AbhishekGhosh1998 commited on
Commit
45cc5cc
·
verified ·
1 Parent(s): 44fccba

granularity results granularity_next_b200/tiera_topk_trace @ 2026-09-17T03:35:04Z from hp-aiinfra-interactive-aghosh-8-xzthl-8xb200-moe-cg-0

Browse files
Files changed (17) hide show
  1. .gitattributes +3 -0
  2. granularity_next_b200/tiera_topk_trace/hf_watcher.log +3 -0
  3. granularity_next_b200/tiera_topk_trace/study.log +16 -0
  4. granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep +3 -0
  5. granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/run.env +102 -0
  6. granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/train.log +1176 -0
  7. granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/verify_trace_ranks.log +6 -0
  8. granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep +3 -0
  9. granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/run.env +102 -0
  10. granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/train.log +0 -0
  11. granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/verify_trace_ranks.log +6 -0
  12. granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep +3 -0
  13. granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/run.env +102 -0
  14. granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/train.log +1268 -0
  15. granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/results.tsv +1 -0
  16. granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/trace_extension_build.log +6 -0
  17. granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/trace_results.tsv +3 -0
.gitattributes CHANGED
@@ -91,3 +91,6 @@ granularity_next_b200/tiera_topk_trace/echo/tierbtrace/tierbtrace_gbs16/G4_K6__e
91
  granularity_next_b200/tiera_topk_trace/echo/tierbtrace/tierbtrace_gbs16/G4_K8__echo__matched__seq4096__gbs16__cap1p109375/nsys/G4_K8__echo__matched__seq4096__gbs16__cap1p109375.nsys-rep filter=lfs diff=lfs merge=lfs -text
92
  granularity_next_b200/tiera_topk_trace/echo/tierbtrace/tierbtrace_gbs16/G4_K10__echo__matched__seq4096__gbs16__cap1p1/nsys/G4_K10__echo__matched__seq4096__gbs16__cap1p1.nsys-rep filter=lfs diff=lfs merge=lfs -text
93
  granularity_next_b200/tiera_topk_trace/echo/tierbtrace/tierbtrace_gbs16/G4_K16__echo__matched__seq4096__gbs16__cap1p08203125/nsys/G4_K16__echo__matched__seq4096__gbs16__cap1p08203125.nsys-rep filter=lfs diff=lfs merge=lfs -text
 
 
 
 
91
  granularity_next_b200/tiera_topk_trace/echo/tierbtrace/tierbtrace_gbs16/G4_K8__echo__matched__seq4096__gbs16__cap1p109375/nsys/G4_K8__echo__matched__seq4096__gbs16__cap1p109375.nsys-rep filter=lfs diff=lfs merge=lfs -text
92
  granularity_next_b200/tiera_topk_trace/echo/tierbtrace/tierbtrace_gbs16/G4_K10__echo__matched__seq4096__gbs16__cap1p1/nsys/G4_K10__echo__matched__seq4096__gbs16__cap1p1.nsys-rep filter=lfs diff=lfs merge=lfs -text
93
  granularity_next_b200/tiera_topk_trace/echo/tierbtrace/tierbtrace_gbs16/G4_K16__echo__matched__seq4096__gbs16__cap1p08203125/nsys/G4_K16__echo__matched__seq4096__gbs16__cap1p08203125.nsys-rep filter=lfs diff=lfs merge=lfs -text
94
+ granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep filter=lfs diff=lfs merge=lfs -text
95
+ granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep filter=lfs diff=lfs merge=lfs -text
96
+ granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep filter=lfs diff=lfs merge=lfs -text
granularity_next_b200/tiera_topk_trace/hf_watcher.log CHANGED
@@ -28,3 +28,6 @@ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
28
  uploaded: /workspace/granularity_next_b200/tiera_topk_trace -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tiera_topk_trace
29
  https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
30
  [2026-09-17T03:25:01+00:00] backup pass
 
 
 
 
28
  uploaded: /workspace/granularity_next_b200/tiera_topk_trace -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tiera_topk_trace
29
  https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
30
  [2026-09-17T03:25:01+00:00] backup pass
31
+ uploaded: /workspace/granularity_next_b200/tiera_topk_trace -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tiera_topk_trace
32
+ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
33
+ [2026-09-17T03:35:04+00:00] backup pass
granularity_next_b200/tiera_topk_trace/study.log CHANGED
@@ -104,3 +104,19 @@ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
104
  results -> /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace
105
  ==============================================================
106
  [03:22:42] building ultraep marker extension (log: /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/trace_extension_build.log)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
104
  results -> /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace
105
  ==============================================================
106
  [03:22:42] building ultraep marker extension (log: /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/trace_extension_build.log)
107
+ [03:26:51] marker extension ready: _ultraep_trace_C.cpython-312-x86_64-linux-gnu.so
108
+ [03:26:51] TRACE-ONLY G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625 (ranks: 0 1 2 3 4 5 6 7)
109
+ [guard] iter 8/10 (70s)
110
+ [guard] no progress for 60s at iter 8/10 (past profile window); treating as done
111
+ [guard] nsys stop (flushing report)
112
+ [guard] tearing down process group (outcome=STALLED)
113
+ [03:29:49] trace verified -> G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep (11M, all 8 GPUs + ultraep markers)
114
+ [03:29:49] TRACE-ONLY G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875 (ranks: 0 1 2 3 4 5 6 7)
115
+ [guard] iter 7/10 (70s)
116
+ [guard] iter 8/10 (80s)
117
+ [guard] no progress for 60s at iter 8/10 (past profile window); treating as done
118
+ [guard] nsys stop (flushing report)
119
+ [guard] tearing down process group (outcome=STALLED)
120
+ [03:32:56] trace verified -> G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep (12M, all 8 GPUs + ultraep markers)
121
+ [03:32:56] TRACE-ONLY G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375 (ranks: 0 1 2 3 4 5 6 7)
122
+ [guard] iter 8/10 (70s)
granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:16d8b184e95b4cc82e363b6b774e38586cdc69ef72bb35e09082e92a2190d7f6
3
+ size 10833175
granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/run.env ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=ultraep
2
+ config=G2_K2
3
+ family=tierA
4
+ source_model=
5
+ standard_num_layers=
6
+ base_rung=G2
7
+ repo=/workspace/Megatron-UltraEP
8
+ branch=ultraep-bracing-regions-through-kernels
9
+ commit=03d57ef7
10
+ num_layers=16
11
+ hidden=4096
12
+ ffn_hidden=14336
13
+ moe_ffn_hidden=3584
14
+ num_experts=32
15
+ topk=2
16
+ experts_per_rank=4
17
+ expert_params=44040192
18
+ heads=32
19
+ kv_groups=8
20
+ kv_channels=
21
+ qk_layernorm=0
22
+ shared_ffn=0
23
+ shared_expert_overlap=0
24
+ moe_token_dispatcher=flex
25
+ mla=0
26
+ q_lora_rank=
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=0
32
+ moe_layers=16
33
+ rotary_base=1000000
34
+ rotary_percent=1.0
35
+ rotary_scaling=
36
+ mscale=
37
+ mscale_all_dim=
38
+ qkv_bias=0
39
+ router_balancing=aux_loss
40
+ aux_loss_coeff=1e-2
41
+ router_pre_softmax=0
42
+ router_fusion=0
43
+ router_dtype=fp32
44
+ router_groups=
45
+ router_group_topk=
46
+ router_score_function=softmax
47
+ router_expert_bias=0
48
+ router_bias_update_rate=
49
+ router_topk_scaling_factor=
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=32768
64
+ init_method_std=0.01
65
+ device_initiated_grouped_gemm=1
66
+ norm_epsilon=1e-5
67
+ seq_length=4096
68
+ micro_batch_size=1
69
+ global_batch_size=16
70
+ num_microbatches=2
71
+ assignments_per_rank=8192
72
+ mean_tokens_per_expert=2048
73
+ slot_policy=matched
74
+ slots_per_rank=4
75
+ capacity_mode=static
76
+ capacity=1.15625
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=256
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/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/train.log ADDED
@@ -0,0 +1,1176 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ==============================================================
2
+ G2_K2 / ultraep (slots/rank=4 policy=matched)
3
+ E=32 K=2 H'=3584 m=4
4
+ L=16 seq=4096 gbs=16 (2 microbatches)
5
+ cap_mode=static cap=1.15625 cg=1 recompute=none
6
+ memory_dump=none snapshots=/workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/memory_snapshots/snapshot.pickle
7
+ repo=/workspace/Megatron-UltraEP @ 03d57ef7
8
+ ==============================================================
9
+ W0917 03:26:59.698000 948326 torch/distributed/run.py:766]
10
+ W0917 03:26:59.698000 948326 torch/distributed/run.py:766] *****************************************
11
+ W0917 03:26:59.698000 948326 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
12
+ W0917 03:26:59.698000 948326 torch/distributed/run.py:766] *****************************************
13
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
14
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
15
+ 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
16
+ WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer
17
+ Number of virtual stages per pipeline stage: None
18
+ accumulate and all-reduce gradients in fp32 for bfloat16 data type.
19
+ using torch.bfloat16 for parameters ...
20
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
21
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
22
+ ------------------------ arguments ------------------------
23
+ account_for_embedding_in_pipeline_split ......... False
24
+ account_for_loss_in_pipeline_split .............. False
25
+ accumulate_allreduce_grads_in_fp32 .............. True
26
+ adam_beta1 ...................................... 0.9
27
+ adam_beta2 ...................................... 0.999
28
+ adam_eps ........................................ 1e-08
29
+ add_bias_linear ................................. False
30
+ add_position_embedding .......................... False
31
+ add_qkv_bias .................................... False
32
+ adlr_autoresume ................................. False
33
+ adlr_autoresume_interval ........................ 1000
34
+ align_grad_reduce ............................... True
35
+ align_param_gather .............................. False
36
+ app_tag_run_name ................................ None
37
+ app_tag_run_version ............................. 0.0.0
38
+ apply_layernorm_1p .............................. False
39
+ apply_query_key_layer_scaling ................... False
40
+ apply_residual_connection_post_layernorm ........ False
41
+ apply_rope_fusion ............................... True
42
+ async_save ...................................... None
43
+ async_tensor_model_parallel_allreduce ........... True
44
+ attention_backend ............................... AttnBackend.auto
45
+ attention_dropout ............................... 0.0
46
+ attention_softmax_in_fp32 ....................... False
47
+ auto_detect_ckpt_format ......................... False
48
+ barrier_with_L1_time ............................ True
49
+ bert_binary_head ................................ True
50
+ bert_embedder_type .............................. megatron
51
+ bert_load ....................................... None
52
+ bf16 ............................................ True
53
+ bias_dropout_fusion ............................. True
54
+ bias_gelu_fusion ................................ False
55
+ bias_swiglu_fusion .............................. True
56
+ biencoder_projection_dim ........................ 0
57
+ biencoder_shared_query_context_model ............ False
58
+ block_data_path ................................. None
59
+ calc_ft_timeouts ................................ False
60
+ calculate_per_token_loss ........................ False
61
+ check_for_large_grads ........................... False
62
+ check_for_nan_in_loss_and_grad .................. False
63
+ check_for_spiky_loss ............................ False
64
+ check_weight_hash_across_dp_replicas_interval ... None
65
+ ckpt_assume_constant_structure .................. False
66
+ ckpt_convert_format ............................. None
67
+ ckpt_convert_save ............................... None
68
+ ckpt_convert_update_legacy_dist_opt_format ...... False
69
+ ckpt_format ..................................... torch_dist
70
+ ckpt_fully_parallel_load ........................ False
71
+ ckpt_fully_parallel_save ........................ True
72
+ ckpt_fully_parallel_save_deprecated ............. False
73
+ ckpt_step ....................................... None
74
+ classes_fraction ................................ 1.0
75
+ clip_grad ....................................... 1.0
76
+ clone_scatter_output_in_embedding ............... True
77
+ config_logger_dir ...............................
78
+ consumed_train_samples .......................... 0
79
+ consumed_valid_samples .......................... 0
80
+ context_parallel_size ........................... 1
81
+ cp_comm_type .................................... ['p2p']
82
+ create_attention_mask_in_dataloader ............. True
83
+ cross_entropy_fusion_impl ....................... native
84
+ cross_entropy_loss_fusion ....................... False
85
+ cuda_graph_scope ................................ full_iteration
86
+ cuda_graph_warmup_steps ......................... 3
87
+ data_args_path .................................. None
88
+ data_cache_path ................................. None
89
+ data_parallel_random_init ....................... False
90
+ data_parallel_sharding_strategy ................. no_shard
91
+ data_parallel_size .............................. 8
92
+ data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document']
93
+ data_per_class_fraction ......................... 1.0
94
+ data_sharding ................................... True
95
+ dataloader_type ................................. single
96
+ ddp_average_in_collective ....................... False
97
+ ddp_bucket_size ................................. None
98
+ ddp_num_buckets ................................. None
99
+ ddp_pad_buckets_for_high_nccl_busbw ............. False
100
+ decoder_first_pipeline_num_layers ............... None
101
+ decoder_last_pipeline_num_layers ................ None
102
+ decoder_num_layers .............................. None
103
+ decoder_seq_length .............................. None
104
+ decoupled_lr .................................... None
105
+ decoupled_min_lr ................................ None
106
+ decrease_batch_size_if_needed ................... False
107
+ defer_embedding_wgrad_compute ................... False
108
+ delay_wgrad_compute ............................. False
109
+ deprecated_use_mcore_models ..................... True
110
+ deterministic_mode .............................. False
111
+ dino_bottleneck_size ............................ 256
112
+ dino_freeze_last_layer .......................... 1
113
+ dino_head_hidden_size ........................... 2048
114
+ dino_local_crops_number ......................... 10
115
+ dino_local_img_size ............................. 96
116
+ dino_norm_last_layer ............................ False
117
+ dino_teacher_temp ............................... 0.07
118
+ dino_warmup_teacher_temp ........................ 0.04
119
+ dino_warmup_teacher_temp_epochs ................. 30
120
+ disable_bf16_reduced_precision_matmul ........... False
121
+ disable_mamba_mem_eff_path ...................... False
122
+ disable_straggler_on_startup .................... False
123
+ dist_ckpt_format_deprecated ..................... None
124
+ dist_ckpt_strictness ............................ assume_ok_unexpected
125
+ distribute_saved_activations .................... False
126
+ distributed_backend ............................. nccl
127
+ distributed_timeout_minutes ..................... 10
128
+ embedding_init_method_std ....................... None
129
+ embedding_path .................................. None
130
+ empty_unused_memory_level ....................... 0
131
+ enable_cuda_graph ............................... True
132
+ enable_experimental ............................. False
133
+ enable_ft_package ............................... False
134
+ enable_gloo_process_groups ...................... True
135
+ enable_msc ...................................... True
136
+ enable_one_logger ............................... True
137
+ encoder_num_layers .............................. 16
138
+ encoder_seq_length .............................. 4096
139
+ end_weight_decay ................................ 0.1
140
+ eod_mask_loss ................................... False
141
+ error_injection_rate ............................ 0
142
+ error_injection_type ............................ transient_error
143
+ eval_interval ................................... 1000000
144
+ eval_iters ...................................... 0
145
+ evidence_data_path .............................. None
146
+ exit_duration_in_mins ........................... None
147
+ exit_interval ................................... None
148
+ exit_on_missing_checkpoint ...................... False
149
+ exit_signal_handler ............................. False
150
+ exp_avg_dtype ................................... torch.float32
151
+ exp_avg_sq_dtype ................................ torch.float32
152
+ expert_model_parallel_size ...................... 8
153
+ expert_tensor_parallel_size ..................... 1
154
+ export_force_local_attention .................... False
155
+ export_kd_cfg ................................... None
156
+ export_kd_teacher_ckpt_format ................... None
157
+ export_kd_teacher_load .......................... None
158
+ export_kv_cache_quant ........................... False
159
+ export_legacy_megatron .......................... False
160
+ export_model_type ............................... GPTModel
161
+ export_moe_apply_probs_on_input ................. False
162
+ export_qk_l2_norm ............................... False
163
+ export_quant_cfg ................................ None
164
+ export_real_quant_cfg ........................... None
165
+ export_te_mcore_model ........................... False
166
+ external_cuda_graph ............................. False
167
+ ffn_hidden_size ................................. 14336
168
+ finetune ........................................ False
169
+ finetune_data_split ............................. train
170
+ finetune_hf_dataset ............................. None
171
+ first_last_layers_bf16 .......................... False
172
+ flash_decode .................................... False
173
+ fp16 ............................................ False
174
+ fp16_lm_cross_entropy ........................... False
175
+ fp32_residual_connection ........................ False
176
+ fp8 ............................................. e4m3
177
+ fp8_amax_compute_algo ........................... most_recent
178
+ fp8_amax_history_len ............................ 1
179
+ fp8_interval .................................... 1
180
+ fp8_margin ...................................... 0
181
+ fp8_param_gather ................................ False
182
+ fp8_recipe ...................................... mxfp8
183
+ fp8_wgrad ....................................... True
184
+ fsdp_double_buffer .............................. False
185
+ global_batch_size ............................... 16
186
+ grad_reduce_in_bf16 ............................. False
187
+ gradient_accumulation_fusion .................... True
188
+ gradient_reduce_div_fusion ...................... True
189
+ group_query_attention ........................... True
190
+ head_lr_mult .................................... 1.0
191
+ heterogeneous_layers_config_encoded_json ........ None
192
+ heterogeneous_layers_config_path ................ None
193
+ hidden_dropout .................................. 0.0
194
+ hidden_size ..................................... 4096
195
+ hierarchical_context_parallel_sizes ............. None
196
+ high_priority_stream_groups ..................... []
197
+ hybrid_attention_ratio .......................... 0.0
198
+ hybrid_mlp_ratio ................................ 0.0
199
+ hybrid_override_pattern ......................... None
200
+ hysteresis ...................................... 2
201
+ ict_head_size ................................... None
202
+ ict_load ........................................ None
203
+ img_h ........................................... 224
204
+ img_w ........................................... 224
205
+ indexer_batch_size .............................. 128
206
+ indexer_log_interval ............................ 1000
207
+ inference_batch_times_seqlen_threshold .......... -1
208
+ inference_dynamic_batching ...................... False
209
+ inference_dynamic_batching_buffer_guaranteed_fraction 0.2
210
+ inference_dynamic_batching_buffer_overflow_factor None
211
+ inference_dynamic_batching_buffer_size_gb ....... 40.0
212
+ inference_dynamic_batching_chunk_size ........... 256
213
+ inference_dynamic_batching_max_requests_override None
214
+ inference_dynamic_batching_max_tokens_override .. None
215
+ inference_dynamic_batching_num_cuda_graphs ...... 16
216
+ inference_max_batch_size ........................ 8
217
+ inference_max_seq_length ........................ 2560
218
+ inference_rng_tracker ........................... False
219
+ init_method_std ................................. 0.01
220
+ init_method_xavier_uniform ...................... False
221
+ init_model_with_meta_device ..................... False
222
+ initial_loss_scale .............................. 4294967296
223
+ inprocess_active_world_size ..................... 8
224
+ inprocess_barrier_timeout ....................... 120
225
+ inprocess_completion_timeout .................... 120
226
+ inprocess_empty_cuda_cache ...................... False
227
+ inprocess_granularity ........................... node
228
+ inprocess_hard_timeout .......................... 90
229
+ inprocess_heartbeat_interval .................... 30
230
+ inprocess_heartbeat_timeout ..................... 60
231
+ inprocess_last_call_wait ........................ 1
232
+ inprocess_max_iterations ........................ None
233
+ inprocess_monitor_process_interval .............. 1.0
234
+ inprocess_monitor_thread_interval ............... 1.0
235
+ inprocess_progress_watchdog_interval ............ 1.0
236
+ inprocess_restart ............................... False
237
+ inprocess_soft_timeout .......................... 60
238
+ inprocess_termination_grace_time ................ 1
239
+ is_hybrid_model ................................. False
240
+ iter_per_epoch .................................. 1250
241
+ iterations_to_skip .............................. []
242
+ keep_fp8_transpose_cache_when_using_custom_fsdp . False
243
+ kitchen_config_file ............................. None
244
+ kitchen_recipe_number ........................... None
245
+ kv_channels ..................................... 128
246
+ kv_lora_rank .................................... 32
247
+ lazy_mpu_init ................................... None
248
+ load ............................................ None
249
+ load_main_params_from_ckpt ...................... None
250
+ load_model_opt_format ........................... False
251
+ local_rank ...................................... 0
252
+ log_energy ...................................... False
253
+ log_interval .................................... 1
254
+ log_loss_scale_to_tensorboard ................... True
255
+ log_memory_to_tensorboard ....................... False
256
+ log_num_zeros_in_grad ........................... False
257
+ log_params_norm ................................. False
258
+ log_progress .................................... False
259
+ log_straggler ................................... False
260
+ log_throughput .................................. True
261
+ log_timers_to_tensorboard ....................... False
262
+ log_validation_ppl_to_tensorboard ............... False
263
+ log_world_size_to_tensorboard ................... False
264
+ logging_level ................................... None
265
+ loss_scale ...................................... None
266
+ loss_scale_window ............................... 1000
267
+ lr .............................................. 0.0001
268
+ lr_decay_iters .................................. 8
269
+ lr_decay_samples ................................ None
270
+ lr_decay_style .................................. cosine
271
+ lr_warmup_fraction .............................. None
272
+ lr_warmup_init .................................. 0.0
273
+ lr_warmup_iters ................................. 5
274
+ lr_warmup_samples ............................... 0
275
+ lr_wsd_decay_iters .............................. None
276
+ lr_wsd_decay_samples ............................ None
277
+ lr_wsd_decay_style .............................. exponential
278
+ main_grads_dtype ................................ torch.float32
279
+ main_params_dtype ............................... torch.float32
280
+ make_vocab_size_divisible_by .................... 128
281
+ mamba_head_dim .................................. 64
282
+ mamba_num_groups ................................ 8
283
+ mamba_num_heads ................................. None
284
+ mamba_state_dim ................................. 128
285
+ manual_gc ....................................... False
286
+ manual_gc_eval .................................. True
287
+ manual_gc_interval .............................. 0
288
+ mask_factor ..................................... 1.0
289
+ mask_prob ....................................... 0.15
290
+ mask_type ....................................... random
291
+ masked_softmax_fusion ........................... False
292
+ max_position_embeddings ......................... 32768
293
+ max_tokens_to_oom ............................... 12000
294
+ memory_snapshot_path ............................ snapshot.pickle
295
+ merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt
296
+ micro_batch_size ................................ 1
297
+ microbatch_group_size_per_vp_stage .............. None
298
+ mid_level_dataset_surplus ....................... 0.005
299
+ min_loss_scale .................................. 1.0
300
+ min_lr .......................................... 1e-05
301
+ mlp_chunks_for_prefill .......................... 1
302
+ mmap_bin_files .................................. True
303
+ mock_data ....................................... False
304
+ moe_apply_probs_on_input ........................ False
305
+ moe_aux_loss_coeff .............................. 0.01
306
+ moe_capacity_monitor ............................ True
307
+ moe_deepep_num_sms .............................. 20
308
+ moe_echo_enable_random_offloading ............... False
309
+ moe_echo_expert_dispatch_overlap ................ False
310
+ moe_echo_expert_dispatcher_type ................. hybridep
311
+ moe_echo_recompute_expert_dispatch .............. False
312
+ moe_enable_deepep ............................... False
313
+ moe_enable_echo ................................. False
314
+ moe_enable_hybridep ............................. True
315
+ moe_enable_ultraep .............................. True
316
+ moe_expert_capacity_factor ...................... None
317
+ moe_extended_tp ................................. False
318
+ moe_ffn_hidden_size ............................. 3584
319
+ moe_grouped_gemm ................................ True
320
+ moe_hybridep_num_sms ............................ 16
321
+ moe_input_jitter_eps ............................ None
322
+ moe_layer_freq .................................. 1
323
+ moe_layer_recompute ............................. False
324
+ moe_num_echo_experts ............................ None
325
+ moe_num_redundant_experts_per_rank .............. 4
326
+ moe_pad_expert_input_to_capacity ................ False
327
+ moe_per_layer_logging ........................... False
328
+ moe_permute_fusion .............................. False
329
+ moe_received_token_capacity ..................... 1.15625
330
+ moe_router_bias_update_rate ..................... 0.001
331
+ moe_router_dtype ................................ fp32
332
+ moe_router_enable_expert_bias ................... False
333
+ moe_router_force_load_balancing ................. False
334
+ moe_router_group_topk ........................... None
335
+ moe_router_load_balancing_type .................. aux_loss
336
+ moe_router_num_groups ........................... None
337
+ moe_router_padding_for_fp8 ...................... False
338
+ moe_router_pre_softmax .......................... False
339
+ moe_router_score_function ....................... softmax
340
+ moe_router_topk ................................. 2
341
+ moe_router_topk_scaling_factor .................. None
342
+ moe_shared_expert_intermediate_size ............. None
343
+ moe_shared_expert_overlap ....................... False
344
+ moe_token_dispatcher_type ....................... flex
345
+ moe_token_drop_policy ........................... probs
346
+ moe_ultraep_max_microbatches .................... None
347
+ moe_ultraep_overlap_weight_sync ................. True
348
+ moe_ultraep_weight_sync_join .................... before_dispatch
349
+ moe_upcycling_granularity ....................... 1
350
+ moe_use_device_initiated_grouped_gemm ........... True
351
+ moe_use_legacy_grouped_gemm ..................... False
352
+ moe_use_upcycling ............................... False
353
+ moe_z_loss_coeff ................................ None
354
+ mrope_section ................................... None
355
+ mscale .......................................... 1.0
356
+ mscale_all_dim .................................. 0.0
357
+ mtp_loss_scaling_factor ......................... 0.1
358
+ mtp_num_layers .................................. None
359
+ multi_latent_attention .......................... False
360
+ nccl_all_reduce_for_prefill ..................... False
361
+ nccl_communicator_config_path ................... None
362
+ nccl_ub ......................................... False
363
+ no_load_optim ................................... None
364
+ no_load_rng ..................................... None
365
+ no_persist_layer_norm ........................... False
366
+ no_rope_freq .................................... None
367
+ no_save_optim ................................... None
368
+ no_save_rng ..................................... None
369
+ non_persistent_ckpt_type ........................ None
370
+ non_persistent_global_ckpt_dir .................. None
371
+ non_persistent_local_ckpt_algo .................. fully_parallel
372
+ non_persistent_local_ckpt_dir ................... None
373
+ non_persistent_save_interval .................... None
374
+ norm_epsilon .................................... 1e-05
375
+ normalization ................................... RMSNorm
376
+ num_attention_heads ............................. 32
377
+ num_channels .................................... 3
378
+ num_classes ..................................... 1000
379
+ num_dataset_builder_threads ..................... 1
380
+ num_distributed_optimizer_instances ............. 1
381
+ num_experts ..................................... 32
382
+ num_layers ...................................... 16
383
+ num_layers_at_end_in_bf16 ....................... 1
384
+ num_layers_at_start_in_bf16 ..................... 1
385
+ num_layers_per_virtual_pipeline_stage ........... None
386
+ num_query_groups ................................ 8
387
+ num_virtual_stages_per_pipeline_rank ............ None
388
+ num_workers ..................................... 2
389
+ object_storage_cache_path ....................... None
390
+ one_logger_async ................................ False
391
+ one_logger_project .............................. megatron-lm
392
+ one_logger_run_name ............................. None
393
+ onnx_safe ....................................... None
394
+ openai_gelu ..................................... False
395
+ optimizer ....................................... adam
396
+ optimizer_cpu_offload ........................... False
397
+ optimizer_offload_fraction ...................... 1.0
398
+ output_bert_embeddings .......................... False
399
+ overlap_cpu_optimizer_d2h_h2d ................... False
400
+ overlap_grad_reduce ............................. True
401
+ overlap_moe_expert_parallel_comm ................ False
402
+ overlap_p2p_comm ................................ False
403
+ overlap_p2p_comm_warmup_flush ................... False
404
+ overlap_param_gather ............................ True
405
+ overlap_param_gather_with_optimizer_step ........ False
406
+ override_opt_param_scheduler .................... False
407
+ params_dtype .................................... torch.bfloat16
408
+ patch_dim ....................................... 16
409
+ per_split_data_args_path ........................ None
410
+ perform_initialization .......................... True
411
+ pin_cpu_grads ................................... True
412
+ pin_cpu_params .................................. True
413
+ pipeline_model_parallel_comm_backend ............ None
414
+ pipeline_model_parallel_layout .................. None
415
+ pipeline_model_parallel_size .................... 1
416
+ position_embedding_type ......................... rope
417
+ pretrained_checkpoint ........................... None
418
+ profile ......................................... True
419
+ profile_ranks ................................... [0, 1, 2, 3, 4, 5, 6, 7]
420
+ profile_step_end ................................ 8
421
+ profile_step_start .............................. 6
422
+ q_lora_rank ..................................... None
423
+ qk_head_dim ..................................... 128
424
+ qk_l2_norm ...................................... False
425
+ qk_layernorm .................................... False
426
+ qk_pos_emb_head_dim ............................. 64
427
+ query_in_block_prob ............................. 0.1
428
+ rampup_batch_size ............................... None
429
+ rank ............................................ 0
430
+ recompute_granularity ........................... None
431
+ recompute_method ................................ None
432
+ recompute_modules ............................... None
433
+ recompute_num_layers ............................ None
434
+ record_memory_history ........................... False
435
+ relative_attention_max_distance ................. 128
436
+ relative_attention_num_buckets .................. 32
437
+ replication ..................................... False
438
+ replication_factor .............................. 2
439
+ replication_jump ................................ None
440
+ rerun_mode ...................................... validate_results
441
+ reset_attention_mask ............................ False
442
+ reset_position_ids .............................. False
443
+ result_rejected_tracker_filename ................ None
444
+ retriever_report_topk_accuracies ................ []
445
+ retriever_score_scaling ......................... False
446
+ retriever_seq_length ............................ 256
447
+ retro_add_retriever ............................. False
448
+ retro_attention_gate ............................ 1
449
+ retro_cyclic_train_iters ........................ None
450
+ retro_encoder_attention_dropout ................. 0.1
451
+ retro_encoder_hidden_dropout .................... 0.1
452
+ retro_encoder_layers ............................ 2
453
+ retro_num_neighbors ............................. 2
454
+ retro_num_retrieved_chunks ...................... 2
455
+ retro_project_dir ............................... None
456
+ retro_verify_neighbor_count ..................... True
457
+ reuse_grad_buf_for_mxfp8_param_ag ............... False
458
+ rope_scaling_factor ............................. 8.0
459
+ rope_type ....................................... None
460
+ rotary_base ..................................... 1000000
461
+ rotary_interleaved .............................. False
462
+ rotary_percent .................................. 1.0
463
+ rotary_scaling_factor ........................... 1.0
464
+ rotary_seq_len_interpolation_factor ............. None
465
+ run_workload_inspector_server ................... False
466
+ sample_rate ..................................... 1.0
467
+ save ............................................ None
468
+ save_interval ................................... None
469
+ save_retain_interval ............................ None
470
+ scatter_gather_tensors_in_pipeline .............. True
471
+ seed ............................................ 1234
472
+ seq_length ...................................... 4096
473
+ sequence_parallel ............................... False
474
+ sft ............................................. False
475
+ sft_tokenizer_prompt_format ..................... nemotron-h-aligned
476
+ sgd_momentum .................................... 0.9
477
+ short_seq_prob .................................. 0.1
478
+ skip_train ...................................... False
479
+ skipped_train_samples ........................... 0
480
+ spec ............................................ None
481
+ split ........................................... 100,0,0
482
+ squared_relu .................................... False
483
+ start_weight_decay .............................. 0.1
484
+ straggler_ctrlr_port ............................ 65535
485
+ straggler_minmax_count .......................... 1
486
+ suggested_communication_unit_size ............... None
487
+ swiglu .......................................... True
488
+ swin_backbone_type .............................. tiny
489
+ symmetric_ar_type ............................... None
490
+ te_rng_tracker .................................. True
491
+ tensor_model_parallel_size ...................... 1
492
+ tensorboard_dir ................................. None
493
+ tensorboard_log_interval ........................ 1
494
+ tensorboard_queue_size .......................... 1000
495
+ test_data_path .................................. None
496
+ test_mode ....................................... False
497
+ tiktoken_num_special_tokens ..................... 1000
498
+ tiktoken_pattern ................................ None
499
+ tiktoken_special_tokens ......................... None
500
+ timing_log_level ................................ 0
501
+ timing_log_option ............................... minmax
502
+ titles_data_path ................................ None
503
+ tokenizer_model ................................. None
504
+ tokenizer_type .................................. GPT2BPETokenizer
505
+ torch_fsdp2_reshard_after_forward ............... True
506
+ tp_comm_bootstrap_backend ....................... nccl
507
+ tp_comm_bulk_dgrad .............................. True
508
+ tp_comm_bulk_wgrad .............................. True
509
+ tp_comm_overlap ................................. False
510
+ tp_comm_overlap_ag .............................. True
511
+ tp_comm_overlap_cfg ............................. None
512
+ tp_comm_overlap_rs .............................. True
513
+ tp_comm_overlap_rs_dgrad ........................ False
514
+ tp_comm_split_ag ................................ True
515
+ tp_comm_split_rs ................................ True
516
+ train_data_path ................................. None
517
+ train_iters ..................................... 10
518
+ train_samples ................................... None
519
+ train_sync_interval ............................. None
520
+ transformer_impl ................................ transformer_engine
521
+ transformer_pipeline_model_parallel_size ........ 1
522
+ untie_embeddings_and_output_weights ............. True
523
+ use_checkpoint_args ............................. False
524
+ use_checkpoint_opt_param_scheduler .............. False
525
+ use_cpu_initialization .......................... None
526
+ use_custom_fsdp ................................. False
527
+ use_dist_ckpt ................................... True
528
+ use_dist_ckpt_deprecated ........................ False
529
+ use_distributed_optimizer ....................... True
530
+ use_flash_attn .................................. False
531
+ use_legacy_models ............................... False
532
+ use_mp_args_from_checkpoint_args ................ False
533
+ use_one_sent_docs ............................... False
534
+ use_persistent_ckpt_worker ...................... False
535
+ use_precision_aware_optimizer ................... False
536
+ use_pytorch_profiler ............................ False
537
+ use_ring_exchange_p2p ........................... False
538
+ use_rope_scaling ................................ False
539
+ use_rotary_position_embeddings .................. False
540
+ use_sharp ....................................... False
541
+ use_tokenizer_model_from_checkpoint_args ........ True
542
+ use_torch_fsdp2 ................................. False
543
+ use_torch_optimizer_for_cpu_offload ............. False
544
+ use_tp_pp_dp_mapping ............................ False
545
+ v_head_dim ...................................... 128
546
+ valid_data_path ................................. None
547
+ variable_seq_lengths ............................ False
548
+ virtual_pipeline_model_parallel_size ............ None
549
+ vision_backbone_type ............................ vit
550
+ vision_pretraining .............................. False
551
+ vision_pretraining_type ......................... classify
552
+ vocab_extra_ids ................................. 0
553
+ vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json
554
+ vocab_size ...................................... None
555
+ wandb_exp_name ..................................
556
+ wandb_project ...................................
557
+ wandb_save_dir ..................................
558
+ weight_decay .................................... 0.1
559
+ weight_decay_incr_style ......................... constant
560
+ wgrad_deferral_limit ............................ 0
561
+ world_size ...................................... 8
562
+ yaml_cfg ........................................ None
563
+ -------------------- end of arguments ---------------------
564
+ INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 2
565
+ > building GPT2BPETokenizer tokenizer ...
566
+ > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304)
567
+ WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS
568
+ > initializing torch distributed ...
569
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
570
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
571
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
572
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
573
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
574
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
575
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
576
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
577
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
578
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
579
+ 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
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
+ [Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
583
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
584
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
585
+ [Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
586
+ [Gloo] Rank 7[Gloo] Rank is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
587
+ 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
588
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
589
+ [Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
590
+ [Gloo] Rank 0 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 [Gloo] Rank 45 is connected to is connected to 77 peer ranks. peer ranks. Expected number of connected peer ranks is : 7Expected number of connected peer ranks is : 7
593
+
594
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
595
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
596
+ [Gloo] Rank [Gloo] Rank 76 is connected to is connected to 77 peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : 77
597
+
598
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
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 : 0
603
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
604
+ > initialized tensor model parallel with size 1
605
+ > initialized pipeline model parallel with size 1
606
+ > setting random seeds to 1234 ...
607
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
608
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
609
+ [rank3]:[W917 03:27:12.759074705 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
610
+ [rank1]:[W917 03:27:12.759291933 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
611
+ [rank5]:[W917 03:27:12.759731800 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 5] using GPU 5 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
612
+ [rank6]:[W917 03:27:12.760213509 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 6] using GPU 6 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
613
+ [rank2]:[W917 03:27:12.763662209 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 2] using GPU 2 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
614
+ [rank4]:[W917 03:27:12.763770125 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 4] using GPU 4 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
615
+ [rank7]:[W917 03:27:12.763810340 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
616
+ > compiling dataset index builder ...
617
+ make: Entering directory '/workspace/Megatron-UltraEP/megatron/core/datasets'
618
+ make: Nothing to be done for 'default'.
619
+ make: Leaving directory '/workspace/Megatron-UltraEP/megatron/core/datasets'
620
+ >>> done with dataset index builder. Compilation time: 0.437 seconds
621
+ WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations.
622
+ > compiling and loading fused kernels ...
623
+ [rank0]:[W917 03:27:13.386048103 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 0] using GPU 0 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
624
+ >>> done with compiling and loading fused kernels. Compilation time: 4.013 seconds
625
+ time to initialize megatron (seconds): 9.439
626
+ [after megatron is initialized] datetime: 2026-09-17 03:27:19
627
+ building GPT model ...
628
+ /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.
629
+ warnings.warn(
630
+ /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.
631
+ warnings.warn(
632
+ /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.
633
+ warnings.warn(
634
+ /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.
635
+ warnings.warn(
636
+ /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.
637
+ warnings.warn(
638
+ /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.
639
+ warnings.warn(
640
+ /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.
641
+ warnings.warn(
642
+ /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.
643
+ warnings.warn(
644
+ [INFO] Use MNNVL fabric: False
645
+ [WARN] MAX_NUM_NVL_PEERS is not set. Using detected value 8.
646
+ NVSHMEM v3.7.2
647
+ /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.
648
+
649
+ /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.
650
+
651
+ /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.
652
+
653
+ /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.
654
+
655
+ /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.
656
+
657
+ /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.
658
+
659
+ /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.
660
+
661
+ /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.
662
+
663
+ UltraEP: NVL domain size <= 8; forcing weight_sync plan mode to direct (was adaptive).
664
+ > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 6722555904
665
+ 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)
666
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 18
667
+ Params for bucket 1 (206045184 elements, 206045184 padded size):
668
+ module.output_layer.weight
669
+ Params for bucket 2 (42086400 elements, 42086400 padded size):
670
+ module.decoder.layers.15.self_attention.linear_qkv.weight
671
+ module.decoder.layers.15.pre_mlp_layernorm.weight
672
+ module.decoder.final_layernorm.weight
673
+ module.decoder.layers.15.self_attention.linear_qkv.layer_norm_weight
674
+ module.decoder.layers.15.mlp.router.weight
675
+ module.decoder.layers.15.self_attention.linear_proj.weight
676
+ Params for bucket 3 (42082304 elements, 42082304 padded size):
677
+ module.decoder.layers.14.self_attention.linear_qkv.weight
678
+ module.decoder.layers.14.pre_mlp_layernorm.weight
679
+ module.decoder.layers.14.mlp.router.weight
680
+ module.decoder.layers.14.self_attention.linear_proj.weight
681
+ module.decoder.layers.14.self_attention.linear_qkv.layer_norm_weight
682
+ Params for bucket 4 (42082304 elements, 42082304 padded size):
683
+ module.decoder.layers.13.pre_mlp_layernorm.weight
684
+ module.decoder.layers.13.self_attention.linear_qkv.weight
685
+ module.decoder.layers.13.mlp.router.weight
686
+ module.decoder.layers.13.self_attention.linear_proj.weight
687
+ module.decoder.layers.13.self_attention.linear_qkv.layer_norm_weight
688
+ Params for bucket 5 (42082304 elements, 42082304 padded size):
689
+ module.decoder.layers.12.self_attention.linear_qkv.weight
690
+ module.decoder.layers.12.pre_mlp_layernorm.weight
691
+ module.decoder.layers.12.mlp.router.weight
692
+ module.decoder.layers.12.self_attention.linear_proj.weight
693
+ module.decoder.layers.12.self_attention.linear_qkv.layer_norm_weight
694
+ Params for bucket 6 (42082304 elements, 42082304 padded size):
695
+ module.decoder.layers.11.self_attention.linear_qkv.weight
696
+ module.decoder.layers.11.pre_mlp_layernorm.weight
697
+ module.decoder.layers.11.mlp.router.weight
698
+ module.decoder.layers.11.self_attention.linear_proj.weight
699
+ module.decoder.layers.11.self_attention.linear_qkv.layer_norm_weight
700
+ Params for bucket 7 (42082304 elements, 42082304 padded size):
701
+ module.decoder.layers.10.self_attention.linear_qkv.weight
702
+ module.decoder.layers.10.pre_mlp_layernorm.weight
703
+ module.decoder.layers.10.mlp.router.weight
704
+ module.decoder.layers.10.self_attention.linear_proj.weight
705
+ module.decoder.layers.10.self_attention.linear_qkv.layer_norm_weight
706
+ Params for bucket 8 (42082304 elements, 42082304 padded size):
707
+ module.decoder.layers.9.self_attention.linear_qkv.weight
708
+ module.decoder.layers.9.pre_mlp_layernorm.weight
709
+ module.decoder.layers.9.mlp.router.weight
710
+ module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight
711
+ module.decoder.layers.9.self_attention.linear_proj.weight
712
+ Params for bucket 9 (42082304 elements, 42082304 padded size):
713
+ module.decoder.layers.8.self_attention.linear_qkv.weight
714
+ module.decoder.layers.8.pre_mlp_layernorm.weight
715
+ module.decoder.layers.8.mlp.router.weight
716
+ module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight
717
+ module.decoder.layers.8.self_attention.linear_proj.weight
718
+ Params for bucket 10 (42082304 elements, 42082304 padded size):
719
+ module.decoder.layers.7.mlp.router.weight
720
+ module.decoder.layers.7.self_attention.linear_proj.weight
721
+ module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight
722
+ module.decoder.layers.7.self_attention.linear_qkv.weight
723
+ module.decoder.layers.7.pre_mlp_layernorm.weight
724
+ Params for bucket 11 (42082304 elements, 42082304 padded size):
725
+ module.decoder.layers.6.mlp.router.weight
726
+ module.decoder.layers.6.self_attention.linear_proj.weight
727
+ module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight
728
+ module.decoder.layers.6.self_attention.linear_qkv.weight
729
+ module.decoder.layers.6.pre_mlp_layernorm.weight
730
+ Params for bucket 12 (42082304 elements, 42082304 padded size):
731
+ module.decoder.layers.5.self_attention.linear_qkv.weight
732
+ module.decoder.layers.5.self_attention.linear_proj.weight
733
+ module.decoder.layers.5.pre_mlp_layernorm.weight
734
+ module.decoder.layers.5.mlp.router.weight
735
+ module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight
736
+ Params for bucket 13 (42082304 elements, 42082304 padded size):
737
+ module.decoder.layers.4.mlp.router.weight
738
+ module.decoder.layers.4.self_attention.linear_proj.weight
739
+ module.decoder.layers.4.self_attention.linear_qkv.layer_norm_weight
740
+ module.decoder.layers.4.self_attention.linear_qkv.weight
741
+ module.decoder.layers.4.pre_mlp_layernorm.weight
742
+ Params for bucket 14 (42082304 elements, 42082304 padded size):
743
+ module.decoder.layers.3.mlp.router.weight
744
+ module.decoder.layers.3.self_attention.linear_proj.weight
745
+ module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight
746
+ module.decoder.layers.3.self_attention.linear_qkv.weight
747
+ module.decoder.layers.3.pre_mlp_layernorm.weight
748
+ Params for bucket 15 (42082304 elements, 42082304 padded size):
749
+ module.decoder.layers.2.mlp.router.weight
750
+ module.decoder.layers.2.self_attention.linear_proj.weight
751
+ module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight
752
+ module.decoder.layers.2.pre_mlp_layernorm.weight
753
+ module.decoder.layers.2.self_attention.linear_qkv.weight
754
+ Params for bucket 16 (42082304 elements, 42082304 padded size):
755
+ module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight
756
+ module.decoder.layers.1.self_attention.linear_proj.weight
757
+ module.decoder.layers.1.mlp.router.weight
758
+ module.decoder.layers.1.self_attention.linear_qkv.weight
759
+ module.decoder.layers.1.pre_mlp_layernorm.weight
760
+ Params for bucket 17 (42082304 elements, 42082304 padded size):
761
+ module.decoder.layers.0.pre_mlp_layernorm.weight
762
+ module.decoder.layers.0.self_attention.linear_proj.weight
763
+ module.decoder.layers.0.mlp.router.weight
764
+ module.decoder.layers.0.self_attention.linear_qkv.weight
765
+ module.decoder.layers.0.self_attention.linear_qkv.layer_norm_weight
766
+ Params for bucket 18 (206045184 elements, 206045184 padded size):
767
+ module.embedding.word_embeddings.weight
768
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 56
769
+ Params for bucket 1 (44040192 elements, 44040192 padded size):
770
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight3
771
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight1
772
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight2
773
+ Params for bucket 2 (44040192 elements, 44040192 padded size):
774
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight0
775
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight3
776
+ Params for bucket 3 (58720256 elements, 58720256 padded size):
777
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight2
778
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight1
779
+ Params for bucket 4 (44040192 elements, 44040192 padded size):
780
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight0
781
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight3
782
+ Params for bucket 5 (44040192 elements, 44040192 padded size):
783
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight2
784
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight0
785
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight1
786
+ Params for bucket 6 (58720256 elements, 58720256 padded size):
787
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight3
788
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight2
789
+ Params for bucket 7 (58720256 elements, 58720256 padded size):
790
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight1
791
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight0
792
+ Params for bucket 8 (44040192 elements, 44040192 padded size):
793
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight3
794
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight1
795
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight2
796
+ Params for bucket 9 (44040192 elements, 44040192 padded size):
797
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight0
798
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight3
799
+ Params for bucket 10 (58720256 elements, 58720256 padded size):
800
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight2
801
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight1
802
+ Params for bucket 11 (44040192 elements, 44040192 padded size):
803
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight0
804
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight3
805
+ Params for bucket 12 (44040192 elements, 44040192 padded size):
806
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight2
807
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight0
808
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight1
809
+ Params for bucket 13 (58720256 elements, 58720256 padded size):
810
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight3
811
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight2
812
+ Params for bucket 14 (58720256 elements, 58720256 padded size):
813
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight1
814
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight0
815
+ Params for bucket 15 (44040192 elements, 44040192 padded size):
816
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight3
817
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight1
818
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight2
819
+ Params for bucket 16 (44040192 elements, 44040192 padded size):
820
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight0
821
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight3
822
+ Params for bucket 17 (58720256 elements, 58720256 padded size):
823
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight2
824
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight1
825
+ Params for bucket 18 (44040192 elements, 44040192 padded size):
826
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight0
827
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight3
828
+ Params for bucket 19 (44040192 elements, 44040192 padded size):
829
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight2
830
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight1
831
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight0
832
+ Params for bucket 20 (58720256 elements, 58720256 padded size):
833
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight3
834
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight2
835
+ Params for bucket 21 (58720256 elements, 58720256 padded size):
836
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight1
837
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight0
838
+ Params for bucket 22 (44040192 elements, 44040192 padded size):
839
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight3
840
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight1
841
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight2
842
+ Params for bucket 23 (44040192 elements, 44040192 padded size):
843
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight0
844
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight3
845
+ Params for bucket 24 (58720256 elements, 58720256 padded size):
846
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight2
847
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight1
848
+ Params for bucket 25 (44040192 elements, 44040192 padded size):
849
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight0
850
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight3
851
+ Params for bucket 26 (44040192 elements, 44040192 padded size):
852
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight2
853
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight1
854
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight0
855
+ Params for bucket 27 (58720256 elements, 58720256 padded size):
856
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight3
857
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight2
858
+ Params for bucket 28 (58720256 elements, 58720256 padded size):
859
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight1
860
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight0
861
+ Params for bucket 29 (44040192 elements, 44040192 padded size):
862
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight3
863
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight1
864
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight2
865
+ Params for bucket 30 (44040192 elements, 44040192 padded size):
866
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight0
867
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight3
868
+ Params for bucket 31 (58720256 elements, 58720256 padded size):
869
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight2
870
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight1
871
+ Params for bucket 32 (44040192 elements, 44040192 padded size):
872
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight0
873
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight3
874
+ Params for bucket 33 (44040192 elements, 44040192 padded size):
875
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight2
876
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight1
877
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight0
878
+ Params for bucket 34 (58720256 elements, 58720256 padded size):
879
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight3
880
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight2
881
+ Params for bucket 35 (58720256 elements, 58720256 padded size):
882
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight1
883
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight0
884
+ Params for bucket 36 (44040192 elements, 44040192 padded size):
885
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight3
886
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight1
887
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight2
888
+ Params for bucket 37 (44040192 elements, 44040192 padded size):
889
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight0
890
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight3
891
+ Params for bucket 38 (58720256 elements, 58720256 padded size):
892
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight2
893
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight1
894
+ Params for bucket 39 (44040192 elements, 44040192 padded size):
895
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight0
896
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight3
897
+ Params for bucket 40 (44040192 elements, 44040192 padded size):
898
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight2
899
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight1
900
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight0
901
+ Params for bucket 41 (58720256 elements, 58720256 padded size):
902
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight3
903
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight2
904
+ Params for bucket 42 (58720256 elements, 58720256 padded size):
905
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight1
906
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight0
907
+ Params for bucket 43 (44040192 elements, 44040192 padded size):
908
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight3
909
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight1
910
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight2
911
+ Params for bucket 44 (44040192 elements, 44040192 padded size):
912
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight0
913
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight3
914
+ Params for bucket 45 (58720256 elements, 58720256 padded size):
915
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight2
916
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight1
917
+ Params for bucket 46 (44040192 elements, 44040192 padded size):
918
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight0
919
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight3
920
+ Params for bucket 47 (44040192 elements, 44040192 padded size):
921
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight2
922
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight0
923
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight1
924
+ Params for bucket 48 (58720256 elements, 58720256 padded size):
925
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight3
926
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight2
927
+ Params for bucket 49 (58720256 elements, 58720256 padded size):
928
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight1
929
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight0
930
+ Params for bucket 50 (44040192 elements, 44040192 padded size):
931
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight3
932
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight1
933
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight2
934
+ Params for bucket 51 (44040192 elements, 44040192 padded size):
935
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight0
936
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight3
937
+ Params for bucket 52 (58720256 elements, 58720256 padded size):
938
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight2
939
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight1
940
+ Params for bucket 53 (44040192 elements, 44040192 padded size):
941
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight0
942
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight3
943
+ Params for bucket 54 (44040192 elements, 44040192 padded size):
944
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight2
945
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight0
946
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight1
947
+ Params for bucket 55 (58720256 elements, 58720256 padded size):
948
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight3
949
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight2
950
+ Params for bucket 56 (58720256 elements, 58720256 padded size):
951
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight1
952
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight0
953
+ 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 0x7de4a8c9d520>, config_logger_dir='')
954
+ INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine
955
+ [after model, optimizer, and learning rate scheduler are built] datetime: 2026-09-17 03:27:22
956
+ > building train, validation, and test datasets ...
957
+ > datasets target sizes (minimum size):
958
+ train: 160
959
+ validation: 0
960
+ test: 0
961
+ INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None]
962
+ > building train, validation, and test datasets for GPT ...
963
+ INFO:megatron.core.datasets.blended_megatron_dataset_builder:Building GPTDataset splits with sizes=(160, 0, 0) and config=GPTDatasetConfig(random_seed=1234, sequence_length=4096, blend=(['/workspace/data/wikitext/wikitext103_gpt2_text_document'], None), blend_per_split=None, split='100,0,0', split_matrix=[(0, 1.0), None, None], num_dataset_builder_threads=1, path_to_cache=None, mmap_bin_files=True, mock=False, tokenizer=<megatron.training.tokenizer.tokenizer._GPT2BPETokenizer object at 0x7de4abd3b290>, 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)
964
+ INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx
965
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths
966
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers
967
+ INFO:megatron.core.datasets.indexed_dataset: Extract the document indices
968
+ INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029
969
+ INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350
970
+ INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices
971
+ INFO:megatron.core.datasets.gpt_dataset: Load the document index from 68684a515e118253a8572c4517cf2ffe-GPTDataset-train-document_index.npy
972
+ INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 68684a515e118253a8572c4517cf2ffe-GPTDataset-train-sample_index.npy
973
+ INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 68684a515e118253a8572c4517cf2ffe-GPTDataset-train-shuffle_index.npy
974
+ INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073
975
+ > finished creating GPT datasets ...
976
+ [after dataloaders are built] datetime: 2026-09-17 03:27:22
977
+ done with setup ...
978
+ (min, max) time across ranks (ms):
979
+ model-and-optimizer-setup ......................: (3347.95, 3351.05)
980
+ train/valid/test-data-iterators-setup ..........: (113.53, 139.33)
981
+ training ...
982
+ Overwriting rerun_state_machine.current_iteration from -1 to 0...
983
+ [before the start of training step] datetime: 2026-09-17 03:27:23
984
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
985
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
986
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
987
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
988
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
989
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
990
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
991
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
992
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
993
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
994
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
995
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
996
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
997
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
998
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
999
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1000
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1001
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1002
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1003
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1004
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1005
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1006
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1007
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1008
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1009
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1010
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1011
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1012
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1013
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1014
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1015
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1016
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1017
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1018
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1019
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1020
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1021
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1022
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1023
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1024
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1025
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1026
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1027
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1028
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1029
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1030
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1031
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1032
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1033
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1034
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1035
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1036
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1037
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1038
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1039
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1040
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1041
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1042
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1043
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1044
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1045
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1046
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1047
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1048
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1049
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1050
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1051
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1052
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1053
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1054
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1055
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1056
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1057
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1058
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1059
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1060
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1061
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1062
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1063
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1064
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1065
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1066
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1067
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1068
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1069
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1070
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1071
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1072
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1073
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1074
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1075
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1076
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1077
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1078
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1079
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1080
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1081
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1082
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1083
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1084
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1085
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1086
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1087
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1088
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1089
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1090
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1091
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1092
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1093
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1094
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1095
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1096
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1097
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1098
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1099
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1100
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1101
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1102
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1103
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1104
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1105
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1106
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1107
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1108
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1109
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1110
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1111
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1112
+ [2026-09-17 03:27:52] iteration 1/ 10 | consumed samples: 16 | elapsed time per iteration (ms): 29813.9 | throughput per GPU (TFLOP/s/GPU): 4.2 | learning rate: 2.000000E-05 | global batch size: 16 | lm loss: 1.102902E+01 | loss scale: 1.0 | grad norm: 44.186 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 8832 | min safe capacity: 1.0781 |
1113
+ Number of parameters in transformer block in billions: 23.22
1114
+ Number of parameters in embedding layers in billions: 0.41
1115
+ Total number of parameters in billions: 23.63
1116
+ Number of parameters in most loaded shard in billions: 23.6320
1117
+ Theoretical memory footprints: weight and optimizer=169029.43 MB
1118
+ [Rank 0] (after 1 iterations) memory (MB) | allocated: 58124.0458984375 | max allocated: 65718.16796875 | reserved: 79458.0 | max reserved: 79458.0
1119
+ [2026-09-17 03:27:53] iteration 2/ 10 | consumed samples: 32 | elapsed time per iteration (ms): 586.2 | throughput per GPU (TFLOP/s/GPU): 214.2 | learning rate: 4.000000E-05 | global batch size: 16 | lm loss: 1.101823E+01 | loss scale: 1.0 | grad norm: 45.729 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 8832 | min safe capacity: 1.0781 |
1120
+ [2026-09-17 03:27:53] iteration 3/ 10 | consumed samples: 48 | elapsed time per iteration (ms): 434.7 | throughput per GPU (TFLOP/s/GPU): 288.9 | learning rate: 6.000000E-05 | global batch size: 16 | lm loss: 9.972256E+00 | loss scale: 1.0 | grad norm: 157.296 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 8832 | min safe capacity: 1.0781 |
1121
+ Capture CUDA graph for training!!!
1122
+ CUDA graph capture done!!!
1123
+ [2026-09-17 03:27:57] iteration 4/ 10 | consumed samples: 64 | elapsed time per iteration (ms): 3115.4 | throughput per GPU (TFLOP/s/GPU): 40.3 | learning rate: 8.000000E-05 | global batch size: 16 | lm loss: 1.311298E+01 | loss scale: 1.0 | grad norm: 41.251 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 8832 | min safe capacity: 1.0781 |
1124
+ [2026-09-17 03:27:57] iteration 5/ 10 | consumed samples: 80 | elapsed time per iteration (ms): 323.9 | throughput per GPU (TFLOP/s/GPU): 387.7 | learning rate: 1.000000E-04 | global batch size: 16 | lm loss: 1.040733E+01 | loss scale: 1.0 | grad norm: 25.625 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 8832 | min safe capacity: 1.0781 |
1125
+ [2026-09-17 03:27:57] iteration 6/ 10 | consumed samples: 96 | elapsed time per iteration (ms): 333.2 | throughput per GPU (TFLOP/s/GPU): 376.9 | learning rate: 7.750000E-05 | global batch size: 16 | lm loss: 1.025381E+01 | loss scale: 1.0 | grad norm: 19.172 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 8832 | min safe capacity: 1.0781 |
1126
+ [2026-09-17 03:28:00] iteration 7/ 10 | consumed samples: 112 | elapsed time per iteration (ms): 3161.7 | throughput per GPU (TFLOP/s/GPU): 39.7 | learning rate: 3.250000E-05 | global batch size: 16 | lm loss: 1.002502E+01 | loss scale: 1.0 | grad norm: 8.664 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 8832 | min safe capacity: 1.0781 |
1127
+ [2026-09-17 03:28:01] iteration 8/ 10 | consumed samples: 128 | elapsed time per iteration (ms): 342.7 | throughput per GPU (TFLOP/s/GPU): 366.4 | learning rate: 1.000000E-05 | global batch size: 16 | lm loss: 9.464363E+00 | loss scale: 1.0 | grad norm: 8.108 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 8832 | min safe capacity: 1.0781 |
1128
+ Capture range started in the application.
1129
+ Capture range ended in the application.
1130
+ Generating '/tmp/nsys-report-2929.qdstrm'
1131
+
1132
+
1133
+ Collection stop is not allowed in this state.
1134
+ W0917 03:29:32.760000 948326 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers
1135
+ W0917 03:29:32.761000 948326 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 948649 closing signal SIGTERM
1136
+ Generated:
1137
+ /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep
1138
+ W0917 03:29:32.765000 948326 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 948650 closing signal SIGTERM
1139
+ W0917 03:29:32.772000 948326 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 948651 closing signal SIGTERM
1140
+ W0917 03:29:32.773000 948326 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 948652 closing signal SIGTERM
1141
+ W0917 03:29:32.777000 948326 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 948653 closing signal SIGTERM
1142
+ W0917 03:29:32.783000 948326 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 948654 closing signal SIGTERM
1143
+ W0917 03:29:32.794000 948326 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 948655 closing signal SIGTERM
1144
+ W0917 03:29:32.804000 948326 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 948656 closing signal SIGTERM
1145
+ Traceback (most recent call last):
1146
+ File "/usr/local/bin/torchrun", line 8, in <module>
1147
+ sys.exit(main())
1148
+ ^^^^^^
1149
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
1150
+ return f(*args, **kwargs)
1151
+ ^^^^^^^^^^^^^^^^^^
1152
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 892, in main
1153
+ run(args)
1154
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 883, in run
1155
+ elastic_launch(
1156
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
1157
+ return launch_agent(self._config, self._entrypoint, list(args))
1158
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1159
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 261, in launch_agent
1160
+ result = agent.run()
1161
+ ^^^^^^^^^^^
1162
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper
1163
+ result = f(*args, **kwargs)
1164
+ ^^^^^^^^^^^^^^^^^^
1165
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 711, in run
1166
+ result = self._invoke_run(role)
1167
+ ^^^^^^^^^^^^^^^^^^^^^^
1168
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 870, in _invoke_run
1169
+ time.sleep(monitor_interval)
1170
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 84, in _terminate_process_handler
1171
+ raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval)
1172
+ torch.distributed.elastic.multiprocessing.api.SignalException: Process 948326 got signal: 15
1173
+ The target application terminated. One or more process it created re-parented.
1174
+ Waiting for termination of re-parented processes.
1175
+ Use the `--wait` option to modify this behavior.
1176
+ There are no active sessions.
granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/verify_trace_ranks.log ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ report : /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep
2
+ kernels : 143382
3
+ devices : 8 of 8 -> [0, 1, 2, 3, 4, 5, 6, 7]
4
+ processes: 1
5
+ markers : 26624 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/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b72936061abb85f8abaf8716a77beb00af3bf28399bb4c5824e9a1ef123b6169
3
+ size 12226799
granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/run.env ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=ultraep
2
+ config=G3_K2
3
+ family=tierA
4
+ source_model=
5
+ standard_num_layers=
6
+ base_rung=G3
7
+ repo=/workspace/Megatron-UltraEP
8
+ branch=ultraep-bracing-regions-through-kernels
9
+ commit=03d57ef7
10
+ num_layers=16
11
+ hidden=4096
12
+ ffn_hidden=14336
13
+ moe_ffn_hidden=1792
14
+ num_experts=64
15
+ topk=2
16
+ experts_per_rank=8
17
+ expert_params=22020096
18
+ heads=32
19
+ kv_groups=8
20
+ kv_channels=
21
+ qk_layernorm=0
22
+ shared_ffn=0
23
+ shared_expert_overlap=0
24
+ moe_token_dispatcher=flex
25
+ mla=0
26
+ q_lora_rank=
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=0
32
+ moe_layers=16
33
+ rotary_base=1000000
34
+ rotary_percent=1.0
35
+ rotary_scaling=
36
+ mscale=
37
+ mscale_all_dim=
38
+ qkv_bias=0
39
+ router_balancing=aux_loss
40
+ aux_loss_coeff=1e-2
41
+ router_pre_softmax=0
42
+ router_fusion=0
43
+ router_dtype=fp32
44
+ router_groups=
45
+ router_group_topk=
46
+ router_score_function=softmax
47
+ router_expert_bias=0
48
+ router_bias_update_rate=
49
+ router_topk_scaling_factor=
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=32768
64
+ init_method_std=0.01
65
+ device_initiated_grouped_gemm=1
66
+ norm_epsilon=1e-5
67
+ seq_length=4096
68
+ micro_batch_size=1
69
+ global_batch_size=16
70
+ num_microbatches=2
71
+ assignments_per_rank=8192
72
+ mean_tokens_per_expert=1024
73
+ slot_policy=matched
74
+ slots_per_rank=8
75
+ capacity_mode=static
76
+ capacity=1.1875
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/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/train.log ADDED
The diff for this file is too large to render. See raw diff
 
granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/verify_trace_ranks.log ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ report : /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep
2
+ kernels : 157846
3
+ devices : 8 of 8 -> [0, 1, 2, 3, 4, 5, 6, 7]
4
+ processes: 1
5
+ markers : 26624 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/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375.nsys-rep ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ccb36de5c1df5199ea5ef7350c20642dfcc3518f1cf7ce829c2b7a31d96b14b2
3
+ size 12191295
granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/run.env ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=ultraep
2
+ config=G3_K8
3
+ family=tierA
4
+ source_model=
5
+ standard_num_layers=
6
+ base_rung=G3
7
+ repo=/workspace/Megatron-UltraEP
8
+ branch=ultraep-bracing-regions-through-kernels
9
+ commit=03d57ef7
10
+ num_layers=16
11
+ hidden=4096
12
+ ffn_hidden=14336
13
+ moe_ffn_hidden=1792
14
+ num_experts=64
15
+ topk=8
16
+ experts_per_rank=8
17
+ expert_params=22020096
18
+ heads=32
19
+ kv_groups=8
20
+ kv_channels=
21
+ qk_layernorm=0
22
+ shared_ffn=0
23
+ shared_expert_overlap=0
24
+ moe_token_dispatcher=flex
25
+ mla=0
26
+ q_lora_rank=
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=0
32
+ moe_layers=16
33
+ rotary_base=1000000
34
+ rotary_percent=1.0
35
+ rotary_scaling=
36
+ mscale=
37
+ mscale_all_dim=
38
+ qkv_bias=0
39
+ router_balancing=aux_loss
40
+ aux_loss_coeff=1e-2
41
+ router_pre_softmax=0
42
+ router_fusion=0
43
+ router_dtype=fp32
44
+ router_groups=
45
+ router_group_topk=
46
+ router_score_function=softmax
47
+ router_expert_bias=0
48
+ router_bias_update_rate=
49
+ router_topk_scaling_factor=
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=32768
64
+ init_method_std=0.01
65
+ device_initiated_grouped_gemm=1
66
+ norm_epsilon=1e-5
67
+ seq_length=4096
68
+ micro_batch_size=1
69
+ global_batch_size=16
70
+ num_microbatches=2
71
+ assignments_per_rank=32768
72
+ mean_tokens_per_expert=4096
73
+ slot_policy=matched
74
+ slots_per_rank=8
75
+ capacity_mode=static
76
+ capacity=1.09375
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=512
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/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/train.log ADDED
@@ -0,0 +1,1268 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ==============================================================
2
+ G3_K8 / ultraep (slots/rank=8 policy=matched)
3
+ E=64 K=8 H'=1792 m=8
4
+ L=16 seq=4096 gbs=16 (2 microbatches)
5
+ cap_mode=static cap=1.09375 cg=1 recompute=none
6
+ memory_dump=none snapshots=/workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K8__ultraep__matched__seq4096__gbs16__cap1p09375/nsys/memory_snapshots/snapshot.pickle
7
+ repo=/workspace/Megatron-UltraEP @ 03d57ef7
8
+ ==============================================================
9
+ W0917 03:33:05.052000 969559 torch/distributed/run.py:766]
10
+ W0917 03:33:05.052000 969559 torch/distributed/run.py:766] *****************************************
11
+ W0917 03:33:05.052000 969559 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
12
+ W0917 03:33:05.052000 969559 torch/distributed/run.py:766] *****************************************
13
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
14
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
15
+ 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
16
+ WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer
17
+ Number of virtual stages per pipeline stage: None
18
+ accumulate and all-reduce gradients in fp32 for bfloat16 data type.
19
+ using torch.bfloat16 for parameters ...
20
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
21
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
22
+ ------------------------ arguments ------------------------
23
+ account_for_embedding_in_pipeline_split ......... False
24
+ account_for_loss_in_pipeline_split .............. False
25
+ accumulate_allreduce_grads_in_fp32 .............. True
26
+ adam_beta1 ...................................... 0.9
27
+ adam_beta2 ...................................... 0.999
28
+ adam_eps ........................................ 1e-08
29
+ add_bias_linear ................................. False
30
+ add_position_embedding .......................... False
31
+ add_qkv_bias .................................... False
32
+ adlr_autoresume ................................. False
33
+ adlr_autoresume_interval ........................ 1000
34
+ align_grad_reduce ............................... True
35
+ align_param_gather .............................. False
36
+ app_tag_run_name ................................ None
37
+ app_tag_run_version ............................. 0.0.0
38
+ apply_layernorm_1p .............................. False
39
+ apply_query_key_layer_scaling ................... False
40
+ apply_residual_connection_post_layernorm ........ False
41
+ apply_rope_fusion ............................... True
42
+ async_save ...................................... None
43
+ async_tensor_model_parallel_allreduce ........... True
44
+ attention_backend ............................... AttnBackend.auto
45
+ attention_dropout ............................... 0.0
46
+ attention_softmax_in_fp32 ....................... False
47
+ auto_detect_ckpt_format ......................... False
48
+ barrier_with_L1_time ............................ True
49
+ bert_binary_head ................................ True
50
+ bert_embedder_type .............................. megatron
51
+ bert_load ....................................... None
52
+ bf16 ............................................ True
53
+ bias_dropout_fusion ............................. True
54
+ bias_gelu_fusion ................................ False
55
+ bias_swiglu_fusion .............................. True
56
+ biencoder_projection_dim ........................ 0
57
+ biencoder_shared_query_context_model ............ False
58
+ block_data_path ................................. None
59
+ calc_ft_timeouts ................................ False
60
+ calculate_per_token_loss ........................ False
61
+ check_for_large_grads ........................... False
62
+ check_for_nan_in_loss_and_grad .................. False
63
+ check_for_spiky_loss ............................ False
64
+ check_weight_hash_across_dp_replicas_interval ... None
65
+ ckpt_assume_constant_structure .................. False
66
+ ckpt_convert_format ............................. None
67
+ ckpt_convert_save ............................... None
68
+ ckpt_convert_update_legacy_dist_opt_format ...... False
69
+ ckpt_format ..................................... torch_dist
70
+ ckpt_fully_parallel_load ........................ False
71
+ ckpt_fully_parallel_save ........................ True
72
+ ckpt_fully_parallel_save_deprecated ............. False
73
+ ckpt_step ....................................... None
74
+ classes_fraction ................................ 1.0
75
+ clip_grad ....................................... 1.0
76
+ clone_scatter_output_in_embedding ............... True
77
+ config_logger_dir ...............................
78
+ consumed_train_samples .......................... 0
79
+ consumed_valid_samples .......................... 0
80
+ context_parallel_size ........................... 1
81
+ cp_comm_type .................................... ['p2p']
82
+ create_attention_mask_in_dataloader ............. True
83
+ cross_entropy_fusion_impl ....................... native
84
+ cross_entropy_loss_fusion ....................... False
85
+ cuda_graph_scope ................................ full_iteration
86
+ cuda_graph_warmup_steps ......................... 3
87
+ data_args_path .................................. None
88
+ data_cache_path ................................. None
89
+ data_parallel_random_init ....................... False
90
+ data_parallel_sharding_strategy ................. no_shard
91
+ data_parallel_size .............................. 8
92
+ data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document']
93
+ data_per_class_fraction ......................... 1.0
94
+ data_sharding ................................... True
95
+ dataloader_type ................................. single
96
+ ddp_average_in_collective ....................... False
97
+ ddp_bucket_size ................................. None
98
+ ddp_num_buckets ................................. None
99
+ ddp_pad_buckets_for_high_nccl_busbw ............. False
100
+ decoder_first_pipeline_num_layers ............... None
101
+ decoder_last_pipeline_num_layers ................ None
102
+ decoder_num_layers .............................. None
103
+ decoder_seq_length .............................. None
104
+ decoupled_lr .................................... None
105
+ decoupled_min_lr ................................ None
106
+ decrease_batch_size_if_needed ................... False
107
+ defer_embedding_wgrad_compute ................... False
108
+ delay_wgrad_compute ............................. False
109
+ deprecated_use_mcore_models ..................... True
110
+ deterministic_mode .............................. False
111
+ dino_bottleneck_size ............................ 256
112
+ dino_freeze_last_layer .......................... 1
113
+ dino_head_hidden_size ........................... 2048
114
+ dino_local_crops_number ......................... 10
115
+ dino_local_img_size ............................. 96
116
+ dino_norm_last_layer ............................ False
117
+ dino_teacher_temp ............................... 0.07
118
+ dino_warmup_teacher_temp ........................ 0.04
119
+ dino_warmup_teacher_temp_epochs ................. 30
120
+ disable_bf16_reduced_precision_matmul ........... False
121
+ disable_mamba_mem_eff_path ...................... False
122
+ disable_straggler_on_startup .................... False
123
+ dist_ckpt_format_deprecated ..................... None
124
+ dist_ckpt_strictness ............................ assume_ok_unexpected
125
+ distribute_saved_activations .................... False
126
+ distributed_backend ............................. nccl
127
+ distributed_timeout_minutes ..................... 10
128
+ embedding_init_method_std ....................... None
129
+ embedding_path .................................. None
130
+ empty_unused_memory_level ....................... 0
131
+ enable_cuda_graph ............................... True
132
+ enable_experimental ............................. False
133
+ enable_ft_package ............................... False
134
+ enable_gloo_process_groups ...................... True
135
+ enable_msc ...................................... True
136
+ enable_one_logger ............................... True
137
+ encoder_num_layers .............................. 16
138
+ encoder_seq_length .............................. 4096
139
+ end_weight_decay ................................ 0.1
140
+ eod_mask_loss ................................... False
141
+ error_injection_rate ............................ 0
142
+ error_injection_type ............................ transient_error
143
+ eval_interval ................................... 1000000
144
+ eval_iters ...................................... 0
145
+ evidence_data_path .............................. None
146
+ exit_duration_in_mins ........................... None
147
+ exit_interval ................................... None
148
+ exit_on_missing_checkpoint ...................... False
149
+ exit_signal_handler ............................. False
150
+ exp_avg_dtype ................................... torch.float32
151
+ exp_avg_sq_dtype ................................ torch.float32
152
+ expert_model_parallel_size ...................... 8
153
+ expert_tensor_parallel_size ..................... 1
154
+ export_force_local_attention .................... False
155
+ export_kd_cfg ................................... None
156
+ export_kd_teacher_ckpt_format ................... None
157
+ export_kd_teacher_load .......................... None
158
+ export_kv_cache_quant ........................... False
159
+ export_legacy_megatron .......................... False
160
+ export_model_type ............................... GPTModel
161
+ export_moe_apply_probs_on_input ................. False
162
+ export_qk_l2_norm ............................... False
163
+ export_quant_cfg ................................ None
164
+ export_real_quant_cfg ........................... None
165
+ export_te_mcore_model ........................... False
166
+ external_cuda_graph ............................. False
167
+ ffn_hidden_size ................................. 14336
168
+ finetune ........................................ False
169
+ finetune_data_split ............................. train
170
+ finetune_hf_dataset ............................. None
171
+ first_last_layers_bf16 .......................... False
172
+ flash_decode .................................... False
173
+ fp16 ............................................ False
174
+ fp16_lm_cross_entropy ........................... False
175
+ fp32_residual_connection ........................ False
176
+ fp8 ............................................. e4m3
177
+ fp8_amax_compute_algo ........................... most_recent
178
+ fp8_amax_history_len ............................ 1
179
+ fp8_interval .................................... 1
180
+ fp8_margin ...................................... 0
181
+ fp8_param_gather ................................ False
182
+ fp8_recipe ...................................... mxfp8
183
+ fp8_wgrad ....................................... True
184
+ fsdp_double_buffer .............................. False
185
+ global_batch_size ............................... 16
186
+ grad_reduce_in_bf16 ............................. False
187
+ gradient_accumulation_fusion .................... True
188
+ gradient_reduce_div_fusion ...................... True
189
+ group_query_attention ........................... True
190
+ head_lr_mult .................................... 1.0
191
+ heterogeneous_layers_config_encoded_json ........ None
192
+ heterogeneous_layers_config_path ................ None
193
+ hidden_dropout .................................. 0.0
194
+ hidden_size ..................................... 4096
195
+ hierarchical_context_parallel_sizes ............. None
196
+ high_priority_stream_groups ..................... []
197
+ hybrid_attention_ratio .......................... 0.0
198
+ hybrid_mlp_ratio ................................ 0.0
199
+ hybrid_override_pattern ......................... None
200
+ hysteresis ...................................... 2
201
+ ict_head_size ................................... None
202
+ ict_load ........................................ None
203
+ img_h ........................................... 224
204
+ img_w ........................................... 224
205
+ indexer_batch_size .............................. 128
206
+ indexer_log_interval ............................ 1000
207
+ inference_batch_times_seqlen_threshold .......... -1
208
+ inference_dynamic_batching ...................... False
209
+ inference_dynamic_batching_buffer_guaranteed_fraction 0.2
210
+ inference_dynamic_batching_buffer_overflow_factor None
211
+ inference_dynamic_batching_buffer_size_gb ....... 40.0
212
+ inference_dynamic_batching_chunk_size ........... 256
213
+ inference_dynamic_batching_max_requests_override None
214
+ inference_dynamic_batching_max_tokens_override .. None
215
+ inference_dynamic_batching_num_cuda_graphs ...... 16
216
+ inference_max_batch_size ........................ 8
217
+ inference_max_seq_length ........................ 2560
218
+ inference_rng_tracker ........................... False
219
+ init_method_std ................................. 0.01
220
+ init_method_xavier_uniform ...................... False
221
+ init_model_with_meta_device ..................... False
222
+ initial_loss_scale .............................. 4294967296
223
+ inprocess_active_world_size ..................... 8
224
+ inprocess_barrier_timeout ....................... 120
225
+ inprocess_completion_timeout .................... 120
226
+ inprocess_empty_cuda_cache ...................... False
227
+ inprocess_granularity ........................... node
228
+ inprocess_hard_timeout .......................... 90
229
+ inprocess_heartbeat_interval .................... 30
230
+ inprocess_heartbeat_timeout ..................... 60
231
+ inprocess_last_call_wait ........................ 1
232
+ inprocess_max_iterations ........................ None
233
+ inprocess_monitor_process_interval .............. 1.0
234
+ inprocess_monitor_thread_interval ............... 1.0
235
+ inprocess_progress_watchdog_interval ............ 1.0
236
+ inprocess_restart ............................... False
237
+ inprocess_soft_timeout .......................... 60
238
+ inprocess_termination_grace_time ................ 1
239
+ is_hybrid_model ................................. False
240
+ iter_per_epoch .................................. 1250
241
+ iterations_to_skip .............................. []
242
+ keep_fp8_transpose_cache_when_using_custom_fsdp . False
243
+ kitchen_config_file ............................. None
244
+ kitchen_recipe_number ........................... None
245
+ kv_channels ..................................... 128
246
+ kv_lora_rank .................................... 32
247
+ lazy_mpu_init ................................... None
248
+ load ............................................ None
249
+ load_main_params_from_ckpt ...................... None
250
+ load_model_opt_format ........................... False
251
+ local_rank ...................................... 0
252
+ log_energy ...................................... False
253
+ log_interval .................................... 1
254
+ log_loss_scale_to_tensorboard ................... True
255
+ log_memory_to_tensorboard ....................... False
256
+ log_num_zeros_in_grad ........................... False
257
+ log_params_norm ................................. False
258
+ log_progress .................................... False
259
+ log_straggler ................................... False
260
+ log_throughput .................................. True
261
+ log_timers_to_tensorboard ....................... False
262
+ log_validation_ppl_to_tensorboard ............... False
263
+ log_world_size_to_tensorboard ................... False
264
+ logging_level ................................... None
265
+ loss_scale ...................................... None
266
+ loss_scale_window ............................... 1000
267
+ lr .............................................. 0.0001
268
+ lr_decay_iters .................................. 8
269
+ lr_decay_samples ................................ None
270
+ lr_decay_style .................................. cosine
271
+ lr_warmup_fraction .............................. None
272
+ lr_warmup_init .................................. 0.0
273
+ lr_warmup_iters ................................. 5
274
+ lr_warmup_samples ............................... 0
275
+ lr_wsd_decay_iters .............................. None
276
+ lr_wsd_decay_samples ............................ None
277
+ lr_wsd_decay_style .............................. exponential
278
+ main_grads_dtype ................................ torch.float32
279
+ main_params_dtype ............................... torch.float32
280
+ make_vocab_size_divisible_by .................... 128
281
+ mamba_head_dim .................................. 64
282
+ mamba_num_groups ................................ 8
283
+ mamba_num_heads ................................. None
284
+ mamba_state_dim ................................. 128
285
+ manual_gc ....................................... False
286
+ manual_gc_eval .................................. True
287
+ manual_gc_interval .............................. 0
288
+ mask_factor ..................................... 1.0
289
+ mask_prob ....................................... 0.15
290
+ mask_type ....................................... random
291
+ masked_softmax_fusion ........................... False
292
+ max_position_embeddings ......................... 32768
293
+ max_tokens_to_oom ............................... 12000
294
+ memory_snapshot_path ............................ snapshot.pickle
295
+ merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt
296
+ micro_batch_size ................................ 1
297
+ microbatch_group_size_per_vp_stage .............. None
298
+ mid_level_dataset_surplus ....................... 0.005
299
+ min_loss_scale .................................. 1.0
300
+ min_lr .......................................... 1e-05
301
+ mlp_chunks_for_prefill .......................... 1
302
+ mmap_bin_files .................................. True
303
+ mock_data ....................................... False
304
+ moe_apply_probs_on_input ........................ False
305
+ moe_aux_loss_coeff .............................. 0.01
306
+ moe_capacity_monitor ............................ True
307
+ moe_deepep_num_sms .............................. 20
308
+ moe_echo_enable_random_offloading ............... False
309
+ moe_echo_expert_dispatch_overlap ................ False
310
+ moe_echo_expert_dispatcher_type ................. hybridep
311
+ moe_echo_recompute_expert_dispatch .............. False
312
+ moe_enable_deepep ............................... False
313
+ moe_enable_echo ................................. False
314
+ moe_enable_hybridep ............................. True
315
+ moe_enable_ultraep .............................. True
316
+ moe_expert_capacity_factor ...................... None
317
+ moe_extended_tp ................................. False
318
+ moe_ffn_hidden_size ............................. 1792
319
+ moe_grouped_gemm ................................ True
320
+ moe_hybridep_num_sms ............................ 16
321
+ moe_input_jitter_eps ............................ None
322
+ moe_layer_freq .................................. 1
323
+ moe_layer_recompute ............................. False
324
+ moe_num_echo_experts ............................ None
325
+ moe_num_redundant_experts_per_rank .............. 8
326
+ moe_pad_expert_input_to_capacity ................ False
327
+ moe_per_layer_logging ........................... False
328
+ moe_permute_fusion .............................. False
329
+ moe_received_token_capacity ..................... 1.09375
330
+ moe_router_bias_update_rate ..................... 0.001
331
+ moe_router_dtype ................................ fp32
332
+ moe_router_enable_expert_bias ................... False
333
+ moe_router_force_load_balancing ................. False
334
+ moe_router_group_topk ........................... None
335
+ moe_router_load_balancing_type .................. aux_loss
336
+ moe_router_num_groups ........................... None
337
+ moe_router_padding_for_fp8 ...................... False
338
+ moe_router_pre_softmax .......................... False
339
+ moe_router_score_function ....................... softmax
340
+ moe_router_topk ................................. 8
341
+ moe_router_topk_scaling_factor .................. None
342
+ moe_shared_expert_intermediate_size ............. None
343
+ moe_shared_expert_overlap ....................... False
344
+ moe_token_dispatcher_type ....................... flex
345
+ moe_token_drop_policy ........................... probs
346
+ moe_ultraep_max_microbatches .................... None
347
+ moe_ultraep_overlap_weight_sync ................. True
348
+ moe_ultraep_weight_sync_join .................... before_dispatch
349
+ moe_upcycling_granularity ....................... 1
350
+ moe_use_device_initiated_grouped_gemm ........... True
351
+ moe_use_legacy_grouped_gemm ..................... False
352
+ moe_use_upcycling ............................... False
353
+ moe_z_loss_coeff ................................ None
354
+ mrope_section ................................... None
355
+ mscale .......................................... 1.0
356
+ mscale_all_dim .................................. 0.0
357
+ mtp_loss_scaling_factor ......................... 0.1
358
+ mtp_num_layers .................................. None
359
+ multi_latent_attention .......................... False
360
+ nccl_all_reduce_for_prefill ..................... False
361
+ nccl_communicator_config_path ................... None
362
+ nccl_ub ......................................... False
363
+ no_load_optim ................................... None
364
+ no_load_rng ..................................... None
365
+ no_persist_layer_norm ........................... False
366
+ no_rope_freq .................................... None
367
+ no_save_optim ................................... None
368
+ no_save_rng ..................................... None
369
+ non_persistent_ckpt_type ........................ None
370
+ non_persistent_global_ckpt_dir .................. None
371
+ non_persistent_local_ckpt_algo .................. fully_parallel
372
+ non_persistent_local_ckpt_dir ................... None
373
+ non_persistent_save_interval .................... None
374
+ norm_epsilon .................................... 1e-05
375
+ normalization ................................... RMSNorm
376
+ num_attention_heads ............................. 32
377
+ num_channels .................................... 3
378
+ num_classes ..................................... 1000
379
+ num_dataset_builder_threads ..................... 1
380
+ num_distributed_optimizer_instances ............. 1
381
+ num_experts ..................................... 64
382
+ num_layers ...................................... 16
383
+ num_layers_at_end_in_bf16 ....................... 1
384
+ num_layers_at_start_in_bf16 ..................... 1
385
+ num_layers_per_virtual_pipeline_stage ........... None
386
+ num_query_groups ................................ 8
387
+ num_virtual_stages_per_pipeline_rank ............ None
388
+ num_workers ..................................... 2
389
+ object_storage_cache_path ....................... None
390
+ one_logger_async ................................ False
391
+ one_logger_project .............................. megatron-lm
392
+ one_logger_run_name ............................. None
393
+ onnx_safe ....................................... None
394
+ openai_gelu ..................................... False
395
+ optimizer ....................................... adam
396
+ optimizer_cpu_offload ........................... False
397
+ optimizer_offload_fraction ...................... 1.0
398
+ output_bert_embeddings .......................... False
399
+ overlap_cpu_optimizer_d2h_h2d ................... False
400
+ overlap_grad_reduce ............................. True
401
+ overlap_moe_expert_parallel_comm ................ False
402
+ overlap_p2p_comm ................................ False
403
+ overlap_p2p_comm_warmup_flush ................... False
404
+ overlap_param_gather ............................ True
405
+ overlap_param_gather_with_optimizer_step ........ False
406
+ override_opt_param_scheduler .................... False
407
+ params_dtype .................................... torch.bfloat16
408
+ patch_dim ....................................... 16
409
+ per_split_data_args_path ........................ None
410
+ perform_initialization .......................... True
411
+ pin_cpu_grads ................................... True
412
+ pin_cpu_params .................................. True
413
+ pipeline_model_parallel_comm_backend ............ None
414
+ pipeline_model_parallel_layout .................. None
415
+ pipeline_model_parallel_size .................... 1
416
+ position_embedding_type ......................... rope
417
+ pretrained_checkpoint ........................... None
418
+ profile ......................................... True
419
+ profile_ranks ................................... [0, 1, 2, 3, 4, 5, 6, 7]
420
+ profile_step_end ................................ 8
421
+ profile_step_start .............................. 6
422
+ q_lora_rank ..................................... None
423
+ qk_head_dim ..................................... 128
424
+ qk_l2_norm ...................................... False
425
+ qk_layernorm .................................... False
426
+ qk_pos_emb_head_dim ............................. 64
427
+ query_in_block_prob ............................. 0.1
428
+ rampup_batch_size ............................... None
429
+ rank ............................................ 0
430
+ recompute_granularity ........................... None
431
+ recompute_method ................................ None
432
+ recompute_modules ............................... None
433
+ recompute_num_layers ............................ None
434
+ record_memory_history ........................... False
435
+ relative_attention_max_distance ................. 128
436
+ relative_attention_num_buckets .................. 32
437
+ replication ..................................... False
438
+ replication_factor .............................. 2
439
+ replication_jump ................................ None
440
+ rerun_mode ...................................... validate_results
441
+ reset_attention_mask ............................ False
442
+ reset_position_ids .............................. False
443
+ result_rejected_tracker_filename ................ None
444
+ retriever_report_topk_accuracies ................ []
445
+ retriever_score_scaling ......................... False
446
+ retriever_seq_length ............................ 256
447
+ retro_add_retriever ............................. False
448
+ retro_attention_gate ............................ 1
449
+ retro_cyclic_train_iters ........................ None
450
+ retro_encoder_attention_dropout ................. 0.1
451
+ retro_encoder_hidden_dropout .................... 0.1
452
+ retro_encoder_layers ............................ 2
453
+ retro_num_neighbors ............................. 2
454
+ retro_num_retrieved_chunks ...................... 2
455
+ retro_project_dir ............................... None
456
+ retro_verify_neighbor_count ..................... True
457
+ reuse_grad_buf_for_mxfp8_param_ag ............... False
458
+ rope_scaling_factor ............................. 8.0
459
+ rope_type ....................................... None
460
+ rotary_base ..................................... 1000000
461
+ rotary_interleaved .............................. False
462
+ rotary_percent .................................. 1.0
463
+ rotary_scaling_factor ........................... 1.0
464
+ rotary_seq_len_interpolation_factor ............. None
465
+ run_workload_inspector_server ................... False
466
+ sample_rate ..................................... 1.0
467
+ save ............................................ None
468
+ save_interval ................................... None
469
+ save_retain_interval ............................ None
470
+ scatter_gather_tensors_in_pipeline .............. True
471
+ seed ............................................ 1234
472
+ seq_length ...................................... 4096
473
+ sequence_parallel ............................... False
474
+ sft ............................................. False
475
+ sft_tokenizer_prompt_format ..................... nemotron-h-aligned
476
+ sgd_momentum .................................... 0.9
477
+ short_seq_prob .................................. 0.1
478
+ skip_train ...................................... False
479
+ skipped_train_samples ........................... 0
480
+ spec ............................................ None
481
+ split ........................................... 100,0,0
482
+ squared_relu .................................... False
483
+ start_weight_decay .............................. 0.1
484
+ straggler_ctrlr_port ............................ 65535
485
+ straggler_minmax_count .......................... 1
486
+ suggested_communication_unit_size ............... None
487
+ swiglu .......................................... True
488
+ swin_backbone_type .............................. tiny
489
+ symmetric_ar_type ............................... None
490
+ te_rng_tracker .................................. True
491
+ tensor_model_parallel_size ...................... 1
492
+ tensorboard_dir ................................. None
493
+ tensorboard_log_interval ........................ 1
494
+ tensorboard_queue_size .......................... 1000
495
+ test_data_path .................................. None
496
+ test_mode ....................................... False
497
+ tiktoken_num_special_tokens ..................... 1000
498
+ tiktoken_pattern ................................ None
499
+ tiktoken_special_tokens ......................... None
500
+ timing_log_level ................................ 0
501
+ timing_log_option ............................... minmax
502
+ titles_data_path ................................ None
503
+ tokenizer_model ................................. None
504
+ tokenizer_type .................................. GPT2BPETokenizer
505
+ torch_fsdp2_reshard_after_forward ............... True
506
+ tp_comm_bootstrap_backend ....................... nccl
507
+ tp_comm_bulk_dgrad .............................. True
508
+ tp_comm_bulk_wgrad .............................. True
509
+ tp_comm_overlap ................................. False
510
+ tp_comm_overlap_ag .............................. True
511
+ tp_comm_overlap_cfg ............................. None
512
+ tp_comm_overlap_rs .............................. True
513
+ tp_comm_overlap_rs_dgrad ........................ False
514
+ tp_comm_split_ag ................................ True
515
+ tp_comm_split_rs ................................ True
516
+ train_data_path ................................. None
517
+ train_iters ..................................... 10
518
+ train_samples ................................... None
519
+ train_sync_interval ............................. None
520
+ transformer_impl ................................ transformer_engine
521
+ transformer_pipeline_model_parallel_size ........ 1
522
+ untie_embeddings_and_output_weights ............. True
523
+ use_checkpoint_args ............................. False
524
+ use_checkpoint_opt_param_scheduler .............. False
525
+ use_cpu_initialization .......................... None
526
+ use_custom_fsdp ................................. False
527
+ use_dist_ckpt ................................... True
528
+ use_dist_ckpt_deprecated ........................ False
529
+ use_distributed_optimizer ....................... True
530
+ use_flash_attn .................................. False
531
+ use_legacy_models ............................... False
532
+ use_mp_args_from_checkpoint_args ................ False
533
+ use_one_sent_docs ............................... False
534
+ use_persistent_ckpt_worker ...................... False
535
+ use_precision_aware_optimizer ................... False
536
+ use_pytorch_profiler ............................ False
537
+ use_ring_exchange_p2p ........................... False
538
+ use_rope_scaling ................................ False
539
+ use_rotary_position_embeddings .................. False
540
+ use_sharp ....................................... False
541
+ use_tokenizer_model_from_checkpoint_args ........ True
542
+ use_torch_fsdp2 ................................. False
543
+ use_torch_optimizer_for_cpu_offload ............. False
544
+ use_tp_pp_dp_mapping ............................ False
545
+ v_head_dim ...................................... 128
546
+ valid_data_path ................................. None
547
+ variable_seq_lengths ............................ False
548
+ virtual_pipeline_model_parallel_size ............ None
549
+ vision_backbone_type ............................ vit
550
+ vision_pretraining .............................. False
551
+ vision_pretraining_type ......................... classify
552
+ vocab_extra_ids ................................. 0
553
+ vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json
554
+ vocab_size ...................................... None
555
+ wandb_exp_name ..................................
556
+ wandb_project ...................................
557
+ wandb_save_dir ..................................
558
+ weight_decay .................................... 0.1
559
+ weight_decay_incr_style ......................... constant
560
+ wgrad_deferral_limit ............................ 0
561
+ world_size ...................................... 8
562
+ yaml_cfg ........................................ None
563
+ -------------------- end of arguments ---------------------
564
+ INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 2
565
+ > building GPT2BPETokenizer tokenizer ...
566
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
567
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
568
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
569
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
570
+ /workspace/Megatron-UltraEP/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
571
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
572
+ > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304)
573
+ WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS
574
+ > initializing torch distributed ...
575
+ 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
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
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
583
+ [Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
584
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
585
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
586
+ [Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
587
+ [Gloo] Rank 4 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 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
590
+ [Gloo] Rank 0 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 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
593
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
594
+ [Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7[Gloo] Rank
595
+ 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
596
+ [Gloo] Rank [Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 67 is connected to
597
+ 7 peer ranks. Expected number of connected peer ranks is : 7
598
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
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 : 0
603
+ [Gloo] Rank 0 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
+ > initialized tensor model parallel with size 1
607
+ > initialized pipeline model parallel with size 1
608
+ > setting random seeds to 1234 ...
609
+ [rank5]:[W917 03:33:17.749694598 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 5] using GPU 5 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
610
+ [rank1]:[W917 03:33:17.749736865 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
611
+ [rank3]:[W917 03:33:17.749771189 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
612
+ [rank2]:[W917 03:33:17.749792831 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 2] using GPU 2 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
613
+ [rank7]:[W917 03:33:17.753345276 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
614
+ [rank4]:[W917 03:33:17.754063808 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 4] using GPU 4 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
615
+ [rank6]:[W917 03:33:17.754496723 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 6] using GPU 6 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
616
+ > compiling dataset index builder ...
617
+ make: Entering directory '/workspace/Megatron-UltraEP/megatron/core/datasets'
618
+ make: Nothing to be done for 'default'.
619
+ make: Leaving directory '/workspace/Megatron-UltraEP/megatron/core/datasets'
620
+ >>> done with dataset index builder. Compilation time: 0.429 seconds
621
+ WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations.
622
+ > compiling and loading fused kernels ...
623
+ [rank0]:[W917 03:33:18.359796766 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 0] using GPU 0 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
624
+ >>> done with compiling and loading fused kernels. Compilation time: 3.438 seconds
625
+ time to initialize megatron (seconds): 8.825
626
+ [after megatron is initialized] datetime: 2026-09-17 03:33:23
627
+ building GPT model ...
628
+ /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.
629
+ warnings.warn(
630
+ /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.
631
+ warnings.warn(
632
+ /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.
633
+ warnings.warn(
634
+ /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.
635
+ warnings.warn(
636
+ /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.
637
+ warnings.warn(
638
+ /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.
639
+ warnings.warn(
640
+ /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.
641
+ warnings.warn(
642
+ /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.
643
+ warnings.warn(
644
+ [INFO] Use MNNVL fabric: False
645
+ [WARN] MAX_NUM_NVL_PEERS is not set. Using detected value 8.
646
+ NVSHMEM v3.7.2
647
+ /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.
648
+
649
+ /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.
650
+
651
+ /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.
652
+
653
+ /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.
654
+
655
+ /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.
656
+
657
+ /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.
658
+
659
+ /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.
660
+
661
+ /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.
662
+
663
+ UltraEP: NVL domain size <= 8; forcing weight_sync plan mode to direct (was adaptive).
664
+ > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 6724653056
665
+ 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)
666
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 18
667
+ Params for bucket 1 (206045184 elements, 206045184 padded size):
668
+ module.output_layer.weight
669
+ Params for bucket 2 (42217472 elements, 42217472 padded size):
670
+ module.decoder.final_layernorm.weight
671
+ module.decoder.layers.15.self_attention.linear_qkv.weight
672
+ module.decoder.layers.15.pre_mlp_layernorm.weight
673
+ module.decoder.layers.15.mlp.router.weight
674
+ module.decoder.layers.15.self_attention.linear_qkv.layer_norm_weight
675
+ module.decoder.layers.15.self_attention.linear_proj.weight
676
+ Params for bucket 3 (42213376 elements, 42213376 padded size):
677
+ module.decoder.layers.14.self_attention.linear_qkv.weight
678
+ module.decoder.layers.14.pre_mlp_layernorm.weight
679
+ module.decoder.layers.14.mlp.router.weight
680
+ module.decoder.layers.14.self_attention.linear_qkv.layer_norm_weight
681
+ module.decoder.layers.14.self_attention.linear_proj.weight
682
+ Params for bucket 4 (42213376 elements, 42213376 padded size):
683
+ module.decoder.layers.13.pre_mlp_layernorm.weight
684
+ module.decoder.layers.13.self_attention.linear_qkv.weight
685
+ module.decoder.layers.13.mlp.router.weight
686
+ module.decoder.layers.13.self_attention.linear_qkv.layer_norm_weight
687
+ module.decoder.layers.13.self_attention.linear_proj.weight
688
+ Params for bucket 5 (42213376 elements, 42213376 padded size):
689
+ module.decoder.layers.12.mlp.router.weight
690
+ module.decoder.layers.12.self_attention.linear_proj.weight
691
+ module.decoder.layers.12.self_attention.linear_qkv.layer_norm_weight
692
+ module.decoder.layers.12.pre_mlp_layernorm.weight
693
+ module.decoder.layers.12.self_attention.linear_qkv.weight
694
+ Params for bucket 6 (42213376 elements, 42213376 padded size):
695
+ module.decoder.layers.11.mlp.router.weight
696
+ module.decoder.layers.11.self_attention.linear_proj.weight
697
+ module.decoder.layers.11.self_attention.linear_qkv.layer_norm_weight
698
+ module.decoder.layers.11.self_attention.linear_qkv.weight
699
+ module.decoder.layers.11.pre_mlp_layernorm.weight
700
+ Params for bucket 7 (42213376 elements, 42213376 padded size):
701
+ module.decoder.layers.10.mlp.router.weight
702
+ module.decoder.layers.10.self_attention.linear_proj.weight
703
+ module.decoder.layers.10.self_attention.linear_qkv.layer_norm_weight
704
+ module.decoder.layers.10.pre_mlp_layernorm.weight
705
+ module.decoder.layers.10.self_attention.linear_qkv.weight
706
+ Params for bucket 8 (42213376 elements, 42213376 padded size):
707
+ module.decoder.layers.9.pre_mlp_layernorm.weight
708
+ module.decoder.layers.9.self_attention.linear_qkv.weight
709
+ module.decoder.layers.9.mlp.router.weight
710
+ module.decoder.layers.9.self_attention.linear_proj.weight
711
+ module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight
712
+ Params for bucket 9 (42213376 elements, 42213376 padded size):
713
+ module.decoder.layers.8.self_attention.linear_qkv.weight
714
+ module.decoder.layers.8.pre_mlp_layernorm.weight
715
+ module.decoder.layers.8.mlp.router.weight
716
+ module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight
717
+ module.decoder.layers.8.self_attention.linear_proj.weight
718
+ Params for bucket 10 (42213376 elements, 42213376 padded size):
719
+ module.decoder.layers.7.mlp.router.weight
720
+ module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight
721
+ module.decoder.layers.7.self_attention.linear_proj.weight
722
+ module.decoder.layers.7.self_attention.linear_qkv.weight
723
+ module.decoder.layers.7.pre_mlp_layernorm.weight
724
+ Params for bucket 11 (42213376 elements, 42213376 padded size):
725
+ module.decoder.layers.6.mlp.router.weight
726
+ module.decoder.layers.6.self_attention.linear_proj.weight
727
+ module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight
728
+ module.decoder.layers.6.self_attention.linear_qkv.weight
729
+ module.decoder.layers.6.pre_mlp_layernorm.weight
730
+ Params for bucket 12 (42213376 elements, 42213376 padded size):
731
+ module.decoder.layers.5.self_attention.linear_qkv.weight
732
+ module.decoder.layers.5.self_attention.linear_proj.weight
733
+ module.decoder.layers.5.pre_mlp_layernorm.weight
734
+ module.decoder.layers.5.mlp.router.weight
735
+ module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight
736
+ Params for bucket 13 (42213376 elements, 42213376 padded size):
737
+ module.decoder.layers.4.self_attention.linear_qkv.weight
738
+ module.decoder.layers.4.pre_mlp_layernorm.weight
739
+ module.decoder.layers.4.mlp.router.weight
740
+ module.decoder.layers.4.self_attention.linear_proj.weight
741
+ module.decoder.layers.4.self_attention.linear_qkv.layer_norm_weight
742
+ Params for bucket 14 (42213376 elements, 42213376 padded size):
743
+ module.decoder.layers.3.mlp.router.weight
744
+ module.decoder.layers.3.self_attention.linear_proj.weight
745
+ module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight
746
+ module.decoder.layers.3.pre_mlp_layernorm.weight
747
+ module.decoder.layers.3.self_attention.linear_qkv.weight
748
+ Params for bucket 15 (42213376 elements, 42213376 padded size):
749
+ module.decoder.layers.2.mlp.router.weight
750
+ module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight
751
+ module.decoder.layers.2.self_attention.linear_proj.weight
752
+ module.decoder.layers.2.self_attention.linear_qkv.weight
753
+ module.decoder.layers.2.pre_mlp_layernorm.weight
754
+ Params for bucket 16 (42213376 elements, 42213376 padded size):
755
+ module.decoder.layers.1.mlp.router.weight
756
+ module.decoder.layers.1.self_attention.linear_proj.weight
757
+ module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight
758
+ module.decoder.layers.1.self_attention.linear_qkv.weight
759
+ module.decoder.layers.1.pre_mlp_layernorm.weight
760
+ Params for bucket 17 (42213376 elements, 42213376 padded size):
761
+ module.decoder.layers.0.pre_mlp_layernorm.weight
762
+ module.decoder.layers.0.self_attention.linear_proj.weight
763
+ module.decoder.layers.0.self_attention.linear_qkv.weight
764
+ module.decoder.layers.0.mlp.router.weight
765
+ module.decoder.layers.0.self_attention.linear_qkv.layer_norm_weight
766
+ Params for bucket 18 (206045184 elements, 206045184 padded size):
767
+ module.embedding.word_embeddings.weight
768
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 64
769
+ Params for bucket 1 (44040192 elements, 44040192 padded size):
770
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight6
771
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight4
772
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight2
773
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight7
774
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight5
775
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight3
776
+ Params for bucket 2 (44040192 elements, 44040192 padded size):
777
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight1
778
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight6
779
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight0
780
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight7
781
+ Params for bucket 3 (44040192 elements, 44040192 padded size):
782
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight5
783
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight3
784
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight4
785
+ Params for bucket 4 (44040192 elements, 44040192 padded size):
786
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight2
787
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight0
788
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight1
789
+ Params for bucket 5 (44040192 elements, 44040192 padded size):
790
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight6
791
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight4
792
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight2
793
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight7
794
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight3
795
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight5
796
+ Params for bucket 6 (44040192 elements, 44040192 padded size):
797
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight1
798
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight0
799
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight6
800
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight7
801
+ Params for bucket 7 (44040192 elements, 44040192 padded size):
802
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight5
803
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight4
804
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight3
805
+ Params for bucket 8 (44040192 elements, 44040192 padded size):
806
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight2
807
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight1
808
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight0
809
+ Params for bucket 9 (44040192 elements, 44040192 padded size):
810
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight7
811
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight5
812
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight3
813
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight4
814
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight6
815
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight2
816
+ Params for bucket 10 (44040192 elements, 44040192 padded size):
817
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight1
818
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight0
819
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight6
820
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight7
821
+ Params for bucket 11 (44040192 elements, 44040192 padded size):
822
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight5
823
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight3
824
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight4
825
+ Params for bucket 12 (44040192 elements, 44040192 padded size):
826
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight2
827
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight0
828
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight1
829
+ Params for bucket 13 (44040192 elements, 44040192 padded size):
830
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight6
831
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight4
832
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight2
833
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight7
834
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight5
835
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight3
836
+ Params for bucket 14 (44040192 elements, 44040192 padded size):
837
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight1
838
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight7
839
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight6
840
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight0
841
+ Params for bucket 15 (44040192 elements, 44040192 padded size):
842
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight5
843
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight4
844
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight3
845
+ Params for bucket 16 (44040192 elements, 44040192 padded size):
846
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight2
847
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight1
848
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight0
849
+ Params for bucket 17 (44040192 elements, 44040192 padded size):
850
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight7
851
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight5
852
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight3
853
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight4
854
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight6
855
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight2
856
+ Params for bucket 18 (44040192 elements, 44040192 padded size):
857
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight1
858
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight7
859
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight6
860
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight0
861
+ Params for bucket 19 (44040192 elements, 44040192 padded size):
862
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight5
863
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight4
864
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight3
865
+ Params for bucket 20 (44040192 elements, 44040192 padded size):
866
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight2
867
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight1
868
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight0
869
+ Params for bucket 21 (44040192 elements, 44040192 padded size):
870
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight7
871
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight3
872
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight5
873
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight6
874
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight4
875
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight2
876
+ Params for bucket 22 (44040192 elements, 44040192 padded size):
877
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight1
878
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight0
879
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight7
880
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight6
881
+ Params for bucket 23 (44040192 elements, 44040192 padded size):
882
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight5
883
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight3
884
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight4
885
+ Params for bucket 24 (44040192 elements, 44040192 padded size):
886
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight2
887
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight0
888
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight1
889
+ Params for bucket 25 (44040192 elements, 44040192 padded size):
890
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight7
891
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight3
892
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight5
893
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight4
894
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight6
895
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight2
896
+ Params for bucket 26 (44040192 elements, 44040192 padded size):
897
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight1
898
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight0
899
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight7
900
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight6
901
+ Params for bucket 27 (44040192 elements, 44040192 padded size):
902
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight5
903
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight3
904
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight4
905
+ Params for bucket 28 (44040192 elements, 44040192 padded size):
906
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight2
907
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight0
908
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight1
909
+ Params for bucket 29 (44040192 elements, 44040192 padded size):
910
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight4
911
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight6
912
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight2
913
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight7
914
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight5
915
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight3
916
+ Params for bucket 30 (44040192 elements, 44040192 padded size):
917
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight1
918
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight7
919
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight0
920
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight6
921
+ Params for bucket 31 (44040192 elements, 44040192 padded size):
922
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight5
923
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight4
924
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight3
925
+ Params for bucket 32 (44040192 elements, 44040192 padded size):
926
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight2
927
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight0
928
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight1
929
+ Params for bucket 33 (44040192 elements, 44040192 padded size):
930
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight4
931
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight6
932
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight2
933
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight7
934
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight5
935
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight3
936
+ Params for bucket 34 (44040192 elements, 44040192 padded size):
937
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight1
938
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight7
939
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight6
940
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight0
941
+ Params for bucket 35 (44040192 elements, 44040192 padded size):
942
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight5
943
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight4
944
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight3
945
+ Params for bucket 36 (44040192 elements, 44040192 padded size):
946
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight2
947
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight0
948
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight1
949
+ Params for bucket 37 (44040192 elements, 44040192 padded size):
950
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight7
951
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight5
952
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight3
953
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight6
954
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight4
955
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight2
956
+ Params for bucket 38 (44040192 elements, 44040192 padded size):
957
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight1
958
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight6
959
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight0
960
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight7
961
+ Params for bucket 39 (44040192 elements, 44040192 padded size):
962
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight5
963
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight3
964
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight4
965
+ Params for bucket 40 (44040192 elements, 44040192 padded size):
966
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight2
967
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight0
968
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight1
969
+ Params for bucket 41 (44040192 elements, 44040192 padded size):
970
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight7
971
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight3
972
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight5
973
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight6
974
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight4
975
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight2
976
+ Params for bucket 42 (44040192 elements, 44040192 padded size):
977
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight1
978
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight7
979
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight0
980
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight6
981
+ Params for bucket 43 (44040192 elements, 44040192 padded size):
982
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight5
983
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight4
984
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight3
985
+ Params for bucket 44 (44040192 elements, 44040192 padded size):
986
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight2
987
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight0
988
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight1
989
+ Params for bucket 45 (44040192 elements, 44040192 padded size):
990
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight7
991
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight3
992
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight5
993
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight4
994
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight6
995
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight2
996
+ Params for bucket 46 (44040192 elements, 44040192 padded size):
997
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight1
998
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight0
999
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight7
1000
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight6
1001
+ Params for bucket 47 (44040192 elements, 44040192 padded size):
1002
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight5
1003
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight3
1004
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight4
1005
+ Params for bucket 48 (44040192 elements, 44040192 padded size):
1006
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight2
1007
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight0
1008
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight1
1009
+ Params for bucket 49 (44040192 elements, 44040192 padded size):
1010
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight7
1011
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight3
1012
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight5
1013
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight6
1014
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight4
1015
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight2
1016
+ Params for bucket 50 (44040192 elements, 44040192 padded size):
1017
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight1
1018
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight0
1019
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight7
1020
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight6
1021
+ Params for bucket 51 (44040192 elements, 44040192 padded size):
1022
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight5
1023
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight3
1024
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight4
1025
+ Params for bucket 52 (44040192 elements, 44040192 padded size):
1026
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight2
1027
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight0
1028
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight1
1029
+ Params for bucket 53 (44040192 elements, 44040192 padded size):
1030
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight7
1031
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight5
1032
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight3
1033
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight4
1034
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight6
1035
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight2
1036
+ Params for bucket 54 (44040192 elements, 44040192 padded size):
1037
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight1
1038
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight7
1039
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight6
1040
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight0
1041
+ Params for bucket 55 (44040192 elements, 44040192 padded size):
1042
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight5
1043
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight4
1044
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight3
1045
+ Params for bucket 56 (44040192 elements, 44040192 padded size):
1046
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight2
1047
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight0
1048
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight1
1049
+ Params for bucket 57 (44040192 elements, 44040192 padded size):
1050
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight7
1051
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight5
1052
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight3
1053
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight6
1054
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight4
1055
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight2
1056
+ Params for bucket 58 (44040192 elements, 44040192 padded size):
1057
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight1
1058
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight0
1059
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight7
1060
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight6
1061
+ Params for bucket 59 (44040192 elements, 44040192 padded size):
1062
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight5
1063
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight3
1064
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight4
1065
+ Params for bucket 60 (44040192 elements, 44040192 padded size):
1066
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight2
1067
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight0
1068
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight1
1069
+ Params for bucket 61 (44040192 elements, 44040192 padded size):
1070
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight6
1071
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight4
1072
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight2
1073
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight7
1074
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight5
1075
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight3
1076
+ Params for bucket 62 (44040192 elements, 44040192 padded size):
1077
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight1
1078
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight6
1079
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight0
1080
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight7
1081
+ Params for bucket 63 (44040192 elements, 44040192 padded size):
1082
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight5
1083
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight4
1084
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight3
1085
+ Params for bucket 64 (44040192 elements, 44040192 padded size):
1086
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight2
1087
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight0
1088
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight1
1089
+ 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 0x7fb354bfa5a0>, config_logger_dir='')
1090
+ INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine
1091
+ [after model, optimizer, and learning rate scheduler are built] datetime: 2026-09-17 03:33:27
1092
+ > building train, validation, and test datasets ...
1093
+ > datasets target sizes (minimum size):
1094
+ train: 160
1095
+ validation: 0
1096
+ test: 0
1097
+ INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None]
1098
+ > building train, validation, and test datasets for GPT ...
1099
+ INFO:megatron.core.datasets.blended_megatron_dataset_builder:Building GPTDataset splits with sizes=(160, 0, 0) and config=GPTDatasetConfig(random_seed=1234, sequence_length=4096, blend=(['/workspace/data/wikitext/wikitext103_gpt2_text_document'], None), blend_per_split=None, split='100,0,0', split_matrix=[(0, 1.0), None, None], num_dataset_builder_threads=1, path_to_cache=None, mmap_bin_files=True, mock=False, tokenizer=<megatron.training.tokenizer.tokenizer._GPT2BPETokenizer object at 0x7fb358c192b0>, 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)
1100
+ INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx
1101
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths
1102
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers
1103
+ INFO:megatron.core.datasets.indexed_dataset: Extract the document indices
1104
+ INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029
1105
+ INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350
1106
+ INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices
1107
+ INFO:megatron.core.datasets.gpt_dataset: Load the document index from 68684a515e118253a8572c4517cf2ffe-GPTDataset-train-document_index.npy
1108
+ INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 68684a515e118253a8572c4517cf2ffe-GPTDataset-train-sample_index.npy
1109
+ INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 68684a515e118253a8572c4517cf2ffe-GPTDataset-train-shuffle_index.npy
1110
+ INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073
1111
+ > finished creating GPT datasets ...
1112
+ [after dataloaders are built] datetime: 2026-09-17 03:33:27
1113
+ done with setup ...
1114
+ (min, max) time across ranks (ms):
1115
+ model-and-optimizer-setup ......................: (3188.55, 3194.23)
1116
+ train/valid/test-data-iterators-setup ..........: (107.22, 135.77)
1117
+ training ...
1118
+ Overwriting rerun_state_machine.current_iteration from -1 to 0...
1119
+ [before the start of training step] datetime: 2026-09-17 03:33:27
1120
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1121
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1122
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1123
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1124
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1125
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1126
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1127
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1128
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1129
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1130
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1131
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1132
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1133
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1134
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1135
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1136
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1137
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1138
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1139
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1140
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1141
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1142
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1143
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1144
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1145
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1146
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1147
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1148
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1149
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1150
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1151
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1152
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1153
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1154
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1155
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1156
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1157
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1158
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1159
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1160
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1161
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1162
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1163
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1164
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1165
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1166
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1167
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1168
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1169
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1170
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1171
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1172
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1173
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1174
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1175
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1176
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1177
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1178
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1179
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1180
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1181
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1182
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1183
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1184
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1185
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1186
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1187
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1188
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1189
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1190
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1191
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1192
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1193
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1194
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1195
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1196
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1197
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1198
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1199
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1200
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1201
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1202
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1203
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1204
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1205
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1206
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1207
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1208
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1209
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1210
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1211
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1212
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1213
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1214
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:482: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1215
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1216
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1217
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1218
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1219
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1220
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1221
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1222
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1223
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1224
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1225
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1226
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1227
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1228
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1229
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1230
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:334: UserWarning: The num_dispatched_tokens is deprecated, it will be removed in the future.
1231
+ warnings.warn("The num_dispatched_tokens is deprecated, it will be removed in the future.")
1232
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1233
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1234
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1235
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1236
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1237
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1238
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1239
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1240
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1241
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1242
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1243
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1244
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1245
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1246
+ /workspace/DeepEP/deep_ep/hybrid_ep_buffer.py:336: UserWarning: The use_host_meta is deprecated, it will be removed in the future.
1247
+ warnings.warn("The use_host_meta is deprecated, it will be removed in the future.")
1248
+ [2026-09-17 03:33:57] iteration 1/ 10 | consumed samples: 16 | elapsed time per iteration (ms): 30515.0 | throughput per GPU (TFLOP/s/GPU): 6.4 | learning rate: 2.000000E-05 | global batch size: 16 | lm loss: 1.102374E+01 | loss scale: 1.0 | grad norm: 44.521 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33920 | min safe capacity: 1.0352 |
1249
+ Number of parameters in transformer block in billions: 23.22
1250
+ Number of parameters in embedding layers in billions: 0.41
1251
+ Total number of parameters in billions: 23.63
1252
+ Number of parameters in most loaded shard in billions: 23.6320
1253
+ Theoretical memory footprints: weight and optimizer=169029.43 MB
1254
+ [Rank 0] (after 1 iterations) memory (MB) | allocated: 58125.50390625 | max allocated: 69789.0546875 | reserved: 76118.0 | max reserved: 76118.0
1255
+ [2026-09-17 03:33:58] iteration 2/ 10 | consumed samples: 32 | elapsed time per iteration (ms): 1083.3 | throughput per GPU (TFLOP/s/GPU): 179.9 | learning rate: 4.000000E-05 | global batch size: 16 | lm loss: 1.101671E+01 | loss scale: 1.0 | grad norm: 46.188 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33920 | min safe capacity: 1.0352 |
1256
+ [2026-09-17 03:33:59] iteration 3/ 10 | consumed samples: 48 | elapsed time per iteration (ms): 437.1 | throughput per GPU (TFLOP/s/GPU): 445.7 | learning rate: 6.000000E-05 | global batch size: 16 | lm loss: 1.046371E+01 | loss scale: 1.0 | grad norm: 211.813 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33920 | min safe capacity: 1.0352 |
1257
+ Capture CUDA graph for training!!!
1258
+ CUDA graph capture done!!!
1259
+ [2026-09-17 03:34:02] iteration 4/ 10 | consumed samples: 64 | elapsed time per iteration (ms): 3593.9 | throughput per GPU (TFLOP/s/GPU): 54.2 | learning rate: 8.000000E-05 | global batch size: 16 | lm loss: 1.314799E+01 | loss scale: 1.0 | grad norm: 39.078 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33920 | min safe capacity: 1.0352 |
1260
+ [2026-09-17 03:34:03] iteration 5/ 10 | consumed samples: 80 | elapsed time per iteration (ms): 360.8 | throughput per GPU (TFLOP/s/GPU): 540.0 | learning rate: 1.000000E-04 | global batch size: 16 | lm loss: 1.038756E+01 | loss scale: 1.0 | grad norm: 17.511 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 33920 | min safe capacity: 1.0352 |
1261
+ [2026-09-17 03:34:03] iteration 6/ 10 | consumed samples: 96 | elapsed time per iteration (ms): 383.9 | throughput per GPU (TFLOP/s/GPU): 507.6 | learning rate: 7.750000E-05 | global batch size: 16 | lm loss: 1.033255E+01 | loss scale: 1.0 | grad norm: 29.428 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34176 | min safe capacity: 1.0430 |
1262
+ [2026-09-17 03:34:06] iteration 7/ 10 | consumed samples: 112 | elapsed time per iteration (ms): 3303.0 | throughput per GPU (TFLOP/s/GPU): 59.0 | learning rate: 3.250000E-05 | global batch size: 16 | lm loss: 1.003726E+01 | loss scale: 1.0 | grad norm: 7.175 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34176 | min safe capacity: 1.0430 |
1263
+ [2026-09-17 03:34:07] iteration 8/ 10 | consumed samples: 128 | elapsed time per iteration (ms): 371.2 | throughput per GPU (TFLOP/s/GPU): 524.8 | learning rate: 1.000000E-05 | global batch size: 16 | lm loss: 9.514786E+00 | loss scale: 1.0 | grad norm: 5.913 | number of skipped iterations: 0 | number of nan iterations: 0 | peak recv tokens: 34176 | min safe capacity: 1.0430 |
1264
+ Capture range started in the application.
1265
+ Capture range ended in the application.
1266
+ Generating '/tmp/nsys-report-2835.qdstrm'
1267
+
1268
+
granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/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/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/trace_extension_build.log CHANGED
@@ -66,3 +66,9 @@ Emitting ninja build file /workspace/Megatron-UltraEP/build/temp.linux-x86_64-cp
66
  Compiling objects...
67
  Allowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N)
68
  [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
 
 
 
 
 
 
 
66
  Compiling objects...
67
  Allowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N)
68
  [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
69
+ [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
70
+ creating build/lib.linux-x86_64-cpython-312/megatron/core/transformer/moe
71
+ 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
72
+ copying build/lib.linux-x86_64-cpython-312/megatron/core/datasets/helpers_cpp.cpython-312-x86_64-linux-gnu.so -> megatron/core/datasets
73
+ 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
74
+ 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/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/trace_results.tsv ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ model backend capacity status report async_overlap_json
2
+ G2_K2 ultraep 1.15625 OK /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625/nsys/G2_K2__ultraep__matched__seq4096__gbs16__cap1p15625.nsys-rep
3
+ G3_K2 ultraep 1.1875 OK /workspace/granularity_next_b200/tiera_topk_trace/ultraep/tierbtrace/tierbtrace_gbs16/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875/nsys/G3_K2__ultraep__matched__seq4096__gbs16__cap1p1875.nsys-rep