 20 (55050240 elements, 55050240 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight7 + module.decoder.layers.5.mlp.experts.linear_fc1.weight5 + module.decoder.layers.5.mlp.experts.linear_fc1.weight6 +Params for bucket 21 (55050240 elements, 55050240 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight4 + module.decoder.layers.5.mlp.experts.linear_fc1.weight2 + module.decoder.layers.5.mlp.experts.linear_fc1.weight3 +Params for bucket 22 (45875200 elements, 45875200 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight1 + module.decoder.layers.4.mlp.experts.linear_fc2.weight7 + module.decoder.layers.5.mlp.experts.linear_fc1.weight0 +Params for bucket 23 (45875200 elements, 45875200 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight3 + module.decoder.layers.4.mlp.experts.linear_fc2.weight5 + module.decoder.layers.4.mlp.experts.linear_fc2.weight6 + module.decoder.layers.4.mlp.experts.linear_fc2.weight4 + module.decoder.layers.4.mlp.experts.linear_fc2.weight2 +Params for bucket 24 (55050240 elements, 55050240 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight1 + module.decoder.layers.4.mlp.experts.linear_fc2.weight0 + module.decoder.layers.4.mlp.experts.linear_fc1.weight6 + module.decoder.layers.4.mlp.experts.linear_fc1.weight7 +Params for bucket 25 (55050240 elements, 55050240 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight5 + module.decoder.layers.4.mlp.experts.linear_fc1.weight4 + module.decoder.layers.4.mlp.experts.linear_fc1.weight3 +Params for bucket 26 (55050240 elements, 55050240 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight2 + module.decoder.layers.4.mlp.experts.linear_fc1.weight0 + module.decoder.layers.4.mlp.experts.linear_fc1.weight1 +Params for bucket 27 (45875200 elements, 45875200 padded size): + module.decoder.layers.3.mlp.experts.linear_fc2.weight6 + module.decoder.layers.3.mlp.experts.linear_fc2.weight4 + module.decoder.layers.3.mlp.experts.linear_fc2.weight7 + module.decoder.layers.3.mlp.experts.linear_fc2.weight3 + module.decoder.layers.3.mlp.experts.linear_fc2.weight5 +Params for bucket 28 (45875200 elements, 45875200 padded size): + module.decoder.layers.3.mlp.experts.linear_fc2.weight2 + module.decoder.layers.3.mlp.experts.linear_fc1.weight7 + module.decoder.layers.3.mlp.experts.linear_fc2.weight0 + module.decoder.layers.3.mlp.experts.linear_fc2.weight1 +Params for bucket 29 (55050240 elements, 55050240 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight6 + module.decoder.layers.3.mlp.experts.linear_fc1.weight4 + module.decoder.layers.3.mlp.experts.linear_fc1.weight5 +Params for bucket 30 (55050240 elements, 55050240 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight3 + module.decoder.layers.3.mlp.experts.linear_fc1.weight1 + module.decoder.layers.3.mlp.experts.linear_fc1.weight2 +Params for bucket 31 (45875200 elements, 45875200 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight0 + module.decoder.layers.2.mlp.experts.linear_fc2.weight6 + module.decoder.layers.2.mlp.experts.linear_fc2.weight7 + module.decoder.layers.2.mlp.experts.linear_fc2.weight5 +Params for bucket 32 (45875200 elements, 45875200 padded size): + module.decoder.layers.2.mlp.experts.linear_fc2.weight3 + module.decoder.layers.2.mlp.experts.linear_fc2.weight1 + module.decoder.layers.2.mlp.experts.linear_fc2.weight4 + module.decoder.layers.2.mlp.experts.linear_fc2.weight0 + module.decoder.layers.2.mlp.experts.linear_fc2.weight2 +Params for bucket 33 (55050240 elements, 55050240 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight7 + module.decoder.layers.2.mlp.experts.linear_fc1.weight6 + module.decoder.layers.2.mlp.experts.linear_fc1.weight5 +Params for bucket 34 (55050240 elements, 55050240 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight2 + module.decoder.layers.2.mlp.experts.linear_fc1.weight3 +Params for bucket 35 (45875200 elements, 45875200 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight1 + module.decoder.layers.1.mlp.experts.linear_fc2.weight7 + module.decoder.layers.2.mlp.experts.linear_fc1.weight0 +Params for bucket 36 (45875200 elements, 45875200 padded size): + module.decoder.layers.1.mlp.experts.linear_fc2.weight6 + module.decoder.layers.1.mlp.experts.linear_fc2.weight2 + module.decoder.layers.1.mlp.experts.linear_fc2.weight4 + module.decoder.layers.1.mlp.experts.linear_fc2.weight3 + module.decoder.layers.1.mlp.experts.linear_fc2.weight5 +Params for bucket 37 (55050240 elements, 55050240 padded size): + module.decoder.layers.1.mlp.experts.linear_fc2.weight1 + module.decoder.layers.1.mlp.experts.linear_fc1.weight6 + module.decoder.layers.1.mlp.experts.linear_fc2.weight0 + module.decoder.layers.1.mlp.experts.linear_fc1.weight7 +Params for bucket 38 (55050240 elements, 55050240 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight5 + module.decoder.layers.1.mlp.experts.linear_fc1.weight3 + module.decoder.layers.1.mlp.experts.linear_fc1.weight4 +Params for bucket 39 (55050240 elements, 55050240 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight0 + module.decoder.layers.1.mlp.experts.linear_fc1.weight1 +Params for bucket 40 (45875200 elements, 45875200 padded size): + module.decoder.layers.0.mlp.experts.linear_fc2.weight7 + module.decoder.layers.0.mlp.experts.linear_fc2.weight3 + module.decoder.layers.0.mlp.experts.linear_fc2.weight5 + module.decoder.layers.0.mlp.experts.linear_fc2.weight4 + module.decoder.layers.0.mlp.experts.linear_fc2.weight6 +Params for bucket 41 (45875200 elements, 45875200 padded size): + module.decoder.layers.0.mlp.experts.linear_fc2.weight2 + module.decoder.layers.0.mlp.experts.linear_fc2.weight1 + module.decoder.layers.0.mlp.experts.linear_fc2.weight0 + module.decoder.layers.0.mlp.experts.linear_fc1.weight7 +Params for bucket 42 (55050240 elements, 55050240 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight6 + module.decoder.layers.0.mlp.experts.linear_fc1.weight5 + module.decoder.layers.0.mlp.experts.linear_fc1.weight4 +Params for bucket 43 (55050240 elements, 55050240 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight3 + module.decoder.layers.0.mlp.experts.linear_fc1.weight2 + module.decoder.layers.0.mlp.experts.linear_fc1.weight1 +Params for bucket 44 (18350080 elements, 18350080 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight0 +INFO:megatron.core.optimizer:Setting up optimizer with config OptimizerConfig(optimizer='adam', lr=0.0001, min_lr=1e-05, decoupled_lr=None, decoupled_min_lr=None, weight_decay=0.1, fp8_recipe='mxfp8', fp16=False, bf16=True, reuse_grad_buf_for_mxfp8_param_ag=False, params_dtype=torch.bfloat16, use_precision_aware_optimizer=False, store_param_remainders=True, main_grads_dtype=torch.float32, main_params_dtype=torch.float32, exp_avg_dtype=torch.float32, exp_avg_sq_dtype=torch.float32, loss_scale=None, initial_loss_scale=4294967296, min_loss_scale=1.0, loss_scale_window=1000, hysteresis=2, adam_beta1=0.9, adam_beta2=0.999, adam_eps=1e-08, sgd_momentum=0.9, use_distributed_optimizer=True, overlap_param_gather_with_optimizer_step=False, optimizer_cpu_offload=False, optimizer_offload_fraction=1.0, use_torch_optimizer_for_cpu_offload=False, overlap_cpu_optimizer_d2h_h2d=False, pin_cpu_grads=True, pin_cpu_params=True, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=, config_logger_dir='') +INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine +[after model, optimizer, and learning rate scheduler are built] datetime: 2026-08-26 18:39:12 +> building train, validation, and test datasets ... + > datasets target sizes (minimum size): + train: 96 + 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=(96, 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 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-document_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-sample_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-shuffle_index.npy +INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073 +> finished creating GPT datasets ... +[after dataloaders are built] datetime: 2026-08-26 18:39:12 +done with setup ... +(min, max) time across ranks (ms): + model-and-optimizer-setup ......................: (3467.51, 3475.61) + train/valid/test-data-iterators-setup ..........: (71.25, 91.10) +training ... +Overwriting rerun_state_machine.current_iteration from -1 to 0... +[before the start of training step] datetime: 2026-08-26 18:39:12 +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") + [2026-08-26 18:39:37] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 25184.3 | throughput per GPU (TFLOP/s/GPU): 4.9 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.097922E+01 | loss scale: 1.0 | grad norm: 29.878 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33664 | min safe capacity: 1.0273 | +Number of parameters in transformer block in billions: 20.11 +Number of parameters in embedding layers in billions: 0.36 +Total number of parameters in billions: 20.47 +Number of parameters in most loaded shard in billions: 20.4724 +Theoretical memory footprints: weight and optimizer=146430.14 MB +[Rank 0] (after 1 iterations) memory (MB) | allocated: 64278.42138671875 | max allocated: 65496.45947265625 | reserved: 73306.0 | max reserved: 73306.0 +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") + [2026-08-26 18:39:37] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 329.5 | throughput per GPU (TFLOP/s/GPU): 374.8 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.097874E+01 | loss scale: 1.0 | grad norm: 30.618 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33792 | min safe capacity: 1.0312 | + [2026-08-26 18:39:38] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 215.4 | throughput per GPU (TFLOP/s/GPU): 573.3 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 9.539364E+00 | loss scale: 1.0 | grad norm: 32.415 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33792 | min safe capacity: 1.0312 | +Capture CUDA graph for training!!! +CUDA graph capture done!!! + [2026-08-26 18:39:39] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 1731.4 | throughput per GPU (TFLOP/s/GPU): 71.3 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 1.109718E+01 | loss scale: 1.0 | grad norm: 107.119 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33920 | min safe capacity: 1.0352 | + [2026-08-26 18:39:40] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 198.9 | throughput per GPU (TFLOP/s/GPU): 620.7 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.000232E+01 | loss scale: 1.0 | grad norm: 26.383 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34048 | min safe capacity: 1.0391 | + [2026-08-26 18:39:40] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 192.7 | throughput per GPU (TFLOP/s/GPU): 640.8 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 1.049848E+01 | loss scale: 1.0 | grad norm: 8.728 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34048 | min safe capacity: 1.0391 | + [2026-08-26 18:39:40] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 184.4 | throughput per GPU (TFLOP/s/GPU): 669.7 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 9.443075E+00 | loss scale: 1.0 | grad norm: 5.409 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34048 | min safe capacity: 1.0391 | + [2026-08-26 18:39:40] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 190.9 | throughput per GPU (TFLOP/s/GPU): 646.7 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.965197E+00 | loss scale: 1.0 | grad norm: 4.552 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34048 | min safe capacity: 1.0391 | + [2026-08-26 18:39:40] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 193.1 | throughput per GPU (TFLOP/s/GPU): 639.4 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.749504E+00 | loss scale: 1.0 | grad norm: 4.299 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34048 | min safe capacity: 1.0391 | + [2026-08-26 18:39:41] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 197.5 | throughput per GPU (TFLOP/s/GPU): 625.3 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.689064E+00 | loss scale: 1.0 | grad norm: 4.427 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34048 | min safe capacity: 1.0391 | + [2026-08-26 18:39:41] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 190.9 | throughput per GPU (TFLOP/s/GPU): 646.8 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.628358E+00 | loss scale: 1.0 | grad norm: 4.348 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34048 | min safe capacity: 1.0391 | + [2026-08-26 18:39:41] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 199.0 | throughput per GPU (TFLOP/s/GPU): 620.6 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.552548E+00 | loss scale: 1.0 | grad norm: 4.285 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34048 | min safe capacity: 1.0391 | +[after training is done] datetime: 2026-08-26 18:39:41 +[rank0]:[W826 18:39:42.511122045 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +W0826 18:39:43.309000 538192 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers +W0826 18:39:43.310000 538192 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 538278 closing signal SIGTERM +W0826 18:39:43.310000 538192 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 538279 closing signal SIGTERM +W0826 18:39:43.316000 538192 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 538280 closing signal SIGTERM +W0826 18:39:43.318000 538192 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 538281 closing signal SIGTERM +W0826 18:39:43.319000 538192 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 538282 closing signal SIGTERM +W0826 18:39:43.323000 538192 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 538283 closing signal SIGTERM +W0826 18:39:43.323000 538192 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 538284 closing signal SIGTERM +W0826 18:39:43.327000 538192 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 538285 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 538192 got signal: 15 diff --git a/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep2/mem_samples.csv b/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep2/mem_samples.csv new file mode 100644 index 0000000000000000000000000000000000000000..c9c7b5893cb73d190ed97a2f9a544df263ab443c --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep2/mem_samples.csv @@ -0,0 +1,24 @@ +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 1 +1, 1 +2, 1 +3, 1 +4, 1 +5, 1 +6, 1 +7, 1 diff --git a/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep2/run.env b/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep2/run.env new file mode 100644 index 0000000000000000000000000000000000000000..f42d1791470a05fc183c10bb34677f755bbb4e2c --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep2/run.env @@ -0,0 +1,61 @@ +backend=ultraep +config=QWEN2_57B +family=tierB +repo=/workspace/Megatron-UltraEP +branch=ultraep-bracing-regions-through-kernels +commit=03d57ef7 +num_layers=10 +hidden=3584 +ffn_hidden=2560 +moe_ffn_hidden=2560 +num_experts=64 +topk=8 +experts_per_rank=8 +expert_params=27525120 +heads=28 +kv_groups=4 +kv_channels= +qk_layernorm=0 +shared_ffn=20480 +shared_expert_overlap=0 +mla=0 +dense_first_layers=0 +moe_layers=10 +rotary_base=1000000 +rotary_scaling= +qkv_bias=1 +router_balancing=aux_loss +aux_loss_coeff=1e-2 +router_pre_softmax=1 +router_fusion=0 +router_dtype=fp32 +device_initiated_grouped_gemm=1 +norm_epsilon=1e-6 +seq_length=4096 +micro_batch_size=1 +global_batch_size=8 +num_microbatches=1 +assignments_per_rank=32768 +mean_tokens_per_expert=4096 +slot_policy=matched +slots_per_rank=8 +capacity_mode=static +capacity=1.08984375 +cuda_graph=1 +probe_mode=0 +async_grad_reduce=0 +branch_role=timing +memory_dump_mode=none +memory_snapshot_path= +cuda_graph_warmup_steps=3 +recompute=none +force_balance=0 +train_iters=12 +ep_size=8 +echo_dispatch_overlap=1 +echo_dispatch_recompute=0 +ultraep_quota_min_tokens=512 +ultraep_weight_sync_join= +profile_enable=0 +kernel_trace=0 +fp8_recipe=mxfp8 diff --git a/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep2/train.log b/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep2/train.log new file mode 100644 index 0000000000000000000000000000000000000000..9852faa5fd14eaaae208b238985765ed2b82a3aa --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep2/train.log @@ -0,0 +1,12 @@ +============================================================== + QWEN2_57B / ultraep (slots/rank=8 policy=matched) + E=64 K=8 H'=2560 m=8 + L=10 seq=4096 gbs=8 (1 microbatches) + cap_mode=static cap=1.08984375 cg=1 recompute=none + memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep2/memory_snapshots/snapshot.pickle + repo=/workspace/Megatron-UltraEP @ 03d57ef7 +============================================================== +W0826 18:39:49.837000 541113 torch/distributed/run.py:766] +W0826 18:39:49.837000 541113 torch/distributed/run.py:766] ***************************************** +W0826 18:39:49.837000 541113 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0826 18:39:49.837000 541113 torch/distributed/run.py:766] ***************************************** diff --git a/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep2/mem_samples.csv b/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep2/mem_samples.csv index a908f71af305e0df0d6a071e4dca61bfa8666726..685c001dd78ebae9801c63f8950de95d4f5bc983 100644 --- a/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep2/mem_samples.csv +++ b/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep2/mem_samples.csv @@ -174,3 +174,59 @@ 5, 121231 6, 121231 7, 121231 +0, 93203 +1, 93173 +2, 94093 +3, 94115 +4, 94319 +5, 106529 +6, 96785 +7, 96701 +0, 121957 +1, 121963 +2, 121963 +3, 121963 +4, 122229 +5, 121963 +6, 121963 +7, 121963 +0, 121957 +1, 121963 +2, 121963 +3, 121963 +4, 122229 +5, 121963 +6, 121963 +7, 121963 +0, 116655 +1, 116341 +2, 116405 +3, 116987 +4, 116865 +5, 116725 +6, 116987 +7, 116987 +0, 115233 +1, 114919 +2, 114983 +3, 115565 +4, 115243 +5, 114599 +6, 114983 +7, 115565 +0, 4 +1, 4 +2, 4 +3, 113885 +4, 4 +5, 4 +6, 4 +7, 113885 +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 diff --git a/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep2/train.log b/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep2/train.log index 6dc0afad65c0b4ab9506bee3a9f33d7df5e5c5ca..7729b887593db3ce90a47e6d2c58f9dc7813abd4 100644 --- a/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep2/train.log +++ b/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep2/train.log @@ -1396,3 +1396,53 @@ Theoretical memory footprints: weight and optimizer=216428.96 MB [2026-08-26 18:29:45] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 855.4 | throughput per GPU (TFLOP/s/GPU): 211.4 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.117607E+01 | loss scale: 1.0 | grad norm: 44.342 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67712 | min safe capacity: 1.0332 | [2026-08-26 18:29:45] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 309.0 | throughput per GPU (TFLOP/s/GPU): 585.2 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 9.686025E+00 | loss scale: 1.0 | grad norm: 62.884 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67968 | min safe capacity: 1.0371 | Capture CUDA graph for training!!! +CUDA graph capture done!!! + [2026-08-26 18:29:48] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 2799.6 | throughput per GPU (TFLOP/s/GPU): 64.6 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 1.338096E+01 | loss scale: 1.0 | grad norm: 20.302 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67968 | min safe capacity: 1.0371 | + [2026-08-26 18:29:48] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 299.9 | throughput per GPU (TFLOP/s/GPU): 603.0 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.049768E+01 | loss scale: 1.0 | grad norm: 95.165 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67968 | min safe capacity: 1.0371 | + [2026-08-26 18:29:49] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 299.0 | throughput per GPU (TFLOP/s/GPU): 604.9 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 1.044046E+01 | loss scale: 1.0 | grad norm: 51.769 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67968 | min safe capacity: 1.0371 | + [2026-08-26 18:29:49] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 300.2 | throughput per GPU (TFLOP/s/GPU): 602.4 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 1.027005E+01 | loss scale: 1.0 | grad norm: 6.321 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67968 | min safe capacity: 1.0371 | + [2026-08-26 18:29:49] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 297.8 | throughput per GPU (TFLOP/s/GPU): 607.2 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.808748E+00 | loss scale: 1.0 | grad norm: 5.544 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67968 | min safe capacity: 1.0371 | + [2026-08-26 18:29:50] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 296.9 | throughput per GPU (TFLOP/s/GPU): 609.0 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.545315E+00 | loss scale: 1.0 | grad norm: 5.293 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67968 | min safe capacity: 1.0371 | + [2026-08-26 18:29:50] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 298.0 | throughput per GPU (TFLOP/s/GPU): 606.8 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.446304E+00 | loss scale: 1.0 | grad norm: 5.022 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67968 | min safe capacity: 1.0371 | + [2026-08-26 18:29:50] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 298.5 | throughput per GPU (TFLOP/s/GPU): 605.8 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.324305E+00 | loss scale: 1.0 | grad norm: 4.703 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67968 | min safe capacity: 1.0371 | + [2026-08-26 18:29:51] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 296.2 | throughput per GPU (TFLOP/s/GPU): 610.5 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.210243E+00 | loss scale: 1.0 | grad norm: 4.466 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67968 | min safe capacity: 1.0371 | +[after training is done] datetime: 2026-08-26 18:29:51 +[rank0]:[W826 18:29:52.290646458 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank6]:[W826 18:29:53.486765266 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank4]:[W826 18:29:53.488339453 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank2]:[W826 18:29:53.489308748 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank3]:[W826 18:29:53.490253156 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank5]:[W826 18:29:53.020287247 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank1]:[W826 18:29:54.066830618 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank7]:[W826 18:29:54.353518346 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +W0826 18:29:59.556000 513115 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers +W0826 18:29:59.557000 513115 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 513207 closing signal SIGTERM +W0826 18:29:59.557000 513115 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 513211 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 513115 got signal: 15 diff --git a/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep3/mem_samples.csv b/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep3/mem_samples.csv new file mode 100644 index 0000000000000000000000000000000000000000..0569136b8d45665d744e7b18a4dd9cff5e7f4ed7 --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep3/mem_samples.csv @@ -0,0 +1,232 @@ +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 2 +1, 1 +2, 1 +3, 1 +4, 1 +5, 1 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 1 +5, 1 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 4 +5, 4 +6, 4 +7, 4 +0, 821 +1, 819 +2, 821 +3, 821 +4, 821 +5, 821 +6, 821 +7, 821 +0, 1537 +1, 1537 +2, 1537 +3, 1537 +4, 1537 +5, 1537 +6, 1537 +7, 1537 +0, 2551 +1, 2551 +2, 2551 +3, 2551 +4, 2551 +5, 2551 +6, 2551 +7, 2551 +0, 2959 +1, 2959 +2, 2959 +3, 2959 +4, 2959 +5, 2959 +6, 2959 +7, 2959 +0, 51537 +1, 51537 +2, 51537 +3, 51537 +4, 51249 +5, 51537 +6, 51537 +7, 51537 +0, 53019 +1, 53017 +2, 53017 +3, 53017 +4, 52729 +5, 53017 +6, 53017 +7, 53017 +0, 54651 +1, 54651 +2, 54651 +3, 54651 +4, 54363 +5, 54651 +6, 54651 +7, 54651 +0, 55207 +1, 55207 +2, 55207 +3, 55207 +4, 54919 +5, 55207 +6, 55207 +7, 55207 +0, 55207 +1, 55207 +2, 55207 +3, 55207 +4, 54919 +5, 55207 +6, 55207 +7, 55207 +0, 56875 +1, 56875 +2, 56875 +3, 56875 +4, 56587 +5, 56875 +6, 56875 +7, 56875 +0, 88753 +1, 88809 +2, 88809 +3, 88809 +4, 88793 +5, 88809 +6, 88809 +7, 88809 +0, 89465 +1, 89521 +2, 89521 +3, 89521 +4, 89505 +5, 89521 +6, 89521 +7, 89521 +0, 89465 +1, 89521 +2, 89521 +3, 89521 +4, 89505 +5, 89521 +6, 89521 +7, 89521 +0, 89465 +1, 89521 +2, 89521 +3, 89521 +4, 89505 +5, 89521 +6, 89521 +7, 89521 +0, 89771 +1, 89921 +2, 89861 +3, 89883 +4, 89877 +5, 89815 +6, 89873 +7, 89859 +0, 94923 +1, 94971 +2, 95021 +3, 94997 +4, 95137 +5, 94989 +6, 94969 +7, 95005 +0, 121033 +1, 111455 +2, 109251 +3, 111165 +4, 107317 +5, 111767 +6, 108277 +7, 121231 +0, 88409 +1, 93185 +2, 91351 +3, 90235 +4, 94155 +5, 90137 +6, 93185 +7, 90371 +0, 121957 +1, 121963 +2, 121963 +3, 121963 +4, 122229 +5, 121963 +6, 121963 +7, 121963 +0, 121957 +1, 121963 +2, 121963 +3, 121963 +4, 122229 +5, 121963 +6, 121963 +7, 121963 +0, 116719 +1, 116987 +2, 116987 +3, 116661 +4, 116671 +5, 116987 +6, 116987 +7, 116987 +0, 114651 +1, 114983 +2, 115565 +3, 115303 +4, 115831 +5, 115303 +6, 115565 +7, 115437 +0, 4 +1, 113885 +2, 113885 +3, 4 +4, 114151 +5, 113883 +6, 114273 +7, 113883 +0, 1 +1, 1 +2, 1 +3, 1 +4, 1 +5, 3 +6, 114185 +7, 4 diff --git a/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep3/run.env b/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep3/run.env new file mode 100644 index 0000000000000000000000000000000000000000..e190605bf52f7b834c17eb866ed94cbe99a7b798 --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep3/run.env @@ -0,0 +1,61 @@ +backend=ultraep +config=QWEN3_235B +family=tierB +repo=/workspace/Megatron-UltraEP +branch=ultraep-bracing-regions-through-kernels +commit=03d57ef7 +num_layers=12 +hidden=4096 +ffn_hidden=12288 +moe_ffn_hidden=1536 +num_experts=128 +topk=8 +experts_per_rank=16 +expert_params=18874368 +heads=64 +kv_groups=4 +kv_channels=128 +qk_layernorm=1 +shared_ffn=0 +shared_expert_overlap=0 +mla=0 +dense_first_layers=0 +moe_layers=12 +rotary_base=1000000 +rotary_scaling= +qkv_bias=0 +router_balancing=aux_loss +aux_loss_coeff=1e-3 +router_pre_softmax=0 +router_fusion=0 +router_dtype=fp32 +device_initiated_grouped_gemm=1 +norm_epsilon=1e-6 +seq_length=8192 +micro_batch_size=1 +global_batch_size=8 +num_microbatches=1 +assignments_per_rank=65536 +mean_tokens_per_expert=4096 +slot_policy=matched +slots_per_rank=16 +capacity_mode=static +capacity=1.083984375 +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=12 +ep_size=8 +echo_dispatch_overlap=1 +echo_dispatch_recompute=0 +ultraep_quota_min_tokens=512 +ultraep_weight_sync_join= +profile_enable=0 +kernel_trace=0 +fp8_recipe=mxfp8 diff --git a/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep3/train.log b/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep3/train.log new file mode 100644 index 0000000000000000000000000000000000000000..937280edb6472d092ec3bb24540c9f0bd6857885 --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep3/train.log @@ -0,0 +1,1450 @@ +[18:30:02] WARN: ultraep-bracing-regions-through-kernels lacks --moe-router-fusion; QWEN3_235B runs unfused (timing only, routing identical) +============================================================== + QWEN3_235B / ultraep (slots/rank=16 policy=matched) + E=128 K=8 H'=1536 m=16 + L=12 seq=8192 gbs=8 (1 microbatches) + cap_mode=static cap=1.083984375 cg=1 recompute=none + memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep3/memory_snapshots/snapshot.pickle + repo=/workspace/Megatron-UltraEP @ 03d57ef7 +============================================================== +W0826 18:30:04.073000 516444 torch/distributed/run.py:766] +W0826 18:30:04.073000 516444 torch/distributed/run.py:766] ***************************************** +W0826 18:30:04.073000 516444 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0826 18:30:04.073000 516444 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") +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") +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 .............................. 12 + encoder_seq_length .............................. 8192 + end_weight_decay ................................ 0.1 + eod_mask_loss ................................... False + error_injection_rate ............................ 0 + error_injection_type ............................ transient_error + eval_interval ................................... 1000000 + eval_iters ...................................... 0 + evidence_data_path .............................. None + exit_duration_in_mins ........................... None + exit_interval ................................... None + exit_on_missing_checkpoint ...................... False + exit_signal_handler ............................. False + exp_avg_dtype ................................... torch.float32 + exp_avg_sq_dtype ................................ torch.float32 + expert_model_parallel_size ...................... 8 + expert_tensor_parallel_size ..................... 1 + export_force_local_attention .................... False + export_kd_cfg ................................... None + export_kd_teacher_ckpt_format ................... None + export_kd_teacher_load .......................... None + export_kv_cache_quant ........................... False + export_legacy_megatron .......................... False + export_model_type ............................... GPTModel + export_moe_apply_probs_on_input ................. False + export_qk_l2_norm ............................... False + export_quant_cfg ................................ None + export_real_quant_cfg ........................... None + export_te_mcore_model ........................... False + external_cuda_graph ............................. False + ffn_hidden_size ................................. 12288 + finetune ........................................ False + finetune_data_split ............................. train + finetune_hf_dataset ............................. None + first_last_layers_bf16 .......................... False + flash_decode .................................... False + fp16 ............................................ False + fp16_lm_cross_entropy ........................... False + fp32_residual_connection ........................ False + fp8 ............................................. e4m3 + fp8_amax_compute_algo ........................... most_recent + fp8_amax_history_len ............................ 1 + fp8_interval .................................... 1 + fp8_margin ...................................... 0 + fp8_param_gather ................................ False + fp8_recipe ...................................... mxfp8 + fp8_wgrad ....................................... True + fsdp_double_buffer .............................. False + global_batch_size ............................... 8 + grad_reduce_in_bf16 ............................. False + gradient_accumulation_fusion .................... True + gradient_reduce_div_fusion ...................... True + group_query_attention ........................... True + head_lr_mult .................................... 1.0 + heterogeneous_layers_config_encoded_json ........ None + heterogeneous_layers_config_path ................ None + hidden_dropout .................................. 0.0 + hidden_size ..................................... 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 .............................. 0.001 + moe_capacity_monitor ............................ True + moe_deepep_num_sms .............................. 20 + moe_echo_enable_random_offloading ............... False + moe_echo_expert_dispatch_overlap ................ False + moe_echo_expert_dispatcher_type ................. hybridep + moe_echo_recompute_expert_dispatch .............. False + moe_enable_deepep ............................... False + moe_enable_echo ................................. False + moe_enable_hybridep ............................. True + moe_enable_ultraep .............................. True + moe_expert_capacity_factor ...................... None + moe_extended_tp ................................. False + moe_ffn_hidden_size ............................. 1536 + 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 .............. 16 + moe_pad_expert_input_to_capacity ................ False + moe_per_layer_logging ........................... False + moe_permute_fusion .............................. False + moe_received_token_capacity ..................... 1.083984375 + 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-06 + normalization ................................... RMSNorm + num_attention_heads ............................. 64 + num_channels .................................... 3 + num_classes ..................................... 1000 + num_dataset_builder_threads ..................... 1 + num_distributed_optimizer_instances ............. 1 + num_experts ..................................... 128 + num_layers ...................................... 12 + num_layers_at_end_in_bf16 ....................... 1 + num_layers_at_start_in_bf16 ..................... 1 + num_layers_per_virtual_pipeline_stage ........... None + num_query_groups ................................ 4 + num_virtual_stages_per_pipeline_rank ............ None + num_workers ..................................... 2 + object_storage_cache_path ....................... None + one_logger_async ................................ False + one_logger_project .............................. megatron-lm + one_logger_run_name ............................. None + onnx_safe ....................................... None + openai_gelu ..................................... False + optimizer ....................................... adam + optimizer_cpu_offload ........................... False + optimizer_offload_fraction ...................... 1.0 + output_bert_embeddings .......................... False + overlap_cpu_optimizer_d2h_h2d ................... False + overlap_grad_reduce ............................. True + overlap_moe_expert_parallel_comm ................ False + overlap_p2p_comm ................................ False + overlap_p2p_comm_warmup_flush ................... False + overlap_param_gather ............................ True + overlap_param_gather_with_optimizer_step ........ False + override_opt_param_scheduler .................... False + params_dtype .................................... torch.bfloat16 + patch_dim ....................................... 16 + per_split_data_args_path ........................ None + perform_initialization .......................... True + pin_cpu_grads ................................... True + pin_cpu_params .................................. True + pipeline_model_parallel_comm_backend ............ None + pipeline_model_parallel_layout .................. None + pipeline_model_parallel_size .................... 1 + position_embedding_type ......................... rope + pretrained_checkpoint ........................... None + profile ......................................... False + profile_ranks ................................... [0] + profile_step_end ................................ 12 + profile_step_start .............................. 10 + q_lora_rank ..................................... None + qk_head_dim ..................................... 128 + qk_l2_norm ...................................... False + qk_layernorm .................................... True + qk_pos_emb_head_dim ............................. 64 + query_in_block_prob ............................. 0.1 + rampup_batch_size ............................... None + rank ............................................ 0 + recompute_granularity ........................... None + recompute_method ................................ None + recompute_modules ............................... None + recompute_num_layers ............................ None + record_memory_history ........................... False + relative_attention_max_distance ................. 128 + relative_attention_num_buckets .................. 32 + replication ..................................... False + replication_factor .............................. 2 + replication_jump ................................ None + rerun_mode ...................................... validate_results + reset_attention_mask ............................ False + reset_position_ids .............................. False + result_rejected_tracker_filename ................ None + retriever_report_topk_accuracies ................ [] + retriever_score_scaling ......................... False + retriever_seq_length ............................ 256 + retro_add_retriever ............................. False + retro_attention_gate ............................ 1 + retro_cyclic_train_iters ........................ None + retro_encoder_attention_dropout ................. 0.1 + retro_encoder_hidden_dropout .................... 0.1 + retro_encoder_layers ............................ 2 + retro_num_neighbors ............................. 2 + retro_num_retrieved_chunks ...................... 2 + retro_project_dir ............................... None + retro_verify_neighbor_count ..................... True + reuse_grad_buf_for_mxfp8_param_ag ............... False + rope_scaling_factor ............................. 8.0 + rope_type ....................................... None + rotary_base ..................................... 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 ...................................... 8192 + sequence_parallel ............................... False + sft ............................................. False + sft_tokenizer_prompt_format ..................... nemotron-h-aligned + sgd_momentum .................................... 0.9 + short_seq_prob .................................. 0.1 + skip_train ...................................... False + skipped_train_samples ........................... 0 + spec ............................................ None + split ........................................... 100,0,0 + squared_relu .................................... False + start_weight_decay .............................. 0.1 + straggler_ctrlr_port ............................ 65535 + straggler_minmax_count .......................... 1 + suggested_communication_unit_size ............... None + swiglu .......................................... True + swin_backbone_type .............................. tiny + symmetric_ar_type ............................... None + te_rng_tracker .................................. True + tensor_model_parallel_size ...................... 1 + tensorboard_dir ................................. None + tensorboard_log_interval ........................ 1 + tensorboard_queue_size .......................... 1000 + test_data_path .................................. None + test_mode ....................................... False + tiktoken_num_special_tokens ..................... 1000 + tiktoken_pattern ................................ None + tiktoken_special_tokens ......................... None + timing_log_level ................................ 0 + timing_log_option ............................... minmax + titles_data_path ................................ None + tokenizer_model ................................. None + tokenizer_type .................................. GPT2BPETokenizer + torch_fsdp2_reshard_after_forward ............... True + tp_comm_bootstrap_backend ....................... nccl + tp_comm_bulk_dgrad .............................. True + tp_comm_bulk_wgrad .............................. True + tp_comm_overlap ................................. False + tp_comm_overlap_ag .............................. True + tp_comm_overlap_cfg ............................. None + tp_comm_overlap_rs .............................. True + tp_comm_overlap_rs_dgrad ........................ False + tp_comm_split_ag ................................ True + tp_comm_split_rs ................................ True + train_data_path ................................. None + train_iters ..................................... 12 + train_samples ................................... None + train_sync_interval ............................. None + transformer_impl ................................ transformer_engine + transformer_pipeline_model_parallel_size ........ 1 + untie_embeddings_and_output_weights ............. True + use_checkpoint_args ............................. False + use_checkpoint_opt_param_scheduler .............. False + use_cpu_initialization .......................... None + use_custom_fsdp ................................. False + use_dist_ckpt ................................... True + use_dist_ckpt_deprecated ........................ False + use_distributed_optimizer ....................... True + use_flash_attn .................................. False + use_legacy_models ............................... False + use_mp_args_from_checkpoint_args ................ False + use_one_sent_docs ............................... False + use_persistent_ckpt_worker ...................... False + use_precision_aware_optimizer ................... False + use_pytorch_profiler ............................ False + use_ring_exchange_p2p ........................... False + use_rope_scaling ................................ False + use_rotary_position_embeddings .................. False + use_sharp ....................................... False + use_tokenizer_model_from_checkpoint_args ........ True + use_torch_fsdp2 ................................. False + use_torch_optimizer_for_cpu_offload ............. False + use_tp_pp_dp_mapping ............................ False + v_head_dim ...................................... 128 + valid_data_path ................................. None + variable_seq_lengths ............................ False + virtual_pipeline_model_parallel_size ............ None + vision_backbone_type ............................ vit + vision_pretraining .............................. False + vision_pretraining_type ......................... classify + vocab_extra_ids ................................. 0 + vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json + vocab_size ...................................... None + wandb_exp_name .................................. + wandb_project ................................... + wandb_save_dir .................................. + weight_decay .................................... 0.1 + weight_decay_incr_style ......................... constant + wgrad_deferral_limit ............................ 0 + world_size ...................................... 8 + yaml_cfg ........................................ None +-------------------- end of arguments --------------------- +INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1 +> building GPT2BPETokenizer tokenizer ... +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") + > 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") +[Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank [Gloo] Rank [Gloo] Rank 351 is connected to is connected to is connected to 777 peer ranks. peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : Expected number of connected peer ranks is : 777 + + +[Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank [Gloo] Rank 7 is connected to 67 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 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 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank 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 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 +> initialized tensor model parallel with size 1 +> initialized pipeline model parallel with size 1 +> setting random seeds to 1234 ... +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[rank4]:[W826 18:30:22.147183612 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]:[W826 18:30:22.147865238 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. +[rank6]:[W826 18:30:22.147891014 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. +[rank1]:[W826 18:30:22.150628720 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank5]:[W826 18:30:22.151517228 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 5] using GPU 5 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank3]:[W826 18:30:22.153405982 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank2]:[W826 18:30:22.156368272 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: 0.034 seconds +WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations. +> compiling and loading fused kernels ... +[rank0]:[W826 18:30:22.302368908 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: 3.725 seconds +time to initialize megatron (seconds): 7.080 +[after megatron is initialized] datetime: 2026-08-26 18:30:27 +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): 8521882624 +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: 14 +Params for bucket 1 (206045184 elements, 206045184 padded size): + module.output_layer.weight +Params for bucket 2 (71840000 elements, 71840000 padded size): + module.decoder.final_layernorm.weight + module.decoder.layers.11.mlp.router.weight + module.decoder.layers.11.self_attention.k_layernorm.weight + module.decoder.layers.11.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.11.self_attention.linear_proj.weight + module.decoder.layers.11.self_attention.q_layernorm.weight + module.decoder.layers.11.pre_mlp_layernorm.weight + module.decoder.layers.11.self_attention.linear_qkv.weight +Params for bucket 3 (71835904 elements, 71835904 padded size): + module.decoder.layers.10.mlp.router.weight + module.decoder.layers.10.self_attention.k_layernorm.weight + module.decoder.layers.10.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.10.self_attention.linear_proj.weight + module.decoder.layers.10.self_attention.linear_qkv.weight + module.decoder.layers.10.self_attention.q_layernorm.weight + module.decoder.layers.10.pre_mlp_layernorm.weight +Params for bucket 4 (71835904 elements, 71835904 padded size): + module.decoder.layers.9.mlp.router.weight + module.decoder.layers.9.self_attention.k_layernorm.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.q_layernorm.weight + module.decoder.layers.9.self_attention.linear_qkv.weight + module.decoder.layers.9.pre_mlp_layernorm.weight +Params for bucket 5 (71835904 elements, 71835904 padded size): + module.decoder.layers.8.self_attention.linear_qkv.weight + module.decoder.layers.8.pre_mlp_layernorm.weight + module.decoder.layers.8.self_attention.q_layernorm.weight + module.decoder.layers.8.mlp.router.weight + module.decoder.layers.8.self_attention.k_layernorm.weight + module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.8.self_attention.linear_proj.weight +Params for bucket 6 (71835904 elements, 71835904 padded size): + module.decoder.layers.7.mlp.router.weight + module.decoder.layers.7.self_attention.k_layernorm.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.q_layernorm.weight + module.decoder.layers.7.pre_mlp_layernorm.weight + module.decoder.layers.7.self_attention.linear_qkv.weight +Params for bucket 7 (71835904 elements, 71835904 padded size): + module.decoder.layers.6.mlp.router.weight + module.decoder.layers.6.self_attention.k_layernorm.weight + module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.6.self_attention.linear_proj.weight + module.decoder.layers.6.self_attention.linear_qkv.weight + module.decoder.layers.6.pre_mlp_layernorm.weight + module.decoder.layers.6.self_attention.q_layernorm.weight +Params for bucket 8 (71835904 elements, 71835904 padded size): + module.decoder.layers.5.mlp.router.weight + module.decoder.layers.5.self_attention.k_layernorm.weight + module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.5.self_attention.q_layernorm.weight + module.decoder.layers.5.self_attention.linear_qkv.weight + module.decoder.layers.5.pre_mlp_layernorm.weight + module.decoder.layers.5.self_attention.linear_proj.weight +Params for bucket 9 (71835904 elements, 71835904 padded size): + module.decoder.layers.4.mlp.router.weight + module.decoder.layers.4.self_attention.k_layernorm.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.q_layernorm.weight + module.decoder.layers.4.pre_mlp_layernorm.weight + module.decoder.layers.4.self_attention.linear_qkv.weight +Params for bucket 10 (71835904 elements, 71835904 padded size): + module.decoder.layers.3.mlp.router.weight + module.decoder.layers.3.self_attention.k_layernorm.weight + module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.3.self_attention.linear_proj.weight + module.decoder.layers.3.pre_mlp_layernorm.weight + module.decoder.layers.3.self_attention.linear_qkv.weight + module.decoder.layers.3.self_attention.q_layernorm.weight +Params for bucket 11 (71835904 elements, 71835904 padded size): + module.decoder.layers.2.mlp.router.weight + module.decoder.layers.2.self_attention.k_layernorm.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.pre_mlp_layernorm.weight + module.decoder.layers.2.self_attention.linear_qkv.weight + module.decoder.layers.2.self_attention.q_layernorm.weight +Params for bucket 12 (71835904 elements, 71835904 padded size): + module.decoder.layers.1.mlp.router.weight + module.decoder.layers.1.self_attention.k_layernorm.weight + module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.1.self_attention.linear_proj.weight + module.decoder.layers.1.self_attention.linear_qkv.weight + module.decoder.layers.1.pre_mlp_layernorm.weight + module.decoder.layers.1.self_attention.q_layernorm.weight +Params for bucket 13 (71835904 elements, 71835904 padded size): + module.decoder.layers.0.mlp.router.weight + module.decoder.layers.0.self_attention.q_layernorm.weight + module.decoder.layers.0.self_attention.linear_qkv.weight + module.decoder.layers.0.self_attention.linear_proj.weight + module.decoder.layers.0.self_attention.k_layernorm.weight + module.decoder.layers.0.pre_mlp_layernorm.weight + module.decoder.layers.0.self_attention.linear_qkv.layer_norm_weight +Params for bucket 14 (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: 76 +Params for bucket 1 (44040192 elements, 44040192 padded size): + module.decoder.layers.11.mlp.experts.linear_fc2.weight15 + module.decoder.layers.11.mlp.experts.linear_fc2.weight11 + module.decoder.layers.11.mlp.experts.linear_fc2.weight13 + module.decoder.layers.11.mlp.experts.linear_fc2.weight9 + module.decoder.layers.11.mlp.experts.linear_fc2.weight12 + module.decoder.layers.11.mlp.experts.linear_fc2.weight14 + module.decoder.layers.11.mlp.experts.linear_fc2.weight10 +Params for bucket 2 (44040192 elements, 44040192 padded size): + module.decoder.layers.11.mlp.experts.linear_fc2.weight5 + module.decoder.layers.11.mlp.experts.linear_fc2.weight7 + module.decoder.layers.11.mlp.experts.linear_fc2.weight3 + module.decoder.layers.11.mlp.experts.linear_fc2.weight8 + module.decoder.layers.11.mlp.experts.linear_fc2.weight6 + module.decoder.layers.11.mlp.experts.linear_fc2.weight4 + module.decoder.layers.11.mlp.experts.linear_fc2.weight2 +Params for bucket 3 (50331648 elements, 50331648 padded size): + module.decoder.layers.11.mlp.experts.linear_fc2.weight1 + module.decoder.layers.11.mlp.experts.linear_fc1.weight14 + module.decoder.layers.11.mlp.experts.linear_fc1.weight15 + module.decoder.layers.11.mlp.experts.linear_fc2.weight0 + module.decoder.layers.11.mlp.experts.linear_fc1.weight13 +Params for bucket 4 (50331648 elements, 50331648 padded size): + module.decoder.layers.11.mlp.experts.linear_fc1.weight12 + module.decoder.layers.11.mlp.experts.linear_fc1.weight9 + module.decoder.layers.11.mlp.experts.linear_fc1.weight10 + module.decoder.layers.11.mlp.experts.linear_fc1.weight11 +Params for bucket 5 (50331648 elements, 50331648 padded size): + module.decoder.layers.11.mlp.experts.linear_fc1.weight8 + module.decoder.layers.11.mlp.experts.linear_fc1.weight5 + module.decoder.layers.11.mlp.experts.linear_fc1.weight6 + module.decoder.layers.11.mlp.experts.linear_fc1.weight7 +Params for bucket 6 (50331648 elements, 50331648 padded size): + module.decoder.layers.11.mlp.experts.linear_fc1.weight4 + module.decoder.layers.11.mlp.experts.linear_fc1.weight2 + module.decoder.layers.11.mlp.experts.linear_fc1.weight3 + module.decoder.layers.11.mlp.experts.linear_fc1.weight1 +Params for bucket 7 (44040192 elements, 44040192 padded size): + module.decoder.layers.11.mlp.experts.linear_fc1.weight0 + module.decoder.layers.10.mlp.experts.linear_fc2.weight15 + module.decoder.layers.10.mlp.experts.linear_fc2.weight13 + module.decoder.layers.10.mlp.experts.linear_fc2.weight11 + module.decoder.layers.10.mlp.experts.linear_fc2.weight12 + module.decoder.layers.10.mlp.experts.linear_fc2.weight14 +Params for bucket 8 (44040192 elements, 44040192 padded size): + module.decoder.layers.10.mlp.experts.linear_fc2.weight9 + module.decoder.layers.10.mlp.experts.linear_fc2.weight7 + module.decoder.layers.10.mlp.experts.linear_fc2.weight5 + module.decoder.layers.10.mlp.experts.linear_fc2.weight10 + module.decoder.layers.10.mlp.experts.linear_fc2.weight6 + module.decoder.layers.10.mlp.experts.linear_fc2.weight8 + module.decoder.layers.10.mlp.experts.linear_fc2.weight4 +Params for bucket 9 (50331648 elements, 50331648 padded size): + module.decoder.layers.10.mlp.experts.linear_fc2.weight3 + module.decoder.layers.10.mlp.experts.linear_fc2.weight1 + module.decoder.layers.10.mlp.experts.linear_fc1.weight14 + module.decoder.layers.10.mlp.experts.linear_fc1.weight15 + module.decoder.layers.10.mlp.experts.linear_fc2.weight0 + module.decoder.layers.10.mlp.experts.linear_fc2.weight2 +Params for bucket 10 (50331648 elements, 50331648 padded size): + module.decoder.layers.10.mlp.experts.linear_fc1.weight13 + module.decoder.layers.10.mlp.experts.linear_fc1.weight12 + module.decoder.layers.10.mlp.experts.linear_fc1.weight10 + module.decoder.layers.10.mlp.experts.linear_fc1.weight11 +Params for bucket 11 (50331648 elements, 50331648 padded size): + module.decoder.layers.10.mlp.experts.linear_fc1.weight9 + module.decoder.layers.10.mlp.experts.linear_fc1.weight6 + module.decoder.layers.10.mlp.experts.linear_fc1.weight7 + module.decoder.layers.10.mlp.experts.linear_fc1.weight8 +Params for bucket 12 (50331648 elements, 50331648 padded size): + module.decoder.layers.10.mlp.experts.linear_fc1.weight5 + module.decoder.layers.10.mlp.experts.linear_fc1.weight2 + module.decoder.layers.10.mlp.experts.linear_fc1.weight3 + module.decoder.layers.10.mlp.experts.linear_fc1.weight4 +Params for bucket 13 (44040192 elements, 44040192 padded size): + module.decoder.layers.9.mlp.experts.linear_fc2.weight15 + module.decoder.layers.9.mlp.experts.linear_fc2.weight13 + module.decoder.layers.10.mlp.experts.linear_fc1.weight0 + module.decoder.layers.10.mlp.experts.linear_fc1.weight1 + module.decoder.layers.9.mlp.experts.linear_fc2.weight14 +Params for bucket 14 (44040192 elements, 44040192 padded size): + module.decoder.layers.9.mlp.experts.linear_fc2.weight9 + module.decoder.layers.9.mlp.experts.linear_fc2.weight11 + module.decoder.layers.9.mlp.experts.linear_fc2.weight7 + module.decoder.layers.9.mlp.experts.linear_fc2.weight12 + module.decoder.layers.9.mlp.experts.linear_fc2.weight8 + module.decoder.layers.9.mlp.experts.linear_fc2.weight10 + module.decoder.layers.9.mlp.experts.linear_fc2.weight6 +Params for bucket 15 (50331648 elements, 50331648 padded size): + module.decoder.layers.9.mlp.experts.linear_fc2.weight5 + module.decoder.layers.9.mlp.experts.linear_fc2.weight3 + module.decoder.layers.9.mlp.experts.linear_fc2.weight1 + module.decoder.layers.9.mlp.experts.linear_fc2.weight4 + module.decoder.layers.9.mlp.experts.linear_fc2.weight2 + module.decoder.layers.9.mlp.experts.linear_fc2.weight0 + module.decoder.layers.9.mlp.experts.linear_fc1.weight15 +Params for bucket 16 (50331648 elements, 50331648 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight14 + module.decoder.layers.9.mlp.experts.linear_fc1.weight12 + module.decoder.layers.9.mlp.experts.linear_fc1.weight13 + module.decoder.layers.9.mlp.experts.linear_fc1.weight11 +Params for bucket 17 (50331648 elements, 50331648 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight10 + module.decoder.layers.9.mlp.experts.linear_fc1.weight9 + module.decoder.layers.9.mlp.experts.linear_fc1.weight7 + module.decoder.layers.9.mlp.experts.linear_fc1.weight8 +Params for bucket 18 (50331648 elements, 50331648 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight6 + module.decoder.layers.9.mlp.experts.linear_fc1.weight3 + module.decoder.layers.9.mlp.experts.linear_fc1.weight4 + module.decoder.layers.9.mlp.experts.linear_fc1.weight5 +Params for bucket 19 (44040192 elements, 44040192 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight2 + module.decoder.layers.8.mlp.experts.linear_fc2.weight15 + module.decoder.layers.9.mlp.experts.linear_fc1.weight0 + module.decoder.layers.9.mlp.experts.linear_fc1.weight1 +Params for bucket 20 (44040192 elements, 44040192 padded size): + module.decoder.layers.8.mlp.experts.linear_fc2.weight14 + module.decoder.layers.8.mlp.experts.linear_fc2.weight12 + module.decoder.layers.8.mlp.experts.linear_fc2.weight10 + module.decoder.layers.8.mlp.experts.linear_fc2.weight8 + module.decoder.layers.8.mlp.experts.linear_fc2.weight13 + module.decoder.layers.8.mlp.experts.linear_fc2.weight11 + module.decoder.layers.8.mlp.experts.linear_fc2.weight9 +Params for bucket 21 (44040192 elements, 44040192 padded size): + 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 + 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.weight1 +Params for bucket 22 (44040192 elements, 44040192 padded size): + module.decoder.layers.8.mlp.experts.linear_fc2.weight0 + module.decoder.layers.8.mlp.experts.linear_fc1.weight15 + module.decoder.layers.8.mlp.experts.linear_fc1.weight13 + module.decoder.layers.8.mlp.experts.linear_fc1.weight14 +Params for bucket 23 (50331648 elements, 50331648 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight12 + module.decoder.layers.8.mlp.experts.linear_fc1.weight9 + module.decoder.layers.8.mlp.experts.linear_fc1.weight10 + module.decoder.layers.8.mlp.experts.linear_fc1.weight11 +Params for bucket 24 (50331648 elements, 50331648 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight8 + module.decoder.layers.8.mlp.experts.linear_fc1.weight5 + module.decoder.layers.8.mlp.experts.linear_fc1.weight6 + module.decoder.layers.8.mlp.experts.linear_fc1.weight7 +Params for bucket 25 (50331648 elements, 50331648 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight4 + module.decoder.layers.8.mlp.experts.linear_fc1.weight2 + module.decoder.layers.8.mlp.experts.linear_fc1.weight3 + module.decoder.layers.8.mlp.experts.linear_fc1.weight1 +Params for bucket 26 (44040192 elements, 44040192 padded size): + module.decoder.layers.7.mlp.experts.linear_fc2.weight12 + module.decoder.layers.7.mlp.experts.linear_fc2.weight14 + module.decoder.layers.8.mlp.experts.linear_fc1.weight0 + module.decoder.layers.7.mlp.experts.linear_fc2.weight13 + module.decoder.layers.7.mlp.experts.linear_fc2.weight15 + module.decoder.layers.7.mlp.experts.linear_fc2.weight11 +Params for bucket 27 (44040192 elements, 44040192 padded size): + module.decoder.layers.7.mlp.experts.linear_fc2.weight10 + module.decoder.layers.7.mlp.experts.linear_fc2.weight6 + module.decoder.layers.7.mlp.experts.linear_fc2.weight8 + module.decoder.layers.7.mlp.experts.linear_fc2.weight4 + module.decoder.layers.7.mlp.experts.linear_fc2.weight9 + module.decoder.layers.7.mlp.experts.linear_fc2.weight7 + module.decoder.layers.7.mlp.experts.linear_fc2.weight5 +Params for bucket 28 (50331648 elements, 50331648 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight15 + module.decoder.layers.7.mlp.experts.linear_fc2.weight0 + module.decoder.layers.7.mlp.experts.linear_fc2.weight2 + 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_fc1.weight14 +Params for bucket 29 (50331648 elements, 50331648 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight13 + module.decoder.layers.7.mlp.experts.linear_fc1.weight12 + module.decoder.layers.7.mlp.experts.linear_fc1.weight10 + module.decoder.layers.7.mlp.experts.linear_fc1.weight11 +Params for bucket 30 (50331648 elements, 50331648 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight9 + module.decoder.layers.7.mlp.experts.linear_fc1.weight6 + module.decoder.layers.7.mlp.experts.linear_fc1.weight7 + module.decoder.layers.7.mlp.experts.linear_fc1.weight8 +Params for bucket 31 (50331648 elements, 50331648 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 + module.decoder.layers.7.mlp.experts.linear_fc1.weight2 +Params for bucket 32 (44040192 elements, 44040192 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight1 + module.decoder.layers.6.mlp.experts.linear_fc2.weight14 + module.decoder.layers.7.mlp.experts.linear_fc1.weight0 + module.decoder.layers.6.mlp.experts.linear_fc2.weight15 + module.decoder.layers.6.mlp.experts.linear_fc2.weight13 +Params for bucket 33 (44040192 elements, 44040192 padded size): + module.decoder.layers.6.mlp.experts.linear_fc2.weight12 + module.decoder.layers.6.mlp.experts.linear_fc2.weight8 + module.decoder.layers.6.mlp.experts.linear_fc2.weight10 + module.decoder.layers.6.mlp.experts.linear_fc2.weight6 + module.decoder.layers.6.mlp.experts.linear_fc2.weight11 + module.decoder.layers.6.mlp.experts.linear_fc2.weight9 + module.decoder.layers.6.mlp.experts.linear_fc2.weight7 +Params for bucket 34 (50331648 elements, 50331648 padded size): + module.decoder.layers.6.mlp.experts.linear_fc2.weight4 + module.decoder.layers.6.mlp.experts.linear_fc2.weight2 + module.decoder.layers.6.mlp.experts.linear_fc2.weight0 + module.decoder.layers.6.mlp.experts.linear_fc1.weight15 + module.decoder.layers.6.mlp.experts.linear_fc2.weight5 + module.decoder.layers.6.mlp.experts.linear_fc2.weight1 + module.decoder.layers.6.mlp.experts.linear_fc2.weight3 +Params for bucket 35 (50331648 elements, 50331648 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight14 + module.decoder.layers.6.mlp.experts.linear_fc1.weight13 + module.decoder.layers.6.mlp.experts.linear_fc1.weight11 + module.decoder.layers.6.mlp.experts.linear_fc1.weight12 +Params for bucket 36 (50331648 elements, 50331648 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight10 + module.decoder.layers.6.mlp.experts.linear_fc1.weight7 + module.decoder.layers.6.mlp.experts.linear_fc1.weight8 + module.decoder.layers.6.mlp.experts.linear_fc1.weight9 +Params for bucket 37 (50331648 elements, 50331648 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight6 + module.decoder.layers.6.mlp.experts.linear_fc1.weight4 + module.decoder.layers.6.mlp.experts.linear_fc1.weight5 + module.decoder.layers.6.mlp.experts.linear_fc1.weight3 +Params for bucket 38 (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.5.mlp.experts.linear_fc2.weight15 + module.decoder.layers.6.mlp.experts.linear_fc1.weight0 +Params for bucket 39 (44040192 elements, 44040192 padded size): + module.decoder.layers.5.mlp.experts.linear_fc2.weight13 + module.decoder.layers.5.mlp.experts.linear_fc2.weight11 + module.decoder.layers.5.mlp.experts.linear_fc2.weight9 + module.decoder.layers.5.mlp.experts.linear_fc2.weight14 + module.decoder.layers.5.mlp.experts.linear_fc2.weight10 + module.decoder.layers.5.mlp.experts.linear_fc2.weight12 + module.decoder.layers.5.mlp.experts.linear_fc2.weight8 +Params for bucket 40 (44040192 elements, 44040192 padded size): + 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 + module.decoder.layers.5.mlp.experts.linear_fc2.weight1 + module.decoder.layers.5.mlp.experts.linear_fc2.weight4 + module.decoder.layers.5.mlp.experts.linear_fc2.weight6 + module.decoder.layers.5.mlp.experts.linear_fc2.weight2 +Params for bucket 41 (44040192 elements, 44040192 padded size): + module.decoder.layers.5.mlp.experts.linear_fc2.weight0 + module.decoder.layers.5.mlp.experts.linear_fc1.weight15 + module.decoder.layers.5.mlp.experts.linear_fc1.weight13 + module.decoder.layers.5.mlp.experts.linear_fc1.weight14 +Params for bucket 42 (50331648 elements, 50331648 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight12 + module.decoder.layers.5.mlp.experts.linear_fc1.weight9 + module.decoder.layers.5.mlp.experts.linear_fc1.weight10 + module.decoder.layers.5.mlp.experts.linear_fc1.weight11 +Params for bucket 43 (50331648 elements, 50331648 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight8 + module.decoder.layers.5.mlp.experts.linear_fc1.weight6 + module.decoder.layers.5.mlp.experts.linear_fc1.weight5 + module.decoder.layers.5.mlp.experts.linear_fc1.weight7 +Params for bucket 44 (50331648 elements, 50331648 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight4 + module.decoder.layers.5.mlp.experts.linear_fc1.weight3 + module.decoder.layers.5.mlp.experts.linear_fc1.weight2 + 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.weight12 + module.decoder.layers.4.mlp.experts.linear_fc2.weight14 + module.decoder.layers.5.mlp.experts.linear_fc1.weight0 + module.decoder.layers.4.mlp.experts.linear_fc2.weight15 + module.decoder.layers.4.mlp.experts.linear_fc2.weight13 + module.decoder.layers.4.mlp.experts.linear_fc2.weight11 +Params for bucket 46 (44040192 elements, 44040192 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight10 + module.decoder.layers.4.mlp.experts.linear_fc2.weight8 + module.decoder.layers.4.mlp.experts.linear_fc2.weight6 + module.decoder.layers.4.mlp.experts.linear_fc2.weight4 + module.decoder.layers.4.mlp.experts.linear_fc2.weight9 + module.decoder.layers.4.mlp.experts.linear_fc2.weight7 + module.decoder.layers.4.mlp.experts.linear_fc2.weight5 +Params for bucket 47 (50331648 elements, 50331648 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight2 + module.decoder.layers.4.mlp.experts.linear_fc2.weight0 + module.decoder.layers.4.mlp.experts.linear_fc1.weight14 + module.decoder.layers.4.mlp.experts.linear_fc2.weight3 + module.decoder.layers.4.mlp.experts.linear_fc1.weight15 + module.decoder.layers.4.mlp.experts.linear_fc2.weight1 +Params for bucket 48 (50331648 elements, 50331648 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight13 + module.decoder.layers.4.mlp.experts.linear_fc1.weight10 + module.decoder.layers.4.mlp.experts.linear_fc1.weight11 + module.decoder.layers.4.mlp.experts.linear_fc1.weight12 +Params for bucket 49 (50331648 elements, 50331648 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight9 + module.decoder.layers.4.mlp.experts.linear_fc1.weight7 + module.decoder.layers.4.mlp.experts.linear_fc1.weight8 + module.decoder.layers.4.mlp.experts.linear_fc1.weight6 +Params for bucket 50 (50331648 elements, 50331648 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.weight2 + module.decoder.layers.4.mlp.experts.linear_fc1.weight3 +Params for bucket 51 (44040192 elements, 44040192 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight0 + module.decoder.layers.3.mlp.experts.linear_fc2.weight15 + module.decoder.layers.3.mlp.experts.linear_fc2.weight13 + module.decoder.layers.4.mlp.experts.linear_fc1.weight1 + module.decoder.layers.3.mlp.experts.linear_fc2.weight14 +Params for bucket 52 (44040192 elements, 44040192 padded size): + module.decoder.layers.3.mlp.experts.linear_fc2.weight11 + module.decoder.layers.3.mlp.experts.linear_fc2.weight9 + module.decoder.layers.3.mlp.experts.linear_fc2.weight7 + module.decoder.layers.3.mlp.experts.linear_fc2.weight12 + module.decoder.layers.3.mlp.experts.linear_fc2.weight8 + module.decoder.layers.3.mlp.experts.linear_fc2.weight10 + module.decoder.layers.3.mlp.experts.linear_fc2.weight6 +Params for bucket 53 (50331648 elements, 50331648 padded size): + module.decoder.layers.3.mlp.experts.linear_fc2.weight5 + module.decoder.layers.3.mlp.experts.linear_fc2.weight1 + module.decoder.layers.3.mlp.experts.linear_fc2.weight3 + module.decoder.layers.3.mlp.experts.linear_fc2.weight2 + module.decoder.layers.3.mlp.experts.linear_fc2.weight4 + module.decoder.layers.3.mlp.experts.linear_fc2.weight0 + module.decoder.layers.3.mlp.experts.linear_fc1.weight15 +Params for bucket 54 (50331648 elements, 50331648 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight14 + module.decoder.layers.3.mlp.experts.linear_fc1.weight11 + module.decoder.layers.3.mlp.experts.linear_fc1.weight12 + module.decoder.layers.3.mlp.experts.linear_fc1.weight13 +Params for bucket 55 (50331648 elements, 50331648 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight10 + module.decoder.layers.3.mlp.experts.linear_fc1.weight7 + module.decoder.layers.3.mlp.experts.linear_fc1.weight8 + module.decoder.layers.3.mlp.experts.linear_fc1.weight9 +Params for bucket 56 (50331648 elements, 50331648 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight6 + module.decoder.layers.3.mlp.experts.linear_fc1.weight4 + module.decoder.layers.3.mlp.experts.linear_fc1.weight5 + module.decoder.layers.3.mlp.experts.linear_fc1.weight3 +Params for bucket 57 (44040192 elements, 44040192 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight2 + module.decoder.layers.3.mlp.experts.linear_fc1.weight1 + module.decoder.layers.2.mlp.experts.linear_fc2.weight15 + module.decoder.layers.3.mlp.experts.linear_fc1.weight0 +Params for bucket 58 (44040192 elements, 44040192 padded size): + module.decoder.layers.2.mlp.experts.linear_fc2.weight14 + module.decoder.layers.2.mlp.experts.linear_fc2.weight10 + module.decoder.layers.2.mlp.experts.linear_fc2.weight12 + module.decoder.layers.2.mlp.experts.linear_fc2.weight8 + module.decoder.layers.2.mlp.experts.linear_fc2.weight11 + module.decoder.layers.2.mlp.experts.linear_fc2.weight13 + module.decoder.layers.2.mlp.experts.linear_fc2.weight9 +Params for bucket 59 (44040192 elements, 44040192 padded size): + 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 + module.decoder.layers.2.mlp.experts.linear_fc2.weight7 + module.decoder.layers.2.mlp.experts.linear_fc2.weight5 + module.decoder.layers.2.mlp.experts.linear_fc2.weight3 + module.decoder.layers.2.mlp.experts.linear_fc2.weight1 +Params for bucket 60 (44040192 elements, 44040192 padded size): + module.decoder.layers.2.mlp.experts.linear_fc2.weight0 + module.decoder.layers.2.mlp.experts.linear_fc1.weight13 + module.decoder.layers.2.mlp.experts.linear_fc1.weight14 + module.decoder.layers.2.mlp.experts.linear_fc1.weight15 +Params for bucket 61 (50331648 elements, 50331648 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight12 + module.decoder.layers.2.mlp.experts.linear_fc1.weight10 + module.decoder.layers.2.mlp.experts.linear_fc1.weight11 + module.decoder.layers.2.mlp.experts.linear_fc1.weight9 +Params for bucket 62 (50331648 elements, 50331648 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight8 + module.decoder.layers.2.mlp.experts.linear_fc1.weight7 + module.decoder.layers.2.mlp.experts.linear_fc1.weight5 + module.decoder.layers.2.mlp.experts.linear_fc1.weight6 +Params for bucket 63 (50331648 elements, 50331648 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight3 + module.decoder.layers.2.mlp.experts.linear_fc1.weight1 + module.decoder.layers.2.mlp.experts.linear_fc1.weight2 +Params for bucket 64 (44040192 elements, 44040192 padded size): + module.decoder.layers.1.mlp.experts.linear_fc2.weight12 + module.decoder.layers.1.mlp.experts.linear_fc2.weight14 + module.decoder.layers.2.mlp.experts.linear_fc1.weight0 + module.decoder.layers.1.mlp.experts.linear_fc2.weight15 + module.decoder.layers.1.mlp.experts.linear_fc2.weight13 + module.decoder.layers.1.mlp.experts.linear_fc2.weight11 +Params for bucket 65 (44040192 elements, 44040192 padded size): + module.decoder.layers.1.mlp.experts.linear_fc2.weight10 + module.decoder.layers.1.mlp.experts.linear_fc2.weight8 + 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.weight9 + module.decoder.layers.1.mlp.experts.linear_fc2.weight7 + module.decoder.layers.1.mlp.experts.linear_fc2.weight5 +Params for bucket 66 (50331648 elements, 50331648 padded size): + module.decoder.layers.1.mlp.experts.linear_fc2.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight15 + module.decoder.layers.1.mlp.experts.linear_fc2.weight0 + module.decoder.layers.1.mlp.experts.linear_fc2.weight3 + module.decoder.layers.1.mlp.experts.linear_fc2.weight1 + module.decoder.layers.1.mlp.experts.linear_fc1.weight14 +Params for bucket 67 (50331648 elements, 50331648 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight13 + module.decoder.layers.1.mlp.experts.linear_fc1.weight12 + module.decoder.layers.1.mlp.experts.linear_fc1.weight11 + module.decoder.layers.1.mlp.experts.linear_fc1.weight10 +Params for bucket 68 (50331648 elements, 50331648 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight9 + module.decoder.layers.1.mlp.experts.linear_fc1.weight6 + module.decoder.layers.1.mlp.experts.linear_fc1.weight8 + module.decoder.layers.1.mlp.experts.linear_fc1.weight7 +Params for bucket 69 (50331648 elements, 50331648 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight5 + module.decoder.layers.1.mlp.experts.linear_fc1.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight3 + module.decoder.layers.1.mlp.experts.linear_fc1.weight4 +Params for bucket 70 (44040192 elements, 44040192 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight1 + module.decoder.layers.0.mlp.experts.linear_fc2.weight13 + module.decoder.layers.0.mlp.experts.linear_fc2.weight15 + module.decoder.layers.0.mlp.experts.linear_fc2.weight14 + module.decoder.layers.1.mlp.experts.linear_fc1.weight0 +Params for bucket 71 (44040192 elements, 44040192 padded size): + module.decoder.layers.0.mlp.experts.linear_fc2.weight11 + module.decoder.layers.0.mlp.experts.linear_fc2.weight9 + module.decoder.layers.0.mlp.experts.linear_fc2.weight7 + module.decoder.layers.0.mlp.experts.linear_fc2.weight12 + module.decoder.layers.0.mlp.experts.linear_fc2.weight10 + module.decoder.layers.0.mlp.experts.linear_fc2.weight8 + module.decoder.layers.0.mlp.experts.linear_fc2.weight6 +Params for bucket 72 (50331648 elements, 50331648 padded size): + module.decoder.layers.0.mlp.experts.linear_fc2.weight5 + module.decoder.layers.0.mlp.experts.linear_fc2.weight1 + module.decoder.layers.0.mlp.experts.linear_fc2.weight3 + module.decoder.layers.0.mlp.experts.linear_fc2.weight2 + module.decoder.layers.0.mlp.experts.linear_fc2.weight4 + module.decoder.layers.0.mlp.experts.linear_fc2.weight0 + module.decoder.layers.0.mlp.experts.linear_fc1.weight15 +Params for bucket 73 (50331648 elements, 50331648 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight14 + module.decoder.layers.0.mlp.experts.linear_fc1.weight12 + module.decoder.layers.0.mlp.experts.linear_fc1.weight11 + module.decoder.layers.0.mlp.experts.linear_fc1.weight13 +Params for bucket 74 (50331648 elements, 50331648 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight10 + module.decoder.layers.0.mlp.experts.linear_fc1.weight8 + module.decoder.layers.0.mlp.experts.linear_fc1.weight9 + module.decoder.layers.0.mlp.experts.linear_fc1.weight7 +Params for bucket 75 (50331648 elements, 50331648 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight6 + 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 76 (37748736 elements, 37748736 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: 2026-08-26 18:30:31 +> building train, validation, and test datasets ... + > datasets target sizes (minimum size): + train: 96 + 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=(96, 0, 0) and config=GPTDatasetConfig(random_seed=1234, sequence_length=8192, blend=(['/workspace/data/wikitext/wikitext103_gpt2_text_document'], None), blend_per_split=None, split='100,0,0', split_matrix=[(0, 1.0), None, None], num_dataset_builder_threads=1, path_to_cache=None, mmap_bin_files=True, mock=False, tokenizer=, mid_level_dataset_surplus=0.005, reset_position_ids=False, reset_attention_mask=False, eod_mask_loss=False, create_attention_mask=True, drop_last_partial_validation_sequence=True, add_extra_token_to_sequence=True, object_storage_cache_path=None) +INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx +INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths +INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers +INFO:megatron.core.datasets.indexed_dataset: Extract the document indices +INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029 +INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350 +INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices +INFO:megatron.core.datasets.gpt_dataset: Load the document index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-document_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the sample index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-sample_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-shuffle_index.npy +INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 14536 +> finished creating GPT datasets ... +[after dataloaders are built] datetime: 2026-08-26 18:30:31 +done with setup ... +training ... +(min, max) time across ranks (ms): + model-and-optimizer-setup ......................: (3722.10, 3731.13) + train/valid/test-data-iterators-setup ..........: (70.29, 84.78) +Overwriting rerun_state_machine.current_iteration from -1 to 0... +[before the start of training step] datetime: 2026-08-26 18:30:31 +/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: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: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: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:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") + [2026-08-26 18:30:58] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 26287.4 | throughput per GPU (TFLOP/s/GPU): 6.9 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.116001E+01 | loss scale: 1.0 | grad norm: 44.806 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67456 | min safe capacity: 1.0293 | +Number of parameters in transformer block in billions: 29.85 +Number of parameters in embedding layers in billions: 0.41 +Total number of parameters in billions: 30.26 +Number of parameters in most loaded shard in billions: 30.2590 +Theoretical memory footprints: weight and optimizer=216428.96 MB +[Rank 0] (after 1 iterations) memory (MB) | allocated: 73714.98388671875 | max allocated: 79407.14892578125 | reserved: 85290.0 | max reserved: 85290.0 +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") + [2026-08-26 18:30:58] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 862.6 | throughput per GPU (TFLOP/s/GPU): 209.6 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.117607E+01 | loss scale: 1.0 | grad norm: 44.340 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67712 | min safe capacity: 1.0332 | + [2026-08-26 18:30:59] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 308.8 | throughput per GPU (TFLOP/s/GPU): 585.6 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 9.685719E+00 | loss scale: 1.0 | grad norm: 62.699 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67840 | min safe capacity: 1.0352 | +Capture CUDA graph for training!!! +CUDA graph capture done!!! + [2026-08-26 18:31:02] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 2723.4 | throughput per GPU (TFLOP/s/GPU): 66.4 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 1.338039E+01 | loss scale: 1.0 | grad norm: 20.324 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67840 | min safe capacity: 1.0352 | + [2026-08-26 18:31:02] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 299.7 | throughput per GPU (TFLOP/s/GPU): 603.3 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.052909E+01 | loss scale: 1.0 | grad norm: 95.617 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67840 | min safe capacity: 1.0352 | + [2026-08-26 18:31:02] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 299.0 | throughput per GPU (TFLOP/s/GPU): 604.9 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 1.026443E+01 | loss scale: 1.0 | grad norm: 40.877 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67840 | min safe capacity: 1.0352 | + [2026-08-26 18:31:02] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 299.4 | throughput per GPU (TFLOP/s/GPU): 604.0 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 1.026034E+01 | loss scale: 1.0 | grad norm: 6.254 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67840 | min safe capacity: 1.0352 | + [2026-08-26 18:31:03] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 298.7 | throughput per GPU (TFLOP/s/GPU): 605.4 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.796303E+00 | loss scale: 1.0 | grad norm: 5.588 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67840 | min safe capacity: 1.0352 | + [2026-08-26 18:31:03] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 297.5 | throughput per GPU (TFLOP/s/GPU): 607.8 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.526175E+00 | loss scale: 1.0 | grad norm: 5.463 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67840 | min safe capacity: 1.0352 | + [2026-08-26 18:31:03] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 298.8 | throughput per GPU (TFLOP/s/GPU): 605.1 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.426507E+00 | loss scale: 1.0 | grad norm: 5.309 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67840 | min safe capacity: 1.0352 | + [2026-08-26 18:31:04] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 297.4 | throughput per GPU (TFLOP/s/GPU): 608.0 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.297106E+00 | loss scale: 1.0 | grad norm: 5.090 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67840 | min safe capacity: 1.0352 | + [2026-08-26 18:31:04] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 295.7 | throughput per GPU (TFLOP/s/GPU): 611.5 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.173257E+00 | loss scale: 1.0 | grad norm: 4.744 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 67840 | min safe capacity: 1.0352 | +[after training is done] datetime: 2026-08-26 18:31:04 +[rank0]:[W826 18:31:05.678672686 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank4]:[W826 18:31:07.179634767 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank3]:[W826 18:31:07.261871250 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank7]:[W826 18:31:07.264503336 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank5]:[W826 18:31:07.265107183 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank6]:[W826 18:31:07.266238746 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank1]:[W826 18:31:07.267545786 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank2]:[W826 18:31:07.832804412 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +W0826 18:31:12.769000 516444 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers +W0826 18:31:12.770000 516444 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 516535 closing signal SIGTERM +W0826 18:31:12.770000 516444 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 516537 closing signal SIGTERM +W0826 18:31:12.771000 516444 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 516539 closing signal SIGTERM +W0826 18:31:12.771000 516444 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 516540 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 516444 got signal: 15 diff --git a/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep1/mem_samples.csv b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep1/mem_samples.csv new file mode 100644 index 0000000000000000000000000000000000000000..98e56fd4aa9e9889b3d2612756d39f6ccfb5b24d --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep1/mem_samples.csv @@ -0,0 +1,240 @@ +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 1 +1, 1 +2, 1 +3, 1 +4, 1 +5, 1 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 1 +5, 1 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 4 +5, 4 +6, 4 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 4 +5, 4 +6, 4 +7, 4 +0, 27 +1, 633 +2, 633 +3, 633 +4, 633 +5, 633 +6, 633 +7, 725 +0, 905 +1, 917 +2, 931 +3, 937 +4, 913 +5, 921 +6, 913 +7, 917 +0, 1831 +1, 1831 +2, 1831 +3, 1831 +4, 1831 +5, 1831 +6, 1831 +7, 1831 +0, 2147 +1, 2159 +2, 2137 +3, 2151 +4, 2133 +5, 2131 +6, 2129 +7, 2129 +0, 11671 +1, 11671 +2, 11611 +3, 11811 +4, 11811 +5, 11811 +6, 12011 +7, 11611 +0, 46961 +1, 46953 +2, 47007 +3, 46939 +4, 47003 +5, 47011 +6, 46973 +7, 46991 +0, 48353 +1, 48395 +2, 48345 +3, 48357 +4, 48381 +5, 48395 +6, 48405 +7, 48375 +0, 48647 +1, 48649 +2, 48659 +3, 48621 +4, 48645 +5, 48659 +6, 48649 +7, 48659 +0, 48791 +1, 48793 +2, 48803 +3, 48765 +4, 48789 +5, 48803 +6, 48793 +7, 48803 +0, 49447 +1, 49447 +2, 49451 +3, 49421 +4, 49419 +5, 49451 +6, 49405 +7, 49451 +0, 56991 +1, 57093 +2, 57049 +3, 57129 +4, 57045 +5, 57133 +6, 57051 +7, 57133 +0, 85219 +1, 85241 +2, 85217 +3, 85233 +4, 85233 +5, 85217 +6, 85217 +7, 85217 +0, 85607 +1, 85629 +2, 85605 +3, 85621 +4, 85621 +5, 85605 +6, 85605 +7, 85605 +0, 85607 +1, 85629 +2, 85605 +3, 85621 +4, 85621 +5, 85605 +6, 85605 +7, 85605 +0, 85845 +1, 85849 +2, 85821 +3, 85933 +4, 85895 +5, 85881 +6, 85865 +7, 85881 +0, 86657 +1, 86677 +2, 86655 +3, 86671 +4, 86669 +5, 86655 +6, 86655 +7, 86655 +0, 87423 +1, 87371 +2, 87409 +3, 87395 +4, 87407 +5, 87373 +6, 87357 +7, 87343 +0, 116721 +1, 116759 +2, 116699 +3, 116727 +4, 116771 +5, 116699 +6, 116743 +7, 116699 +0, 116981 +1, 116881 +2, 117931 +3, 117141 +4, 118163 +5, 120025 +6, 117225 +7, 120025 +0, 120103 +1, 120129 +2, 120099 +3, 119829 +4, 120121 +5, 120099 +6, 120101 +7, 120099 +0, 120103 +1, 120129 +2, 120099 +3, 119829 +4, 120121 +5, 120099 +6, 120101 +7, 120099 +0, 117099 +1, 117125 +2, 117095 +3, 116825 +4, 117117 +5, 117095 +6, 117097 +7, 117095 +0, 772 +1, 115703 +2, 772 +3, 115083 +4, 115695 +5, 116061 +6, 115675 +7, 115673 +0, 641 +1, 1 +2, 1 +3, 1 +4, 1 +5, 4 +6, 132 +7, 115623 diff --git a/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep1/run.env b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep1/run.env new file mode 100644 index 0000000000000000000000000000000000000000..3213d7fbe624a565f15157e5d63615da960f6547 --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep1/run.env @@ -0,0 +1,61 @@ +backend=ultraep +config=QWEN3_30B +family=tierB +repo=/workspace/Megatron-UltraEP +branch=ultraep-bracing-regions-through-kernels +commit=03d57ef7 +num_layers=48 +hidden=2048 +ffn_hidden=6144 +moe_ffn_hidden=768 +num_experts=128 +topk=8 +experts_per_rank=16 +expert_params=4718592 +heads=32 +kv_groups=4 +kv_channels=128 +qk_layernorm=1 +shared_ffn=0 +shared_expert_overlap=0 +mla=0 +dense_first_layers=0 +moe_layers=48 +rotary_base=1000000 +rotary_scaling= +qkv_bias=0 +router_balancing=aux_loss +aux_loss_coeff=1e-3 +router_pre_softmax=0 +router_fusion=0 +router_dtype=fp32 +device_initiated_grouped_gemm=1 +norm_epsilon=1e-6 +seq_length=4096 +micro_batch_size=1 +global_batch_size=8 +num_microbatches=1 +assignments_per_rank=32768 +mean_tokens_per_expert=2048 +slot_policy=matched +slots_per_rank=16 +capacity_mode=static +capacity=1.11328125 +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=12 +ep_size=8 +echo_dispatch_overlap=1 +echo_dispatch_recompute=0 +ultraep_quota_min_tokens=256 +ultraep_weight_sync_join= +profile_enable=0 +kernel_trace=0 +fp8_recipe=mxfp8 diff --git a/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep1/train.log b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep1/train.log new file mode 100644 index 0000000000000000000000000000000000000000..11c867e0761c8d4822ae7a3983586ebeef089030 --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep1/train.log @@ -0,0 +1,2878 @@ +[18:31:14] WARN: ultraep-bracing-regions-through-kernels lacks --moe-router-fusion; QWEN3_30B runs unfused (timing only, routing identical) +============================================================== + QWEN3_30B / ultraep (slots/rank=16 policy=matched) + E=128 K=8 H'=768 m=16 + L=48 seq=4096 gbs=8 (1 microbatches) + cap_mode=static cap=1.11328125 cg=1 recompute=none + memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep1/memory_snapshots/snapshot.pickle + repo=/workspace/Megatron-UltraEP @ 03d57ef7 +============================================================== +W0826 18:31:16.307000 519501 torch/distributed/run.py:766] +W0826 18:31:16.307000 519501 torch/distributed/run.py:766] ***************************************** +W0826 18:31:16.307000 519501 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0826 18:31:16.307000 519501 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") +/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 .............................. 48 + 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 ................................. 6144 + finetune ........................................ False + finetune_data_split ............................. train + finetune_hf_dataset ............................. None + first_last_layers_bf16 .......................... False + flash_decode .................................... False + fp16 ............................................ False + fp16_lm_cross_entropy ........................... False + fp32_residual_connection ........................ False + fp8 ............................................. e4m3 + fp8_amax_compute_algo ........................... most_recent + fp8_amax_history_len ............................ 1 + fp8_interval .................................... 1 + fp8_margin ...................................... 0 + fp8_param_gather ................................ False + fp8_recipe ...................................... mxfp8 + fp8_wgrad ....................................... True + fsdp_double_buffer .............................. False + global_batch_size ............................... 8 + grad_reduce_in_bf16 ............................. False + gradient_accumulation_fusion .................... True + gradient_reduce_div_fusion ...................... True + group_query_attention ........................... True + head_lr_mult .................................... 1.0 + heterogeneous_layers_config_encoded_json ........ None + heterogeneous_layers_config_path ................ None + hidden_dropout .................................. 0.0 + hidden_size ..................................... 2048 + hierarchical_context_parallel_sizes ............. None + high_priority_stream_groups ..................... [] + hybrid_attention_ratio .......................... 0.0 + hybrid_mlp_ratio ................................ 0.0 + hybrid_override_pattern ......................... None + hysteresis ...................................... 2 + ict_head_size ................................... None + ict_load ........................................ None + img_h ........................................... 224 + img_w ........................................... 224 + indexer_batch_size .............................. 128 + indexer_log_interval ............................ 1000 + inference_batch_times_seqlen_threshold .......... -1 + inference_dynamic_batching ...................... False + inference_dynamic_batching_buffer_guaranteed_fraction 0.2 + inference_dynamic_batching_buffer_overflow_factor None + inference_dynamic_batching_buffer_size_gb ....... 40.0 + inference_dynamic_batching_chunk_size ........... 256 + inference_dynamic_batching_max_requests_override None + inference_dynamic_batching_max_tokens_override .. None + inference_dynamic_batching_num_cuda_graphs ...... 16 + inference_max_batch_size ........................ 8 + inference_max_seq_length ........................ 2560 + inference_rng_tracker ........................... False + init_method_std ................................. 0.01 + init_method_xavier_uniform ...................... False + init_model_with_meta_device ..................... False + initial_loss_scale .............................. 4294967296 + inprocess_active_world_size ..................... 8 + inprocess_barrier_timeout ....................... 120 + inprocess_completion_timeout .................... 120 + inprocess_empty_cuda_cache ...................... False + inprocess_granularity ........................... node + inprocess_hard_timeout .......................... 90 + inprocess_heartbeat_interval .................... 30 + inprocess_heartbeat_timeout ..................... 60 + inprocess_last_call_wait ........................ 1 + inprocess_max_iterations ........................ None + inprocess_monitor_process_interval .............. 1.0 + inprocess_monitor_thread_interval ............... 1.0 + inprocess_progress_watchdog_interval ............ 1.0 + inprocess_restart ............................... False + inprocess_soft_timeout .......................... 60 + inprocess_termination_grace_time ................ 1 + is_hybrid_model ................................. False + iter_per_epoch .................................. 1250 + iterations_to_skip .............................. [] + keep_fp8_transpose_cache_when_using_custom_fsdp . False + kitchen_config_file ............................. None + kitchen_recipe_number ........................... None + kv_channels ..................................... 128 + kv_lora_rank .................................... 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 .............................. 0.001 + moe_capacity_monitor ............................ True + moe_deepep_num_sms .............................. 20 + moe_echo_enable_random_offloading ............... False + moe_echo_expert_dispatch_overlap ................ False + moe_echo_expert_dispatcher_type ................. hybridep + moe_echo_recompute_expert_dispatch .............. False + moe_enable_deepep ............................... False + moe_enable_echo ................................. False + moe_enable_hybridep ............................. True + moe_enable_ultraep .............................. True + moe_expert_capacity_factor ...................... None + moe_extended_tp ................................. False + moe_ffn_hidden_size ............................. 768 + 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 .............. 16 + moe_pad_expert_input_to_capacity ................ False + moe_per_layer_logging ........................... False + moe_permute_fusion .............................. False + moe_received_token_capacity ..................... 1.11328125 + 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-06 + normalization ................................... RMSNorm + num_attention_heads ............................. 32 + num_channels .................................... 3 + num_classes ..................................... 1000 + num_dataset_builder_threads ..................... 1 + num_distributed_optimizer_instances ............. 1 + num_experts ..................................... 128 + num_layers ...................................... 48 + num_layers_at_end_in_bf16 ....................... 1 + num_layers_at_start_in_bf16 ..................... 1 + num_layers_per_virtual_pipeline_stage ........... None + num_query_groups ................................ 4 + num_virtual_stages_per_pipeline_rank ............ None + num_workers ..................................... 2 + object_storage_cache_path ....................... None + one_logger_async ................................ False + one_logger_project .............................. megatron-lm + one_logger_run_name ............................. None + onnx_safe ....................................... None + openai_gelu ..................................... False + optimizer ....................................... adam + optimizer_cpu_offload ........................... False + optimizer_offload_fraction ...................... 1.0 + output_bert_embeddings .......................... False + overlap_cpu_optimizer_d2h_h2d ................... False + overlap_grad_reduce ............................. True + overlap_moe_expert_parallel_comm ................ False + overlap_p2p_comm ................................ False + overlap_p2p_comm_warmup_flush ................... False + overlap_param_gather ............................ True + overlap_param_gather_with_optimizer_step ........ False + override_opt_param_scheduler .................... False + params_dtype .................................... torch.bfloat16 + patch_dim ....................................... 16 + per_split_data_args_path ........................ None + perform_initialization .......................... True + pin_cpu_grads ................................... True + pin_cpu_params .................................. True + pipeline_model_parallel_comm_backend ............ None + pipeline_model_parallel_layout .................. None + pipeline_model_parallel_size .................... 1 + position_embedding_type ......................... rope + pretrained_checkpoint ........................... None + profile ......................................... False + profile_ranks ................................... [0] + profile_step_end ................................ 12 + profile_step_start .............................. 10 + q_lora_rank ..................................... None + qk_head_dim ..................................... 128 + qk_l2_norm ...................................... False + qk_layernorm .................................... True + qk_pos_emb_head_dim ............................. 64 + query_in_block_prob ............................. 0.1 + rampup_batch_size ............................... None + rank ............................................ 0 + recompute_granularity ........................... None + recompute_method ................................ None + recompute_modules ............................... None + recompute_num_layers ............................ None + record_memory_history ........................... False + relative_attention_max_distance ................. 128 + relative_attention_num_buckets .................. 32 + replication ..................................... False + replication_factor .............................. 2 + replication_jump ................................ None + rerun_mode ...................................... validate_results + reset_attention_mask ............................ False + reset_position_ids .............................. False + result_rejected_tracker_filename ................ None + retriever_report_topk_accuracies ................ [] + retriever_score_scaling ......................... False + retriever_seq_length ............................ 256 + retro_add_retriever ............................. False + retro_attention_gate ............................ 1 + retro_cyclic_train_iters ........................ None + retro_encoder_attention_dropout ................. 0.1 + retro_encoder_hidden_dropout .................... 0.1 + retro_encoder_layers ............................ 2 + retro_num_neighbors ............................. 2 + retro_num_retrieved_chunks ...................... 2 + retro_project_dir ............................... None + retro_verify_neighbor_count ..................... True + reuse_grad_buf_for_mxfp8_param_ag ............... False + rope_scaling_factor ............................. 8.0 + rope_type ....................................... None + rotary_base ..................................... 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 ..................................... 12 + train_samples ................................... None + train_sync_interval ............................. None + transformer_impl ................................ transformer_engine + transformer_pipeline_model_parallel_size ........ 1 + untie_embeddings_and_output_weights ............. True + use_checkpoint_args ............................. False + use_checkpoint_opt_param_scheduler .............. False + use_cpu_initialization .......................... None + use_custom_fsdp ................................. False + use_dist_ckpt ................................... True + use_dist_ckpt_deprecated ........................ False + use_distributed_optimizer ....................... True + use_flash_attn .................................. False + use_legacy_models ............................... False + use_mp_args_from_checkpoint_args ................ False + use_one_sent_docs ............................... False + use_persistent_ckpt_worker ...................... False + use_precision_aware_optimizer ................... False + use_pytorch_profiler ............................ False + use_ring_exchange_p2p ........................... False + use_rope_scaling ................................ False + use_rotary_position_embeddings .................. False + use_sharp ....................................... False + use_tokenizer_model_from_checkpoint_args ........ True + use_torch_fsdp2 ................................. False + use_torch_optimizer_for_cpu_offload ............. False + use_tp_pp_dp_mapping ............................ False + v_head_dim ...................................... 128 + valid_data_path ................................. None + variable_seq_lengths ............................ False + virtual_pipeline_model_parallel_size ............ None + vision_backbone_type ............................ vit + vision_pretraining .............................. False + vision_pretraining_type ......................... classify + vocab_extra_ids ................................. 0 + vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json + vocab_size ...................................... None + wandb_exp_name .................................. + wandb_project ................................... + wandb_save_dir .................................. + weight_decay .................................... 0.1 + weight_decay_incr_style ......................... constant + wgrad_deferral_limit ............................ 0 + world_size ...................................... 8 + yaml_cfg ........................................ None +-------------------- end of arguments --------------------- +INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1 +> building GPT2BPETokenizer tokenizer ... +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +WARNING: one_logger package is required to enable e2e metrics tracking. please go to https://confluence.nvidia.com/display/MLWFO/Package+Repositories for details to install it + > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304) +WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS +> initializing torch distributed ... +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") +[Gloo] Rank 2 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 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 6 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 2 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 [Gloo] Rank [Gloo] Rank [Gloo] Rank [Gloo] Rank [Gloo] Rank 137546 is connected to is connected to is connected to is connected to is connected to is connected to 777777 peer ranks. peer ranks. peer ranks. peer ranks. peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : Expected number of connected peer ranks is : Expected number of connected peer ranks is : Expected number of connected peer ranks is : Expected number of connected peer ranks is : 777777 + + + + + +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +> initialized tensor model parallel with size 1 +> initialized pipeline model parallel with size 1 +> setting random seeds to 1234 ... +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : [Gloo] Rank 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 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[rank4]:[W826 18:31:34.510614802 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. +[rank1]:[W826 18:31:34.510718415 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank5]:[W826 18:31:34.510756245 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. +[rank2]:[W826 18:31:34.510779915 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 2] using GPU 2 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank6]:[W826 18:31:34.510846878 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 6] using GPU 6 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank7]:[W826 18:31:34.510953080 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. +[rank3]:[W826 18:31:34.517245505 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. +> 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: 0.035 seconds +WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations. +> compiling and loading fused kernels ... +[rank0]:[W826 18:31:34.680397160 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: 4.248 seconds +time to initialize megatron (seconds): 7.353 +[after megatron is initialized] datetime: 2026-08-26 18:31:40 +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): 8372566016 +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: 21 +Params for bucket 1 (103022592 elements, 103022592 padded size): + module.output_layer.weight +Params for bucket 2 (49033984 elements, 49033984 padded size): + module.decoder.layers.47.self_attention.linear_proj.weight + module.decoder.layers.47.self_attention.linear_qkv.weight + module.decoder.layers.46.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.45.mlp.router.weight + module.decoder.layers.45.self_attention.k_layernorm.weight + module.decoder.layers.45.pre_mlp_layernorm.weight + module.decoder.layers.45.self_attention.q_layernorm.weight + module.decoder.layers.47.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.46.mlp.router.weight + module.decoder.layers.46.self_attention.k_layernorm.weight + module.decoder.layers.46.pre_mlp_layernorm.weight + module.decoder.layers.46.self_attention.q_layernorm.weight + module.decoder.layers.45.self_attention.linear_qkv.weight + module.decoder.final_layernorm.weight + module.decoder.layers.47.self_attention.k_layernorm.weight + module.decoder.layers.47.mlp.router.weight + module.decoder.layers.46.self_attention.linear_proj.weight + module.decoder.layers.47.self_attention.q_layernorm.weight + module.decoder.layers.47.pre_mlp_layernorm.weight + module.decoder.layers.46.self_attention.linear_qkv.weight +Params for bucket 3 (46672384 elements, 46672384 padded size): + module.decoder.layers.45.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.45.self_attention.linear_proj.weight + module.decoder.layers.44.mlp.router.weight + module.decoder.layers.44.self_attention.k_layernorm.weight + module.decoder.layers.44.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.44.self_attention.linear_proj.weight + module.decoder.layers.43.mlp.router.weight + module.decoder.layers.43.self_attention.k_layernorm.weight + module.decoder.layers.43.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.43.self_attention.linear_proj.weight + module.decoder.layers.44.pre_mlp_layernorm.weight + module.decoder.layers.44.self_attention.q_layernorm.weight + module.decoder.layers.44.self_attention.linear_qkv.weight + module.decoder.layers.43.pre_mlp_layernorm.weight + module.decoder.layers.43.self_attention.q_layernorm.weight + module.decoder.layers.43.self_attention.linear_qkv.weight +Params for bucket 4 (49031936 elements, 49031936 padded size): + module.decoder.layers.42.mlp.router.weight + module.decoder.layers.42.pre_mlp_layernorm.weight + module.decoder.layers.41.self_attention.linear_qkv.weight + module.decoder.layers.42.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.42.self_attention.k_layernorm.weight + module.decoder.layers.41.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.41.self_attention.linear_proj.weight + module.decoder.layers.40.mlp.router.weight + module.decoder.layers.40.self_attention.k_layernorm.weight + module.decoder.layers.42.self_attention.q_layernorm.weight + module.decoder.layers.40.pre_mlp_layernorm.weight + module.decoder.layers.40.self_attention.q_layernorm.weight + module.decoder.layers.42.self_attention.linear_proj.weight + module.decoder.layers.41.mlp.router.weight + module.decoder.layers.41.self_attention.k_layernorm.weight + module.decoder.layers.42.self_attention.linear_qkv.weight + module.decoder.layers.41.pre_mlp_layernorm.weight + module.decoder.layers.41.self_attention.q_layernorm.weight + module.decoder.layers.40.self_attention.linear_qkv.weight +Params for bucket 5 (46672384 elements, 46672384 padded size): + module.decoder.layers.40.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.40.self_attention.linear_proj.weight + module.decoder.layers.39.mlp.router.weight + module.decoder.layers.39.self_attention.k_layernorm.weight + module.decoder.layers.39.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.39.self_attention.linear_proj.weight + module.decoder.layers.38.mlp.router.weight + module.decoder.layers.38.self_attention.k_layernorm.weight + module.decoder.layers.38.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.38.self_attention.linear_proj.weight + module.decoder.layers.39.pre_mlp_layernorm.weight + module.decoder.layers.39.self_attention.q_layernorm.weight + module.decoder.layers.39.self_attention.linear_qkv.weight + module.decoder.layers.38.pre_mlp_layernorm.weight + module.decoder.layers.38.self_attention.q_layernorm.weight + module.decoder.layers.38.self_attention.linear_qkv.weight +Params for bucket 6 (49031936 elements, 49031936 padded size): + module.decoder.layers.37.self_attention.linear_proj.weight + module.decoder.layers.37.self_attention.linear_qkv.weight + module.decoder.layers.36.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.35.mlp.router.weight + module.decoder.layers.35.self_attention.k_layernorm.weight + module.decoder.layers.35.pre_mlp_layernorm.weight + module.decoder.layers.35.self_attention.q_layernorm.weight + module.decoder.layers.37.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.36.mlp.router.weight + module.decoder.layers.36.self_attention.k_layernorm.weight + module.decoder.layers.36.pre_mlp_layernorm.weight + module.decoder.layers.36.self_attention.q_layernorm.weight + module.decoder.layers.35.self_attention.linear_qkv.weight + module.decoder.layers.37.mlp.router.weight + module.decoder.layers.37.self_attention.k_layernorm.weight + module.decoder.layers.36.self_attention.linear_proj.weight + module.decoder.layers.37.pre_mlp_layernorm.weight + module.decoder.layers.37.self_attention.q_layernorm.weight + module.decoder.layers.36.self_attention.linear_qkv.weight +Params for bucket 7 (46672384 elements, 46672384 padded size): + module.decoder.layers.35.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.34.self_attention.k_layernorm.weight + module.decoder.layers.35.self_attention.linear_proj.weight + module.decoder.layers.34.mlp.router.weight + module.decoder.layers.34.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.34.self_attention.linear_proj.weight + module.decoder.layers.33.mlp.router.weight + module.decoder.layers.33.self_attention.k_layernorm.weight + module.decoder.layers.33.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.33.self_attention.linear_proj.weight + module.decoder.layers.34.pre_mlp_layernorm.weight + module.decoder.layers.34.self_attention.q_layernorm.weight + module.decoder.layers.34.self_attention.linear_qkv.weight + module.decoder.layers.33.pre_mlp_layernorm.weight + module.decoder.layers.33.self_attention.q_layernorm.weight + module.decoder.layers.33.self_attention.linear_qkv.weight +Params for bucket 8 (49031936 elements, 49031936 padded size): + module.decoder.layers.31.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.30.mlp.router.weight + module.decoder.layers.30.self_attention.k_layernorm.weight + module.decoder.layers.30.pre_mlp_layernorm.weight + module.decoder.layers.30.self_attention.q_layernorm.weight + module.decoder.layers.32.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.31.mlp.router.weight + module.decoder.layers.31.self_attention.k_layernorm.weight + module.decoder.layers.31.pre_mlp_layernorm.weight + module.decoder.layers.31.self_attention.q_layernorm.weight + module.decoder.layers.30.self_attention.linear_qkv.weight + module.decoder.layers.32.mlp.router.weight + module.decoder.layers.32.self_attention.k_layernorm.weight + module.decoder.layers.31.self_attention.linear_proj.weight + module.decoder.layers.32.pre_mlp_layernorm.weight + module.decoder.layers.32.self_attention.q_layernorm.weight + module.decoder.layers.31.self_attention.linear_qkv.weight + module.decoder.layers.32.self_attention.linear_proj.weight + module.decoder.layers.32.self_attention.linear_qkv.weight +Params for bucket 9 (46672384 elements, 46672384 padded size): + module.decoder.layers.30.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.29.self_attention.k_layernorm.weight + module.decoder.layers.30.self_attention.linear_proj.weight + module.decoder.layers.29.mlp.router.weight + module.decoder.layers.29.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.29.self_attention.linear_proj.weight + module.decoder.layers.28.mlp.router.weight + module.decoder.layers.28.self_attention.k_layernorm.weight + module.decoder.layers.28.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.28.self_attention.linear_proj.weight + module.decoder.layers.29.pre_mlp_layernorm.weight + module.decoder.layers.29.self_attention.q_layernorm.weight + module.decoder.layers.29.self_attention.linear_qkv.weight + module.decoder.layers.28.pre_mlp_layernorm.weight + module.decoder.layers.28.self_attention.q_layernorm.weight + module.decoder.layers.28.self_attention.linear_qkv.weight +Params for bucket 10 (49031936 elements, 49031936 padded size): + module.decoder.layers.27.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.26.mlp.router.weight + module.decoder.layers.26.self_attention.k_layernorm.weight + module.decoder.layers.26.pre_mlp_layernorm.weight + module.decoder.layers.26.self_attention.q_layernorm.weight + module.decoder.layers.25.self_attention.linear_qkv.weight + module.decoder.layers.27.mlp.router.weight + module.decoder.layers.27.self_attention.k_layernorm.weight + module.decoder.layers.26.self_attention.linear_proj.weight + module.decoder.layers.27.self_attention.q_layernorm.weight + module.decoder.layers.27.pre_mlp_layernorm.weight + module.decoder.layers.26.self_attention.linear_qkv.weight + module.decoder.layers.27.self_attention.linear_proj.weight + module.decoder.layers.27.self_attention.linear_qkv.weight + module.decoder.layers.26.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.25.mlp.router.weight + module.decoder.layers.25.self_attention.k_layernorm.weight + module.decoder.layers.25.pre_mlp_layernorm.weight + module.decoder.layers.25.self_attention.q_layernorm.weight +Params for bucket 11 (46672384 elements, 46672384 padded size): + module.decoder.layers.25.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.25.self_attention.linear_proj.weight + module.decoder.layers.24.mlp.router.weight + module.decoder.layers.24.self_attention.k_layernorm.weight + module.decoder.layers.24.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.24.self_attention.linear_proj.weight + module.decoder.layers.23.pre_mlp_layernorm.weight + module.decoder.layers.23.self_attention.q_layernorm.weight + module.decoder.layers.23.self_attention.linear_qkv.weight + module.decoder.layers.24.pre_mlp_layernorm.weight + module.decoder.layers.24.self_attention.q_layernorm.weight + module.decoder.layers.24.self_attention.linear_qkv.weight + module.decoder.layers.23.mlp.router.weight + module.decoder.layers.23.self_attention.k_layernorm.weight + module.decoder.layers.23.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.23.self_attention.linear_proj.weight +Params for bucket 12 (49031936 elements, 49031936 padded size): + module.decoder.layers.22.self_attention.k_layernorm.weight + module.decoder.layers.20.pre_mlp_layernorm.weight + module.decoder.layers.20.self_attention.q_layernorm.weight + module.decoder.layers.22.pre_mlp_layernorm.weight + module.decoder.layers.22.self_attention.q_layernorm.weight + module.decoder.layers.21.self_attention.k_layernorm.weight + module.decoder.layers.22.self_attention.linear_proj.weight + module.decoder.layers.21.pre_mlp_layernorm.weight + module.decoder.layers.21.self_attention.q_layernorm.weight + module.decoder.layers.20.self_attention.linear_qkv.weight + module.decoder.layers.22.self_attention.linear_qkv.weight + module.decoder.layers.21.self_attention.linear_proj.weight + module.decoder.layers.21.self_attention.linear_qkv.weight + module.decoder.layers.22.mlp.router.weight + module.decoder.layers.22.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.21.mlp.router.weight + module.decoder.layers.21.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.20.mlp.router.weight + module.decoder.layers.20.self_attention.k_layernorm.weight +Params for bucket 13 (46672384 elements, 46672384 padded size): + module.decoder.layers.19.self_attention.k_layernorm.weight + module.decoder.layers.19.mlp.router.weight + module.decoder.layers.19.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.19.self_attention.linear_proj.weight + module.decoder.layers.18.mlp.router.weight + module.decoder.layers.18.self_attention.k_layernorm.weight + module.decoder.layers.18.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.18.self_attention.linear_proj.weight + module.decoder.layers.20.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.20.self_attention.linear_proj.weight + module.decoder.layers.19.pre_mlp_layernorm.weight + module.decoder.layers.19.self_attention.q_layernorm.weight + module.decoder.layers.19.self_attention.linear_qkv.weight + module.decoder.layers.18.pre_mlp_layernorm.weight + module.decoder.layers.18.self_attention.q_layernorm.weight + module.decoder.layers.18.self_attention.linear_qkv.weight +Params for bucket 14 (49031936 elements, 49031936 padded size): + module.decoder.layers.17.self_attention.linear_proj.weight + module.decoder.layers.17.self_attention.linear_qkv.weight + module.decoder.layers.16.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.15.mlp.router.weight + module.decoder.layers.15.self_attention.k_layernorm.weight + module.decoder.layers.15.pre_mlp_layernorm.weight + module.decoder.layers.15.self_attention.q_layernorm.weight + module.decoder.layers.17.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.16.mlp.router.weight + module.decoder.layers.16.self_attention.k_layernorm.weight + module.decoder.layers.16.pre_mlp_layernorm.weight + module.decoder.layers.16.self_attention.q_layernorm.weight + module.decoder.layers.15.self_attention.linear_qkv.weight + module.decoder.layers.17.mlp.router.weight + module.decoder.layers.17.self_attention.k_layernorm.weight + module.decoder.layers.16.self_attention.linear_proj.weight + module.decoder.layers.17.pre_mlp_layernorm.weight + module.decoder.layers.17.self_attention.q_layernorm.weight + module.decoder.layers.16.self_attention.linear_qkv.weight +Params for bucket 15 (46672384 elements, 46672384 padded size): + module.decoder.layers.15.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.14.pre_mlp_layernorm.weight + module.decoder.layers.15.self_attention.linear_proj.weight + module.decoder.layers.14.self_attention.q_layernorm.weight + module.decoder.layers.14.self_attention.linear_qkv.weight + module.decoder.layers.13.pre_mlp_layernorm.weight + module.decoder.layers.13.self_attention.q_layernorm.weight + module.decoder.layers.13.self_attention.linear_qkv.weight + module.decoder.layers.14.self_attention.k_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 + module.decoder.layers.13.mlp.router.weight + module.decoder.layers.13.self_attention.k_layernorm.weight + module.decoder.layers.13.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.13.self_attention.linear_proj.weight +Params for bucket 16 (49031936 elements, 49031936 padded size): + module.decoder.layers.12.self_attention.linear_proj.weight + module.decoder.layers.10.pre_mlp_layernorm.weight + module.decoder.layers.10.self_attention.q_layernorm.weight + module.decoder.layers.12.self_attention.linear_qkv.weight + module.decoder.layers.11.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.10.self_attention.linear_qkv.weight + module.decoder.layers.12.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.11.mlp.router.weight + module.decoder.layers.11.self_attention.k_layernorm.weight + module.decoder.layers.11.pre_mlp_layernorm.weight + module.decoder.layers.11.self_attention.q_layernorm.weight + module.decoder.layers.12.mlp.router.weight + module.decoder.layers.12.self_attention.k_layernorm.weight + module.decoder.layers.11.self_attention.linear_proj.weight + module.decoder.layers.12.pre_mlp_layernorm.weight + module.decoder.layers.12.self_attention.q_layernorm.weight + module.decoder.layers.11.self_attention.linear_qkv.weight + module.decoder.layers.10.mlp.router.weight + module.decoder.layers.10.self_attention.k_layernorm.weight +Params for bucket 17 (46672384 elements, 46672384 padded size): + module.decoder.layers.9.pre_mlp_layernorm.weight + module.decoder.layers.9.self_attention.q_layernorm.weight + module.decoder.layers.9.self_attention.linear_qkv.weight + module.decoder.layers.8.pre_mlp_layernorm.weight + module.decoder.layers.8.self_attention.q_layernorm.weight + module.decoder.layers.8.self_attention.linear_qkv.weight + module.decoder.layers.10.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.10.self_attention.linear_proj.weight + module.decoder.layers.9.mlp.router.weight + module.decoder.layers.9.self_attention.k_layernorm.weight + module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.9.self_attention.linear_proj.weight + module.decoder.layers.8.mlp.router.weight + module.decoder.layers.8.self_attention.k_layernorm.weight + module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.8.self_attention.linear_proj.weight +Params for bucket 18 (49031936 elements, 49031936 padded size): + module.decoder.layers.7.self_attention.linear_qkv.weight + module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.5.mlp.router.weight + module.decoder.layers.5.self_attention.k_layernorm.weight + module.decoder.layers.5.pre_mlp_layernorm.weight + module.decoder.layers.5.self_attention.q_layernorm.weight + module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.6.mlp.router.weight + module.decoder.layers.6.self_attention.k_layernorm.weight + module.decoder.layers.6.pre_mlp_layernorm.weight + module.decoder.layers.6.self_attention.q_layernorm.weight + module.decoder.layers.5.self_attention.linear_qkv.weight + module.decoder.layers.7.mlp.router.weight + module.decoder.layers.7.self_attention.k_layernorm.weight + module.decoder.layers.6.self_attention.linear_proj.weight + module.decoder.layers.7.self_attention.q_layernorm.weight + module.decoder.layers.7.pre_mlp_layernorm.weight + module.decoder.layers.6.self_attention.linear_qkv.weight + module.decoder.layers.7.self_attention.linear_proj.weight +Params for bucket 19 (46672384 elements, 46672384 padded size): + module.decoder.layers.4.self_attention.k_layernorm.weight + module.decoder.layers.5.self_attention.linear_proj.weight + 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.3.mlp.router.weight + module.decoder.layers.3.self_attention.k_layernorm.weight + module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.3.self_attention.linear_proj.weight + module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.4.pre_mlp_layernorm.weight + module.decoder.layers.4.self_attention.q_layernorm.weight + module.decoder.layers.4.self_attention.linear_qkv.weight + module.decoder.layers.3.pre_mlp_layernorm.weight + module.decoder.layers.3.self_attention.q_layernorm.weight + module.decoder.layers.3.self_attention.linear_qkv.weight +Params for bucket 20 (49031936 elements, 49031936 padded size): + module.decoder.layers.0.mlp.router.weight + module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.1.mlp.router.weight + module.decoder.layers.1.self_attention.k_layernorm.weight + module.decoder.layers.1.pre_mlp_layernorm.weight + module.decoder.layers.1.self_attention.q_layernorm.weight + module.decoder.layers.2.mlp.router.weight + module.decoder.layers.2.self_attention.k_layernorm.weight + module.decoder.layers.1.self_attention.linear_proj.weight + module.decoder.layers.2.self_attention.q_layernorm.weight + module.decoder.layers.2.pre_mlp_layernorm.weight + module.decoder.layers.1.self_attention.linear_qkv.weight + module.decoder.layers.0.self_attention.k_layernorm.weight + module.decoder.layers.2.self_attention.linear_proj.weight + module.decoder.layers.2.self_attention.linear_qkv.weight + module.decoder.layers.0.pre_mlp_layernorm.weight + module.decoder.layers.0.self_attention.q_layernorm.weight + module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.0.self_attention.linear_qkv.weight +Params for bucket 21 (111413248 elements, 111413248 padded size): + module.decoder.layers.0.self_attention.linear_qkv.layer_norm_weight + module.embedding.word_embeddings.weight + module.decoder.layers.0.self_attention.linear_proj.weight +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 89 +Params for bucket 1 (40894464 elements, 40894464 padded size): + module.decoder.layers.47.mlp.experts.linear_fc2.weight15 + module.decoder.layers.47.mlp.experts.linear_fc2.weight7 + module.decoder.layers.47.mlp.experts.linear_fc2.weight12 + module.decoder.layers.47.mlp.experts.linear_fc2.weight4 + module.decoder.layers.47.mlp.experts.linear_fc1.weight15 + module.decoder.layers.47.mlp.experts.linear_fc2.weight9 + module.decoder.layers.47.mlp.experts.linear_fc2.weight1 + module.decoder.layers.47.mlp.experts.linear_fc1.weight12 + module.decoder.layers.47.mlp.experts.linear_fc2.weight14 + module.decoder.layers.47.mlp.experts.linear_fc2.weight6 + module.decoder.layers.47.mlp.experts.linear_fc2.weight11 + module.decoder.layers.47.mlp.experts.linear_fc2.weight3 + module.decoder.layers.47.mlp.experts.linear_fc1.weight14 + module.decoder.layers.47.mlp.experts.linear_fc2.weight8 + module.decoder.layers.47.mlp.experts.linear_fc2.weight0 + module.decoder.layers.47.mlp.experts.linear_fc1.weight11 + module.decoder.layers.47.mlp.experts.linear_fc2.weight13 + module.decoder.layers.47.mlp.experts.linear_fc2.weight5 + module.decoder.layers.47.mlp.experts.linear_fc2.weight10 + module.decoder.layers.47.mlp.experts.linear_fc2.weight2 + module.decoder.layers.47.mlp.experts.linear_fc1.weight13 +Params for bucket 2 (40894464 elements, 40894464 padded size): + module.decoder.layers.47.mlp.experts.linear_fc1.weight10 + module.decoder.layers.47.mlp.experts.linear_fc1.weight6 + module.decoder.layers.47.mlp.experts.linear_fc1.weight8 + module.decoder.layers.47.mlp.experts.linear_fc1.weight4 + module.decoder.layers.47.mlp.experts.linear_fc1.weight2 + module.decoder.layers.47.mlp.experts.linear_fc1.weight0 + module.decoder.layers.46.mlp.experts.linear_fc2.weight15 + module.decoder.layers.46.mlp.experts.linear_fc2.weight13 + module.decoder.layers.47.mlp.experts.linear_fc1.weight9 + module.decoder.layers.47.mlp.experts.linear_fc1.weight7 + module.decoder.layers.47.mlp.experts.linear_fc1.weight5 + module.decoder.layers.47.mlp.experts.linear_fc1.weight3 + module.decoder.layers.47.mlp.experts.linear_fc1.weight1 + module.decoder.layers.46.mlp.experts.linear_fc2.weight14 + module.decoder.layers.46.mlp.experts.linear_fc2.weight12 +Params for bucket 3 (40894464 elements, 40894464 padded size): + module.decoder.layers.46.mlp.experts.linear_fc2.weight9 + module.decoder.layers.46.mlp.experts.linear_fc2.weight1 + module.decoder.layers.46.mlp.experts.linear_fc1.weight12 + module.decoder.layers.46.mlp.experts.linear_fc2.weight6 + module.decoder.layers.46.mlp.experts.linear_fc1.weight9 + module.decoder.layers.46.mlp.experts.linear_fc2.weight11 + module.decoder.layers.46.mlp.experts.linear_fc2.weight3 + module.decoder.layers.46.mlp.experts.linear_fc1.weight14 + module.decoder.layers.46.mlp.experts.linear_fc2.weight8 + module.decoder.layers.46.mlp.experts.linear_fc2.weight0 + module.decoder.layers.46.mlp.experts.linear_fc1.weight11 + module.decoder.layers.46.mlp.experts.linear_fc2.weight5 + module.decoder.layers.46.mlp.experts.linear_fc2.weight10 + module.decoder.layers.46.mlp.experts.linear_fc2.weight2 + module.decoder.layers.46.mlp.experts.linear_fc1.weight13 + module.decoder.layers.46.mlp.experts.linear_fc2.weight7 + module.decoder.layers.46.mlp.experts.linear_fc1.weight10 + module.decoder.layers.46.mlp.experts.linear_fc2.weight4 + module.decoder.layers.46.mlp.experts.linear_fc1.weight15 +Params for bucket 4 (40894464 elements, 40894464 padded size): + module.decoder.layers.46.mlp.experts.linear_fc1.weight8 + module.decoder.layers.46.mlp.experts.linear_fc1.weight6 + module.decoder.layers.46.mlp.experts.linear_fc1.weight4 + module.decoder.layers.46.mlp.experts.linear_fc1.weight2 + module.decoder.layers.46.mlp.experts.linear_fc1.weight0 + module.decoder.layers.45.mlp.experts.linear_fc2.weight15 + module.decoder.layers.45.mlp.experts.linear_fc2.weight13 + module.decoder.layers.45.mlp.experts.linear_fc2.weight11 + module.decoder.layers.45.mlp.experts.linear_fc2.weight9 + module.decoder.layers.46.mlp.experts.linear_fc1.weight7 + module.decoder.layers.46.mlp.experts.linear_fc1.weight5 + module.decoder.layers.46.mlp.experts.linear_fc1.weight3 + module.decoder.layers.46.mlp.experts.linear_fc1.weight1 + module.decoder.layers.45.mlp.experts.linear_fc2.weight14 + module.decoder.layers.45.mlp.experts.linear_fc2.weight12 + module.decoder.layers.45.mlp.experts.linear_fc2.weight10 + module.decoder.layers.45.mlp.experts.linear_fc2.weight8 +Params for bucket 5 (40894464 elements, 40894464 padded size): + module.decoder.layers.45.mlp.experts.linear_fc2.weight7 + module.decoder.layers.45.mlp.experts.linear_fc2.weight3 + module.decoder.layers.45.mlp.experts.linear_fc2.weight5 + module.decoder.layers.45.mlp.experts.linear_fc2.weight1 + module.decoder.layers.45.mlp.experts.linear_fc1.weight14 + module.decoder.layers.45.mlp.experts.linear_fc1.weight12 + module.decoder.layers.45.mlp.experts.linear_fc1.weight10 + module.decoder.layers.45.mlp.experts.linear_fc1.weight8 + module.decoder.layers.45.mlp.experts.linear_fc2.weight4 + module.decoder.layers.45.mlp.experts.linear_fc2.weight6 + module.decoder.layers.45.mlp.experts.linear_fc2.weight2 + module.decoder.layers.45.mlp.experts.linear_fc2.weight0 + module.decoder.layers.45.mlp.experts.linear_fc1.weight15 + module.decoder.layers.45.mlp.experts.linear_fc1.weight13 + module.decoder.layers.45.mlp.experts.linear_fc1.weight11 + module.decoder.layers.45.mlp.experts.linear_fc1.weight9 + module.decoder.layers.45.mlp.experts.linear_fc1.weight7 +Params for bucket 6 (40894464 elements, 40894464 padded size): + module.decoder.layers.45.mlp.experts.linear_fc1.weight6 + module.decoder.layers.44.mlp.experts.linear_fc2.weight13 + module.decoder.layers.44.mlp.experts.linear_fc2.weight5 + module.decoder.layers.45.mlp.experts.linear_fc1.weight3 + module.decoder.layers.44.mlp.experts.linear_fc2.weight10 + module.decoder.layers.45.mlp.experts.linear_fc1.weight0 + module.decoder.layers.44.mlp.experts.linear_fc2.weight15 + module.decoder.layers.44.mlp.experts.linear_fc2.weight7 + module.decoder.layers.45.mlp.experts.linear_fc1.weight5 + module.decoder.layers.44.mlp.experts.linear_fc2.weight12 + module.decoder.layers.44.mlp.experts.linear_fc2.weight4 + module.decoder.layers.45.mlp.experts.linear_fc1.weight2 + module.decoder.layers.44.mlp.experts.linear_fc2.weight9 + module.decoder.layers.44.mlp.experts.linear_fc2.weight14 + module.decoder.layers.44.mlp.experts.linear_fc2.weight6 + module.decoder.layers.45.mlp.experts.linear_fc1.weight4 + module.decoder.layers.44.mlp.experts.linear_fc2.weight11 + module.decoder.layers.45.mlp.experts.linear_fc1.weight1 + module.decoder.layers.44.mlp.experts.linear_fc2.weight8 +Params for bucket 7 (40894464 elements, 40894464 padded size): + module.decoder.layers.44.mlp.experts.linear_fc2.weight3 + module.decoder.layers.44.mlp.experts.linear_fc2.weight1 + module.decoder.layers.44.mlp.experts.linear_fc1.weight14 + module.decoder.layers.44.mlp.experts.linear_fc1.weight12 + module.decoder.layers.44.mlp.experts.linear_fc1.weight10 + module.decoder.layers.44.mlp.experts.linear_fc1.weight8 + module.decoder.layers.44.mlp.experts.linear_fc1.weight6 + module.decoder.layers.44.mlp.experts.linear_fc2.weight0 + module.decoder.layers.44.mlp.experts.linear_fc2.weight2 + module.decoder.layers.44.mlp.experts.linear_fc1.weight15 + module.decoder.layers.44.mlp.experts.linear_fc1.weight13 + module.decoder.layers.44.mlp.experts.linear_fc1.weight11 + module.decoder.layers.44.mlp.experts.linear_fc1.weight9 + module.decoder.layers.44.mlp.experts.linear_fc1.weight7 + module.decoder.layers.44.mlp.experts.linear_fc1.weight5 +Params for bucket 8 (40894464 elements, 40894464 padded size): + module.decoder.layers.44.mlp.experts.linear_fc1.weight0 + module.decoder.layers.43.mlp.experts.linear_fc2.weight15 + module.decoder.layers.43.mlp.experts.linear_fc2.weight7 + module.decoder.layers.43.mlp.experts.linear_fc2.weight12 + module.decoder.layers.43.mlp.experts.linear_fc2.weight4 + module.decoder.layers.44.mlp.experts.linear_fc1.weight2 + module.decoder.layers.43.mlp.experts.linear_fc2.weight9 + module.decoder.layers.43.mlp.experts.linear_fc2.weight1 + module.decoder.layers.43.mlp.experts.linear_fc2.weight14 + module.decoder.layers.43.mlp.experts.linear_fc2.weight6 + module.decoder.layers.44.mlp.experts.linear_fc1.weight4 + module.decoder.layers.43.mlp.experts.linear_fc2.weight11 + module.decoder.layers.43.mlp.experts.linear_fc2.weight3 + module.decoder.layers.44.mlp.experts.linear_fc1.weight1 + module.decoder.layers.43.mlp.experts.linear_fc2.weight8 + module.decoder.layers.43.mlp.experts.linear_fc2.weight0 + module.decoder.layers.43.mlp.experts.linear_fc2.weight13 + module.decoder.layers.43.mlp.experts.linear_fc2.weight5 + module.decoder.layers.44.mlp.experts.linear_fc1.weight3 + module.decoder.layers.43.mlp.experts.linear_fc2.weight10 + module.decoder.layers.43.mlp.experts.linear_fc2.weight2 +Params for bucket 9 (40894464 elements, 40894464 padded size): + module.decoder.layers.43.mlp.experts.linear_fc1.weight14 + module.decoder.layers.43.mlp.experts.linear_fc1.weight12 + module.decoder.layers.43.mlp.experts.linear_fc1.weight10 + module.decoder.layers.43.mlp.experts.linear_fc1.weight8 + module.decoder.layers.43.mlp.experts.linear_fc1.weight6 + module.decoder.layers.43.mlp.experts.linear_fc1.weight4 + module.decoder.layers.43.mlp.experts.linear_fc1.weight15 + module.decoder.layers.43.mlp.experts.linear_fc1.weight11 + module.decoder.layers.43.mlp.experts.linear_fc1.weight13 + module.decoder.layers.43.mlp.experts.linear_fc1.weight9 + module.decoder.layers.43.mlp.experts.linear_fc1.weight7 + module.decoder.layers.43.mlp.experts.linear_fc1.weight5 + module.decoder.layers.43.mlp.experts.linear_fc1.weight3 +Params for bucket 10 (40894464 elements, 40894464 padded size): + module.decoder.layers.43.mlp.experts.linear_fc1.weight2 + module.decoder.layers.42.mlp.experts.linear_fc2.weight9 + module.decoder.layers.42.mlp.experts.linear_fc2.weight1 + module.decoder.layers.42.mlp.experts.linear_fc2.weight14 + module.decoder.layers.42.mlp.experts.linear_fc2.weight6 + module.decoder.layers.42.mlp.experts.linear_fc2.weight11 + module.decoder.layers.42.mlp.experts.linear_fc2.weight3 + module.decoder.layers.42.mlp.experts.linear_fc1.weight14 + module.decoder.layers.43.mlp.experts.linear_fc1.weight1 + module.decoder.layers.42.mlp.experts.linear_fc2.weight8 + module.decoder.layers.42.mlp.experts.linear_fc2.weight0 + module.decoder.layers.42.mlp.experts.linear_fc2.weight13 + module.decoder.layers.42.mlp.experts.linear_fc2.weight5 + module.decoder.layers.42.mlp.experts.linear_fc2.weight10 + module.decoder.layers.42.mlp.experts.linear_fc2.weight2 + module.decoder.layers.42.mlp.experts.linear_fc2.weight15 + module.decoder.layers.43.mlp.experts.linear_fc1.weight0 + module.decoder.layers.42.mlp.experts.linear_fc2.weight7 + module.decoder.layers.42.mlp.experts.linear_fc2.weight12 + module.decoder.layers.42.mlp.experts.linear_fc2.weight4 + module.decoder.layers.42.mlp.experts.linear_fc1.weight15 +Params for bucket 11 (40894464 elements, 40894464 padded size): + module.decoder.layers.42.mlp.experts.linear_fc1.weight12 + module.decoder.layers.42.mlp.experts.linear_fc1.weight10 + module.decoder.layers.42.mlp.experts.linear_fc1.weight8 + module.decoder.layers.42.mlp.experts.linear_fc1.weight6 + module.decoder.layers.42.mlp.experts.linear_fc1.weight4 + module.decoder.layers.42.mlp.experts.linear_fc1.weight2 + module.decoder.layers.42.mlp.experts.linear_fc1.weight13 + module.decoder.layers.42.mlp.experts.linear_fc1.weight11 + module.decoder.layers.42.mlp.experts.linear_fc1.weight9 + module.decoder.layers.42.mlp.experts.linear_fc1.weight7 + module.decoder.layers.42.mlp.experts.linear_fc1.weight5 + module.decoder.layers.42.mlp.experts.linear_fc1.weight3 + module.decoder.layers.42.mlp.experts.linear_fc1.weight1 +Params for bucket 12 (40894464 elements, 40894464 padded size): + module.decoder.layers.41.mlp.experts.linear_fc2.weight11 + module.decoder.layers.41.mlp.experts.linear_fc2.weight3 + module.decoder.layers.41.mlp.experts.linear_fc1.weight14 + module.decoder.layers.41.mlp.experts.linear_fc2.weight8 + module.decoder.layers.41.mlp.experts.linear_fc2.weight0 + module.decoder.layers.41.mlp.experts.linear_fc2.weight13 + module.decoder.layers.41.mlp.experts.linear_fc2.weight5 + module.decoder.layers.41.mlp.experts.linear_fc2.weight10 + module.decoder.layers.41.mlp.experts.linear_fc2.weight2 + module.decoder.layers.41.mlp.experts.linear_fc1.weight13 + module.decoder.layers.42.mlp.experts.linear_fc1.weight0 + module.decoder.layers.41.mlp.experts.linear_fc2.weight15 + module.decoder.layers.41.mlp.experts.linear_fc2.weight7 + module.decoder.layers.41.mlp.experts.linear_fc2.weight12 + module.decoder.layers.41.mlp.experts.linear_fc2.weight4 + module.decoder.layers.41.mlp.experts.linear_fc1.weight15 + module.decoder.layers.41.mlp.experts.linear_fc2.weight9 + module.decoder.layers.41.mlp.experts.linear_fc2.weight1 + module.decoder.layers.41.mlp.experts.linear_fc1.weight12 + module.decoder.layers.41.mlp.experts.linear_fc2.weight14 + module.decoder.layers.41.mlp.experts.linear_fc2.weight6 +Params for bucket 13 (40894464 elements, 40894464 padded size): + module.decoder.layers.41.mlp.experts.linear_fc1.weight8 + module.decoder.layers.41.mlp.experts.linear_fc1.weight10 + module.decoder.layers.41.mlp.experts.linear_fc1.weight6 + module.decoder.layers.41.mlp.experts.linear_fc1.weight4 + module.decoder.layers.41.mlp.experts.linear_fc1.weight2 + module.decoder.layers.41.mlp.experts.linear_fc1.weight0 + module.decoder.layers.40.mlp.experts.linear_fc2.weight15 + module.decoder.layers.41.mlp.experts.linear_fc1.weight11 + module.decoder.layers.41.mlp.experts.linear_fc1.weight7 + module.decoder.layers.41.mlp.experts.linear_fc1.weight9 + module.decoder.layers.41.mlp.experts.linear_fc1.weight5 + module.decoder.layers.41.mlp.experts.linear_fc1.weight3 + module.decoder.layers.41.mlp.experts.linear_fc1.weight1 + module.decoder.layers.40.mlp.experts.linear_fc2.weight14 +Params for bucket 14 (40894464 elements, 40894464 padded size): + module.decoder.layers.40.mlp.experts.linear_fc2.weight13 + module.decoder.layers.40.mlp.experts.linear_fc2.weight5 + module.decoder.layers.40.mlp.experts.linear_fc2.weight10 + module.decoder.layers.40.mlp.experts.linear_fc2.weight2 + module.decoder.layers.40.mlp.experts.linear_fc1.weight13 + module.decoder.layers.40.mlp.experts.linear_fc2.weight7 + module.decoder.layers.40.mlp.experts.linear_fc1.weight10 + module.decoder.layers.40.mlp.experts.linear_fc2.weight12 + module.decoder.layers.40.mlp.experts.linear_fc2.weight4 + module.decoder.layers.40.mlp.experts.linear_fc1.weight15 + module.decoder.layers.40.mlp.experts.linear_fc2.weight9 + module.decoder.layers.40.mlp.experts.linear_fc2.weight1 + module.decoder.layers.40.mlp.experts.linear_fc1.weight12 + module.decoder.layers.40.mlp.experts.linear_fc2.weight6 + module.decoder.layers.40.mlp.experts.linear_fc2.weight11 + module.decoder.layers.40.mlp.experts.linear_fc2.weight3 + module.decoder.layers.40.mlp.experts.linear_fc1.weight14 + module.decoder.layers.40.mlp.experts.linear_fc2.weight8 + module.decoder.layers.40.mlp.experts.linear_fc2.weight0 + module.decoder.layers.40.mlp.experts.linear_fc1.weight11 +Params for bucket 15 (40894464 elements, 40894464 padded size): + module.decoder.layers.40.mlp.experts.linear_fc1.weight8 + module.decoder.layers.40.mlp.experts.linear_fc1.weight6 + module.decoder.layers.40.mlp.experts.linear_fc1.weight4 + module.decoder.layers.40.mlp.experts.linear_fc1.weight2 + module.decoder.layers.40.mlp.experts.linear_fc1.weight0 + module.decoder.layers.39.mlp.experts.linear_fc2.weight15 + module.decoder.layers.39.mlp.experts.linear_fc2.weight13 + module.decoder.layers.39.mlp.experts.linear_fc2.weight11 + module.decoder.layers.40.mlp.experts.linear_fc1.weight9 + module.decoder.layers.40.mlp.experts.linear_fc1.weight5 + module.decoder.layers.40.mlp.experts.linear_fc1.weight7 + module.decoder.layers.40.mlp.experts.linear_fc1.weight3 + module.decoder.layers.40.mlp.experts.linear_fc1.weight1 + module.decoder.layers.39.mlp.experts.linear_fc2.weight14 + module.decoder.layers.39.mlp.experts.linear_fc2.weight12 + module.decoder.layers.39.mlp.experts.linear_fc2.weight10 +Params for bucket 16 (40894464 elements, 40894464 padded size): + module.decoder.layers.39.mlp.experts.linear_fc2.weight9 + module.decoder.layers.39.mlp.experts.linear_fc2.weight7 + module.decoder.layers.39.mlp.experts.linear_fc2.weight5 + module.decoder.layers.39.mlp.experts.linear_fc2.weight3 + module.decoder.layers.39.mlp.experts.linear_fc2.weight1 + module.decoder.layers.39.mlp.experts.linear_fc1.weight14 + module.decoder.layers.39.mlp.experts.linear_fc1.weight12 + module.decoder.layers.39.mlp.experts.linear_fc1.weight10 + module.decoder.layers.39.mlp.experts.linear_fc1.weight8 + module.decoder.layers.39.mlp.experts.linear_fc2.weight6 + module.decoder.layers.39.mlp.experts.linear_fc2.weight8 + module.decoder.layers.39.mlp.experts.linear_fc2.weight4 + module.decoder.layers.39.mlp.experts.linear_fc2.weight2 + module.decoder.layers.39.mlp.experts.linear_fc2.weight0 + module.decoder.layers.39.mlp.experts.linear_fc1.weight15 + module.decoder.layers.39.mlp.experts.linear_fc1.weight13 + module.decoder.layers.39.mlp.experts.linear_fc1.weight11 + module.decoder.layers.39.mlp.experts.linear_fc1.weight9 +Params for bucket 17 (40894464 elements, 40894464 padded size): + module.decoder.layers.39.mlp.experts.linear_fc1.weight6 + module.decoder.layers.39.mlp.experts.linear_fc1.weight4 + module.decoder.layers.39.mlp.experts.linear_fc1.weight2 + module.decoder.layers.39.mlp.experts.linear_fc1.weight0 + module.decoder.layers.38.mlp.experts.linear_fc2.weight15 + module.decoder.layers.38.mlp.experts.linear_fc2.weight13 + module.decoder.layers.38.mlp.experts.linear_fc2.weight11 + module.decoder.layers.38.mlp.experts.linear_fc2.weight9 + module.decoder.layers.38.mlp.experts.linear_fc2.weight7 + module.decoder.layers.39.mlp.experts.linear_fc1.weight7 + module.decoder.layers.39.mlp.experts.linear_fc1.weight5 + module.decoder.layers.39.mlp.experts.linear_fc1.weight3 + module.decoder.layers.39.mlp.experts.linear_fc1.weight1 + module.decoder.layers.38.mlp.experts.linear_fc2.weight14 + module.decoder.layers.38.mlp.experts.linear_fc2.weight12 + module.decoder.layers.38.mlp.experts.linear_fc2.weight10 + module.decoder.layers.38.mlp.experts.linear_fc2.weight8 + module.decoder.layers.38.mlp.experts.linear_fc2.weight6 +Params for bucket 18 (40894464 elements, 40894464 padded size): + module.decoder.layers.38.mlp.experts.linear_fc2.weight5 + module.decoder.layers.38.mlp.experts.linear_fc2.weight3 + module.decoder.layers.38.mlp.experts.linear_fc2.weight1 + module.decoder.layers.38.mlp.experts.linear_fc1.weight14 + module.decoder.layers.38.mlp.experts.linear_fc1.weight12 + module.decoder.layers.38.mlp.experts.linear_fc1.weight10 + module.decoder.layers.38.mlp.experts.linear_fc1.weight8 + module.decoder.layers.38.mlp.experts.linear_fc1.weight6 + module.decoder.layers.38.mlp.experts.linear_fc2.weight4 + module.decoder.layers.38.mlp.experts.linear_fc2.weight2 + module.decoder.layers.38.mlp.experts.linear_fc2.weight0 + module.decoder.layers.38.mlp.experts.linear_fc1.weight15 + module.decoder.layers.38.mlp.experts.linear_fc1.weight13 + module.decoder.layers.38.mlp.experts.linear_fc1.weight11 + module.decoder.layers.38.mlp.experts.linear_fc1.weight9 + module.decoder.layers.38.mlp.experts.linear_fc1.weight7 +Params for bucket 19 (40894464 elements, 40894464 padded size): + module.decoder.layers.38.mlp.experts.linear_fc1.weight4 + module.decoder.layers.37.mlp.experts.linear_fc2.weight11 + module.decoder.layers.37.mlp.experts.linear_fc2.weight3 + module.decoder.layers.38.mlp.experts.linear_fc1.weight1 + module.decoder.layers.37.mlp.experts.linear_fc2.weight8 + module.decoder.layers.37.mlp.experts.linear_fc2.weight13 + module.decoder.layers.37.mlp.experts.linear_fc2.weight5 + module.decoder.layers.38.mlp.experts.linear_fc1.weight3 + module.decoder.layers.37.mlp.experts.linear_fc2.weight10 + module.decoder.layers.37.mlp.experts.linear_fc2.weight2 + module.decoder.layers.38.mlp.experts.linear_fc1.weight0 + module.decoder.layers.37.mlp.experts.linear_fc2.weight15 + module.decoder.layers.37.mlp.experts.linear_fc2.weight7 + module.decoder.layers.38.mlp.experts.linear_fc1.weight5 + module.decoder.layers.37.mlp.experts.linear_fc2.weight12 + module.decoder.layers.37.mlp.experts.linear_fc2.weight4 + module.decoder.layers.38.mlp.experts.linear_fc1.weight2 + module.decoder.layers.37.mlp.experts.linear_fc2.weight9 + module.decoder.layers.37.mlp.experts.linear_fc2.weight14 + module.decoder.layers.37.mlp.experts.linear_fc2.weight6 +Params for bucket 20 (40894464 elements, 40894464 padded size): + module.decoder.layers.37.mlp.experts.linear_fc2.weight1 + module.decoder.layers.37.mlp.experts.linear_fc1.weight14 + module.decoder.layers.37.mlp.experts.linear_fc1.weight12 + module.decoder.layers.37.mlp.experts.linear_fc1.weight10 + module.decoder.layers.37.mlp.experts.linear_fc1.weight8 + module.decoder.layers.37.mlp.experts.linear_fc1.weight6 + module.decoder.layers.37.mlp.experts.linear_fc1.weight4 + module.decoder.layers.37.mlp.experts.linear_fc2.weight0 + module.decoder.layers.37.mlp.experts.linear_fc1.weight13 + module.decoder.layers.37.mlp.experts.linear_fc1.weight15 + module.decoder.layers.37.mlp.experts.linear_fc1.weight11 + module.decoder.layers.37.mlp.experts.linear_fc1.weight9 + module.decoder.layers.37.mlp.experts.linear_fc1.weight7 + module.decoder.layers.37.mlp.experts.linear_fc1.weight5 +Params for bucket 21 (40894464 elements, 40894464 padded size): + module.decoder.layers.36.mlp.experts.linear_fc2.weight13 + module.decoder.layers.36.mlp.experts.linear_fc2.weight5 + module.decoder.layers.37.mlp.experts.linear_fc1.weight3 + module.decoder.layers.36.mlp.experts.linear_fc2.weight10 + module.decoder.layers.36.mlp.experts.linear_fc2.weight2 + module.decoder.layers.37.mlp.experts.linear_fc1.weight0 + module.decoder.layers.36.mlp.experts.linear_fc2.weight15 + module.decoder.layers.36.mlp.experts.linear_fc2.weight7 + module.decoder.layers.36.mlp.experts.linear_fc2.weight12 + module.decoder.layers.36.mlp.experts.linear_fc2.weight4 + module.decoder.layers.36.mlp.experts.linear_fc1.weight15 + module.decoder.layers.37.mlp.experts.linear_fc1.weight2 + module.decoder.layers.36.mlp.experts.linear_fc2.weight9 + module.decoder.layers.36.mlp.experts.linear_fc2.weight1 + module.decoder.layers.36.mlp.experts.linear_fc2.weight14 + module.decoder.layers.36.mlp.experts.linear_fc2.weight6 + module.decoder.layers.36.mlp.experts.linear_fc2.weight11 + module.decoder.layers.36.mlp.experts.linear_fc2.weight3 + module.decoder.layers.37.mlp.experts.linear_fc1.weight1 + module.decoder.layers.36.mlp.experts.linear_fc2.weight8 + module.decoder.layers.36.mlp.experts.linear_fc2.weight0 +Params for bucket 22 (40894464 elements, 40894464 padded size): + module.decoder.layers.36.mlp.experts.linear_fc1.weight14 + module.decoder.layers.36.mlp.experts.linear_fc1.weight10 + module.decoder.layers.36.mlp.experts.linear_fc1.weight12 + module.decoder.layers.36.mlp.experts.linear_fc1.weight8 + module.decoder.layers.36.mlp.experts.linear_fc1.weight6 + module.decoder.layers.36.mlp.experts.linear_fc1.weight4 + module.decoder.layers.36.mlp.experts.linear_fc1.weight2 + module.decoder.layers.36.mlp.experts.linear_fc1.weight13 + module.decoder.layers.36.mlp.experts.linear_fc1.weight11 + module.decoder.layers.36.mlp.experts.linear_fc1.weight9 + module.decoder.layers.36.mlp.experts.linear_fc1.weight7 + module.decoder.layers.36.mlp.experts.linear_fc1.weight5 + module.decoder.layers.36.mlp.experts.linear_fc1.weight3 +Params for bucket 23 (40894464 elements, 40894464 padded size): + module.decoder.layers.36.mlp.experts.linear_fc1.weight0 + module.decoder.layers.35.mlp.experts.linear_fc2.weight15 + module.decoder.layers.35.mlp.experts.linear_fc2.weight7 + module.decoder.layers.35.mlp.experts.linear_fc2.weight12 + module.decoder.layers.35.mlp.experts.linear_fc2.weight4 + module.decoder.layers.35.mlp.experts.linear_fc1.weight15 + module.decoder.layers.35.mlp.experts.linear_fc2.weight9 + module.decoder.layers.35.mlp.experts.linear_fc2.weight1 + module.decoder.layers.35.mlp.experts.linear_fc2.weight14 + module.decoder.layers.35.mlp.experts.linear_fc2.weight6 + module.decoder.layers.35.mlp.experts.linear_fc2.weight11 + module.decoder.layers.35.mlp.experts.linear_fc2.weight3 + module.decoder.layers.35.mlp.experts.linear_fc1.weight14 + module.decoder.layers.36.mlp.experts.linear_fc1.weight1 + module.decoder.layers.35.mlp.experts.linear_fc2.weight8 + module.decoder.layers.35.mlp.experts.linear_fc2.weight0 + module.decoder.layers.35.mlp.experts.linear_fc2.weight13 + module.decoder.layers.35.mlp.experts.linear_fc2.weight5 + module.decoder.layers.35.mlp.experts.linear_fc2.weight10 + module.decoder.layers.35.mlp.experts.linear_fc2.weight2 + module.decoder.layers.35.mlp.experts.linear_fc1.weight13 +Params for bucket 24 (40894464 elements, 40894464 padded size): + module.decoder.layers.35.mlp.experts.linear_fc1.weight12 + module.decoder.layers.35.mlp.experts.linear_fc1.weight10 + module.decoder.layers.35.mlp.experts.linear_fc1.weight8 + module.decoder.layers.35.mlp.experts.linear_fc1.weight6 + module.decoder.layers.35.mlp.experts.linear_fc1.weight4 + module.decoder.layers.35.mlp.experts.linear_fc1.weight2 + module.decoder.layers.35.mlp.experts.linear_fc1.weight0 + module.decoder.layers.35.mlp.experts.linear_fc1.weight11 + module.decoder.layers.35.mlp.experts.linear_fc1.weight9 + module.decoder.layers.35.mlp.experts.linear_fc1.weight7 + module.decoder.layers.35.mlp.experts.linear_fc1.weight5 + module.decoder.layers.35.mlp.experts.linear_fc1.weight3 + module.decoder.layers.35.mlp.experts.linear_fc1.weight1 +Params for bucket 25 (40894464 elements, 40894464 padded size): + module.decoder.layers.34.mlp.experts.linear_fc2.weight9 + module.decoder.layers.34.mlp.experts.linear_fc2.weight1 + module.decoder.layers.34.mlp.experts.linear_fc1.weight12 + module.decoder.layers.34.mlp.experts.linear_fc2.weight6 + module.decoder.layers.34.mlp.experts.linear_fc2.weight15 + module.decoder.layers.34.mlp.experts.linear_fc2.weight11 + module.decoder.layers.34.mlp.experts.linear_fc2.weight3 + module.decoder.layers.34.mlp.experts.linear_fc1.weight14 + module.decoder.layers.34.mlp.experts.linear_fc2.weight8 + module.decoder.layers.34.mlp.experts.linear_fc2.weight0 + module.decoder.layers.34.mlp.experts.linear_fc1.weight11 + module.decoder.layers.34.mlp.experts.linear_fc2.weight13 + module.decoder.layers.34.mlp.experts.linear_fc2.weight5 + module.decoder.layers.34.mlp.experts.linear_fc2.weight14 + module.decoder.layers.34.mlp.experts.linear_fc2.weight10 + module.decoder.layers.34.mlp.experts.linear_fc2.weight2 + module.decoder.layers.34.mlp.experts.linear_fc1.weight13 + module.decoder.layers.34.mlp.experts.linear_fc2.weight7 + module.decoder.layers.34.mlp.experts.linear_fc2.weight12 + module.decoder.layers.34.mlp.experts.linear_fc2.weight4 + module.decoder.layers.34.mlp.experts.linear_fc1.weight15 +Params for bucket 26 (40894464 elements, 40894464 padded size): + module.decoder.layers.34.mlp.experts.linear_fc1.weight10 + module.decoder.layers.34.mlp.experts.linear_fc1.weight8 + module.decoder.layers.34.mlp.experts.linear_fc1.weight6 + module.decoder.layers.34.mlp.experts.linear_fc1.weight4 + module.decoder.layers.34.mlp.experts.linear_fc1.weight2 + module.decoder.layers.34.mlp.experts.linear_fc1.weight0 + module.decoder.layers.33.mlp.experts.linear_fc2.weight15 + module.decoder.layers.33.mlp.experts.linear_fc2.weight13 + module.decoder.layers.34.mlp.experts.linear_fc1.weight7 + module.decoder.layers.34.mlp.experts.linear_fc1.weight9 + module.decoder.layers.34.mlp.experts.linear_fc1.weight5 + module.decoder.layers.34.mlp.experts.linear_fc1.weight3 + module.decoder.layers.34.mlp.experts.linear_fc1.weight1 + module.decoder.layers.33.mlp.experts.linear_fc2.weight14 + module.decoder.layers.33.mlp.experts.linear_fc2.weight12 +Params for bucket 27 (40894464 elements, 40894464 padded size): + module.decoder.layers.33.mlp.experts.linear_fc2.weight3 + module.decoder.layers.33.mlp.experts.linear_fc1.weight14 + module.decoder.layers.33.mlp.experts.linear_fc2.weight0 + module.decoder.layers.33.mlp.experts.linear_fc1.weight11 + module.decoder.layers.33.mlp.experts.linear_fc2.weight9 + module.decoder.layers.33.mlp.experts.linear_fc2.weight5 + module.decoder.layers.33.mlp.experts.linear_fc2.weight2 + module.decoder.layers.33.mlp.experts.linear_fc1.weight13 + module.decoder.layers.33.mlp.experts.linear_fc2.weight11 + module.decoder.layers.33.mlp.experts.linear_fc2.weight7 + module.decoder.layers.33.mlp.experts.linear_fc1.weight10 + module.decoder.layers.33.mlp.experts.linear_fc2.weight8 + module.decoder.layers.33.mlp.experts.linear_fc2.weight4 + module.decoder.layers.33.mlp.experts.linear_fc1.weight15 + module.decoder.layers.33.mlp.experts.linear_fc2.weight1 + module.decoder.layers.33.mlp.experts.linear_fc1.weight12 + module.decoder.layers.33.mlp.experts.linear_fc2.weight10 + module.decoder.layers.33.mlp.experts.linear_fc2.weight6 + module.decoder.layers.33.mlp.experts.linear_fc1.weight9 +Params for bucket 28 (40894464 elements, 40894464 padded size): + module.decoder.layers.33.mlp.experts.linear_fc1.weight8 + module.decoder.layers.33.mlp.experts.linear_fc1.weight4 + module.decoder.layers.33.mlp.experts.linear_fc1.weight6 + module.decoder.layers.33.mlp.experts.linear_fc1.weight2 + module.decoder.layers.33.mlp.experts.linear_fc1.weight0 + module.decoder.layers.32.mlp.experts.linear_fc2.weight15 + module.decoder.layers.32.mlp.experts.linear_fc2.weight13 + module.decoder.layers.32.mlp.experts.linear_fc2.weight11 + module.decoder.layers.32.mlp.experts.linear_fc2.weight9 + module.decoder.layers.33.mlp.experts.linear_fc1.weight5 + module.decoder.layers.33.mlp.experts.linear_fc1.weight7 + module.decoder.layers.33.mlp.experts.linear_fc1.weight3 + module.decoder.layers.33.mlp.experts.linear_fc1.weight1 + module.decoder.layers.32.mlp.experts.linear_fc2.weight14 + module.decoder.layers.32.mlp.experts.linear_fc2.weight12 + module.decoder.layers.32.mlp.experts.linear_fc2.weight10 + module.decoder.layers.32.mlp.experts.linear_fc2.weight8 +Params for bucket 29 (40894464 elements, 40894464 padded size): + module.decoder.layers.32.mlp.experts.linear_fc2.weight7 + module.decoder.layers.32.mlp.experts.linear_fc2.weight5 + module.decoder.layers.32.mlp.experts.linear_fc2.weight3 + module.decoder.layers.32.mlp.experts.linear_fc2.weight1 + module.decoder.layers.32.mlp.experts.linear_fc1.weight14 + module.decoder.layers.32.mlp.experts.linear_fc1.weight12 + module.decoder.layers.32.mlp.experts.linear_fc1.weight10 + module.decoder.layers.32.mlp.experts.linear_fc1.weight8 + module.decoder.layers.32.mlp.experts.linear_fc2.weight4 + module.decoder.layers.32.mlp.experts.linear_fc2.weight6 + module.decoder.layers.32.mlp.experts.linear_fc2.weight2 + module.decoder.layers.32.mlp.experts.linear_fc2.weight0 + module.decoder.layers.32.mlp.experts.linear_fc1.weight15 + module.decoder.layers.32.mlp.experts.linear_fc1.weight13 + module.decoder.layers.32.mlp.experts.linear_fc1.weight11 + module.decoder.layers.32.mlp.experts.linear_fc1.weight9 + module.decoder.layers.32.mlp.experts.linear_fc1.weight7 +Params for bucket 30 (40894464 elements, 40894464 padded size): + module.decoder.layers.32.mlp.experts.linear_fc1.weight0 + module.decoder.layers.31.mlp.experts.linear_fc2.weight11 + module.decoder.layers.32.mlp.experts.linear_fc1.weight5 + module.decoder.layers.31.mlp.experts.linear_fc2.weight8 + module.decoder.layers.32.mlp.experts.linear_fc1.weight2 + module.decoder.layers.31.mlp.experts.linear_fc2.weight13 + module.decoder.layers.31.mlp.experts.linear_fc2.weight5 + module.decoder.layers.31.mlp.experts.linear_fc2.weight10 + module.decoder.layers.32.mlp.experts.linear_fc1.weight4 + module.decoder.layers.31.mlp.experts.linear_fc2.weight15 + module.decoder.layers.31.mlp.experts.linear_fc2.weight7 + module.decoder.layers.32.mlp.experts.linear_fc1.weight1 + module.decoder.layers.31.mlp.experts.linear_fc2.weight12 + module.decoder.layers.31.mlp.experts.linear_fc2.weight4 + module.decoder.layers.32.mlp.experts.linear_fc1.weight6 + module.decoder.layers.31.mlp.experts.linear_fc2.weight9 + module.decoder.layers.32.mlp.experts.linear_fc1.weight3 + module.decoder.layers.31.mlp.experts.linear_fc2.weight14 + module.decoder.layers.31.mlp.experts.linear_fc2.weight6 +Params for bucket 31 (40894464 elements, 40894464 padded size): + module.decoder.layers.31.mlp.experts.linear_fc2.weight3 + module.decoder.layers.31.mlp.experts.linear_fc2.weight1 + module.decoder.layers.31.mlp.experts.linear_fc1.weight14 + module.decoder.layers.31.mlp.experts.linear_fc1.weight12 + module.decoder.layers.31.mlp.experts.linear_fc1.weight10 + module.decoder.layers.31.mlp.experts.linear_fc1.weight8 + module.decoder.layers.31.mlp.experts.linear_fc1.weight6 + module.decoder.layers.31.mlp.experts.linear_fc2.weight2 + module.decoder.layers.31.mlp.experts.linear_fc2.weight0 + module.decoder.layers.31.mlp.experts.linear_fc1.weight15 + module.decoder.layers.31.mlp.experts.linear_fc1.weight13 + module.decoder.layers.31.mlp.experts.linear_fc1.weight11 + module.decoder.layers.31.mlp.experts.linear_fc1.weight9 + module.decoder.layers.31.mlp.experts.linear_fc1.weight7 + module.decoder.layers.31.mlp.experts.linear_fc1.weight5 +Params for bucket 32 (40894464 elements, 40894464 padded size): + module.decoder.layers.31.mlp.experts.linear_fc1.weight2 + module.decoder.layers.30.mlp.experts.linear_fc2.weight13 + module.decoder.layers.30.mlp.experts.linear_fc2.weight5 + module.decoder.layers.30.mlp.experts.linear_fc2.weight10 + module.decoder.layers.30.mlp.experts.linear_fc2.weight2 + module.decoder.layers.31.mlp.experts.linear_fc1.weight4 + module.decoder.layers.30.mlp.experts.linear_fc2.weight15 + module.decoder.layers.30.mlp.experts.linear_fc2.weight7 + module.decoder.layers.31.mlp.experts.linear_fc1.weight1 + module.decoder.layers.30.mlp.experts.linear_fc2.weight12 + module.decoder.layers.30.mlp.experts.linear_fc2.weight4 + module.decoder.layers.30.mlp.experts.linear_fc2.weight9 + module.decoder.layers.30.mlp.experts.linear_fc2.weight1 + module.decoder.layers.31.mlp.experts.linear_fc1.weight3 + module.decoder.layers.30.mlp.experts.linear_fc2.weight14 + module.decoder.layers.30.mlp.experts.linear_fc2.weight6 + module.decoder.layers.31.mlp.experts.linear_fc1.weight0 + module.decoder.layers.30.mlp.experts.linear_fc2.weight11 + module.decoder.layers.30.mlp.experts.linear_fc2.weight3 + module.decoder.layers.30.mlp.experts.linear_fc2.weight8 + module.decoder.layers.30.mlp.experts.linear_fc2.weight0 +Params for bucket 33 (40894464 elements, 40894464 padded size): + module.decoder.layers.30.mlp.experts.linear_fc1.weight12 + module.decoder.layers.30.mlp.experts.linear_fc1.weight14 + module.decoder.layers.30.mlp.experts.linear_fc1.weight10 + module.decoder.layers.30.mlp.experts.linear_fc1.weight8 + module.decoder.layers.30.mlp.experts.linear_fc1.weight6 + module.decoder.layers.30.mlp.experts.linear_fc1.weight4 + module.decoder.layers.30.mlp.experts.linear_fc1.weight15 + module.decoder.layers.30.mlp.experts.linear_fc1.weight13 + module.decoder.layers.30.mlp.experts.linear_fc1.weight11 + module.decoder.layers.30.mlp.experts.linear_fc1.weight9 + module.decoder.layers.30.mlp.experts.linear_fc1.weight7 + module.decoder.layers.30.mlp.experts.linear_fc1.weight5 + module.decoder.layers.30.mlp.experts.linear_fc1.weight3 +Params for bucket 34 (40894464 elements, 40894464 padded size): + module.decoder.layers.29.mlp.experts.linear_fc2.weight15 + module.decoder.layers.29.mlp.experts.linear_fc2.weight7 + module.decoder.layers.29.mlp.experts.linear_fc1.weight14 + module.decoder.layers.30.mlp.experts.linear_fc1.weight1 + module.decoder.layers.29.mlp.experts.linear_fc2.weight12 + module.decoder.layers.29.mlp.experts.linear_fc2.weight4 + module.decoder.layers.29.mlp.experts.linear_fc2.weight9 + module.decoder.layers.29.mlp.experts.linear_fc2.weight1 + module.decoder.layers.29.mlp.experts.linear_fc2.weight14 + module.decoder.layers.29.mlp.experts.linear_fc2.weight6 + module.decoder.layers.30.mlp.experts.linear_fc1.weight0 + module.decoder.layers.29.mlp.experts.linear_fc2.weight11 + module.decoder.layers.29.mlp.experts.linear_fc2.weight3 + module.decoder.layers.29.mlp.experts.linear_fc2.weight8 + module.decoder.layers.29.mlp.experts.linear_fc2.weight0 + module.decoder.layers.29.mlp.experts.linear_fc1.weight15 + module.decoder.layers.30.mlp.experts.linear_fc1.weight2 + module.decoder.layers.29.mlp.experts.linear_fc2.weight13 + module.decoder.layers.29.mlp.experts.linear_fc2.weight5 + module.decoder.layers.29.mlp.experts.linear_fc2.weight10 + module.decoder.layers.29.mlp.experts.linear_fc2.weight2 +Params for bucket 35 (40894464 elements, 40894464 padded size): + module.decoder.layers.29.mlp.experts.linear_fc1.weight10 + module.decoder.layers.29.mlp.experts.linear_fc1.weight12 + module.decoder.layers.29.mlp.experts.linear_fc1.weight8 + module.decoder.layers.29.mlp.experts.linear_fc1.weight6 + module.decoder.layers.29.mlp.experts.linear_fc1.weight4 + module.decoder.layers.29.mlp.experts.linear_fc1.weight2 + module.decoder.layers.29.mlp.experts.linear_fc1.weight13 + module.decoder.layers.29.mlp.experts.linear_fc1.weight9 + module.decoder.layers.29.mlp.experts.linear_fc1.weight11 + module.decoder.layers.29.mlp.experts.linear_fc1.weight7 + module.decoder.layers.29.mlp.experts.linear_fc1.weight5 + module.decoder.layers.29.mlp.experts.linear_fc1.weight3 + module.decoder.layers.29.mlp.experts.linear_fc1.weight1 +Params for bucket 36 (40894464 elements, 40894464 padded size): + module.decoder.layers.28.mlp.experts.linear_fc2.weight9 + module.decoder.layers.28.mlp.experts.linear_fc2.weight1 + module.decoder.layers.28.mlp.experts.linear_fc2.weight14 + module.decoder.layers.28.mlp.experts.linear_fc2.weight6 + module.decoder.layers.28.mlp.experts.linear_fc1.weight13 + module.decoder.layers.29.mlp.experts.linear_fc1.weight0 + module.decoder.layers.28.mlp.experts.linear_fc2.weight11 + module.decoder.layers.28.mlp.experts.linear_fc2.weight3 + module.decoder.layers.28.mlp.experts.linear_fc2.weight8 + module.decoder.layers.28.mlp.experts.linear_fc2.weight0 + module.decoder.layers.28.mlp.experts.linear_fc1.weight15 + module.decoder.layers.28.mlp.experts.linear_fc2.weight13 + module.decoder.layers.28.mlp.experts.linear_fc2.weight5 + module.decoder.layers.28.mlp.experts.linear_fc1.weight12 + module.decoder.layers.28.mlp.experts.linear_fc2.weight10 + module.decoder.layers.28.mlp.experts.linear_fc2.weight2 + module.decoder.layers.28.mlp.experts.linear_fc2.weight15 + module.decoder.layers.28.mlp.experts.linear_fc2.weight7 + module.decoder.layers.28.mlp.experts.linear_fc1.weight14 + module.decoder.layers.28.mlp.experts.linear_fc2.weight12 + module.decoder.layers.28.mlp.experts.linear_fc2.weight4 +Params for bucket 37 (40894464 elements, 40894464 padded size): + module.decoder.layers.28.mlp.experts.linear_fc1.weight10 + module.decoder.layers.28.mlp.experts.linear_fc1.weight8 + module.decoder.layers.28.mlp.experts.linear_fc1.weight6 + module.decoder.layers.28.mlp.experts.linear_fc1.weight4 + module.decoder.layers.28.mlp.experts.linear_fc1.weight2 + module.decoder.layers.28.mlp.experts.linear_fc1.weight0 + module.decoder.layers.27.mlp.experts.linear_fc2.weight15 + module.decoder.layers.28.mlp.experts.linear_fc1.weight11 + module.decoder.layers.28.mlp.experts.linear_fc1.weight7 + module.decoder.layers.28.mlp.experts.linear_fc1.weight9 + module.decoder.layers.28.mlp.experts.linear_fc1.weight5 + module.decoder.layers.28.mlp.experts.linear_fc1.weight3 + module.decoder.layers.28.mlp.experts.linear_fc1.weight1 + module.decoder.layers.27.mlp.experts.linear_fc2.weight14 +Params for bucket 38 (40894464 elements, 40894464 padded size): + module.decoder.layers.27.mlp.experts.linear_fc2.weight11 + module.decoder.layers.27.mlp.experts.linear_fc2.weight3 + module.decoder.layers.27.mlp.experts.linear_fc1.weight10 + module.decoder.layers.27.mlp.experts.linear_fc2.weight8 + module.decoder.layers.27.mlp.experts.linear_fc2.weight0 + module.decoder.layers.27.mlp.experts.linear_fc1.weight15 + module.decoder.layers.27.mlp.experts.linear_fc2.weight13 + module.decoder.layers.27.mlp.experts.linear_fc2.weight5 + module.decoder.layers.27.mlp.experts.linear_fc1.weight12 + module.decoder.layers.27.mlp.experts.linear_fc2.weight10 + module.decoder.layers.27.mlp.experts.linear_fc2.weight2 + module.decoder.layers.27.mlp.experts.linear_fc2.weight7 + module.decoder.layers.27.mlp.experts.linear_fc1.weight14 + module.decoder.layers.27.mlp.experts.linear_fc2.weight12 + module.decoder.layers.27.mlp.experts.linear_fc2.weight4 + module.decoder.layers.27.mlp.experts.linear_fc1.weight11 + module.decoder.layers.27.mlp.experts.linear_fc2.weight9 + module.decoder.layers.27.mlp.experts.linear_fc2.weight1 + module.decoder.layers.27.mlp.experts.linear_fc2.weight6 + module.decoder.layers.27.mlp.experts.linear_fc1.weight13 +Params for bucket 39 (40894464 elements, 40894464 padded size): + module.decoder.layers.27.mlp.experts.linear_fc1.weight8 + module.decoder.layers.27.mlp.experts.linear_fc1.weight6 + module.decoder.layers.27.mlp.experts.linear_fc1.weight4 + module.decoder.layers.27.mlp.experts.linear_fc1.weight2 + module.decoder.layers.27.mlp.experts.linear_fc1.weight0 + module.decoder.layers.26.mlp.experts.linear_fc2.weight15 + module.decoder.layers.26.mlp.experts.linear_fc2.weight13 + module.decoder.layers.26.mlp.experts.linear_fc2.weight11 + module.decoder.layers.27.mlp.experts.linear_fc1.weight9 + module.decoder.layers.27.mlp.experts.linear_fc1.weight7 + module.decoder.layers.27.mlp.experts.linear_fc1.weight5 + module.decoder.layers.27.mlp.experts.linear_fc1.weight3 + module.decoder.layers.27.mlp.experts.linear_fc1.weight1 + module.decoder.layers.26.mlp.experts.linear_fc2.weight14 + module.decoder.layers.26.mlp.experts.linear_fc2.weight12 + module.decoder.layers.26.mlp.experts.linear_fc2.weight10 +Params for bucket 40 (40894464 elements, 40894464 padded size): + module.decoder.layers.26.mlp.experts.linear_fc2.weight9 + module.decoder.layers.26.mlp.experts.linear_fc2.weight7 + module.decoder.layers.26.mlp.experts.linear_fc2.weight5 + module.decoder.layers.26.mlp.experts.linear_fc2.weight3 + module.decoder.layers.26.mlp.experts.linear_fc2.weight1 + module.decoder.layers.26.mlp.experts.linear_fc1.weight14 + module.decoder.layers.26.mlp.experts.linear_fc1.weight12 + module.decoder.layers.26.mlp.experts.linear_fc1.weight10 + module.decoder.layers.26.mlp.experts.linear_fc1.weight8 + module.decoder.layers.26.mlp.experts.linear_fc2.weight8 + module.decoder.layers.26.mlp.experts.linear_fc2.weight6 + module.decoder.layers.26.mlp.experts.linear_fc2.weight4 + module.decoder.layers.26.mlp.experts.linear_fc2.weight2 + module.decoder.layers.26.mlp.experts.linear_fc2.weight0 + module.decoder.layers.26.mlp.experts.linear_fc1.weight15 + module.decoder.layers.26.mlp.experts.linear_fc1.weight13 + module.decoder.layers.26.mlp.experts.linear_fc1.weight11 + module.decoder.layers.26.mlp.experts.linear_fc1.weight9 +Params for bucket 41 (40894464 elements, 40894464 padded size): + module.decoder.layers.26.mlp.experts.linear_fc1.weight4 + module.decoder.layers.26.mlp.experts.linear_fc1.weight6 + module.decoder.layers.26.mlp.experts.linear_fc1.weight2 + module.decoder.layers.26.mlp.experts.linear_fc1.weight0 + module.decoder.layers.25.mlp.experts.linear_fc2.weight15 + module.decoder.layers.25.mlp.experts.linear_fc2.weight13 + module.decoder.layers.25.mlp.experts.linear_fc2.weight11 + module.decoder.layers.25.mlp.experts.linear_fc2.weight9 + module.decoder.layers.25.mlp.experts.linear_fc2.weight7 + module.decoder.layers.26.mlp.experts.linear_fc1.weight7 + module.decoder.layers.26.mlp.experts.linear_fc1.weight5 + module.decoder.layers.26.mlp.experts.linear_fc1.weight3 + module.decoder.layers.26.mlp.experts.linear_fc1.weight1 + module.decoder.layers.25.mlp.experts.linear_fc2.weight14 + module.decoder.layers.25.mlp.experts.linear_fc2.weight12 + module.decoder.layers.25.mlp.experts.linear_fc2.weight10 + module.decoder.layers.25.mlp.experts.linear_fc2.weight8 + module.decoder.layers.25.mlp.experts.linear_fc2.weight6 +Params for bucket 42 (40894464 elements, 40894464 padded size): + module.decoder.layers.25.mlp.experts.linear_fc2.weight5 + module.decoder.layers.25.mlp.experts.linear_fc2.weight1 + module.decoder.layers.25.mlp.experts.linear_fc2.weight3 + module.decoder.layers.25.mlp.experts.linear_fc1.weight14 + module.decoder.layers.25.mlp.experts.linear_fc1.weight12 + module.decoder.layers.25.mlp.experts.linear_fc1.weight10 + module.decoder.layers.25.mlp.experts.linear_fc1.weight8 + module.decoder.layers.25.mlp.experts.linear_fc1.weight6 + module.decoder.layers.25.mlp.experts.linear_fc2.weight4 + module.decoder.layers.25.mlp.experts.linear_fc2.weight2 + module.decoder.layers.25.mlp.experts.linear_fc2.weight0 + module.decoder.layers.25.mlp.experts.linear_fc1.weight15 + module.decoder.layers.25.mlp.experts.linear_fc1.weight13 + module.decoder.layers.25.mlp.experts.linear_fc1.weight11 + module.decoder.layers.25.mlp.experts.linear_fc1.weight9 + module.decoder.layers.25.mlp.experts.linear_fc1.weight7 +Params for bucket 43 (40894464 elements, 40894464 padded size): + module.decoder.layers.24.mlp.experts.linear_fc2.weight9 + module.decoder.layers.25.mlp.experts.linear_fc1.weight3 + module.decoder.layers.24.mlp.experts.linear_fc2.weight14 + module.decoder.layers.24.mlp.experts.linear_fc2.weight6 + module.decoder.layers.25.mlp.experts.linear_fc1.weight0 + module.decoder.layers.24.mlp.experts.linear_fc2.weight11 + module.decoder.layers.24.mlp.experts.linear_fc2.weight3 + module.decoder.layers.25.mlp.experts.linear_fc1.weight5 + module.decoder.layers.24.mlp.experts.linear_fc2.weight8 + module.decoder.layers.25.mlp.experts.linear_fc1.weight2 + module.decoder.layers.24.mlp.experts.linear_fc2.weight13 + module.decoder.layers.24.mlp.experts.linear_fc2.weight5 + module.decoder.layers.24.mlp.experts.linear_fc2.weight10 + module.decoder.layers.24.mlp.experts.linear_fc2.weight2 + module.decoder.layers.25.mlp.experts.linear_fc1.weight4 + module.decoder.layers.24.mlp.experts.linear_fc2.weight15 + module.decoder.layers.24.mlp.experts.linear_fc2.weight7 + module.decoder.layers.25.mlp.experts.linear_fc1.weight1 + module.decoder.layers.24.mlp.experts.linear_fc2.weight12 + module.decoder.layers.24.mlp.experts.linear_fc2.weight4 +Params for bucket 44 (40894464 elements, 40894464 padded size): + module.decoder.layers.24.mlp.experts.linear_fc2.weight1 + module.decoder.layers.24.mlp.experts.linear_fc1.weight14 + module.decoder.layers.24.mlp.experts.linear_fc1.weight12 + module.decoder.layers.24.mlp.experts.linear_fc1.weight10 + module.decoder.layers.24.mlp.experts.linear_fc1.weight8 + module.decoder.layers.24.mlp.experts.linear_fc1.weight6 + module.decoder.layers.24.mlp.experts.linear_fc1.weight4 + module.decoder.layers.24.mlp.experts.linear_fc1.weight15 + module.decoder.layers.24.mlp.experts.linear_fc1.weight13 + module.decoder.layers.24.mlp.experts.linear_fc2.weight0 + module.decoder.layers.24.mlp.experts.linear_fc1.weight11 + module.decoder.layers.24.mlp.experts.linear_fc1.weight9 + module.decoder.layers.24.mlp.experts.linear_fc1.weight7 + module.decoder.layers.24.mlp.experts.linear_fc1.weight5 +Params for bucket 45 (40894464 elements, 40894464 padded size): + module.decoder.layers.24.mlp.experts.linear_fc1.weight0 + module.decoder.layers.23.mlp.experts.linear_fc2.weight10 + module.decoder.layers.23.mlp.experts.linear_fc2.weight2 + module.decoder.layers.23.mlp.experts.linear_fc2.weight15 + module.decoder.layers.23.mlp.experts.linear_fc2.weight7 + module.decoder.layers.24.mlp.experts.linear_fc1.weight2 + module.decoder.layers.23.mlp.experts.linear_fc2.weight12 + module.decoder.layers.23.mlp.experts.linear_fc2.weight4 + module.decoder.layers.23.mlp.experts.linear_fc1.weight15 + module.decoder.layers.23.mlp.experts.linear_fc2.weight9 + module.decoder.layers.23.mlp.experts.linear_fc2.weight1 + module.decoder.layers.23.mlp.experts.linear_fc2.weight14 + module.decoder.layers.23.mlp.experts.linear_fc2.weight6 + module.decoder.layers.24.mlp.experts.linear_fc1.weight1 + module.decoder.layers.23.mlp.experts.linear_fc2.weight11 + module.decoder.layers.23.mlp.experts.linear_fc2.weight3 + module.decoder.layers.23.mlp.experts.linear_fc2.weight8 + module.decoder.layers.23.mlp.experts.linear_fc2.weight0 + module.decoder.layers.24.mlp.experts.linear_fc1.weight3 + module.decoder.layers.23.mlp.experts.linear_fc2.weight13 + module.decoder.layers.23.mlp.experts.linear_fc2.weight5 +Params for bucket 46 (40894464 elements, 40894464 padded size): + module.decoder.layers.23.mlp.experts.linear_fc1.weight11 + module.decoder.layers.23.mlp.experts.linear_fc1.weight13 + module.decoder.layers.23.mlp.experts.linear_fc1.weight9 + module.decoder.layers.23.mlp.experts.linear_fc1.weight7 + module.decoder.layers.23.mlp.experts.linear_fc1.weight5 + module.decoder.layers.23.mlp.experts.linear_fc1.weight3 + module.decoder.layers.23.mlp.experts.linear_fc1.weight14 + module.decoder.layers.23.mlp.experts.linear_fc1.weight10 + module.decoder.layers.23.mlp.experts.linear_fc1.weight12 + module.decoder.layers.23.mlp.experts.linear_fc1.weight8 + module.decoder.layers.23.mlp.experts.linear_fc1.weight6 + module.decoder.layers.23.mlp.experts.linear_fc1.weight4 + module.decoder.layers.23.mlp.experts.linear_fc1.weight2 +Params for bucket 47 (40894464 elements, 40894464 padded size): + module.decoder.layers.23.mlp.experts.linear_fc1.weight1 + module.decoder.layers.22.mlp.experts.linear_fc2.weight12 + module.decoder.layers.22.mlp.experts.linear_fc2.weight4 + module.decoder.layers.22.mlp.experts.linear_fc1.weight15 + module.decoder.layers.22.mlp.experts.linear_fc2.weight9 + module.decoder.layers.22.mlp.experts.linear_fc2.weight1 + module.decoder.layers.22.mlp.experts.linear_fc2.weight14 + module.decoder.layers.22.mlp.experts.linear_fc2.weight6 + module.decoder.layers.23.mlp.experts.linear_fc1.weight0 + module.decoder.layers.22.mlp.experts.linear_fc2.weight11 + module.decoder.layers.22.mlp.experts.linear_fc2.weight3 + module.decoder.layers.22.mlp.experts.linear_fc1.weight14 + module.decoder.layers.22.mlp.experts.linear_fc2.weight8 + module.decoder.layers.22.mlp.experts.linear_fc2.weight0 + module.decoder.layers.22.mlp.experts.linear_fc2.weight13 + module.decoder.layers.22.mlp.experts.linear_fc2.weight5 + module.decoder.layers.22.mlp.experts.linear_fc2.weight10 + module.decoder.layers.22.mlp.experts.linear_fc2.weight2 + module.decoder.layers.22.mlp.experts.linear_fc1.weight13 + module.decoder.layers.22.mlp.experts.linear_fc2.weight15 + module.decoder.layers.22.mlp.experts.linear_fc2.weight7 +Params for bucket 48 (40894464 elements, 40894464 padded size): + module.decoder.layers.22.mlp.experts.linear_fc1.weight11 + module.decoder.layers.22.mlp.experts.linear_fc1.weight9 + module.decoder.layers.22.mlp.experts.linear_fc1.weight7 + module.decoder.layers.22.mlp.experts.linear_fc1.weight5 + module.decoder.layers.22.mlp.experts.linear_fc1.weight3 + module.decoder.layers.22.mlp.experts.linear_fc1.weight1 + module.decoder.layers.22.mlp.experts.linear_fc1.weight12 + module.decoder.layers.22.mlp.experts.linear_fc1.weight10 + module.decoder.layers.22.mlp.experts.linear_fc1.weight8 + module.decoder.layers.22.mlp.experts.linear_fc1.weight6 + module.decoder.layers.22.mlp.experts.linear_fc1.weight4 + module.decoder.layers.22.mlp.experts.linear_fc1.weight2 + module.decoder.layers.22.mlp.experts.linear_fc1.weight0 +Params for bucket 49 (40894464 elements, 40894464 padded size): + module.decoder.layers.21.mlp.experts.linear_fc2.weight15 + module.decoder.layers.21.mlp.experts.linear_fc2.weight7 + module.decoder.layers.21.mlp.experts.linear_fc2.weight12 + module.decoder.layers.21.mlp.experts.linear_fc2.weight4 + module.decoder.layers.21.mlp.experts.linear_fc1.weight15 + module.decoder.layers.21.mlp.experts.linear_fc2.weight9 + module.decoder.layers.21.mlp.experts.linear_fc2.weight1 + module.decoder.layers.21.mlp.experts.linear_fc1.weight12 + module.decoder.layers.21.mlp.experts.linear_fc2.weight14 + module.decoder.layers.21.mlp.experts.linear_fc2.weight6 + module.decoder.layers.21.mlp.experts.linear_fc2.weight11 + module.decoder.layers.21.mlp.experts.linear_fc2.weight3 + module.decoder.layers.21.mlp.experts.linear_fc1.weight14 + module.decoder.layers.21.mlp.experts.linear_fc2.weight8 + module.decoder.layers.21.mlp.experts.linear_fc2.weight0 + module.decoder.layers.21.mlp.experts.linear_fc1.weight11 + module.decoder.layers.21.mlp.experts.linear_fc2.weight13 + module.decoder.layers.21.mlp.experts.linear_fc2.weight5 + module.decoder.layers.21.mlp.experts.linear_fc2.weight10 + module.decoder.layers.21.mlp.experts.linear_fc2.weight2 + module.decoder.layers.21.mlp.experts.linear_fc1.weight13 +Params for bucket 50 (40894464 elements, 40894464 padded size): + module.decoder.layers.21.mlp.experts.linear_fc1.weight10 + module.decoder.layers.21.mlp.experts.linear_fc1.weight8 + module.decoder.layers.21.mlp.experts.linear_fc1.weight6 + module.decoder.layers.21.mlp.experts.linear_fc1.weight4 + module.decoder.layers.21.mlp.experts.linear_fc1.weight2 + module.decoder.layers.21.mlp.experts.linear_fc1.weight0 + module.decoder.layers.20.mlp.experts.linear_fc2.weight14 + module.decoder.layers.20.mlp.experts.linear_fc2.weight12 + module.decoder.layers.21.mlp.experts.linear_fc1.weight9 + module.decoder.layers.21.mlp.experts.linear_fc1.weight7 + module.decoder.layers.21.mlp.experts.linear_fc1.weight5 + module.decoder.layers.21.mlp.experts.linear_fc1.weight3 + module.decoder.layers.21.mlp.experts.linear_fc1.weight1 + module.decoder.layers.20.mlp.experts.linear_fc2.weight15 + module.decoder.layers.20.mlp.experts.linear_fc2.weight13 +Params for bucket 51 (40894464 elements, 40894464 padded size): + module.decoder.layers.20.mlp.experts.linear_fc2.weight10 + module.decoder.layers.20.mlp.experts.linear_fc2.weight2 + module.decoder.layers.20.mlp.experts.linear_fc1.weight13 + module.decoder.layers.20.mlp.experts.linear_fc2.weight7 + module.decoder.layers.20.mlp.experts.linear_fc1.weight10 + module.decoder.layers.20.mlp.experts.linear_fc2.weight4 + module.decoder.layers.20.mlp.experts.linear_fc1.weight15 + module.decoder.layers.20.mlp.experts.linear_fc2.weight9 + module.decoder.layers.20.mlp.experts.linear_fc2.weight1 + module.decoder.layers.20.mlp.experts.linear_fc1.weight12 + module.decoder.layers.20.mlp.experts.linear_fc2.weight6 + module.decoder.layers.20.mlp.experts.linear_fc1.weight9 + module.decoder.layers.20.mlp.experts.linear_fc2.weight11 + module.decoder.layers.20.mlp.experts.linear_fc2.weight3 + module.decoder.layers.20.mlp.experts.linear_fc1.weight14 + module.decoder.layers.20.mlp.experts.linear_fc2.weight8 + module.decoder.layers.20.mlp.experts.linear_fc2.weight0 + module.decoder.layers.20.mlp.experts.linear_fc1.weight11 + module.decoder.layers.20.mlp.experts.linear_fc2.weight5 +Params for bucket 52 (40894464 elements, 40894464 padded size): + module.decoder.layers.20.mlp.experts.linear_fc1.weight7 + module.decoder.layers.20.mlp.experts.linear_fc1.weight5 + module.decoder.layers.20.mlp.experts.linear_fc1.weight3 + module.decoder.layers.20.mlp.experts.linear_fc1.weight1 + module.decoder.layers.19.mlp.experts.linear_fc2.weight14 + module.decoder.layers.19.mlp.experts.linear_fc2.weight12 + module.decoder.layers.19.mlp.experts.linear_fc2.weight10 + module.decoder.layers.19.mlp.experts.linear_fc2.weight8 + module.decoder.layers.20.mlp.experts.linear_fc1.weight8 + module.decoder.layers.20.mlp.experts.linear_fc1.weight4 + module.decoder.layers.20.mlp.experts.linear_fc1.weight6 + module.decoder.layers.20.mlp.experts.linear_fc1.weight2 + module.decoder.layers.20.mlp.experts.linear_fc1.weight0 + module.decoder.layers.19.mlp.experts.linear_fc2.weight15 + module.decoder.layers.19.mlp.experts.linear_fc2.weight13 + module.decoder.layers.19.mlp.experts.linear_fc2.weight11 + module.decoder.layers.19.mlp.experts.linear_fc2.weight9 +Params for bucket 53 (40894464 elements, 40894464 padded size): + module.decoder.layers.19.mlp.experts.linear_fc2.weight4 + module.decoder.layers.19.mlp.experts.linear_fc2.weight6 + module.decoder.layers.19.mlp.experts.linear_fc2.weight2 + module.decoder.layers.19.mlp.experts.linear_fc2.weight0 + module.decoder.layers.19.mlp.experts.linear_fc1.weight14 + module.decoder.layers.19.mlp.experts.linear_fc1.weight12 + module.decoder.layers.19.mlp.experts.linear_fc1.weight10 + module.decoder.layers.19.mlp.experts.linear_fc1.weight8 + module.decoder.layers.19.mlp.experts.linear_fc2.weight7 + module.decoder.layers.19.mlp.experts.linear_fc2.weight5 + module.decoder.layers.19.mlp.experts.linear_fc2.weight3 + module.decoder.layers.19.mlp.experts.linear_fc2.weight1 + module.decoder.layers.19.mlp.experts.linear_fc1.weight15 + module.decoder.layers.19.mlp.experts.linear_fc1.weight13 + module.decoder.layers.19.mlp.experts.linear_fc1.weight11 + module.decoder.layers.19.mlp.experts.linear_fc1.weight9 + module.decoder.layers.19.mlp.experts.linear_fc1.weight7 +Params for bucket 54 (40894464 elements, 40894464 padded size): + module.decoder.layers.19.mlp.experts.linear_fc1.weight6 + module.decoder.layers.18.mlp.experts.linear_fc2.weight13 + module.decoder.layers.18.mlp.experts.linear_fc2.weight5 + module.decoder.layers.19.mlp.experts.linear_fc1.weight3 + module.decoder.layers.18.mlp.experts.linear_fc2.weight10 + module.decoder.layers.19.mlp.experts.linear_fc1.weight0 + module.decoder.layers.18.mlp.experts.linear_fc2.weight15 + module.decoder.layers.18.mlp.experts.linear_fc2.weight7 + module.decoder.layers.19.mlp.experts.linear_fc1.weight5 + module.decoder.layers.18.mlp.experts.linear_fc2.weight12 + module.decoder.layers.18.mlp.experts.linear_fc2.weight4 + module.decoder.layers.19.mlp.experts.linear_fc1.weight2 + module.decoder.layers.18.mlp.experts.linear_fc2.weight9 + module.decoder.layers.18.mlp.experts.linear_fc2.weight14 + module.decoder.layers.18.mlp.experts.linear_fc2.weight6 + module.decoder.layers.19.mlp.experts.linear_fc1.weight4 + module.decoder.layers.18.mlp.experts.linear_fc2.weight11 + module.decoder.layers.19.mlp.experts.linear_fc1.weight1 + module.decoder.layers.18.mlp.experts.linear_fc2.weight8 +Params for bucket 55 (40894464 elements, 40894464 padded size): + module.decoder.layers.18.mlp.experts.linear_fc2.weight3 + module.decoder.layers.18.mlp.experts.linear_fc2.weight1 + module.decoder.layers.18.mlp.experts.linear_fc1.weight14 + module.decoder.layers.18.mlp.experts.linear_fc1.weight12 + module.decoder.layers.18.mlp.experts.linear_fc1.weight10 + module.decoder.layers.18.mlp.experts.linear_fc1.weight8 + module.decoder.layers.18.mlp.experts.linear_fc1.weight6 + module.decoder.layers.18.mlp.experts.linear_fc2.weight2 + module.decoder.layers.18.mlp.experts.linear_fc1.weight15 + module.decoder.layers.18.mlp.experts.linear_fc2.weight0 + module.decoder.layers.18.mlp.experts.linear_fc1.weight13 + module.decoder.layers.18.mlp.experts.linear_fc1.weight11 + module.decoder.layers.18.mlp.experts.linear_fc1.weight9 + module.decoder.layers.18.mlp.experts.linear_fc1.weight7 + module.decoder.layers.18.mlp.experts.linear_fc1.weight5 +Params for bucket 56 (40894464 elements, 40894464 padded size): + module.decoder.layers.18.mlp.experts.linear_fc1.weight0 + module.decoder.layers.17.mlp.experts.linear_fc2.weight15 + module.decoder.layers.17.mlp.experts.linear_fc2.weight7 + module.decoder.layers.17.mlp.experts.linear_fc2.weight12 + module.decoder.layers.17.mlp.experts.linear_fc2.weight4 + module.decoder.layers.18.mlp.experts.linear_fc1.weight2 + module.decoder.layers.17.mlp.experts.linear_fc2.weight9 + module.decoder.layers.17.mlp.experts.linear_fc2.weight1 + module.decoder.layers.17.mlp.experts.linear_fc2.weight14 + module.decoder.layers.17.mlp.experts.linear_fc2.weight6 + module.decoder.layers.18.mlp.experts.linear_fc1.weight4 + module.decoder.layers.17.mlp.experts.linear_fc2.weight11 + module.decoder.layers.17.mlp.experts.linear_fc2.weight3 + module.decoder.layers.18.mlp.experts.linear_fc1.weight1 + module.decoder.layers.17.mlp.experts.linear_fc2.weight8 + module.decoder.layers.17.mlp.experts.linear_fc2.weight0 + module.decoder.layers.17.mlp.experts.linear_fc2.weight13 + module.decoder.layers.17.mlp.experts.linear_fc2.weight5 + module.decoder.layers.18.mlp.experts.linear_fc1.weight3 + module.decoder.layers.17.mlp.experts.linear_fc2.weight10 + module.decoder.layers.17.mlp.experts.linear_fc2.weight2 +Params for bucket 57 (40894464 elements, 40894464 padded size): + module.decoder.layers.17.mlp.experts.linear_fc1.weight14 + module.decoder.layers.17.mlp.experts.linear_fc1.weight12 + module.decoder.layers.17.mlp.experts.linear_fc1.weight10 + module.decoder.layers.17.mlp.experts.linear_fc1.weight8 + module.decoder.layers.17.mlp.experts.linear_fc1.weight6 + module.decoder.layers.17.mlp.experts.linear_fc1.weight4 + module.decoder.layers.17.mlp.experts.linear_fc1.weight15 + module.decoder.layers.17.mlp.experts.linear_fc1.weight11 + module.decoder.layers.17.mlp.experts.linear_fc1.weight13 + module.decoder.layers.17.mlp.experts.linear_fc1.weight9 + module.decoder.layers.17.mlp.experts.linear_fc1.weight7 + module.decoder.layers.17.mlp.experts.linear_fc1.weight5 + module.decoder.layers.17.mlp.experts.linear_fc1.weight3 +Params for bucket 58 (40894464 elements, 40894464 padded size): + module.decoder.layers.17.mlp.experts.linear_fc1.weight2 + module.decoder.layers.16.mlp.experts.linear_fc2.weight9 + module.decoder.layers.16.mlp.experts.linear_fc2.weight1 + module.decoder.layers.16.mlp.experts.linear_fc2.weight14 + module.decoder.layers.16.mlp.experts.linear_fc2.weight6 + module.decoder.layers.16.mlp.experts.linear_fc2.weight11 + module.decoder.layers.16.mlp.experts.linear_fc2.weight3 + module.decoder.layers.16.mlp.experts.linear_fc1.weight14 + module.decoder.layers.17.mlp.experts.linear_fc1.weight1 + module.decoder.layers.16.mlp.experts.linear_fc2.weight8 + module.decoder.layers.16.mlp.experts.linear_fc2.weight0 + module.decoder.layers.16.mlp.experts.linear_fc2.weight13 + module.decoder.layers.16.mlp.experts.linear_fc2.weight5 + module.decoder.layers.16.mlp.experts.linear_fc2.weight10 + module.decoder.layers.16.mlp.experts.linear_fc2.weight2 + module.decoder.layers.16.mlp.experts.linear_fc2.weight15 + module.decoder.layers.17.mlp.experts.linear_fc1.weight0 + module.decoder.layers.16.mlp.experts.linear_fc2.weight7 + module.decoder.layers.16.mlp.experts.linear_fc2.weight12 + module.decoder.layers.16.mlp.experts.linear_fc2.weight4 + module.decoder.layers.16.mlp.experts.linear_fc1.weight15 +Params for bucket 59 (40894464 elements, 40894464 padded size): + module.decoder.layers.16.mlp.experts.linear_fc1.weight10 + module.decoder.layers.16.mlp.experts.linear_fc1.weight12 + module.decoder.layers.16.mlp.experts.linear_fc1.weight8 + module.decoder.layers.16.mlp.experts.linear_fc1.weight6 + module.decoder.layers.16.mlp.experts.linear_fc1.weight4 + module.decoder.layers.16.mlp.experts.linear_fc1.weight2 + module.decoder.layers.16.mlp.experts.linear_fc1.weight13 + module.decoder.layers.16.mlp.experts.linear_fc1.weight11 + module.decoder.layers.16.mlp.experts.linear_fc1.weight9 + module.decoder.layers.16.mlp.experts.linear_fc1.weight7 + module.decoder.layers.16.mlp.experts.linear_fc1.weight5 + module.decoder.layers.16.mlp.experts.linear_fc1.weight3 + module.decoder.layers.16.mlp.experts.linear_fc1.weight1 +Params for bucket 60 (40894464 elements, 40894464 padded size): + module.decoder.layers.15.mlp.experts.linear_fc2.weight11 + module.decoder.layers.15.mlp.experts.linear_fc2.weight3 + module.decoder.layers.15.mlp.experts.linear_fc1.weight14 + module.decoder.layers.15.mlp.experts.linear_fc2.weight8 + module.decoder.layers.15.mlp.experts.linear_fc2.weight0 + module.decoder.layers.15.mlp.experts.linear_fc2.weight13 + module.decoder.layers.15.mlp.experts.linear_fc2.weight5 + module.decoder.layers.15.mlp.experts.linear_fc2.weight10 + module.decoder.layers.15.mlp.experts.linear_fc2.weight2 + module.decoder.layers.15.mlp.experts.linear_fc1.weight13 + module.decoder.layers.16.mlp.experts.linear_fc1.weight0 + module.decoder.layers.15.mlp.experts.linear_fc2.weight15 + module.decoder.layers.15.mlp.experts.linear_fc2.weight7 + module.decoder.layers.15.mlp.experts.linear_fc2.weight12 + module.decoder.layers.15.mlp.experts.linear_fc2.weight4 + module.decoder.layers.15.mlp.experts.linear_fc1.weight15 + module.decoder.layers.15.mlp.experts.linear_fc2.weight9 + module.decoder.layers.15.mlp.experts.linear_fc2.weight1 + module.decoder.layers.15.mlp.experts.linear_fc1.weight12 + module.decoder.layers.15.mlp.experts.linear_fc2.weight14 + module.decoder.layers.15.mlp.experts.linear_fc2.weight6 +Params for bucket 61 (40894464 elements, 40894464 padded size): + module.decoder.layers.15.mlp.experts.linear_fc1.weight8 + module.decoder.layers.15.mlp.experts.linear_fc1.weight10 + module.decoder.layers.15.mlp.experts.linear_fc1.weight6 + module.decoder.layers.15.mlp.experts.linear_fc1.weight4 + module.decoder.layers.15.mlp.experts.linear_fc1.weight2 + module.decoder.layers.15.mlp.experts.linear_fc1.weight0 + module.decoder.layers.14.mlp.experts.linear_fc2.weight15 + module.decoder.layers.15.mlp.experts.linear_fc1.weight11 + module.decoder.layers.15.mlp.experts.linear_fc1.weight7 + module.decoder.layers.15.mlp.experts.linear_fc1.weight9 + 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.weight1 + module.decoder.layers.14.mlp.experts.linear_fc2.weight14 +Params for bucket 62 (40894464 elements, 40894464 padded size): + module.decoder.layers.14.mlp.experts.linear_fc2.weight13 + module.decoder.layers.14.mlp.experts.linear_fc2.weight5 + module.decoder.layers.14.mlp.experts.linear_fc1.weight15 + module.decoder.layers.14.mlp.experts.linear_fc2.weight10 + module.decoder.layers.14.mlp.experts.linear_fc2.weight2 + module.decoder.layers.14.mlp.experts.linear_fc1.weight12 + module.decoder.layers.14.mlp.experts.linear_fc2.weight7 + module.decoder.layers.14.mlp.experts.linear_fc2.weight12 + module.decoder.layers.14.mlp.experts.linear_fc2.weight4 + module.decoder.layers.14.mlp.experts.linear_fc1.weight14 + module.decoder.layers.14.mlp.experts.linear_fc2.weight9 + module.decoder.layers.14.mlp.experts.linear_fc2.weight1 + module.decoder.layers.14.mlp.experts.linear_fc1.weight11 + module.decoder.layers.14.mlp.experts.linear_fc2.weight6 + module.decoder.layers.14.mlp.experts.linear_fc2.weight11 + module.decoder.layers.14.mlp.experts.linear_fc2.weight3 + module.decoder.layers.14.mlp.experts.linear_fc1.weight13 + module.decoder.layers.14.mlp.experts.linear_fc2.weight8 + module.decoder.layers.14.mlp.experts.linear_fc2.weight0 + module.decoder.layers.14.mlp.experts.linear_fc1.weight10 +Params for bucket 63 (40894464 elements, 40894464 padded size): + module.decoder.layers.14.mlp.experts.linear_fc1.weight9 + module.decoder.layers.14.mlp.experts.linear_fc1.weight7 + module.decoder.layers.14.mlp.experts.linear_fc1.weight5 + module.decoder.layers.14.mlp.experts.linear_fc1.weight3 + module.decoder.layers.14.mlp.experts.linear_fc1.weight1 + module.decoder.layers.13.mlp.experts.linear_fc2.weight14 + module.decoder.layers.13.mlp.experts.linear_fc2.weight12 + module.decoder.layers.13.mlp.experts.linear_fc2.weight10 + module.decoder.layers.14.mlp.experts.linear_fc1.weight8 + module.decoder.layers.14.mlp.experts.linear_fc1.weight6 + module.decoder.layers.14.mlp.experts.linear_fc1.weight4 + module.decoder.layers.14.mlp.experts.linear_fc1.weight2 + module.decoder.layers.14.mlp.experts.linear_fc1.weight0 + module.decoder.layers.13.mlp.experts.linear_fc2.weight15 + module.decoder.layers.13.mlp.experts.linear_fc2.weight13 + module.decoder.layers.13.mlp.experts.linear_fc2.weight11 +Params for bucket 64 (40894464 elements, 40894464 padded size): + module.decoder.layers.13.mlp.experts.linear_fc2.weight8 + module.decoder.layers.13.mlp.experts.linear_fc2.weight6 + module.decoder.layers.13.mlp.experts.linear_fc2.weight4 + module.decoder.layers.13.mlp.experts.linear_fc2.weight2 + module.decoder.layers.13.mlp.experts.linear_fc2.weight0 + module.decoder.layers.13.mlp.experts.linear_fc1.weight15 + module.decoder.layers.13.mlp.experts.linear_fc1.weight13 + module.decoder.layers.13.mlp.experts.linear_fc1.weight11 + module.decoder.layers.13.mlp.experts.linear_fc1.weight9 + module.decoder.layers.13.mlp.experts.linear_fc2.weight9 + module.decoder.layers.13.mlp.experts.linear_fc2.weight5 + module.decoder.layers.13.mlp.experts.linear_fc2.weight7 + module.decoder.layers.13.mlp.experts.linear_fc2.weight3 + module.decoder.layers.13.mlp.experts.linear_fc2.weight1 + module.decoder.layers.13.mlp.experts.linear_fc1.weight14 + module.decoder.layers.13.mlp.experts.linear_fc1.weight12 + module.decoder.layers.13.mlp.experts.linear_fc1.weight10 + module.decoder.layers.13.mlp.experts.linear_fc1.weight8 +Params for bucket 65 (40894464 elements, 40894464 padded size): + module.decoder.layers.13.mlp.experts.linear_fc1.weight7 + module.decoder.layers.13.mlp.experts.linear_fc1.weight5 + module.decoder.layers.13.mlp.experts.linear_fc1.weight3 + module.decoder.layers.13.mlp.experts.linear_fc1.weight1 + module.decoder.layers.12.mlp.experts.linear_fc2.weight14 + module.decoder.layers.12.mlp.experts.linear_fc2.weight12 + module.decoder.layers.12.mlp.experts.linear_fc2.weight10 + module.decoder.layers.12.mlp.experts.linear_fc2.weight8 + module.decoder.layers.12.mlp.experts.linear_fc2.weight6 + module.decoder.layers.13.mlp.experts.linear_fc1.weight4 + module.decoder.layers.13.mlp.experts.linear_fc1.weight6 + module.decoder.layers.13.mlp.experts.linear_fc1.weight2 + module.decoder.layers.13.mlp.experts.linear_fc1.weight0 + module.decoder.layers.12.mlp.experts.linear_fc2.weight15 + module.decoder.layers.12.mlp.experts.linear_fc2.weight13 + module.decoder.layers.12.mlp.experts.linear_fc2.weight11 + module.decoder.layers.12.mlp.experts.linear_fc2.weight9 + module.decoder.layers.12.mlp.experts.linear_fc2.weight7 +Params for bucket 66 (40894464 elements, 40894464 padded size): + module.decoder.layers.12.mlp.experts.linear_fc2.weight2 + module.decoder.layers.12.mlp.experts.linear_fc2.weight4 + module.decoder.layers.12.mlp.experts.linear_fc2.weight0 + module.decoder.layers.12.mlp.experts.linear_fc1.weight14 + module.decoder.layers.12.mlp.experts.linear_fc1.weight12 + module.decoder.layers.12.mlp.experts.linear_fc1.weight10 + module.decoder.layers.12.mlp.experts.linear_fc1.weight8 + module.decoder.layers.12.mlp.experts.linear_fc1.weight6 + 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.weight1 + module.decoder.layers.12.mlp.experts.linear_fc1.weight15 + module.decoder.layers.12.mlp.experts.linear_fc1.weight13 + module.decoder.layers.12.mlp.experts.linear_fc1.weight11 + module.decoder.layers.12.mlp.experts.linear_fc1.weight9 + module.decoder.layers.12.mlp.experts.linear_fc1.weight7 +Params for bucket 67 (40894464 elements, 40894464 padded size): + module.decoder.layers.12.mlp.experts.linear_fc1.weight2 + module.decoder.layers.11.mlp.experts.linear_fc2.weight9 + module.decoder.layers.11.mlp.experts.linear_fc2.weight14 + module.decoder.layers.11.mlp.experts.linear_fc2.weight6 + module.decoder.layers.12.mlp.experts.linear_fc1.weight4 + module.decoder.layers.11.mlp.experts.linear_fc2.weight11 + module.decoder.layers.11.mlp.experts.linear_fc2.weight3 + module.decoder.layers.12.mlp.experts.linear_fc1.weight1 + module.decoder.layers.11.mlp.experts.linear_fc2.weight8 + module.decoder.layers.11.mlp.experts.linear_fc2.weight13 + module.decoder.layers.11.mlp.experts.linear_fc2.weight5 + module.decoder.layers.12.mlp.experts.linear_fc1.weight3 + module.decoder.layers.11.mlp.experts.linear_fc2.weight10 + module.decoder.layers.11.mlp.experts.linear_fc2.weight2 + module.decoder.layers.12.mlp.experts.linear_fc1.weight0 + module.decoder.layers.11.mlp.experts.linear_fc2.weight15 + module.decoder.layers.11.mlp.experts.linear_fc2.weight7 + module.decoder.layers.12.mlp.experts.linear_fc1.weight5 + module.decoder.layers.11.mlp.experts.linear_fc2.weight12 + module.decoder.layers.11.mlp.experts.linear_fc2.weight4 +Params for bucket 68 (40894464 elements, 40894464 padded size): + module.decoder.layers.11.mlp.experts.linear_fc2.weight1 + module.decoder.layers.11.mlp.experts.linear_fc1.weight14 + module.decoder.layers.11.mlp.experts.linear_fc1.weight12 + module.decoder.layers.11.mlp.experts.linear_fc1.weight10 + module.decoder.layers.11.mlp.experts.linear_fc1.weight8 + module.decoder.layers.11.mlp.experts.linear_fc1.weight6 + module.decoder.layers.11.mlp.experts.linear_fc1.weight4 + module.decoder.layers.11.mlp.experts.linear_fc1.weight15 + module.decoder.layers.11.mlp.experts.linear_fc2.weight0 + module.decoder.layers.11.mlp.experts.linear_fc1.weight13 + module.decoder.layers.11.mlp.experts.linear_fc1.weight11 + module.decoder.layers.11.mlp.experts.linear_fc1.weight9 + module.decoder.layers.11.mlp.experts.linear_fc1.weight7 + module.decoder.layers.11.mlp.experts.linear_fc1.weight5 +Params for bucket 69 (40894464 elements, 40894464 padded size): + module.decoder.layers.10.mlp.experts.linear_fc2.weight10 + module.decoder.layers.10.mlp.experts.linear_fc2.weight2 + module.decoder.layers.11.mlp.experts.linear_fc1.weight1 + module.decoder.layers.10.mlp.experts.linear_fc2.weight15 + module.decoder.layers.10.mlp.experts.linear_fc2.weight7 + module.decoder.layers.10.mlp.experts.linear_fc2.weight12 + module.decoder.layers.10.mlp.experts.linear_fc2.weight4 + module.decoder.layers.10.mlp.experts.linear_fc1.weight15 + module.decoder.layers.11.mlp.experts.linear_fc1.weight3 + module.decoder.layers.10.mlp.experts.linear_fc2.weight9 + module.decoder.layers.10.mlp.experts.linear_fc2.weight1 + module.decoder.layers.11.mlp.experts.linear_fc1.weight0 + module.decoder.layers.10.mlp.experts.linear_fc2.weight14 + module.decoder.layers.10.mlp.experts.linear_fc2.weight6 + module.decoder.layers.10.mlp.experts.linear_fc2.weight11 + module.decoder.layers.10.mlp.experts.linear_fc2.weight3 + module.decoder.layers.11.mlp.experts.linear_fc1.weight2 + module.decoder.layers.10.mlp.experts.linear_fc2.weight8 + module.decoder.layers.10.mlp.experts.linear_fc2.weight0 + module.decoder.layers.10.mlp.experts.linear_fc2.weight13 + module.decoder.layers.10.mlp.experts.linear_fc2.weight5 +Params for bucket 70 (40894464 elements, 40894464 padded size): + module.decoder.layers.10.mlp.experts.linear_fc1.weight13 + module.decoder.layers.10.mlp.experts.linear_fc1.weight11 + module.decoder.layers.10.mlp.experts.linear_fc1.weight9 + module.decoder.layers.10.mlp.experts.linear_fc1.weight7 + 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.weight14 + module.decoder.layers.10.mlp.experts.linear_fc1.weight12 + module.decoder.layers.10.mlp.experts.linear_fc1.weight10 + module.decoder.layers.10.mlp.experts.linear_fc1.weight8 + module.decoder.layers.10.mlp.experts.linear_fc1.weight6 + module.decoder.layers.10.mlp.experts.linear_fc1.weight4 + module.decoder.layers.10.mlp.experts.linear_fc1.weight2 +Params for bucket 71 (40894464 elements, 40894464 padded size): + module.decoder.layers.9.mlp.experts.linear_fc2.weight12 + module.decoder.layers.9.mlp.experts.linear_fc2.weight4 + module.decoder.layers.9.mlp.experts.linear_fc1.weight15 + module.decoder.layers.9.mlp.experts.linear_fc2.weight9 + module.decoder.layers.9.mlp.experts.linear_fc2.weight1 + module.decoder.layers.9.mlp.experts.linear_fc2.weight14 + module.decoder.layers.9.mlp.experts.linear_fc2.weight6 + module.decoder.layers.9.mlp.experts.linear_fc2.weight11 + module.decoder.layers.9.mlp.experts.linear_fc2.weight3 + module.decoder.layers.9.mlp.experts.linear_fc1.weight14 + module.decoder.layers.10.mlp.experts.linear_fc1.weight1 + module.decoder.layers.9.mlp.experts.linear_fc2.weight8 + module.decoder.layers.9.mlp.experts.linear_fc2.weight0 + module.decoder.layers.9.mlp.experts.linear_fc2.weight13 + module.decoder.layers.9.mlp.experts.linear_fc2.weight5 + module.decoder.layers.9.mlp.experts.linear_fc2.weight10 + module.decoder.layers.9.mlp.experts.linear_fc2.weight2 + module.decoder.layers.9.mlp.experts.linear_fc1.weight13 + module.decoder.layers.9.mlp.experts.linear_fc2.weight15 + module.decoder.layers.10.mlp.experts.linear_fc1.weight0 + module.decoder.layers.9.mlp.experts.linear_fc2.weight7 +Params for bucket 72 (40894464 elements, 40894464 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight11 + module.decoder.layers.9.mlp.experts.linear_fc1.weight9 + module.decoder.layers.9.mlp.experts.linear_fc1.weight7 + 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.weight1 + module.decoder.layers.9.mlp.experts.linear_fc1.weight12 + module.decoder.layers.9.mlp.experts.linear_fc1.weight8 + module.decoder.layers.9.mlp.experts.linear_fc1.weight10 + module.decoder.layers.9.mlp.experts.linear_fc1.weight6 + module.decoder.layers.9.mlp.experts.linear_fc1.weight4 + module.decoder.layers.9.mlp.experts.linear_fc1.weight2 + module.decoder.layers.9.mlp.experts.linear_fc1.weight0 +Params for bucket 73 (40894464 elements, 40894464 padded size): + module.decoder.layers.8.mlp.experts.linear_fc2.weight14 + module.decoder.layers.8.mlp.experts.linear_fc2.weight6 + module.decoder.layers.8.mlp.experts.linear_fc2.weight11 + module.decoder.layers.8.mlp.experts.linear_fc2.weight3 + module.decoder.layers.8.mlp.experts.linear_fc1.weight14 + module.decoder.layers.8.mlp.experts.linear_fc2.weight8 + module.decoder.layers.8.mlp.experts.linear_fc2.weight0 + module.decoder.layers.8.mlp.experts.linear_fc1.weight11 + module.decoder.layers.8.mlp.experts.linear_fc2.weight13 + module.decoder.layers.8.mlp.experts.linear_fc2.weight5 + module.decoder.layers.8.mlp.experts.linear_fc2.weight10 + module.decoder.layers.8.mlp.experts.linear_fc2.weight2 + module.decoder.layers.8.mlp.experts.linear_fc1.weight13 + module.decoder.layers.8.mlp.experts.linear_fc2.weight15 + module.decoder.layers.8.mlp.experts.linear_fc2.weight7 + module.decoder.layers.8.mlp.experts.linear_fc2.weight12 + module.decoder.layers.8.mlp.experts.linear_fc2.weight4 + module.decoder.layers.8.mlp.experts.linear_fc1.weight15 + module.decoder.layers.8.mlp.experts.linear_fc2.weight9 + module.decoder.layers.8.mlp.experts.linear_fc2.weight1 + module.decoder.layers.8.mlp.experts.linear_fc1.weight12 +Params for bucket 74 (40894464 elements, 40894464 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight9 + module.decoder.layers.8.mlp.experts.linear_fc1.weight7 + module.decoder.layers.8.mlp.experts.linear_fc1.weight5 + module.decoder.layers.8.mlp.experts.linear_fc1.weight3 + module.decoder.layers.8.mlp.experts.linear_fc1.weight1 + module.decoder.layers.7.mlp.experts.linear_fc2.weight14 + module.decoder.layers.7.mlp.experts.linear_fc2.weight12 + module.decoder.layers.8.mlp.experts.linear_fc1.weight10 + module.decoder.layers.8.mlp.experts.linear_fc1.weight8 + module.decoder.layers.8.mlp.experts.linear_fc1.weight6 + module.decoder.layers.8.mlp.experts.linear_fc1.weight4 + module.decoder.layers.8.mlp.experts.linear_fc1.weight2 + module.decoder.layers.8.mlp.experts.linear_fc1.weight0 + module.decoder.layers.7.mlp.experts.linear_fc2.weight15 + module.decoder.layers.7.mlp.experts.linear_fc2.weight13 +Params for bucket 75 (40894464 elements, 40894464 padded size): + module.decoder.layers.7.mlp.experts.linear_fc2.weight8 + module.decoder.layers.7.mlp.experts.linear_fc2.weight0 + module.decoder.layers.7.mlp.experts.linear_fc1.weight11 + module.decoder.layers.7.mlp.experts.linear_fc2.weight5 + module.decoder.layers.7.mlp.experts.linear_fc2.weight10 + module.decoder.layers.7.mlp.experts.linear_fc2.weight2 + module.decoder.layers.7.mlp.experts.linear_fc1.weight13 + module.decoder.layers.7.mlp.experts.linear_fc2.weight7 + module.decoder.layers.7.mlp.experts.linear_fc1.weight10 + module.decoder.layers.7.mlp.experts.linear_fc2.weight4 + module.decoder.layers.7.mlp.experts.linear_fc1.weight15 + module.decoder.layers.7.mlp.experts.linear_fc2.weight9 + module.decoder.layers.7.mlp.experts.linear_fc2.weight1 + module.decoder.layers.7.mlp.experts.linear_fc1.weight12 + module.decoder.layers.7.mlp.experts.linear_fc2.weight6 + module.decoder.layers.7.mlp.experts.linear_fc1.weight9 + module.decoder.layers.7.mlp.experts.linear_fc2.weight11 + module.decoder.layers.7.mlp.experts.linear_fc2.weight3 + module.decoder.layers.7.mlp.experts.linear_fc1.weight14 +Params for bucket 76 (40894464 elements, 40894464 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight5 + module.decoder.layers.7.mlp.experts.linear_fc1.weight7 + module.decoder.layers.7.mlp.experts.linear_fc1.weight3 + module.decoder.layers.7.mlp.experts.linear_fc1.weight1 + module.decoder.layers.6.mlp.experts.linear_fc2.weight14 + module.decoder.layers.6.mlp.experts.linear_fc2.weight12 + module.decoder.layers.6.mlp.experts.linear_fc2.weight10 + module.decoder.layers.6.mlp.experts.linear_fc2.weight8 + module.decoder.layers.7.mlp.experts.linear_fc1.weight8 + module.decoder.layers.7.mlp.experts.linear_fc1.weight4 + module.decoder.layers.7.mlp.experts.linear_fc1.weight6 + module.decoder.layers.7.mlp.experts.linear_fc1.weight2 + module.decoder.layers.7.mlp.experts.linear_fc1.weight0 + module.decoder.layers.6.mlp.experts.linear_fc2.weight15 + module.decoder.layers.6.mlp.experts.linear_fc2.weight13 + module.decoder.layers.6.mlp.experts.linear_fc2.weight11 + module.decoder.layers.6.mlp.experts.linear_fc2.weight9 +Params for bucket 77 (40894464 elements, 40894464 padded size): + module.decoder.layers.6.mlp.experts.linear_fc2.weight6 + module.decoder.layers.6.mlp.experts.linear_fc2.weight4 + module.decoder.layers.6.mlp.experts.linear_fc2.weight2 + module.decoder.layers.6.mlp.experts.linear_fc2.weight0 + module.decoder.layers.6.mlp.experts.linear_fc1.weight15 + module.decoder.layers.6.mlp.experts.linear_fc1.weight13 + module.decoder.layers.6.mlp.experts.linear_fc1.weight11 + module.decoder.layers.6.mlp.experts.linear_fc1.weight9 + module.decoder.layers.6.mlp.experts.linear_fc1.weight7 + module.decoder.layers.6.mlp.experts.linear_fc2.weight7 + module.decoder.layers.6.mlp.experts.linear_fc2.weight3 + module.decoder.layers.6.mlp.experts.linear_fc2.weight5 + module.decoder.layers.6.mlp.experts.linear_fc2.weight1 + module.decoder.layers.6.mlp.experts.linear_fc1.weight14 + module.decoder.layers.6.mlp.experts.linear_fc1.weight12 + module.decoder.layers.6.mlp.experts.linear_fc1.weight10 + module.decoder.layers.6.mlp.experts.linear_fc1.weight8 +Params for bucket 78 (40894464 elements, 40894464 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight5 + module.decoder.layers.5.mlp.experts.linear_fc2.weight11 + module.decoder.layers.6.mlp.experts.linear_fc1.weight2 + module.decoder.layers.5.mlp.experts.linear_fc2.weight8 + module.decoder.layers.5.mlp.experts.linear_fc2.weight13 + module.decoder.layers.5.mlp.experts.linear_fc2.weight5 + module.decoder.layers.6.mlp.experts.linear_fc1.weight4 + module.decoder.layers.5.mlp.experts.linear_fc2.weight10 + module.decoder.layers.6.mlp.experts.linear_fc1.weight1 + module.decoder.layers.5.mlp.experts.linear_fc2.weight15 + module.decoder.layers.5.mlp.experts.linear_fc2.weight7 + module.decoder.layers.6.mlp.experts.linear_fc1.weight6 + module.decoder.layers.5.mlp.experts.linear_fc2.weight12 + module.decoder.layers.5.mlp.experts.linear_fc2.weight4 + module.decoder.layers.6.mlp.experts.linear_fc1.weight3 + module.decoder.layers.5.mlp.experts.linear_fc2.weight9 + module.decoder.layers.6.mlp.experts.linear_fc1.weight0 + module.decoder.layers.5.mlp.experts.linear_fc2.weight14 + module.decoder.layers.5.mlp.experts.linear_fc2.weight6 +Params for bucket 79 (40894464 elements, 40894464 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_fc1.weight14 + module.decoder.layers.5.mlp.experts.linear_fc1.weight12 + module.decoder.layers.5.mlp.experts.linear_fc1.weight10 + module.decoder.layers.5.mlp.experts.linear_fc1.weight8 + module.decoder.layers.5.mlp.experts.linear_fc1.weight6 + module.decoder.layers.5.mlp.experts.linear_fc2.weight2 + module.decoder.layers.5.mlp.experts.linear_fc1.weight15 + module.decoder.layers.5.mlp.experts.linear_fc2.weight0 + module.decoder.layers.5.mlp.experts.linear_fc1.weight13 + module.decoder.layers.5.mlp.experts.linear_fc1.weight11 + module.decoder.layers.5.mlp.experts.linear_fc1.weight9 + module.decoder.layers.5.mlp.experts.linear_fc1.weight7 + module.decoder.layers.5.mlp.experts.linear_fc1.weight5 +Params for bucket 80 (40894464 elements, 40894464 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight15 + module.decoder.layers.4.mlp.experts.linear_fc2.weight7 + module.decoder.layers.5.mlp.experts.linear_fc1.weight3 + module.decoder.layers.4.mlp.experts.linear_fc2.weight12 + module.decoder.layers.4.mlp.experts.linear_fc2.weight4 + module.decoder.layers.5.mlp.experts.linear_fc1.weight0 + module.decoder.layers.4.mlp.experts.linear_fc2.weight9 + module.decoder.layers.4.mlp.experts.linear_fc2.weight1 + module.decoder.layers.4.mlp.experts.linear_fc2.weight14 + module.decoder.layers.4.mlp.experts.linear_fc2.weight6 + module.decoder.layers.5.mlp.experts.linear_fc1.weight2 + module.decoder.layers.4.mlp.experts.linear_fc2.weight11 + module.decoder.layers.4.mlp.experts.linear_fc2.weight3 + module.decoder.layers.4.mlp.experts.linear_fc2.weight8 + module.decoder.layers.4.mlp.experts.linear_fc2.weight0 + module.decoder.layers.5.mlp.experts.linear_fc1.weight4 + module.decoder.layers.4.mlp.experts.linear_fc2.weight13 + module.decoder.layers.4.mlp.experts.linear_fc2.weight5 + module.decoder.layers.5.mlp.experts.linear_fc1.weight1 + module.decoder.layers.4.mlp.experts.linear_fc2.weight10 + module.decoder.layers.4.mlp.experts.linear_fc2.weight2 +Params for bucket 81 (40894464 elements, 40894464 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight14 + module.decoder.layers.4.mlp.experts.linear_fc1.weight12 + module.decoder.layers.4.mlp.experts.linear_fc1.weight10 + module.decoder.layers.4.mlp.experts.linear_fc1.weight8 + module.decoder.layers.4.mlp.experts.linear_fc1.weight6 + module.decoder.layers.4.mlp.experts.linear_fc1.weight4 + module.decoder.layers.4.mlp.experts.linear_fc1.weight15 + module.decoder.layers.4.mlp.experts.linear_fc1.weight13 + module.decoder.layers.4.mlp.experts.linear_fc1.weight11 + module.decoder.layers.4.mlp.experts.linear_fc1.weight9 + module.decoder.layers.4.mlp.experts.linear_fc1.weight7 + module.decoder.layers.4.mlp.experts.linear_fc1.weight5 + module.decoder.layers.4.mlp.experts.linear_fc1.weight3 +Params for bucket 82 (40894464 elements, 40894464 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight2 + module.decoder.layers.3.mlp.experts.linear_fc2.weight9 + module.decoder.layers.3.mlp.experts.linear_fc2.weight1 + module.decoder.layers.3.mlp.experts.linear_fc2.weight14 + module.decoder.layers.3.mlp.experts.linear_fc2.weight6 + module.decoder.layers.3.mlp.experts.linear_fc2.weight11 + module.decoder.layers.3.mlp.experts.linear_fc2.weight3 + module.decoder.layers.3.mlp.experts.linear_fc1.weight14 + module.decoder.layers.4.mlp.experts.linear_fc1.weight1 + module.decoder.layers.3.mlp.experts.linear_fc2.weight8 + module.decoder.layers.3.mlp.experts.linear_fc2.weight0 + module.decoder.layers.3.mlp.experts.linear_fc2.weight13 + module.decoder.layers.3.mlp.experts.linear_fc2.weight5 + module.decoder.layers.3.mlp.experts.linear_fc2.weight10 + module.decoder.layers.3.mlp.experts.linear_fc2.weight2 + module.decoder.layers.4.mlp.experts.linear_fc1.weight0 + module.decoder.layers.3.mlp.experts.linear_fc2.weight15 + module.decoder.layers.3.mlp.experts.linear_fc2.weight7 + module.decoder.layers.3.mlp.experts.linear_fc2.weight12 + module.decoder.layers.3.mlp.experts.linear_fc2.weight4 + module.decoder.layers.3.mlp.experts.linear_fc1.weight15 +Params for bucket 83 (40894464 elements, 40894464 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight10 + module.decoder.layers.3.mlp.experts.linear_fc1.weight12 + module.decoder.layers.3.mlp.experts.linear_fc1.weight8 + module.decoder.layers.3.mlp.experts.linear_fc1.weight6 + module.decoder.layers.3.mlp.experts.linear_fc1.weight4 + module.decoder.layers.3.mlp.experts.linear_fc1.weight2 + module.decoder.layers.3.mlp.experts.linear_fc1.weight13 + module.decoder.layers.3.mlp.experts.linear_fc1.weight11 + module.decoder.layers.3.mlp.experts.linear_fc1.weight9 + module.decoder.layers.3.mlp.experts.linear_fc1.weight7 + 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.weight1 +Params for bucket 84 (40894464 elements, 40894464 padded size): + module.decoder.layers.2.mlp.experts.linear_fc2.weight11 + module.decoder.layers.2.mlp.experts.linear_fc2.weight3 + module.decoder.layers.2.mlp.experts.linear_fc1.weight14 + module.decoder.layers.2.mlp.experts.linear_fc2.weight8 + module.decoder.layers.2.mlp.experts.linear_fc2.weight0 + module.decoder.layers.2.mlp.experts.linear_fc2.weight13 + module.decoder.layers.2.mlp.experts.linear_fc2.weight5 + module.decoder.layers.2.mlp.experts.linear_fc2.weight10 + module.decoder.layers.2.mlp.experts.linear_fc2.weight2 + module.decoder.layers.2.mlp.experts.linear_fc1.weight13 + module.decoder.layers.3.mlp.experts.linear_fc1.weight0 + module.decoder.layers.2.mlp.experts.linear_fc2.weight15 + module.decoder.layers.2.mlp.experts.linear_fc2.weight7 + module.decoder.layers.2.mlp.experts.linear_fc2.weight12 + module.decoder.layers.2.mlp.experts.linear_fc2.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight15 + module.decoder.layers.2.mlp.experts.linear_fc2.weight9 + module.decoder.layers.2.mlp.experts.linear_fc2.weight1 + module.decoder.layers.2.mlp.experts.linear_fc1.weight12 + module.decoder.layers.2.mlp.experts.linear_fc2.weight14 + module.decoder.layers.2.mlp.experts.linear_fc2.weight6 +Params for bucket 85 (40894464 elements, 40894464 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight8 + module.decoder.layers.2.mlp.experts.linear_fc1.weight10 + module.decoder.layers.2.mlp.experts.linear_fc1.weight6 + module.decoder.layers.2.mlp.experts.linear_fc1.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight2 + module.decoder.layers.2.mlp.experts.linear_fc1.weight0 + module.decoder.layers.1.mlp.experts.linear_fc2.weight14 + module.decoder.layers.2.mlp.experts.linear_fc1.weight11 + module.decoder.layers.2.mlp.experts.linear_fc1.weight9 + module.decoder.layers.2.mlp.experts.linear_fc1.weight7 + module.decoder.layers.2.mlp.experts.linear_fc1.weight5 + module.decoder.layers.2.mlp.experts.linear_fc1.weight3 + module.decoder.layers.2.mlp.experts.linear_fc1.weight1 + module.decoder.layers.1.mlp.experts.linear_fc2.weight15 +Params for bucket 86 (40894464 elements, 40894464 padded size): + module.decoder.layers.1.mlp.experts.linear_fc2.weight12 + module.decoder.layers.1.mlp.experts.linear_fc2.weight4 + module.decoder.layers.1.mlp.experts.linear_fc1.weight15 + module.decoder.layers.1.mlp.experts.linear_fc2.weight9 + module.decoder.layers.1.mlp.experts.linear_fc2.weight1 + module.decoder.layers.1.mlp.experts.linear_fc1.weight12 + module.decoder.layers.1.mlp.experts.linear_fc2.weight6 + module.decoder.layers.1.mlp.experts.linear_fc2.weight11 + module.decoder.layers.1.mlp.experts.linear_fc2.weight3 + module.decoder.layers.1.mlp.experts.linear_fc1.weight14 + module.decoder.layers.1.mlp.experts.linear_fc2.weight8 + module.decoder.layers.1.mlp.experts.linear_fc2.weight0 + module.decoder.layers.1.mlp.experts.linear_fc1.weight11 + module.decoder.layers.1.mlp.experts.linear_fc2.weight13 + module.decoder.layers.1.mlp.experts.linear_fc2.weight5 + module.decoder.layers.1.mlp.experts.linear_fc2.weight10 + module.decoder.layers.1.mlp.experts.linear_fc2.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight13 + module.decoder.layers.1.mlp.experts.linear_fc2.weight7 + module.decoder.layers.1.mlp.experts.linear_fc1.weight10 +Params for bucket 87 (40894464 elements, 40894464 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight9 + module.decoder.layers.1.mlp.experts.linear_fc1.weight7 + 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.weight1 + module.decoder.layers.0.mlp.experts.linear_fc2.weight14 + module.decoder.layers.0.mlp.experts.linear_fc2.weight12 + module.decoder.layers.0.mlp.experts.linear_fc2.weight10 + module.decoder.layers.1.mlp.experts.linear_fc1.weight8 + module.decoder.layers.1.mlp.experts.linear_fc1.weight6 + module.decoder.layers.1.mlp.experts.linear_fc1.weight4 + module.decoder.layers.1.mlp.experts.linear_fc1.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight0 + module.decoder.layers.0.mlp.experts.linear_fc2.weight15 + module.decoder.layers.0.mlp.experts.linear_fc2.weight13 + module.decoder.layers.0.mlp.experts.linear_fc2.weight11 +Params for bucket 88 (40894464 elements, 40894464 padded size): + module.decoder.layers.0.mlp.experts.linear_fc2.weight6 + module.decoder.layers.0.mlp.experts.linear_fc2.weight8 + 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.weight0 + module.decoder.layers.0.mlp.experts.linear_fc1.weight14 + module.decoder.layers.0.mlp.experts.linear_fc1.weight12 + module.decoder.layers.0.mlp.experts.linear_fc1.weight10 + module.decoder.layers.0.mlp.experts.linear_fc1.weight8 + module.decoder.layers.0.mlp.experts.linear_fc2.weight9 + module.decoder.layers.0.mlp.experts.linear_fc2.weight5 + module.decoder.layers.0.mlp.experts.linear_fc2.weight7 + module.decoder.layers.0.mlp.experts.linear_fc2.weight3 + module.decoder.layers.0.mlp.experts.linear_fc2.weight1 + module.decoder.layers.0.mlp.experts.linear_fc1.weight15 + module.decoder.layers.0.mlp.experts.linear_fc1.weight13 + module.decoder.layers.0.mlp.experts.linear_fc1.weight11 + module.decoder.layers.0.mlp.experts.linear_fc1.weight9 +Params for bucket 89 (25165824 elements, 25165824 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight4 + module.decoder.layers.0.mlp.experts.linear_fc1.weight6 + 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.weight7 + 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.weight1 +INFO:megatron.core.optimizer:Setting up optimizer with config OptimizerConfig(optimizer='adam', lr=0.0001, min_lr=1e-05, decoupled_lr=None, decoupled_min_lr=None, weight_decay=0.1, fp8_recipe='mxfp8', fp16=False, bf16=True, reuse_grad_buf_for_mxfp8_param_ag=False, params_dtype=torch.bfloat16, use_precision_aware_optimizer=False, store_param_remainders=True, main_grads_dtype=torch.float32, main_params_dtype=torch.float32, exp_avg_dtype=torch.float32, exp_avg_sq_dtype=torch.float32, loss_scale=None, initial_loss_scale=4294967296, min_loss_scale=1.0, loss_scale_window=1000, hysteresis=2, adam_beta1=0.9, adam_beta2=0.999, adam_eps=1e-08, sgd_momentum=0.9, use_distributed_optimizer=True, overlap_param_gather_with_optimizer_step=False, optimizer_cpu_offload=False, optimizer_offload_fraction=1.0, use_torch_optimizer_for_cpu_offload=False, overlap_cpu_optimizer_d2h_h2d=False, pin_cpu_grads=True, pin_cpu_params=True, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=, config_logger_dir='') +INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine +[after model, optimizer, and learning rate scheduler are built] datetime: 2026-08-26 18:31:45 +> building train, validation, and test datasets ... + > datasets target sizes (minimum size): + train: 96 + 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=(96, 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 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-document_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-sample_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-shuffle_index.npy +INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073 +> finished creating GPT datasets ... +[after dataloaders are built] datetime: 2026-08-26 18:31:45 +done with setup ... +(min, max) time across ranks (ms): + model-and-optimizer-setup ......................: (4588.56, 4604.44) + train/valid/test-data-iterators-setup ..........: (68.06, 82.90) +training ... +Overwriting rerun_state_machine.current_iteration from -1 to 0... +[before the start of training step] datetime: 2026-08-26 18:31:45 +/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:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") + [2026-08-26 18:32:12] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 27009.1 | throughput per GPU (TFLOP/s/GPU): 3.3 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.095286E+01 | loss scale: 1.0 | grad norm: 54.365 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | +Number of parameters in transformer block in billions: 29.90 +Number of parameters in embedding layers in billions: 0.21 +Total number of parameters in billions: 30.10 +Number of parameters in most loaded shard in billions: 30.1034 +Theoretical memory footprints: weight and optimizer=215316.59 MB +[Rank 0] (after 1 iterations) memory (MB) | allocated: 72365.84375 | max allocated: 78646.15576171875 | reserved: 79250.0 | max reserved: 79250.0 +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") + [2026-08-26 18:32:13] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 1170.4 | throughput per GPU (TFLOP/s/GPU): 76.1 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.095473E+01 | loss scale: 1.0 | grad norm: 56.503 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:32:14] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 923.8 | throughput per GPU (TFLOP/s/GPU): 96.5 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 1.008747E+01 | loss scale: 1.0 | grad norm: 180.743 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | +Capture CUDA graph for training!!! +CUDA graph capture done!!! + [2026-08-26 18:32:17] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 2938.4 | throughput per GPU (TFLOP/s/GPU): 30.3 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 9.604395E+00 | loss scale: 1.0 | grad norm: 8.738 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:32:18] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 357.2 | throughput per GPU (TFLOP/s/GPU): 249.5 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 9.307559E+00 | loss scale: 1.0 | grad norm: 5.715 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:32:18] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 355.5 | throughput per GPU (TFLOP/s/GPU): 250.7 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 9.181961E+00 | loss scale: 1.0 | grad norm: 3.720 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:32:18] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 353.5 | throughput per GPU (TFLOP/s/GPU): 252.1 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 9.030100E+00 | loss scale: 1.0 | grad norm: 3.447 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:32:19] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 353.4 | throughput per GPU (TFLOP/s/GPU): 252.2 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.855101E+00 | loss scale: 1.0 | grad norm: 3.239 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:32:19] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 352.4 | throughput per GPU (TFLOP/s/GPU): 252.9 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.799657E+00 | loss scale: 1.0 | grad norm: 2.982 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:32:19] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 351.9 | throughput per GPU (TFLOP/s/GPU): 253.2 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.760715E+00 | loss scale: 1.0 | grad norm: 2.970 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:32:20] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 352.2 | throughput per GPU (TFLOP/s/GPU): 253.0 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.752829E+00 | loss scale: 1.0 | grad norm: 2.821 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:32:20] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 351.4 | throughput per GPU (TFLOP/s/GPU): 253.6 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.726992E+00 | loss scale: 1.0 | grad norm: 2.686 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | +[after training is done] datetime: 2026-08-26 18:32:20 +[rank0]:[W826 18:32:21.793817868 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank2]:[W826 18:32:23.058821644 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank6]:[W826 18:32:23.077364618 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank4]:[W826 18:32:23.078672645 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank3]:[W826 18:32:23.080710808 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank7]:[W826 18:32:23.081042434 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank1]:[W826 18:32:23.081154191 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank5]:[W826 18:32:23.875018085 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +W0826 18:32:25.004000 519501 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers +W0826 18:32:25.005000 519501 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 519590 closing signal SIGTERM +W0826 18:32:25.005000 519501 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 519591 closing signal SIGTERM +W0826 18:32:25.008000 519501 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 519592 closing signal SIGTERM +W0826 18:32:25.009000 519501 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 519593 closing signal SIGTERM +W0826 18:32:25.010000 519501 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 519594 closing signal SIGTERM +W0826 18:32:25.011000 519501 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 519595 closing signal SIGTERM +W0826 18:32:25.013000 519501 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 519596 closing signal SIGTERM +W0826 18:32:25.015000 519501 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 519597 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 519501 got signal: 15 diff --git a/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep2/mem_samples.csv b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep2/mem_samples.csv new file mode 100644 index 0000000000000000000000000000000000000000..8f5cd379b2718b6e85e42cb5ad60fef5615114eb --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep2/mem_samples.csv @@ -0,0 +1,232 @@ +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 1 +1, 1 +2, 1 +3, 1 +4, 1 +5, 1 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 1 +5, 1 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 4 +5, 4 +6, 4 +7, 4 +0, 4 +1, 4 +2, 4 +3, 4 +4, 35 +5, 4 +6, 4 +7, 4 +0, 821 +1, 821 +2, 821 +3, 821 +4, 821 +5, 821 +6, 821 +7, 821 +0, 1831 +1, 1831 +2, 1831 +3, 1831 +4, 1831 +5, 1831 +6, 1831 +7, 1831 +0, 2171 +1, 2195 +2, 2191 +3, 2183 +4, 2183 +5, 2205 +6, 2195 +7, 2201 +0, 10591 +1, 10631 +2, 10431 +3, 10451 +4, 10571 +5, 10631 +6, 10251 +7, 10631 +0, 47051 +1, 47059 +2, 47015 +3, 47027 +4, 47003 +5, 47077 +6, 47039 +7, 47055 +0, 48477 +1, 48479 +2, 48489 +3, 48451 +4, 48475 +5, 48489 +6, 48479 +7, 48489 +0, 48647 +1, 48649 +2, 48659 +3, 48621 +4, 48645 +5, 48659 +6, 48649 +7, 48659 +0, 48791 +1, 48793 +2, 48803 +3, 48765 +4, 48789 +5, 48803 +6, 48793 +7, 48803 +0, 49447 +1, 49447 +2, 49451 +3, 49421 +4, 49419 +5, 49451 +6, 49405 +7, 49451 +0, 59355 +1, 59503 +2, 59501 +3, 59389 +4, 59387 +5, 59559 +6, 59539 +7, 59529 +0, 85147 +1, 85155 +2, 85155 +3, 85175 +4, 85145 +5, 85125 +6, 85149 +7, 85123 +0, 85607 +1, 85629 +2, 85605 +3, 85621 +4, 85621 +5, 85605 +6, 85605 +7, 85605 +0, 85607 +1, 85629 +2, 85605 +3, 85621 +4, 85621 +5, 85605 +6, 85605 +7, 85605 +0, 85869 +1, 85925 +2, 85843 +3, 85871 +4, 85901 +5, 85861 +6, 85871 +7, 85867 +0, 86919 +1, 86911 +2, 86927 +3, 86911 +4, 86865 +5, 86861 +6, 86871 +7, 86871 +0, 107195 +1, 107519 +2, 107885 +3, 107893 +4, 106799 +5, 107437 +6, 107361 +7, 107721 +0, 106561 +1, 106635 +2, 106819 +3, 106783 +4, 105027 +5, 105727 +6, 102315 +7, 105615 +0, 120103 +1, 120129 +2, 120099 +3, 119829 +4, 120121 +5, 120099 +6, 120101 +7, 120099 +0, 120103 +1, 120129 +2, 120099 +3, 119829 +4, 120121 +5, 120099 +6, 120101 +7, 120099 +0, 120103 +1, 120129 +2, 120099 +3, 119829 +4, 120121 +5, 120099 +6, 120101 +7, 120099 +0, 117099 +1, 116799 +2, 116423 +3, 116179 +4, 116535 +5, 117095 +6, 116835 +7, 117095 +0, 115239 +1, 114481 +2, 114445 +3, 114757 +4, 115003 +5, 115135 +6, 115093 +7, 115305 +0, 1 +1, 4 +2, 4 +3, 4 +4, 68 +5, 115027 +6, 324 +7, 4 diff --git a/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep2/run.env b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep2/run.env new file mode 100644 index 0000000000000000000000000000000000000000..3213d7fbe624a565f15157e5d63615da960f6547 --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep2/run.env @@ -0,0 +1,61 @@ +backend=ultraep +config=QWEN3_30B +family=tierB +repo=/workspace/Megatron-UltraEP +branch=ultraep-bracing-regions-through-kernels +commit=03d57ef7 +num_layers=48 +hidden=2048 +ffn_hidden=6144 +moe_ffn_hidden=768 +num_experts=128 +topk=8 +experts_per_rank=16 +expert_params=4718592 +heads=32 +kv_groups=4 +kv_channels=128 +qk_layernorm=1 +shared_ffn=0 +shared_expert_overlap=0 +mla=0 +dense_first_layers=0 +moe_layers=48 +rotary_base=1000000 +rotary_scaling= +qkv_bias=0 +router_balancing=aux_loss +aux_loss_coeff=1e-3 +router_pre_softmax=0 +router_fusion=0 +router_dtype=fp32 +device_initiated_grouped_gemm=1 +norm_epsilon=1e-6 +seq_length=4096 +micro_batch_size=1 +global_batch_size=8 +num_microbatches=1 +assignments_per_rank=32768 +mean_tokens_per_expert=2048 +slot_policy=matched +slots_per_rank=16 +capacity_mode=static +capacity=1.11328125 +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=12 +ep_size=8 +echo_dispatch_overlap=1 +echo_dispatch_recompute=0 +ultraep_quota_min_tokens=256 +ultraep_weight_sync_join= +profile_enable=0 +kernel_trace=0 +fp8_recipe=mxfp8 diff --git a/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep2/train.log b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep2/train.log new file mode 100644 index 0000000000000000000000000000000000000000..469735f23656488c1c67349626e7e0de1794f543 --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep2/train.log @@ -0,0 +1,2878 @@ +[18:32:30] WARN: ultraep-bracing-regions-through-kernels lacks --moe-router-fusion; QWEN3_30B runs unfused (timing only, routing identical) +============================================================== + QWEN3_30B / ultraep (slots/rank=16 policy=matched) + E=128 K=8 H'=768 m=16 + L=48 seq=4096 gbs=8 (1 microbatches) + cap_mode=static cap=1.11328125 cg=1 recompute=none + memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep2/memory_snapshots/snapshot.pickle + repo=/workspace/Megatron-UltraEP @ 03d57ef7 +============================================================== +W0826 18:32:31.532000 522564 torch/distributed/run.py:766] +W0826 18:32:31.532000 522564 torch/distributed/run.py:766] ***************************************** +W0826 18:32:31.532000 522564 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0826 18:32:31.532000 522564 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") +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 .............................. 48 + 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 ................................. 6144 + finetune ........................................ False + finetune_data_split ............................. train + finetune_hf_dataset ............................. None + first_last_layers_bf16 .......................... False + flash_decode .................................... False + fp16 ............................................ False + fp16_lm_cross_entropy ........................... False + fp32_residual_connection ........................ False + fp8 ............................................. e4m3 + fp8_amax_compute_algo ........................... most_recent + fp8_amax_history_len ............................ 1 + fp8_interval .................................... 1 + fp8_margin ...................................... 0 + fp8_param_gather ................................ False + fp8_recipe ...................................... mxfp8 + fp8_wgrad ....................................... True + fsdp_double_buffer .............................. False + global_batch_size ............................... 8 + grad_reduce_in_bf16 ............................. False + gradient_accumulation_fusion .................... True + gradient_reduce_div_fusion ...................... True + group_query_attention ........................... True + head_lr_mult .................................... 1.0 + heterogeneous_layers_config_encoded_json ........ None + heterogeneous_layers_config_path ................ None + hidden_dropout .................................. 0.0 + hidden_size ..................................... 2048 + hierarchical_context_parallel_sizes ............. None + high_priority_stream_groups ..................... [] + hybrid_attention_ratio .......................... 0.0 + hybrid_mlp_ratio ................................ 0.0 + hybrid_override_pattern ......................... None + hysteresis ...................................... 2 + ict_head_size ................................... None + ict_load ........................................ None + img_h ........................................... 224 + img_w ........................................... 224 + indexer_batch_size .............................. 128 + indexer_log_interval ............................ 1000 + inference_batch_times_seqlen_threshold .......... -1 + inference_dynamic_batching ...................... False + inference_dynamic_batching_buffer_guaranteed_fraction 0.2 + inference_dynamic_batching_buffer_overflow_factor None + inference_dynamic_batching_buffer_size_gb ....... 40.0 + inference_dynamic_batching_chunk_size ........... 256 + inference_dynamic_batching_max_requests_override None + inference_dynamic_batching_max_tokens_override .. None + inference_dynamic_batching_num_cuda_graphs ...... 16 + inference_max_batch_size ........................ 8 + inference_max_seq_length ........................ 2560 + inference_rng_tracker ........................... False + init_method_std ................................. 0.01 + init_method_xavier_uniform ...................... False + init_model_with_meta_device ..................... False + initial_loss_scale .............................. 4294967296 + inprocess_active_world_size ..................... 8 + inprocess_barrier_timeout ....................... 120 + inprocess_completion_timeout .................... 120 + inprocess_empty_cuda_cache ...................... False + inprocess_granularity ........................... node + inprocess_hard_timeout .......................... 90 + inprocess_heartbeat_interval .................... 30 + inprocess_heartbeat_timeout ..................... 60 + inprocess_last_call_wait ........................ 1 + inprocess_max_iterations ........................ None + inprocess_monitor_process_interval .............. 1.0 + inprocess_monitor_thread_interval ............... 1.0 + inprocess_progress_watchdog_interval ............ 1.0 + inprocess_restart ............................... False + inprocess_soft_timeout .......................... 60 + inprocess_termination_grace_time ................ 1 + is_hybrid_model ................................. False + iter_per_epoch .................................. 1250 + iterations_to_skip .............................. [] + keep_fp8_transpose_cache_when_using_custom_fsdp . False + kitchen_config_file ............................. None + kitchen_recipe_number ........................... None + kv_channels ..................................... 128 + kv_lora_rank .................................... 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 .............................. 0.001 + moe_capacity_monitor ............................ True + moe_deepep_num_sms .............................. 20 + moe_echo_enable_random_offloading ............... False + moe_echo_expert_dispatch_overlap ................ False + moe_echo_expert_dispatcher_type ................. hybridep + moe_echo_recompute_expert_dispatch .............. False + moe_enable_deepep ............................... False + moe_enable_echo ................................. False + moe_enable_hybridep ............................. True + moe_enable_ultraep .............................. True + moe_expert_capacity_factor ...................... None + moe_extended_tp ................................. False + moe_ffn_hidden_size ............................. 768 + 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 .............. 16 + moe_pad_expert_input_to_capacity ................ False + moe_per_layer_logging ........................... False + moe_permute_fusion .............................. False + moe_received_token_capacity ..................... 1.11328125 + 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-06 + normalization ................................... RMSNorm + num_attention_heads ............................. 32 + num_channels .................................... 3 + num_classes ..................................... 1000 + num_dataset_builder_threads ..................... 1 + num_distributed_optimizer_instances ............. 1 + num_experts ..................................... 128 + num_layers ...................................... 48 + num_layers_at_end_in_bf16 ....................... 1 + num_layers_at_start_in_bf16 ..................... 1 + num_layers_per_virtual_pipeline_stage ........... None + num_query_groups ................................ 4 + num_virtual_stages_per_pipeline_rank ............ None + num_workers ..................................... 2 + object_storage_cache_path ....................... None + one_logger_async ................................ False + one_logger_project .............................. megatron-lm + one_logger_run_name ............................. None + onnx_safe ....................................... None + openai_gelu ..................................... False + optimizer ....................................... adam + optimizer_cpu_offload ........................... False + optimizer_offload_fraction ...................... 1.0 + output_bert_embeddings .......................... False + overlap_cpu_optimizer_d2h_h2d ................... False + overlap_grad_reduce ............................. True + overlap_moe_expert_parallel_comm ................ False + overlap_p2p_comm ................................ False + overlap_p2p_comm_warmup_flush ................... False + overlap_param_gather ............................ True + overlap_param_gather_with_optimizer_step ........ False + override_opt_param_scheduler .................... False + params_dtype .................................... torch.bfloat16 + patch_dim ....................................... 16 + per_split_data_args_path ........................ None + perform_initialization .......................... True + pin_cpu_grads ................................... True + pin_cpu_params .................................. True + pipeline_model_parallel_comm_backend ............ None + pipeline_model_parallel_layout .................. None + pipeline_model_parallel_size .................... 1 + position_embedding_type ......................... rope + pretrained_checkpoint ........................... None + profile ......................................... False + profile_ranks ................................... [0] + profile_step_end ................................ 12 + profile_step_start .............................. 10 + q_lora_rank ..................................... None + qk_head_dim ..................................... 128 + qk_l2_norm ...................................... False + qk_layernorm .................................... True + qk_pos_emb_head_dim ............................. 64 + query_in_block_prob ............................. 0.1 + rampup_batch_size ............................... None + rank ............................................ 0 + recompute_granularity ........................... None + recompute_method ................................ None + recompute_modules ............................... None + recompute_num_layers ............................ None + record_memory_history ........................... False + relative_attention_max_distance ................. 128 + relative_attention_num_buckets .................. 32 + replication ..................................... False + replication_factor .............................. 2 + replication_jump ................................ None + rerun_mode ...................................... validate_results + reset_attention_mask ............................ False + reset_position_ids .............................. False + result_rejected_tracker_filename ................ None + retriever_report_topk_accuracies ................ [] + retriever_score_scaling ......................... False + retriever_seq_length ............................ 256 + retro_add_retriever ............................. False + retro_attention_gate ............................ 1 + retro_cyclic_train_iters ........................ None + retro_encoder_attention_dropout ................. 0.1 + retro_encoder_hidden_dropout .................... 0.1 + retro_encoder_layers ............................ 2 + retro_num_neighbors ............................. 2 + retro_num_retrieved_chunks ...................... 2 + retro_project_dir ............................... None + retro_verify_neighbor_count ..................... True + reuse_grad_buf_for_mxfp8_param_ag ............... False + rope_scaling_factor ............................. 8.0 + rope_type ....................................... None + rotary_base ..................................... 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 ..................................... 12 + train_samples ................................... None + train_sync_interval ............................. None + transformer_impl ................................ transformer_engine + transformer_pipeline_model_parallel_size ........ 1 + untie_embeddings_and_output_weights ............. True + use_checkpoint_args ............................. False + use_checkpoint_opt_param_scheduler .............. False + use_cpu_initialization .......................... None + use_custom_fsdp ................................. False + use_dist_ckpt ................................... True + use_dist_ckpt_deprecated ........................ False + use_distributed_optimizer ....................... True + use_flash_attn .................................. False + use_legacy_models ............................... False + use_mp_args_from_checkpoint_args ................ False + use_one_sent_docs ............................... False + use_persistent_ckpt_worker ...................... False + use_precision_aware_optimizer ................... False + use_pytorch_profiler ............................ False + use_ring_exchange_p2p ........................... False + use_rope_scaling ................................ False + use_rotary_position_embeddings .................. False + use_sharp ....................................... False + use_tokenizer_model_from_checkpoint_args ........ True + use_torch_fsdp2 ................................. False + use_torch_optimizer_for_cpu_offload ............. False + use_tp_pp_dp_mapping ............................ False + v_head_dim ...................................... 128 + valid_data_path ................................. None + variable_seq_lengths ............................ False + virtual_pipeline_model_parallel_size ............ None + vision_backbone_type ............................ vit + vision_pretraining .............................. False + vision_pretraining_type ......................... classify + vocab_extra_ids ................................. 0 + vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json + vocab_size ...................................... None + wandb_exp_name .................................. + wandb_project ................................... + wandb_save_dir .................................. + weight_decay .................................... 0.1 + weight_decay_incr_style ......................... constant + wgrad_deferral_limit ............................ 0 + world_size ...................................... 8 + yaml_cfg ........................................ None +-------------------- end of arguments --------------------- +INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1 +> building GPT2BPETokenizer tokenizer ... + > 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") +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 [Gloo] Rank 12 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 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank [Gloo] Rank 5 is connected to 77 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 + peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank [Gloo] Rank [Gloo] Rank 436 is connected to is connected to is connected to 777 peer ranks. peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : Expected number of connected peer ranks is : 777 + + +[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 42 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 5 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 6 is connected to 77 is connected to peer ranks. 7Expected number of connected peer ranks is : peer ranks. 7Expected number of connected peer ranks is : +7 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +> initialized tensor model parallel with size 1 +> initialized pipeline model parallel with size 1 +> setting random seeds to 1234 ... +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0[Gloo] Rank is connected to [Gloo] Rank 000 is connected to is connected to peer ranks. 00Expected number of connected peer ranks is : peer ranks. peer ranks. 0Expected number of connected peer ranks is : Expected number of connected peer ranks is : +00 + +[Gloo] Rank [Gloo] Rank 00 is connected to is connected to 00 peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : 00 + +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[rank2]:[W826 18:32:49.540260328 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. +[rank3]:[W826 18:32:49.547926965 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. +[rank5]:[W826 18:32:49.548015638 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. +[rank4]:[W826 18:32:49.548017416 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. +[rank1]:[W826 18:32:49.548044556 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. +[rank7]:[W826 18:32:49.548075492 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. +[rank6]:[W826 18:32:49.549891096 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: 0.034 seconds +WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations. +> compiling and loading fused kernels ... +[rank0]:[W826 18:32:49.698693148 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: 3.803 seconds +time to initialize megatron (seconds): 7.259 +[after megatron is initialized] datetime: 2026-08-26 18:32:55 +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): 8372566016 +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: 21 +Params for bucket 1 (103022592 elements, 103022592 padded size): + module.output_layer.weight +Params for bucket 2 (49033984 elements, 49033984 padded size): + module.decoder.layers.46.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.45.self_attention.linear_qkv.weight + module.decoder.layers.47.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.46.mlp.router.weight + module.decoder.layers.46.self_attention.k_layernorm.weight + module.decoder.layers.46.pre_mlp_layernorm.weight + module.decoder.layers.46.self_attention.q_layernorm.weight + module.decoder.final_layernorm.weight + module.decoder.layers.47.mlp.router.weight + module.decoder.layers.47.self_attention.k_layernorm.weight + module.decoder.layers.46.self_attention.linear_proj.weight + module.decoder.layers.47.self_attention.q_layernorm.weight + module.decoder.layers.47.pre_mlp_layernorm.weight + module.decoder.layers.46.self_attention.linear_qkv.weight + module.decoder.layers.45.mlp.router.weight + module.decoder.layers.45.self_attention.k_layernorm.weight + module.decoder.layers.47.self_attention.linear_proj.weight + module.decoder.layers.45.pre_mlp_layernorm.weight + module.decoder.layers.45.self_attention.q_layernorm.weight + module.decoder.layers.47.self_attention.linear_qkv.weight +Params for bucket 3 (46672384 elements, 46672384 padded size): + module.decoder.layers.44.pre_mlp_layernorm.weight + module.decoder.layers.44.self_attention.q_layernorm.weight + module.decoder.layers.44.self_attention.linear_qkv.weight + module.decoder.layers.43.pre_mlp_layernorm.weight + module.decoder.layers.43.self_attention.q_layernorm.weight + module.decoder.layers.43.self_attention.linear_qkv.weight + module.decoder.layers.45.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.44.mlp.router.weight + module.decoder.layers.44.self_attention.k_layernorm.weight + module.decoder.layers.45.self_attention.linear_proj.weight + module.decoder.layers.44.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.44.self_attention.linear_proj.weight + module.decoder.layers.43.mlp.router.weight + module.decoder.layers.43.self_attention.k_layernorm.weight + module.decoder.layers.43.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.43.self_attention.linear_proj.weight +Params for bucket 4 (49031936 elements, 49031936 padded size): + module.decoder.layers.41.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.40.mlp.router.weight + module.decoder.layers.40.self_attention.k_layernorm.weight + module.decoder.layers.42.self_attention.q_layernorm.weight + module.decoder.layers.40.pre_mlp_layernorm.weight + module.decoder.layers.40.self_attention.q_layernorm.weight + module.decoder.layers.42.self_attention.linear_proj.weight + module.decoder.layers.41.mlp.router.weight + module.decoder.layers.41.self_attention.k_layernorm.weight + module.decoder.layers.42.self_attention.linear_qkv.weight + module.decoder.layers.41.pre_mlp_layernorm.weight + module.decoder.layers.41.self_attention.q_layernorm.weight + module.decoder.layers.40.self_attention.linear_qkv.weight + module.decoder.layers.42.mlp.router.weight + module.decoder.layers.42.self_attention.k_layernorm.weight + module.decoder.layers.41.self_attention.linear_proj.weight + module.decoder.layers.42.pre_mlp_layernorm.weight + module.decoder.layers.41.self_attention.linear_qkv.weight + module.decoder.layers.42.self_attention.linear_qkv.layer_norm_weight +Params for bucket 5 (46672384 elements, 46672384 padded size): + module.decoder.layers.39.pre_mlp_layernorm.weight + module.decoder.layers.39.self_attention.q_layernorm.weight + module.decoder.layers.39.self_attention.linear_qkv.weight + module.decoder.layers.38.pre_mlp_layernorm.weight + module.decoder.layers.38.self_attention.q_layernorm.weight + module.decoder.layers.38.self_attention.linear_qkv.weight + module.decoder.layers.40.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.40.self_attention.linear_proj.weight + module.decoder.layers.39.mlp.router.weight + module.decoder.layers.39.self_attention.k_layernorm.weight + module.decoder.layers.39.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.39.self_attention.linear_proj.weight + module.decoder.layers.38.mlp.router.weight + module.decoder.layers.38.self_attention.k_layernorm.weight + module.decoder.layers.38.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.38.self_attention.linear_proj.weight +Params for bucket 6 (49031936 elements, 49031936 padded size): + module.decoder.layers.35.pre_mlp_layernorm.weight + module.decoder.layers.35.self_attention.q_layernorm.weight + module.decoder.layers.37.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.36.mlp.router.weight + module.decoder.layers.36.self_attention.k_layernorm.weight + module.decoder.layers.36.pre_mlp_layernorm.weight + module.decoder.layers.36.self_attention.q_layernorm.weight + module.decoder.layers.35.self_attention.linear_qkv.weight + module.decoder.layers.37.mlp.router.weight + module.decoder.layers.37.self_attention.k_layernorm.weight + module.decoder.layers.36.self_attention.linear_proj.weight + module.decoder.layers.37.pre_mlp_layernorm.weight + module.decoder.layers.37.self_attention.q_layernorm.weight + module.decoder.layers.36.self_attention.linear_qkv.weight + module.decoder.layers.37.self_attention.linear_proj.weight + module.decoder.layers.37.self_attention.linear_qkv.weight + module.decoder.layers.36.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.35.mlp.router.weight + module.decoder.layers.35.self_attention.k_layernorm.weight +Params for bucket 7 (46672384 elements, 46672384 padded size): + module.decoder.layers.34.pre_mlp_layernorm.weight + module.decoder.layers.34.self_attention.q_layernorm.weight + module.decoder.layers.34.self_attention.linear_qkv.weight + module.decoder.layers.33.pre_mlp_layernorm.weight + module.decoder.layers.33.self_attention.q_layernorm.weight + module.decoder.layers.33.self_attention.linear_qkv.weight + module.decoder.layers.35.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.34.self_attention.k_layernorm.weight + module.decoder.layers.35.self_attention.linear_proj.weight + module.decoder.layers.34.mlp.router.weight + module.decoder.layers.34.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.34.self_attention.linear_proj.weight + module.decoder.layers.33.mlp.router.weight + module.decoder.layers.33.self_attention.k_layernorm.weight + module.decoder.layers.33.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.33.self_attention.linear_proj.weight +Params for bucket 8 (49031936 elements, 49031936 padded size): + module.decoder.layers.31.pre_mlp_layernorm.weight + module.decoder.layers.31.self_attention.q_layernorm.weight + module.decoder.layers.30.self_attention.linear_qkv.weight + module.decoder.layers.32.mlp.router.weight + module.decoder.layers.32.self_attention.k_layernorm.weight + module.decoder.layers.31.self_attention.linear_proj.weight + module.decoder.layers.32.self_attention.q_layernorm.weight + module.decoder.layers.32.pre_mlp_layernorm.weight + module.decoder.layers.31.self_attention.linear_qkv.weight + module.decoder.layers.32.self_attention.linear_proj.weight + module.decoder.layers.32.self_attention.linear_qkv.weight + module.decoder.layers.31.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.30.mlp.router.weight + module.decoder.layers.30.self_attention.k_layernorm.weight + module.decoder.layers.30.pre_mlp_layernorm.weight + module.decoder.layers.30.self_attention.q_layernorm.weight + module.decoder.layers.32.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.31.mlp.router.weight + module.decoder.layers.31.self_attention.k_layernorm.weight +Params for bucket 9 (46672384 elements, 46672384 padded size): + module.decoder.layers.29.pre_mlp_layernorm.weight + module.decoder.layers.29.self_attention.q_layernorm.weight + module.decoder.layers.29.self_attention.linear_qkv.weight + module.decoder.layers.28.pre_mlp_layernorm.weight + module.decoder.layers.28.self_attention.q_layernorm.weight + module.decoder.layers.28.self_attention.linear_qkv.weight + module.decoder.layers.30.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.29.mlp.router.weight + module.decoder.layers.29.self_attention.k_layernorm.weight + module.decoder.layers.30.self_attention.linear_proj.weight + module.decoder.layers.29.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.29.self_attention.linear_proj.weight + module.decoder.layers.28.mlp.router.weight + module.decoder.layers.28.self_attention.k_layernorm.weight + module.decoder.layers.28.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.28.self_attention.linear_proj.weight +Params for bucket 10 (49031936 elements, 49031936 padded size): + module.decoder.layers.27.pre_mlp_layernorm.weight + module.decoder.layers.27.self_attention.q_layernorm.weight + module.decoder.layers.26.self_attention.linear_qkv.weight + module.decoder.layers.27.self_attention.linear_proj.weight + module.decoder.layers.27.self_attention.linear_qkv.weight + module.decoder.layers.26.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.25.mlp.router.weight + module.decoder.layers.25.self_attention.k_layernorm.weight + module.decoder.layers.25.pre_mlp_layernorm.weight + module.decoder.layers.25.self_attention.q_layernorm.weight + module.decoder.layers.27.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.26.mlp.router.weight + module.decoder.layers.26.self_attention.k_layernorm.weight + module.decoder.layers.26.pre_mlp_layernorm.weight + module.decoder.layers.26.self_attention.q_layernorm.weight + module.decoder.layers.25.self_attention.linear_qkv.weight + module.decoder.layers.27.mlp.router.weight + module.decoder.layers.27.self_attention.k_layernorm.weight + module.decoder.layers.26.self_attention.linear_proj.weight +Params for bucket 11 (46672384 elements, 46672384 padded size): + module.decoder.layers.24.pre_mlp_layernorm.weight + module.decoder.layers.24.self_attention.q_layernorm.weight + module.decoder.layers.24.self_attention.linear_qkv.weight + module.decoder.layers.23.pre_mlp_layernorm.weight + module.decoder.layers.23.self_attention.q_layernorm.weight + module.decoder.layers.23.self_attention.linear_qkv.weight + module.decoder.layers.25.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.24.self_attention.k_layernorm.weight + module.decoder.layers.25.self_attention.linear_proj.weight + module.decoder.layers.24.mlp.router.weight + module.decoder.layers.24.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.24.self_attention.linear_proj.weight + module.decoder.layers.23.mlp.router.weight + module.decoder.layers.23.self_attention.k_layernorm.weight + module.decoder.layers.23.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.23.self_attention.linear_proj.weight +Params for bucket 12 (49031936 elements, 49031936 padded size): + module.decoder.layers.22.self_attention.k_layernorm.weight + module.decoder.layers.20.pre_mlp_layernorm.weight + module.decoder.layers.20.self_attention.q_layernorm.weight + module.decoder.layers.22.pre_mlp_layernorm.weight + module.decoder.layers.22.self_attention.q_layernorm.weight + module.decoder.layers.21.self_attention.k_layernorm.weight + module.decoder.layers.22.self_attention.linear_proj.weight + module.decoder.layers.21.pre_mlp_layernorm.weight + module.decoder.layers.21.self_attention.q_layernorm.weight + module.decoder.layers.20.self_attention.linear_qkv.weight + module.decoder.layers.22.self_attention.linear_qkv.weight + module.decoder.layers.21.self_attention.linear_proj.weight + module.decoder.layers.21.self_attention.linear_qkv.weight + module.decoder.layers.22.mlp.router.weight + module.decoder.layers.22.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.21.mlp.router.weight + module.decoder.layers.21.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.20.mlp.router.weight + module.decoder.layers.20.self_attention.k_layernorm.weight +Params for bucket 13 (46672384 elements, 46672384 padded size): + module.decoder.layers.19.pre_mlp_layernorm.weight + module.decoder.layers.19.self_attention.q_layernorm.weight + module.decoder.layers.19.self_attention.linear_qkv.weight + module.decoder.layers.18.pre_mlp_layernorm.weight + module.decoder.layers.18.self_attention.q_layernorm.weight + module.decoder.layers.18.self_attention.linear_qkv.weight + module.decoder.layers.20.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.20.self_attention.linear_proj.weight + module.decoder.layers.19.mlp.router.weight + module.decoder.layers.19.self_attention.k_layernorm.weight + module.decoder.layers.19.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.19.self_attention.linear_proj.weight + module.decoder.layers.18.mlp.router.weight + module.decoder.layers.18.self_attention.k_layernorm.weight + module.decoder.layers.18.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.18.self_attention.linear_proj.weight +Params for bucket 14 (49031936 elements, 49031936 padded size): + module.decoder.layers.16.pre_mlp_layernorm.weight + module.decoder.layers.16.self_attention.q_layernorm.weight + module.decoder.layers.15.pre_mlp_layernorm.weight + module.decoder.layers.15.self_attention.q_layernorm.weight + module.decoder.layers.17.mlp.router.weight + module.decoder.layers.17.self_attention.k_layernorm.weight + module.decoder.layers.16.self_attention.linear_proj.weight + module.decoder.layers.17.pre_mlp_layernorm.weight + module.decoder.layers.17.self_attention.q_layernorm.weight + module.decoder.layers.16.self_attention.linear_qkv.weight + module.decoder.layers.15.self_attention.linear_qkv.weight + module.decoder.layers.17.self_attention.linear_proj.weight + module.decoder.layers.17.self_attention.linear_qkv.weight + module.decoder.layers.16.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.17.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.16.mlp.router.weight + module.decoder.layers.16.self_attention.k_layernorm.weight + module.decoder.layers.15.mlp.router.weight + module.decoder.layers.15.self_attention.k_layernorm.weight +Params for bucket 15 (46672384 elements, 46672384 padded size): + module.decoder.layers.14.pre_mlp_layernorm.weight + module.decoder.layers.14.self_attention.q_layernorm.weight + module.decoder.layers.14.self_attention.linear_qkv.weight + module.decoder.layers.13.pre_mlp_layernorm.weight + module.decoder.layers.13.self_attention.q_layernorm.weight + module.decoder.layers.13.self_attention.linear_qkv.weight + module.decoder.layers.15.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.15.self_attention.linear_proj.weight + module.decoder.layers.14.mlp.router.weight + module.decoder.layers.14.self_attention.k_layernorm.weight + module.decoder.layers.14.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.14.self_attention.linear_proj.weight + module.decoder.layers.13.mlp.router.weight + module.decoder.layers.13.self_attention.k_layernorm.weight + module.decoder.layers.13.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.13.self_attention.linear_proj.weight +Params for bucket 16 (49031936 elements, 49031936 padded size): + module.decoder.layers.10.pre_mlp_layernorm.weight + module.decoder.layers.10.self_attention.q_layernorm.weight + module.decoder.layers.12.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.11.mlp.router.weight + module.decoder.layers.11.self_attention.k_layernorm.weight + module.decoder.layers.11.pre_mlp_layernorm.weight + module.decoder.layers.11.self_attention.q_layernorm.weight + module.decoder.layers.10.self_attention.linear_qkv.weight + module.decoder.layers.12.mlp.router.weight + module.decoder.layers.12.self_attention.k_layernorm.weight + module.decoder.layers.11.self_attention.linear_proj.weight + module.decoder.layers.12.self_attention.q_layernorm.weight + module.decoder.layers.12.pre_mlp_layernorm.weight + module.decoder.layers.11.self_attention.linear_qkv.weight + module.decoder.layers.12.self_attention.linear_proj.weight + module.decoder.layers.12.self_attention.linear_qkv.weight + module.decoder.layers.11.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.10.mlp.router.weight + module.decoder.layers.10.self_attention.k_layernorm.weight +Params for bucket 17 (46672384 elements, 46672384 padded size): + module.decoder.layers.9.pre_mlp_layernorm.weight + module.decoder.layers.9.self_attention.q_layernorm.weight + module.decoder.layers.9.self_attention.linear_qkv.weight + module.decoder.layers.8.pre_mlp_layernorm.weight + module.decoder.layers.8.self_attention.q_layernorm.weight + module.decoder.layers.8.self_attention.linear_qkv.weight + module.decoder.layers.10.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.10.self_attention.linear_proj.weight + module.decoder.layers.9.mlp.router.weight + module.decoder.layers.9.self_attention.k_layernorm.weight + module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.9.self_attention.linear_proj.weight + module.decoder.layers.8.mlp.router.weight + module.decoder.layers.8.self_attention.k_layernorm.weight + module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.8.self_attention.linear_proj.weight +Params for bucket 18 (49031936 elements, 49031936 padded size): + module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.6.mlp.router.weight + module.decoder.layers.6.self_attention.k_layernorm.weight + module.decoder.layers.6.pre_mlp_layernorm.weight + module.decoder.layers.6.self_attention.q_layernorm.weight + module.decoder.layers.5.mlp.router.weight + module.decoder.layers.5.self_attention.k_layernorm.weight + module.decoder.layers.7.mlp.router.weight + module.decoder.layers.7.self_attention.k_layernorm.weight + module.decoder.layers.6.self_attention.linear_proj.weight + module.decoder.layers.5.pre_mlp_layernorm.weight + module.decoder.layers.5.self_attention.q_layernorm.weight + module.decoder.layers.7.self_attention.q_layernorm.weight + module.decoder.layers.7.pre_mlp_layernorm.weight + module.decoder.layers.6.self_attention.linear_qkv.weight + module.decoder.layers.7.self_attention.linear_proj.weight + module.decoder.layers.5.self_attention.linear_qkv.weight + module.decoder.layers.7.self_attention.linear_qkv.weight + module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight +Params for bucket 19 (46672384 elements, 46672384 padded size): + module.decoder.layers.4.self_attention.k_layernorm.weight + module.decoder.layers.5.self_attention.linear_proj.weight + 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.3.mlp.router.weight + module.decoder.layers.3.self_attention.k_layernorm.weight + module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.3.self_attention.linear_proj.weight + module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.4.pre_mlp_layernorm.weight + module.decoder.layers.4.self_attention.q_layernorm.weight + module.decoder.layers.4.self_attention.linear_qkv.weight + module.decoder.layers.3.pre_mlp_layernorm.weight + module.decoder.layers.3.self_attention.q_layernorm.weight + module.decoder.layers.3.self_attention.linear_qkv.weight +Params for bucket 20 (49031936 elements, 49031936 padded size): + module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.1.mlp.router.weight + module.decoder.layers.1.self_attention.k_layernorm.weight + module.decoder.layers.1.pre_mlp_layernorm.weight + module.decoder.layers.1.self_attention.q_layernorm.weight + module.decoder.layers.0.self_attention.k_layernorm.weight + module.decoder.layers.2.mlp.router.weight + module.decoder.layers.2.self_attention.k_layernorm.weight + module.decoder.layers.1.self_attention.linear_proj.weight + module.decoder.layers.0.mlp.router.weight + module.decoder.layers.2.self_attention.q_layernorm.weight + module.decoder.layers.2.pre_mlp_layernorm.weight + module.decoder.layers.1.self_attention.linear_qkv.weight + module.decoder.layers.0.pre_mlp_layernorm.weight + module.decoder.layers.2.self_attention.linear_proj.weight + module.decoder.layers.0.self_attention.q_layernorm.weight + module.decoder.layers.2.self_attention.linear_qkv.weight + module.decoder.layers.0.self_attention.linear_qkv.weight + module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight +Params for bucket 21 (111413248 elements, 111413248 padded size): + module.decoder.layers.0.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.0.self_attention.linear_proj.weight + module.embedding.word_embeddings.weight +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 89 +Params for bucket 1 (40894464 elements, 40894464 padded size): + module.decoder.layers.47.mlp.experts.linear_fc2.weight9 + module.decoder.layers.47.mlp.experts.linear_fc2.weight1 + module.decoder.layers.47.mlp.experts.linear_fc1.weight12 + module.decoder.layers.47.mlp.experts.linear_fc2.weight14 + module.decoder.layers.47.mlp.experts.linear_fc2.weight6 + module.decoder.layers.47.mlp.experts.linear_fc2.weight11 + module.decoder.layers.47.mlp.experts.linear_fc2.weight3 + module.decoder.layers.47.mlp.experts.linear_fc1.weight14 + module.decoder.layers.47.mlp.experts.linear_fc2.weight8 + module.decoder.layers.47.mlp.experts.linear_fc2.weight0 + module.decoder.layers.47.mlp.experts.linear_fc1.weight11 + module.decoder.layers.47.mlp.experts.linear_fc2.weight13 + module.decoder.layers.47.mlp.experts.linear_fc2.weight5 + module.decoder.layers.47.mlp.experts.linear_fc2.weight10 + module.decoder.layers.47.mlp.experts.linear_fc2.weight2 + module.decoder.layers.47.mlp.experts.linear_fc1.weight13 + module.decoder.layers.47.mlp.experts.linear_fc2.weight15 + module.decoder.layers.47.mlp.experts.linear_fc2.weight7 + module.decoder.layers.47.mlp.experts.linear_fc2.weight12 + module.decoder.layers.47.mlp.experts.linear_fc2.weight4 + module.decoder.layers.47.mlp.experts.linear_fc1.weight15 +Params for bucket 2 (40894464 elements, 40894464 padded size): + module.decoder.layers.47.mlp.experts.linear_fc1.weight10 + module.decoder.layers.47.mlp.experts.linear_fc1.weight8 + module.decoder.layers.47.mlp.experts.linear_fc1.weight6 + module.decoder.layers.47.mlp.experts.linear_fc1.weight4 + module.decoder.layers.47.mlp.experts.linear_fc1.weight2 + module.decoder.layers.47.mlp.experts.linear_fc1.weight0 + module.decoder.layers.46.mlp.experts.linear_fc2.weight15 + module.decoder.layers.46.mlp.experts.linear_fc2.weight13 + module.decoder.layers.47.mlp.experts.linear_fc1.weight7 + module.decoder.layers.47.mlp.experts.linear_fc1.weight9 + module.decoder.layers.47.mlp.experts.linear_fc1.weight5 + module.decoder.layers.47.mlp.experts.linear_fc1.weight3 + module.decoder.layers.47.mlp.experts.linear_fc1.weight1 + module.decoder.layers.46.mlp.experts.linear_fc2.weight14 + module.decoder.layers.46.mlp.experts.linear_fc2.weight12 +Params for bucket 3 (40894464 elements, 40894464 padded size): + module.decoder.layers.46.mlp.experts.linear_fc2.weight11 + module.decoder.layers.46.mlp.experts.linear_fc2.weight3 + module.decoder.layers.46.mlp.experts.linear_fc1.weight14 + module.decoder.layers.46.mlp.experts.linear_fc2.weight8 + module.decoder.layers.46.mlp.experts.linear_fc2.weight0 + module.decoder.layers.46.mlp.experts.linear_fc1.weight11 + module.decoder.layers.46.mlp.experts.linear_fc2.weight5 + module.decoder.layers.46.mlp.experts.linear_fc2.weight10 + module.decoder.layers.46.mlp.experts.linear_fc2.weight2 + module.decoder.layers.46.mlp.experts.linear_fc1.weight13 + module.decoder.layers.46.mlp.experts.linear_fc2.weight7 + module.decoder.layers.46.mlp.experts.linear_fc1.weight10 + module.decoder.layers.46.mlp.experts.linear_fc2.weight4 + module.decoder.layers.46.mlp.experts.linear_fc1.weight15 + module.decoder.layers.46.mlp.experts.linear_fc2.weight9 + module.decoder.layers.46.mlp.experts.linear_fc2.weight1 + module.decoder.layers.46.mlp.experts.linear_fc1.weight12 + module.decoder.layers.46.mlp.experts.linear_fc2.weight6 + module.decoder.layers.46.mlp.experts.linear_fc1.weight9 +Params for bucket 4 (40894464 elements, 40894464 padded size): + module.decoder.layers.46.mlp.experts.linear_fc1.weight8 + module.decoder.layers.46.mlp.experts.linear_fc1.weight4 + module.decoder.layers.46.mlp.experts.linear_fc1.weight6 + module.decoder.layers.46.mlp.experts.linear_fc1.weight2 + module.decoder.layers.46.mlp.experts.linear_fc1.weight0 + module.decoder.layers.45.mlp.experts.linear_fc2.weight14 + module.decoder.layers.45.mlp.experts.linear_fc2.weight12 + module.decoder.layers.45.mlp.experts.linear_fc2.weight10 + module.decoder.layers.45.mlp.experts.linear_fc2.weight8 + module.decoder.layers.46.mlp.experts.linear_fc1.weight5 + module.decoder.layers.46.mlp.experts.linear_fc1.weight7 + module.decoder.layers.46.mlp.experts.linear_fc1.weight3 + module.decoder.layers.46.mlp.experts.linear_fc1.weight1 + module.decoder.layers.45.mlp.experts.linear_fc2.weight15 + module.decoder.layers.45.mlp.experts.linear_fc2.weight13 + module.decoder.layers.45.mlp.experts.linear_fc2.weight11 + module.decoder.layers.45.mlp.experts.linear_fc2.weight9 +Params for bucket 5 (40894464 elements, 40894464 padded size): + module.decoder.layers.45.mlp.experts.linear_fc2.weight4 + module.decoder.layers.45.mlp.experts.linear_fc2.weight6 + module.decoder.layers.45.mlp.experts.linear_fc2.weight2 + module.decoder.layers.45.mlp.experts.linear_fc2.weight0 + module.decoder.layers.45.mlp.experts.linear_fc1.weight15 + module.decoder.layers.45.mlp.experts.linear_fc1.weight13 + module.decoder.layers.45.mlp.experts.linear_fc1.weight11 + module.decoder.layers.45.mlp.experts.linear_fc1.weight9 + module.decoder.layers.45.mlp.experts.linear_fc1.weight7 + module.decoder.layers.45.mlp.experts.linear_fc2.weight7 + module.decoder.layers.45.mlp.experts.linear_fc2.weight5 + module.decoder.layers.45.mlp.experts.linear_fc2.weight3 + module.decoder.layers.45.mlp.experts.linear_fc2.weight1 + module.decoder.layers.45.mlp.experts.linear_fc1.weight14 + module.decoder.layers.45.mlp.experts.linear_fc1.weight12 + module.decoder.layers.45.mlp.experts.linear_fc1.weight10 + module.decoder.layers.45.mlp.experts.linear_fc1.weight8 +Params for bucket 6 (40894464 elements, 40894464 padded size): + module.decoder.layers.44.mlp.experts.linear_fc2.weight14 + module.decoder.layers.44.mlp.experts.linear_fc2.weight6 + module.decoder.layers.45.mlp.experts.linear_fc1.weight4 + module.decoder.layers.44.mlp.experts.linear_fc2.weight11 + module.decoder.layers.45.mlp.experts.linear_fc1.weight1 + module.decoder.layers.44.mlp.experts.linear_fc2.weight8 + module.decoder.layers.45.mlp.experts.linear_fc1.weight6 + module.decoder.layers.44.mlp.experts.linear_fc2.weight13 + module.decoder.layers.44.mlp.experts.linear_fc2.weight5 + module.decoder.layers.45.mlp.experts.linear_fc1.weight3 + module.decoder.layers.44.mlp.experts.linear_fc2.weight10 + module.decoder.layers.45.mlp.experts.linear_fc1.weight0 + module.decoder.layers.44.mlp.experts.linear_fc2.weight15 + module.decoder.layers.44.mlp.experts.linear_fc2.weight7 + module.decoder.layers.45.mlp.experts.linear_fc1.weight5 + module.decoder.layers.44.mlp.experts.linear_fc2.weight12 + module.decoder.layers.44.mlp.experts.linear_fc2.weight4 + module.decoder.layers.45.mlp.experts.linear_fc1.weight2 + module.decoder.layers.44.mlp.experts.linear_fc2.weight9 +Params for bucket 7 (40894464 elements, 40894464 padded size): + module.decoder.layers.44.mlp.experts.linear_fc2.weight2 + module.decoder.layers.44.mlp.experts.linear_fc1.weight15 + module.decoder.layers.44.mlp.experts.linear_fc2.weight0 + module.decoder.layers.44.mlp.experts.linear_fc1.weight13 + module.decoder.layers.44.mlp.experts.linear_fc1.weight11 + module.decoder.layers.44.mlp.experts.linear_fc1.weight9 + module.decoder.layers.44.mlp.experts.linear_fc1.weight7 + module.decoder.layers.44.mlp.experts.linear_fc1.weight5 + module.decoder.layers.44.mlp.experts.linear_fc2.weight3 + module.decoder.layers.44.mlp.experts.linear_fc2.weight1 + module.decoder.layers.44.mlp.experts.linear_fc1.weight14 + module.decoder.layers.44.mlp.experts.linear_fc1.weight12 + module.decoder.layers.44.mlp.experts.linear_fc1.weight10 + module.decoder.layers.44.mlp.experts.linear_fc1.weight8 + module.decoder.layers.44.mlp.experts.linear_fc1.weight6 +Params for bucket 8 (40894464 elements, 40894464 padded size): + module.decoder.layers.44.mlp.experts.linear_fc1.weight1 + module.decoder.layers.43.mlp.experts.linear_fc2.weight15 + module.decoder.layers.43.mlp.experts.linear_fc2.weight7 + module.decoder.layers.43.mlp.experts.linear_fc2.weight12 + module.decoder.layers.43.mlp.experts.linear_fc2.weight4 + module.decoder.layers.44.mlp.experts.linear_fc1.weight3 + module.decoder.layers.43.mlp.experts.linear_fc2.weight9 + module.decoder.layers.43.mlp.experts.linear_fc2.weight1 + module.decoder.layers.44.mlp.experts.linear_fc1.weight0 + module.decoder.layers.43.mlp.experts.linear_fc2.weight14 + module.decoder.layers.43.mlp.experts.linear_fc2.weight6 + module.decoder.layers.43.mlp.experts.linear_fc2.weight11 + module.decoder.layers.43.mlp.experts.linear_fc2.weight3 + module.decoder.layers.44.mlp.experts.linear_fc1.weight2 + module.decoder.layers.43.mlp.experts.linear_fc2.weight8 + module.decoder.layers.43.mlp.experts.linear_fc2.weight0 + module.decoder.layers.43.mlp.experts.linear_fc2.weight13 + module.decoder.layers.43.mlp.experts.linear_fc2.weight5 + module.decoder.layers.44.mlp.experts.linear_fc1.weight4 + module.decoder.layers.43.mlp.experts.linear_fc2.weight10 + module.decoder.layers.43.mlp.experts.linear_fc2.weight2 +Params for bucket 9 (40894464 elements, 40894464 padded size): + module.decoder.layers.43.mlp.experts.linear_fc1.weight15 + module.decoder.layers.43.mlp.experts.linear_fc1.weight13 + module.decoder.layers.43.mlp.experts.linear_fc1.weight11 + module.decoder.layers.43.mlp.experts.linear_fc1.weight9 + module.decoder.layers.43.mlp.experts.linear_fc1.weight7 + module.decoder.layers.43.mlp.experts.linear_fc1.weight5 + module.decoder.layers.43.mlp.experts.linear_fc1.weight3 + module.decoder.layers.43.mlp.experts.linear_fc1.weight14 + module.decoder.layers.43.mlp.experts.linear_fc1.weight12 + module.decoder.layers.43.mlp.experts.linear_fc1.weight10 + module.decoder.layers.43.mlp.experts.linear_fc1.weight8 + module.decoder.layers.43.mlp.experts.linear_fc1.weight6 + module.decoder.layers.43.mlp.experts.linear_fc1.weight4 +Params for bucket 10 (40894464 elements, 40894464 padded size): + module.decoder.layers.43.mlp.experts.linear_fc1.weight1 + module.decoder.layers.42.mlp.experts.linear_fc2.weight8 + module.decoder.layers.42.mlp.experts.linear_fc2.weight0 + module.decoder.layers.42.mlp.experts.linear_fc2.weight13 + module.decoder.layers.42.mlp.experts.linear_fc2.weight5 + module.decoder.layers.42.mlp.experts.linear_fc2.weight10 + module.decoder.layers.42.mlp.experts.linear_fc2.weight2 + module.decoder.layers.43.mlp.experts.linear_fc1.weight0 + module.decoder.layers.42.mlp.experts.linear_fc2.weight15 + module.decoder.layers.42.mlp.experts.linear_fc2.weight7 + module.decoder.layers.42.mlp.experts.linear_fc2.weight12 + module.decoder.layers.42.mlp.experts.linear_fc2.weight4 + module.decoder.layers.42.mlp.experts.linear_fc1.weight15 + module.decoder.layers.43.mlp.experts.linear_fc1.weight2 + module.decoder.layers.42.mlp.experts.linear_fc2.weight9 + module.decoder.layers.42.mlp.experts.linear_fc2.weight1 + module.decoder.layers.42.mlp.experts.linear_fc2.weight14 + module.decoder.layers.42.mlp.experts.linear_fc2.weight6 + module.decoder.layers.42.mlp.experts.linear_fc2.weight11 + module.decoder.layers.42.mlp.experts.linear_fc2.weight3 + module.decoder.layers.42.mlp.experts.linear_fc1.weight14 +Params for bucket 11 (40894464 elements, 40894464 padded size): + module.decoder.layers.42.mlp.experts.linear_fc1.weight13 + module.decoder.layers.42.mlp.experts.linear_fc1.weight11 + module.decoder.layers.42.mlp.experts.linear_fc1.weight9 + module.decoder.layers.42.mlp.experts.linear_fc1.weight7 + module.decoder.layers.42.mlp.experts.linear_fc1.weight5 + module.decoder.layers.42.mlp.experts.linear_fc1.weight3 + module.decoder.layers.42.mlp.experts.linear_fc1.weight1 + module.decoder.layers.42.mlp.experts.linear_fc1.weight10 + module.decoder.layers.42.mlp.experts.linear_fc1.weight12 + module.decoder.layers.42.mlp.experts.linear_fc1.weight8 + module.decoder.layers.42.mlp.experts.linear_fc1.weight6 + module.decoder.layers.42.mlp.experts.linear_fc1.weight4 + module.decoder.layers.42.mlp.experts.linear_fc1.weight2 +Params for bucket 12 (40894464 elements, 40894464 padded size): + module.decoder.layers.41.mlp.experts.linear_fc2.weight10 + module.decoder.layers.41.mlp.experts.linear_fc2.weight2 + module.decoder.layers.41.mlp.experts.linear_fc1.weight13 + module.decoder.layers.42.mlp.experts.linear_fc1.weight0 + module.decoder.layers.41.mlp.experts.linear_fc2.weight15 + module.decoder.layers.41.mlp.experts.linear_fc2.weight7 + module.decoder.layers.41.mlp.experts.linear_fc2.weight12 + module.decoder.layers.41.mlp.experts.linear_fc2.weight4 + module.decoder.layers.41.mlp.experts.linear_fc1.weight15 + module.decoder.layers.41.mlp.experts.linear_fc2.weight9 + module.decoder.layers.41.mlp.experts.linear_fc2.weight1 + module.decoder.layers.41.mlp.experts.linear_fc1.weight12 + module.decoder.layers.41.mlp.experts.linear_fc2.weight14 + module.decoder.layers.41.mlp.experts.linear_fc2.weight6 + module.decoder.layers.41.mlp.experts.linear_fc2.weight11 + module.decoder.layers.41.mlp.experts.linear_fc2.weight3 + module.decoder.layers.41.mlp.experts.linear_fc1.weight14 + module.decoder.layers.41.mlp.experts.linear_fc2.weight8 + module.decoder.layers.41.mlp.experts.linear_fc2.weight0 + module.decoder.layers.41.mlp.experts.linear_fc2.weight13 + module.decoder.layers.41.mlp.experts.linear_fc2.weight5 +Params for bucket 13 (40894464 elements, 40894464 padded size): + module.decoder.layers.41.mlp.experts.linear_fc1.weight11 + module.decoder.layers.41.mlp.experts.linear_fc1.weight7 + module.decoder.layers.41.mlp.experts.linear_fc1.weight9 + module.decoder.layers.41.mlp.experts.linear_fc1.weight5 + module.decoder.layers.41.mlp.experts.linear_fc1.weight3 + module.decoder.layers.41.mlp.experts.linear_fc1.weight1 + module.decoder.layers.40.mlp.experts.linear_fc2.weight14 + module.decoder.layers.41.mlp.experts.linear_fc1.weight10 + module.decoder.layers.41.mlp.experts.linear_fc1.weight8 + module.decoder.layers.41.mlp.experts.linear_fc1.weight6 + module.decoder.layers.41.mlp.experts.linear_fc1.weight4 + module.decoder.layers.41.mlp.experts.linear_fc1.weight2 + module.decoder.layers.41.mlp.experts.linear_fc1.weight0 + module.decoder.layers.40.mlp.experts.linear_fc2.weight15 +Params for bucket 14 (40894464 elements, 40894464 padded size): + module.decoder.layers.40.mlp.experts.linear_fc2.weight12 + module.decoder.layers.40.mlp.experts.linear_fc2.weight4 + module.decoder.layers.40.mlp.experts.linear_fc1.weight15 + module.decoder.layers.40.mlp.experts.linear_fc2.weight9 + module.decoder.layers.40.mlp.experts.linear_fc2.weight1 + module.decoder.layers.40.mlp.experts.linear_fc1.weight12 + module.decoder.layers.40.mlp.experts.linear_fc2.weight6 + module.decoder.layers.40.mlp.experts.linear_fc2.weight11 + module.decoder.layers.40.mlp.experts.linear_fc2.weight3 + module.decoder.layers.40.mlp.experts.linear_fc1.weight14 + module.decoder.layers.40.mlp.experts.linear_fc2.weight8 + module.decoder.layers.40.mlp.experts.linear_fc2.weight0 + module.decoder.layers.40.mlp.experts.linear_fc1.weight11 + module.decoder.layers.40.mlp.experts.linear_fc2.weight13 + module.decoder.layers.40.mlp.experts.linear_fc2.weight5 + module.decoder.layers.40.mlp.experts.linear_fc2.weight10 + module.decoder.layers.40.mlp.experts.linear_fc2.weight2 + module.decoder.layers.40.mlp.experts.linear_fc1.weight13 + module.decoder.layers.40.mlp.experts.linear_fc2.weight7 + module.decoder.layers.40.mlp.experts.linear_fc1.weight10 +Params for bucket 15 (40894464 elements, 40894464 padded size): + module.decoder.layers.40.mlp.experts.linear_fc1.weight9 + module.decoder.layers.40.mlp.experts.linear_fc1.weight7 + module.decoder.layers.40.mlp.experts.linear_fc1.weight5 + module.decoder.layers.40.mlp.experts.linear_fc1.weight3 + module.decoder.layers.40.mlp.experts.linear_fc1.weight1 + module.decoder.layers.39.mlp.experts.linear_fc2.weight14 + module.decoder.layers.39.mlp.experts.linear_fc2.weight12 + module.decoder.layers.39.mlp.experts.linear_fc2.weight10 + module.decoder.layers.40.mlp.experts.linear_fc1.weight8 + module.decoder.layers.40.mlp.experts.linear_fc1.weight6 + module.decoder.layers.40.mlp.experts.linear_fc1.weight4 + module.decoder.layers.40.mlp.experts.linear_fc1.weight2 + module.decoder.layers.40.mlp.experts.linear_fc1.weight0 + module.decoder.layers.39.mlp.experts.linear_fc2.weight15 + module.decoder.layers.39.mlp.experts.linear_fc2.weight13 + module.decoder.layers.39.mlp.experts.linear_fc2.weight11 +Params for bucket 16 (40894464 elements, 40894464 padded size): + module.decoder.layers.39.mlp.experts.linear_fc2.weight8 + module.decoder.layers.39.mlp.experts.linear_fc2.weight6 + module.decoder.layers.39.mlp.experts.linear_fc2.weight4 + module.decoder.layers.39.mlp.experts.linear_fc2.weight2 + module.decoder.layers.39.mlp.experts.linear_fc2.weight0 + module.decoder.layers.39.mlp.experts.linear_fc1.weight15 + module.decoder.layers.39.mlp.experts.linear_fc1.weight13 + module.decoder.layers.39.mlp.experts.linear_fc1.weight11 + module.decoder.layers.39.mlp.experts.linear_fc1.weight9 + module.decoder.layers.39.mlp.experts.linear_fc2.weight9 + module.decoder.layers.39.mlp.experts.linear_fc2.weight5 + module.decoder.layers.39.mlp.experts.linear_fc2.weight7 + module.decoder.layers.39.mlp.experts.linear_fc2.weight3 + module.decoder.layers.39.mlp.experts.linear_fc2.weight1 + module.decoder.layers.39.mlp.experts.linear_fc1.weight14 + module.decoder.layers.39.mlp.experts.linear_fc1.weight12 + module.decoder.layers.39.mlp.experts.linear_fc1.weight10 + module.decoder.layers.39.mlp.experts.linear_fc1.weight8 +Params for bucket 17 (40894464 elements, 40894464 padded size): + module.decoder.layers.39.mlp.experts.linear_fc1.weight7 + module.decoder.layers.39.mlp.experts.linear_fc1.weight5 + module.decoder.layers.39.mlp.experts.linear_fc1.weight3 + module.decoder.layers.39.mlp.experts.linear_fc1.weight1 + module.decoder.layers.38.mlp.experts.linear_fc2.weight14 + module.decoder.layers.38.mlp.experts.linear_fc2.weight12 + module.decoder.layers.38.mlp.experts.linear_fc2.weight10 + module.decoder.layers.38.mlp.experts.linear_fc2.weight8 + module.decoder.layers.38.mlp.experts.linear_fc2.weight6 + module.decoder.layers.39.mlp.experts.linear_fc1.weight4 + module.decoder.layers.39.mlp.experts.linear_fc1.weight6 + module.decoder.layers.39.mlp.experts.linear_fc1.weight2 + module.decoder.layers.39.mlp.experts.linear_fc1.weight0 + module.decoder.layers.38.mlp.experts.linear_fc2.weight15 + module.decoder.layers.38.mlp.experts.linear_fc2.weight13 + module.decoder.layers.38.mlp.experts.linear_fc2.weight11 + module.decoder.layers.38.mlp.experts.linear_fc2.weight9 + module.decoder.layers.38.mlp.experts.linear_fc2.weight7 +Params for bucket 18 (40894464 elements, 40894464 padded size): + module.decoder.layers.38.mlp.experts.linear_fc2.weight2 + module.decoder.layers.38.mlp.experts.linear_fc2.weight4 + module.decoder.layers.38.mlp.experts.linear_fc2.weight0 + module.decoder.layers.38.mlp.experts.linear_fc1.weight15 + module.decoder.layers.38.mlp.experts.linear_fc1.weight13 + module.decoder.layers.38.mlp.experts.linear_fc1.weight11 + module.decoder.layers.38.mlp.experts.linear_fc1.weight9 + module.decoder.layers.38.mlp.experts.linear_fc1.weight7 + module.decoder.layers.38.mlp.experts.linear_fc2.weight5 + module.decoder.layers.38.mlp.experts.linear_fc2.weight1 + module.decoder.layers.38.mlp.experts.linear_fc2.weight3 + module.decoder.layers.38.mlp.experts.linear_fc1.weight14 + module.decoder.layers.38.mlp.experts.linear_fc1.weight12 + module.decoder.layers.38.mlp.experts.linear_fc1.weight10 + module.decoder.layers.38.mlp.experts.linear_fc1.weight8 + module.decoder.layers.38.mlp.experts.linear_fc1.weight6 +Params for bucket 19 (40894464 elements, 40894464 padded size): + module.decoder.layers.38.mlp.experts.linear_fc1.weight3 + module.decoder.layers.37.mlp.experts.linear_fc2.weight10 + module.decoder.layers.37.mlp.experts.linear_fc2.weight2 + module.decoder.layers.38.mlp.experts.linear_fc1.weight0 + module.decoder.layers.37.mlp.experts.linear_fc2.weight15 + module.decoder.layers.37.mlp.experts.linear_fc2.weight7 + module.decoder.layers.38.mlp.experts.linear_fc1.weight5 + module.decoder.layers.37.mlp.experts.linear_fc2.weight12 + module.decoder.layers.37.mlp.experts.linear_fc2.weight4 + module.decoder.layers.38.mlp.experts.linear_fc1.weight2 + module.decoder.layers.37.mlp.experts.linear_fc2.weight9 + module.decoder.layers.37.mlp.experts.linear_fc2.weight14 + module.decoder.layers.37.mlp.experts.linear_fc2.weight6 + module.decoder.layers.38.mlp.experts.linear_fc1.weight4 + module.decoder.layers.37.mlp.experts.linear_fc2.weight11 + module.decoder.layers.37.mlp.experts.linear_fc2.weight3 + module.decoder.layers.38.mlp.experts.linear_fc1.weight1 + module.decoder.layers.37.mlp.experts.linear_fc2.weight8 + module.decoder.layers.37.mlp.experts.linear_fc2.weight13 + module.decoder.layers.37.mlp.experts.linear_fc2.weight5 +Params for bucket 20 (40894464 elements, 40894464 padded size): + module.decoder.layers.37.mlp.experts.linear_fc1.weight15 + module.decoder.layers.37.mlp.experts.linear_fc2.weight0 + module.decoder.layers.37.mlp.experts.linear_fc1.weight13 + module.decoder.layers.37.mlp.experts.linear_fc1.weight11 + module.decoder.layers.37.mlp.experts.linear_fc1.weight9 + module.decoder.layers.37.mlp.experts.linear_fc1.weight7 + module.decoder.layers.37.mlp.experts.linear_fc1.weight5 + module.decoder.layers.37.mlp.experts.linear_fc2.weight1 + module.decoder.layers.37.mlp.experts.linear_fc1.weight14 + module.decoder.layers.37.mlp.experts.linear_fc1.weight12 + module.decoder.layers.37.mlp.experts.linear_fc1.weight10 + module.decoder.layers.37.mlp.experts.linear_fc1.weight8 + module.decoder.layers.37.mlp.experts.linear_fc1.weight6 + module.decoder.layers.37.mlp.experts.linear_fc1.weight4 +Params for bucket 21 (40894464 elements, 40894464 padded size): + module.decoder.layers.36.mlp.experts.linear_fc2.weight12 + module.decoder.layers.36.mlp.experts.linear_fc2.weight4 + module.decoder.layers.36.mlp.experts.linear_fc1.weight15 + module.decoder.layers.37.mlp.experts.linear_fc1.weight2 + module.decoder.layers.36.mlp.experts.linear_fc2.weight9 + module.decoder.layers.36.mlp.experts.linear_fc2.weight1 + module.decoder.layers.36.mlp.experts.linear_fc2.weight14 + module.decoder.layers.36.mlp.experts.linear_fc2.weight6 + module.decoder.layers.36.mlp.experts.linear_fc2.weight11 + module.decoder.layers.36.mlp.experts.linear_fc2.weight3 + module.decoder.layers.37.mlp.experts.linear_fc1.weight1 + module.decoder.layers.36.mlp.experts.linear_fc2.weight8 + module.decoder.layers.36.mlp.experts.linear_fc2.weight0 + module.decoder.layers.36.mlp.experts.linear_fc2.weight13 + module.decoder.layers.36.mlp.experts.linear_fc2.weight5 + module.decoder.layers.37.mlp.experts.linear_fc1.weight3 + module.decoder.layers.36.mlp.experts.linear_fc2.weight10 + module.decoder.layers.36.mlp.experts.linear_fc2.weight2 + module.decoder.layers.36.mlp.experts.linear_fc2.weight15 + module.decoder.layers.37.mlp.experts.linear_fc1.weight0 + module.decoder.layers.36.mlp.experts.linear_fc2.weight7 +Params for bucket 22 (40894464 elements, 40894464 padded size): + module.decoder.layers.36.mlp.experts.linear_fc1.weight13 + module.decoder.layers.36.mlp.experts.linear_fc1.weight11 + module.decoder.layers.36.mlp.experts.linear_fc1.weight9 + module.decoder.layers.36.mlp.experts.linear_fc1.weight7 + module.decoder.layers.36.mlp.experts.linear_fc1.weight5 + module.decoder.layers.36.mlp.experts.linear_fc1.weight3 + module.decoder.layers.36.mlp.experts.linear_fc1.weight14 + module.decoder.layers.36.mlp.experts.linear_fc1.weight12 + module.decoder.layers.36.mlp.experts.linear_fc1.weight10 + module.decoder.layers.36.mlp.experts.linear_fc1.weight8 + module.decoder.layers.36.mlp.experts.linear_fc1.weight6 + module.decoder.layers.36.mlp.experts.linear_fc1.weight4 + module.decoder.layers.36.mlp.experts.linear_fc1.weight2 +Params for bucket 23 (40894464 elements, 40894464 padded size): + module.decoder.layers.35.mlp.experts.linear_fc2.weight14 + module.decoder.layers.35.mlp.experts.linear_fc2.weight6 + module.decoder.layers.35.mlp.experts.linear_fc2.weight11 + module.decoder.layers.35.mlp.experts.linear_fc2.weight3 + module.decoder.layers.35.mlp.experts.linear_fc1.weight14 + module.decoder.layers.36.mlp.experts.linear_fc1.weight1 + module.decoder.layers.35.mlp.experts.linear_fc2.weight8 + module.decoder.layers.35.mlp.experts.linear_fc2.weight0 + module.decoder.layers.35.mlp.experts.linear_fc2.weight13 + module.decoder.layers.35.mlp.experts.linear_fc2.weight5 + module.decoder.layers.35.mlp.experts.linear_fc2.weight10 + module.decoder.layers.35.mlp.experts.linear_fc2.weight2 + module.decoder.layers.35.mlp.experts.linear_fc1.weight13 + module.decoder.layers.36.mlp.experts.linear_fc1.weight0 + module.decoder.layers.35.mlp.experts.linear_fc2.weight15 + module.decoder.layers.35.mlp.experts.linear_fc2.weight7 + module.decoder.layers.35.mlp.experts.linear_fc2.weight12 + module.decoder.layers.35.mlp.experts.linear_fc2.weight4 + module.decoder.layers.35.mlp.experts.linear_fc1.weight15 + module.decoder.layers.35.mlp.experts.linear_fc2.weight9 + module.decoder.layers.35.mlp.experts.linear_fc2.weight1 +Params for bucket 24 (40894464 elements, 40894464 padded size): + module.decoder.layers.35.mlp.experts.linear_fc1.weight9 + module.decoder.layers.35.mlp.experts.linear_fc1.weight11 + module.decoder.layers.35.mlp.experts.linear_fc1.weight7 + module.decoder.layers.35.mlp.experts.linear_fc1.weight5 + module.decoder.layers.35.mlp.experts.linear_fc1.weight3 + module.decoder.layers.35.mlp.experts.linear_fc1.weight1 + module.decoder.layers.35.mlp.experts.linear_fc1.weight12 + module.decoder.layers.35.mlp.experts.linear_fc1.weight8 + module.decoder.layers.35.mlp.experts.linear_fc1.weight10 + module.decoder.layers.35.mlp.experts.linear_fc1.weight6 + module.decoder.layers.35.mlp.experts.linear_fc1.weight4 + module.decoder.layers.35.mlp.experts.linear_fc1.weight2 + module.decoder.layers.35.mlp.experts.linear_fc1.weight0 +Params for bucket 25 (40894464 elements, 40894464 padded size): + module.decoder.layers.34.mlp.experts.linear_fc2.weight8 + module.decoder.layers.34.mlp.experts.linear_fc2.weight0 + module.decoder.layers.34.mlp.experts.linear_fc1.weight11 + module.decoder.layers.34.mlp.experts.linear_fc2.weight5 + module.decoder.layers.34.mlp.experts.linear_fc2.weight14 + module.decoder.layers.34.mlp.experts.linear_fc2.weight10 + module.decoder.layers.34.mlp.experts.linear_fc2.weight2 + module.decoder.layers.34.mlp.experts.linear_fc1.weight13 + module.decoder.layers.34.mlp.experts.linear_fc2.weight7 + module.decoder.layers.34.mlp.experts.linear_fc2.weight12 + module.decoder.layers.34.mlp.experts.linear_fc2.weight4 + module.decoder.layers.34.mlp.experts.linear_fc1.weight15 + module.decoder.layers.34.mlp.experts.linear_fc2.weight13 + module.decoder.layers.34.mlp.experts.linear_fc2.weight9 + module.decoder.layers.34.mlp.experts.linear_fc2.weight1 + module.decoder.layers.34.mlp.experts.linear_fc1.weight12 + module.decoder.layers.34.mlp.experts.linear_fc2.weight6 + module.decoder.layers.34.mlp.experts.linear_fc2.weight15 + module.decoder.layers.34.mlp.experts.linear_fc2.weight11 + module.decoder.layers.34.mlp.experts.linear_fc2.weight3 + module.decoder.layers.34.mlp.experts.linear_fc1.weight14 +Params for bucket 26 (40894464 elements, 40894464 padded size): + module.decoder.layers.34.mlp.experts.linear_fc1.weight7 + module.decoder.layers.34.mlp.experts.linear_fc1.weight9 + module.decoder.layers.34.mlp.experts.linear_fc1.weight5 + module.decoder.layers.34.mlp.experts.linear_fc1.weight3 + module.decoder.layers.34.mlp.experts.linear_fc1.weight1 + module.decoder.layers.33.mlp.experts.linear_fc2.weight14 + module.decoder.layers.33.mlp.experts.linear_fc2.weight12 + module.decoder.layers.34.mlp.experts.linear_fc1.weight10 + module.decoder.layers.34.mlp.experts.linear_fc1.weight6 + module.decoder.layers.34.mlp.experts.linear_fc1.weight8 + module.decoder.layers.34.mlp.experts.linear_fc1.weight4 + module.decoder.layers.34.mlp.experts.linear_fc1.weight2 + module.decoder.layers.34.mlp.experts.linear_fc1.weight0 + module.decoder.layers.33.mlp.experts.linear_fc2.weight15 + module.decoder.layers.33.mlp.experts.linear_fc2.weight13 +Params for bucket 27 (40894464 elements, 40894464 padded size): + module.decoder.layers.33.mlp.experts.linear_fc2.weight2 + module.decoder.layers.33.mlp.experts.linear_fc1.weight13 + module.decoder.layers.33.mlp.experts.linear_fc2.weight11 + module.decoder.layers.33.mlp.experts.linear_fc1.weight10 + module.decoder.layers.33.mlp.experts.linear_fc2.weight8 + module.decoder.layers.33.mlp.experts.linear_fc2.weight4 + module.decoder.layers.33.mlp.experts.linear_fc1.weight15 + module.decoder.layers.33.mlp.experts.linear_fc2.weight1 + module.decoder.layers.33.mlp.experts.linear_fc1.weight12 + module.decoder.layers.33.mlp.experts.linear_fc2.weight10 + module.decoder.layers.33.mlp.experts.linear_fc2.weight6 + module.decoder.layers.33.mlp.experts.linear_fc1.weight9 + module.decoder.layers.33.mlp.experts.linear_fc2.weight7 + module.decoder.layers.33.mlp.experts.linear_fc2.weight3 + module.decoder.layers.33.mlp.experts.linear_fc1.weight14 + module.decoder.layers.33.mlp.experts.linear_fc2.weight0 + module.decoder.layers.33.mlp.experts.linear_fc1.weight11 + module.decoder.layers.33.mlp.experts.linear_fc2.weight9 + module.decoder.layers.33.mlp.experts.linear_fc2.weight5 +Params for bucket 28 (40894464 elements, 40894464 padded size): + module.decoder.layers.33.mlp.experts.linear_fc1.weight7 + module.decoder.layers.33.mlp.experts.linear_fc1.weight5 + module.decoder.layers.33.mlp.experts.linear_fc1.weight3 + module.decoder.layers.33.mlp.experts.linear_fc1.weight1 + module.decoder.layers.32.mlp.experts.linear_fc2.weight14 + module.decoder.layers.32.mlp.experts.linear_fc2.weight12 + module.decoder.layers.32.mlp.experts.linear_fc2.weight10 + module.decoder.layers.32.mlp.experts.linear_fc2.weight8 + module.decoder.layers.33.mlp.experts.linear_fc1.weight8 + module.decoder.layers.33.mlp.experts.linear_fc1.weight4 + module.decoder.layers.33.mlp.experts.linear_fc1.weight6 + module.decoder.layers.33.mlp.experts.linear_fc1.weight2 + module.decoder.layers.33.mlp.experts.linear_fc1.weight0 + module.decoder.layers.32.mlp.experts.linear_fc2.weight15 + module.decoder.layers.32.mlp.experts.linear_fc2.weight13 + module.decoder.layers.32.mlp.experts.linear_fc2.weight11 + module.decoder.layers.32.mlp.experts.linear_fc2.weight9 +Params for bucket 29 (40894464 elements, 40894464 padded size): + module.decoder.layers.32.mlp.experts.linear_fc2.weight6 + module.decoder.layers.32.mlp.experts.linear_fc2.weight4 + module.decoder.layers.32.mlp.experts.linear_fc2.weight2 + module.decoder.layers.32.mlp.experts.linear_fc2.weight0 + module.decoder.layers.32.mlp.experts.linear_fc1.weight15 + module.decoder.layers.32.mlp.experts.linear_fc1.weight13 + module.decoder.layers.32.mlp.experts.linear_fc1.weight11 + module.decoder.layers.32.mlp.experts.linear_fc1.weight9 + module.decoder.layers.32.mlp.experts.linear_fc1.weight7 + module.decoder.layers.32.mlp.experts.linear_fc2.weight7 + module.decoder.layers.32.mlp.experts.linear_fc2.weight5 + module.decoder.layers.32.mlp.experts.linear_fc2.weight3 + module.decoder.layers.32.mlp.experts.linear_fc2.weight1 + module.decoder.layers.32.mlp.experts.linear_fc1.weight14 + module.decoder.layers.32.mlp.experts.linear_fc1.weight12 + module.decoder.layers.32.mlp.experts.linear_fc1.weight10 + module.decoder.layers.32.mlp.experts.linear_fc1.weight8 +Params for bucket 30 (40894464 elements, 40894464 padded size): + module.decoder.layers.31.mlp.experts.linear_fc2.weight10 + module.decoder.layers.32.mlp.experts.linear_fc1.weight4 + module.decoder.layers.31.mlp.experts.linear_fc2.weight15 + module.decoder.layers.31.mlp.experts.linear_fc2.weight7 + module.decoder.layers.32.mlp.experts.linear_fc1.weight1 + module.decoder.layers.31.mlp.experts.linear_fc2.weight12 + module.decoder.layers.31.mlp.experts.linear_fc2.weight4 + module.decoder.layers.32.mlp.experts.linear_fc1.weight6 + module.decoder.layers.31.mlp.experts.linear_fc2.weight9 + module.decoder.layers.32.mlp.experts.linear_fc1.weight3 + module.decoder.layers.31.mlp.experts.linear_fc2.weight14 + module.decoder.layers.31.mlp.experts.linear_fc2.weight6 + module.decoder.layers.32.mlp.experts.linear_fc1.weight0 + module.decoder.layers.31.mlp.experts.linear_fc2.weight11 + module.decoder.layers.32.mlp.experts.linear_fc1.weight5 + module.decoder.layers.31.mlp.experts.linear_fc2.weight8 + module.decoder.layers.32.mlp.experts.linear_fc1.weight2 + module.decoder.layers.31.mlp.experts.linear_fc2.weight13 + module.decoder.layers.31.mlp.experts.linear_fc2.weight5 +Params for bucket 31 (40894464 elements, 40894464 padded size): + module.decoder.layers.31.mlp.experts.linear_fc2.weight0 + module.decoder.layers.31.mlp.experts.linear_fc1.weight15 + module.decoder.layers.31.mlp.experts.linear_fc2.weight2 + module.decoder.layers.31.mlp.experts.linear_fc1.weight13 + module.decoder.layers.31.mlp.experts.linear_fc1.weight11 + module.decoder.layers.31.mlp.experts.linear_fc1.weight9 + module.decoder.layers.31.mlp.experts.linear_fc1.weight7 + module.decoder.layers.31.mlp.experts.linear_fc1.weight5 + module.decoder.layers.31.mlp.experts.linear_fc2.weight3 + module.decoder.layers.31.mlp.experts.linear_fc2.weight1 + module.decoder.layers.31.mlp.experts.linear_fc1.weight14 + module.decoder.layers.31.mlp.experts.linear_fc1.weight12 + module.decoder.layers.31.mlp.experts.linear_fc1.weight10 + module.decoder.layers.31.mlp.experts.linear_fc1.weight8 + module.decoder.layers.31.mlp.experts.linear_fc1.weight6 +Params for bucket 32 (40894464 elements, 40894464 padded size): + module.decoder.layers.31.mlp.experts.linear_fc1.weight1 + module.decoder.layers.30.mlp.experts.linear_fc2.weight12 + module.decoder.layers.30.mlp.experts.linear_fc2.weight4 + module.decoder.layers.30.mlp.experts.linear_fc2.weight9 + module.decoder.layers.30.mlp.experts.linear_fc2.weight1 + module.decoder.layers.31.mlp.experts.linear_fc1.weight3 + module.decoder.layers.30.mlp.experts.linear_fc2.weight14 + module.decoder.layers.30.mlp.experts.linear_fc2.weight6 + module.decoder.layers.31.mlp.experts.linear_fc1.weight0 + module.decoder.layers.30.mlp.experts.linear_fc2.weight11 + module.decoder.layers.30.mlp.experts.linear_fc2.weight3 + module.decoder.layers.30.mlp.experts.linear_fc2.weight8 + module.decoder.layers.30.mlp.experts.linear_fc2.weight0 + module.decoder.layers.31.mlp.experts.linear_fc1.weight2 + module.decoder.layers.30.mlp.experts.linear_fc2.weight13 + module.decoder.layers.30.mlp.experts.linear_fc2.weight5 + module.decoder.layers.30.mlp.experts.linear_fc2.weight10 + module.decoder.layers.30.mlp.experts.linear_fc2.weight2 + module.decoder.layers.31.mlp.experts.linear_fc1.weight4 + module.decoder.layers.30.mlp.experts.linear_fc2.weight15 + module.decoder.layers.30.mlp.experts.linear_fc2.weight7 +Params for bucket 33 (40894464 elements, 40894464 padded size): + module.decoder.layers.30.mlp.experts.linear_fc1.weight15 + module.decoder.layers.30.mlp.experts.linear_fc1.weight11 + module.decoder.layers.30.mlp.experts.linear_fc1.weight13 + module.decoder.layers.30.mlp.experts.linear_fc1.weight9 + module.decoder.layers.30.mlp.experts.linear_fc1.weight7 + module.decoder.layers.30.mlp.experts.linear_fc1.weight5 + module.decoder.layers.30.mlp.experts.linear_fc1.weight3 + module.decoder.layers.30.mlp.experts.linear_fc1.weight12 + module.decoder.layers.30.mlp.experts.linear_fc1.weight14 + module.decoder.layers.30.mlp.experts.linear_fc1.weight10 + module.decoder.layers.30.mlp.experts.linear_fc1.weight8 + module.decoder.layers.30.mlp.experts.linear_fc1.weight6 + module.decoder.layers.30.mlp.experts.linear_fc1.weight4 +Params for bucket 34 (40894464 elements, 40894464 padded size): + module.decoder.layers.29.mlp.experts.linear_fc2.weight14 + module.decoder.layers.29.mlp.experts.linear_fc2.weight6 + module.decoder.layers.30.mlp.experts.linear_fc1.weight0 + module.decoder.layers.29.mlp.experts.linear_fc2.weight11 + module.decoder.layers.29.mlp.experts.linear_fc2.weight3 + module.decoder.layers.29.mlp.experts.linear_fc2.weight8 + module.decoder.layers.29.mlp.experts.linear_fc2.weight0 + module.decoder.layers.29.mlp.experts.linear_fc1.weight15 + module.decoder.layers.30.mlp.experts.linear_fc1.weight2 + module.decoder.layers.29.mlp.experts.linear_fc2.weight13 + module.decoder.layers.29.mlp.experts.linear_fc2.weight5 + module.decoder.layers.29.mlp.experts.linear_fc2.weight10 + module.decoder.layers.29.mlp.experts.linear_fc2.weight2 + module.decoder.layers.29.mlp.experts.linear_fc2.weight15 + module.decoder.layers.29.mlp.experts.linear_fc2.weight7 + module.decoder.layers.29.mlp.experts.linear_fc1.weight14 + module.decoder.layers.30.mlp.experts.linear_fc1.weight1 + module.decoder.layers.29.mlp.experts.linear_fc2.weight12 + module.decoder.layers.29.mlp.experts.linear_fc2.weight4 + module.decoder.layers.29.mlp.experts.linear_fc2.weight9 + module.decoder.layers.29.mlp.experts.linear_fc2.weight1 +Params for bucket 35 (40894464 elements, 40894464 padded size): + module.decoder.layers.29.mlp.experts.linear_fc1.weight13 + module.decoder.layers.29.mlp.experts.linear_fc1.weight9 + module.decoder.layers.29.mlp.experts.linear_fc1.weight11 + module.decoder.layers.29.mlp.experts.linear_fc1.weight7 + module.decoder.layers.29.mlp.experts.linear_fc1.weight5 + module.decoder.layers.29.mlp.experts.linear_fc1.weight3 + module.decoder.layers.29.mlp.experts.linear_fc1.weight1 + module.decoder.layers.29.mlp.experts.linear_fc1.weight12 + module.decoder.layers.29.mlp.experts.linear_fc1.weight10 + module.decoder.layers.29.mlp.experts.linear_fc1.weight8 + module.decoder.layers.29.mlp.experts.linear_fc1.weight6 + module.decoder.layers.29.mlp.experts.linear_fc1.weight4 + module.decoder.layers.29.mlp.experts.linear_fc1.weight2 +Params for bucket 36 (40894464 elements, 40894464 padded size): + module.decoder.layers.28.mlp.experts.linear_fc2.weight8 + module.decoder.layers.28.mlp.experts.linear_fc2.weight0 + module.decoder.layers.28.mlp.experts.linear_fc1.weight15 + module.decoder.layers.28.mlp.experts.linear_fc2.weight13 + module.decoder.layers.28.mlp.experts.linear_fc2.weight5 + module.decoder.layers.28.mlp.experts.linear_fc1.weight12 + module.decoder.layers.28.mlp.experts.linear_fc2.weight10 + module.decoder.layers.28.mlp.experts.linear_fc2.weight2 + module.decoder.layers.28.mlp.experts.linear_fc2.weight15 + module.decoder.layers.28.mlp.experts.linear_fc2.weight7 + module.decoder.layers.28.mlp.experts.linear_fc1.weight14 + module.decoder.layers.28.mlp.experts.linear_fc2.weight12 + module.decoder.layers.28.mlp.experts.linear_fc2.weight4 + module.decoder.layers.28.mlp.experts.linear_fc2.weight9 + module.decoder.layers.28.mlp.experts.linear_fc2.weight1 + module.decoder.layers.28.mlp.experts.linear_fc2.weight14 + module.decoder.layers.28.mlp.experts.linear_fc2.weight6 + module.decoder.layers.28.mlp.experts.linear_fc1.weight13 + module.decoder.layers.29.mlp.experts.linear_fc1.weight0 + module.decoder.layers.28.mlp.experts.linear_fc2.weight11 + module.decoder.layers.28.mlp.experts.linear_fc2.weight3 +Params for bucket 37 (40894464 elements, 40894464 padded size): + module.decoder.layers.28.mlp.experts.linear_fc1.weight11 + module.decoder.layers.28.mlp.experts.linear_fc1.weight9 + module.decoder.layers.28.mlp.experts.linear_fc1.weight7 + module.decoder.layers.28.mlp.experts.linear_fc1.weight5 + module.decoder.layers.28.mlp.experts.linear_fc1.weight3 + module.decoder.layers.28.mlp.experts.linear_fc1.weight1 + module.decoder.layers.27.mlp.experts.linear_fc2.weight14 + module.decoder.layers.28.mlp.experts.linear_fc1.weight10 + module.decoder.layers.28.mlp.experts.linear_fc1.weight8 + module.decoder.layers.28.mlp.experts.linear_fc1.weight6 + module.decoder.layers.28.mlp.experts.linear_fc1.weight4 + module.decoder.layers.28.mlp.experts.linear_fc1.weight2 + module.decoder.layers.28.mlp.experts.linear_fc1.weight0 + module.decoder.layers.27.mlp.experts.linear_fc2.weight15 +Params for bucket 38 (40894464 elements, 40894464 padded size): + module.decoder.layers.27.mlp.experts.linear_fc2.weight10 + module.decoder.layers.27.mlp.experts.linear_fc2.weight2 + module.decoder.layers.27.mlp.experts.linear_fc2.weight7 + module.decoder.layers.27.mlp.experts.linear_fc1.weight14 + module.decoder.layers.27.mlp.experts.linear_fc2.weight12 + module.decoder.layers.27.mlp.experts.linear_fc2.weight4 + module.decoder.layers.27.mlp.experts.linear_fc1.weight11 + module.decoder.layers.27.mlp.experts.linear_fc2.weight9 + module.decoder.layers.27.mlp.experts.linear_fc2.weight1 + module.decoder.layers.27.mlp.experts.linear_fc2.weight6 + module.decoder.layers.27.mlp.experts.linear_fc1.weight13 + module.decoder.layers.27.mlp.experts.linear_fc2.weight11 + module.decoder.layers.27.mlp.experts.linear_fc2.weight3 + module.decoder.layers.27.mlp.experts.linear_fc1.weight10 + module.decoder.layers.27.mlp.experts.linear_fc2.weight8 + module.decoder.layers.27.mlp.experts.linear_fc2.weight0 + module.decoder.layers.27.mlp.experts.linear_fc1.weight15 + module.decoder.layers.27.mlp.experts.linear_fc2.weight13 + module.decoder.layers.27.mlp.experts.linear_fc2.weight5 + module.decoder.layers.27.mlp.experts.linear_fc1.weight12 +Params for bucket 39 (40894464 elements, 40894464 padded size): + module.decoder.layers.27.mlp.experts.linear_fc1.weight9 + module.decoder.layers.27.mlp.experts.linear_fc1.weight7 + module.decoder.layers.27.mlp.experts.linear_fc1.weight5 + module.decoder.layers.27.mlp.experts.linear_fc1.weight3 + module.decoder.layers.27.mlp.experts.linear_fc1.weight1 + module.decoder.layers.26.mlp.experts.linear_fc2.weight14 + module.decoder.layers.26.mlp.experts.linear_fc2.weight12 + module.decoder.layers.26.mlp.experts.linear_fc2.weight10 + module.decoder.layers.27.mlp.experts.linear_fc1.weight6 + module.decoder.layers.27.mlp.experts.linear_fc1.weight8 + module.decoder.layers.27.mlp.experts.linear_fc1.weight4 + module.decoder.layers.27.mlp.experts.linear_fc1.weight2 + module.decoder.layers.27.mlp.experts.linear_fc1.weight0 + module.decoder.layers.26.mlp.experts.linear_fc2.weight15 + module.decoder.layers.26.mlp.experts.linear_fc2.weight13 + module.decoder.layers.26.mlp.experts.linear_fc2.weight11 +Params for bucket 40 (40894464 elements, 40894464 padded size): + module.decoder.layers.26.mlp.experts.linear_fc2.weight6 + module.decoder.layers.26.mlp.experts.linear_fc2.weight8 + module.decoder.layers.26.mlp.experts.linear_fc2.weight4 + module.decoder.layers.26.mlp.experts.linear_fc2.weight2 + module.decoder.layers.26.mlp.experts.linear_fc2.weight0 + module.decoder.layers.26.mlp.experts.linear_fc1.weight15 + module.decoder.layers.26.mlp.experts.linear_fc1.weight13 + module.decoder.layers.26.mlp.experts.linear_fc1.weight11 + module.decoder.layers.26.mlp.experts.linear_fc1.weight9 + module.decoder.layers.26.mlp.experts.linear_fc2.weight9 + module.decoder.layers.26.mlp.experts.linear_fc2.weight7 + module.decoder.layers.26.mlp.experts.linear_fc2.weight5 + module.decoder.layers.26.mlp.experts.linear_fc2.weight3 + module.decoder.layers.26.mlp.experts.linear_fc2.weight1 + module.decoder.layers.26.mlp.experts.linear_fc1.weight14 + module.decoder.layers.26.mlp.experts.linear_fc1.weight12 + module.decoder.layers.26.mlp.experts.linear_fc1.weight10 + module.decoder.layers.26.mlp.experts.linear_fc1.weight8 +Params for bucket 41 (40894464 elements, 40894464 padded size): + module.decoder.layers.26.mlp.experts.linear_fc1.weight7 + module.decoder.layers.26.mlp.experts.linear_fc1.weight3 + module.decoder.layers.26.mlp.experts.linear_fc1.weight5 + module.decoder.layers.26.mlp.experts.linear_fc1.weight1 + module.decoder.layers.25.mlp.experts.linear_fc2.weight14 + module.decoder.layers.25.mlp.experts.linear_fc2.weight12 + module.decoder.layers.25.mlp.experts.linear_fc2.weight10 + module.decoder.layers.25.mlp.experts.linear_fc2.weight8 + module.decoder.layers.25.mlp.experts.linear_fc2.weight6 + module.decoder.layers.26.mlp.experts.linear_fc1.weight4 + module.decoder.layers.26.mlp.experts.linear_fc1.weight6 + module.decoder.layers.26.mlp.experts.linear_fc1.weight2 + module.decoder.layers.26.mlp.experts.linear_fc1.weight0 + module.decoder.layers.25.mlp.experts.linear_fc2.weight15 + module.decoder.layers.25.mlp.experts.linear_fc2.weight13 + module.decoder.layers.25.mlp.experts.linear_fc2.weight11 + module.decoder.layers.25.mlp.experts.linear_fc2.weight9 + module.decoder.layers.25.mlp.experts.linear_fc2.weight7 +Params for bucket 42 (40894464 elements, 40894464 padded size): + module.decoder.layers.25.mlp.experts.linear_fc2.weight4 + module.decoder.layers.25.mlp.experts.linear_fc2.weight2 + module.decoder.layers.25.mlp.experts.linear_fc2.weight0 + module.decoder.layers.25.mlp.experts.linear_fc1.weight15 + module.decoder.layers.25.mlp.experts.linear_fc1.weight13 + module.decoder.layers.25.mlp.experts.linear_fc1.weight11 + module.decoder.layers.25.mlp.experts.linear_fc1.weight9 + module.decoder.layers.25.mlp.experts.linear_fc1.weight7 + module.decoder.layers.25.mlp.experts.linear_fc2.weight5 + module.decoder.layers.25.mlp.experts.linear_fc2.weight3 + module.decoder.layers.25.mlp.experts.linear_fc2.weight1 + module.decoder.layers.25.mlp.experts.linear_fc1.weight14 + module.decoder.layers.25.mlp.experts.linear_fc1.weight12 + module.decoder.layers.25.mlp.experts.linear_fc1.weight10 + module.decoder.layers.25.mlp.experts.linear_fc1.weight8 + module.decoder.layers.25.mlp.experts.linear_fc1.weight6 +Params for bucket 43 (40894464 elements, 40894464 padded size): + module.decoder.layers.25.mlp.experts.linear_fc1.weight5 + module.decoder.layers.24.mlp.experts.linear_fc2.weight8 + module.decoder.layers.25.mlp.experts.linear_fc1.weight2 + module.decoder.layers.24.mlp.experts.linear_fc2.weight13 + module.decoder.layers.24.mlp.experts.linear_fc2.weight5 + module.decoder.layers.24.mlp.experts.linear_fc2.weight10 + module.decoder.layers.24.mlp.experts.linear_fc2.weight2 + module.decoder.layers.25.mlp.experts.linear_fc1.weight4 + module.decoder.layers.24.mlp.experts.linear_fc2.weight15 + module.decoder.layers.24.mlp.experts.linear_fc2.weight7 + module.decoder.layers.25.mlp.experts.linear_fc1.weight1 + module.decoder.layers.24.mlp.experts.linear_fc2.weight12 + module.decoder.layers.24.mlp.experts.linear_fc2.weight4 + module.decoder.layers.24.mlp.experts.linear_fc2.weight9 + module.decoder.layers.25.mlp.experts.linear_fc1.weight3 + module.decoder.layers.24.mlp.experts.linear_fc2.weight14 + module.decoder.layers.24.mlp.experts.linear_fc2.weight6 + module.decoder.layers.25.mlp.experts.linear_fc1.weight0 + module.decoder.layers.24.mlp.experts.linear_fc2.weight11 + module.decoder.layers.24.mlp.experts.linear_fc2.weight3 +Params for bucket 44 (40894464 elements, 40894464 padded size): + module.decoder.layers.24.mlp.experts.linear_fc2.weight0 + module.decoder.layers.24.mlp.experts.linear_fc1.weight15 + module.decoder.layers.24.mlp.experts.linear_fc1.weight13 + module.decoder.layers.24.mlp.experts.linear_fc1.weight11 + module.decoder.layers.24.mlp.experts.linear_fc1.weight9 + module.decoder.layers.24.mlp.experts.linear_fc1.weight7 + module.decoder.layers.24.mlp.experts.linear_fc1.weight5 + module.decoder.layers.24.mlp.experts.linear_fc2.weight1 + module.decoder.layers.24.mlp.experts.linear_fc1.weight14 + module.decoder.layers.24.mlp.experts.linear_fc1.weight12 + module.decoder.layers.24.mlp.experts.linear_fc1.weight10 + module.decoder.layers.24.mlp.experts.linear_fc1.weight8 + module.decoder.layers.24.mlp.experts.linear_fc1.weight6 + module.decoder.layers.24.mlp.experts.linear_fc1.weight4 +Params for bucket 45 (40894464 elements, 40894464 padded size): + module.decoder.layers.23.mlp.experts.linear_fc2.weight10 + module.decoder.layers.23.mlp.experts.linear_fc2.weight2 + module.decoder.layers.23.mlp.experts.linear_fc2.weight15 + module.decoder.layers.23.mlp.experts.linear_fc2.weight7 + module.decoder.layers.24.mlp.experts.linear_fc1.weight1 + module.decoder.layers.23.mlp.experts.linear_fc2.weight12 + module.decoder.layers.23.mlp.experts.linear_fc2.weight4 + module.decoder.layers.23.mlp.experts.linear_fc1.weight15 + module.decoder.layers.23.mlp.experts.linear_fc2.weight9 + module.decoder.layers.23.mlp.experts.linear_fc2.weight1 + module.decoder.layers.24.mlp.experts.linear_fc1.weight3 + module.decoder.layers.23.mlp.experts.linear_fc2.weight14 + module.decoder.layers.23.mlp.experts.linear_fc2.weight6 + module.decoder.layers.24.mlp.experts.linear_fc1.weight0 + module.decoder.layers.23.mlp.experts.linear_fc2.weight11 + module.decoder.layers.23.mlp.experts.linear_fc2.weight3 + module.decoder.layers.23.mlp.experts.linear_fc2.weight8 + module.decoder.layers.23.mlp.experts.linear_fc2.weight0 + module.decoder.layers.24.mlp.experts.linear_fc1.weight2 + module.decoder.layers.23.mlp.experts.linear_fc2.weight13 + module.decoder.layers.23.mlp.experts.linear_fc2.weight5 +Params for bucket 46 (40894464 elements, 40894464 padded size): + module.decoder.layers.23.mlp.experts.linear_fc1.weight11 + module.decoder.layers.23.mlp.experts.linear_fc1.weight13 + module.decoder.layers.23.mlp.experts.linear_fc1.weight9 + module.decoder.layers.23.mlp.experts.linear_fc1.weight7 + module.decoder.layers.23.mlp.experts.linear_fc1.weight5 + module.decoder.layers.23.mlp.experts.linear_fc1.weight3 + module.decoder.layers.23.mlp.experts.linear_fc1.weight14 + module.decoder.layers.23.mlp.experts.linear_fc1.weight10 + module.decoder.layers.23.mlp.experts.linear_fc1.weight12 + module.decoder.layers.23.mlp.experts.linear_fc1.weight8 + module.decoder.layers.23.mlp.experts.linear_fc1.weight6 + module.decoder.layers.23.mlp.experts.linear_fc1.weight4 + module.decoder.layers.23.mlp.experts.linear_fc1.weight2 +Params for bucket 47 (40894464 elements, 40894464 padded size): + module.decoder.layers.23.mlp.experts.linear_fc1.weight1 + module.decoder.layers.22.mlp.experts.linear_fc2.weight12 + module.decoder.layers.22.mlp.experts.linear_fc2.weight4 + module.decoder.layers.22.mlp.experts.linear_fc1.weight15 + module.decoder.layers.22.mlp.experts.linear_fc2.weight9 + module.decoder.layers.22.mlp.experts.linear_fc2.weight1 + module.decoder.layers.22.mlp.experts.linear_fc2.weight14 + module.decoder.layers.22.mlp.experts.linear_fc2.weight6 + module.decoder.layers.23.mlp.experts.linear_fc1.weight0 + module.decoder.layers.22.mlp.experts.linear_fc2.weight11 + module.decoder.layers.22.mlp.experts.linear_fc2.weight3 + module.decoder.layers.22.mlp.experts.linear_fc1.weight14 + module.decoder.layers.22.mlp.experts.linear_fc2.weight8 + module.decoder.layers.22.mlp.experts.linear_fc2.weight0 + module.decoder.layers.22.mlp.experts.linear_fc2.weight13 + module.decoder.layers.22.mlp.experts.linear_fc2.weight5 + module.decoder.layers.22.mlp.experts.linear_fc2.weight10 + module.decoder.layers.22.mlp.experts.linear_fc2.weight2 + module.decoder.layers.22.mlp.experts.linear_fc1.weight13 + module.decoder.layers.22.mlp.experts.linear_fc2.weight15 + module.decoder.layers.22.mlp.experts.linear_fc2.weight7 +Params for bucket 48 (40894464 elements, 40894464 padded size): + module.decoder.layers.22.mlp.experts.linear_fc1.weight11 + module.decoder.layers.22.mlp.experts.linear_fc1.weight9 + module.decoder.layers.22.mlp.experts.linear_fc1.weight7 + module.decoder.layers.22.mlp.experts.linear_fc1.weight5 + module.decoder.layers.22.mlp.experts.linear_fc1.weight3 + module.decoder.layers.22.mlp.experts.linear_fc1.weight1 + module.decoder.layers.22.mlp.experts.linear_fc1.weight12 + module.decoder.layers.22.mlp.experts.linear_fc1.weight10 + module.decoder.layers.22.mlp.experts.linear_fc1.weight8 + module.decoder.layers.22.mlp.experts.linear_fc1.weight6 + module.decoder.layers.22.mlp.experts.linear_fc1.weight4 + module.decoder.layers.22.mlp.experts.linear_fc1.weight2 + module.decoder.layers.22.mlp.experts.linear_fc1.weight0 +Params for bucket 49 (40894464 elements, 40894464 padded size): + module.decoder.layers.21.mlp.experts.linear_fc2.weight15 + module.decoder.layers.21.mlp.experts.linear_fc2.weight7 + module.decoder.layers.21.mlp.experts.linear_fc2.weight12 + module.decoder.layers.21.mlp.experts.linear_fc2.weight4 + module.decoder.layers.21.mlp.experts.linear_fc1.weight15 + module.decoder.layers.21.mlp.experts.linear_fc2.weight9 + module.decoder.layers.21.mlp.experts.linear_fc2.weight1 + module.decoder.layers.21.mlp.experts.linear_fc1.weight12 + module.decoder.layers.21.mlp.experts.linear_fc2.weight14 + module.decoder.layers.21.mlp.experts.linear_fc2.weight6 + module.decoder.layers.21.mlp.experts.linear_fc2.weight11 + module.decoder.layers.21.mlp.experts.linear_fc2.weight3 + module.decoder.layers.21.mlp.experts.linear_fc1.weight14 + module.decoder.layers.21.mlp.experts.linear_fc2.weight8 + module.decoder.layers.21.mlp.experts.linear_fc2.weight0 + module.decoder.layers.21.mlp.experts.linear_fc1.weight11 + module.decoder.layers.21.mlp.experts.linear_fc2.weight13 + module.decoder.layers.21.mlp.experts.linear_fc2.weight5 + module.decoder.layers.21.mlp.experts.linear_fc2.weight10 + module.decoder.layers.21.mlp.experts.linear_fc2.weight2 + module.decoder.layers.21.mlp.experts.linear_fc1.weight13 +Params for bucket 50 (40894464 elements, 40894464 padded size): + module.decoder.layers.21.mlp.experts.linear_fc1.weight10 + module.decoder.layers.21.mlp.experts.linear_fc1.weight8 + module.decoder.layers.21.mlp.experts.linear_fc1.weight6 + module.decoder.layers.21.mlp.experts.linear_fc1.weight4 + module.decoder.layers.21.mlp.experts.linear_fc1.weight2 + module.decoder.layers.21.mlp.experts.linear_fc1.weight0 + module.decoder.layers.20.mlp.experts.linear_fc2.weight14 + module.decoder.layers.20.mlp.experts.linear_fc2.weight12 + module.decoder.layers.21.mlp.experts.linear_fc1.weight9 + module.decoder.layers.21.mlp.experts.linear_fc1.weight7 + module.decoder.layers.21.mlp.experts.linear_fc1.weight5 + module.decoder.layers.21.mlp.experts.linear_fc1.weight3 + module.decoder.layers.21.mlp.experts.linear_fc1.weight1 + module.decoder.layers.20.mlp.experts.linear_fc2.weight15 + module.decoder.layers.20.mlp.experts.linear_fc2.weight13 +Params for bucket 51 (40894464 elements, 40894464 padded size): + module.decoder.layers.20.mlp.experts.linear_fc2.weight10 + module.decoder.layers.20.mlp.experts.linear_fc2.weight2 + module.decoder.layers.20.mlp.experts.linear_fc1.weight13 + module.decoder.layers.20.mlp.experts.linear_fc2.weight7 + module.decoder.layers.20.mlp.experts.linear_fc1.weight10 + module.decoder.layers.20.mlp.experts.linear_fc2.weight4 + module.decoder.layers.20.mlp.experts.linear_fc1.weight15 + module.decoder.layers.20.mlp.experts.linear_fc2.weight9 + module.decoder.layers.20.mlp.experts.linear_fc2.weight1 + module.decoder.layers.20.mlp.experts.linear_fc1.weight12 + module.decoder.layers.20.mlp.experts.linear_fc2.weight6 + module.decoder.layers.20.mlp.experts.linear_fc1.weight9 + module.decoder.layers.20.mlp.experts.linear_fc2.weight11 + module.decoder.layers.20.mlp.experts.linear_fc2.weight3 + module.decoder.layers.20.mlp.experts.linear_fc1.weight14 + module.decoder.layers.20.mlp.experts.linear_fc2.weight8 + module.decoder.layers.20.mlp.experts.linear_fc2.weight0 + module.decoder.layers.20.mlp.experts.linear_fc1.weight11 + module.decoder.layers.20.mlp.experts.linear_fc2.weight5 +Params for bucket 52 (40894464 elements, 40894464 padded size): + module.decoder.layers.20.mlp.experts.linear_fc1.weight7 + module.decoder.layers.20.mlp.experts.linear_fc1.weight5 + module.decoder.layers.20.mlp.experts.linear_fc1.weight3 + module.decoder.layers.20.mlp.experts.linear_fc1.weight1 + module.decoder.layers.19.mlp.experts.linear_fc2.weight14 + module.decoder.layers.19.mlp.experts.linear_fc2.weight12 + module.decoder.layers.19.mlp.experts.linear_fc2.weight10 + module.decoder.layers.19.mlp.experts.linear_fc2.weight8 + module.decoder.layers.20.mlp.experts.linear_fc1.weight8 + module.decoder.layers.20.mlp.experts.linear_fc1.weight4 + module.decoder.layers.20.mlp.experts.linear_fc1.weight6 + module.decoder.layers.20.mlp.experts.linear_fc1.weight2 + module.decoder.layers.20.mlp.experts.linear_fc1.weight0 + module.decoder.layers.19.mlp.experts.linear_fc2.weight15 + module.decoder.layers.19.mlp.experts.linear_fc2.weight13 + module.decoder.layers.19.mlp.experts.linear_fc2.weight11 + module.decoder.layers.19.mlp.experts.linear_fc2.weight9 +Params for bucket 53 (40894464 elements, 40894464 padded size): + module.decoder.layers.19.mlp.experts.linear_fc2.weight4 + module.decoder.layers.19.mlp.experts.linear_fc2.weight6 + module.decoder.layers.19.mlp.experts.linear_fc2.weight2 + module.decoder.layers.19.mlp.experts.linear_fc2.weight0 + module.decoder.layers.19.mlp.experts.linear_fc1.weight15 + module.decoder.layers.19.mlp.experts.linear_fc1.weight13 + module.decoder.layers.19.mlp.experts.linear_fc1.weight11 + module.decoder.layers.19.mlp.experts.linear_fc1.weight9 + module.decoder.layers.19.mlp.experts.linear_fc1.weight7 + module.decoder.layers.19.mlp.experts.linear_fc2.weight7 + module.decoder.layers.19.mlp.experts.linear_fc2.weight5 + module.decoder.layers.19.mlp.experts.linear_fc2.weight3 + module.decoder.layers.19.mlp.experts.linear_fc2.weight1 + module.decoder.layers.19.mlp.experts.linear_fc1.weight14 + module.decoder.layers.19.mlp.experts.linear_fc1.weight12 + module.decoder.layers.19.mlp.experts.linear_fc1.weight10 + module.decoder.layers.19.mlp.experts.linear_fc1.weight8 +Params for bucket 54 (40894464 elements, 40894464 padded size): + module.decoder.layers.18.mlp.experts.linear_fc2.weight14 + module.decoder.layers.18.mlp.experts.linear_fc2.weight6 + module.decoder.layers.19.mlp.experts.linear_fc1.weight4 + module.decoder.layers.18.mlp.experts.linear_fc2.weight11 + module.decoder.layers.19.mlp.experts.linear_fc1.weight1 + module.decoder.layers.18.mlp.experts.linear_fc2.weight8 + module.decoder.layers.19.mlp.experts.linear_fc1.weight6 + module.decoder.layers.18.mlp.experts.linear_fc2.weight13 + module.decoder.layers.18.mlp.experts.linear_fc2.weight5 + module.decoder.layers.19.mlp.experts.linear_fc1.weight3 + module.decoder.layers.18.mlp.experts.linear_fc2.weight10 + module.decoder.layers.19.mlp.experts.linear_fc1.weight0 + module.decoder.layers.18.mlp.experts.linear_fc2.weight15 + module.decoder.layers.18.mlp.experts.linear_fc2.weight7 + module.decoder.layers.19.mlp.experts.linear_fc1.weight5 + module.decoder.layers.18.mlp.experts.linear_fc2.weight12 + module.decoder.layers.18.mlp.experts.linear_fc2.weight4 + module.decoder.layers.19.mlp.experts.linear_fc1.weight2 + module.decoder.layers.18.mlp.experts.linear_fc2.weight9 +Params for bucket 55 (40894464 elements, 40894464 padded size): + module.decoder.layers.18.mlp.experts.linear_fc2.weight2 + module.decoder.layers.18.mlp.experts.linear_fc1.weight15 + module.decoder.layers.18.mlp.experts.linear_fc2.weight0 + module.decoder.layers.18.mlp.experts.linear_fc1.weight13 + module.decoder.layers.18.mlp.experts.linear_fc1.weight11 + module.decoder.layers.18.mlp.experts.linear_fc1.weight9 + module.decoder.layers.18.mlp.experts.linear_fc1.weight7 + module.decoder.layers.18.mlp.experts.linear_fc1.weight5 + module.decoder.layers.18.mlp.experts.linear_fc2.weight3 + module.decoder.layers.18.mlp.experts.linear_fc2.weight1 + module.decoder.layers.18.mlp.experts.linear_fc1.weight14 + module.decoder.layers.18.mlp.experts.linear_fc1.weight12 + module.decoder.layers.18.mlp.experts.linear_fc1.weight10 + module.decoder.layers.18.mlp.experts.linear_fc1.weight8 + module.decoder.layers.18.mlp.experts.linear_fc1.weight6 +Params for bucket 56 (40894464 elements, 40894464 padded size): + module.decoder.layers.18.mlp.experts.linear_fc1.weight1 + module.decoder.layers.17.mlp.experts.linear_fc2.weight8 + module.decoder.layers.17.mlp.experts.linear_fc2.weight0 + module.decoder.layers.17.mlp.experts.linear_fc2.weight13 + module.decoder.layers.17.mlp.experts.linear_fc2.weight5 + module.decoder.layers.18.mlp.experts.linear_fc1.weight3 + module.decoder.layers.17.mlp.experts.linear_fc2.weight10 + module.decoder.layers.17.mlp.experts.linear_fc2.weight2 + module.decoder.layers.18.mlp.experts.linear_fc1.weight0 + module.decoder.layers.17.mlp.experts.linear_fc2.weight15 + module.decoder.layers.17.mlp.experts.linear_fc2.weight7 + module.decoder.layers.17.mlp.experts.linear_fc2.weight12 + module.decoder.layers.17.mlp.experts.linear_fc2.weight4 + module.decoder.layers.18.mlp.experts.linear_fc1.weight2 + module.decoder.layers.17.mlp.experts.linear_fc2.weight9 + module.decoder.layers.17.mlp.experts.linear_fc2.weight1 + module.decoder.layers.17.mlp.experts.linear_fc2.weight14 + module.decoder.layers.17.mlp.experts.linear_fc2.weight6 + module.decoder.layers.18.mlp.experts.linear_fc1.weight4 + module.decoder.layers.17.mlp.experts.linear_fc2.weight11 + module.decoder.layers.17.mlp.experts.linear_fc2.weight3 +Params for bucket 57 (40894464 elements, 40894464 padded size): + module.decoder.layers.17.mlp.experts.linear_fc1.weight15 + module.decoder.layers.17.mlp.experts.linear_fc1.weight11 + module.decoder.layers.17.mlp.experts.linear_fc1.weight13 + module.decoder.layers.17.mlp.experts.linear_fc1.weight9 + module.decoder.layers.17.mlp.experts.linear_fc1.weight7 + module.decoder.layers.17.mlp.experts.linear_fc1.weight5 + module.decoder.layers.17.mlp.experts.linear_fc1.weight3 + module.decoder.layers.17.mlp.experts.linear_fc1.weight12 + module.decoder.layers.17.mlp.experts.linear_fc1.weight14 + module.decoder.layers.17.mlp.experts.linear_fc1.weight10 + module.decoder.layers.17.mlp.experts.linear_fc1.weight8 + module.decoder.layers.17.mlp.experts.linear_fc1.weight6 + module.decoder.layers.17.mlp.experts.linear_fc1.weight4 +Params for bucket 58 (40894464 elements, 40894464 padded size): + module.decoder.layers.16.mlp.experts.linear_fc2.weight10 + module.decoder.layers.16.mlp.experts.linear_fc2.weight2 + module.decoder.layers.16.mlp.experts.linear_fc2.weight15 + module.decoder.layers.17.mlp.experts.linear_fc1.weight0 + module.decoder.layers.16.mlp.experts.linear_fc2.weight7 + module.decoder.layers.16.mlp.experts.linear_fc2.weight12 + module.decoder.layers.16.mlp.experts.linear_fc2.weight4 + module.decoder.layers.16.mlp.experts.linear_fc1.weight15 + module.decoder.layers.17.mlp.experts.linear_fc1.weight2 + module.decoder.layers.16.mlp.experts.linear_fc2.weight9 + module.decoder.layers.16.mlp.experts.linear_fc2.weight1 + module.decoder.layers.16.mlp.experts.linear_fc2.weight14 + module.decoder.layers.16.mlp.experts.linear_fc2.weight6 + module.decoder.layers.16.mlp.experts.linear_fc2.weight11 + module.decoder.layers.16.mlp.experts.linear_fc2.weight3 + module.decoder.layers.16.mlp.experts.linear_fc1.weight14 + module.decoder.layers.17.mlp.experts.linear_fc1.weight1 + module.decoder.layers.16.mlp.experts.linear_fc2.weight8 + module.decoder.layers.16.mlp.experts.linear_fc2.weight0 + module.decoder.layers.16.mlp.experts.linear_fc2.weight13 + module.decoder.layers.16.mlp.experts.linear_fc2.weight5 +Params for bucket 59 (40894464 elements, 40894464 padded size): + module.decoder.layers.16.mlp.experts.linear_fc1.weight13 + module.decoder.layers.16.mlp.experts.linear_fc1.weight11 + module.decoder.layers.16.mlp.experts.linear_fc1.weight9 + module.decoder.layers.16.mlp.experts.linear_fc1.weight7 + module.decoder.layers.16.mlp.experts.linear_fc1.weight5 + module.decoder.layers.16.mlp.experts.linear_fc1.weight3 + module.decoder.layers.16.mlp.experts.linear_fc1.weight1 + module.decoder.layers.16.mlp.experts.linear_fc1.weight12 + module.decoder.layers.16.mlp.experts.linear_fc1.weight10 + module.decoder.layers.16.mlp.experts.linear_fc1.weight8 + module.decoder.layers.16.mlp.experts.linear_fc1.weight6 + module.decoder.layers.16.mlp.experts.linear_fc1.weight4 + module.decoder.layers.16.mlp.experts.linear_fc1.weight2 +Params for bucket 60 (40894464 elements, 40894464 padded size): + module.decoder.layers.15.mlp.experts.linear_fc2.weight11 + module.decoder.layers.15.mlp.experts.linear_fc2.weight3 + module.decoder.layers.15.mlp.experts.linear_fc1.weight13 + module.decoder.layers.15.mlp.experts.linear_fc2.weight8 + module.decoder.layers.15.mlp.experts.linear_fc2.weight0 + module.decoder.layers.15.mlp.experts.linear_fc2.weight13 + module.decoder.layers.15.mlp.experts.linear_fc2.weight5 + module.decoder.layers.15.mlp.experts.linear_fc1.weight15 + module.decoder.layers.15.mlp.experts.linear_fc2.weight10 + module.decoder.layers.15.mlp.experts.linear_fc2.weight2 + module.decoder.layers.15.mlp.experts.linear_fc1.weight12 + module.decoder.layers.15.mlp.experts.linear_fc2.weight15 + module.decoder.layers.15.mlp.experts.linear_fc2.weight7 + module.decoder.layers.15.mlp.experts.linear_fc2.weight12 + module.decoder.layers.15.mlp.experts.linear_fc2.weight4 + module.decoder.layers.15.mlp.experts.linear_fc1.weight14 + module.decoder.layers.15.mlp.experts.linear_fc2.weight9 + module.decoder.layers.15.mlp.experts.linear_fc2.weight1 + module.decoder.layers.16.mlp.experts.linear_fc1.weight0 + module.decoder.layers.15.mlp.experts.linear_fc2.weight14 + module.decoder.layers.15.mlp.experts.linear_fc2.weight6 +Params for bucket 61 (40894464 elements, 40894464 padded size): + module.decoder.layers.15.mlp.experts.linear_fc1.weight11 + module.decoder.layers.15.mlp.experts.linear_fc1.weight7 + module.decoder.layers.15.mlp.experts.linear_fc1.weight9 + 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.weight1 + module.decoder.layers.14.mlp.experts.linear_fc2.weight14 + module.decoder.layers.15.mlp.experts.linear_fc1.weight10 + module.decoder.layers.15.mlp.experts.linear_fc1.weight8 + module.decoder.layers.15.mlp.experts.linear_fc1.weight6 + module.decoder.layers.15.mlp.experts.linear_fc1.weight4 + module.decoder.layers.15.mlp.experts.linear_fc1.weight2 + module.decoder.layers.15.mlp.experts.linear_fc1.weight0 + module.decoder.layers.14.mlp.experts.linear_fc2.weight15 +Params for bucket 62 (40894464 elements, 40894464 padded size): + module.decoder.layers.14.mlp.experts.linear_fc2.weight12 + module.decoder.layers.14.mlp.experts.linear_fc2.weight4 + module.decoder.layers.14.mlp.experts.linear_fc1.weight15 + module.decoder.layers.14.mlp.experts.linear_fc2.weight9 + module.decoder.layers.14.mlp.experts.linear_fc2.weight1 + module.decoder.layers.14.mlp.experts.linear_fc1.weight12 + module.decoder.layers.14.mlp.experts.linear_fc2.weight6 + module.decoder.layers.14.mlp.experts.linear_fc2.weight11 + module.decoder.layers.14.mlp.experts.linear_fc2.weight3 + module.decoder.layers.14.mlp.experts.linear_fc1.weight14 + module.decoder.layers.14.mlp.experts.linear_fc2.weight8 + module.decoder.layers.14.mlp.experts.linear_fc2.weight0 + module.decoder.layers.14.mlp.experts.linear_fc1.weight11 + module.decoder.layers.14.mlp.experts.linear_fc2.weight13 + module.decoder.layers.14.mlp.experts.linear_fc2.weight5 + module.decoder.layers.14.mlp.experts.linear_fc2.weight10 + module.decoder.layers.14.mlp.experts.linear_fc2.weight2 + module.decoder.layers.14.mlp.experts.linear_fc1.weight13 + module.decoder.layers.14.mlp.experts.linear_fc2.weight7 + module.decoder.layers.14.mlp.experts.linear_fc1.weight10 +Params for bucket 63 (40894464 elements, 40894464 padded size): + module.decoder.layers.14.mlp.experts.linear_fc1.weight9 + module.decoder.layers.14.mlp.experts.linear_fc1.weight7 + module.decoder.layers.14.mlp.experts.linear_fc1.weight5 + module.decoder.layers.14.mlp.experts.linear_fc1.weight3 + module.decoder.layers.14.mlp.experts.linear_fc1.weight1 + module.decoder.layers.13.mlp.experts.linear_fc2.weight14 + module.decoder.layers.13.mlp.experts.linear_fc2.weight12 + module.decoder.layers.13.mlp.experts.linear_fc2.weight10 + module.decoder.layers.14.mlp.experts.linear_fc1.weight8 + module.decoder.layers.14.mlp.experts.linear_fc1.weight6 + module.decoder.layers.14.mlp.experts.linear_fc1.weight4 + module.decoder.layers.14.mlp.experts.linear_fc1.weight2 + module.decoder.layers.14.mlp.experts.linear_fc1.weight0 + module.decoder.layers.13.mlp.experts.linear_fc2.weight15 + module.decoder.layers.13.mlp.experts.linear_fc2.weight13 + module.decoder.layers.13.mlp.experts.linear_fc2.weight11 +Params for bucket 64 (40894464 elements, 40894464 padded size): + module.decoder.layers.13.mlp.experts.linear_fc2.weight8 + module.decoder.layers.13.mlp.experts.linear_fc2.weight6 + module.decoder.layers.13.mlp.experts.linear_fc2.weight4 + module.decoder.layers.13.mlp.experts.linear_fc2.weight2 + module.decoder.layers.13.mlp.experts.linear_fc2.weight0 + module.decoder.layers.13.mlp.experts.linear_fc1.weight15 + module.decoder.layers.13.mlp.experts.linear_fc1.weight13 + module.decoder.layers.13.mlp.experts.linear_fc1.weight11 + module.decoder.layers.13.mlp.experts.linear_fc1.weight9 + module.decoder.layers.13.mlp.experts.linear_fc2.weight9 + module.decoder.layers.13.mlp.experts.linear_fc2.weight5 + module.decoder.layers.13.mlp.experts.linear_fc2.weight7 + module.decoder.layers.13.mlp.experts.linear_fc2.weight3 + module.decoder.layers.13.mlp.experts.linear_fc2.weight1 + module.decoder.layers.13.mlp.experts.linear_fc1.weight14 + module.decoder.layers.13.mlp.experts.linear_fc1.weight12 + module.decoder.layers.13.mlp.experts.linear_fc1.weight10 + module.decoder.layers.13.mlp.experts.linear_fc1.weight8 +Params for bucket 65 (40894464 elements, 40894464 padded size): + module.decoder.layers.13.mlp.experts.linear_fc1.weight7 + module.decoder.layers.13.mlp.experts.linear_fc1.weight5 + module.decoder.layers.13.mlp.experts.linear_fc1.weight3 + module.decoder.layers.13.mlp.experts.linear_fc1.weight1 + module.decoder.layers.12.mlp.experts.linear_fc2.weight14 + module.decoder.layers.12.mlp.experts.linear_fc2.weight12 + module.decoder.layers.12.mlp.experts.linear_fc2.weight10 + module.decoder.layers.12.mlp.experts.linear_fc2.weight8 + module.decoder.layers.12.mlp.experts.linear_fc2.weight6 + module.decoder.layers.13.mlp.experts.linear_fc1.weight4 + module.decoder.layers.13.mlp.experts.linear_fc1.weight6 + module.decoder.layers.13.mlp.experts.linear_fc1.weight2 + module.decoder.layers.13.mlp.experts.linear_fc1.weight0 + module.decoder.layers.12.mlp.experts.linear_fc2.weight15 + module.decoder.layers.12.mlp.experts.linear_fc2.weight13 + module.decoder.layers.12.mlp.experts.linear_fc2.weight11 + module.decoder.layers.12.mlp.experts.linear_fc2.weight9 + module.decoder.layers.12.mlp.experts.linear_fc2.weight7 +Params for bucket 66 (40894464 elements, 40894464 padded size): + module.decoder.layers.12.mlp.experts.linear_fc2.weight2 + module.decoder.layers.12.mlp.experts.linear_fc2.weight4 + module.decoder.layers.12.mlp.experts.linear_fc2.weight0 + module.decoder.layers.12.mlp.experts.linear_fc1.weight15 + module.decoder.layers.12.mlp.experts.linear_fc1.weight13 + module.decoder.layers.12.mlp.experts.linear_fc1.weight11 + module.decoder.layers.12.mlp.experts.linear_fc1.weight9 + module.decoder.layers.12.mlp.experts.linear_fc1.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.weight1 + module.decoder.layers.12.mlp.experts.linear_fc1.weight14 + module.decoder.layers.12.mlp.experts.linear_fc1.weight12 + module.decoder.layers.12.mlp.experts.linear_fc1.weight10 + module.decoder.layers.12.mlp.experts.linear_fc1.weight8 + module.decoder.layers.12.mlp.experts.linear_fc1.weight6 +Params for bucket 67 (40894464 elements, 40894464 padded size): + module.decoder.layers.12.mlp.experts.linear_fc1.weight1 + module.decoder.layers.11.mlp.experts.linear_fc2.weight8 + module.decoder.layers.11.mlp.experts.linear_fc2.weight13 + module.decoder.layers.11.mlp.experts.linear_fc2.weight5 + module.decoder.layers.12.mlp.experts.linear_fc1.weight3 + module.decoder.layers.11.mlp.experts.linear_fc2.weight10 + module.decoder.layers.11.mlp.experts.linear_fc2.weight2 + module.decoder.layers.12.mlp.experts.linear_fc1.weight0 + module.decoder.layers.11.mlp.experts.linear_fc2.weight15 + module.decoder.layers.11.mlp.experts.linear_fc2.weight7 + module.decoder.layers.12.mlp.experts.linear_fc1.weight5 + module.decoder.layers.11.mlp.experts.linear_fc2.weight12 + module.decoder.layers.11.mlp.experts.linear_fc2.weight4 + module.decoder.layers.12.mlp.experts.linear_fc1.weight2 + module.decoder.layers.11.mlp.experts.linear_fc2.weight9 + module.decoder.layers.11.mlp.experts.linear_fc2.weight14 + module.decoder.layers.11.mlp.experts.linear_fc2.weight6 + module.decoder.layers.12.mlp.experts.linear_fc1.weight4 + module.decoder.layers.11.mlp.experts.linear_fc2.weight11 + module.decoder.layers.11.mlp.experts.linear_fc2.weight3 +Params for bucket 68 (40894464 elements, 40894464 padded size): + module.decoder.layers.11.mlp.experts.linear_fc2.weight0 + module.decoder.layers.11.mlp.experts.linear_fc1.weight13 + module.decoder.layers.11.mlp.experts.linear_fc1.weight15 + module.decoder.layers.11.mlp.experts.linear_fc1.weight11 + module.decoder.layers.11.mlp.experts.linear_fc1.weight9 + module.decoder.layers.11.mlp.experts.linear_fc1.weight7 + module.decoder.layers.11.mlp.experts.linear_fc1.weight5 + module.decoder.layers.11.mlp.experts.linear_fc2.weight1 + module.decoder.layers.11.mlp.experts.linear_fc1.weight14 + module.decoder.layers.11.mlp.experts.linear_fc1.weight12 + module.decoder.layers.11.mlp.experts.linear_fc1.weight10 + module.decoder.layers.11.mlp.experts.linear_fc1.weight8 + module.decoder.layers.11.mlp.experts.linear_fc1.weight6 + module.decoder.layers.11.mlp.experts.linear_fc1.weight4 +Params for bucket 69 (40894464 elements, 40894464 padded size): + module.decoder.layers.11.mlp.experts.linear_fc1.weight3 + module.decoder.layers.10.mlp.experts.linear_fc2.weight10 + module.decoder.layers.10.mlp.experts.linear_fc2.weight2 + module.decoder.layers.11.mlp.experts.linear_fc1.weight0 + module.decoder.layers.10.mlp.experts.linear_fc2.weight15 + module.decoder.layers.10.mlp.experts.linear_fc2.weight7 + module.decoder.layers.10.mlp.experts.linear_fc2.weight12 + module.decoder.layers.10.mlp.experts.linear_fc2.weight4 + module.decoder.layers.10.mlp.experts.linear_fc1.weight15 + module.decoder.layers.11.mlp.experts.linear_fc1.weight2 + module.decoder.layers.10.mlp.experts.linear_fc2.weight9 + module.decoder.layers.10.mlp.experts.linear_fc2.weight1 + module.decoder.layers.10.mlp.experts.linear_fc2.weight14 + module.decoder.layers.10.mlp.experts.linear_fc2.weight6 + module.decoder.layers.10.mlp.experts.linear_fc2.weight11 + module.decoder.layers.10.mlp.experts.linear_fc2.weight3 + module.decoder.layers.11.mlp.experts.linear_fc1.weight1 + module.decoder.layers.10.mlp.experts.linear_fc2.weight8 + module.decoder.layers.10.mlp.experts.linear_fc2.weight0 + module.decoder.layers.10.mlp.experts.linear_fc2.weight13 + module.decoder.layers.10.mlp.experts.linear_fc2.weight5 +Params for bucket 70 (40894464 elements, 40894464 padded size): + module.decoder.layers.10.mlp.experts.linear_fc1.weight13 + module.decoder.layers.10.mlp.experts.linear_fc1.weight11 + module.decoder.layers.10.mlp.experts.linear_fc1.weight9 + module.decoder.layers.10.mlp.experts.linear_fc1.weight7 + 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.weight14 + module.decoder.layers.10.mlp.experts.linear_fc1.weight10 + module.decoder.layers.10.mlp.experts.linear_fc1.weight12 + module.decoder.layers.10.mlp.experts.linear_fc1.weight8 + module.decoder.layers.10.mlp.experts.linear_fc1.weight6 + module.decoder.layers.10.mlp.experts.linear_fc1.weight4 + module.decoder.layers.10.mlp.experts.linear_fc1.weight2 +Params for bucket 71 (40894464 elements, 40894464 padded size): + module.decoder.layers.9.mlp.experts.linear_fc2.weight12 + module.decoder.layers.9.mlp.experts.linear_fc2.weight4 + module.decoder.layers.9.mlp.experts.linear_fc1.weight15 + module.decoder.layers.9.mlp.experts.linear_fc2.weight9 + module.decoder.layers.9.mlp.experts.linear_fc2.weight1 + module.decoder.layers.9.mlp.experts.linear_fc2.weight14 + module.decoder.layers.9.mlp.experts.linear_fc2.weight6 + module.decoder.layers.9.mlp.experts.linear_fc2.weight11 + module.decoder.layers.9.mlp.experts.linear_fc2.weight3 + module.decoder.layers.9.mlp.experts.linear_fc1.weight14 + module.decoder.layers.10.mlp.experts.linear_fc1.weight1 + module.decoder.layers.9.mlp.experts.linear_fc2.weight8 + module.decoder.layers.9.mlp.experts.linear_fc2.weight0 + module.decoder.layers.9.mlp.experts.linear_fc2.weight13 + module.decoder.layers.9.mlp.experts.linear_fc2.weight5 + module.decoder.layers.9.mlp.experts.linear_fc2.weight10 + module.decoder.layers.9.mlp.experts.linear_fc2.weight2 + module.decoder.layers.9.mlp.experts.linear_fc1.weight13 + module.decoder.layers.9.mlp.experts.linear_fc2.weight15 + module.decoder.layers.10.mlp.experts.linear_fc1.weight0 + module.decoder.layers.9.mlp.experts.linear_fc2.weight7 +Params for bucket 72 (40894464 elements, 40894464 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight11 + module.decoder.layers.9.mlp.experts.linear_fc1.weight9 + module.decoder.layers.9.mlp.experts.linear_fc1.weight7 + 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.weight1 + module.decoder.layers.9.mlp.experts.linear_fc1.weight12 + module.decoder.layers.9.mlp.experts.linear_fc1.weight8 + module.decoder.layers.9.mlp.experts.linear_fc1.weight10 + module.decoder.layers.9.mlp.experts.linear_fc1.weight6 + module.decoder.layers.9.mlp.experts.linear_fc1.weight4 + module.decoder.layers.9.mlp.experts.linear_fc1.weight2 + module.decoder.layers.9.mlp.experts.linear_fc1.weight0 +Params for bucket 73 (40894464 elements, 40894464 padded size): + module.decoder.layers.8.mlp.experts.linear_fc2.weight14 + module.decoder.layers.8.mlp.experts.linear_fc2.weight6 + module.decoder.layers.8.mlp.experts.linear_fc2.weight11 + module.decoder.layers.8.mlp.experts.linear_fc2.weight3 + module.decoder.layers.8.mlp.experts.linear_fc1.weight14 + module.decoder.layers.8.mlp.experts.linear_fc2.weight8 + module.decoder.layers.8.mlp.experts.linear_fc2.weight0 + module.decoder.layers.8.mlp.experts.linear_fc1.weight11 + module.decoder.layers.8.mlp.experts.linear_fc2.weight13 + module.decoder.layers.8.mlp.experts.linear_fc2.weight5 + module.decoder.layers.8.mlp.experts.linear_fc2.weight10 + module.decoder.layers.8.mlp.experts.linear_fc2.weight2 + module.decoder.layers.8.mlp.experts.linear_fc1.weight13 + module.decoder.layers.8.mlp.experts.linear_fc2.weight15 + module.decoder.layers.8.mlp.experts.linear_fc2.weight7 + module.decoder.layers.8.mlp.experts.linear_fc2.weight12 + module.decoder.layers.8.mlp.experts.linear_fc2.weight4 + module.decoder.layers.8.mlp.experts.linear_fc1.weight15 + module.decoder.layers.8.mlp.experts.linear_fc2.weight9 + module.decoder.layers.8.mlp.experts.linear_fc2.weight1 + module.decoder.layers.8.mlp.experts.linear_fc1.weight12 +Params for bucket 74 (40894464 elements, 40894464 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight9 + module.decoder.layers.8.mlp.experts.linear_fc1.weight7 + module.decoder.layers.8.mlp.experts.linear_fc1.weight5 + module.decoder.layers.8.mlp.experts.linear_fc1.weight3 + module.decoder.layers.8.mlp.experts.linear_fc1.weight1 + module.decoder.layers.7.mlp.experts.linear_fc2.weight15 + module.decoder.layers.7.mlp.experts.linear_fc2.weight13 + module.decoder.layers.8.mlp.experts.linear_fc1.weight10 + module.decoder.layers.8.mlp.experts.linear_fc1.weight8 + module.decoder.layers.8.mlp.experts.linear_fc1.weight6 + module.decoder.layers.8.mlp.experts.linear_fc1.weight4 + module.decoder.layers.8.mlp.experts.linear_fc1.weight2 + module.decoder.layers.8.mlp.experts.linear_fc1.weight0 + module.decoder.layers.7.mlp.experts.linear_fc2.weight14 + module.decoder.layers.7.mlp.experts.linear_fc2.weight12 +Params for bucket 75 (40894464 elements, 40894464 padded size): + module.decoder.layers.7.mlp.experts.linear_fc2.weight7 + module.decoder.layers.7.mlp.experts.linear_fc1.weight10 + module.decoder.layers.7.mlp.experts.linear_fc2.weight4 + module.decoder.layers.7.mlp.experts.linear_fc1.weight15 + module.decoder.layers.7.mlp.experts.linear_fc2.weight9 + module.decoder.layers.7.mlp.experts.linear_fc2.weight1 + module.decoder.layers.7.mlp.experts.linear_fc1.weight12 + module.decoder.layers.7.mlp.experts.linear_fc2.weight6 + module.decoder.layers.7.mlp.experts.linear_fc1.weight9 + module.decoder.layers.7.mlp.experts.linear_fc2.weight11 + module.decoder.layers.7.mlp.experts.linear_fc2.weight3 + module.decoder.layers.7.mlp.experts.linear_fc1.weight14 + module.decoder.layers.7.mlp.experts.linear_fc2.weight8 + module.decoder.layers.7.mlp.experts.linear_fc2.weight0 + module.decoder.layers.7.mlp.experts.linear_fc1.weight11 + module.decoder.layers.7.mlp.experts.linear_fc2.weight5 + module.decoder.layers.7.mlp.experts.linear_fc2.weight10 + module.decoder.layers.7.mlp.experts.linear_fc2.weight2 + module.decoder.layers.7.mlp.experts.linear_fc1.weight13 +Params for bucket 76 (40894464 elements, 40894464 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight8 + module.decoder.layers.7.mlp.experts.linear_fc1.weight4 + module.decoder.layers.7.mlp.experts.linear_fc1.weight6 + module.decoder.layers.7.mlp.experts.linear_fc1.weight2 + module.decoder.layers.7.mlp.experts.linear_fc1.weight0 + module.decoder.layers.6.mlp.experts.linear_fc2.weight15 + module.decoder.layers.6.mlp.experts.linear_fc2.weight13 + module.decoder.layers.6.mlp.experts.linear_fc2.weight11 + module.decoder.layers.6.mlp.experts.linear_fc2.weight9 + module.decoder.layers.7.mlp.experts.linear_fc1.weight7 + module.decoder.layers.7.mlp.experts.linear_fc1.weight5 + module.decoder.layers.7.mlp.experts.linear_fc1.weight3 + module.decoder.layers.7.mlp.experts.linear_fc1.weight1 + module.decoder.layers.6.mlp.experts.linear_fc2.weight14 + module.decoder.layers.6.mlp.experts.linear_fc2.weight12 + module.decoder.layers.6.mlp.experts.linear_fc2.weight10 + module.decoder.layers.6.mlp.experts.linear_fc2.weight8 +Params for bucket 77 (40894464 elements, 40894464 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.weight1 + module.decoder.layers.6.mlp.experts.linear_fc1.weight14 + module.decoder.layers.6.mlp.experts.linear_fc1.weight12 + module.decoder.layers.6.mlp.experts.linear_fc1.weight10 + module.decoder.layers.6.mlp.experts.linear_fc1.weight8 + module.decoder.layers.6.mlp.experts.linear_fc2.weight6 + module.decoder.layers.6.mlp.experts.linear_fc2.weight4 + module.decoder.layers.6.mlp.experts.linear_fc2.weight2 + module.decoder.layers.6.mlp.experts.linear_fc2.weight0 + module.decoder.layers.6.mlp.experts.linear_fc1.weight15 + module.decoder.layers.6.mlp.experts.linear_fc1.weight13 + module.decoder.layers.6.mlp.experts.linear_fc1.weight11 + module.decoder.layers.6.mlp.experts.linear_fc1.weight9 + module.decoder.layers.6.mlp.experts.linear_fc1.weight7 +Params for bucket 78 (40894464 elements, 40894464 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight4 + module.decoder.layers.5.mlp.experts.linear_fc2.weight11 + module.decoder.layers.6.mlp.experts.linear_fc1.weight1 + module.decoder.layers.5.mlp.experts.linear_fc2.weight8 + module.decoder.layers.6.mlp.experts.linear_fc1.weight6 + module.decoder.layers.5.mlp.experts.linear_fc2.weight13 + module.decoder.layers.5.mlp.experts.linear_fc2.weight5 + module.decoder.layers.6.mlp.experts.linear_fc1.weight3 + module.decoder.layers.5.mlp.experts.linear_fc2.weight10 + module.decoder.layers.6.mlp.experts.linear_fc1.weight0 + module.decoder.layers.5.mlp.experts.linear_fc2.weight15 + module.decoder.layers.5.mlp.experts.linear_fc2.weight7 + module.decoder.layers.6.mlp.experts.linear_fc1.weight5 + module.decoder.layers.5.mlp.experts.linear_fc2.weight12 + module.decoder.layers.5.mlp.experts.linear_fc2.weight4 + module.decoder.layers.6.mlp.experts.linear_fc1.weight2 + module.decoder.layers.5.mlp.experts.linear_fc2.weight9 + module.decoder.layers.5.mlp.experts.linear_fc2.weight14 + module.decoder.layers.5.mlp.experts.linear_fc2.weight6 +Params for bucket 79 (40894464 elements, 40894464 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_fc1.weight14 + module.decoder.layers.5.mlp.experts.linear_fc1.weight12 + module.decoder.layers.5.mlp.experts.linear_fc1.weight10 + module.decoder.layers.5.mlp.experts.linear_fc1.weight8 + module.decoder.layers.5.mlp.experts.linear_fc1.weight6 + module.decoder.layers.5.mlp.experts.linear_fc2.weight2 + module.decoder.layers.5.mlp.experts.linear_fc1.weight15 + module.decoder.layers.5.mlp.experts.linear_fc2.weight0 + module.decoder.layers.5.mlp.experts.linear_fc1.weight13 + module.decoder.layers.5.mlp.experts.linear_fc1.weight11 + module.decoder.layers.5.mlp.experts.linear_fc1.weight9 + module.decoder.layers.5.mlp.experts.linear_fc1.weight7 + module.decoder.layers.5.mlp.experts.linear_fc1.weight5 +Params for bucket 80 (40894464 elements, 40894464 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight15 + module.decoder.layers.4.mlp.experts.linear_fc2.weight7 + module.decoder.layers.5.mlp.experts.linear_fc1.weight3 + module.decoder.layers.4.mlp.experts.linear_fc2.weight12 + module.decoder.layers.4.mlp.experts.linear_fc2.weight4 + module.decoder.layers.5.mlp.experts.linear_fc1.weight0 + module.decoder.layers.4.mlp.experts.linear_fc2.weight9 + module.decoder.layers.4.mlp.experts.linear_fc2.weight1 + module.decoder.layers.4.mlp.experts.linear_fc2.weight14 + module.decoder.layers.4.mlp.experts.linear_fc2.weight6 + module.decoder.layers.5.mlp.experts.linear_fc1.weight2 + module.decoder.layers.4.mlp.experts.linear_fc2.weight11 + module.decoder.layers.4.mlp.experts.linear_fc2.weight3 + module.decoder.layers.4.mlp.experts.linear_fc2.weight8 + module.decoder.layers.4.mlp.experts.linear_fc2.weight0 + module.decoder.layers.5.mlp.experts.linear_fc1.weight4 + module.decoder.layers.4.mlp.experts.linear_fc2.weight13 + module.decoder.layers.4.mlp.experts.linear_fc2.weight5 + module.decoder.layers.5.mlp.experts.linear_fc1.weight1 + module.decoder.layers.4.mlp.experts.linear_fc2.weight10 + module.decoder.layers.4.mlp.experts.linear_fc2.weight2 +Params for bucket 81 (40894464 elements, 40894464 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight14 + module.decoder.layers.4.mlp.experts.linear_fc1.weight12 + module.decoder.layers.4.mlp.experts.linear_fc1.weight10 + module.decoder.layers.4.mlp.experts.linear_fc1.weight8 + module.decoder.layers.4.mlp.experts.linear_fc1.weight6 + module.decoder.layers.4.mlp.experts.linear_fc1.weight4 + module.decoder.layers.4.mlp.experts.linear_fc1.weight15 + module.decoder.layers.4.mlp.experts.linear_fc1.weight13 + module.decoder.layers.4.mlp.experts.linear_fc1.weight11 + module.decoder.layers.4.mlp.experts.linear_fc1.weight9 + module.decoder.layers.4.mlp.experts.linear_fc1.weight7 + module.decoder.layers.4.mlp.experts.linear_fc1.weight5 + module.decoder.layers.4.mlp.experts.linear_fc1.weight3 +Params for bucket 82 (40894464 elements, 40894464 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight2 + module.decoder.layers.3.mlp.experts.linear_fc2.weight9 + module.decoder.layers.3.mlp.experts.linear_fc2.weight1 + module.decoder.layers.3.mlp.experts.linear_fc2.weight14 + module.decoder.layers.3.mlp.experts.linear_fc2.weight6 + module.decoder.layers.3.mlp.experts.linear_fc2.weight11 + module.decoder.layers.3.mlp.experts.linear_fc2.weight3 + module.decoder.layers.3.mlp.experts.linear_fc1.weight14 + module.decoder.layers.4.mlp.experts.linear_fc1.weight1 + module.decoder.layers.3.mlp.experts.linear_fc2.weight8 + module.decoder.layers.3.mlp.experts.linear_fc2.weight0 + module.decoder.layers.3.mlp.experts.linear_fc2.weight13 + module.decoder.layers.3.mlp.experts.linear_fc2.weight5 + module.decoder.layers.3.mlp.experts.linear_fc2.weight10 + module.decoder.layers.3.mlp.experts.linear_fc2.weight2 + module.decoder.layers.3.mlp.experts.linear_fc2.weight15 + module.decoder.layers.4.mlp.experts.linear_fc1.weight0 + module.decoder.layers.3.mlp.experts.linear_fc2.weight7 + module.decoder.layers.3.mlp.experts.linear_fc2.weight12 + module.decoder.layers.3.mlp.experts.linear_fc2.weight4 + module.decoder.layers.3.mlp.experts.linear_fc1.weight15 +Params for bucket 83 (40894464 elements, 40894464 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight10 + module.decoder.layers.3.mlp.experts.linear_fc1.weight12 + module.decoder.layers.3.mlp.experts.linear_fc1.weight8 + module.decoder.layers.3.mlp.experts.linear_fc1.weight6 + module.decoder.layers.3.mlp.experts.linear_fc1.weight4 + module.decoder.layers.3.mlp.experts.linear_fc1.weight2 + module.decoder.layers.3.mlp.experts.linear_fc1.weight13 + module.decoder.layers.3.mlp.experts.linear_fc1.weight11 + module.decoder.layers.3.mlp.experts.linear_fc1.weight9 + module.decoder.layers.3.mlp.experts.linear_fc1.weight7 + 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.weight1 +Params for bucket 84 (40894464 elements, 40894464 padded size): + module.decoder.layers.2.mlp.experts.linear_fc2.weight10 + module.decoder.layers.2.mlp.experts.linear_fc2.weight2 + module.decoder.layers.2.mlp.experts.linear_fc1.weight13 + module.decoder.layers.2.mlp.experts.linear_fc2.weight15 + module.decoder.layers.2.mlp.experts.linear_fc2.weight7 + module.decoder.layers.2.mlp.experts.linear_fc2.weight12 + module.decoder.layers.2.mlp.experts.linear_fc2.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight15 + module.decoder.layers.2.mlp.experts.linear_fc2.weight9 + module.decoder.layers.2.mlp.experts.linear_fc2.weight1 + module.decoder.layers.2.mlp.experts.linear_fc1.weight12 + module.decoder.layers.3.mlp.experts.linear_fc1.weight0 + module.decoder.layers.2.mlp.experts.linear_fc2.weight14 + module.decoder.layers.2.mlp.experts.linear_fc2.weight6 + module.decoder.layers.2.mlp.experts.linear_fc2.weight11 + module.decoder.layers.2.mlp.experts.linear_fc2.weight3 + module.decoder.layers.2.mlp.experts.linear_fc1.weight14 + module.decoder.layers.2.mlp.experts.linear_fc2.weight8 + module.decoder.layers.2.mlp.experts.linear_fc2.weight0 + module.decoder.layers.2.mlp.experts.linear_fc2.weight13 + module.decoder.layers.2.mlp.experts.linear_fc2.weight5 +Params for bucket 85 (40894464 elements, 40894464 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight11 + module.decoder.layers.2.mlp.experts.linear_fc1.weight7 + module.decoder.layers.2.mlp.experts.linear_fc1.weight9 + module.decoder.layers.2.mlp.experts.linear_fc1.weight5 + module.decoder.layers.2.mlp.experts.linear_fc1.weight3 + module.decoder.layers.2.mlp.experts.linear_fc1.weight1 + module.decoder.layers.1.mlp.experts.linear_fc2.weight14 + module.decoder.layers.2.mlp.experts.linear_fc1.weight10 + module.decoder.layers.2.mlp.experts.linear_fc1.weight8 + module.decoder.layers.2.mlp.experts.linear_fc1.weight6 + module.decoder.layers.2.mlp.experts.linear_fc1.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight2 + module.decoder.layers.2.mlp.experts.linear_fc1.weight0 + module.decoder.layers.1.mlp.experts.linear_fc2.weight15 +Params for bucket 86 (40894464 elements, 40894464 padded size): + module.decoder.layers.1.mlp.experts.linear_fc2.weight12 + module.decoder.layers.1.mlp.experts.linear_fc2.weight4 + module.decoder.layers.1.mlp.experts.linear_fc1.weight15 + module.decoder.layers.1.mlp.experts.linear_fc2.weight9 + module.decoder.layers.1.mlp.experts.linear_fc2.weight1 + module.decoder.layers.1.mlp.experts.linear_fc1.weight12 + module.decoder.layers.1.mlp.experts.linear_fc2.weight6 + module.decoder.layers.1.mlp.experts.linear_fc2.weight11 + module.decoder.layers.1.mlp.experts.linear_fc2.weight3 + module.decoder.layers.1.mlp.experts.linear_fc1.weight14 + module.decoder.layers.1.mlp.experts.linear_fc2.weight8 + module.decoder.layers.1.mlp.experts.linear_fc2.weight0 + module.decoder.layers.1.mlp.experts.linear_fc1.weight11 + module.decoder.layers.1.mlp.experts.linear_fc2.weight13 + module.decoder.layers.1.mlp.experts.linear_fc2.weight5 + module.decoder.layers.1.mlp.experts.linear_fc2.weight10 + module.decoder.layers.1.mlp.experts.linear_fc2.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight13 + module.decoder.layers.1.mlp.experts.linear_fc2.weight7 + module.decoder.layers.1.mlp.experts.linear_fc1.weight10 +Params for bucket 87 (40894464 elements, 40894464 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight9 + module.decoder.layers.1.mlp.experts.linear_fc1.weight7 + 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.weight1 + module.decoder.layers.0.mlp.experts.linear_fc2.weight14 + module.decoder.layers.0.mlp.experts.linear_fc2.weight12 + module.decoder.layers.0.mlp.experts.linear_fc2.weight10 + module.decoder.layers.1.mlp.experts.linear_fc1.weight8 + module.decoder.layers.1.mlp.experts.linear_fc1.weight6 + module.decoder.layers.1.mlp.experts.linear_fc1.weight4 + module.decoder.layers.1.mlp.experts.linear_fc1.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight0 + module.decoder.layers.0.mlp.experts.linear_fc2.weight15 + module.decoder.layers.0.mlp.experts.linear_fc2.weight13 + module.decoder.layers.0.mlp.experts.linear_fc2.weight11 +Params for bucket 88 (40894464 elements, 40894464 padded size): + module.decoder.layers.0.mlp.experts.linear_fc2.weight8 + 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.weight0 + module.decoder.layers.0.mlp.experts.linear_fc1.weight15 + module.decoder.layers.0.mlp.experts.linear_fc1.weight13 + module.decoder.layers.0.mlp.experts.linear_fc1.weight11 + module.decoder.layers.0.mlp.experts.linear_fc1.weight9 + module.decoder.layers.0.mlp.experts.linear_fc2.weight9 + 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 + module.decoder.layers.0.mlp.experts.linear_fc2.weight1 + module.decoder.layers.0.mlp.experts.linear_fc1.weight14 + module.decoder.layers.0.mlp.experts.linear_fc1.weight12 + module.decoder.layers.0.mlp.experts.linear_fc1.weight10 + module.decoder.layers.0.mlp.experts.linear_fc1.weight8 +Params for bucket 89 (25165824 elements, 25165824 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight7 + module.decoder.layers.0.mlp.experts.linear_fc1.weight3 + module.decoder.layers.0.mlp.experts.linear_fc1.weight5 + module.decoder.layers.0.mlp.experts.linear_fc1.weight1 + module.decoder.layers.0.mlp.experts.linear_fc1.weight6 + module.decoder.layers.0.mlp.experts.linear_fc1.weight4 + module.decoder.layers.0.mlp.experts.linear_fc1.weight2 + module.decoder.layers.0.mlp.experts.linear_fc1.weight0 +INFO:megatron.core.optimizer:Setting up optimizer with config OptimizerConfig(optimizer='adam', lr=0.0001, min_lr=1e-05, decoupled_lr=None, decoupled_min_lr=None, weight_decay=0.1, fp8_recipe='mxfp8', fp16=False, bf16=True, reuse_grad_buf_for_mxfp8_param_ag=False, params_dtype=torch.bfloat16, use_precision_aware_optimizer=False, store_param_remainders=True, main_grads_dtype=torch.float32, main_params_dtype=torch.float32, exp_avg_dtype=torch.float32, exp_avg_sq_dtype=torch.float32, loss_scale=None, initial_loss_scale=4294967296, min_loss_scale=1.0, loss_scale_window=1000, hysteresis=2, adam_beta1=0.9, adam_beta2=0.999, adam_eps=1e-08, sgd_momentum=0.9, use_distributed_optimizer=True, overlap_param_gather_with_optimizer_step=False, optimizer_cpu_offload=False, optimizer_offload_fraction=1.0, use_torch_optimizer_for_cpu_offload=False, overlap_cpu_optimizer_d2h_h2d=False, pin_cpu_grads=True, pin_cpu_params=True, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=, config_logger_dir='') +INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine +[after model, optimizer, and learning rate scheduler are built] datetime: 2026-08-26 18:33:00 +> building train, validation, and test datasets ... + > datasets target sizes (minimum size): + train: 96 + 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=(96, 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 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-document_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-sample_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-shuffle_index.npy +INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073 +> finished creating GPT datasets ... +[after dataloaders are built] datetime: 2026-08-26 18:33:00 +done with setup ... +(min, max) time across ranks (ms): + model-and-optimizer-setup ......................: (4648.09, 4665.34) + train/valid/test-data-iterators-setup ..........: (67.19, 84.85) +training ... +Overwriting rerun_state_machine.current_iteration from -1 to 0... +[before the start of training step] datetime: 2026-08-26 18:33:00 +/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: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: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:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") + [2026-08-26 18:33:26] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 26428.6 | throughput per GPU (TFLOP/s/GPU): 3.4 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.095286E+01 | loss scale: 1.0 | grad norm: 54.366 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 |Number of parameters in transformer block in billions: 29.90 + +Number of parameters in embedding layers in billions: 0.21 +Total number of parameters in billions: 30.10 +Number of parameters in most loaded shard in billions: 30.1034 +Theoretical memory footprints: weight and optimizer=215316.59 MB +[Rank 0] (after 1 iterations) memory (MB) | allocated: 72365.84375 | max allocated: 78646.15576171875 | reserved: 79250.0 | max reserved: 79250.0 +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") + [2026-08-26 18:33:27] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 1135.6 | throughput per GPU (TFLOP/s/GPU): 78.5 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.095473E+01 | loss scale: 1.0 | grad norm: 56.502 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:33:28] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 929.9 | throughput per GPU (TFLOP/s/GPU): 95.8 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 1.008671E+01 | loss scale: 1.0 | grad norm: 180.233 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | +Capture CUDA graph for training!!! +CUDA graph capture done!!! + [2026-08-26 18:33:31] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 3039.0 | throughput per GPU (TFLOP/s/GPU): 29.3 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 9.604192E+00 | loss scale: 1.0 | grad norm: 8.718 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:33:32] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 358.4 | throughput per GPU (TFLOP/s/GPU): 248.7 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 9.307684E+00 | loss scale: 1.0 | grad norm: 6.781 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:33:32] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 356.8 | throughput per GPU (TFLOP/s/GPU): 249.8 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 9.149645E+00 | loss scale: 1.0 | grad norm: 3.068 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:33:32] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 354.3 | throughput per GPU (TFLOP/s/GPU): 251.6 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 8.997829E+00 | loss scale: 1.0 | grad norm: 3.397 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:33:33] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 353.7 | throughput per GPU (TFLOP/s/GPU): 251.9 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.841221E+00 | loss scale: 1.0 | grad norm: 3.065 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:33:33] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 353.5 | throughput per GPU (TFLOP/s/GPU): 252.1 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.793358E+00 | loss scale: 1.0 | grad norm: 2.779 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:33:33] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 355.1 | throughput per GPU (TFLOP/s/GPU): 251.0 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.752890E+00 | loss scale: 1.0 | grad norm: 2.735 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:33:34] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 352.7 | throughput per GPU (TFLOP/s/GPU): 252.7 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.745195E+00 | loss scale: 1.0 | grad norm: 2.555 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:33:34] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 354.4 | throughput per GPU (TFLOP/s/GPU): 251.4 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.722492E+00 | loss scale: 1.0 | grad norm: 2.433 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | +[after training is done] datetime: 2026-08-26 18:33:34 +[rank0]:[W826 18:33:35.829006250 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank1]:[W826 18:33:37.270337014 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank4]:[W826 18:33:37.289257225 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank2]:[W826 18:33:37.289330075 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank3]:[W826 18:33:37.289787519 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank5]:[W826 18:33:37.516105909 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank7]:[W826 18:33:37.518015746 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank6]:[W826 18:33:37.543231823 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +W0826 18:33:40.206000 522564 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers +W0826 18:33:40.206000 522564 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 522653 closing signal SIGTERM +W0826 18:33:40.207000 522564 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 522654 closing signal SIGTERM +W0826 18:33:40.209000 522564 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 522655 closing signal SIGTERM +W0826 18:33:40.210000 522564 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 522656 closing signal SIGTERM +W0826 18:33:40.211000 522564 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 522657 closing signal SIGTERM +W0826 18:33:40.212000 522564 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 522658 closing signal SIGTERM +W0826 18:33:40.214000 522564 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 522659 closing signal SIGTERM +W0826 18:33:40.216000 522564 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 522660 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 522564 got signal: 15 diff --git a/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep3/mem_samples.csv b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep3/mem_samples.csv new file mode 100644 index 0000000000000000000000000000000000000000..ccb6b0a565f10699ce7d75d9f9268b20973cd097 --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep3/mem_samples.csv @@ -0,0 +1,232 @@ +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 0 +1, 0 +2, 0 +3, 0 +4, 0 +5, 0 +6, 0 +7, 0 +0, 1 +1, 1 +2, 1 +3, 1 +4, 1 +5, 1 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 1 +5, 1 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 4 +5, 4 +6, 1 +7, 1 +0, 4 +1, 4 +2, 4 +3, 4 +4, 4 +5, 4 +6, 4 +7, 4 +0, 821 +1, 821 +2, 821 +3, 821 +4, 821 +5, 821 +6, 821 +7, 821 +0, 1831 +1, 1831 +2, 1831 +3, 1831 +4, 1831 +5, 1831 +6, 1831 +7, 1831 +0, 2151 +1, 2161 +2, 2163 +3, 2155 +4, 2153 +5, 2145 +6, 2157 +7, 2141 +0, 9851 +1, 10251 +2, 10051 +3, 10051 +4, 10291 +5, 10131 +6, 9851 +7, 10251 +0, 46999 +1, 47003 +2, 47021 +3, 46939 +4, 47019 +5, 46983 +6, 47007 +7, 46989 +0, 48123 +1, 48145 +2, 48205 +3, 48147 +4, 48211 +5, 48175 +6, 48225 +7, 48205 +0, 48647 +1, 48649 +2, 48659 +3, 48621 +4, 48645 +5, 48659 +6, 48649 +7, 48659 +0, 48791 +1, 48793 +2, 48803 +3, 48765 +4, 48789 +5, 48803 +6, 48793 +7, 48803 +0, 49447 +1, 49447 +2, 49451 +3, 49421 +4, 49419 +5, 49451 +6, 49405 +7, 49451 +0, 56771 +1, 56791 +2, 56769 +3, 56785 +4, 56781 +5, 56767 +6, 56769 +7, 56767 +0, 85175 +1, 85227 +2, 85177 +3, 85201 +4, 85215 +5, 85163 +6, 85171 +7, 85167 +0, 85607 +1, 85629 +2, 85605 +3, 85621 +4, 85621 +5, 85605 +6, 85605 +7, 85605 +0, 85607 +1, 85629 +2, 85605 +3, 85621 +4, 85621 +5, 85605 +6, 85605 +7, 85605 +0, 85907 +1, 85897 +2, 85913 +3, 85923 +4, 85933 +5, 85845 +6, 85929 +7, 85863 +0, 86817 +1, 86879 +2, 86857 +3, 86843 +4, 86847 +5, 86807 +6, 86855 +7, 86871 +0, 108415 +1, 108783 +2, 108069 +3, 108077 +4, 108447 +5, 109085 +6, 108093 +7, 108577 +0, 105469 +1, 105543 +2, 105615 +3, 104923 +4, 105683 +5, 95563 +6, 104547 +7, 103735 +0, 120103 +1, 120129 +2, 120099 +3, 119829 +4, 120121 +5, 120099 +6, 120101 +7, 120099 +0, 120103 +1, 120129 +2, 120099 +3, 119829 +4, 120121 +5, 120099 +6, 120101 +7, 120099 +0, 120103 +1, 120129 +2, 120099 +3, 119829 +4, 120121 +5, 120099 +6, 120101 +7, 120099 +0, 116837 +1, 117125 +2, 117095 +3, 116825 +4, 117117 +5, 117095 +6, 117097 +7, 117095 +0, 115613 +1, 115377 +2, 115091 +3, 115403 +4, 115433 +5, 115673 +6, 115413 +7, 115353 +0, 641 +1, 4 +2, 4 +3, 4 +4, 4 +5, 4 +6, 132 +7, 115289 diff --git a/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep3/run.env b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep3/run.env new file mode 100644 index 0000000000000000000000000000000000000000..3213d7fbe624a565f15157e5d63615da960f6547 --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep3/run.env @@ -0,0 +1,61 @@ +backend=ultraep +config=QWEN3_30B +family=tierB +repo=/workspace/Megatron-UltraEP +branch=ultraep-bracing-regions-through-kernels +commit=03d57ef7 +num_layers=48 +hidden=2048 +ffn_hidden=6144 +moe_ffn_hidden=768 +num_experts=128 +topk=8 +experts_per_rank=16 +expert_params=4718592 +heads=32 +kv_groups=4 +kv_channels=128 +qk_layernorm=1 +shared_ffn=0 +shared_expert_overlap=0 +mla=0 +dense_first_layers=0 +moe_layers=48 +rotary_base=1000000 +rotary_scaling= +qkv_bias=0 +router_balancing=aux_loss +aux_loss_coeff=1e-3 +router_pre_softmax=0 +router_fusion=0 +router_dtype=fp32 +device_initiated_grouped_gemm=1 +norm_epsilon=1e-6 +seq_length=4096 +micro_batch_size=1 +global_batch_size=8 +num_microbatches=1 +assignments_per_rank=32768 +mean_tokens_per_expert=2048 +slot_policy=matched +slots_per_rank=16 +capacity_mode=static +capacity=1.11328125 +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=12 +ep_size=8 +echo_dispatch_overlap=1 +echo_dispatch_recompute=0 +ultraep_quota_min_tokens=256 +ultraep_weight_sync_join= +profile_enable=0 +kernel_trace=0 +fp8_recipe=mxfp8 diff --git a/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep3/train.log b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep3/train.log new file mode 100644 index 0000000000000000000000000000000000000000..4eb65e34f528572e2e4b84c78572f7e8dcda9f7f --- /dev/null +++ b/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep3/train.log @@ -0,0 +1,2878 @@ +[18:33:45] WARN: ultraep-bracing-regions-through-kernels lacks --moe-router-fusion; QWEN3_30B runs unfused (timing only, routing identical) +============================================================== + QWEN3_30B / ultraep (slots/rank=16 policy=matched) + E=128 K=8 H'=768 m=16 + L=48 seq=4096 gbs=8 (1 microbatches) + cap_mode=static cap=1.11328125 cg=1 recompute=none + memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep3/memory_snapshots/snapshot.pickle + repo=/workspace/Megatron-UltraEP @ 03d57ef7 +============================================================== +W0826 18:33:46.730000 525651 torch/distributed/run.py:766] +W0826 18:33:46.730000 525651 torch/distributed/run.py:766] ***************************************** +W0826 18:33:46.730000 525651 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0826 18:33:46.730000 525651 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") +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") +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 .............................. 48 + 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 ................................. 6144 + finetune ........................................ False + finetune_data_split ............................. train + finetune_hf_dataset ............................. None + first_last_layers_bf16 .......................... False + flash_decode .................................... False + fp16 ............................................ False + fp16_lm_cross_entropy ........................... False + fp32_residual_connection ........................ False + fp8 ............................................. e4m3 + fp8_amax_compute_algo ........................... most_recent + fp8_amax_history_len ............................ 1 + fp8_interval .................................... 1 + fp8_margin ...................................... 0 + fp8_param_gather ................................ False + fp8_recipe ...................................... mxfp8 + fp8_wgrad ....................................... True + fsdp_double_buffer .............................. False + global_batch_size ............................... 8 + grad_reduce_in_bf16 ............................. False + gradient_accumulation_fusion .................... True + gradient_reduce_div_fusion ...................... True + group_query_attention ........................... True + head_lr_mult .................................... 1.0 + heterogeneous_layers_config_encoded_json ........ None + heterogeneous_layers_config_path ................ None + hidden_dropout .................................. 0.0 + hidden_size ..................................... 2048 + hierarchical_context_parallel_sizes ............. None + high_priority_stream_groups ..................... [] + hybrid_attention_ratio .......................... 0.0 + hybrid_mlp_ratio ................................ 0.0 + hybrid_override_pattern ......................... None + hysteresis ...................................... 2 + ict_head_size ................................... None + ict_load ........................................ None + img_h ........................................... 224 + img_w ........................................... 224 + indexer_batch_size .............................. 128 + indexer_log_interval ............................ 1000 + inference_batch_times_seqlen_threshold .......... -1 + inference_dynamic_batching ...................... False + inference_dynamic_batching_buffer_guaranteed_fraction 0.2 + inference_dynamic_batching_buffer_overflow_factor None + inference_dynamic_batching_buffer_size_gb ....... 40.0 + inference_dynamic_batching_chunk_size ........... 256 + inference_dynamic_batching_max_requests_override None + inference_dynamic_batching_max_tokens_override .. None + inference_dynamic_batching_num_cuda_graphs ...... 16 + inference_max_batch_size ........................ 8 + inference_max_seq_length ........................ 2560 + inference_rng_tracker ........................... False + init_method_std ................................. 0.01 + init_method_xavier_uniform ...................... False + init_model_with_meta_device ..................... False + initial_loss_scale .............................. 4294967296 + inprocess_active_world_size ..................... 8 + inprocess_barrier_timeout ....................... 120 + inprocess_completion_timeout .................... 120 + inprocess_empty_cuda_cache ...................... False + inprocess_granularity ........................... node + inprocess_hard_timeout .......................... 90 + inprocess_heartbeat_interval .................... 30 + inprocess_heartbeat_timeout ..................... 60 + inprocess_last_call_wait ........................ 1 + inprocess_max_iterations ........................ None + inprocess_monitor_process_interval .............. 1.0 + inprocess_monitor_thread_interval ............... 1.0 + inprocess_progress_watchdog_interval ............ 1.0 + inprocess_restart ............................... False + inprocess_soft_timeout .......................... 60 + inprocess_termination_grace_time ................ 1 + is_hybrid_model ................................. False + iter_per_epoch .................................. 1250 + iterations_to_skip .............................. [] + keep_fp8_transpose_cache_when_using_custom_fsdp . False + kitchen_config_file ............................. None + kitchen_recipe_number ........................... None + kv_channels ..................................... 128 + kv_lora_rank .................................... 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 .............................. 0.001 + moe_capacity_monitor ............................ True + moe_deepep_num_sms .............................. 20 + moe_echo_enable_random_offloading ............... False + moe_echo_expert_dispatch_overlap ................ False + moe_echo_expert_dispatcher_type ................. hybridep + moe_echo_recompute_expert_dispatch .............. False + moe_enable_deepep ............................... False + moe_enable_echo ................................. False + moe_enable_hybridep ............................. True + moe_enable_ultraep .............................. True + moe_expert_capacity_factor ...................... None + moe_extended_tp ................................. False + moe_ffn_hidden_size ............................. 768 + 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 .............. 16 + moe_pad_expert_input_to_capacity ................ False + moe_per_layer_logging ........................... False + moe_permute_fusion .............................. False + moe_received_token_capacity ..................... 1.11328125 + 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-06 + normalization ................................... RMSNorm + num_attention_heads ............................. 32 + num_channels .................................... 3 + num_classes ..................................... 1000 + num_dataset_builder_threads ..................... 1 + num_distributed_optimizer_instances ............. 1 + num_experts ..................................... 128 + num_layers ...................................... 48 + num_layers_at_end_in_bf16 ....................... 1 + num_layers_at_start_in_bf16 ..................... 1 + num_layers_per_virtual_pipeline_stage ........... None + num_query_groups ................................ 4 + num_virtual_stages_per_pipeline_rank ............ None + num_workers ..................................... 2 + object_storage_cache_path ....................... None + one_logger_async ................................ False + one_logger_project .............................. megatron-lm + one_logger_run_name ............................. None + onnx_safe ....................................... None + openai_gelu ..................................... False + optimizer ....................................... adam + optimizer_cpu_offload ........................... False + optimizer_offload_fraction ...................... 1.0 + output_bert_embeddings .......................... False + overlap_cpu_optimizer_d2h_h2d ................... False + overlap_grad_reduce ............................. True + overlap_moe_expert_parallel_comm ................ False + overlap_p2p_comm ................................ False + overlap_p2p_comm_warmup_flush ................... False + overlap_param_gather ............................ True + overlap_param_gather_with_optimizer_step ........ False + override_opt_param_scheduler .................... False + params_dtype .................................... torch.bfloat16 + patch_dim ....................................... 16 + per_split_data_args_path ........................ None + perform_initialization .......................... True + pin_cpu_grads ................................... True + pin_cpu_params .................................. True + pipeline_model_parallel_comm_backend ............ None + pipeline_model_parallel_layout .................. None + pipeline_model_parallel_size .................... 1 + position_embedding_type ......................... rope + pretrained_checkpoint ........................... None + profile ......................................... False + profile_ranks ................................... [0] + profile_step_end ................................ 12 + profile_step_start .............................. 10 + q_lora_rank ..................................... None + qk_head_dim ..................................... 128 + qk_l2_norm ...................................... False + qk_layernorm .................................... True + qk_pos_emb_head_dim ............................. 64 + query_in_block_prob ............................. 0.1 + rampup_batch_size ............................... None + rank ............................................ 0 + recompute_granularity ........................... None + recompute_method ................................ None + recompute_modules ............................... None + recompute_num_layers ............................ None + record_memory_history ........................... False + relative_attention_max_distance ................. 128 + relative_attention_num_buckets .................. 32 + replication ..................................... False + replication_factor .............................. 2 + replication_jump ................................ None + rerun_mode ...................................... validate_results + reset_attention_mask ............................ False + reset_position_ids .............................. False + result_rejected_tracker_filename ................ None + retriever_report_topk_accuracies ................ [] + retriever_score_scaling ......................... False + retriever_seq_length ............................ 256 + retro_add_retriever ............................. False + retro_attention_gate ............................ 1 + retro_cyclic_train_iters ........................ None + retro_encoder_attention_dropout ................. 0.1 + retro_encoder_hidden_dropout .................... 0.1 + retro_encoder_layers ............................ 2 + retro_num_neighbors ............................. 2 + retro_num_retrieved_chunks ...................... 2 + retro_project_dir ............................... None + retro_verify_neighbor_count ..................... True + reuse_grad_buf_for_mxfp8_param_ag ............... False + rope_scaling_factor ............................. 8.0 + rope_type ....................................... None + rotary_base ..................................... 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 ..................................... 12 + train_samples ................................... None + train_sync_interval ............................. None + transformer_impl ................................ transformer_engine + transformer_pipeline_model_parallel_size ........ 1 + untie_embeddings_and_output_weights ............. True + use_checkpoint_args ............................. False + use_checkpoint_opt_param_scheduler .............. False + use_cpu_initialization .......................... None + use_custom_fsdp ................................. False + use_dist_ckpt ................................... True + use_dist_ckpt_deprecated ........................ False + use_distributed_optimizer ....................... True + use_flash_attn .................................. False + use_legacy_models ............................... False + use_mp_args_from_checkpoint_args ................ False + use_one_sent_docs ............................... False + use_persistent_ckpt_worker ...................... False + use_precision_aware_optimizer ................... False + use_pytorch_profiler ............................ False + use_ring_exchange_p2p ........................... False + use_rope_scaling ................................ False + use_rotary_position_embeddings .................. False + use_sharp ....................................... False + use_tokenizer_model_from_checkpoint_args ........ True + use_torch_fsdp2 ................................. False + use_torch_optimizer_for_cpu_offload ............. False + use_tp_pp_dp_mapping ............................ False + v_head_dim ...................................... 128 + valid_data_path ................................. None + variable_seq_lengths ............................ False + virtual_pipeline_model_parallel_size ............ None + vision_backbone_type ............................ vit + vision_pretraining .............................. False + vision_pretraining_type ......................... classify + vocab_extra_ids ................................. 0 + vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json + vocab_size ...................................... None + wandb_exp_name .................................. + wandb_project ................................... + wandb_save_dir .................................. + weight_decay .................................... 0.1 + weight_decay_incr_style ......................... constant + wgrad_deferral_limit ............................ 0 + world_size ...................................... 8 + yaml_cfg ........................................ None +-------------------- end of arguments --------------------- +INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1 +> building GPT2BPETokenizer tokenizer ... +/workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled + warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled") + > 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") +[Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank [Gloo] Rank [Gloo] Rank 10 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 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 [Gloo] Rank 6 is connected to 7 peer ranks. 7Expected number of connected peer ranks is : is connected to 77 + peer ranks. Expected number of connected peer ranks is : 7 +2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank [Gloo] Rank 0 is connected to 27 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 1 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 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7 +[Gloo] Rank [Gloo] Rank 67 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 +> initialized tensor model parallel with size 1 +> initialized pipeline model parallel with size 1 +> setting random seeds to 1234 ... +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[Gloo] Rank [Gloo] Rank 00 is connected to is connected to 00 peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : 00 + +[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0 +[rank3]:[W826 18:34:04.732281644 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. +[rank6]:[W826 18:34:04.732323370 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. +[rank1]:[W826 18:34:04.732480099 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. +[rank2]:[W826 18:34:04.732493964 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]:[W826 18:34:04.732758629 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device. +[rank5]:[W826 18:34:04.734574423 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. +[rank4]:[W826 18:34:04.740566038 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. +> 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: 0.035 seconds +WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations. +> compiling and loading fused kernels ... +[rank0]:[W826 18:34:04.888063339 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: 3.779 seconds +time to initialize megatron (seconds): 6.979 +[after megatron is initialized] datetime: 2026-08-26 18:34:10 +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): 8372566016 +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: 21 +Params for bucket 1 (103022592 elements, 103022592 padded size): + module.output_layer.weight +Params for bucket 2 (49033984 elements, 49033984 padded size): + module.decoder.layers.47.pre_mlp_layernorm.weight + module.decoder.layers.47.self_attention.q_layernorm.weight + module.decoder.layers.46.self_attention.linear_qkv.weight + module.decoder.layers.47.self_attention.linear_proj.weight + module.decoder.layers.47.self_attention.linear_qkv.weight + module.decoder.layers.46.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.45.mlp.router.weight + module.decoder.layers.45.self_attention.k_layernorm.weight + module.decoder.layers.45.pre_mlp_layernorm.weight + module.decoder.layers.45.self_attention.q_layernorm.weight + module.decoder.layers.47.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.46.mlp.router.weight + module.decoder.layers.46.self_attention.k_layernorm.weight + module.decoder.layers.46.pre_mlp_layernorm.weight + module.decoder.layers.46.self_attention.q_layernorm.weight + module.decoder.layers.45.self_attention.linear_qkv.weight + module.decoder.final_layernorm.weight + module.decoder.layers.47.self_attention.k_layernorm.weight + module.decoder.layers.47.mlp.router.weight + module.decoder.layers.46.self_attention.linear_proj.weight +Params for bucket 3 (46672384 elements, 46672384 padded size): + module.decoder.layers.44.pre_mlp_layernorm.weight + module.decoder.layers.44.self_attention.q_layernorm.weight + module.decoder.layers.44.self_attention.linear_qkv.weight + module.decoder.layers.43.pre_mlp_layernorm.weight + module.decoder.layers.43.self_attention.q_layernorm.weight + module.decoder.layers.43.self_attention.linear_qkv.weight + module.decoder.layers.45.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.44.self_attention.k_layernorm.weight + module.decoder.layers.45.self_attention.linear_proj.weight + module.decoder.layers.44.mlp.router.weight + module.decoder.layers.44.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.44.self_attention.linear_proj.weight + module.decoder.layers.43.mlp.router.weight + module.decoder.layers.43.self_attention.k_layernorm.weight + module.decoder.layers.43.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.43.self_attention.linear_proj.weight +Params for bucket 4 (49031936 elements, 49031936 padded size): + module.decoder.layers.42.self_attention.linear_qkv.weight + module.decoder.layers.41.pre_mlp_layernorm.weight + module.decoder.layers.40.self_attention.linear_qkv.weight + module.decoder.layers.42.mlp.router.weight + module.decoder.layers.42.pre_mlp_layernorm.weight + module.decoder.layers.41.self_attention.linear_qkv.weight + module.decoder.layers.42.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.41.self_attention.q_layernorm.weight + module.decoder.layers.42.self_attention.k_layernorm.weight + module.decoder.layers.41.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.41.self_attention.linear_proj.weight + module.decoder.layers.40.mlp.router.weight + module.decoder.layers.40.self_attention.k_layernorm.weight + module.decoder.layers.42.self_attention.q_layernorm.weight + module.decoder.layers.40.pre_mlp_layernorm.weight + module.decoder.layers.40.self_attention.q_layernorm.weight + module.decoder.layers.42.self_attention.linear_proj.weight + module.decoder.layers.41.mlp.router.weight + module.decoder.layers.41.self_attention.k_layernorm.weight +Params for bucket 5 (46672384 elements, 46672384 padded size): + module.decoder.layers.39.pre_mlp_layernorm.weight + module.decoder.layers.39.self_attention.q_layernorm.weight + module.decoder.layers.39.self_attention.linear_qkv.weight + module.decoder.layers.38.pre_mlp_layernorm.weight + module.decoder.layers.38.self_attention.q_layernorm.weight + module.decoder.layers.38.self_attention.linear_qkv.weight + module.decoder.layers.40.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.40.self_attention.linear_proj.weight + module.decoder.layers.39.mlp.router.weight + module.decoder.layers.39.self_attention.k_layernorm.weight + module.decoder.layers.39.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.39.self_attention.linear_proj.weight + module.decoder.layers.38.mlp.router.weight + module.decoder.layers.38.self_attention.k_layernorm.weight + module.decoder.layers.38.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.38.self_attention.linear_proj.weight +Params for bucket 6 (49031936 elements, 49031936 padded size): + module.decoder.layers.36.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.35.mlp.router.weight + module.decoder.layers.35.self_attention.k_layernorm.weight + module.decoder.layers.35.pre_mlp_layernorm.weight + module.decoder.layers.35.self_attention.q_layernorm.weight + module.decoder.layers.37.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.36.mlp.router.weight + module.decoder.layers.36.self_attention.k_layernorm.weight + module.decoder.layers.36.pre_mlp_layernorm.weight + module.decoder.layers.36.self_attention.q_layernorm.weight + module.decoder.layers.35.self_attention.linear_qkv.weight + module.decoder.layers.37.mlp.router.weight + module.decoder.layers.37.self_attention.k_layernorm.weight + module.decoder.layers.36.self_attention.linear_proj.weight + module.decoder.layers.37.pre_mlp_layernorm.weight + module.decoder.layers.37.self_attention.q_layernorm.weight + module.decoder.layers.36.self_attention.linear_qkv.weight + module.decoder.layers.37.self_attention.linear_proj.weight + module.decoder.layers.37.self_attention.linear_qkv.weight +Params for bucket 7 (46672384 elements, 46672384 padded size): + module.decoder.layers.35.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.34.mlp.router.weight + module.decoder.layers.35.self_attention.linear_proj.weight + module.decoder.layers.34.self_attention.k_layernorm.weight + module.decoder.layers.34.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.34.self_attention.linear_proj.weight + module.decoder.layers.33.mlp.router.weight + module.decoder.layers.33.self_attention.k_layernorm.weight + module.decoder.layers.33.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.33.self_attention.linear_proj.weight + module.decoder.layers.34.pre_mlp_layernorm.weight + module.decoder.layers.34.self_attention.q_layernorm.weight + module.decoder.layers.34.self_attention.linear_qkv.weight + module.decoder.layers.33.pre_mlp_layernorm.weight + module.decoder.layers.33.self_attention.q_layernorm.weight + module.decoder.layers.33.self_attention.linear_qkv.weight +Params for bucket 8 (49031936 elements, 49031936 padded size): + module.decoder.layers.32.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.31.self_attention.linear_qkv.weight + module.decoder.layers.30.mlp.router.weight + module.decoder.layers.30.self_attention.k_layernorm.weight + module.decoder.layers.30.pre_mlp_layernorm.weight + module.decoder.layers.30.self_attention.q_layernorm.weight + module.decoder.layers.32.mlp.router.weight + module.decoder.layers.32.self_attention.k_layernorm.weight + module.decoder.layers.32.pre_mlp_layernorm.weight + module.decoder.layers.32.self_attention.q_layernorm.weight + module.decoder.layers.31.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.30.self_attention.linear_qkv.weight + module.decoder.layers.32.self_attention.linear_proj.weight + module.decoder.layers.32.self_attention.linear_qkv.weight + module.decoder.layers.31.mlp.router.weight + module.decoder.layers.31.self_attention.k_layernorm.weight + module.decoder.layers.31.pre_mlp_layernorm.weight + module.decoder.layers.31.self_attention.q_layernorm.weight + module.decoder.layers.31.self_attention.linear_proj.weight +Params for bucket 9 (46672384 elements, 46672384 padded size): + module.decoder.layers.30.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.29.self_attention.k_layernorm.weight + module.decoder.layers.30.self_attention.linear_proj.weight + module.decoder.layers.29.mlp.router.weight + module.decoder.layers.29.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.29.self_attention.linear_proj.weight + module.decoder.layers.28.mlp.router.weight + module.decoder.layers.28.self_attention.k_layernorm.weight + module.decoder.layers.28.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.28.self_attention.linear_proj.weight + module.decoder.layers.29.pre_mlp_layernorm.weight + module.decoder.layers.29.self_attention.q_layernorm.weight + module.decoder.layers.29.self_attention.linear_qkv.weight + module.decoder.layers.28.pre_mlp_layernorm.weight + module.decoder.layers.28.self_attention.q_layernorm.weight + module.decoder.layers.28.self_attention.linear_qkv.weight +Params for bucket 10 (49031936 elements, 49031936 padded size): + module.decoder.layers.27.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.26.mlp.router.weight + module.decoder.layers.26.self_attention.k_layernorm.weight + module.decoder.layers.26.pre_mlp_layernorm.weight + module.decoder.layers.26.self_attention.q_layernorm.weight + module.decoder.layers.25.self_attention.linear_qkv.weight + module.decoder.layers.27.mlp.router.weight + module.decoder.layers.27.self_attention.k_layernorm.weight + module.decoder.layers.26.self_attention.linear_proj.weight + module.decoder.layers.27.self_attention.q_layernorm.weight + module.decoder.layers.27.pre_mlp_layernorm.weight + module.decoder.layers.26.self_attention.linear_qkv.weight + module.decoder.layers.27.self_attention.linear_proj.weight + module.decoder.layers.27.self_attention.linear_qkv.weight + module.decoder.layers.26.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.25.mlp.router.weight + module.decoder.layers.25.self_attention.k_layernorm.weight + module.decoder.layers.25.pre_mlp_layernorm.weight + module.decoder.layers.25.self_attention.q_layernorm.weight +Params for bucket 11 (46672384 elements, 46672384 padded size): + module.decoder.layers.25.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.25.self_attention.linear_proj.weight + module.decoder.layers.24.mlp.router.weight + module.decoder.layers.24.self_attention.k_layernorm.weight + module.decoder.layers.24.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.24.self_attention.linear_proj.weight + module.decoder.layers.23.mlp.router.weight + module.decoder.layers.23.self_attention.k_layernorm.weight + module.decoder.layers.23.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.23.self_attention.linear_proj.weight + module.decoder.layers.24.pre_mlp_layernorm.weight + module.decoder.layers.24.self_attention.q_layernorm.weight + module.decoder.layers.24.self_attention.linear_qkv.weight + module.decoder.layers.23.pre_mlp_layernorm.weight + module.decoder.layers.23.self_attention.q_layernorm.weight + module.decoder.layers.23.self_attention.linear_qkv.weight +Params for bucket 12 (49031936 elements, 49031936 padded size): + module.decoder.layers.22.mlp.router.weight + module.decoder.layers.22.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.21.mlp.router.weight + module.decoder.layers.21.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.20.mlp.router.weight + module.decoder.layers.20.self_attention.k_layernorm.weight + module.decoder.layers.22.self_attention.k_layernorm.weight + module.decoder.layers.20.pre_mlp_layernorm.weight + module.decoder.layers.20.self_attention.q_layernorm.weight + module.decoder.layers.22.self_attention.q_layernorm.weight + module.decoder.layers.22.pre_mlp_layernorm.weight + module.decoder.layers.21.self_attention.k_layernorm.weight + module.decoder.layers.22.self_attention.linear_proj.weight + module.decoder.layers.21.pre_mlp_layernorm.weight + module.decoder.layers.21.self_attention.q_layernorm.weight + module.decoder.layers.20.self_attention.linear_qkv.weight + module.decoder.layers.22.self_attention.linear_qkv.weight + module.decoder.layers.21.self_attention.linear_proj.weight + module.decoder.layers.21.self_attention.linear_qkv.weight +Params for bucket 13 (46672384 elements, 46672384 padded size): + module.decoder.layers.20.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.19.pre_mlp_layernorm.weight + module.decoder.layers.20.self_attention.linear_proj.weight + module.decoder.layers.19.self_attention.q_layernorm.weight + module.decoder.layers.19.self_attention.linear_qkv.weight + module.decoder.layers.18.pre_mlp_layernorm.weight + module.decoder.layers.18.self_attention.q_layernorm.weight + module.decoder.layers.18.self_attention.linear_qkv.weight + module.decoder.layers.19.mlp.router.weight + module.decoder.layers.19.self_attention.k_layernorm.weight + module.decoder.layers.19.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.19.self_attention.linear_proj.weight + module.decoder.layers.18.mlp.router.weight + module.decoder.layers.18.self_attention.k_layernorm.weight + module.decoder.layers.18.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.18.self_attention.linear_proj.weight +Params for bucket 14 (49031936 elements, 49031936 padded size): + module.decoder.layers.17.self_attention.linear_proj.weight + module.decoder.layers.17.self_attention.linear_qkv.weight + module.decoder.layers.16.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.15.mlp.router.weight + module.decoder.layers.15.self_attention.k_layernorm.weight + module.decoder.layers.15.pre_mlp_layernorm.weight + module.decoder.layers.15.self_attention.q_layernorm.weight + module.decoder.layers.17.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.16.mlp.router.weight + module.decoder.layers.16.self_attention.k_layernorm.weight + module.decoder.layers.16.pre_mlp_layernorm.weight + module.decoder.layers.16.self_attention.q_layernorm.weight + module.decoder.layers.15.self_attention.linear_qkv.weight + module.decoder.layers.17.mlp.router.weight + module.decoder.layers.17.self_attention.k_layernorm.weight + module.decoder.layers.16.self_attention.linear_proj.weight + module.decoder.layers.17.pre_mlp_layernorm.weight + module.decoder.layers.17.self_attention.q_layernorm.weight + module.decoder.layers.16.self_attention.linear_qkv.weight +Params for bucket 15 (46672384 elements, 46672384 padded size): + module.decoder.layers.15.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.14.mlp.router.weight + module.decoder.layers.14.self_attention.k_layernorm.weight + module.decoder.layers.15.self_attention.linear_proj.weight + module.decoder.layers.14.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.14.self_attention.linear_proj.weight + module.decoder.layers.13.mlp.router.weight + module.decoder.layers.13.self_attention.k_layernorm.weight + module.decoder.layers.13.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.13.self_attention.linear_proj.weight + module.decoder.layers.14.pre_mlp_layernorm.weight + module.decoder.layers.14.self_attention.q_layernorm.weight + module.decoder.layers.14.self_attention.linear_qkv.weight + module.decoder.layers.13.pre_mlp_layernorm.weight + module.decoder.layers.13.self_attention.q_layernorm.weight + module.decoder.layers.13.self_attention.linear_qkv.weight +Params for bucket 16 (49031936 elements, 49031936 padded size): + module.decoder.layers.11.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.10.mlp.router.weight + module.decoder.layers.10.self_attention.k_layernorm.weight + module.decoder.layers.10.pre_mlp_layernorm.weight + module.decoder.layers.10.self_attention.q_layernorm.weight + module.decoder.layers.12.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.11.mlp.router.weight + module.decoder.layers.11.self_attention.k_layernorm.weight + module.decoder.layers.11.pre_mlp_layernorm.weight + module.decoder.layers.11.self_attention.q_layernorm.weight + module.decoder.layers.10.self_attention.linear_qkv.weight + module.decoder.layers.12.mlp.router.weight + module.decoder.layers.12.self_attention.k_layernorm.weight + module.decoder.layers.11.self_attention.linear_proj.weight + module.decoder.layers.12.self_attention.q_layernorm.weight + module.decoder.layers.12.pre_mlp_layernorm.weight + module.decoder.layers.11.self_attention.linear_qkv.weight + module.decoder.layers.12.self_attention.linear_proj.weight + module.decoder.layers.12.self_attention.linear_qkv.weight +Params for bucket 17 (46672384 elements, 46672384 padded size): + module.decoder.layers.10.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.10.self_attention.linear_proj.weight + module.decoder.layers.9.pre_mlp_layernorm.weight + module.decoder.layers.9.self_attention.q_layernorm.weight + module.decoder.layers.9.self_attention.linear_qkv.weight + module.decoder.layers.8.pre_mlp_layernorm.weight + module.decoder.layers.8.self_attention.q_layernorm.weight + module.decoder.layers.8.self_attention.linear_qkv.weight + module.decoder.layers.9.mlp.router.weight + module.decoder.layers.9.self_attention.k_layernorm.weight + module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.9.self_attention.linear_proj.weight + module.decoder.layers.8.mlp.router.weight + module.decoder.layers.8.self_attention.k_layernorm.weight + module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.8.self_attention.linear_proj.weight +Params for bucket 18 (49031936 elements, 49031936 padded size): + module.decoder.layers.7.mlp.router.weight + module.decoder.layers.7.self_attention.k_layernorm.weight + module.decoder.layers.6.self_attention.linear_proj.weight + module.decoder.layers.7.pre_mlp_layernorm.weight + module.decoder.layers.7.self_attention.q_layernorm.weight + module.decoder.layers.6.self_attention.linear_qkv.weight + module.decoder.layers.7.self_attention.linear_proj.weight + module.decoder.layers.7.self_attention.linear_qkv.weight + module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.5.mlp.router.weight + module.decoder.layers.5.self_attention.k_layernorm.weight + module.decoder.layers.5.pre_mlp_layernorm.weight + module.decoder.layers.5.self_attention.q_layernorm.weight + module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.6.mlp.router.weight + module.decoder.layers.6.self_attention.k_layernorm.weight + module.decoder.layers.6.pre_mlp_layernorm.weight + module.decoder.layers.6.self_attention.q_layernorm.weight + module.decoder.layers.5.self_attention.linear_qkv.weight +Params for bucket 19 (46672384 elements, 46672384 padded size): + module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.4.self_attention.k_layernorm.weight + 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.3.mlp.router.weight + module.decoder.layers.3.self_attention.k_layernorm.weight + module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.3.self_attention.linear_proj.weight + module.decoder.layers.5.self_attention.linear_proj.weight + module.decoder.layers.4.pre_mlp_layernorm.weight + module.decoder.layers.4.self_attention.q_layernorm.weight + module.decoder.layers.4.self_attention.linear_qkv.weight + module.decoder.layers.3.pre_mlp_layernorm.weight + module.decoder.layers.3.self_attention.q_layernorm.weight + module.decoder.layers.3.self_attention.linear_qkv.weight +Params for bucket 20 (49031936 elements, 49031936 padded size): + module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight + module.decoder.layers.1.self_attention.linear_qkv.weight + module.decoder.layers.0.mlp.router.weight + module.decoder.layers.0.self_attention.linear_qkv.weight + module.decoder.layers.0.pre_mlp_layernorm.weight + module.decoder.layers.2.mlp.router.weight + module.decoder.layers.2.self_attention.k_layernorm.weight + module.decoder.layers.2.pre_mlp_layernorm.weight + module.decoder.layers.2.self_attention.q_layernorm.weight + module.decoder.layers.1.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.1.mlp.router.weight + module.decoder.layers.1.self_attention.k_layernorm.weight + module.decoder.layers.0.self_attention.q_layernorm.weight + module.decoder.layers.1.pre_mlp_layernorm.weight + module.decoder.layers.1.self_attention.q_layernorm.weight + module.decoder.layers.1.self_attention.linear_proj.weight + module.decoder.layers.0.self_attention.k_layernorm.weight +Params for bucket 21 (111413248 elements, 111413248 padded size): + module.decoder.layers.0.self_attention.linear_qkv.layer_norm_weight + module.embedding.word_embeddings.weight + module.decoder.layers.0.self_attention.linear_proj.weight +INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 89 +Params for bucket 1 (40894464 elements, 40894464 padded size): + module.decoder.layers.47.mlp.experts.linear_fc2.weight10 + module.decoder.layers.47.mlp.experts.linear_fc2.weight2 + module.decoder.layers.47.mlp.experts.linear_fc1.weight13 + module.decoder.layers.47.mlp.experts.linear_fc2.weight15 + module.decoder.layers.47.mlp.experts.linear_fc2.weight7 + module.decoder.layers.47.mlp.experts.linear_fc2.weight12 + module.decoder.layers.47.mlp.experts.linear_fc2.weight4 + module.decoder.layers.47.mlp.experts.linear_fc1.weight15 + module.decoder.layers.47.mlp.experts.linear_fc2.weight9 + module.decoder.layers.47.mlp.experts.linear_fc2.weight1 + module.decoder.layers.47.mlp.experts.linear_fc1.weight12 + module.decoder.layers.47.mlp.experts.linear_fc2.weight14 + module.decoder.layers.47.mlp.experts.linear_fc2.weight6 + module.decoder.layers.47.mlp.experts.linear_fc2.weight11 + module.decoder.layers.47.mlp.experts.linear_fc2.weight3 + module.decoder.layers.47.mlp.experts.linear_fc1.weight14 + module.decoder.layers.47.mlp.experts.linear_fc2.weight8 + module.decoder.layers.47.mlp.experts.linear_fc2.weight0 + module.decoder.layers.47.mlp.experts.linear_fc1.weight11 + module.decoder.layers.47.mlp.experts.linear_fc2.weight13 + module.decoder.layers.47.mlp.experts.linear_fc2.weight5 +Params for bucket 2 (40894464 elements, 40894464 padded size): + module.decoder.layers.47.mlp.experts.linear_fc1.weight9 + module.decoder.layers.47.mlp.experts.linear_fc1.weight7 + module.decoder.layers.47.mlp.experts.linear_fc1.weight5 + module.decoder.layers.47.mlp.experts.linear_fc1.weight3 + module.decoder.layers.47.mlp.experts.linear_fc1.weight1 + module.decoder.layers.46.mlp.experts.linear_fc2.weight14 + module.decoder.layers.46.mlp.experts.linear_fc2.weight12 + module.decoder.layers.47.mlp.experts.linear_fc1.weight10 + module.decoder.layers.47.mlp.experts.linear_fc1.weight6 + module.decoder.layers.47.mlp.experts.linear_fc1.weight8 + module.decoder.layers.47.mlp.experts.linear_fc1.weight4 + module.decoder.layers.47.mlp.experts.linear_fc1.weight2 + module.decoder.layers.47.mlp.experts.linear_fc1.weight0 + module.decoder.layers.46.mlp.experts.linear_fc2.weight15 + module.decoder.layers.46.mlp.experts.linear_fc2.weight13 +Params for bucket 3 (40894464 elements, 40894464 padded size): + module.decoder.layers.46.mlp.experts.linear_fc2.weight4 + module.decoder.layers.46.mlp.experts.linear_fc1.weight15 + module.decoder.layers.46.mlp.experts.linear_fc2.weight9 + module.decoder.layers.46.mlp.experts.linear_fc2.weight1 + module.decoder.layers.46.mlp.experts.linear_fc1.weight12 + module.decoder.layers.46.mlp.experts.linear_fc2.weight6 + module.decoder.layers.46.mlp.experts.linear_fc1.weight9 + module.decoder.layers.46.mlp.experts.linear_fc2.weight11 + module.decoder.layers.46.mlp.experts.linear_fc2.weight3 + module.decoder.layers.46.mlp.experts.linear_fc1.weight14 + module.decoder.layers.46.mlp.experts.linear_fc2.weight8 + module.decoder.layers.46.mlp.experts.linear_fc2.weight0 + module.decoder.layers.46.mlp.experts.linear_fc1.weight11 + module.decoder.layers.46.mlp.experts.linear_fc2.weight5 + module.decoder.layers.46.mlp.experts.linear_fc2.weight10 + module.decoder.layers.46.mlp.experts.linear_fc2.weight2 + module.decoder.layers.46.mlp.experts.linear_fc1.weight13 + module.decoder.layers.46.mlp.experts.linear_fc2.weight7 + module.decoder.layers.46.mlp.experts.linear_fc1.weight10 +Params for bucket 4 (40894464 elements, 40894464 padded size): + module.decoder.layers.46.mlp.experts.linear_fc1.weight7 + module.decoder.layers.46.mlp.experts.linear_fc1.weight5 + module.decoder.layers.46.mlp.experts.linear_fc1.weight3 + module.decoder.layers.46.mlp.experts.linear_fc1.weight1 + module.decoder.layers.45.mlp.experts.linear_fc2.weight14 + module.decoder.layers.45.mlp.experts.linear_fc2.weight12 + module.decoder.layers.45.mlp.experts.linear_fc2.weight10 + module.decoder.layers.45.mlp.experts.linear_fc2.weight8 + module.decoder.layers.46.mlp.experts.linear_fc1.weight8 + module.decoder.layers.46.mlp.experts.linear_fc1.weight6 + module.decoder.layers.46.mlp.experts.linear_fc1.weight4 + module.decoder.layers.46.mlp.experts.linear_fc1.weight2 + module.decoder.layers.46.mlp.experts.linear_fc1.weight0 + module.decoder.layers.45.mlp.experts.linear_fc2.weight15 + module.decoder.layers.45.mlp.experts.linear_fc2.weight13 + module.decoder.layers.45.mlp.experts.linear_fc2.weight11 + module.decoder.layers.45.mlp.experts.linear_fc2.weight9 +Params for bucket 5 (40894464 elements, 40894464 padded size): + module.decoder.layers.45.mlp.experts.linear_fc2.weight6 + module.decoder.layers.45.mlp.experts.linear_fc2.weight4 + module.decoder.layers.45.mlp.experts.linear_fc2.weight2 + module.decoder.layers.45.mlp.experts.linear_fc2.weight0 + module.decoder.layers.45.mlp.experts.linear_fc1.weight15 + module.decoder.layers.45.mlp.experts.linear_fc1.weight13 + module.decoder.layers.45.mlp.experts.linear_fc1.weight11 + module.decoder.layers.45.mlp.experts.linear_fc1.weight9 + module.decoder.layers.45.mlp.experts.linear_fc1.weight7 + module.decoder.layers.45.mlp.experts.linear_fc2.weight7 + module.decoder.layers.45.mlp.experts.linear_fc2.weight5 + module.decoder.layers.45.mlp.experts.linear_fc2.weight3 + module.decoder.layers.45.mlp.experts.linear_fc2.weight1 + module.decoder.layers.45.mlp.experts.linear_fc1.weight14 + module.decoder.layers.45.mlp.experts.linear_fc1.weight12 + module.decoder.layers.45.mlp.experts.linear_fc1.weight10 + module.decoder.layers.45.mlp.experts.linear_fc1.weight8 +Params for bucket 6 (40894464 elements, 40894464 padded size): + module.decoder.layers.45.mlp.experts.linear_fc1.weight1 + module.decoder.layers.44.mlp.experts.linear_fc2.weight8 + module.decoder.layers.45.mlp.experts.linear_fc1.weight6 + module.decoder.layers.44.mlp.experts.linear_fc2.weight13 + module.decoder.layers.44.mlp.experts.linear_fc2.weight5 + module.decoder.layers.45.mlp.experts.linear_fc1.weight3 + module.decoder.layers.44.mlp.experts.linear_fc2.weight10 + module.decoder.layers.45.mlp.experts.linear_fc1.weight0 + module.decoder.layers.44.mlp.experts.linear_fc2.weight15 + module.decoder.layers.44.mlp.experts.linear_fc2.weight7 + module.decoder.layers.45.mlp.experts.linear_fc1.weight5 + module.decoder.layers.44.mlp.experts.linear_fc2.weight12 + module.decoder.layers.44.mlp.experts.linear_fc2.weight4 + module.decoder.layers.45.mlp.experts.linear_fc1.weight2 + module.decoder.layers.44.mlp.experts.linear_fc2.weight9 + module.decoder.layers.44.mlp.experts.linear_fc2.weight14 + module.decoder.layers.44.mlp.experts.linear_fc2.weight6 + module.decoder.layers.45.mlp.experts.linear_fc1.weight4 + module.decoder.layers.44.mlp.experts.linear_fc2.weight11 +Params for bucket 7 (40894464 elements, 40894464 padded size): + module.decoder.layers.44.mlp.experts.linear_fc2.weight0 + module.decoder.layers.44.mlp.experts.linear_fc2.weight2 + module.decoder.layers.44.mlp.experts.linear_fc1.weight15 + module.decoder.layers.44.mlp.experts.linear_fc1.weight13 + module.decoder.layers.44.mlp.experts.linear_fc1.weight11 + module.decoder.layers.44.mlp.experts.linear_fc1.weight9 + module.decoder.layers.44.mlp.experts.linear_fc1.weight7 + module.decoder.layers.44.mlp.experts.linear_fc1.weight5 + module.decoder.layers.44.mlp.experts.linear_fc2.weight3 + module.decoder.layers.44.mlp.experts.linear_fc2.weight1 + module.decoder.layers.44.mlp.experts.linear_fc1.weight14 + module.decoder.layers.44.mlp.experts.linear_fc1.weight12 + module.decoder.layers.44.mlp.experts.linear_fc1.weight10 + module.decoder.layers.44.mlp.experts.linear_fc1.weight8 + module.decoder.layers.44.mlp.experts.linear_fc1.weight6 +Params for bucket 8 (40894464 elements, 40894464 padded size): + module.decoder.layers.44.mlp.experts.linear_fc1.weight3 + module.decoder.layers.43.mlp.experts.linear_fc2.weight10 + module.decoder.layers.43.mlp.experts.linear_fc2.weight2 + module.decoder.layers.44.mlp.experts.linear_fc1.weight0 + module.decoder.layers.43.mlp.experts.linear_fc2.weight15 + module.decoder.layers.43.mlp.experts.linear_fc2.weight7 + module.decoder.layers.43.mlp.experts.linear_fc2.weight12 + module.decoder.layers.43.mlp.experts.linear_fc2.weight4 + module.decoder.layers.44.mlp.experts.linear_fc1.weight2 + module.decoder.layers.43.mlp.experts.linear_fc2.weight9 + module.decoder.layers.43.mlp.experts.linear_fc2.weight1 + module.decoder.layers.43.mlp.experts.linear_fc2.weight14 + module.decoder.layers.43.mlp.experts.linear_fc2.weight6 + module.decoder.layers.44.mlp.experts.linear_fc1.weight4 + module.decoder.layers.43.mlp.experts.linear_fc2.weight11 + module.decoder.layers.43.mlp.experts.linear_fc2.weight3 + module.decoder.layers.44.mlp.experts.linear_fc1.weight1 + module.decoder.layers.43.mlp.experts.linear_fc2.weight8 + module.decoder.layers.43.mlp.experts.linear_fc2.weight0 + module.decoder.layers.43.mlp.experts.linear_fc2.weight13 + module.decoder.layers.43.mlp.experts.linear_fc2.weight5 +Params for bucket 9 (40894464 elements, 40894464 padded size): + module.decoder.layers.43.mlp.experts.linear_fc1.weight15 + module.decoder.layers.43.mlp.experts.linear_fc1.weight13 + module.decoder.layers.43.mlp.experts.linear_fc1.weight11 + module.decoder.layers.43.mlp.experts.linear_fc1.weight9 + module.decoder.layers.43.mlp.experts.linear_fc1.weight7 + module.decoder.layers.43.mlp.experts.linear_fc1.weight5 + module.decoder.layers.43.mlp.experts.linear_fc1.weight3 + module.decoder.layers.43.mlp.experts.linear_fc1.weight14 + module.decoder.layers.43.mlp.experts.linear_fc1.weight12 + module.decoder.layers.43.mlp.experts.linear_fc1.weight10 + module.decoder.layers.43.mlp.experts.linear_fc1.weight8 + module.decoder.layers.43.mlp.experts.linear_fc1.weight6 + module.decoder.layers.43.mlp.experts.linear_fc1.weight4 +Params for bucket 10 (40894464 elements, 40894464 padded size): + module.decoder.layers.42.mlp.experts.linear_fc2.weight12 + module.decoder.layers.42.mlp.experts.linear_fc2.weight4 + module.decoder.layers.42.mlp.experts.linear_fc1.weight15 + module.decoder.layers.43.mlp.experts.linear_fc1.weight2 + module.decoder.layers.42.mlp.experts.linear_fc2.weight9 + module.decoder.layers.42.mlp.experts.linear_fc2.weight1 + module.decoder.layers.42.mlp.experts.linear_fc2.weight14 + module.decoder.layers.42.mlp.experts.linear_fc2.weight6 + module.decoder.layers.42.mlp.experts.linear_fc2.weight11 + module.decoder.layers.42.mlp.experts.linear_fc2.weight3 + module.decoder.layers.42.mlp.experts.linear_fc1.weight14 + module.decoder.layers.43.mlp.experts.linear_fc1.weight1 + module.decoder.layers.42.mlp.experts.linear_fc2.weight8 + module.decoder.layers.42.mlp.experts.linear_fc2.weight0 + module.decoder.layers.42.mlp.experts.linear_fc2.weight13 + module.decoder.layers.42.mlp.experts.linear_fc2.weight5 + module.decoder.layers.42.mlp.experts.linear_fc2.weight10 + module.decoder.layers.42.mlp.experts.linear_fc2.weight2 + module.decoder.layers.42.mlp.experts.linear_fc2.weight15 + module.decoder.layers.43.mlp.experts.linear_fc1.weight0 + module.decoder.layers.42.mlp.experts.linear_fc2.weight7 +Params for bucket 11 (40894464 elements, 40894464 padded size): + module.decoder.layers.42.mlp.experts.linear_fc1.weight13 + module.decoder.layers.42.mlp.experts.linear_fc1.weight11 + module.decoder.layers.42.mlp.experts.linear_fc1.weight9 + module.decoder.layers.42.mlp.experts.linear_fc1.weight7 + module.decoder.layers.42.mlp.experts.linear_fc1.weight5 + module.decoder.layers.42.mlp.experts.linear_fc1.weight3 + module.decoder.layers.42.mlp.experts.linear_fc1.weight1 + module.decoder.layers.42.mlp.experts.linear_fc1.weight12 + module.decoder.layers.42.mlp.experts.linear_fc1.weight10 + module.decoder.layers.42.mlp.experts.linear_fc1.weight8 + module.decoder.layers.42.mlp.experts.linear_fc1.weight6 + module.decoder.layers.42.mlp.experts.linear_fc1.weight4 + module.decoder.layers.42.mlp.experts.linear_fc1.weight2 +Params for bucket 12 (40894464 elements, 40894464 padded size): + module.decoder.layers.41.mlp.experts.linear_fc2.weight14 + module.decoder.layers.41.mlp.experts.linear_fc2.weight6 + module.decoder.layers.41.mlp.experts.linear_fc2.weight11 + module.decoder.layers.41.mlp.experts.linear_fc2.weight3 + module.decoder.layers.41.mlp.experts.linear_fc1.weight14 + module.decoder.layers.41.mlp.experts.linear_fc2.weight8 + module.decoder.layers.41.mlp.experts.linear_fc2.weight0 + module.decoder.layers.41.mlp.experts.linear_fc2.weight13 + module.decoder.layers.41.mlp.experts.linear_fc2.weight5 + module.decoder.layers.41.mlp.experts.linear_fc2.weight10 + module.decoder.layers.41.mlp.experts.linear_fc2.weight2 + module.decoder.layers.41.mlp.experts.linear_fc1.weight13 + module.decoder.layers.42.mlp.experts.linear_fc1.weight0 + module.decoder.layers.41.mlp.experts.linear_fc2.weight15 + module.decoder.layers.41.mlp.experts.linear_fc2.weight7 + module.decoder.layers.41.mlp.experts.linear_fc2.weight12 + module.decoder.layers.41.mlp.experts.linear_fc2.weight4 + module.decoder.layers.41.mlp.experts.linear_fc1.weight15 + module.decoder.layers.41.mlp.experts.linear_fc2.weight9 + module.decoder.layers.41.mlp.experts.linear_fc2.weight1 + module.decoder.layers.41.mlp.experts.linear_fc1.weight12 +Params for bucket 13 (40894464 elements, 40894464 padded size): + module.decoder.layers.41.mlp.experts.linear_fc1.weight11 + module.decoder.layers.41.mlp.experts.linear_fc1.weight7 + module.decoder.layers.41.mlp.experts.linear_fc1.weight9 + module.decoder.layers.41.mlp.experts.linear_fc1.weight5 + module.decoder.layers.41.mlp.experts.linear_fc1.weight3 + module.decoder.layers.41.mlp.experts.linear_fc1.weight1 + module.decoder.layers.40.mlp.experts.linear_fc2.weight14 + module.decoder.layers.41.mlp.experts.linear_fc1.weight8 + module.decoder.layers.41.mlp.experts.linear_fc1.weight10 + module.decoder.layers.41.mlp.experts.linear_fc1.weight6 + module.decoder.layers.41.mlp.experts.linear_fc1.weight4 + module.decoder.layers.41.mlp.experts.linear_fc1.weight2 + module.decoder.layers.41.mlp.experts.linear_fc1.weight0 + module.decoder.layers.40.mlp.experts.linear_fc2.weight15 +Params for bucket 14 (40894464 elements, 40894464 padded size): + module.decoder.layers.40.mlp.experts.linear_fc2.weight8 + module.decoder.layers.40.mlp.experts.linear_fc2.weight0 + module.decoder.layers.40.mlp.experts.linear_fc1.weight11 + module.decoder.layers.40.mlp.experts.linear_fc2.weight13 + module.decoder.layers.40.mlp.experts.linear_fc2.weight5 + module.decoder.layers.40.mlp.experts.linear_fc2.weight10 + module.decoder.layers.40.mlp.experts.linear_fc2.weight2 + module.decoder.layers.40.mlp.experts.linear_fc1.weight13 + module.decoder.layers.40.mlp.experts.linear_fc2.weight7 + module.decoder.layers.40.mlp.experts.linear_fc1.weight10 + module.decoder.layers.40.mlp.experts.linear_fc2.weight12 + module.decoder.layers.40.mlp.experts.linear_fc2.weight4 + module.decoder.layers.40.mlp.experts.linear_fc1.weight15 + module.decoder.layers.40.mlp.experts.linear_fc2.weight9 + module.decoder.layers.40.mlp.experts.linear_fc2.weight1 + module.decoder.layers.40.mlp.experts.linear_fc1.weight12 + module.decoder.layers.40.mlp.experts.linear_fc2.weight6 + module.decoder.layers.40.mlp.experts.linear_fc2.weight11 + module.decoder.layers.40.mlp.experts.linear_fc2.weight3 + module.decoder.layers.40.mlp.experts.linear_fc1.weight14 +Params for bucket 15 (40894464 elements, 40894464 padded size): + module.decoder.layers.40.mlp.experts.linear_fc1.weight9 + module.decoder.layers.40.mlp.experts.linear_fc1.weight5 + module.decoder.layers.40.mlp.experts.linear_fc1.weight7 + module.decoder.layers.40.mlp.experts.linear_fc1.weight3 + module.decoder.layers.40.mlp.experts.linear_fc1.weight1 + module.decoder.layers.39.mlp.experts.linear_fc2.weight14 + module.decoder.layers.39.mlp.experts.linear_fc2.weight12 + module.decoder.layers.39.mlp.experts.linear_fc2.weight10 + module.decoder.layers.40.mlp.experts.linear_fc1.weight8 + module.decoder.layers.40.mlp.experts.linear_fc1.weight6 + module.decoder.layers.40.mlp.experts.linear_fc1.weight4 + module.decoder.layers.40.mlp.experts.linear_fc1.weight2 + module.decoder.layers.40.mlp.experts.linear_fc1.weight0 + module.decoder.layers.39.mlp.experts.linear_fc2.weight15 + module.decoder.layers.39.mlp.experts.linear_fc2.weight13 + module.decoder.layers.39.mlp.experts.linear_fc2.weight11 +Params for bucket 16 (40894464 elements, 40894464 padded size): + module.decoder.layers.39.mlp.experts.linear_fc2.weight6 + module.decoder.layers.39.mlp.experts.linear_fc2.weight8 + module.decoder.layers.39.mlp.experts.linear_fc2.weight4 + module.decoder.layers.39.mlp.experts.linear_fc2.weight2 + module.decoder.layers.39.mlp.experts.linear_fc2.weight0 + module.decoder.layers.39.mlp.experts.linear_fc1.weight15 + module.decoder.layers.39.mlp.experts.linear_fc1.weight13 + module.decoder.layers.39.mlp.experts.linear_fc1.weight11 + module.decoder.layers.39.mlp.experts.linear_fc1.weight9 + module.decoder.layers.39.mlp.experts.linear_fc2.weight9 + module.decoder.layers.39.mlp.experts.linear_fc2.weight5 + module.decoder.layers.39.mlp.experts.linear_fc2.weight7 + module.decoder.layers.39.mlp.experts.linear_fc2.weight3 + module.decoder.layers.39.mlp.experts.linear_fc2.weight1 + module.decoder.layers.39.mlp.experts.linear_fc1.weight14 + module.decoder.layers.39.mlp.experts.linear_fc1.weight12 + module.decoder.layers.39.mlp.experts.linear_fc1.weight10 + module.decoder.layers.39.mlp.experts.linear_fc1.weight8 +Params for bucket 17 (40894464 elements, 40894464 padded size): + module.decoder.layers.39.mlp.experts.linear_fc1.weight7 + module.decoder.layers.39.mlp.experts.linear_fc1.weight5 + module.decoder.layers.39.mlp.experts.linear_fc1.weight3 + module.decoder.layers.39.mlp.experts.linear_fc1.weight1 + module.decoder.layers.38.mlp.experts.linear_fc2.weight14 + module.decoder.layers.38.mlp.experts.linear_fc2.weight12 + module.decoder.layers.38.mlp.experts.linear_fc2.weight10 + module.decoder.layers.38.mlp.experts.linear_fc2.weight8 + module.decoder.layers.38.mlp.experts.linear_fc2.weight6 + module.decoder.layers.39.mlp.experts.linear_fc1.weight6 + module.decoder.layers.39.mlp.experts.linear_fc1.weight4 + module.decoder.layers.39.mlp.experts.linear_fc1.weight2 + module.decoder.layers.39.mlp.experts.linear_fc1.weight0 + module.decoder.layers.38.mlp.experts.linear_fc2.weight15 + module.decoder.layers.38.mlp.experts.linear_fc2.weight13 + module.decoder.layers.38.mlp.experts.linear_fc2.weight11 + module.decoder.layers.38.mlp.experts.linear_fc2.weight9 + module.decoder.layers.38.mlp.experts.linear_fc2.weight7 +Params for bucket 18 (40894464 elements, 40894464 padded size): + module.decoder.layers.38.mlp.experts.linear_fc2.weight2 + module.decoder.layers.38.mlp.experts.linear_fc2.weight4 + module.decoder.layers.38.mlp.experts.linear_fc2.weight0 + module.decoder.layers.38.mlp.experts.linear_fc1.weight15 + module.decoder.layers.38.mlp.experts.linear_fc1.weight13 + module.decoder.layers.38.mlp.experts.linear_fc1.weight11 + module.decoder.layers.38.mlp.experts.linear_fc1.weight9 + module.decoder.layers.38.mlp.experts.linear_fc1.weight7 + module.decoder.layers.38.mlp.experts.linear_fc2.weight5 + module.decoder.layers.38.mlp.experts.linear_fc2.weight3 + module.decoder.layers.38.mlp.experts.linear_fc2.weight1 + module.decoder.layers.38.mlp.experts.linear_fc1.weight14 + module.decoder.layers.38.mlp.experts.linear_fc1.weight12 + module.decoder.layers.38.mlp.experts.linear_fc1.weight10 + module.decoder.layers.38.mlp.experts.linear_fc1.weight8 + module.decoder.layers.38.mlp.experts.linear_fc1.weight6 +Params for bucket 19 (40894464 elements, 40894464 padded size): + module.decoder.layers.37.mlp.experts.linear_fc2.weight13 + module.decoder.layers.37.mlp.experts.linear_fc2.weight5 + module.decoder.layers.38.mlp.experts.linear_fc1.weight4 + module.decoder.layers.37.mlp.experts.linear_fc2.weight10 + module.decoder.layers.37.mlp.experts.linear_fc2.weight2 + module.decoder.layers.38.mlp.experts.linear_fc1.weight1 + module.decoder.layers.37.mlp.experts.linear_fc2.weight15 + module.decoder.layers.37.mlp.experts.linear_fc2.weight7 + module.decoder.layers.37.mlp.experts.linear_fc2.weight12 + module.decoder.layers.37.mlp.experts.linear_fc2.weight4 + module.decoder.layers.38.mlp.experts.linear_fc1.weight3 + module.decoder.layers.37.mlp.experts.linear_fc2.weight9 + module.decoder.layers.38.mlp.experts.linear_fc1.weight0 + module.decoder.layers.37.mlp.experts.linear_fc2.weight14 + module.decoder.layers.37.mlp.experts.linear_fc2.weight6 + module.decoder.layers.38.mlp.experts.linear_fc1.weight5 + module.decoder.layers.37.mlp.experts.linear_fc2.weight11 + module.decoder.layers.37.mlp.experts.linear_fc2.weight3 + module.decoder.layers.38.mlp.experts.linear_fc1.weight2 + module.decoder.layers.37.mlp.experts.linear_fc2.weight8 +Params for bucket 20 (40894464 elements, 40894464 padded size): + module.decoder.layers.37.mlp.experts.linear_fc2.weight1 + module.decoder.layers.37.mlp.experts.linear_fc1.weight14 + module.decoder.layers.37.mlp.experts.linear_fc1.weight12 + module.decoder.layers.37.mlp.experts.linear_fc1.weight10 + module.decoder.layers.37.mlp.experts.linear_fc1.weight8 + module.decoder.layers.37.mlp.experts.linear_fc1.weight6 + module.decoder.layers.37.mlp.experts.linear_fc1.weight4 + module.decoder.layers.37.mlp.experts.linear_fc1.weight15 + module.decoder.layers.37.mlp.experts.linear_fc2.weight0 + module.decoder.layers.37.mlp.experts.linear_fc1.weight13 + module.decoder.layers.37.mlp.experts.linear_fc1.weight11 + module.decoder.layers.37.mlp.experts.linear_fc1.weight9 + module.decoder.layers.37.mlp.experts.linear_fc1.weight7 + module.decoder.layers.37.mlp.experts.linear_fc1.weight5 +Params for bucket 21 (40894464 elements, 40894464 padded size): + module.decoder.layers.36.mlp.experts.linear_fc2.weight15 + module.decoder.layers.37.mlp.experts.linear_fc1.weight0 + module.decoder.layers.36.mlp.experts.linear_fc2.weight7 + module.decoder.layers.36.mlp.experts.linear_fc2.weight12 + module.decoder.layers.36.mlp.experts.linear_fc2.weight4 + module.decoder.layers.36.mlp.experts.linear_fc1.weight15 + module.decoder.layers.37.mlp.experts.linear_fc1.weight2 + module.decoder.layers.36.mlp.experts.linear_fc2.weight9 + module.decoder.layers.36.mlp.experts.linear_fc2.weight1 + module.decoder.layers.36.mlp.experts.linear_fc2.weight14 + module.decoder.layers.36.mlp.experts.linear_fc2.weight6 + module.decoder.layers.36.mlp.experts.linear_fc2.weight11 + module.decoder.layers.36.mlp.experts.linear_fc2.weight3 + module.decoder.layers.37.mlp.experts.linear_fc1.weight1 + module.decoder.layers.36.mlp.experts.linear_fc2.weight8 + module.decoder.layers.36.mlp.experts.linear_fc2.weight0 + module.decoder.layers.36.mlp.experts.linear_fc2.weight13 + module.decoder.layers.36.mlp.experts.linear_fc2.weight5 + module.decoder.layers.37.mlp.experts.linear_fc1.weight3 + module.decoder.layers.36.mlp.experts.linear_fc2.weight10 + module.decoder.layers.36.mlp.experts.linear_fc2.weight2 +Params for bucket 22 (40894464 elements, 40894464 padded size): + module.decoder.layers.36.mlp.experts.linear_fc1.weight14 + module.decoder.layers.36.mlp.experts.linear_fc1.weight12 + module.decoder.layers.36.mlp.experts.linear_fc1.weight10 + module.decoder.layers.36.mlp.experts.linear_fc1.weight8 + module.decoder.layers.36.mlp.experts.linear_fc1.weight6 + module.decoder.layers.36.mlp.experts.linear_fc1.weight4 + module.decoder.layers.36.mlp.experts.linear_fc1.weight2 + module.decoder.layers.36.mlp.experts.linear_fc1.weight13 + module.decoder.layers.36.mlp.experts.linear_fc1.weight11 + module.decoder.layers.36.mlp.experts.linear_fc1.weight9 + module.decoder.layers.36.mlp.experts.linear_fc1.weight7 + module.decoder.layers.36.mlp.experts.linear_fc1.weight5 + module.decoder.layers.36.mlp.experts.linear_fc1.weight3 +Params for bucket 23 (40894464 elements, 40894464 padded size): + module.decoder.layers.35.mlp.experts.linear_fc2.weight9 + module.decoder.layers.35.mlp.experts.linear_fc2.weight1 + module.decoder.layers.35.mlp.experts.linear_fc2.weight14 + module.decoder.layers.35.mlp.experts.linear_fc2.weight6 + module.decoder.layers.35.mlp.experts.linear_fc2.weight11 + module.decoder.layers.35.mlp.experts.linear_fc2.weight3 + module.decoder.layers.35.mlp.experts.linear_fc1.weight14 + module.decoder.layers.36.mlp.experts.linear_fc1.weight1 + module.decoder.layers.35.mlp.experts.linear_fc2.weight8 + module.decoder.layers.35.mlp.experts.linear_fc2.weight0 + module.decoder.layers.35.mlp.experts.linear_fc2.weight13 + module.decoder.layers.35.mlp.experts.linear_fc2.weight5 + module.decoder.layers.35.mlp.experts.linear_fc2.weight10 + module.decoder.layers.35.mlp.experts.linear_fc2.weight2 + module.decoder.layers.35.mlp.experts.linear_fc1.weight13 + module.decoder.layers.35.mlp.experts.linear_fc2.weight15 + module.decoder.layers.36.mlp.experts.linear_fc1.weight0 + module.decoder.layers.35.mlp.experts.linear_fc2.weight7 + module.decoder.layers.35.mlp.experts.linear_fc2.weight12 + module.decoder.layers.35.mlp.experts.linear_fc2.weight4 + module.decoder.layers.35.mlp.experts.linear_fc1.weight15 +Params for bucket 24 (40894464 elements, 40894464 padded size): + module.decoder.layers.35.mlp.experts.linear_fc1.weight12 + module.decoder.layers.35.mlp.experts.linear_fc1.weight10 + module.decoder.layers.35.mlp.experts.linear_fc1.weight8 + module.decoder.layers.35.mlp.experts.linear_fc1.weight6 + module.decoder.layers.35.mlp.experts.linear_fc1.weight4 + module.decoder.layers.35.mlp.experts.linear_fc1.weight2 + module.decoder.layers.35.mlp.experts.linear_fc1.weight0 + module.decoder.layers.35.mlp.experts.linear_fc1.weight9 + module.decoder.layers.35.mlp.experts.linear_fc1.weight11 + module.decoder.layers.35.mlp.experts.linear_fc1.weight7 + module.decoder.layers.35.mlp.experts.linear_fc1.weight5 + module.decoder.layers.35.mlp.experts.linear_fc1.weight3 + module.decoder.layers.35.mlp.experts.linear_fc1.weight1 +Params for bucket 25 (40894464 elements, 40894464 padded size): + module.decoder.layers.34.mlp.experts.linear_fc2.weight11 + module.decoder.layers.34.mlp.experts.linear_fc2.weight3 + module.decoder.layers.34.mlp.experts.linear_fc1.weight14 + module.decoder.layers.34.mlp.experts.linear_fc2.weight8 + module.decoder.layers.34.mlp.experts.linear_fc2.weight0 + module.decoder.layers.34.mlp.experts.linear_fc1.weight11 + module.decoder.layers.34.mlp.experts.linear_fc2.weight13 + module.decoder.layers.34.mlp.experts.linear_fc2.weight5 + module.decoder.layers.34.mlp.experts.linear_fc2.weight10 + module.decoder.layers.34.mlp.experts.linear_fc2.weight2 + module.decoder.layers.34.mlp.experts.linear_fc1.weight13 + module.decoder.layers.34.mlp.experts.linear_fc2.weight15 + module.decoder.layers.34.mlp.experts.linear_fc2.weight7 + module.decoder.layers.34.mlp.experts.linear_fc2.weight12 + module.decoder.layers.34.mlp.experts.linear_fc2.weight4 + module.decoder.layers.34.mlp.experts.linear_fc1.weight15 + module.decoder.layers.34.mlp.experts.linear_fc2.weight9 + module.decoder.layers.34.mlp.experts.linear_fc2.weight1 + module.decoder.layers.34.mlp.experts.linear_fc1.weight12 + module.decoder.layers.34.mlp.experts.linear_fc2.weight14 + module.decoder.layers.34.mlp.experts.linear_fc2.weight6 +Params for bucket 26 (40894464 elements, 40894464 padded size): + module.decoder.layers.34.mlp.experts.linear_fc1.weight10 + module.decoder.layers.34.mlp.experts.linear_fc1.weight6 + module.decoder.layers.34.mlp.experts.linear_fc1.weight8 + module.decoder.layers.34.mlp.experts.linear_fc1.weight4 + module.decoder.layers.34.mlp.experts.linear_fc1.weight2 + module.decoder.layers.34.mlp.experts.linear_fc1.weight0 + module.decoder.layers.33.mlp.experts.linear_fc2.weight15 + module.decoder.layers.33.mlp.experts.linear_fc2.weight13 + module.decoder.layers.34.mlp.experts.linear_fc1.weight7 + module.decoder.layers.34.mlp.experts.linear_fc1.weight9 + module.decoder.layers.34.mlp.experts.linear_fc1.weight5 + module.decoder.layers.34.mlp.experts.linear_fc1.weight3 + module.decoder.layers.34.mlp.experts.linear_fc1.weight1 + module.decoder.layers.33.mlp.experts.linear_fc2.weight14 + module.decoder.layers.33.mlp.experts.linear_fc2.weight12 +Params for bucket 27 (40894464 elements, 40894464 padded size): + module.decoder.layers.33.mlp.experts.linear_fc2.weight5 + module.decoder.layers.33.mlp.experts.linear_fc2.weight2 + module.decoder.layers.33.mlp.experts.linear_fc1.weight13 + module.decoder.layers.33.mlp.experts.linear_fc2.weight11 + module.decoder.layers.33.mlp.experts.linear_fc2.weight7 + module.decoder.layers.33.mlp.experts.linear_fc1.weight10 + module.decoder.layers.33.mlp.experts.linear_fc2.weight4 + module.decoder.layers.33.mlp.experts.linear_fc1.weight15 + module.decoder.layers.33.mlp.experts.linear_fc2.weight9 + module.decoder.layers.33.mlp.experts.linear_fc2.weight1 + module.decoder.layers.33.mlp.experts.linear_fc1.weight12 + module.decoder.layers.33.mlp.experts.linear_fc2.weight10 + module.decoder.layers.33.mlp.experts.linear_fc2.weight6 + module.decoder.layers.33.mlp.experts.linear_fc1.weight9 + module.decoder.layers.33.mlp.experts.linear_fc2.weight3 + module.decoder.layers.33.mlp.experts.linear_fc1.weight14 + module.decoder.layers.33.mlp.experts.linear_fc2.weight8 + module.decoder.layers.33.mlp.experts.linear_fc2.weight0 + module.decoder.layers.33.mlp.experts.linear_fc1.weight11 +Params for bucket 28 (40894464 elements, 40894464 padded size): + module.decoder.layers.33.mlp.experts.linear_fc1.weight8 + module.decoder.layers.33.mlp.experts.linear_fc1.weight6 + module.decoder.layers.33.mlp.experts.linear_fc1.weight4 + module.decoder.layers.33.mlp.experts.linear_fc1.weight2 + module.decoder.layers.33.mlp.experts.linear_fc1.weight0 + module.decoder.layers.32.mlp.experts.linear_fc2.weight15 + module.decoder.layers.32.mlp.experts.linear_fc2.weight13 + module.decoder.layers.32.mlp.experts.linear_fc2.weight11 + module.decoder.layers.32.mlp.experts.linear_fc2.weight9 + module.decoder.layers.33.mlp.experts.linear_fc1.weight7 + module.decoder.layers.33.mlp.experts.linear_fc1.weight5 + module.decoder.layers.33.mlp.experts.linear_fc1.weight3 + module.decoder.layers.33.mlp.experts.linear_fc1.weight1 + module.decoder.layers.32.mlp.experts.linear_fc2.weight14 + module.decoder.layers.32.mlp.experts.linear_fc2.weight12 + module.decoder.layers.32.mlp.experts.linear_fc2.weight10 + module.decoder.layers.32.mlp.experts.linear_fc2.weight8 +Params for bucket 29 (40894464 elements, 40894464 padded size): + module.decoder.layers.32.mlp.experts.linear_fc2.weight7 + module.decoder.layers.32.mlp.experts.linear_fc2.weight5 + module.decoder.layers.32.mlp.experts.linear_fc2.weight3 + module.decoder.layers.32.mlp.experts.linear_fc2.weight1 + module.decoder.layers.32.mlp.experts.linear_fc1.weight14 + module.decoder.layers.32.mlp.experts.linear_fc1.weight12 + module.decoder.layers.32.mlp.experts.linear_fc1.weight10 + module.decoder.layers.32.mlp.experts.linear_fc1.weight8 + module.decoder.layers.32.mlp.experts.linear_fc2.weight6 + module.decoder.layers.32.mlp.experts.linear_fc2.weight4 + module.decoder.layers.32.mlp.experts.linear_fc2.weight2 + module.decoder.layers.32.mlp.experts.linear_fc2.weight0 + module.decoder.layers.32.mlp.experts.linear_fc1.weight15 + module.decoder.layers.32.mlp.experts.linear_fc1.weight13 + module.decoder.layers.32.mlp.experts.linear_fc1.weight11 + module.decoder.layers.32.mlp.experts.linear_fc1.weight9 + module.decoder.layers.32.mlp.experts.linear_fc1.weight7 +Params for bucket 30 (40894464 elements, 40894464 padded size): + module.decoder.layers.32.mlp.experts.linear_fc1.weight2 + module.decoder.layers.31.mlp.experts.linear_fc2.weight12 + module.decoder.layers.31.mlp.experts.linear_fc2.weight4 + module.decoder.layers.31.mlp.experts.linear_fc2.weight9 + module.decoder.layers.32.mlp.experts.linear_fc1.weight4 + module.decoder.layers.31.mlp.experts.linear_fc2.weight14 + module.decoder.layers.31.mlp.experts.linear_fc2.weight6 + module.decoder.layers.32.mlp.experts.linear_fc1.weight1 + module.decoder.layers.31.mlp.experts.linear_fc2.weight11 + module.decoder.layers.32.mlp.experts.linear_fc1.weight6 + module.decoder.layers.31.mlp.experts.linear_fc2.weight8 + module.decoder.layers.32.mlp.experts.linear_fc1.weight3 + module.decoder.layers.31.mlp.experts.linear_fc2.weight13 + module.decoder.layers.31.mlp.experts.linear_fc2.weight5 + module.decoder.layers.32.mlp.experts.linear_fc1.weight0 + module.decoder.layers.31.mlp.experts.linear_fc2.weight10 + module.decoder.layers.32.mlp.experts.linear_fc1.weight5 + module.decoder.layers.31.mlp.experts.linear_fc2.weight15 + module.decoder.layers.31.mlp.experts.linear_fc2.weight7 +Params for bucket 31 (40894464 elements, 40894464 padded size): + module.decoder.layers.31.mlp.experts.linear_fc2.weight2 + module.decoder.layers.31.mlp.experts.linear_fc2.weight0 + module.decoder.layers.31.mlp.experts.linear_fc1.weight15 + module.decoder.layers.31.mlp.experts.linear_fc1.weight13 + module.decoder.layers.31.mlp.experts.linear_fc1.weight11 + module.decoder.layers.31.mlp.experts.linear_fc1.weight9 + module.decoder.layers.31.mlp.experts.linear_fc1.weight7 + module.decoder.layers.31.mlp.experts.linear_fc1.weight5 + module.decoder.layers.31.mlp.experts.linear_fc2.weight3 + module.decoder.layers.31.mlp.experts.linear_fc2.weight1 + module.decoder.layers.31.mlp.experts.linear_fc1.weight14 + module.decoder.layers.31.mlp.experts.linear_fc1.weight12 + module.decoder.layers.31.mlp.experts.linear_fc1.weight10 + module.decoder.layers.31.mlp.experts.linear_fc1.weight8 + module.decoder.layers.31.mlp.experts.linear_fc1.weight6 +Params for bucket 32 (40894464 elements, 40894464 padded size): + module.decoder.layers.31.mlp.experts.linear_fc1.weight3 + module.decoder.layers.30.mlp.experts.linear_fc2.weight13 + module.decoder.layers.30.mlp.experts.linear_fc2.weight5 + module.decoder.layers.31.mlp.experts.linear_fc1.weight0 + module.decoder.layers.30.mlp.experts.linear_fc2.weight10 + module.decoder.layers.30.mlp.experts.linear_fc2.weight2 + module.decoder.layers.30.mlp.experts.linear_fc2.weight15 + module.decoder.layers.30.mlp.experts.linear_fc2.weight7 + module.decoder.layers.31.mlp.experts.linear_fc1.weight2 + module.decoder.layers.30.mlp.experts.linear_fc2.weight12 + module.decoder.layers.30.mlp.experts.linear_fc2.weight4 + module.decoder.layers.30.mlp.experts.linear_fc2.weight9 + module.decoder.layers.30.mlp.experts.linear_fc2.weight1 + module.decoder.layers.31.mlp.experts.linear_fc1.weight4 + module.decoder.layers.30.mlp.experts.linear_fc2.weight14 + module.decoder.layers.30.mlp.experts.linear_fc2.weight6 + module.decoder.layers.31.mlp.experts.linear_fc1.weight1 + module.decoder.layers.30.mlp.experts.linear_fc2.weight11 + module.decoder.layers.30.mlp.experts.linear_fc2.weight3 + module.decoder.layers.30.mlp.experts.linear_fc2.weight8 + module.decoder.layers.30.mlp.experts.linear_fc2.weight0 +Params for bucket 33 (40894464 elements, 40894464 padded size): + module.decoder.layers.30.mlp.experts.linear_fc1.weight12 + module.decoder.layers.30.mlp.experts.linear_fc1.weight14 + module.decoder.layers.30.mlp.experts.linear_fc1.weight10 + module.decoder.layers.30.mlp.experts.linear_fc1.weight8 + module.decoder.layers.30.mlp.experts.linear_fc1.weight6 + module.decoder.layers.30.mlp.experts.linear_fc1.weight4 + module.decoder.layers.30.mlp.experts.linear_fc1.weight15 + module.decoder.layers.30.mlp.experts.linear_fc1.weight13 + module.decoder.layers.30.mlp.experts.linear_fc1.weight11 + module.decoder.layers.30.mlp.experts.linear_fc1.weight9 + module.decoder.layers.30.mlp.experts.linear_fc1.weight7 + module.decoder.layers.30.mlp.experts.linear_fc1.weight5 + module.decoder.layers.30.mlp.experts.linear_fc1.weight3 +Params for bucket 34 (40894464 elements, 40894464 padded size): + module.decoder.layers.29.mlp.experts.linear_fc2.weight15 + module.decoder.layers.29.mlp.experts.linear_fc2.weight7 + module.decoder.layers.29.mlp.experts.linear_fc1.weight14 + module.decoder.layers.30.mlp.experts.linear_fc1.weight1 + module.decoder.layers.29.mlp.experts.linear_fc2.weight12 + module.decoder.layers.29.mlp.experts.linear_fc2.weight4 + module.decoder.layers.29.mlp.experts.linear_fc2.weight9 + module.decoder.layers.29.mlp.experts.linear_fc2.weight1 + module.decoder.layers.29.mlp.experts.linear_fc2.weight14 + module.decoder.layers.29.mlp.experts.linear_fc2.weight6 + module.decoder.layers.30.mlp.experts.linear_fc1.weight0 + module.decoder.layers.29.mlp.experts.linear_fc2.weight11 + module.decoder.layers.29.mlp.experts.linear_fc2.weight3 + module.decoder.layers.29.mlp.experts.linear_fc2.weight8 + module.decoder.layers.29.mlp.experts.linear_fc2.weight0 + module.decoder.layers.29.mlp.experts.linear_fc1.weight15 + module.decoder.layers.30.mlp.experts.linear_fc1.weight2 + module.decoder.layers.29.mlp.experts.linear_fc2.weight13 + module.decoder.layers.29.mlp.experts.linear_fc2.weight5 + module.decoder.layers.29.mlp.experts.linear_fc2.weight10 + module.decoder.layers.29.mlp.experts.linear_fc2.weight2 +Params for bucket 35 (40894464 elements, 40894464 padded size): + module.decoder.layers.29.mlp.experts.linear_fc1.weight10 + module.decoder.layers.29.mlp.experts.linear_fc1.weight12 + module.decoder.layers.29.mlp.experts.linear_fc1.weight8 + module.decoder.layers.29.mlp.experts.linear_fc1.weight6 + module.decoder.layers.29.mlp.experts.linear_fc1.weight4 + module.decoder.layers.29.mlp.experts.linear_fc1.weight2 + module.decoder.layers.29.mlp.experts.linear_fc1.weight13 + module.decoder.layers.29.mlp.experts.linear_fc1.weight9 + module.decoder.layers.29.mlp.experts.linear_fc1.weight11 + module.decoder.layers.29.mlp.experts.linear_fc1.weight7 + module.decoder.layers.29.mlp.experts.linear_fc1.weight5 + module.decoder.layers.29.mlp.experts.linear_fc1.weight3 + module.decoder.layers.29.mlp.experts.linear_fc1.weight1 +Params for bucket 36 (40894464 elements, 40894464 padded size): + module.decoder.layers.28.mlp.experts.linear_fc2.weight9 + module.decoder.layers.28.mlp.experts.linear_fc2.weight1 + module.decoder.layers.28.mlp.experts.linear_fc2.weight14 + module.decoder.layers.28.mlp.experts.linear_fc2.weight6 + module.decoder.layers.28.mlp.experts.linear_fc1.weight13 + module.decoder.layers.29.mlp.experts.linear_fc1.weight0 + module.decoder.layers.28.mlp.experts.linear_fc2.weight11 + module.decoder.layers.28.mlp.experts.linear_fc2.weight3 + module.decoder.layers.28.mlp.experts.linear_fc2.weight8 + module.decoder.layers.28.mlp.experts.linear_fc2.weight0 + module.decoder.layers.28.mlp.experts.linear_fc1.weight15 + module.decoder.layers.28.mlp.experts.linear_fc2.weight13 + module.decoder.layers.28.mlp.experts.linear_fc2.weight5 + module.decoder.layers.28.mlp.experts.linear_fc1.weight12 + module.decoder.layers.28.mlp.experts.linear_fc2.weight10 + module.decoder.layers.28.mlp.experts.linear_fc2.weight2 + module.decoder.layers.28.mlp.experts.linear_fc2.weight15 + module.decoder.layers.28.mlp.experts.linear_fc2.weight7 + module.decoder.layers.28.mlp.experts.linear_fc1.weight14 + module.decoder.layers.28.mlp.experts.linear_fc2.weight12 + module.decoder.layers.28.mlp.experts.linear_fc2.weight4 +Params for bucket 37 (40894464 elements, 40894464 padded size): + module.decoder.layers.28.mlp.experts.linear_fc1.weight10 + module.decoder.layers.28.mlp.experts.linear_fc1.weight8 + module.decoder.layers.28.mlp.experts.linear_fc1.weight6 + module.decoder.layers.28.mlp.experts.linear_fc1.weight4 + module.decoder.layers.28.mlp.experts.linear_fc1.weight2 + module.decoder.layers.28.mlp.experts.linear_fc1.weight0 + module.decoder.layers.27.mlp.experts.linear_fc2.weight15 + module.decoder.layers.28.mlp.experts.linear_fc1.weight11 + module.decoder.layers.28.mlp.experts.linear_fc1.weight7 + module.decoder.layers.28.mlp.experts.linear_fc1.weight9 + module.decoder.layers.28.mlp.experts.linear_fc1.weight5 + module.decoder.layers.28.mlp.experts.linear_fc1.weight3 + module.decoder.layers.28.mlp.experts.linear_fc1.weight1 + module.decoder.layers.27.mlp.experts.linear_fc2.weight14 +Params for bucket 38 (40894464 elements, 40894464 padded size): + module.decoder.layers.27.mlp.experts.linear_fc2.weight11 + module.decoder.layers.27.mlp.experts.linear_fc2.weight3 + module.decoder.layers.27.mlp.experts.linear_fc1.weight10 + module.decoder.layers.27.mlp.experts.linear_fc2.weight8 + module.decoder.layers.27.mlp.experts.linear_fc2.weight0 + module.decoder.layers.27.mlp.experts.linear_fc1.weight15 + module.decoder.layers.27.mlp.experts.linear_fc2.weight13 + module.decoder.layers.27.mlp.experts.linear_fc2.weight5 + module.decoder.layers.27.mlp.experts.linear_fc1.weight12 + module.decoder.layers.27.mlp.experts.linear_fc2.weight10 + module.decoder.layers.27.mlp.experts.linear_fc2.weight2 + module.decoder.layers.27.mlp.experts.linear_fc2.weight7 + module.decoder.layers.27.mlp.experts.linear_fc1.weight14 + module.decoder.layers.27.mlp.experts.linear_fc2.weight12 + module.decoder.layers.27.mlp.experts.linear_fc2.weight4 + module.decoder.layers.27.mlp.experts.linear_fc1.weight11 + module.decoder.layers.27.mlp.experts.linear_fc2.weight9 + module.decoder.layers.27.mlp.experts.linear_fc2.weight1 + module.decoder.layers.27.mlp.experts.linear_fc2.weight6 + module.decoder.layers.27.mlp.experts.linear_fc1.weight13 +Params for bucket 39 (40894464 elements, 40894464 padded size): + module.decoder.layers.27.mlp.experts.linear_fc1.weight8 + module.decoder.layers.27.mlp.experts.linear_fc1.weight6 + module.decoder.layers.27.mlp.experts.linear_fc1.weight4 + module.decoder.layers.27.mlp.experts.linear_fc1.weight2 + module.decoder.layers.27.mlp.experts.linear_fc1.weight0 + module.decoder.layers.26.mlp.experts.linear_fc2.weight15 + module.decoder.layers.26.mlp.experts.linear_fc2.weight13 + module.decoder.layers.26.mlp.experts.linear_fc2.weight11 + module.decoder.layers.27.mlp.experts.linear_fc1.weight9 + module.decoder.layers.27.mlp.experts.linear_fc1.weight7 + module.decoder.layers.27.mlp.experts.linear_fc1.weight5 + module.decoder.layers.27.mlp.experts.linear_fc1.weight3 + module.decoder.layers.27.mlp.experts.linear_fc1.weight1 + module.decoder.layers.26.mlp.experts.linear_fc2.weight14 + module.decoder.layers.26.mlp.experts.linear_fc2.weight12 + module.decoder.layers.26.mlp.experts.linear_fc2.weight10 +Params for bucket 40 (40894464 elements, 40894464 padded size): + module.decoder.layers.26.mlp.experts.linear_fc2.weight9 + module.decoder.layers.26.mlp.experts.linear_fc2.weight7 + module.decoder.layers.26.mlp.experts.linear_fc2.weight5 + module.decoder.layers.26.mlp.experts.linear_fc2.weight3 + module.decoder.layers.26.mlp.experts.linear_fc2.weight1 + module.decoder.layers.26.mlp.experts.linear_fc1.weight14 + module.decoder.layers.26.mlp.experts.linear_fc1.weight12 + module.decoder.layers.26.mlp.experts.linear_fc1.weight10 + module.decoder.layers.26.mlp.experts.linear_fc1.weight8 + module.decoder.layers.26.mlp.experts.linear_fc2.weight8 + module.decoder.layers.26.mlp.experts.linear_fc2.weight6 + module.decoder.layers.26.mlp.experts.linear_fc2.weight4 + module.decoder.layers.26.mlp.experts.linear_fc2.weight2 + module.decoder.layers.26.mlp.experts.linear_fc2.weight0 + module.decoder.layers.26.mlp.experts.linear_fc1.weight15 + module.decoder.layers.26.mlp.experts.linear_fc1.weight13 + module.decoder.layers.26.mlp.experts.linear_fc1.weight11 + module.decoder.layers.26.mlp.experts.linear_fc1.weight9 +Params for bucket 41 (40894464 elements, 40894464 padded size): + module.decoder.layers.26.mlp.experts.linear_fc1.weight4 + module.decoder.layers.26.mlp.experts.linear_fc1.weight6 + module.decoder.layers.26.mlp.experts.linear_fc1.weight2 + module.decoder.layers.26.mlp.experts.linear_fc1.weight0 + module.decoder.layers.25.mlp.experts.linear_fc2.weight15 + module.decoder.layers.25.mlp.experts.linear_fc2.weight13 + module.decoder.layers.25.mlp.experts.linear_fc2.weight11 + module.decoder.layers.25.mlp.experts.linear_fc2.weight9 + module.decoder.layers.25.mlp.experts.linear_fc2.weight7 + module.decoder.layers.26.mlp.experts.linear_fc1.weight7 + module.decoder.layers.26.mlp.experts.linear_fc1.weight5 + module.decoder.layers.26.mlp.experts.linear_fc1.weight3 + module.decoder.layers.26.mlp.experts.linear_fc1.weight1 + module.decoder.layers.25.mlp.experts.linear_fc2.weight14 + module.decoder.layers.25.mlp.experts.linear_fc2.weight12 + module.decoder.layers.25.mlp.experts.linear_fc2.weight10 + module.decoder.layers.25.mlp.experts.linear_fc2.weight8 + module.decoder.layers.25.mlp.experts.linear_fc2.weight6 +Params for bucket 42 (40894464 elements, 40894464 padded size): + module.decoder.layers.25.mlp.experts.linear_fc2.weight5 + module.decoder.layers.25.mlp.experts.linear_fc2.weight1 + module.decoder.layers.25.mlp.experts.linear_fc2.weight3 + module.decoder.layers.25.mlp.experts.linear_fc1.weight14 + module.decoder.layers.25.mlp.experts.linear_fc1.weight12 + module.decoder.layers.25.mlp.experts.linear_fc1.weight10 + module.decoder.layers.25.mlp.experts.linear_fc1.weight8 + module.decoder.layers.25.mlp.experts.linear_fc1.weight6 + module.decoder.layers.25.mlp.experts.linear_fc2.weight4 + module.decoder.layers.25.mlp.experts.linear_fc2.weight2 + module.decoder.layers.25.mlp.experts.linear_fc2.weight0 + module.decoder.layers.25.mlp.experts.linear_fc1.weight15 + module.decoder.layers.25.mlp.experts.linear_fc1.weight13 + module.decoder.layers.25.mlp.experts.linear_fc1.weight11 + module.decoder.layers.25.mlp.experts.linear_fc1.weight9 + module.decoder.layers.25.mlp.experts.linear_fc1.weight7 +Params for bucket 43 (40894464 elements, 40894464 padded size): + module.decoder.layers.24.mlp.experts.linear_fc2.weight9 + module.decoder.layers.25.mlp.experts.linear_fc1.weight3 + module.decoder.layers.24.mlp.experts.linear_fc2.weight14 + module.decoder.layers.24.mlp.experts.linear_fc2.weight6 + module.decoder.layers.25.mlp.experts.linear_fc1.weight0 + module.decoder.layers.24.mlp.experts.linear_fc2.weight11 + module.decoder.layers.24.mlp.experts.linear_fc2.weight3 + module.decoder.layers.25.mlp.experts.linear_fc1.weight5 + module.decoder.layers.24.mlp.experts.linear_fc2.weight8 + module.decoder.layers.25.mlp.experts.linear_fc1.weight2 + module.decoder.layers.24.mlp.experts.linear_fc2.weight13 + module.decoder.layers.24.mlp.experts.linear_fc2.weight5 + module.decoder.layers.24.mlp.experts.linear_fc2.weight10 + module.decoder.layers.24.mlp.experts.linear_fc2.weight2 + module.decoder.layers.25.mlp.experts.linear_fc1.weight4 + module.decoder.layers.24.mlp.experts.linear_fc2.weight15 + module.decoder.layers.24.mlp.experts.linear_fc2.weight7 + module.decoder.layers.25.mlp.experts.linear_fc1.weight1 + module.decoder.layers.24.mlp.experts.linear_fc2.weight12 + module.decoder.layers.24.mlp.experts.linear_fc2.weight4 +Params for bucket 44 (40894464 elements, 40894464 padded size): + module.decoder.layers.24.mlp.experts.linear_fc2.weight1 + module.decoder.layers.24.mlp.experts.linear_fc1.weight14 + module.decoder.layers.24.mlp.experts.linear_fc1.weight12 + module.decoder.layers.24.mlp.experts.linear_fc1.weight10 + module.decoder.layers.24.mlp.experts.linear_fc1.weight8 + module.decoder.layers.24.mlp.experts.linear_fc1.weight6 + module.decoder.layers.24.mlp.experts.linear_fc1.weight4 + module.decoder.layers.24.mlp.experts.linear_fc1.weight15 + module.decoder.layers.24.mlp.experts.linear_fc1.weight13 + module.decoder.layers.24.mlp.experts.linear_fc2.weight0 + module.decoder.layers.24.mlp.experts.linear_fc1.weight11 + module.decoder.layers.24.mlp.experts.linear_fc1.weight9 + module.decoder.layers.24.mlp.experts.linear_fc1.weight7 + module.decoder.layers.24.mlp.experts.linear_fc1.weight5 +Params for bucket 45 (40894464 elements, 40894464 padded size): + module.decoder.layers.24.mlp.experts.linear_fc1.weight0 + module.decoder.layers.23.mlp.experts.linear_fc2.weight11 + module.decoder.layers.23.mlp.experts.linear_fc2.weight3 + module.decoder.layers.23.mlp.experts.linear_fc2.weight8 + module.decoder.layers.23.mlp.experts.linear_fc2.weight0 + module.decoder.layers.24.mlp.experts.linear_fc1.weight2 + module.decoder.layers.23.mlp.experts.linear_fc2.weight13 + module.decoder.layers.23.mlp.experts.linear_fc2.weight5 + module.decoder.layers.23.mlp.experts.linear_fc2.weight10 + module.decoder.layers.23.mlp.experts.linear_fc2.weight2 + module.decoder.layers.23.mlp.experts.linear_fc2.weight15 + module.decoder.layers.23.mlp.experts.linear_fc2.weight7 + module.decoder.layers.24.mlp.experts.linear_fc1.weight1 + module.decoder.layers.23.mlp.experts.linear_fc2.weight12 + module.decoder.layers.23.mlp.experts.linear_fc2.weight4 + module.decoder.layers.23.mlp.experts.linear_fc1.weight15 + module.decoder.layers.23.mlp.experts.linear_fc2.weight9 + module.decoder.layers.23.mlp.experts.linear_fc2.weight1 + module.decoder.layers.24.mlp.experts.linear_fc1.weight3 + module.decoder.layers.23.mlp.experts.linear_fc2.weight14 + module.decoder.layers.23.mlp.experts.linear_fc2.weight6 +Params for bucket 46 (40894464 elements, 40894464 padded size): + module.decoder.layers.23.mlp.experts.linear_fc1.weight14 + module.decoder.layers.23.mlp.experts.linear_fc1.weight12 + module.decoder.layers.23.mlp.experts.linear_fc1.weight10 + module.decoder.layers.23.mlp.experts.linear_fc1.weight8 + module.decoder.layers.23.mlp.experts.linear_fc1.weight6 + module.decoder.layers.23.mlp.experts.linear_fc1.weight4 + module.decoder.layers.23.mlp.experts.linear_fc1.weight2 + module.decoder.layers.23.mlp.experts.linear_fc1.weight13 + module.decoder.layers.23.mlp.experts.linear_fc1.weight11 + module.decoder.layers.23.mlp.experts.linear_fc1.weight9 + module.decoder.layers.23.mlp.experts.linear_fc1.weight7 + module.decoder.layers.23.mlp.experts.linear_fc1.weight5 + module.decoder.layers.23.mlp.experts.linear_fc1.weight3 +Params for bucket 47 (40894464 elements, 40894464 padded size): + module.decoder.layers.22.mlp.experts.linear_fc2.weight13 + module.decoder.layers.22.mlp.experts.linear_fc2.weight5 + module.decoder.layers.22.mlp.experts.linear_fc2.weight10 + module.decoder.layers.22.mlp.experts.linear_fc2.weight2 + module.decoder.layers.22.mlp.experts.linear_fc1.weight13 + module.decoder.layers.22.mlp.experts.linear_fc2.weight15 + module.decoder.layers.22.mlp.experts.linear_fc2.weight7 + module.decoder.layers.23.mlp.experts.linear_fc1.weight1 + module.decoder.layers.22.mlp.experts.linear_fc2.weight12 + module.decoder.layers.22.mlp.experts.linear_fc2.weight4 + module.decoder.layers.22.mlp.experts.linear_fc1.weight15 + module.decoder.layers.22.mlp.experts.linear_fc2.weight9 + module.decoder.layers.22.mlp.experts.linear_fc2.weight1 + module.decoder.layers.22.mlp.experts.linear_fc2.weight14 + module.decoder.layers.22.mlp.experts.linear_fc2.weight6 + module.decoder.layers.23.mlp.experts.linear_fc1.weight0 + module.decoder.layers.22.mlp.experts.linear_fc2.weight11 + module.decoder.layers.22.mlp.experts.linear_fc2.weight3 + module.decoder.layers.22.mlp.experts.linear_fc1.weight14 + module.decoder.layers.22.mlp.experts.linear_fc2.weight8 + module.decoder.layers.22.mlp.experts.linear_fc2.weight0 +Params for bucket 48 (40894464 elements, 40894464 padded size): + module.decoder.layers.22.mlp.experts.linear_fc1.weight12 + module.decoder.layers.22.mlp.experts.linear_fc1.weight10 + module.decoder.layers.22.mlp.experts.linear_fc1.weight8 + module.decoder.layers.22.mlp.experts.linear_fc1.weight6 + module.decoder.layers.22.mlp.experts.linear_fc1.weight4 + module.decoder.layers.22.mlp.experts.linear_fc1.weight2 + module.decoder.layers.22.mlp.experts.linear_fc1.weight0 + module.decoder.layers.22.mlp.experts.linear_fc1.weight11 + module.decoder.layers.22.mlp.experts.linear_fc1.weight9 + module.decoder.layers.22.mlp.experts.linear_fc1.weight7 + module.decoder.layers.22.mlp.experts.linear_fc1.weight5 + module.decoder.layers.22.mlp.experts.linear_fc1.weight3 + module.decoder.layers.22.mlp.experts.linear_fc1.weight1 +Params for bucket 49 (40894464 elements, 40894464 padded size): + module.decoder.layers.21.mlp.experts.linear_fc2.weight8 + module.decoder.layers.21.mlp.experts.linear_fc2.weight0 + module.decoder.layers.21.mlp.experts.linear_fc1.weight11 + module.decoder.layers.21.mlp.experts.linear_fc2.weight13 + module.decoder.layers.21.mlp.experts.linear_fc2.weight5 + module.decoder.layers.21.mlp.experts.linear_fc2.weight10 + module.decoder.layers.21.mlp.experts.linear_fc2.weight2 + module.decoder.layers.21.mlp.experts.linear_fc1.weight13 + module.decoder.layers.21.mlp.experts.linear_fc2.weight15 + module.decoder.layers.21.mlp.experts.linear_fc2.weight7 + module.decoder.layers.21.mlp.experts.linear_fc2.weight12 + module.decoder.layers.21.mlp.experts.linear_fc2.weight4 + module.decoder.layers.21.mlp.experts.linear_fc1.weight15 + module.decoder.layers.21.mlp.experts.linear_fc2.weight9 + module.decoder.layers.21.mlp.experts.linear_fc2.weight1 + module.decoder.layers.21.mlp.experts.linear_fc1.weight12 + module.decoder.layers.21.mlp.experts.linear_fc2.weight14 + module.decoder.layers.21.mlp.experts.linear_fc2.weight6 + module.decoder.layers.21.mlp.experts.linear_fc2.weight11 + module.decoder.layers.21.mlp.experts.linear_fc2.weight3 + module.decoder.layers.21.mlp.experts.linear_fc1.weight14 +Params for bucket 50 (40894464 elements, 40894464 padded size): + module.decoder.layers.21.mlp.experts.linear_fc1.weight9 + module.decoder.layers.21.mlp.experts.linear_fc1.weight7 + module.decoder.layers.21.mlp.experts.linear_fc1.weight5 + module.decoder.layers.21.mlp.experts.linear_fc1.weight3 + module.decoder.layers.21.mlp.experts.linear_fc1.weight1 + module.decoder.layers.20.mlp.experts.linear_fc2.weight15 + module.decoder.layers.20.mlp.experts.linear_fc2.weight13 + module.decoder.layers.21.mlp.experts.linear_fc1.weight10 + module.decoder.layers.21.mlp.experts.linear_fc1.weight6 + module.decoder.layers.21.mlp.experts.linear_fc1.weight8 + module.decoder.layers.21.mlp.experts.linear_fc1.weight4 + module.decoder.layers.21.mlp.experts.linear_fc1.weight2 + module.decoder.layers.21.mlp.experts.linear_fc1.weight0 + module.decoder.layers.20.mlp.experts.linear_fc2.weight14 + module.decoder.layers.20.mlp.experts.linear_fc2.weight12 +Params for bucket 51 (40894464 elements, 40894464 padded size): + module.decoder.layers.20.mlp.experts.linear_fc2.weight11 + module.decoder.layers.20.mlp.experts.linear_fc2.weight3 + module.decoder.layers.20.mlp.experts.linear_fc1.weight14 + module.decoder.layers.20.mlp.experts.linear_fc2.weight8 + module.decoder.layers.20.mlp.experts.linear_fc2.weight0 + module.decoder.layers.20.mlp.experts.linear_fc1.weight11 + module.decoder.layers.20.mlp.experts.linear_fc2.weight5 + module.decoder.layers.20.mlp.experts.linear_fc2.weight10 + module.decoder.layers.20.mlp.experts.linear_fc2.weight2 + module.decoder.layers.20.mlp.experts.linear_fc1.weight13 + module.decoder.layers.20.mlp.experts.linear_fc2.weight7 + module.decoder.layers.20.mlp.experts.linear_fc1.weight10 + module.decoder.layers.20.mlp.experts.linear_fc2.weight4 + module.decoder.layers.20.mlp.experts.linear_fc1.weight15 + module.decoder.layers.20.mlp.experts.linear_fc2.weight9 + module.decoder.layers.20.mlp.experts.linear_fc2.weight1 + module.decoder.layers.20.mlp.experts.linear_fc1.weight12 + module.decoder.layers.20.mlp.experts.linear_fc2.weight6 + module.decoder.layers.20.mlp.experts.linear_fc1.weight9 +Params for bucket 52 (40894464 elements, 40894464 padded size): + module.decoder.layers.20.mlp.experts.linear_fc1.weight8 + module.decoder.layers.20.mlp.experts.linear_fc1.weight4 + module.decoder.layers.20.mlp.experts.linear_fc1.weight6 + module.decoder.layers.20.mlp.experts.linear_fc1.weight2 + module.decoder.layers.20.mlp.experts.linear_fc1.weight0 + module.decoder.layers.19.mlp.experts.linear_fc2.weight15 + module.decoder.layers.19.mlp.experts.linear_fc2.weight13 + module.decoder.layers.19.mlp.experts.linear_fc2.weight11 + module.decoder.layers.19.mlp.experts.linear_fc2.weight9 + module.decoder.layers.20.mlp.experts.linear_fc1.weight5 + module.decoder.layers.20.mlp.experts.linear_fc1.weight7 + module.decoder.layers.20.mlp.experts.linear_fc1.weight3 + module.decoder.layers.20.mlp.experts.linear_fc1.weight1 + module.decoder.layers.19.mlp.experts.linear_fc2.weight14 + module.decoder.layers.19.mlp.experts.linear_fc2.weight12 + module.decoder.layers.19.mlp.experts.linear_fc2.weight10 + module.decoder.layers.19.mlp.experts.linear_fc2.weight8 +Params for bucket 53 (40894464 elements, 40894464 padded size): + module.decoder.layers.19.mlp.experts.linear_fc2.weight7 + module.decoder.layers.19.mlp.experts.linear_fc2.weight5 + module.decoder.layers.19.mlp.experts.linear_fc2.weight3 + module.decoder.layers.19.mlp.experts.linear_fc2.weight1 + module.decoder.layers.19.mlp.experts.linear_fc1.weight15 + module.decoder.layers.19.mlp.experts.linear_fc1.weight13 + module.decoder.layers.19.mlp.experts.linear_fc1.weight11 + module.decoder.layers.19.mlp.experts.linear_fc1.weight9 + module.decoder.layers.19.mlp.experts.linear_fc1.weight7 + module.decoder.layers.19.mlp.experts.linear_fc2.weight6 + module.decoder.layers.19.mlp.experts.linear_fc2.weight4 + module.decoder.layers.19.mlp.experts.linear_fc2.weight2 + module.decoder.layers.19.mlp.experts.linear_fc2.weight0 + module.decoder.layers.19.mlp.experts.linear_fc1.weight14 + module.decoder.layers.19.mlp.experts.linear_fc1.weight12 + module.decoder.layers.19.mlp.experts.linear_fc1.weight10 + module.decoder.layers.19.mlp.experts.linear_fc1.weight8 +Params for bucket 54 (40894464 elements, 40894464 padded size): + module.decoder.layers.18.mlp.experts.linear_fc2.weight14 + module.decoder.layers.18.mlp.experts.linear_fc2.weight6 + module.decoder.layers.19.mlp.experts.linear_fc1.weight4 + module.decoder.layers.18.mlp.experts.linear_fc2.weight11 + module.decoder.layers.19.mlp.experts.linear_fc1.weight1 + module.decoder.layers.18.mlp.experts.linear_fc2.weight8 + module.decoder.layers.19.mlp.experts.linear_fc1.weight6 + module.decoder.layers.18.mlp.experts.linear_fc2.weight13 + module.decoder.layers.18.mlp.experts.linear_fc2.weight5 + module.decoder.layers.19.mlp.experts.linear_fc1.weight3 + module.decoder.layers.18.mlp.experts.linear_fc2.weight10 + module.decoder.layers.19.mlp.experts.linear_fc1.weight0 + module.decoder.layers.18.mlp.experts.linear_fc2.weight15 + module.decoder.layers.18.mlp.experts.linear_fc2.weight7 + module.decoder.layers.19.mlp.experts.linear_fc1.weight5 + module.decoder.layers.18.mlp.experts.linear_fc2.weight12 + module.decoder.layers.18.mlp.experts.linear_fc2.weight4 + module.decoder.layers.19.mlp.experts.linear_fc1.weight2 + module.decoder.layers.18.mlp.experts.linear_fc2.weight9 +Params for bucket 55 (40894464 elements, 40894464 padded size): + module.decoder.layers.18.mlp.experts.linear_fc2.weight2 + module.decoder.layers.18.mlp.experts.linear_fc1.weight15 + module.decoder.layers.18.mlp.experts.linear_fc2.weight0 + module.decoder.layers.18.mlp.experts.linear_fc1.weight13 + module.decoder.layers.18.mlp.experts.linear_fc1.weight11 + module.decoder.layers.18.mlp.experts.linear_fc1.weight9 + module.decoder.layers.18.mlp.experts.linear_fc1.weight7 + module.decoder.layers.18.mlp.experts.linear_fc1.weight5 + module.decoder.layers.18.mlp.experts.linear_fc2.weight3 + module.decoder.layers.18.mlp.experts.linear_fc2.weight1 + module.decoder.layers.18.mlp.experts.linear_fc1.weight14 + module.decoder.layers.18.mlp.experts.linear_fc1.weight12 + module.decoder.layers.18.mlp.experts.linear_fc1.weight10 + module.decoder.layers.18.mlp.experts.linear_fc1.weight8 + module.decoder.layers.18.mlp.experts.linear_fc1.weight6 +Params for bucket 56 (40894464 elements, 40894464 padded size): + module.decoder.layers.18.mlp.experts.linear_fc1.weight1 + module.decoder.layers.17.mlp.experts.linear_fc2.weight8 + module.decoder.layers.17.mlp.experts.linear_fc2.weight0 + module.decoder.layers.17.mlp.experts.linear_fc2.weight13 + module.decoder.layers.17.mlp.experts.linear_fc2.weight5 + module.decoder.layers.18.mlp.experts.linear_fc1.weight3 + module.decoder.layers.17.mlp.experts.linear_fc2.weight10 + module.decoder.layers.17.mlp.experts.linear_fc2.weight2 + module.decoder.layers.18.mlp.experts.linear_fc1.weight0 + module.decoder.layers.17.mlp.experts.linear_fc2.weight15 + module.decoder.layers.17.mlp.experts.linear_fc2.weight7 + module.decoder.layers.17.mlp.experts.linear_fc2.weight12 + module.decoder.layers.17.mlp.experts.linear_fc2.weight4 + module.decoder.layers.18.mlp.experts.linear_fc1.weight2 + module.decoder.layers.17.mlp.experts.linear_fc2.weight9 + module.decoder.layers.17.mlp.experts.linear_fc2.weight1 + module.decoder.layers.17.mlp.experts.linear_fc2.weight14 + module.decoder.layers.17.mlp.experts.linear_fc2.weight6 + module.decoder.layers.18.mlp.experts.linear_fc1.weight4 + module.decoder.layers.17.mlp.experts.linear_fc2.weight11 + module.decoder.layers.17.mlp.experts.linear_fc2.weight3 +Params for bucket 57 (40894464 elements, 40894464 padded size): + module.decoder.layers.17.mlp.experts.linear_fc1.weight14 + module.decoder.layers.17.mlp.experts.linear_fc1.weight12 + module.decoder.layers.17.mlp.experts.linear_fc1.weight10 + module.decoder.layers.17.mlp.experts.linear_fc1.weight8 + module.decoder.layers.17.mlp.experts.linear_fc1.weight6 + module.decoder.layers.17.mlp.experts.linear_fc1.weight4 + module.decoder.layers.17.mlp.experts.linear_fc1.weight15 + module.decoder.layers.17.mlp.experts.linear_fc1.weight11 + module.decoder.layers.17.mlp.experts.linear_fc1.weight13 + module.decoder.layers.17.mlp.experts.linear_fc1.weight9 + module.decoder.layers.17.mlp.experts.linear_fc1.weight7 + module.decoder.layers.17.mlp.experts.linear_fc1.weight5 + module.decoder.layers.17.mlp.experts.linear_fc1.weight3 +Params for bucket 58 (40894464 elements, 40894464 padded size): + module.decoder.layers.17.mlp.experts.linear_fc1.weight2 + module.decoder.layers.16.mlp.experts.linear_fc2.weight9 + module.decoder.layers.16.mlp.experts.linear_fc2.weight1 + module.decoder.layers.16.mlp.experts.linear_fc2.weight14 + module.decoder.layers.16.mlp.experts.linear_fc2.weight6 + module.decoder.layers.16.mlp.experts.linear_fc2.weight11 + module.decoder.layers.16.mlp.experts.linear_fc2.weight3 + module.decoder.layers.16.mlp.experts.linear_fc1.weight14 + module.decoder.layers.17.mlp.experts.linear_fc1.weight1 + module.decoder.layers.16.mlp.experts.linear_fc2.weight8 + module.decoder.layers.16.mlp.experts.linear_fc2.weight0 + module.decoder.layers.16.mlp.experts.linear_fc2.weight13 + module.decoder.layers.16.mlp.experts.linear_fc2.weight5 + module.decoder.layers.16.mlp.experts.linear_fc2.weight10 + module.decoder.layers.16.mlp.experts.linear_fc2.weight2 + module.decoder.layers.17.mlp.experts.linear_fc1.weight0 + module.decoder.layers.16.mlp.experts.linear_fc2.weight15 + module.decoder.layers.16.mlp.experts.linear_fc2.weight7 + module.decoder.layers.16.mlp.experts.linear_fc2.weight12 + module.decoder.layers.16.mlp.experts.linear_fc2.weight4 + module.decoder.layers.16.mlp.experts.linear_fc1.weight15 +Params for bucket 59 (40894464 elements, 40894464 padded size): + module.decoder.layers.16.mlp.experts.linear_fc1.weight10 + module.decoder.layers.16.mlp.experts.linear_fc1.weight12 + module.decoder.layers.16.mlp.experts.linear_fc1.weight8 + module.decoder.layers.16.mlp.experts.linear_fc1.weight6 + module.decoder.layers.16.mlp.experts.linear_fc1.weight4 + module.decoder.layers.16.mlp.experts.linear_fc1.weight2 + module.decoder.layers.16.mlp.experts.linear_fc1.weight13 + module.decoder.layers.16.mlp.experts.linear_fc1.weight11 + module.decoder.layers.16.mlp.experts.linear_fc1.weight9 + module.decoder.layers.16.mlp.experts.linear_fc1.weight7 + module.decoder.layers.16.mlp.experts.linear_fc1.weight5 + module.decoder.layers.16.mlp.experts.linear_fc1.weight3 + module.decoder.layers.16.mlp.experts.linear_fc1.weight1 +Params for bucket 60 (40894464 elements, 40894464 padded size): + module.decoder.layers.15.mlp.experts.linear_fc2.weight11 + module.decoder.layers.15.mlp.experts.linear_fc2.weight3 + module.decoder.layers.15.mlp.experts.linear_fc1.weight14 + module.decoder.layers.15.mlp.experts.linear_fc2.weight8 + module.decoder.layers.15.mlp.experts.linear_fc2.weight0 + module.decoder.layers.15.mlp.experts.linear_fc2.weight13 + module.decoder.layers.15.mlp.experts.linear_fc2.weight5 + module.decoder.layers.15.mlp.experts.linear_fc2.weight10 + module.decoder.layers.15.mlp.experts.linear_fc2.weight2 + module.decoder.layers.15.mlp.experts.linear_fc1.weight13 + module.decoder.layers.16.mlp.experts.linear_fc1.weight0 + module.decoder.layers.15.mlp.experts.linear_fc2.weight15 + module.decoder.layers.15.mlp.experts.linear_fc2.weight7 + module.decoder.layers.15.mlp.experts.linear_fc2.weight12 + module.decoder.layers.15.mlp.experts.linear_fc2.weight4 + module.decoder.layers.15.mlp.experts.linear_fc1.weight15 + module.decoder.layers.15.mlp.experts.linear_fc2.weight9 + module.decoder.layers.15.mlp.experts.linear_fc2.weight1 + module.decoder.layers.15.mlp.experts.linear_fc1.weight12 + module.decoder.layers.15.mlp.experts.linear_fc2.weight14 + module.decoder.layers.15.mlp.experts.linear_fc2.weight6 +Params for bucket 61 (40894464 elements, 40894464 padded size): + module.decoder.layers.15.mlp.experts.linear_fc1.weight8 + module.decoder.layers.15.mlp.experts.linear_fc1.weight10 + module.decoder.layers.15.mlp.experts.linear_fc1.weight6 + module.decoder.layers.15.mlp.experts.linear_fc1.weight4 + module.decoder.layers.15.mlp.experts.linear_fc1.weight2 + module.decoder.layers.15.mlp.experts.linear_fc1.weight0 + module.decoder.layers.14.mlp.experts.linear_fc2.weight15 + module.decoder.layers.15.mlp.experts.linear_fc1.weight11 + module.decoder.layers.15.mlp.experts.linear_fc1.weight9 + module.decoder.layers.15.mlp.experts.linear_fc1.weight7 + module.decoder.layers.15.mlp.experts.linear_fc1.weight5 + module.decoder.layers.15.mlp.experts.linear_fc1.weight3 + module.decoder.layers.15.mlp.experts.linear_fc1.weight1 + module.decoder.layers.14.mlp.experts.linear_fc2.weight14 +Params for bucket 62 (40894464 elements, 40894464 padded size): + module.decoder.layers.14.mlp.experts.linear_fc2.weight13 + module.decoder.layers.14.mlp.experts.linear_fc2.weight5 + module.decoder.layers.14.mlp.experts.linear_fc2.weight10 + module.decoder.layers.14.mlp.experts.linear_fc2.weight2 + module.decoder.layers.14.mlp.experts.linear_fc1.weight13 + module.decoder.layers.14.mlp.experts.linear_fc2.weight7 + module.decoder.layers.14.mlp.experts.linear_fc1.weight10 + module.decoder.layers.14.mlp.experts.linear_fc2.weight12 + module.decoder.layers.14.mlp.experts.linear_fc2.weight4 + module.decoder.layers.14.mlp.experts.linear_fc1.weight15 + module.decoder.layers.14.mlp.experts.linear_fc2.weight9 + module.decoder.layers.14.mlp.experts.linear_fc2.weight1 + module.decoder.layers.14.mlp.experts.linear_fc1.weight12 + module.decoder.layers.14.mlp.experts.linear_fc2.weight6 + module.decoder.layers.14.mlp.experts.linear_fc2.weight11 + module.decoder.layers.14.mlp.experts.linear_fc2.weight3 + module.decoder.layers.14.mlp.experts.linear_fc1.weight14 + module.decoder.layers.14.mlp.experts.linear_fc2.weight8 + module.decoder.layers.14.mlp.experts.linear_fc2.weight0 + module.decoder.layers.14.mlp.experts.linear_fc1.weight11 +Params for bucket 63 (40894464 elements, 40894464 padded size): + module.decoder.layers.14.mlp.experts.linear_fc1.weight8 + module.decoder.layers.14.mlp.experts.linear_fc1.weight6 + module.decoder.layers.14.mlp.experts.linear_fc1.weight4 + module.decoder.layers.14.mlp.experts.linear_fc1.weight2 + module.decoder.layers.14.mlp.experts.linear_fc1.weight0 + module.decoder.layers.13.mlp.experts.linear_fc2.weight15 + module.decoder.layers.13.mlp.experts.linear_fc2.weight13 + module.decoder.layers.13.mlp.experts.linear_fc2.weight11 + module.decoder.layers.14.mlp.experts.linear_fc1.weight9 + module.decoder.layers.14.mlp.experts.linear_fc1.weight5 + module.decoder.layers.14.mlp.experts.linear_fc1.weight7 + module.decoder.layers.14.mlp.experts.linear_fc1.weight3 + module.decoder.layers.14.mlp.experts.linear_fc1.weight1 + module.decoder.layers.13.mlp.experts.linear_fc2.weight14 + module.decoder.layers.13.mlp.experts.linear_fc2.weight12 + module.decoder.layers.13.mlp.experts.linear_fc2.weight10 +Params for bucket 64 (40894464 elements, 40894464 padded size): + module.decoder.layers.13.mlp.experts.linear_fc2.weight9 + module.decoder.layers.13.mlp.experts.linear_fc2.weight5 + module.decoder.layers.13.mlp.experts.linear_fc2.weight7 + module.decoder.layers.13.mlp.experts.linear_fc2.weight3 + module.decoder.layers.13.mlp.experts.linear_fc2.weight1 + module.decoder.layers.13.mlp.experts.linear_fc1.weight14 + module.decoder.layers.13.mlp.experts.linear_fc1.weight12 + module.decoder.layers.13.mlp.experts.linear_fc1.weight10 + module.decoder.layers.13.mlp.experts.linear_fc1.weight8 + module.decoder.layers.13.mlp.experts.linear_fc2.weight6 + module.decoder.layers.13.mlp.experts.linear_fc2.weight8 + module.decoder.layers.13.mlp.experts.linear_fc2.weight4 + module.decoder.layers.13.mlp.experts.linear_fc2.weight2 + module.decoder.layers.13.mlp.experts.linear_fc2.weight0 + module.decoder.layers.13.mlp.experts.linear_fc1.weight15 + module.decoder.layers.13.mlp.experts.linear_fc1.weight13 + module.decoder.layers.13.mlp.experts.linear_fc1.weight11 + module.decoder.layers.13.mlp.experts.linear_fc1.weight9 +Params for bucket 65 (40894464 elements, 40894464 padded size): + module.decoder.layers.13.mlp.experts.linear_fc1.weight6 + module.decoder.layers.13.mlp.experts.linear_fc1.weight4 + module.decoder.layers.13.mlp.experts.linear_fc1.weight2 + module.decoder.layers.13.mlp.experts.linear_fc1.weight0 + module.decoder.layers.12.mlp.experts.linear_fc2.weight15 + module.decoder.layers.12.mlp.experts.linear_fc2.weight13 + module.decoder.layers.12.mlp.experts.linear_fc2.weight11 + module.decoder.layers.12.mlp.experts.linear_fc2.weight9 + module.decoder.layers.12.mlp.experts.linear_fc2.weight7 + module.decoder.layers.13.mlp.experts.linear_fc1.weight7 + module.decoder.layers.13.mlp.experts.linear_fc1.weight3 + module.decoder.layers.13.mlp.experts.linear_fc1.weight5 + module.decoder.layers.13.mlp.experts.linear_fc1.weight1 + module.decoder.layers.12.mlp.experts.linear_fc2.weight14 + module.decoder.layers.12.mlp.experts.linear_fc2.weight12 + module.decoder.layers.12.mlp.experts.linear_fc2.weight10 + module.decoder.layers.12.mlp.experts.linear_fc2.weight8 + module.decoder.layers.12.mlp.experts.linear_fc2.weight6 +Params for bucket 66 (40894464 elements, 40894464 padded size): + 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.weight1 + module.decoder.layers.12.mlp.experts.linear_fc1.weight14 + module.decoder.layers.12.mlp.experts.linear_fc1.weight12 + module.decoder.layers.12.mlp.experts.linear_fc1.weight10 + module.decoder.layers.12.mlp.experts.linear_fc1.weight8 + module.decoder.layers.12.mlp.experts.linear_fc1.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.weight0 + module.decoder.layers.12.mlp.experts.linear_fc1.weight15 + module.decoder.layers.12.mlp.experts.linear_fc1.weight13 + module.decoder.layers.12.mlp.experts.linear_fc1.weight11 + module.decoder.layers.12.mlp.experts.linear_fc1.weight9 + module.decoder.layers.12.mlp.experts.linear_fc1.weight7 +Params for bucket 67 (40894464 elements, 40894464 padded size): + module.decoder.layers.12.mlp.experts.linear_fc1.weight4 + module.decoder.layers.11.mlp.experts.linear_fc2.weight11 + module.decoder.layers.11.mlp.experts.linear_fc2.weight3 + module.decoder.layers.12.mlp.experts.linear_fc1.weight1 + module.decoder.layers.11.mlp.experts.linear_fc2.weight8 + module.decoder.layers.11.mlp.experts.linear_fc2.weight13 + module.decoder.layers.11.mlp.experts.linear_fc2.weight5 + module.decoder.layers.12.mlp.experts.linear_fc1.weight3 + module.decoder.layers.11.mlp.experts.linear_fc2.weight10 + module.decoder.layers.11.mlp.experts.linear_fc2.weight2 + module.decoder.layers.12.mlp.experts.linear_fc1.weight0 + module.decoder.layers.11.mlp.experts.linear_fc2.weight15 + module.decoder.layers.11.mlp.experts.linear_fc2.weight7 + module.decoder.layers.12.mlp.experts.linear_fc1.weight5 + module.decoder.layers.11.mlp.experts.linear_fc2.weight12 + module.decoder.layers.11.mlp.experts.linear_fc2.weight4 + module.decoder.layers.12.mlp.experts.linear_fc1.weight2 + module.decoder.layers.11.mlp.experts.linear_fc2.weight9 + module.decoder.layers.11.mlp.experts.linear_fc2.weight14 + module.decoder.layers.11.mlp.experts.linear_fc2.weight6 +Params for bucket 68 (40894464 elements, 40894464 padded size): + module.decoder.layers.11.mlp.experts.linear_fc2.weight1 + module.decoder.layers.11.mlp.experts.linear_fc1.weight14 + module.decoder.layers.11.mlp.experts.linear_fc1.weight12 + module.decoder.layers.11.mlp.experts.linear_fc1.weight10 + module.decoder.layers.11.mlp.experts.linear_fc1.weight8 + module.decoder.layers.11.mlp.experts.linear_fc1.weight6 + module.decoder.layers.11.mlp.experts.linear_fc1.weight4 + module.decoder.layers.11.mlp.experts.linear_fc2.weight0 + module.decoder.layers.11.mlp.experts.linear_fc1.weight13 + module.decoder.layers.11.mlp.experts.linear_fc1.weight15 + module.decoder.layers.11.mlp.experts.linear_fc1.weight11 + module.decoder.layers.11.mlp.experts.linear_fc1.weight9 + module.decoder.layers.11.mlp.experts.linear_fc1.weight7 + module.decoder.layers.11.mlp.experts.linear_fc1.weight5 +Params for bucket 69 (40894464 elements, 40894464 padded size): + module.decoder.layers.10.mlp.experts.linear_fc2.weight13 + module.decoder.layers.10.mlp.experts.linear_fc2.weight5 + module.decoder.layers.11.mlp.experts.linear_fc1.weight3 + module.decoder.layers.10.mlp.experts.linear_fc2.weight10 + module.decoder.layers.10.mlp.experts.linear_fc2.weight2 + module.decoder.layers.11.mlp.experts.linear_fc1.weight0 + module.decoder.layers.10.mlp.experts.linear_fc2.weight15 + module.decoder.layers.10.mlp.experts.linear_fc2.weight7 + module.decoder.layers.10.mlp.experts.linear_fc2.weight12 + module.decoder.layers.10.mlp.experts.linear_fc2.weight4 + module.decoder.layers.10.mlp.experts.linear_fc1.weight15 + module.decoder.layers.11.mlp.experts.linear_fc1.weight2 + module.decoder.layers.10.mlp.experts.linear_fc2.weight9 + module.decoder.layers.10.mlp.experts.linear_fc2.weight1 + module.decoder.layers.10.mlp.experts.linear_fc2.weight14 + module.decoder.layers.10.mlp.experts.linear_fc2.weight6 + module.decoder.layers.10.mlp.experts.linear_fc2.weight11 + module.decoder.layers.10.mlp.experts.linear_fc2.weight3 + module.decoder.layers.11.mlp.experts.linear_fc1.weight1 + module.decoder.layers.10.mlp.experts.linear_fc2.weight8 + module.decoder.layers.10.mlp.experts.linear_fc2.weight0 +Params for bucket 70 (40894464 elements, 40894464 padded size): + module.decoder.layers.10.mlp.experts.linear_fc1.weight14 + module.decoder.layers.10.mlp.experts.linear_fc1.weight10 + module.decoder.layers.10.mlp.experts.linear_fc1.weight12 + module.decoder.layers.10.mlp.experts.linear_fc1.weight8 + module.decoder.layers.10.mlp.experts.linear_fc1.weight6 + module.decoder.layers.10.mlp.experts.linear_fc1.weight4 + module.decoder.layers.10.mlp.experts.linear_fc1.weight2 + module.decoder.layers.10.mlp.experts.linear_fc1.weight13 + module.decoder.layers.10.mlp.experts.linear_fc1.weight11 + module.decoder.layers.10.mlp.experts.linear_fc1.weight9 + module.decoder.layers.10.mlp.experts.linear_fc1.weight7 + module.decoder.layers.10.mlp.experts.linear_fc1.weight5 + module.decoder.layers.10.mlp.experts.linear_fc1.weight3 +Params for bucket 71 (40894464 elements, 40894464 padded size): + module.decoder.layers.10.mlp.experts.linear_fc1.weight0 + module.decoder.layers.9.mlp.experts.linear_fc2.weight14 + module.decoder.layers.9.mlp.experts.linear_fc2.weight6 + module.decoder.layers.9.mlp.experts.linear_fc2.weight11 + module.decoder.layers.9.mlp.experts.linear_fc2.weight3 + module.decoder.layers.9.mlp.experts.linear_fc1.weight14 + module.decoder.layers.9.mlp.experts.linear_fc2.weight8 + module.decoder.layers.9.mlp.experts.linear_fc2.weight0 + module.decoder.layers.9.mlp.experts.linear_fc2.weight13 + module.decoder.layers.9.mlp.experts.linear_fc2.weight5 + module.decoder.layers.9.mlp.experts.linear_fc2.weight10 + module.decoder.layers.9.mlp.experts.linear_fc2.weight2 + module.decoder.layers.9.mlp.experts.linear_fc1.weight13 + module.decoder.layers.10.mlp.experts.linear_fc1.weight1 + module.decoder.layers.9.mlp.experts.linear_fc2.weight15 + module.decoder.layers.9.mlp.experts.linear_fc2.weight7 + module.decoder.layers.9.mlp.experts.linear_fc2.weight12 + module.decoder.layers.9.mlp.experts.linear_fc2.weight4 + module.decoder.layers.9.mlp.experts.linear_fc1.weight15 + module.decoder.layers.9.mlp.experts.linear_fc2.weight9 + module.decoder.layers.9.mlp.experts.linear_fc2.weight1 +Params for bucket 72 (40894464 elements, 40894464 padded size): + module.decoder.layers.9.mlp.experts.linear_fc1.weight9 + module.decoder.layers.9.mlp.experts.linear_fc1.weight11 + module.decoder.layers.9.mlp.experts.linear_fc1.weight7 + 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.weight1 + module.decoder.layers.9.mlp.experts.linear_fc1.weight12 + module.decoder.layers.9.mlp.experts.linear_fc1.weight10 + module.decoder.layers.9.mlp.experts.linear_fc1.weight8 + module.decoder.layers.9.mlp.experts.linear_fc1.weight6 + module.decoder.layers.9.mlp.experts.linear_fc1.weight4 + module.decoder.layers.9.mlp.experts.linear_fc1.weight2 + module.decoder.layers.9.mlp.experts.linear_fc1.weight0 +Params for bucket 73 (40894464 elements, 40894464 padded size): + module.decoder.layers.8.mlp.experts.linear_fc2.weight8 + module.decoder.layers.8.mlp.experts.linear_fc2.weight0 + module.decoder.layers.8.mlp.experts.linear_fc1.weight11 + module.decoder.layers.8.mlp.experts.linear_fc2.weight13 + module.decoder.layers.8.mlp.experts.linear_fc2.weight5 + module.decoder.layers.8.mlp.experts.linear_fc2.weight10 + module.decoder.layers.8.mlp.experts.linear_fc2.weight2 + module.decoder.layers.8.mlp.experts.linear_fc1.weight13 + module.decoder.layers.8.mlp.experts.linear_fc2.weight15 + module.decoder.layers.8.mlp.experts.linear_fc2.weight7 + module.decoder.layers.8.mlp.experts.linear_fc2.weight12 + module.decoder.layers.8.mlp.experts.linear_fc2.weight4 + module.decoder.layers.8.mlp.experts.linear_fc1.weight15 + module.decoder.layers.8.mlp.experts.linear_fc2.weight9 + module.decoder.layers.8.mlp.experts.linear_fc2.weight1 + module.decoder.layers.8.mlp.experts.linear_fc1.weight12 + module.decoder.layers.8.mlp.experts.linear_fc2.weight14 + module.decoder.layers.8.mlp.experts.linear_fc2.weight6 + module.decoder.layers.8.mlp.experts.linear_fc2.weight11 + module.decoder.layers.8.mlp.experts.linear_fc2.weight3 + module.decoder.layers.8.mlp.experts.linear_fc1.weight14 +Params for bucket 74 (40894464 elements, 40894464 padded size): + module.decoder.layers.8.mlp.experts.linear_fc1.weight7 + module.decoder.layers.8.mlp.experts.linear_fc1.weight9 + module.decoder.layers.8.mlp.experts.linear_fc1.weight5 + module.decoder.layers.8.mlp.experts.linear_fc1.weight3 + module.decoder.layers.8.mlp.experts.linear_fc1.weight1 + module.decoder.layers.7.mlp.experts.linear_fc2.weight15 + module.decoder.layers.7.mlp.experts.linear_fc2.weight13 + module.decoder.layers.8.mlp.experts.linear_fc1.weight10 + module.decoder.layers.8.mlp.experts.linear_fc1.weight6 + module.decoder.layers.8.mlp.experts.linear_fc1.weight8 + module.decoder.layers.8.mlp.experts.linear_fc1.weight4 + module.decoder.layers.8.mlp.experts.linear_fc1.weight2 + module.decoder.layers.8.mlp.experts.linear_fc1.weight0 + module.decoder.layers.7.mlp.experts.linear_fc2.weight14 + module.decoder.layers.7.mlp.experts.linear_fc2.weight12 +Params for bucket 75 (40894464 elements, 40894464 padded size): + module.decoder.layers.7.mlp.experts.linear_fc2.weight9 + module.decoder.layers.7.mlp.experts.linear_fc2.weight1 + module.decoder.layers.7.mlp.experts.linear_fc1.weight12 + module.decoder.layers.7.mlp.experts.linear_fc2.weight6 + module.decoder.layers.7.mlp.experts.linear_fc1.weight9 + module.decoder.layers.7.mlp.experts.linear_fc2.weight11 + module.decoder.layers.7.mlp.experts.linear_fc2.weight3 + module.decoder.layers.7.mlp.experts.linear_fc1.weight14 + module.decoder.layers.7.mlp.experts.linear_fc2.weight8 + module.decoder.layers.7.mlp.experts.linear_fc2.weight0 + module.decoder.layers.7.mlp.experts.linear_fc1.weight11 + module.decoder.layers.7.mlp.experts.linear_fc2.weight5 + module.decoder.layers.7.mlp.experts.linear_fc2.weight10 + module.decoder.layers.7.mlp.experts.linear_fc2.weight2 + module.decoder.layers.7.mlp.experts.linear_fc1.weight13 + module.decoder.layers.7.mlp.experts.linear_fc2.weight7 + module.decoder.layers.7.mlp.experts.linear_fc1.weight10 + module.decoder.layers.7.mlp.experts.linear_fc2.weight4 + module.decoder.layers.7.mlp.experts.linear_fc1.weight15 +Params for bucket 76 (40894464 elements, 40894464 padded size): + module.decoder.layers.7.mlp.experts.linear_fc1.weight8 + module.decoder.layers.7.mlp.experts.linear_fc1.weight6 + module.decoder.layers.7.mlp.experts.linear_fc1.weight4 + module.decoder.layers.7.mlp.experts.linear_fc1.weight2 + module.decoder.layers.7.mlp.experts.linear_fc1.weight0 + module.decoder.layers.6.mlp.experts.linear_fc2.weight15 + module.decoder.layers.6.mlp.experts.linear_fc2.weight13 + module.decoder.layers.6.mlp.experts.linear_fc2.weight11 + module.decoder.layers.6.mlp.experts.linear_fc2.weight9 + module.decoder.layers.7.mlp.experts.linear_fc1.weight7 + module.decoder.layers.7.mlp.experts.linear_fc1.weight5 + module.decoder.layers.7.mlp.experts.linear_fc1.weight3 + module.decoder.layers.7.mlp.experts.linear_fc1.weight1 + module.decoder.layers.6.mlp.experts.linear_fc2.weight14 + module.decoder.layers.6.mlp.experts.linear_fc2.weight12 + module.decoder.layers.6.mlp.experts.linear_fc2.weight10 + module.decoder.layers.6.mlp.experts.linear_fc2.weight8 +Params for bucket 77 (40894464 elements, 40894464 padded size): + module.decoder.layers.6.mlp.experts.linear_fc2.weight7 + module.decoder.layers.6.mlp.experts.linear_fc2.weight3 + module.decoder.layers.6.mlp.experts.linear_fc2.weight5 + module.decoder.layers.6.mlp.experts.linear_fc2.weight1 + module.decoder.layers.6.mlp.experts.linear_fc1.weight14 + module.decoder.layers.6.mlp.experts.linear_fc1.weight12 + module.decoder.layers.6.mlp.experts.linear_fc1.weight10 + module.decoder.layers.6.mlp.experts.linear_fc1.weight8 + 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.weight0 + module.decoder.layers.6.mlp.experts.linear_fc1.weight15 + module.decoder.layers.6.mlp.experts.linear_fc1.weight13 + module.decoder.layers.6.mlp.experts.linear_fc1.weight11 + module.decoder.layers.6.mlp.experts.linear_fc1.weight9 + module.decoder.layers.6.mlp.experts.linear_fc1.weight7 +Params for bucket 78 (40894464 elements, 40894464 padded size): + module.decoder.layers.6.mlp.experts.linear_fc1.weight6 + module.decoder.layers.5.mlp.experts.linear_fc2.weight12 + module.decoder.layers.5.mlp.experts.linear_fc2.weight4 + module.decoder.layers.6.mlp.experts.linear_fc1.weight3 + module.decoder.layers.5.mlp.experts.linear_fc2.weight9 + module.decoder.layers.6.mlp.experts.linear_fc1.weight0 + module.decoder.layers.5.mlp.experts.linear_fc2.weight14 + module.decoder.layers.5.mlp.experts.linear_fc2.weight6 + module.decoder.layers.6.mlp.experts.linear_fc1.weight5 + module.decoder.layers.5.mlp.experts.linear_fc2.weight11 + module.decoder.layers.6.mlp.experts.linear_fc1.weight2 + module.decoder.layers.5.mlp.experts.linear_fc2.weight8 + module.decoder.layers.5.mlp.experts.linear_fc2.weight13 + module.decoder.layers.5.mlp.experts.linear_fc2.weight5 + module.decoder.layers.6.mlp.experts.linear_fc1.weight4 + module.decoder.layers.5.mlp.experts.linear_fc2.weight10 + module.decoder.layers.6.mlp.experts.linear_fc1.weight1 + module.decoder.layers.5.mlp.experts.linear_fc2.weight15 + module.decoder.layers.5.mlp.experts.linear_fc2.weight7 +Params for bucket 79 (40894464 elements, 40894464 padded size): + module.decoder.layers.5.mlp.experts.linear_fc1.weight15 + module.decoder.layers.5.mlp.experts.linear_fc2.weight0 + module.decoder.layers.5.mlp.experts.linear_fc2.weight2 + module.decoder.layers.5.mlp.experts.linear_fc1.weight13 + module.decoder.layers.5.mlp.experts.linear_fc1.weight11 + module.decoder.layers.5.mlp.experts.linear_fc1.weight9 + module.decoder.layers.5.mlp.experts.linear_fc1.weight7 + module.decoder.layers.5.mlp.experts.linear_fc1.weight5 + 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_fc1.weight14 + module.decoder.layers.5.mlp.experts.linear_fc1.weight12 + module.decoder.layers.5.mlp.experts.linear_fc1.weight10 + module.decoder.layers.5.mlp.experts.linear_fc1.weight8 + module.decoder.layers.5.mlp.experts.linear_fc1.weight6 +Params for bucket 80 (40894464 elements, 40894464 padded size): + module.decoder.layers.4.mlp.experts.linear_fc2.weight15 + module.decoder.layers.4.mlp.experts.linear_fc2.weight7 + module.decoder.layers.5.mlp.experts.linear_fc1.weight4 + module.decoder.layers.4.mlp.experts.linear_fc2.weight12 + module.decoder.layers.4.mlp.experts.linear_fc2.weight4 + module.decoder.layers.5.mlp.experts.linear_fc1.weight1 + module.decoder.layers.4.mlp.experts.linear_fc2.weight9 + module.decoder.layers.4.mlp.experts.linear_fc2.weight1 + module.decoder.layers.4.mlp.experts.linear_fc2.weight14 + module.decoder.layers.4.mlp.experts.linear_fc2.weight6 + module.decoder.layers.5.mlp.experts.linear_fc1.weight3 + module.decoder.layers.4.mlp.experts.linear_fc2.weight11 + module.decoder.layers.4.mlp.experts.linear_fc2.weight3 + module.decoder.layers.5.mlp.experts.linear_fc1.weight0 + module.decoder.layers.4.mlp.experts.linear_fc2.weight8 + module.decoder.layers.4.mlp.experts.linear_fc2.weight0 + module.decoder.layers.4.mlp.experts.linear_fc2.weight13 + module.decoder.layers.4.mlp.experts.linear_fc2.weight5 + module.decoder.layers.5.mlp.experts.linear_fc1.weight2 + module.decoder.layers.4.mlp.experts.linear_fc2.weight10 + module.decoder.layers.4.mlp.experts.linear_fc2.weight2 +Params for bucket 81 (40894464 elements, 40894464 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight14 + module.decoder.layers.4.mlp.experts.linear_fc1.weight12 + module.decoder.layers.4.mlp.experts.linear_fc1.weight10 + module.decoder.layers.4.mlp.experts.linear_fc1.weight8 + module.decoder.layers.4.mlp.experts.linear_fc1.weight6 + module.decoder.layers.4.mlp.experts.linear_fc1.weight4 + module.decoder.layers.4.mlp.experts.linear_fc1.weight15 + module.decoder.layers.4.mlp.experts.linear_fc1.weight13 + module.decoder.layers.4.mlp.experts.linear_fc1.weight11 + module.decoder.layers.4.mlp.experts.linear_fc1.weight9 + module.decoder.layers.4.mlp.experts.linear_fc1.weight7 + module.decoder.layers.4.mlp.experts.linear_fc1.weight5 + module.decoder.layers.4.mlp.experts.linear_fc1.weight3 +Params for bucket 82 (40894464 elements, 40894464 padded size): + module.decoder.layers.4.mlp.experts.linear_fc1.weight2 + module.decoder.layers.3.mlp.experts.linear_fc2.weight9 + module.decoder.layers.3.mlp.experts.linear_fc2.weight1 + module.decoder.layers.3.mlp.experts.linear_fc2.weight14 + module.decoder.layers.3.mlp.experts.linear_fc2.weight6 + module.decoder.layers.3.mlp.experts.linear_fc2.weight11 + module.decoder.layers.3.mlp.experts.linear_fc2.weight3 + module.decoder.layers.3.mlp.experts.linear_fc1.weight14 + module.decoder.layers.4.mlp.experts.linear_fc1.weight1 + module.decoder.layers.3.mlp.experts.linear_fc2.weight8 + module.decoder.layers.3.mlp.experts.linear_fc2.weight0 + module.decoder.layers.3.mlp.experts.linear_fc2.weight13 + module.decoder.layers.3.mlp.experts.linear_fc2.weight5 + module.decoder.layers.3.mlp.experts.linear_fc2.weight10 + module.decoder.layers.3.mlp.experts.linear_fc2.weight2 + module.decoder.layers.3.mlp.experts.linear_fc2.weight15 + module.decoder.layers.4.mlp.experts.linear_fc1.weight0 + module.decoder.layers.3.mlp.experts.linear_fc2.weight7 + module.decoder.layers.3.mlp.experts.linear_fc2.weight12 + module.decoder.layers.3.mlp.experts.linear_fc2.weight4 + module.decoder.layers.3.mlp.experts.linear_fc1.weight15 +Params for bucket 83 (40894464 elements, 40894464 padded size): + module.decoder.layers.3.mlp.experts.linear_fc1.weight10 + module.decoder.layers.3.mlp.experts.linear_fc1.weight12 + module.decoder.layers.3.mlp.experts.linear_fc1.weight8 + module.decoder.layers.3.mlp.experts.linear_fc1.weight6 + module.decoder.layers.3.mlp.experts.linear_fc1.weight4 + module.decoder.layers.3.mlp.experts.linear_fc1.weight2 + module.decoder.layers.3.mlp.experts.linear_fc1.weight13 + module.decoder.layers.3.mlp.experts.linear_fc1.weight11 + module.decoder.layers.3.mlp.experts.linear_fc1.weight9 + module.decoder.layers.3.mlp.experts.linear_fc1.weight7 + 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.weight1 +Params for bucket 84 (40894464 elements, 40894464 padded size): + module.decoder.layers.2.mlp.experts.linear_fc2.weight9 + module.decoder.layers.2.mlp.experts.linear_fc2.weight1 + module.decoder.layers.2.mlp.experts.linear_fc2.weight14 + module.decoder.layers.2.mlp.experts.linear_fc2.weight6 + module.decoder.layers.2.mlp.experts.linear_fc1.weight15 + module.decoder.layers.2.mlp.experts.linear_fc2.weight11 + module.decoder.layers.2.mlp.experts.linear_fc2.weight3 + module.decoder.layers.2.mlp.experts.linear_fc1.weight12 + module.decoder.layers.2.mlp.experts.linear_fc2.weight8 + module.decoder.layers.2.mlp.experts.linear_fc2.weight0 + module.decoder.layers.3.mlp.experts.linear_fc1.weight0 + module.decoder.layers.2.mlp.experts.linear_fc2.weight13 + module.decoder.layers.2.mlp.experts.linear_fc2.weight5 + module.decoder.layers.2.mlp.experts.linear_fc1.weight14 + module.decoder.layers.2.mlp.experts.linear_fc2.weight10 + module.decoder.layers.2.mlp.experts.linear_fc2.weight2 + module.decoder.layers.2.mlp.experts.linear_fc2.weight15 + module.decoder.layers.2.mlp.experts.linear_fc2.weight7 + module.decoder.layers.2.mlp.experts.linear_fc2.weight12 + module.decoder.layers.2.mlp.experts.linear_fc2.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight13 +Params for bucket 85 (40894464 elements, 40894464 padded size): + module.decoder.layers.2.mlp.experts.linear_fc1.weight10 + module.decoder.layers.2.mlp.experts.linear_fc1.weight8 + module.decoder.layers.2.mlp.experts.linear_fc1.weight6 + module.decoder.layers.2.mlp.experts.linear_fc1.weight4 + module.decoder.layers.2.mlp.experts.linear_fc1.weight2 + module.decoder.layers.2.mlp.experts.linear_fc1.weight0 + module.decoder.layers.1.mlp.experts.linear_fc2.weight15 + module.decoder.layers.2.mlp.experts.linear_fc1.weight11 + module.decoder.layers.2.mlp.experts.linear_fc1.weight9 + module.decoder.layers.2.mlp.experts.linear_fc1.weight7 + module.decoder.layers.2.mlp.experts.linear_fc1.weight5 + module.decoder.layers.2.mlp.experts.linear_fc1.weight3 + module.decoder.layers.2.mlp.experts.linear_fc1.weight1 + module.decoder.layers.1.mlp.experts.linear_fc2.weight14 +Params for bucket 86 (40894464 elements, 40894464 padded size): + module.decoder.layers.1.mlp.experts.linear_fc2.weight9 + module.decoder.layers.1.mlp.experts.linear_fc2.weight1 + module.decoder.layers.1.mlp.experts.linear_fc1.weight11 + module.decoder.layers.1.mlp.experts.linear_fc2.weight6 + module.decoder.layers.1.mlp.experts.linear_fc2.weight11 + module.decoder.layers.1.mlp.experts.linear_fc2.weight3 + module.decoder.layers.1.mlp.experts.linear_fc1.weight13 + module.decoder.layers.1.mlp.experts.linear_fc2.weight8 + module.decoder.layers.1.mlp.experts.linear_fc2.weight0 + module.decoder.layers.1.mlp.experts.linear_fc1.weight10 + module.decoder.layers.1.mlp.experts.linear_fc2.weight13 + module.decoder.layers.1.mlp.experts.linear_fc2.weight5 + module.decoder.layers.1.mlp.experts.linear_fc1.weight15 + module.decoder.layers.1.mlp.experts.linear_fc2.weight10 + module.decoder.layers.1.mlp.experts.linear_fc2.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight12 + module.decoder.layers.1.mlp.experts.linear_fc2.weight7 + module.decoder.layers.1.mlp.experts.linear_fc2.weight12 + module.decoder.layers.1.mlp.experts.linear_fc2.weight4 + module.decoder.layers.1.mlp.experts.linear_fc1.weight14 +Params for bucket 87 (40894464 elements, 40894464 padded size): + module.decoder.layers.1.mlp.experts.linear_fc1.weight9 + module.decoder.layers.1.mlp.experts.linear_fc1.weight7 + 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.weight1 + module.decoder.layers.0.mlp.experts.linear_fc2.weight15 + module.decoder.layers.0.mlp.experts.linear_fc2.weight13 + module.decoder.layers.0.mlp.experts.linear_fc2.weight11 + module.decoder.layers.1.mlp.experts.linear_fc1.weight6 + module.decoder.layers.1.mlp.experts.linear_fc1.weight8 + module.decoder.layers.1.mlp.experts.linear_fc1.weight4 + module.decoder.layers.1.mlp.experts.linear_fc1.weight2 + module.decoder.layers.1.mlp.experts.linear_fc1.weight0 + module.decoder.layers.0.mlp.experts.linear_fc2.weight14 + module.decoder.layers.0.mlp.experts.linear_fc2.weight12 + module.decoder.layers.0.mlp.experts.linear_fc2.weight10 +Params for bucket 88 (40894464 elements, 40894464 padded size): + module.decoder.layers.0.mlp.experts.linear_fc2.weight9 + 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 + module.decoder.layers.0.mlp.experts.linear_fc2.weight1 + module.decoder.layers.0.mlp.experts.linear_fc1.weight15 + module.decoder.layers.0.mlp.experts.linear_fc1.weight13 + module.decoder.layers.0.mlp.experts.linear_fc1.weight11 + module.decoder.layers.0.mlp.experts.linear_fc1.weight9 + module.decoder.layers.0.mlp.experts.linear_fc2.weight8 + 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.weight0 + module.decoder.layers.0.mlp.experts.linear_fc1.weight14 + module.decoder.layers.0.mlp.experts.linear_fc1.weight12 + module.decoder.layers.0.mlp.experts.linear_fc1.weight10 + module.decoder.layers.0.mlp.experts.linear_fc1.weight8 +Params for bucket 89 (25165824 elements, 25165824 padded size): + module.decoder.layers.0.mlp.experts.linear_fc1.weight7 + 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.weight1 + module.decoder.layers.0.mlp.experts.linear_fc1.weight4 + module.decoder.layers.0.mlp.experts.linear_fc1.weight6 + module.decoder.layers.0.mlp.experts.linear_fc1.weight2 + module.decoder.layers.0.mlp.experts.linear_fc1.weight0 +INFO:megatron.core.optimizer:Setting up optimizer with config OptimizerConfig(optimizer='adam', lr=0.0001, min_lr=1e-05, decoupled_lr=None, decoupled_min_lr=None, weight_decay=0.1, fp8_recipe='mxfp8', fp16=False, bf16=True, reuse_grad_buf_for_mxfp8_param_ag=False, params_dtype=torch.bfloat16, use_precision_aware_optimizer=False, store_param_remainders=True, main_grads_dtype=torch.float32, main_params_dtype=torch.float32, exp_avg_dtype=torch.float32, exp_avg_sq_dtype=torch.float32, loss_scale=None, initial_loss_scale=4294967296, min_loss_scale=1.0, loss_scale_window=1000, hysteresis=2, adam_beta1=0.9, adam_beta2=0.999, adam_eps=1e-08, sgd_momentum=0.9, use_distributed_optimizer=True, overlap_param_gather_with_optimizer_step=False, optimizer_cpu_offload=False, optimizer_offload_fraction=1.0, use_torch_optimizer_for_cpu_offload=False, overlap_cpu_optimizer_d2h_h2d=False, pin_cpu_grads=True, pin_cpu_params=True, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=, config_logger_dir='') +INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine +[after model, optimizer, and learning rate scheduler are built] datetime: 2026-08-26 18:34:15 +> building train, validation, and test datasets ... + > datasets target sizes (minimum size): + train: 96 + 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=(96, 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 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-document_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-sample_index.npy +INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-shuffle_index.npy +INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073 +> finished creating GPT datasets ... +[after dataloaders are built] datetime: 2026-08-26 18:34:15 +done with setup ... +training ... +(min, max) time across ranks (ms): + model-and-optimizer-setup ......................: (4609.48, 4627.77) + train/valid/test-data-iterators-setup ..........: (73.34, 86.83) +Overwriting rerun_state_machine.current_iteration from -1 to 0... +[before the start of training step] datetime: 2026-08-26 18:34:15 +/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:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future. + warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.") + [2026-08-26 18:34:42] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 26858.6 | throughput per GPU (TFLOP/s/GPU): 3.3 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.095286E+01 | loss scale: 1.0 | grad norm: 54.368 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | +Number of parameters in transformer block in billions: 29.90 +Number of parameters in embedding layers in billions: 0.21 +Total number of parameters in billions: 30.10 +Number of parameters in most loaded shard in billions: 30.1034 +Theoretical memory footprints: weight and optimizer=215316.59 MB +[Rank 0] (after 1 iterations) memory (MB) | allocated: 72365.84375 | max allocated: 78646.15576171875 | reserved: 79250.0 | max reserved: 79250.0 +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") +/workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future. + warnings.warn("The use_host_meta is deprecated, it will be removed in the future.") + [2026-08-26 18:34:43] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 1130.6 | throughput per GPU (TFLOP/s/GPU): 78.8 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.095473E+01 | loss scale: 1.0 | grad norm: 56.501 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:34:44] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 935.7 | throughput per GPU (TFLOP/s/GPU): 95.2 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 1.008597E+01 | loss scale: 1.0 | grad norm: 180.006 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | +Capture CUDA graph for training!!! +CUDA graph capture done!!! + [2026-08-26 18:34:47] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 3022.6 | throughput per GPU (TFLOP/s/GPU): 29.5 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 9.604107E+00 | loss scale: 1.0 | grad norm: 8.753 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:34:47] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 360.4 | throughput per GPU (TFLOP/s/GPU): 247.3 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 9.307376E+00 | loss scale: 1.0 | grad norm: 4.102 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:34:48] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 358.1 | throughput per GPU (TFLOP/s/GPU): 248.9 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 9.292999E+00 | loss scale: 1.0 | grad norm: 7.055 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:34:48] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 358.8 | throughput per GPU (TFLOP/s/GPU): 248.4 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 9.114014E+00 | loss scale: 1.0 | grad norm: 3.643 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:34:48] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 357.5 | throughput per GPU (TFLOP/s/GPU): 249.3 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.917605E+00 | loss scale: 1.0 | grad norm: 3.419 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:34:49] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 357.5 | throughput per GPU (TFLOP/s/GPU): 249.3 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.853426E+00 | loss scale: 1.0 | grad norm: 3.281 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:34:49] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 355.9 | throughput per GPU (TFLOP/s/GPU): 250.4 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.812331E+00 | loss scale: 1.0 | grad norm: 3.303 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:34:49] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 354.2 | throughput per GPU (TFLOP/s/GPU): 251.6 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.799855E+00 | loss scale: 1.0 | grad norm: 3.173 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | + [2026-08-26 18:34:50] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 352.7 | throughput per GPU (TFLOP/s/GPU): 252.7 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.769467E+00 | loss scale: 1.0 | grad norm: 3.069 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34560 | min safe capacity: 1.0547 | +[after training is done] datetime: 2026-08-26 18:34:50 +[rank0]:[W826 18:34:51.461828600 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank6]:[W826 18:34:52.939656335 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank1]:[W826 18:34:52.939947669 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank4]:[W826 18:34:52.941637707 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank2]:[W826 18:34:52.942328028 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank3]:[W826 18:34:52.943362085 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank5]:[W826 18:34:53.048786142 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +[rank7]:[W826 18:34:53.057144820 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator()) +W0826 18:34:55.411000 525651 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers +W0826 18:34:55.411000 525651 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 525740 closing signal SIGTERM +W0826 18:34:55.412000 525651 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 525741 closing signal SIGTERM +W0826 18:34:55.414000 525651 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 525742 closing signal SIGTERM +W0826 18:34:55.415000 525651 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 525743 closing signal SIGTERM +W0826 18:34:55.416000 525651 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 525744 closing signal SIGTERM +W0826 18:34:55.417000 525651 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 525745 closing signal SIGTERM +W0826 18:34:55.419000 525651 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 525746 closing signal SIGTERM +W0826 18:34:55.420000 525651 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 525747 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 525651 got signal: 15 diff --git a/ultraep/tierb/tierb_gbs8/results.tsv b/ultraep/tierb/tierb_gbs8/results.tsv index d85c3f6eae5bc896ea80e5dff32c7be03d9367a4..2046ce87d73d5fdbef6951df337e6e882c20cec4 100644 --- a/ultraep/tierb/tierb_gbs8/results.tsv +++ b/ultraep/tierb/tierb_gbs8/results.tsv @@ -1,2 +1,11 @@ config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir QWEN3_235B ultraep matched 16 8192 8 1.083984375 none 1 OK 297.73 295.80 122229 /workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep1 +QWEN3_235B ultraep matched 16 8192 8 1.083984375 none 2 OK 297.91 295.80 122229 /workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep2 +QWEN3_235B ultraep matched 16 8192 8 1.083984375 none 3 OK 297.96 295.70 122229 /workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p083984375/rep3 +QWEN3_30B ultraep matched 16 4096 8 1.11328125 none 1 OK 352.90 351.40 120129 /workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep1 +QWEN3_30B ultraep matched 16 4096 8 1.11328125 none 2 OK 353.63 351.40 120129 /workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep2 +QWEN3_30B ultraep matched 16 4096 8 1.11328125 none 3 OK 354.55 351.40 120129 /workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p11328125/rep3 +DSV2_LITE_L24 ultraep matched 8 8192 8 1.0755208333333333 none 1 OK 579.87 578.60 142231 /workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep1 +DSV2_LITE_L24 ultraep matched 8 8192 8 1.0755208333333333 none 2 OK 580.05 578.50 142231 /workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep2 +DSV2_LITE_L24 ultraep matched 8 8192 8 1.0755208333333333 none 3 OK 580.05 578.50 142231 /workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p0755208333333333/rep3 +QWEN2_57B ultraep matched 8 4096 8 1.08984375 none 1 OK 192.64 184.40 98637 /workspace/tierb_study_20260826_144644/ultraep/tierb/tierb_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p08984375/rep1