AbhishekGhosh1998 commited on
Commit
f17e1ed
·
verified ·
1 Parent(s): 7c70a88

granularity results granularity_next_b200/tierb_models/traces/DEEPSEEK_V3_L4/ultraep/tierbtrace @ 2026-09-21T07:26:08Z from sp-aiinfra-interactive-aghosh-8-mb99k-8xb200-moe-cg-0

Browse files
.gitattributes CHANGED
@@ -110,3 +110,4 @@ granularity_next_b200/tiera_topk_trace/moonep/tierbtrace/tierbtrace_gbs16/G4_K10
110
  granularity_next_b200/tiera_topk_trace/moonep/tierbtrace/tierbtrace_gbs16/G4_K16__moonep__matched__seq4096__gbs16__capnone__async/nsys/G4_K16__moonep__matched__seq4096__gbs16__capnone__async.nsys-rep filter=lfs diff=lfs merge=lfs -text
111
  granularity_next_b200/tiera_topk_trace/moonep/tierbtrace/tierbtrace_gbs16/G4_K8__moonep__matched__seq4096__gbs16__capnone__async/nsys/G4_K8__moonep__matched__seq4096__gbs16__capnone__async.nsys-rep filter=lfs diff=lfs merge=lfs -text
112
  granularity_next_b200/tierb_models/traces/DEEPSEEK_V3_L4/echo/tierbtrace/tierbtrace_gbs8/DEEPSEEK_V3_L4__echo__matched__seq4096__gbs8__cap1p13671875/nsys/DEEPSEEK_V3_L4__echo__matched__seq4096__gbs8__cap1p13671875.nsys-rep filter=lfs diff=lfs merge=lfs -text
 
 
110
  granularity_next_b200/tiera_topk_trace/moonep/tierbtrace/tierbtrace_gbs16/G4_K16__moonep__matched__seq4096__gbs16__capnone__async/nsys/G4_K16__moonep__matched__seq4096__gbs16__capnone__async.nsys-rep filter=lfs diff=lfs merge=lfs -text
111
  granularity_next_b200/tiera_topk_trace/moonep/tierbtrace/tierbtrace_gbs16/G4_K8__moonep__matched__seq4096__gbs16__capnone__async/nsys/G4_K8__moonep__matched__seq4096__gbs16__capnone__async.nsys-rep filter=lfs diff=lfs merge=lfs -text
112
  granularity_next_b200/tierb_models/traces/DEEPSEEK_V3_L4/echo/tierbtrace/tierbtrace_gbs8/DEEPSEEK_V3_L4__echo__matched__seq4096__gbs8__cap1p13671875/nsys/DEEPSEEK_V3_L4__echo__matched__seq4096__gbs8__cap1p13671875.nsys-rep filter=lfs diff=lfs merge=lfs -text
113
+ granularity_next_b200/tierb_models/traces/DEEPSEEK_V3_L4/ultraep/tierbtrace/tierbtrace_gbs8/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875/nsys/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875.nsys-rep filter=lfs diff=lfs merge=lfs -text
granularity_next_b200/tierb_models/traces/DEEPSEEK_V3_L4/ultraep/tierbtrace/tierbtrace_gbs8/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875/nsys/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875.nsys-rep ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:06871820e0b37e6b4280095c8957a20f08122917f270d29f4213199e1f64df37
3
+ size 3501341
granularity_next_b200/tierb_models/traces/DEEPSEEK_V3_L4/ultraep/tierbtrace/tierbtrace_gbs8/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875/nsys/run.env ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=ultraep
2
+ config=DEEPSEEK_V3_L4
3
+ family=tierB
4
+ source_model=DeepSeek-V3
5
+ standard_num_layers=61
6
+ base_rung=
7
+ repo=/workspace/Megatron-UltraEP
8
+ branch=ultraep-bracing-regions-through-kernels
9
+ commit=03d57ef7
10
+ num_layers=4
11
+ hidden=7168
12
+ ffn_hidden=18432
13
+ moe_ffn_hidden=2048
14
+ num_experts=256
15
+ topk=8
16
+ experts_per_rank=32
17
+ expert_params=44040192
18
+ heads=128
19
+ kv_groups=128
20
+ kv_channels=128
21
+ qk_layernorm=1
22
+ shared_ffn=2048
23
+ shared_expert_overlap=0
24
+ moe_token_dispatcher=flex
25
+ mla=1
26
+ q_lora_rank=1536
27
+ kv_lora_rank=512
28
+ qk_head_dim=128
29
+ qk_pos_emb_head_dim=64
30
+ v_head_dim=128
31
+ dense_first_layers=3
32
+ moe_layers=1
33
+ rotary_base=10000
34
+ rotary_percent=1.0
35
+ rotary_scaling=40
36
+ mscale=1.0
37
+ mscale_all_dim=1.0
38
+ qkv_bias=0
39
+ router_balancing=seq_aux_loss
40
+ aux_loss_coeff=1e-4
41
+ router_pre_softmax=0
42
+ router_fusion=0
43
+ router_dtype=fp32
44
+ router_groups=8
45
+ router_group_topk=4
46
+ router_score_function=sigmoid
47
+ router_expert_bias=1
48
+ router_bias_update_rate=1e-3
49
+ router_topk_scaling_factor=2.5
50
+ shared_expert_gate=0
51
+ apply_layernorm_1p=0
52
+ attention_output_gate=0
53
+ linear_attention_type=
54
+ linear_attention_freq=
55
+ linear_conv_kernel_dim=
56
+ linear_key_head_dim=
57
+ linear_value_head_dim=
58
+ linear_num_key_heads=
59
+ linear_num_value_heads=
60
+ mtp_num_layers=0
61
+ mtp_loss_scaling_factor=0.1
62
+ no_weight_decay_cond_type=
63
+ max_position_embeddings=4096
64
+ init_method_std=0.02
65
+ device_initiated_grouped_gemm=1
66
+ norm_epsilon=1e-6
67
+ seq_length=4096
68
+ micro_batch_size=1
69
+ global_batch_size=8
70
+ num_microbatches=1
71
+ assignments_per_rank=32768
72
+ mean_tokens_per_expert=1024
73
+ slot_policy=matched
74
+ slots_per_rank=32
75
+ capacity_mode=static
76
+ capacity=1.13671875
77
+ cuda_graph=1
78
+ probe_mode=0
79
+ async_grad_reduce=0
80
+ branch_role=timing
81
+ memory_dump_mode=none
82
+ memory_snapshot_path=
83
+ cuda_graph_warmup_steps=3
84
+ recompute=none
85
+ force_balance=0
86
+ train_iters=10
87
+ ep_size=8
88
+ echo_dispatch_overlap=1
89
+ echo_dispatch_recompute=0
90
+ ultraep_quota_min_tokens=128
91
+ ultraep_weight_sync_join=
92
+ profile_enable=1
93
+ kernel_trace=1
94
+ fp8_mode=mxfp8
95
+ fp8_recipe=mxfp8
96
+ training_seed=1234
97
+ learning_rate=1e-4
98
+ min_learning_rate=1.0e-5
99
+ lr_warmup_iters=5
100
+ state_dump_enable=0
101
+ state_dump_dir=
102
+ moonep_assert_no_pending_reduce=0
granularity_next_b200/tierb_models/traces/DEEPSEEK_V3_L4/ultraep/tierbtrace/tierbtrace_gbs8/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875/nsys/train.log ADDED
@@ -0,0 +1,1064 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [07:23:52] WARN: ultraep-bracing-regions-through-kernels lacks --moe-router-fusion; DEEPSEEK_V3_L4 runs unfused (timing only, routing identical)
2
+ ==============================================================
3
+ DEEPSEEK_V3_L4 / ultraep (slots/rank=32 policy=matched)
4
+ E=256 K=8 H'=2048 m=32
5
+ L=4 seq=4096 gbs=8 (1 microbatches)
6
+ cap_mode=static cap=1.13671875 cg=1 recompute=none
7
+ memory_dump=none snapshots=/workspace/granularity_next_b200/tierb_models/traces/DEEPSEEK_V3_L4/ultraep/tierbtrace/tierbtrace_gbs8/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875/nsys/memory_snapshots/snapshot.pickle
8
+ repo=/workspace/Megatron-UltraEP @ 03d57ef7
9
+ ==============================================================
10
+ W0921 07:23:54.053000 434771 torch/distributed/run.py:766]
11
+ W0921 07:23:54.053000 434771 torch/distributed/run.py:766] *****************************************
12
+ W0921 07:23:54.053000 434771 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.
13
+ W0921 07:23:54.053000 434771 torch/distributed/run.py:766] *****************************************
14
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
15
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
16
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
17
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
18
+ 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
19
+ WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer
20
+ Number of virtual stages per pipeline stage: None
21
+ accumulate and all-reduce gradients in fp32 for bfloat16 data type.
22
+ using torch.bfloat16 for parameters ...
23
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
24
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
25
+ ------------------------ arguments ------------------------
26
+ account_for_embedding_in_pipeline_split ......... False
27
+ account_for_loss_in_pipeline_split .............. False
28
+ accumulate_allreduce_grads_in_fp32 .............. True
29
+ adam_beta1 ...................................... 0.9
30
+ adam_beta2 ...................................... 0.999
31
+ adam_eps ........................................ 1e-08
32
+ add_bias_linear ................................. False
33
+ add_position_embedding .......................... False
34
+ add_qkv_bias .................................... False
35
+ adlr_autoresume ................................. False
36
+ adlr_autoresume_interval ........................ 1000
37
+ align_grad_reduce ............................... True
38
+ align_param_gather .............................. False
39
+ app_tag_run_name ................................ None
40
+ app_tag_run_version ............................. 0.0.0
41
+ apply_layernorm_1p .............................. False
42
+ apply_query_key_layer_scaling ................... False
43
+ apply_residual_connection_post_layernorm ........ False
44
+ apply_rope_fusion ............................... False
45
+ async_save ...................................... None
46
+ async_tensor_model_parallel_allreduce ........... True
47
+ attention_backend ............................... AttnBackend.auto
48
+ attention_dropout ............................... 0.0
49
+ attention_softmax_in_fp32 ....................... False
50
+ auto_detect_ckpt_format ......................... False
51
+ barrier_with_L1_time ............................ True
52
+ bert_binary_head ................................ True
53
+ bert_embedder_type .............................. megatron
54
+ bert_load ....................................... None
55
+ bf16 ............................................ True
56
+ bias_dropout_fusion ............................. True
57
+ bias_gelu_fusion ................................ False
58
+ bias_swiglu_fusion .............................. True
59
+ biencoder_projection_dim ........................ 0
60
+ biencoder_shared_query_context_model ............ False
61
+ block_data_path ................................. None
62
+ calc_ft_timeouts ................................ False
63
+ calculate_per_token_loss ........................ False
64
+ check_for_large_grads ........................... False
65
+ check_for_nan_in_loss_and_grad .................. False
66
+ check_for_spiky_loss ............................ False
67
+ check_weight_hash_across_dp_replicas_interval ... None
68
+ ckpt_assume_constant_structure .................. False
69
+ ckpt_convert_format ............................. None
70
+ ckpt_convert_save ............................... None
71
+ ckpt_convert_update_legacy_dist_opt_format ...... False
72
+ ckpt_format ..................................... torch_dist
73
+ ckpt_fully_parallel_load ........................ False
74
+ ckpt_fully_parallel_save ........................ True
75
+ ckpt_fully_parallel_save_deprecated ............. False
76
+ ckpt_step ....................................... None
77
+ classes_fraction ................................ 1.0
78
+ clip_grad ....................................... 1.0
79
+ clone_scatter_output_in_embedding ............... True
80
+ config_logger_dir ...............................
81
+ consumed_train_samples .......................... 0
82
+ consumed_valid_samples .......................... 0
83
+ context_parallel_size ........................... 1
84
+ cp_comm_type .................................... ['p2p']
85
+ create_attention_mask_in_dataloader ............. True
86
+ cross_entropy_fusion_impl ....................... native
87
+ cross_entropy_loss_fusion ....................... False
88
+ cuda_graph_scope ................................ full_iteration
89
+ cuda_graph_warmup_steps ......................... 3
90
+ data_args_path .................................. None
91
+ data_cache_path ................................. None
92
+ data_parallel_random_init ....................... False
93
+ data_parallel_sharding_strategy ................. no_shard
94
+ data_parallel_size .............................. 8
95
+ data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document']
96
+ data_per_class_fraction ......................... 1.0
97
+ data_sharding ................................... True
98
+ dataloader_type ................................. single
99
+ ddp_average_in_collective ....................... False
100
+ ddp_bucket_size ................................. None
101
+ ddp_num_buckets ................................. None
102
+ ddp_pad_buckets_for_high_nccl_busbw ............. False
103
+ decoder_first_pipeline_num_layers ............... None
104
+ decoder_last_pipeline_num_layers ................ None
105
+ decoder_num_layers .............................. None
106
+ decoder_seq_length .............................. None
107
+ decoupled_lr .................................... None
108
+ decoupled_min_lr ................................ None
109
+ decrease_batch_size_if_needed ................... False
110
+ defer_embedding_wgrad_compute ................... False
111
+ delay_wgrad_compute ............................. False
112
+ deprecated_use_mcore_models ..................... True
113
+ deterministic_mode .............................. False
114
+ dino_bottleneck_size ............................ 256
115
+ dino_freeze_last_layer .......................... 1
116
+ dino_head_hidden_size ........................... 2048
117
+ dino_local_crops_number ......................... 10
118
+ dino_local_img_size ............................. 96
119
+ dino_norm_last_layer ............................ False
120
+ dino_teacher_temp ............................... 0.07
121
+ dino_warmup_teacher_temp ........................ 0.04
122
+ dino_warmup_teacher_temp_epochs ................. 30
123
+ disable_bf16_reduced_precision_matmul ........... False
124
+ disable_mamba_mem_eff_path ...................... False
125
+ disable_straggler_on_startup .................... False
126
+ dist_ckpt_format_deprecated ..................... None
127
+ dist_ckpt_strictness ............................ assume_ok_unexpected
128
+ distribute_saved_activations .................... False
129
+ distributed_backend ............................. nccl
130
+ distributed_timeout_minutes ..................... 10
131
+ embedding_init_method_std ....................... None
132
+ embedding_path .................................. None
133
+ empty_unused_memory_level ....................... 0
134
+ enable_cuda_graph ............................... True
135
+ enable_experimental ............................. False
136
+ enable_ft_package ............................... False
137
+ enable_gloo_process_groups ...................... True
138
+ enable_msc ...................................... True
139
+ enable_one_logger ............................... True
140
+ encoder_num_layers .............................. 4
141
+ encoder_seq_length .............................. 4096
142
+ end_weight_decay ................................ 0.1
143
+ eod_mask_loss ................................... False
144
+ error_injection_rate ............................ 0
145
+ error_injection_type ............................ transient_error
146
+ eval_interval ................................... 1000000
147
+ eval_iters ...................................... 0
148
+ evidence_data_path .............................. None
149
+ exit_duration_in_mins ........................... None
150
+ exit_interval ................................... None
151
+ exit_on_missing_checkpoint ...................... False
152
+ exit_signal_handler ............................. False
153
+ exp_avg_dtype ................................... torch.float32
154
+ exp_avg_sq_dtype ................................ torch.float32
155
+ expert_model_parallel_size ...................... 8
156
+ expert_tensor_parallel_size ..................... 1
157
+ export_force_local_attention .................... False
158
+ export_kd_cfg ................................... None
159
+ export_kd_teacher_ckpt_format ................... None
160
+ export_kd_teacher_load .......................... None
161
+ export_kv_cache_quant ........................... False
162
+ export_legacy_megatron .......................... False
163
+ export_model_type ............................... GPTModel
164
+ export_moe_apply_probs_on_input ................. False
165
+ export_qk_l2_norm ............................... False
166
+ export_quant_cfg ................................ None
167
+ export_real_quant_cfg ........................... None
168
+ export_te_mcore_model ........................... False
169
+ external_cuda_graph ............................. False
170
+ ffn_hidden_size ................................. 18432
171
+ finetune ........................................ False
172
+ finetune_data_split ............................. train
173
+ finetune_hf_dataset ............................. None
174
+ first_last_layers_bf16 .......................... False
175
+ flash_decode .................................... False
176
+ fp16 ............................................ False
177
+ fp16_lm_cross_entropy ........................... False
178
+ fp32_residual_connection ........................ False
179
+ fp8 ............................................. e4m3
180
+ fp8_amax_compute_algo ........................... most_recent
181
+ fp8_amax_history_len ............................ 1
182
+ fp8_interval .................................... 1
183
+ fp8_margin ...................................... 0
184
+ fp8_param_gather ................................ False
185
+ fp8_recipe ...................................... mxfp8
186
+ fp8_wgrad ....................................... True
187
+ fsdp_double_buffer .............................. False
188
+ global_batch_size ............................... 8
189
+ grad_reduce_in_bf16 ............................. False
190
+ gradient_accumulation_fusion .................... True
191
+ gradient_reduce_div_fusion ...................... True
192
+ group_query_attention ........................... False
193
+ head_lr_mult .................................... 1.0
194
+ heterogeneous_layers_config_encoded_json ........ None
195
+ heterogeneous_layers_config_path ................ None
196
+ hidden_dropout .................................. 0.0
197
+ hidden_size ..................................... 7168
198
+ hierarchical_context_parallel_sizes ............. None
199
+ high_priority_stream_groups ..................... []
200
+ hybrid_attention_ratio .......................... 0.0
201
+ hybrid_mlp_ratio ................................ 0.0
202
+ hybrid_override_pattern ......................... None
203
+ hysteresis ...................................... 2
204
+ ict_head_size ................................... None
205
+ ict_load ........................................ None
206
+ img_h ........................................... 224
207
+ img_w ........................................... 224
208
+ indexer_batch_size .............................. 128
209
+ indexer_log_interval ............................ 1000
210
+ inference_batch_times_seqlen_threshold .......... -1
211
+ inference_dynamic_batching ...................... False
212
+ inference_dynamic_batching_buffer_guaranteed_fraction 0.2
213
+ inference_dynamic_batching_buffer_overflow_factor None
214
+ inference_dynamic_batching_buffer_size_gb ....... 40.0
215
+ inference_dynamic_batching_chunk_size ........... 256
216
+ inference_dynamic_batching_max_requests_override None
217
+ inference_dynamic_batching_max_tokens_override .. None
218
+ inference_dynamic_batching_num_cuda_graphs ...... 16
219
+ inference_max_batch_size ........................ 8
220
+ inference_max_seq_length ........................ 2560
221
+ inference_rng_tracker ........................... False
222
+ init_method_std ................................. 0.02
223
+ init_method_xavier_uniform ...................... False
224
+ init_model_with_meta_device ..................... False
225
+ initial_loss_scale .............................. 4294967296
226
+ inprocess_active_world_size ..................... 8
227
+ inprocess_barrier_timeout ....................... 120
228
+ inprocess_completion_timeout .................... 120
229
+ inprocess_empty_cuda_cache ...................... False
230
+ inprocess_granularity ........................... node
231
+ inprocess_hard_timeout .......................... 90
232
+ inprocess_heartbeat_interval .................... 30
233
+ inprocess_heartbeat_timeout ..................... 60
234
+ inprocess_last_call_wait ........................ 1
235
+ inprocess_max_iterations ........................ None
236
+ inprocess_monitor_process_interval .............. 1.0
237
+ inprocess_monitor_thread_interval ............... 1.0
238
+ inprocess_progress_watchdog_interval ............ 1.0
239
+ inprocess_restart ............................... False
240
+ inprocess_soft_timeout .......................... 60
241
+ inprocess_termination_grace_time ................ 1
242
+ is_hybrid_model ................................. False
243
+ iter_per_epoch .................................. 1250
244
+ iterations_to_skip .............................. []
245
+ keep_fp8_transpose_cache_when_using_custom_fsdp . False
246
+ kitchen_config_file ............................. None
247
+ kitchen_recipe_number ........................... None
248
+ kv_channels ..................................... 128
249
+ kv_lora_rank .................................... 512
250
+ lazy_mpu_init ................................... None
251
+ load ............................................ None
252
+ load_main_params_from_ckpt ...................... None
253
+ load_model_opt_format ........................... False
254
+ local_rank ...................................... 0
255
+ log_energy ...................................... False
256
+ log_interval .................................... 1
257
+ log_loss_scale_to_tensorboard ................... True
258
+ log_memory_to_tensorboard ....................... False
259
+ log_num_zeros_in_grad ........................... False
260
+ log_params_norm ................................. False
261
+ log_progress .................................... False
262
+ log_straggler ................................... False
263
+ log_throughput .................................. True
264
+ log_timers_to_tensorboard ....................... False
265
+ log_validation_ppl_to_tensorboard ............... False
266
+ log_world_size_to_tensorboard ................... False
267
+ logging_level ................................... None
268
+ loss_scale ...................................... None
269
+ loss_scale_window ............................... 1000
270
+ lr .............................................. 0.0001
271
+ lr_decay_iters .................................. 8
272
+ lr_decay_samples ................................ None
273
+ lr_decay_style .................................. cosine
274
+ lr_warmup_fraction .............................. None
275
+ lr_warmup_init .................................. 0.0
276
+ lr_warmup_iters ................................. 5
277
+ lr_warmup_samples ............................... 0
278
+ lr_wsd_decay_iters .............................. None
279
+ lr_wsd_decay_samples ............................ None
280
+ lr_wsd_decay_style .............................. exponential
281
+ main_grads_dtype ................................ torch.float32
282
+ main_params_dtype ............................... torch.float32
283
+ make_vocab_size_divisible_by .................... 128
284
+ mamba_head_dim .................................. 64
285
+ mamba_num_groups ................................ 8
286
+ mamba_num_heads ................................. None
287
+ mamba_state_dim ................................. 128
288
+ manual_gc ....................................... False
289
+ manual_gc_eval .................................. True
290
+ manual_gc_interval .............................. 0
291
+ mask_factor ..................................... 1.0
292
+ mask_prob ....................................... 0.15
293
+ mask_type ....................................... random
294
+ masked_softmax_fusion ........................... False
295
+ max_position_embeddings ......................... 4096
296
+ max_tokens_to_oom ............................... 12000
297
+ memory_snapshot_path ............................ snapshot.pickle
298
+ merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt
299
+ micro_batch_size ................................ 1
300
+ microbatch_group_size_per_vp_stage .............. None
301
+ mid_level_dataset_surplus ....................... 0.005
302
+ min_loss_scale .................................. 1.0
303
+ min_lr .......................................... 1e-05
304
+ mlp_chunks_for_prefill .......................... 1
305
+ mmap_bin_files .................................. True
306
+ mock_data ....................................... False
307
+ moe_apply_probs_on_input ........................ False
308
+ moe_aux_loss_coeff .............................. 0.0001
309
+ moe_capacity_monitor ............................ True
310
+ moe_deepep_num_sms .............................. 20
311
+ moe_echo_enable_random_offloading ............... False
312
+ moe_echo_expert_dispatch_overlap ................ False
313
+ moe_echo_expert_dispatcher_type ................. hybridep
314
+ moe_echo_recompute_expert_dispatch .............. False
315
+ moe_enable_deepep ............................... False
316
+ moe_enable_echo ................................. False
317
+ moe_enable_hybridep ............................. True
318
+ moe_enable_ultraep .............................. True
319
+ moe_expert_capacity_factor ...................... None
320
+ moe_extended_tp ................................. False
321
+ moe_ffn_hidden_size ............................. 2048
322
+ moe_grouped_gemm ................................ True
323
+ moe_hybridep_num_sms ............................ 16
324
+ moe_input_jitter_eps ............................ None
325
+ moe_layer_freq .................................. [0, 0, 0, 1]
326
+ moe_layer_recompute ............................. False
327
+ moe_num_echo_experts ............................ None
328
+ moe_num_redundant_experts_per_rank .............. 32
329
+ moe_pad_expert_input_to_capacity ................ False
330
+ moe_per_layer_logging ........................... False
331
+ moe_permute_fusion .............................. False
332
+ moe_received_token_capacity ..................... 1.13671875
333
+ moe_router_bias_update_rate ..................... 0.001
334
+ moe_router_dtype ................................ fp32
335
+ moe_router_enable_expert_bias ................... True
336
+ moe_router_force_load_balancing ................. False
337
+ moe_router_group_topk ........................... 4
338
+ moe_router_load_balancing_type .................. seq_aux_loss
339
+ moe_router_num_groups ........................... 8
340
+ moe_router_padding_for_fp8 ...................... False
341
+ moe_router_pre_softmax .......................... False
342
+ moe_router_score_function ....................... sigmoid
343
+ moe_router_topk ................................. 8
344
+ moe_router_topk_scaling_factor .................. 2.5
345
+ moe_shared_expert_intermediate_size ............. 2048
346
+ moe_shared_expert_overlap ....................... False
347
+ moe_token_dispatcher_type ....................... flex
348
+ moe_token_drop_policy ........................... probs
349
+ moe_ultraep_max_microbatches .................... None
350
+ moe_ultraep_overlap_weight_sync ................. True
351
+ moe_ultraep_weight_sync_join .................... before_dispatch
352
+ moe_upcycling_granularity ....................... 1
353
+ moe_use_device_initiated_grouped_gemm ........... True
354
+ moe_use_legacy_grouped_gemm ..................... False
355
+ moe_use_upcycling ............................... False
356
+ moe_z_loss_coeff ................................ None
357
+ mrope_section ................................... None
358
+ mscale .......................................... 1.0
359
+ mscale_all_dim .................................. 1.0
360
+ mtp_loss_scaling_factor ......................... 0.1
361
+ mtp_num_layers .................................. None
362
+ multi_latent_attention .......................... True
363
+ nccl_all_reduce_for_prefill ..................... False
364
+ nccl_communicator_config_path ................... None
365
+ nccl_ub ......................................... False
366
+ no_load_optim ................................... None
367
+ no_load_rng ..................................... None
368
+ no_persist_layer_norm ........................... False
369
+ no_rope_freq .................................... None
370
+ no_save_optim ................................... None
371
+ no_save_rng ..................................... None
372
+ non_persistent_ckpt_type ........................ None
373
+ non_persistent_global_ckpt_dir .................. None
374
+ non_persistent_local_ckpt_algo .................. fully_parallel
375
+ non_persistent_local_ckpt_dir ................... None
376
+ non_persistent_save_interval .................... None
377
+ norm_epsilon .................................... 1e-06
378
+ normalization ................................... RMSNorm
379
+ num_attention_heads ............................. 128
380
+ num_channels .................................... 3
381
+ num_classes ..................................... 1000
382
+ num_dataset_builder_threads ..................... 1
383
+ num_distributed_optimizer_instances ............. 1
384
+ num_experts ..................................... 256
385
+ num_layers ...................................... 4
386
+ num_layers_at_end_in_bf16 ....................... 1
387
+ num_layers_at_start_in_bf16 ..................... 1
388
+ num_layers_per_virtual_pipeline_stage ........... None
389
+ num_query_groups ................................ 1
390
+ num_virtual_stages_per_pipeline_rank ............ None
391
+ num_workers ..................................... 2
392
+ object_storage_cache_path ....................... None
393
+ one_logger_async ................................ False
394
+ one_logger_project .............................. megatron-lm
395
+ one_logger_run_name ............................. None
396
+ onnx_safe ....................................... None
397
+ openai_gelu ..................................... False
398
+ optimizer ....................................... adam
399
+ optimizer_cpu_offload ........................... False
400
+ optimizer_offload_fraction ...................... 1.0
401
+ output_bert_embeddings .......................... False
402
+ overlap_cpu_optimizer_d2h_h2d ................... False
403
+ overlap_grad_reduce ............................. True
404
+ overlap_moe_expert_parallel_comm ................ False
405
+ overlap_p2p_comm ................................ False
406
+ overlap_p2p_comm_warmup_flush ................... False
407
+ overlap_param_gather ............................ True
408
+ overlap_param_gather_with_optimizer_step ........ False
409
+ override_opt_param_scheduler .................... False
410
+ params_dtype .................................... torch.bfloat16
411
+ patch_dim ....................................... 16
412
+ per_split_data_args_path ........................ None
413
+ perform_initialization .......................... True
414
+ pin_cpu_grads ................................... True
415
+ pin_cpu_params .................................. True
416
+ pipeline_model_parallel_comm_backend ............ None
417
+ pipeline_model_parallel_layout .................. None
418
+ pipeline_model_parallel_size .................... 1
419
+ position_embedding_type ......................... rope
420
+ pretrained_checkpoint ........................... None
421
+ profile ......................................... True
422
+ profile_ranks ................................... [0, 1, 2, 3, 4, 5, 6, 7]
423
+ profile_step_end ................................ 8
424
+ profile_step_start .............................. 6
425
+ q_lora_rank ..................................... 1536
426
+ qk_head_dim ..................................... 128
427
+ qk_l2_norm ...................................... False
428
+ qk_layernorm .................................... True
429
+ qk_pos_emb_head_dim ............................. 64
430
+ query_in_block_prob ............................. 0.1
431
+ rampup_batch_size ............................... None
432
+ rank ............................................ 0
433
+ recompute_granularity ........................... None
434
+ recompute_method ................................ None
435
+ recompute_modules ............................... None
436
+ recompute_num_layers ............................ None
437
+ record_memory_history ........................... False
438
+ relative_attention_max_distance ................. 128
439
+ relative_attention_num_buckets .................. 32
440
+ replication ..................................... False
441
+ replication_factor .............................. 2
442
+ replication_jump ................................ None
443
+ rerun_mode ...................................... validate_results
444
+ reset_attention_mask ............................ False
445
+ reset_position_ids .............................. False
446
+ result_rejected_tracker_filename ................ None
447
+ retriever_report_topk_accuracies ................ []
448
+ retriever_score_scaling ......................... False
449
+ retriever_seq_length ............................ 256
450
+ retro_add_retriever ............................. False
451
+ retro_attention_gate ............................ 1
452
+ retro_cyclic_train_iters ........................ None
453
+ retro_encoder_attention_dropout ................. 0.1
454
+ retro_encoder_hidden_dropout .................... 0.1
455
+ retro_encoder_layers ............................ 2
456
+ retro_num_neighbors ............................. 2
457
+ retro_num_retrieved_chunks ...................... 2
458
+ retro_project_dir ............................... None
459
+ retro_verify_neighbor_count ..................... True
460
+ reuse_grad_buf_for_mxfp8_param_ag ............... False
461
+ rope_scaling_factor ............................. 8.0
462
+ rope_type ....................................... None
463
+ rotary_base ..................................... 10000
464
+ rotary_interleaved .............................. False
465
+ rotary_percent .................................. 1.0
466
+ rotary_scaling_factor ........................... 40.0
467
+ rotary_seq_len_interpolation_factor ............. None
468
+ run_workload_inspector_server ................... False
469
+ sample_rate ..................................... 1.0
470
+ save ............................................ None
471
+ save_interval ................................... None
472
+ save_retain_interval ............................ None
473
+ scatter_gather_tensors_in_pipeline .............. True
474
+ seed ............................................ 1234
475
+ seq_length ...................................... 4096
476
+ sequence_parallel ............................... False
477
+ sft ............................................. False
478
+ sft_tokenizer_prompt_format ..................... nemotron-h-aligned
479
+ sgd_momentum .................................... 0.9
480
+ short_seq_prob .................................. 0.1
481
+ skip_train ...................................... False
482
+ skipped_train_samples ........................... 0
483
+ spec ............................................ None
484
+ split ........................................... 100,0,0
485
+ squared_relu .................................... False
486
+ start_weight_decay .............................. 0.1
487
+ straggler_ctrlr_port ............................ 65535
488
+ straggler_minmax_count .......................... 1
489
+ suggested_communication_unit_size ............... None
490
+ swiglu .......................................... True
491
+ swin_backbone_type .............................. tiny
492
+ symmetric_ar_type ............................... None
493
+ te_rng_tracker .................................. True
494
+ tensor_model_parallel_size ...................... 1
495
+ tensorboard_dir ................................. None
496
+ tensorboard_log_interval ........................ 1
497
+ tensorboard_queue_size .......................... 1000
498
+ test_data_path .................................. None
499
+ test_mode ....................................... False
500
+ tiktoken_num_special_tokens ..................... 1000
501
+ tiktoken_pattern ................................ None
502
+ tiktoken_special_tokens ......................... None
503
+ timing_log_level ................................ 0
504
+ timing_log_option ............................... minmax
505
+ titles_data_path ................................ None
506
+ tokenizer_model ................................. None
507
+ tokenizer_type .................................. GPT2BPETokenizer
508
+ torch_fsdp2_reshard_after_forward ............... True
509
+ tp_comm_bootstrap_backend ....................... nccl
510
+ tp_comm_bulk_dgrad .............................. True
511
+ tp_comm_bulk_wgrad .............................. True
512
+ tp_comm_overlap ................................. False
513
+ tp_comm_overlap_ag .............................. True
514
+ tp_comm_overlap_cfg ............................. None
515
+ tp_comm_overlap_rs .............................. True
516
+ tp_comm_overlap_rs_dgrad ........................ False
517
+ tp_comm_split_ag ................................ True
518
+ tp_comm_split_rs ................................ True
519
+ train_data_path ................................. None
520
+ train_iters ..................................... 10
521
+ train_samples ................................... None
522
+ train_sync_interval ............................. None
523
+ transformer_impl ................................ transformer_engine
524
+ transformer_pipeline_model_parallel_size ........ 1
525
+ untie_embeddings_and_output_weights ............. True
526
+ use_checkpoint_args ............................. False
527
+ use_checkpoint_opt_param_scheduler .............. False
528
+ use_cpu_initialization .......................... None
529
+ use_custom_fsdp ................................. False
530
+ use_dist_ckpt ................................... True
531
+ use_dist_ckpt_deprecated ........................ False
532
+ use_distributed_optimizer ....................... True
533
+ use_flash_attn .................................. False
534
+ use_legacy_models ............................... False
535
+ use_mp_args_from_checkpoint_args ................ False
536
+ use_one_sent_docs ............................... False
537
+ use_persistent_ckpt_worker ...................... False
538
+ use_precision_aware_optimizer ................... False
539
+ use_pytorch_profiler ............................ False
540
+ use_ring_exchange_p2p ........................... False
541
+ use_rope_scaling ................................ False
542
+ use_rotary_position_embeddings .................. False
543
+ use_sharp ....................................... False
544
+ use_tokenizer_model_from_checkpoint_args ........ True
545
+ use_torch_fsdp2 ................................. False
546
+ use_torch_optimizer_for_cpu_offload ............. False
547
+ use_tp_pp_dp_mapping ............................ False
548
+ v_head_dim ...................................... 128
549
+ valid_data_path ................................. None
550
+ variable_seq_lengths ............................ False
551
+ virtual_pipeline_model_parallel_size ............ None
552
+ vision_backbone_type ............................ vit
553
+ vision_pretraining .............................. False
554
+ vision_pretraining_type ......................... classify
555
+ vocab_extra_ids ................................. 0
556
+ vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json
557
+ vocab_size ...................................... None
558
+ wandb_exp_name ..................................
559
+ wandb_project ...................................
560
+ wandb_save_dir ..................................
561
+ weight_decay .................................... 0.1
562
+ weight_decay_incr_style ......................... constant
563
+ wgrad_deferral_limit ............................ 0
564
+ world_size ...................................... 8
565
+ yaml_cfg ........................................ None
566
+ -------------------- end of arguments ---------------------
567
+ INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1
568
+ > building GPT2BPETokenizer tokenizer ...
569
+ > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304)
570
+ WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS
571
+ > initializing torch distributed ...
572
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
573
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
574
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
575
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
576
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
577
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
578
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
579
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
580
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
581
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
582
+ 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
583
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
584
+ [Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
585
+ [Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
586
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
587
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
588
+ [Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
589
+ [Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
590
+ [Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
591
+ [Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
592
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
593
+ [Gloo] Rank [Gloo] Rank 25 is connected to 7 is connected to peer ranks. 7Expected number of connected peer ranks is : peer ranks. 7Expected number of connected peer ranks is : 7
594
+
595
+ [Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
596
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
597
+ [Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
598
+ [Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
599
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
600
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
601
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
602
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : [Gloo] Rank 00
603
+ is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
604
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
605
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
606
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
607
+ > initialized tensor model parallel with size 1
608
+ > initialized pipeline model parallel with size 1
609
+ > setting random seeds to 1234 ...
610
+ [rank5]:[W921 07:24:07.231242443 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.
611
+ [rank3]:[W921 07:24:07.231288984 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.
612
+ [rank2]:[W921 07:24:07.231304783 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.
613
+ [rank1]:[W921 07:24:07.231347216 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.
614
+ [rank6]:[W921 07:24:07.231409229 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.
615
+ [rank4]:[W921 07:24:07.234086582 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.
616
+ [rank7]:[W921 07:24:07.234116920 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.
617
+ > compiling dataset index builder ...
618
+ make: Entering directory '/workspace/Megatron-UltraEP/megatron/core/datasets'
619
+ make: Nothing to be done for 'default'.
620
+ make: Leaving directory '/workspace/Megatron-UltraEP/megatron/core/datasets'
621
+ >>> done with dataset index builder. Compilation time: 0.434 seconds
622
+ WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations.
623
+ > compiling and loading fused kernels ...
624
+ [rank0]:[W921 07:24:07.853833848 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.
625
+ >>> done with compiling and loading fused kernels. Compilation time: 4.083 seconds
626
+ time to initialize megatron (seconds): 9.455
627
+ [after megatron is initialized] datetime: 2026-09-21 07:24:14
628
+ building GPT model ...
629
+ /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.
630
+ warnings.warn(
631
+ /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.
632
+ warnings.warn(
633
+ /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.
634
+ warnings.warn(
635
+ /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.
636
+ warnings.warn(
637
+ /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.
638
+ warnings.warn(
639
+ /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.
640
+ warnings.warn(
641
+ /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.
642
+ warnings.warn(
643
+ /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.
644
+ warnings.warn(
645
+ [INFO] Use MNNVL fabric: False
646
+ [WARN] MAX_NUM_NVL_PEERS is not set. Using detected value 8.
647
+ NVSHMEM v3.7.2
648
+ /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.
649
+
650
+ /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.
651
+
652
+ /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.
653
+
654
+ /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.
655
+
656
+ /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.
657
+
658
+ /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.
659
+
660
+ /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.
661
+
662
+ /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.
663
+
664
+ UltraEP: NVL domain size <= 8; forcing weight_sync plan mode to direct (was adaptive).
665
+ > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 5523184640
666
+ 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)
667
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 17
668
+ Params for bucket 1 (360579072 elements, 360579072 padded size):
669
+ module.output_layer.weight
670
+ Params for bucket 2 (44047360 elements, 44047360 padded size):
671
+ module.decoder.final_layernorm.weight
672
+ module.decoder.layers.3.mlp.shared_experts.linear_fc2.weight
673
+ module.decoder.layers.3.mlp.shared_experts.linear_fc1.weight
674
+ Params for bucket 3 (60497408 elements, 60497408 padded size):
675
+ module.decoder.layers.3.mlp.router.weight
676
+ module.decoder.layers.3.self_attention.linear_kv_down_proj.weight
677
+ module.decoder.layers.3.self_attention.linear_kv_up_proj.layer_norm_weight
678
+ module.decoder.layers.3.self_attention.linear_q_up_proj.weight
679
+ module.decoder.layers.3.self_attention.linear_kv_up_proj.weight
680
+ module.decoder.layers.3.pre_mlp_layernorm.weight
681
+ Params for bucket 4 (128452096 elements, 128452096 padded size):
682
+ module.decoder.layers.3.self_attention.linear_q_up_proj.layer_norm_weight
683
+ module.decoder.layers.3.self_attention.linear_proj.weight
684
+ module.decoder.layers.3.self_attention.linear_q_down_proj.weight
685
+ Params for bucket 5 (132127744 elements, 132127744 padded size):
686
+ module.decoder.layers.3.input_layernorm.weight
687
+ module.decoder.layers.2.mlp.linear_fc2.weight
688
+ Params for bucket 6 (264241152 elements, 264241152 padded size):
689
+ module.decoder.layers.2.mlp.linear_fc1.weight
690
+ Params for bucket 7 (58662400 elements, 58662400 padded size):
691
+ module.decoder.layers.2.mlp.linear_fc1.layer_norm_weight
692
+ module.decoder.layers.2.self_attention.linear_kv_up_proj.layer_norm_weight
693
+ module.decoder.layers.2.self_attention.linear_kv_down_proj.weight
694
+ module.decoder.layers.2.self_attention.linear_q_up_proj.weight
695
+ module.decoder.layers.2.self_attention.linear_kv_up_proj.weight
696
+ Params for bucket 8 (128452096 elements, 128452096 padded size):
697
+ module.decoder.layers.2.self_attention.linear_q_up_proj.layer_norm_weight
698
+ module.decoder.layers.2.self_attention.linear_proj.weight
699
+ module.decoder.layers.2.self_attention.linear_q_down_proj.weight
700
+ Params for bucket 9 (132127744 elements, 132127744 padded size):
701
+ module.decoder.layers.2.input_layernorm.weight
702
+ module.decoder.layers.1.mlp.linear_fc2.weight
703
+ Params for bucket 10 (264241152 elements, 264241152 padded size):
704
+ module.decoder.layers.1.mlp.linear_fc1.weight
705
+ Params for bucket 11 (58662400 elements, 58662400 padded size):
706
+ module.decoder.layers.1.mlp.linear_fc1.layer_norm_weight
707
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.weight
708
+ module.decoder.layers.1.self_attention.linear_kv_down_proj.weight
709
+ module.decoder.layers.1.self_attention.linear_q_up_proj.weight
710
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.layer_norm_weight
711
+ Params for bucket 12 (128452096 elements, 128452096 padded size):
712
+ module.decoder.layers.1.self_attention.linear_q_up_proj.layer_norm_weight
713
+ module.decoder.layers.1.self_attention.linear_proj.weight
714
+ module.decoder.layers.1.self_attention.linear_q_down_proj.weight
715
+ Params for bucket 13 (132127744 elements, 132127744 padded size):
716
+ module.decoder.layers.1.input_layernorm.weight
717
+ module.decoder.layers.0.mlp.linear_fc2.weight
718
+ Params for bucket 14 (264241152 elements, 264241152 padded size):
719
+ module.decoder.layers.0.mlp.linear_fc1.weight
720
+ Params for bucket 15 (58662400 elements, 58662400 padded size):
721
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.weight
722
+ module.decoder.layers.0.mlp.linear_fc1.layer_norm_weight
723
+ module.decoder.layers.0.self_attention.linear_q_up_proj.weight
724
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.layer_norm_weight
725
+ module.decoder.layers.0.self_attention.linear_kv_down_proj.weight
726
+ Params for bucket 16 (128452096 elements, 128452096 padded size):
727
+ module.decoder.layers.0.self_attention.linear_q_up_proj.layer_norm_weight
728
+ module.decoder.layers.0.self_attention.linear_proj.weight
729
+ module.decoder.layers.0.self_attention.linear_q_down_proj.weight
730
+ Params for bucket 17 (360586240 elements, 360586240 padded size):
731
+ module.decoder.layers.0.input_layernorm.weight
732
+ module.embedding.word_embeddings.weight
733
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 27
734
+ Params for bucket 1 (44040192 elements, 44040192 padded size):
735
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight31
736
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight29
737
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight30
738
+ Params for bucket 2 (44040192 elements, 44040192 padded size):
739
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight28
740
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight26
741
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight27
742
+ Params for bucket 3 (44040192 elements, 44040192 padded size):
743
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight25
744
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight23
745
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight24
746
+ Params for bucket 4 (44040192 elements, 44040192 padded size):
747
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight22
748
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight21
749
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight20
750
+ Params for bucket 5 (44040192 elements, 44040192 padded size):
751
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight19
752
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight18
753
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight17
754
+ Params for bucket 6 (44040192 elements, 44040192 padded size):
755
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight16
756
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight15
757
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight14
758
+ Params for bucket 7 (44040192 elements, 44040192 padded size):
759
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight13
760
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight12
761
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight11
762
+ Params for bucket 8 (44040192 elements, 44040192 padded size):
763
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight10
764
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight9
765
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight8
766
+ Params for bucket 9 (44040192 elements, 44040192 padded size):
767
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight7
768
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight5
769
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight6
770
+ Params for bucket 10 (44040192 elements, 44040192 padded size):
771
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight4
772
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight2
773
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight3
774
+ Params for bucket 11 (58720256 elements, 58720256 padded size):
775
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight1
776
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight31
777
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight0
778
+ Params for bucket 12 (58720256 elements, 58720256 padded size):
779
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight30
780
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight29
781
+ Params for bucket 13 (58720256 elements, 58720256 padded size):
782
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight28
783
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight27
784
+ Params for bucket 14 (58720256 elements, 58720256 padded size):
785
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight26
786
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight25
787
+ Params for bucket 15 (58720256 elements, 58720256 padded size):
788
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight24
789
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight23
790
+ Params for bucket 16 (58720256 elements, 58720256 padded size):
791
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight22
792
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight21
793
+ Params for bucket 17 (58720256 elements, 58720256 padded size):
794
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight20
795
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight19
796
+ Params for bucket 18 (58720256 elements, 58720256 padded size):
797
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight18
798
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight17
799
+ Params for bucket 19 (58720256 elements, 58720256 padded size):
800
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight16
801
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight15
802
+ Params for bucket 20 (58720256 elements, 58720256 padded size):
803
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight14
804
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight13
805
+ Params for bucket 21 (58720256 elements, 58720256 padded size):
806
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight12
807
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight11
808
+ Params for bucket 22 (58720256 elements, 58720256 padded size):
809
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight10
810
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight9
811
+ Params for bucket 23 (58720256 elements, 58720256 padded size):
812
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight8
813
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight7
814
+ Params for bucket 24 (58720256 elements, 58720256 padded size):
815
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight6
816
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight5
817
+ Params for bucket 25 (58720256 elements, 58720256 padded size):
818
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight4
819
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight3
820
+ Params for bucket 26 (58720256 elements, 58720256 padded size):
821
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight2
822
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight1
823
+ Params for bucket 27 (29360128 elements, 29360128 padded size):
824
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight0
825
+ 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=<megatron.core.timers.Timers object at 0x77520de2ae10>, config_logger_dir='')
826
+ INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine
827
+ [after model, optimizer, and learning rate scheduler are built] datetime: 2026-09-21 07:24:17
828
+ > building train, validation, and test datasets ...
829
+ > datasets target sizes (minimum size):
830
+ train: 80
831
+ validation: 0
832
+ test: 0
833
+ INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None]
834
+ > building train, validation, and test datasets for GPT ...
835
+ INFO:megatron.core.datasets.blended_megatron_dataset_builder:Building GPTDataset splits with sizes=(80, 0, 0) and config=GPTDatasetConfig(random_seed=1234, sequence_length=4096, blend=(['/workspace/data/wikitext/wikitext103_gpt2_text_document'], None), blend_per_split=None, split='100,0,0', split_matrix=[(0, 1.0), None, None], num_dataset_builder_threads=1, path_to_cache=None, mmap_bin_files=True, mock=False, tokenizer=<megatron.training.tokenizer.tokenizer._GPT2BPETokenizer object at 0x77520ecf2240>, 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)
836
+ INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx
837
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths
838
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers
839
+ INFO:megatron.core.datasets.indexed_dataset: Extract the document indices
840
+ INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029
841
+ INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350
842
+ INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices
843
+ INFO:megatron.core.datasets.gpt_dataset: Load the document index from 71e99bc7c46bc3412b339a272e0f9cf3-GPTDataset-train-document_index.npy
844
+ INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 71e99bc7c46bc3412b339a272e0f9cf3-GPTDataset-train-sample_index.npy
845
+ INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 71e99bc7c46bc3412b339a272e0f9cf3-GPTDataset-train-shuffle_index.npy
846
+ INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073
847
+ > finished creating GPT datasets ...
848
+ [after dataloaders are built] datetime: 2026-09-21 07:24:17
849
+ done with setup ...
850
+ training ...
851
+ (min, max) time across ranks (ms):
852
+ model-and-optimizer-setup ......................: (3353.51, 3353.67)
853
+ train/valid/test-data-iterators-setup ..........: (106.42, 127.08)
854
+ Overwriting rerun_state_machine.current_iteration from -1 to 0...
855
+ [before the start of training step] datetime: 2026-09-21 07:24:17
856
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
857
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
858
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
859
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
860
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
861
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
862
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
863
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
864
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
865
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
866
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
867
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
868
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
869
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
870
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
871
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
872
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
873
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
874
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
875
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
876
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
877
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
878
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
879
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
880
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
881
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
882
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
883
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
884
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
885
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
886
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
887
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
888
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
889
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
890
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
891
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
892
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
893
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
894
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
895
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
896
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
897
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
898
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
899
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
900
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
901
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
902
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
903
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
904
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
905
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
906
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
907
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
908
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
909
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
910
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
911
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
912
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
913
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
914
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
915
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
916
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
917
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
918
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
919
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
920
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
921
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
922
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
923
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
924
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
925
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
926
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
927
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
928
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
929
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
930
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
931
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
932
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
933
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
934
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
935
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
936
+ [2026-09-21 07:24:49] iteration 1/ 10 | consumed samples: 8 | elapsed time per iteration (ms): 31346.3 | throughput per GPU (TFLOP/s/GPU): 2.4 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.226155E+01 | loss scale: 1.0 | grad norm: 15.011 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34048 | min safe capacity: 1.0391 |
937
+ Number of parameters in transformer block in billions: 13.26
938
+ Number of parameters in embedding layers in billions: 0.72
939
+ Total number of parameters in billions: 13.98
940
+ Number of parameters in most loaded shard in billions: 13.9771
941
+ Theoretical memory footprints: weight and optimizer=99972.23 MB
942
+ [Rank 0] (after 1 iterations) memory (MB) | allocated: 49424.15869140625 | max allocated: 65699.892578125 | reserved: 68070.0 | max reserved: 68070.0
943
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
944
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
945
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
946
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
947
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
948
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
949
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
950
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
951
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
952
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
953
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
954
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
955
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
956
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
957
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
958
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
959
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
960
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
961
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
962
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
963
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
964
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
965
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
966
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
967
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
968
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
969
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
970
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
971
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
972
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
973
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
974
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
975
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
976
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
977
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
978
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
979
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
980
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
981
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
982
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
983
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
984
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
985
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
986
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
987
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
988
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
989
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
990
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
991
+ [2026-09-21 07:24:49] iteration 2/ 10 | consumed samples: 16 | elapsed time per iteration (ms): 265.8 | throughput per GPU (TFLOP/s/GPU): 280.1 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.227770E+01 | loss scale: 1.0 | grad norm: 14.855 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34688 | min safe capacity: 1.0586 |
992
+ [2026-09-21 07:24:49] iteration 3/ 10 | consumed samples: 24 | elapsed time per iteration (ms): 205.5 | throughput per GPU (TFLOP/s/GPU): 362.4 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 1.070660E+01 | loss scale: 1.0 | grad norm: 18.852 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 35200 | min safe capacity: 1.0742 |
993
+ Capture CUDA graph for training!!!
994
+ CUDA graph capture done!!!
995
+ [2026-09-21 07:24:50] iteration 4/ 10 | consumed samples: 32 | elapsed time per iteration (ms): 1081.1 | throughput per GPU (TFLOP/s/GPU): 68.9 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 1.557676E+01 | loss scale: 1.0 | grad norm: 97.972 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 35200 | min safe capacity: 1.0742 |
996
+ [2026-09-21 07:24:50] iteration 5/ 10 | consumed samples: 40 | elapsed time per iteration (ms): 198.7 | throughput per GPU (TFLOP/s/GPU): 374.7 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.446037E+01 | loss scale: 1.0 | grad norm: 97.493 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 35200 | min safe capacity: 1.0742 |
997
+ [2026-09-21 07:24:51] iteration 6/ 10 | consumed samples: 48 | elapsed time per iteration (ms): 194.6 | throughput per GPU (TFLOP/s/GPU): 382.7 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 1.293258E+01 | loss scale: 1.0 | grad norm: 16.031 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 35200 | min safe capacity: 1.0742 |
998
+ [2026-09-21 07:24:54] iteration 7/ 10 | consumed samples: 56 | elapsed time per iteration (ms): 3055.4 | throughput per GPU (TFLOP/s/GPU): 24.4 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 1.697075E+01 | loss scale: 1.0 | grad norm: 101.127 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 35200 | min safe capacity: 1.0742 |
999
+ [2026-09-21 07:24:54] iteration 8/ 10 | consumed samples: 64 | elapsed time per iteration (ms): 199.6 | throughput per GPU (TFLOP/s/GPU): 373.1 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.099243E+01 | loss scale: 1.0 | grad norm: 18.463 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 35200 | min safe capacity: 1.0742 |
1000
+ Capture range started in the application.
1001
+ Capture range ended in the application.
1002
+ Generating '/tmp/nsys-report-cff1.qdstrm'
1003
+
1004
+
1005
+ [2026-09-21 07:25:33] iteration 9/ 10 | consumed samples: 72 | elapsed time per iteration (ms): 39118.8 | throughput per GPU (TFLOP/s/GPU): 1.9 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.045388E+01 | loss scale: 1.0 | grad norm: 15.870 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 35200 | min safe capacity: 1.0742 |
1006
+ [2026-09-21 07:25:33] iteration 10/ 10 | consumed samples: 80 | elapsed time per iteration (ms): 198.2 | throughput per GPU (TFLOP/s/GPU): 375.8 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.014687E+01 | loss scale: 1.0 | grad norm: 11.163 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 35456 | min safe capacity: 1.0820 |
1007
+ [after training is done] datetime: 2026-09-21 07:25:33
1008
+ [rank0]:[W921 07:25:34.830568489 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())
1009
+ [rank1]:[W921 07:25:36.050340507 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())
1010
+ [rank4]:[W921 07:25:36.562427478 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())
1011
+ [rank7]:[W921 07:25:36.580254500 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())
1012
+ [rank3]:[W921 07:25:36.581873163 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())
1013
+ [rank5]:[W921 07:25:36.582359327 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())
1014
+ [rank6]:[W921 07:25:36.622032489 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())
1015
+ Collection stop is not allowed in this state.
1016
+ [rank2]:[W921 07:25:37.101938966 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())
1017
+ [rank1]:[W921 07:25:39.305464287 observer.cpp:116] Warning: Leaked callback handle: 1 (function operator())
1018
+ pure virtual method called
1019
+ terminate called without an active exception
1020
+ [rank0]:[W921 07:25:39.468382058 observer.cpp:116] Warning: Leaked callback handle: 1 (function operator())
1021
+ pure virtual method called
1022
+ terminate called without an active exception
1023
+ W0921 07:25:39.949000 434771 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 434900 closing signal SIGTERM
1024
+ W0921 07:25:39.950000 434771 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 434902 closing signal SIGTERM
1025
+ W0921 07:25:39.950000 434771 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 434903 closing signal SIGTERM
1026
+ W0921 07:25:39.950000 434771 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 434904 closing signal SIGTERM
1027
+ W0921 07:25:39.951000 434771 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 434905 closing signal SIGTERM
1028
+ W0921 07:25:39.951000 434771 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 434906 closing signal SIGTERM
1029
+ W0921 07:25:39.952000 434771 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 434907 closing signal SIGTERM
1030
+ E0921 07:25:41.784000 434771 torch/distributed/elastic/multiprocessing/api.py:874] failed (exitcode: -6) local_rank: 1 (pid: 434901) of binary: /usr/bin/python
1031
+ Traceback (most recent call last):
1032
+ File "/usr/local/bin/torchrun", line 8, in <module>
1033
+ sys.exit(main())
1034
+ ^^^^^^
1035
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
1036
+ return f(*args, **kwargs)
1037
+ ^^^^^^^^^^^^^^^^^^
1038
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 892, in main
1039
+ run(args)
1040
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 883, in run
1041
+ elastic_launch(
1042
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
1043
+ return launch_agent(self._config, self._entrypoint, list(args))
1044
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1045
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 270, in launch_agent
1046
+ raise ChildFailedError(
1047
+ torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
1048
+ ============================================================
1049
+ pretrain_gpt.py FAILED
1050
+ ------------------------------------------------------------
1051
+ Failures:
1052
+ <NO_OTHER_FAILURES>
1053
+ ------------------------------------------------------------
1054
+ Root Cause (first observed failure):
1055
+ [0]:
1056
+ time : 2026-09-21_07:25:39
1057
+ host : sp-aiinfra-interactive-aghosh-8-mb99k-8xb200-moe-cg-0
1058
+ rank : 1 (local_rank: 1)
1059
+ exitcode : -6 (pid: 434901)
1060
+ error_file: <N/A>
1061
+ traceback : Signal 6 (SIGABRT) received by PID 434901
1062
+ ============================================================
1063
+ Generated:
1064
+ /workspace/granularity_next_b200/tierb_models/traces/DEEPSEEK_V3_L4/ultraep/tierbtrace/tierbtrace_gbs8/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875/nsys/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875.nsys-rep
granularity_next_b200/tierb_models/traces/DEEPSEEK_V3_L4/ultraep/tierbtrace/tierbtrace_gbs8/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875/nsys/verify_trace_ranks.log ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ report : /workspace/granularity_next_b200/tierb_models/traces/DEEPSEEK_V3_L4/ultraep/tierbtrace/tierbtrace_gbs8/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875/nsys/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875.nsys-rep
2
+ kernels : 28624
3
+ devices : 8 of 8 -> [0, 1, 2, 3, 4, 5, 6, 7]
4
+ processes: 1
5
+ markers : 896 ultraep_trace__ kernels on devices [0, 1, 2, 3, 4, 5, 6, 7]
6
+ PASS: all ranks present and backend markers emitted
granularity_next_b200/tierb_models/traces/DEEPSEEK_V3_L4/ultraep/tierbtrace/tierbtrace_gbs8/results.tsv ADDED
@@ -0,0 +1 @@
 
 
1
+ config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir
granularity_next_b200/tierb_models/traces/DEEPSEEK_V3_L4/ultraep/tierbtrace/tierbtrace_gbs8/trace_extension_build.log CHANGED
@@ -51,3 +51,8 @@ Emitting ninja build file /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cp
51
  Compiling objects...
52
  Allowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N)
53
  [1/2] c++ -MMD -MF /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cpython-312/megatron/core/transformer/moe/ultraep_trace_ops.o.d -fno-strict-overflow -Wsign-compare -DNDEBUG -g -O2 -Wall -fPIC -I/usr/local/lib/python3.12/dist-packages/torch/include -I/usr/local/lib/python3.12/dist-packages/torch/include/torch/csrc/api/include -I/usr/local/cuda/include -I/usr/include/python3.12 -c -c /workspace/Megatron-UltraEP/megatron/core/transformer/moe/ultraep_trace_ops.cpp -o /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cpython-312/megatron/core/transformer/moe/ultraep_trace_ops.o -O3 -std=c++17 -DTORCH_API_INCLUDE_EXTENSION_H '-DPYBIND11_COMPILER_TYPE="_gcc"' '-DPYBIND11_STDLIB="_libstdcpp"' '-DPYBIND11_BUILD_ABI="_cxxabi1016"' -DTORCH_EXTENSION_NAME=_ultraep_trace_C
 
 
 
 
 
 
51
  Compiling objects...
52
  Allowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N)
53
  [1/2] c++ -MMD -MF /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cpython-312/megatron/core/transformer/moe/ultraep_trace_ops.o.d -fno-strict-overflow -Wsign-compare -DNDEBUG -g -O2 -Wall -fPIC -I/usr/local/lib/python3.12/dist-packages/torch/include -I/usr/local/lib/python3.12/dist-packages/torch/include/torch/csrc/api/include -I/usr/local/cuda/include -I/usr/include/python3.12 -c -c /workspace/Megatron-UltraEP/megatron/core/transformer/moe/ultraep_trace_ops.cpp -o /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cpython-312/megatron/core/transformer/moe/ultraep_trace_ops.o -O3 -std=c++17 -DTORCH_API_INCLUDE_EXTENSION_H '-DPYBIND11_COMPILER_TYPE="_gcc"' '-DPYBIND11_STDLIB="_libstdcpp"' '-DPYBIND11_BUILD_ABI="_cxxabi1016"' -DTORCH_EXTENSION_NAME=_ultraep_trace_C
54
+ [2/2] /usr/local/cuda/bin/nvcc --generate-dependencies-with-compile --dependency-output /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cpython-312/megatron/core/transformer/moe/ultraep_trace_kernels.o.d -I/usr/local/lib/python3.12/dist-packages/torch/include -I/usr/local/lib/python3.12/dist-packages/torch/include/torch/csrc/api/include -I/usr/local/cuda/include -I/usr/include/python3.12 -c -c /workspace/Megatron-UltraEP/megatron/core/transformer/moe/ultraep_trace_kernels.cu -o /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cpython-312/megatron/core/transformer/moe/ultraep_trace_kernels.o -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr --compiler-options ''"'"'-fPIC'"'"'' -O3 -std=c++17 -lineinfo -DTORCH_API_INCLUDE_EXTENSION_H '-DPYBIND11_COMPILER_TYPE="_gcc"' '-DPYBIND11_STDLIB="_libstdcpp"' '-DPYBIND11_BUILD_ABI="_cxxabi1016"' -DTORCH_EXTENSION_NAME=_ultraep_trace_C -gencode=arch=compute_100,code=sm_100 -gencode=arch=compute_120,code=compute_120 -gencode=arch=compute_120,code=sm_120 -gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_80,code=sm_80 -gencode=arch=compute_86,code=sm_86 -gencode=arch=compute_90,code=sm_90
55
+ x86_64-linux-gnu-g++ -fno-strict-overflow -Wsign-compare -DNDEBUG -g -O2 -Wall -shared -Wl,-O1 -Wl,-Bsymbolic-functions /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cpython-312/megatron/core/transformer/moe/ultraep_trace_kernels.o /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cpython-312/megatron/core/transformer/moe/ultraep_trace_ops.o -L/usr/local/lib/python3.12/dist-packages/torch/lib -L/usr/local/cuda/lib64 -L/usr/lib/x86_64-linux-gnu -lc10 -ltorch -ltorch_cpu -ltorch_python -lcudart -lc10_cuda -ltorch_cuda -o build/lib.linux-x86_64-cpython-312/megatron/core/transformer/moe/_ultraep_trace_C.cpython-312-x86_64-linux-gnu.so
56
+ copying build/lib.linux-x86_64-cpython-312/megatron/core/datasets/helpers_cpp.cpython-312-x86_64-linux-gnu.so -> megatron/core/datasets
57
+ copying build/lib.linux-x86_64-cpython-312/megatron/core/transformer/moe/_ultraep_trace_C.cpython-312-x86_64-linux-gnu.so -> megatron/core/transformer/moe
58
+ load check passed: /workspace/Megatron-UltraEP/megatron/core/transformer/moe/_ultraep_trace_C.cpython-312-x86_64-linux-gnu.so -> torch.ops.megatron_ultraep_trace.fwd_ultraep_forward_begin
granularity_next_b200/tierb_models/traces/DEEPSEEK_V3_L4/ultraep/tierbtrace/tierbtrace_gbs8/trace_results.tsv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ model backend capacity status report async_overlap_json
2
+ DEEPSEEK_V3_L4 ultraep 1.13671875 OK /workspace/granularity_next_b200/tierb_models/traces/DEEPSEEK_V3_L4/ultraep/tierbtrace/tierbtrace_gbs8/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875/nsys/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875.nsys-rep