AbhishekGhosh1998 commited on
Commit
0e03f9a
·
verified ·
1 Parent(s): 84e0478

granularity results granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4 @ 2026-09-21T07:02:26Z from sp-aiinfra-interactive-aghosh-8-mb99k-8xb200-moe-cg-0

Browse files
Files changed (19) hide show
  1. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/.completed/echo__tierb__gbs8 +1 -0
  2. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/.completed/moonep__tierbprobe__gbs8 +1 -0
  3. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/.completed/ultraep__tierb__gbs8 +1 -0
  4. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/STATUS.txt +5 -2
  5. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/hf_watcher.log +3 -0
  6. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/manifest.txt +19 -0
  7. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/pip_freeze.txt +317 -0
  8. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/provenance.txt +90 -0
  9. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep1/mem_samples.csv +224 -0
  10. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep1/run.env +105 -0
  11. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep1/train.log +901 -0
  12. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep2/mem_samples.csv +152 -0
  13. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep2/run.env +105 -0
  14. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep2/train.log +840 -0
  15. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/results.tsv +2 -0
  16. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/steps.tsv +3 -0
  17. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/study.log +159 -0
  18. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/tierb_timing_capacities.env +1 -0
  19. granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/tierb_timing_capacities.tsv +2 -0
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/.completed/echo__tierb__gbs8 ADDED
@@ -0,0 +1 @@
 
 
1
+ 2026-09-21T06:56:36+00:00
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/.completed/moonep__tierbprobe__gbs8 ADDED
@@ -0,0 +1 @@
 
 
1
+ 2026-09-21T06:52:36+00:00
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/.completed/ultraep__tierb__gbs8 ADDED
@@ -0,0 +1 @@
 
 
1
+ 2026-09-21T07:00:08+00:00
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/STATUS.txt CHANGED
@@ -1,10 +1,13 @@
1
  study_root: /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4
2
  gbs: 8
3
- updated: 2026-09-21T06:49:55+00:00
4
- current: moonep__tierbprobe__gbs8
5
 
6
  completed steps:
 
7
  echo__tierbprobe__gbs8
 
 
8
  ultraep__tierbprobe__gbs8
9
 
10
  failed steps:
 
1
  study_root: /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4
2
  gbs: 8
3
+ updated: 2026-09-21T07:00:10+00:00
4
+ current: moonep__tierb__gbs8
5
 
6
  completed steps:
7
+ echo__tierb__gbs8
8
  echo__tierbprobe__gbs8
9
+ moonep__tierbprobe__gbs8
10
+ ultraep__tierb__gbs8
11
  ultraep__tierbprobe__gbs8
12
 
13
  failed steps:
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/hf_watcher.log CHANGED
@@ -4,3 +4,6 @@ Watching /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4 -
4
  uploaded: /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4 -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4
5
  https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
6
  [2026-09-21T06:52:23+00:00] backup pass
 
 
 
 
4
  uploaded: /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4 -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4
5
  https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
6
  [2026-09-21T06:52:23+00:00] backup pass
7
+ uploaded: /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4 -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4
8
+ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
9
+ [2026-09-21T07:02:25+00:00] backup pass
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/manifest.txt ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ timestamp: 2026-09-21T07:00:25+00:00
2
+ host: sp-aiinfra-interactive-aghosh-8-mb99k-8xb200-moe-cg-0
3
+ gpu: NVIDIA B200 x8
4
+ driver: 580.95.05
5
+ study_scripts_branch: granularity-moonep
6
+ study_scripts_commit: d29f657
7
+ campaign: moonep
8
+ backend: moonep
9
+ phase: tierb
10
+ gbs: 8
11
+ seq: 4096
12
+ configs: G1 G2 G3 G4
13
+ capacity: 1.25
14
+ repeats: 3
15
+ nsys: 1
16
+ megatron_repo: /workspace/Megatron-MoE-Echo
17
+ megatron_branch_required: moonep_backend_async_grad_reduce
18
+ megatron_branch_actual: moonep_backend_async_grad_reduce
19
+ megatron_commit: 2e68c8ad
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/pip_freeze.txt ADDED
@@ -0,0 +1,317 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ absl-py==2.3.0
2
+ aiohappyeyeballs @ file:///rapids/aiohappyeyeballs-2.6.1-py3-none-any.whl#sha256=f349ba8f4b75cb25c99c5c2d84e997e485204d2902a9597802b0371f09331fb8
3
+ aiohttp @ file:///rapids/aiohttp-3.12.7-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl#sha256=94f98e0e5a49f89b252e115844f756c04fc8050f38252a32a3dd994ce8121f10
4
+ aiosignal @ file:///rapids/aiosignal-1.3.2-py2.py3-none-any.whl#sha256=45cde58e409a301715980c2b01d0c28bdde3770d8290b5eb2173759d9acb31a5
5
+ annotated-types==0.7.0
6
+ anyio==4.9.0
7
+ apex @ file:///opt/pytorch/apex/dist/apex-0.1-cp312-cp312-linux_x86_64.whl#sha256=a1508e4a386362e9c8f46c92ccada9110e421212b964c694235d0fd5a224b170
8
+ argon2-cffi==25.1.0
9
+ argon2-cffi-bindings==21.2.0
10
+ arrow==1.3.0
11
+ asciitree @ file:///rapids/asciitree-0.3.3-py3-none-any.whl#sha256=3e507ccfaeb417823158091fdb9d26b0205f6a59ba680b479a71db65c73d01d4
12
+ asttokens==3.0.0
13
+ astunparse==1.6.3
14
+ async-lru==2.0.5
15
+ attrs==25.3.0
16
+ audioread==3.0.1
17
+ babel==2.17.0
18
+ beautifulsoup4==4.13.4
19
+ black==25.1.0
20
+ bleach==6.2.0
21
+ blis==0.7.11
22
+ cachetools @ file:///rapids/cachetools-6.0.0-py3-none-any.whl#sha256=82e73ba88f7b30228b5507dce1a1f878498fc669d972aef2dde4f3a3c24f103e
23
+ catalogue==2.0.10
24
+ certifi==2025.4.26
25
+ cffi==1.17.1
26
+ charset-normalizer==3.4.2
27
+ click==8.5.0
28
+ cloudpathlib==0.21.1
29
+ cloudpickle @ file:///rapids/cloudpickle-3.1.1-py3-none-any.whl#sha256=c8c5a44295039331ee9dad40ba100a9c7297b6f988e50e87ccdf3765a668350e
30
+ cmake==3.31.6
31
+ comm==0.2.2
32
+ confection==0.1.5
33
+ contourpy==1.3.2
34
+ cuda-bindings @ file:///rapids/cuda_bindings-12.9.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl#sha256=ff0e28d1e34758654b9c961e1f55e4786e49aee6a4dbceaf3cc24c46c672df7e
35
+ cuda-python @ file:///rapids/cuda_python-12.9.0-py3-none-any.whl#sha256=926acba49b2c0a0374c61b7c98f337c085199cf51cdfe4d6423c4129c20547a7
36
+ cudf @ file:///rapids/cudf-25.4.0-cp312-cp312-linux_x86_64.whl#sha256=b779a8e8c0e8577b9b560895958496da211bfac49cdc4f11a49e8d18cb82e348
37
+ cudf-polars @ file:///rapids/cudf_polars-25.4.0-py3-none-any.whl#sha256=227b6c5147e93ea86282c8f7219118f916769fa2ca0cdd931e97a148aa66b869
38
+ cugraph @ file:///rapids/cugraph-25.4.0-cp312-cp312-linux_x86_64.whl#sha256=b837c2dc2f7b235c702e4f8df64479b8fcbb342f848516d5faf0cdc4b5f31f29
39
+ cugraph-service-client @ file:///rapids/cugraph_service_client-25.4.0-py3-none-any.whl#sha256=7e568c3e41154b9752f5173555ffcc282ac52d72343feeb79653b2e2b1232470
40
+ cugraph-service-server @ file:///rapids/cugraph_service_server-25.4.0-py3-none-any.whl#sha256=a2830dcffafcccb16e12ebf85b98f8f247869d38c01ab4dd8664a1e7439d31be
41
+ cuml @ file:///rapids/cuml-25.4.0-cp312-cp312-linux_x86_64.whl#sha256=1e90a464a18b094e24797f1676688c5d7696c4cf97c934504214a53d5aa09aa5
42
+ cupy-cuda12x @ file:///rapids/cupy_cuda12x-13.3.0-cp312-cp312-linux_x86_64.whl#sha256=6a2dfa0aa1e8189a10d1534276c08f1d2f73b747fd28bfbe7b9f4ef66d9af73f
43
+ cuvs @ file:///rapids/cuvs-25.4.0-cp312-cp312-linux_x86_64.whl#sha256=b12f6b9f80e053a3b38542c2947d2516e468c02cad54a13d8b1a4a61d5aa7914
44
+ cycler==0.12.1
45
+ cymem==2.0.11
46
+ Cython==3.1.1
47
+ dask @ file:///rapids/dask-2025.2.0-py3-none-any.whl#sha256=f0fdeef6ceb0a06569d456c9e704f220f7f54e80f3a6ea42ab98cea6bc642b6e
48
+ dask-cuda @ file:///rapids/dask_cuda-25.4.0-py3-none-any.whl#sha256=442d87da0a95d00ca2a262b64c76e785a32ad905c82000fa2227932f81390e19
49
+ dask-cudf @ file:///rapids/dask_cudf-25.4.0-py3-none-any.whl#sha256=5e7cea0b26daea17da6e0bb457435e24460f76ba740358a3a18e239d6015c95c
50
+ datasets==5.0.1
51
+ debugpy==1.8.14
52
+ decorator==5.2.1
53
+ -e git+https://github.com/deepseek-ai/DeepEP.git@34152ae28f80bcc3ee38d7a12cb2ad87cfd4ea72#egg=deep_ep
54
+ defusedxml==0.7.1
55
+ dill==0.4.1
56
+ distributed @ file:///rapids/distributed-2025.2.0-py3-none-any.whl#sha256=368462084be47092d0a835f6c64c9b852da64dc88302b76d5f93bd5131228c67
57
+ distributed-ucxx @ file:///rapids/distributed_ucxx-0.43.0-py3-none-any.whl#sha256=94f60d0ccb0c94ba8dbfa1eaf455874d18aaab17e2ae218c38590e9fdc2d721d
58
+ dm-tree==0.1.9
59
+ einops==0.8.1
60
+ execnet==2.1.1
61
+ executing==2.2.0
62
+ expecttest==0.3.0
63
+ fasteners @ file:///rapids/fasteners-0.19-py3-none-any.whl#sha256=758819cb5d94cdedf4e836988b74de396ceacb8e2794d21f82d131fd9ee77237
64
+ fastjsonschema==2.21.1
65
+ fastrlock @ file:///rapids/fastrlock-0.8.3-cp312-cp312-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_28_x86_64.whl#sha256=5f13ec08f1adb1aa916c384b05ecb7dbebb8df9ea81abd045f60941c6283a670
66
+ filelock==3.18.0
67
+ flash_attn @ file:///opt/pytorch/flash_attn-2.7.4.post1-cp312-cp312-linux_x86_64.whl#sha256=a10455d6c56b387785b7413fd0bbcb392b42b3d4f4b88fda5d86cc8a9ce9182a
68
+ fonttools==4.58.1
69
+ fqdn==1.5.1
70
+ frozenlist @ file:///rapids/frozenlist-1.6.0-cp312-cp312-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl#sha256=4e1be9111cb6756868ac242b3c2bd1f09d9aea09846e4f5c23715e7afb647103
71
+ fsspec==2025.5.1
72
+ gast==0.6.0
73
+ grpcio==1.62.1
74
+ h11==0.16.0
75
+ hf-xet==1.6.0
76
+ httpcore==1.0.9
77
+ httpx==0.28.1
78
+ huggingface_hub==1.32.0
79
+ hypothesis==6.130.8
80
+ idna==3.10
81
+ importlib_metadata==8.7.0
82
+ iniconfig==2.1.0
83
+ intel-openmp==2021.4.0
84
+ ipykernel==6.29.5
85
+ ipython==9.3.0
86
+ ipython_pygments_lexers==1.1.1
87
+ isoduration==20.11.0
88
+ isort==6.0.1
89
+ jedi==0.19.2
90
+ Jinja2==3.1.6
91
+ joblib==1.5.1
92
+ json5==0.12.0
93
+ jsonpointer==3.0.0
94
+ jsonschema==4.24.0
95
+ jsonschema-specifications==2025.4.1
96
+ jupyter-events==0.12.0
97
+ jupyter-lsp==2.2.5
98
+ jupyter_client==8.6.3
99
+ jupyter_core==5.8.1
100
+ jupyter_server==2.16.0
101
+ jupyter_server_terminals==0.5.3
102
+ jupyterlab==4.4.3
103
+ jupyterlab_code_formatter==3.0.2
104
+ jupyterlab_pygments==0.3.0
105
+ jupyterlab_server==2.27.3
106
+ jupyterlab_tensorboard_pro @ file:///builder/jupyterlab_tensorboard_pro-4.0.0-py2.py3-none-any.whl#sha256=6bfcb5ced3358eb72150f4b2ab91312faef2f0ca14d845e8379d8fa52df6902e
107
+ jupytext==1.17.2
108
+ kiwisolver==1.4.8
109
+ kvikio @ file:///rapids/kvikio-25.4.0-cp312-cp312-linux_x86_64.whl#sha256=8b40fd8ea9e4a2974c04b50b7fc26b06094ac807f9a233281d0bafbaa71b1d85
110
+ langcodes==3.5.0
111
+ language_data==1.3.0
112
+ lazy_loader==0.4
113
+ libcudf @ file:///rapids/libcudf-25.4.0-py3-none-linux_x86_64.whl#sha256=238320b4958c1737bba2a609673e2539a78216b811b5498a68e00395d9b021a4
114
+ libcugraph @ file:///rapids/libcugraph-25.4.0-py3-none-linux_x86_64.whl#sha256=29085e901b2f7569ff2653bb1355fe6fad20f2fcaa585f5649da9867d8df766d
115
+ libcuml @ file:///rapids/libcuml-25.4.0-py3-none-linux_x86_64.whl#sha256=8e2e444c72b0e995e15a48da37ff12fb410a47a2baa69ba86ccc9b1a3db91e8c
116
+ libcuvs @ file:///rapids/libcuvs-25.4.0-py3-none-linux_x86_64.whl#sha256=e2f8db61c92b769c590dd752f43439ddd62b2f6592a3c337362b6c1ef220c3b9
117
+ libkvikio @ file:///rapids/libkvikio-25.4.0-py3-none-linux_x86_64.whl#sha256=da8a5713b05fc478bf40529a3e0fb61a2d605bb5a70ca4d461f96a6005dbe8b7
118
+ libraft @ file:///rapids/libraft-25.4.0-py3-none-linux_x86_64.whl#sha256=b921dec005aa4a321dd156776a18376feb47c96a5b67eec125c8a8e19edd1e24
119
+ librmm @ file:///rapids/librmm-25.4.0-py3-none-linux_x86_64.whl#sha256=ff951811bda729d4fa1109ab906225ae1f489a6ec2eeab5d1b674cc34bd4422f
120
+ librmm-cu12 @ file:///rapids/librmm_cu12-25.4.0-py3-none-any.whl#sha256=29296ef2b0e90679a0db280e86015be12d6ab9f4e62a181aa085f62dbadd06b3
121
+ librosa==0.11.0
122
+ libucxx @ file:///rapids/libucxx-0.43.0-py3-none-linux_x86_64.whl#sha256=f15849d7ac2fb68f0fd8998838a13321ddcfb0c594f7e6cbc401a3bcc4750566
123
+ lightning-thunder @ file:///opt/pytorch/lightning-thunder
124
+ lightning-utilities==0.14.3
125
+ lintrunner==0.12.7
126
+ llvmlite==0.42.0
127
+ locket @ file:///rapids/locket-1.0.0-py2.py3-none-any.whl#sha256=b6c819a722f7b6bd955b80781788e4a66a55628b858d347536b7e81325a3a5e3
128
+ looseversion==1.3.0
129
+ marisa-trie==1.2.1
130
+ Markdown==3.8
131
+ markdown-it-py==3.0.0
132
+ MarkupSafe==3.0.2
133
+ matplotlib==3.10.3
134
+ matplotlib-inline==0.1.7
135
+ mdit-py-plugins==0.4.2
136
+ mdurl==0.1.2
137
+ -e git+ssh://git@github.com/sriram-17-17/my-Megatron-MoE-Echo.git@2e68c8ad6bfb5bac5d6eb962fa7d4fbf0a5fb095#egg=megatron_core
138
+ mistune==3.1.3
139
+ mkl==2021.1.1
140
+ mkl-devel==2021.1.1
141
+ mkl-include==2021.1.1
142
+ ml_dtypes==0.6.0
143
+ mock==5.2.0
144
+ -e git+ssh://git@github.com/sriram-17-17/MoonEP.git@d8bc6667fad7177e6e09b637d773fa67f08bae93#egg=moonep
145
+ mpmath==1.3.0
146
+ msgpack==1.1.0
147
+ multidict @ file:///rapids/multidict-6.4.4-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl#sha256=a920f9cf2abdf6e493c519492d892c362007f113c94da4c239ae88429835bad1
148
+ multiprocess==0.70.19
149
+ murmurhash==1.0.13
150
+ mypy_extensions==1.1.0
151
+ nbclient==0.10.2
152
+ nbconvert==7.16.6
153
+ nbformat==5.10.4
154
+ nest-asyncio==1.6.0
155
+ networkx==3.5
156
+ ninja==1.11.1.4
157
+ notebook==7.4.3
158
+ notebook_shim==0.2.4
159
+ numba==0.59.1
160
+ numba-cuda @ file:///rapids/numba_cuda-0.4.0-py3-none-any.whl#sha256=5b95a5f42068771e37142944d44fe844fa87f1b2f89051461a4cb0af91bcc28a
161
+ numcodecs @ file:///rapids/numcodecs-0.13.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl#sha256=da2230484e6102e5fa3cc1a5dd37ca1f92dfbd183d91662074d6f7574e3e8f53
162
+ numpy==1.26.4
163
+ nvdlfw_inspect @ git+https://github.com/NVIDIA/nvidia-dlfw-inspect.git@6b60eb0e675606fb2fbbfcb12667ecbec75eaf87
164
+ nvfuser==0.2.27a0+9bf5aca
165
+ nvidia-cuda-cccl-cu12==12.9.27
166
+ nvidia-cudnn-frontend @ file:///opt/pytorch/lightning-thunder/tmp/cudnn_frontend
167
+ nvidia-cutlass-dsl==4.4.2
168
+ nvidia-cutlass-dsl-libs-base==4.4.2
169
+ nvidia-dali-cuda120==1.50.0
170
+ nvidia-ml-py @ file:///rapids/nvidia_ml_py-12.575.51-py3-none-any.whl#sha256=eb8641800d98ce40a22f479873f34b482e214a7e80349c63be51c3919845446e
171
+ nvidia-modelopt==0.29.0
172
+ nvidia-modelopt-core==0.29.0
173
+ nvidia-nvcomp-cu12==4.2.0.14
174
+ nvidia-nvimgcodec-cu12==0.5.0.13
175
+ nvidia-nvjpeg-cu12==12.4.0.16
176
+ nvidia-nvjpeg2k-cu12==0.8.1.40
177
+ nvidia-nvshmem-cu12==3.7.2
178
+ nvidia-nvtiff-cu12==0.5.0.67
179
+ nvidia-resiliency-ext==0.4.0
180
+ nvtx @ file:///rapids/nvtx-0.2.11-cp312-cp312-linux_x86_64.whl#sha256=73521e37cba42b4db0b3cd68fc5eb814e55c500a61eeed6edcdf6746315ac53e
181
+ nx-cugraph @ file:///rapids/nx_cugraph-25.4.0-py3-none-any.whl#sha256=e625444486cc899a77bd48e7da76a2edd95c0006dd95fbf323745feae7f33111
182
+ onnx @ file:///opt/pytorch/pytorch/third_party/onnx
183
+ onnx-ir==1.0.0
184
+ onnxscript==0.7.2
185
+ opt_einsum==3.4.0
186
+ optree==0.16.0
187
+ overrides==7.7.0
188
+ packaging==23.2
189
+ pandas @ file:///rapids/pandas-2.2.3-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl#sha256=fffb8ae78d8af97f849404f21411c95062db1496aeb3e56f146f0355c9989319
190
+ pandocfilters==1.5.1
191
+ parso==0.8.4
192
+ partd @ file:///rapids/partd-1.4.2-py3-none-any.whl#sha256=978e4ac767ec4ba5b86c6eaa52e5a2a3bc748a2ca839e8cc798f1cc6ce6efb0f
193
+ pathspec==0.12.1
194
+ pexpect==4.9.0
195
+ pillow==11.2.1
196
+ platformdirs==4.3.8
197
+ pluggy==1.6.0
198
+ ply @ file:///rapids/ply-3.11-py2.py3-none-any.whl#sha256=096f9b8350b65ebd2fd1346b12452efe5b9607f7482813ffca50c22722a807ce
199
+ polars @ file:///rapids/polars-1.25.2-cp39-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl#sha256=f7fcbb4f476784384ccda48757fca4e8c2e2c5a0a3aef3717aaf56aee4e30e09
200
+ polygraphy==0.49.20
201
+ pooch==1.8.2
202
+ preshed==3.0.10
203
+ prometheus_client==0.22.1
204
+ prompt_toolkit==3.0.51
205
+ propcache @ file:///rapids/propcache-0.3.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl#sha256=563f9d8c03ad645597b8d010ef4e9eab359faeb11a0a2ac9f7b4bc8c28ebef25
206
+ protobuf==4.24.4
207
+ psutil==7.0.0
208
+ ptyprocess==0.7.0
209
+ PuLP==3.2.1
210
+ pure_eval==0.2.3
211
+ pyarrow==25.0.1
212
+ pybind11==2.13.6
213
+ pybind11_global==2.13.6
214
+ pycocotools @ git+https://github.com/nvidia/cocoapi.git@08350a157b41f0755a536441b5abec89ab934c5b#subdirectory=PythonAPI
215
+ pycparser==2.22
216
+ pydantic==2.11.5
217
+ pydantic_core==2.33.2
218
+ Pygments==2.19.1
219
+ pylibcudf @ file:///rapids/pylibcudf-25.4.0-cp312-cp312-linux_x86_64.whl#sha256=401b8bfd022d06f84e120779ca7c3d01a28b68762aaa8f7873b10d6b1975cf6a
220
+ pylibcugraph @ file:///rapids/pylibcugraph-25.4.0-cp312-cp312-linux_x86_64.whl#sha256=061c7389a88c815b54fd5d3a3584fb87bf693e5602f5be335260d0ba575ba557
221
+ pylibcugraphops @ file:///rapids/pylibcugraphops-25.4.0-cp312-cp312-linux_x86_64.whl#sha256=a25ef4b22166c80718d5c7e5bddb755e81a2fe340fa5becd0ac36ec3edfb208a
222
+ pylibraft @ file:///rapids/pylibraft-25.4.0-cp312-cp312-linux_x86_64.whl#sha256=f2ecd23fe51f9c2caa87bbbe85cb32fcaff4ecfb52ceb9a9991bc50f82782c2e
223
+ pylibwholegraph @ file:///rapids/pylibwholegraph-25.4.0-cp312-cp312-linux_x86_64.whl#sha256=b82913d1e38bda76d8c272fbd8cc9072d987209d3d47cc93a464a37c2925aca2
224
+ pynvjitlink @ file:///rapids/pynvjitlink-0.3.0-cp312-cp312-linux_x86_64.whl#sha256=066caf1314730643d3b48c9847bd65789defdd6a7c78993176eb3d8bbe3f4919
225
+ pynvml @ file:///rapids/pynvml-12.0.0-py3-none-any.whl#sha256=fdff84b62a27dbe98e08e1a647eb77342bef1aebe0878bcd15e99a83fcbecb9e
226
+ pyparsing==3.2.3
227
+ pytest==8.1.1
228
+ pytest-flakefinder==1.1.0
229
+ pytest-rerunfailures==15.1
230
+ pytest-shard==0.1.2
231
+ pytest-xdist==3.7.0
232
+ python-dateutil==2.9.0.post0
233
+ python-hostlist==2.2.1
234
+ python-json-logger==3.3.0
235
+ pytorch-triton @ file:///tmp/dist/pytorch_triton-3.3.0%2Bgit96316ce52.nvinternal-cp312-cp312-linux_x86_64.whl#sha256=6d6921772b9dfd0243c055e0d054de42a113e92cabcfdd1d3e80a5ad9c08cf7d
236
+ pytz @ file:///rapids/pytz-2023.4-py2.py3-none-any.whl#sha256=f90ef520d95e7c46951105338d918664ebfd6f1d995bd7d153127ce90efafa6a
237
+ PyYAML==6.0.2
238
+ pyzmq==26.4.0
239
+ raft-dask @ file:///rapids/raft_dask-25.4.0-cp312-cp312-linux_x86_64.whl#sha256=36cf37ba6ce35454c4729c2cb597a360f22780eefb11d514a16a07137da5eb6c
240
+ rapids-dask-dependency @ file:///rapids/rapids_dask_dependency-25.4.0a0-py3-none-any.whl#sha256=484faca626e7760112a7d8ad3426f5bdd720b609006d8dca87c2607076a348bd
241
+ rapids-logger @ file:///rapids/rapids_logger-0.1.18-py3-none-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl#sha256=3617b2d86291ac6cb957542895f20fc9737ace29773ca90337e6fb04c73bd120
242
+ referencing==0.36.2
243
+ regex==2024.11.6
244
+ requests==2.32.3
245
+ rfc3339-validator==0.1.4
246
+ rfc3986-validator==0.1.1
247
+ rich==14.0.0
248
+ rmm @ file:///rapids/rmm-25.4.0-cp312-cp312-linux_x86_64.whl#sha256=1824bcab3836296666980a997ddcb9c4bca24dfb873012cfa0a7ecea583ca712
249
+ rpds-py==0.25.1
250
+ safetensors==0.5.3
251
+ scikit-learn==1.6.1
252
+ scipy==1.15.3
253
+ Send2Trash==1.8.3
254
+ setuptools==78.1.1
255
+ shellingham==1.5.4
256
+ six==1.16.0
257
+ smart-open==7.1.0
258
+ sniffio==1.3.1
259
+ sortedcontainers==2.4.0
260
+ soundfile==0.13.1
261
+ soupsieve==2.7
262
+ soxr==0.5.0.post1
263
+ spacy==3.7.5
264
+ spacy-legacy==3.0.12
265
+ spacy-loggers==1.0.5
266
+ srsly==2.5.1
267
+ stack-data==0.6.3
268
+ sympy==1.14.0
269
+ tabulate==0.9.0
270
+ tbb==2021.13.1
271
+ tblib @ file:///rapids/tblib-3.1.0-py3-none-any.whl#sha256=670bb4582578134b3d81a84afa1b016128b429f3d48e6cbbaecc9d15675e984e
272
+ tensorboard==2.16.2
273
+ tensorboard-data-server==0.7.2
274
+ tensorrt @ file:///workspace/TensorRT-10.11.0.33/python/tensorrt-10.11.0.33-cp312-none-linux_x86_64.whl#sha256=c2f18f72758a3d54a71901495c8fda9e7094a8bcdd2a5b6b58967fbf05733b3a
275
+ terminado==0.18.1
276
+ thinc==8.2.5
277
+ threadpoolctl==3.6.0
278
+ thriftpy2 @ file:///rapids/thriftpy2-0.5.2-cp312-cp312-linux_x86_64.whl#sha256=c0c7d939bdc5ccf7739154959938445d5fdf72f1d6e12bd563947a36f67e2147
279
+ tinycss2==1.4.0
280
+ toolz @ file:///rapids/toolz-1.0.0-py3-none-any.whl#sha256=292c8f1c4e7516bf9086f8850935c799a874039c8bcf959d47b600e4c44a6236
281
+ torch @ file:///opt/transfer/torch-2.8.0a0%2B5228986c39.nv25.6-29946474-cp312-cp312-linux_x86_64.whl#sha256=7a3c9ff5288d8102a31296b3d62491f3f47c8820a5c9aaefc7dc0fe20e1e8ad5
282
+ torch_tensorrt @ file:///opt/pytorch/torch_tensorrt/dist/torch_tensorrt-2.8.0a0-cp312-cp312-linux_x86_64.whl#sha256=26f4d7a04e7e6e017a5b917b8b98b83720bb2bb0195f48d25f52d567d286bfe1
283
+ torchao @ file:///opt/pytorch/ao
284
+ torchprofile==0.0.4
285
+ torchvision @ file:///opt/transfer/torchvision-0.22.0a0%2B95f10a4e-29946474-cp312-cp312-manylinux_2_28_x86_64.whl#sha256=746a3c674cc61680a86f4a14b374b21af1b5c21f629c42765b40c9494f6b234d
286
+ tornado==6.5.1
287
+ tqdm==4.67.1
288
+ traitlets==5.14.3
289
+ transformer_engine @ file:///pvcdata/aiinfra/b-aghosh/wheels/transformer_engine-2.10.0.dev0%2B10f6128d-cp312-cp312-linux_x86_64.whl#sha256=35cf5f4b83b2baad864e12e512bb422e0929409ac05610812fa521d4942d0189
290
+ treelite @ file:///rapids/treelite-4.4.1-py3-none-linux_x86_64.whl#sha256=174741724205540a1a5b37da81483aeff55d89548f42b3461cf07acac7ae96eb
291
+ typer==0.16.0
292
+ types-dataclasses==0.6.6
293
+ types-python-dateutil==2.9.0.20250516
294
+ typing-inspection==0.4.1
295
+ typing_extensions==4.14.0
296
+ tzdata @ file:///rapids/tzdata-2025.2-py2.py3-none-any.whl#sha256=1a403fada01ff9221ca8044d701868fa132215d84beb92242d9acd2147f667a8
297
+ ucx-py @ file:///rapids/ucx_py-0.43.0-cp312-cp312-linux_x86_64.whl#sha256=4a452fd8d8c983283ada3d49086e3657dc2ee94fbad08efdd708d15ab6e0b73a
298
+ ucxx @ file:///rapids/ucxx-0.43.0-cp312-cp312-linux_x86_64.whl#sha256=f1b4e55cb45c66fe88a2d8f9c2b87751165f3cfec21fbdc9c47c4ddff75e8a5f
299
+ -e git+https://github.com/Dots-Infra/UltraEP.git@94cab099b44fffa99a82fea99e7c12d89cf65e4f#egg=ultra_ep
300
+ uri-template==1.3.0
301
+ urllib3==2.0.7
302
+ wasabi==1.1.3
303
+ wcwidth==0.2.13
304
+ weasel==0.4.1
305
+ webcolors==24.11.1
306
+ webencodings==0.5.1
307
+ websocket-client==1.8.0
308
+ Werkzeug==3.1.3
309
+ wheel==0.45.1
310
+ wrapt==1.17.2
311
+ xdoctest==1.0.2
312
+ xgboost @ file:///rapids/xgboost-2.1.4-py3-none-manylinux_2_39_x86_64.whl#sha256=25eeca27f4048fc491940b3a0b4b5782ee48c33cb3aec059cef71f8da542930b
313
+ xxhash==4.0.1
314
+ yarl @ file:///rapids/yarl-1.20.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl#sha256=bf099e2432131093cc611623e0b0bcc399b8cddd9a91eded8bfb50402ec35018
315
+ zarr @ file:///rapids/zarr-2.18.7-py3-none-any.whl#sha256=ac3dc4033e9ae4e9d7b5e27c97ea3eaf1003cc0a07f010bd83d5134bf8c4b223
316
+ zict @ file:///rapids/zict-3.0.0-py2.py3-none-any.whl#sha256=5796e36bd0e0cc8cf0fbc1ace6a68912611c1dbd74750a3f3026b9b9d6a327ae
317
+ zipp==3.22.0
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/provenance.txt ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ==============================================================
2
+ PROVENANCE: moonep__tierb__gbs8
3
+ ==============================================================
4
+ captured_at : 2026-09-21T07:00:11+00:00
5
+ hostname : sp-aiinfra-interactive-aghosh-8-mb99k-8xb200-moe-cg-0
6
+ kernel : 6.8.0-1043-nvidia
7
+ cwd : /workspace/blackwell_run_scripts
8
+ command : run_ladder.sh --phase tierb --gbs 8 --out /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb --nsys
9
+
10
+ ---- GPU / driver --------------------------------------------
11
+ 0, NVIDIA B200, 10.0, 183359 MiB, 580.95.05
12
+ 1, NVIDIA B200, 10.0, 183359 MiB, 580.95.05
13
+ 2, NVIDIA B200, 10.0, 183359 MiB, 580.95.05
14
+ 3, NVIDIA B200, 10.0, 183359 MiB, 580.95.05
15
+ 4, NVIDIA B200, 10.0, 183359 MiB, 580.95.05
16
+ 5, NVIDIA B200, 10.0, 183359 MiB, 580.95.05
17
+ 6, NVIDIA B200, 10.0, 183359 MiB, 580.95.05
18
+ 7, NVIDIA B200, 10.0, 183359 MiB, 580.95.05
19
+ persistence/clocks:
20
+ 0, 120 MHz, 1965 MHz, 1000.00 W, 30
21
+ 1, 1965 MHz, 1965 MHz, 1000.00 W, 36
22
+ 2, 195 MHz, 1965 MHz, 1000.00 W, 35
23
+ 3, 120 MHz, 1965 MHz, 1000.00 W, 31
24
+ 4, 1965 MHz, 1965 MHz, 1000.00 W, 31
25
+ 5, 157 MHz, 1965 MHz, 1000.00 W, 34
26
+ 6, 240 MHz, 1965 MHz, 1000.00 W, 34
27
+ 7, 202 MHz, 1965 MHz, 1000.00 W, 30
28
+
29
+ NVLink topology:
30
+ GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7 NIC0 NIC1 NIC2 NIC3 NIC4 NIC5 NIC6 NIC7 NIC8 CPU Affinity NUMA Affinity GPU NUMA ID
31
+ GPU0 X NV18 NV18 NV18 NV18 NV18 NV18 NV18 PIX NODE SYS SYS SYS SYS SYS SYS NODE 0-27 0 N/A
32
+ GPU1 NV18 X NV18 NV18 NV18 NV18 NV18 NV18 NODE PIX SYS SYS SYS SYS SYS SYS NODE 0-27 0 N/A
33
+ GPU2 NV18 NV18 X NV18 NV18 NV18 NV18 NV18 SYS SYS PIX NODE SYS SYS SYS SYS SYS 28-55 1 N/A
34
+ GPU3 NV18 NV18 NV18 X NV18 NV18 NV18 NV18 SYS SYS NODE PIX SYS SYS SYS SYS SYS 28-55 1 N/A
35
+ GPU4 NV18 NV18 NV18 NV18 X NV18 NV18 NV18 SYS SYS SYS SYS PIX NODE SYS SYS SYS 56-83 2 N/A
36
+ GPU5 NV18 NV18 NV18 NV18 NV18 X NV18 NV18 SYS SYS SYS SYS NODE PIX SYS SYS SYS 56-83 2 N/A
37
+ GPU6 NV18 NV18 NV18 NV18 NV18 NV18 X NV18 SYS SYS SYS SYS SYS SYS PIX NODE SYS 84-111 3 N/A
38
+ GPU7 NV18 NV18 NV18 NV18 NV18 NV18 NV18 X SYS SYS SYS SYS SYS SYS NODE PIX SYS 84-111 3 N/A
39
+ NIC0 PIX NODE SYS SYS SYS SYS SYS SYS X NODE SYS SYS SYS SYS SYS SYS NODE
40
+ NIC1 NODE PIX SYS SYS SYS SYS SYS SYS NODE X SYS SYS SYS SYS SYS SYS NODE
41
+ NIC2 SYS SYS PIX NODE SYS SYS SYS SYS SYS SYS X NODE SYS SYS SYS SYS SYS
42
+ NIC3 SYS SYS NODE PIX SYS SYS SYS SYS SYS SYS NODE X SYS SYS SYS SYS SYS
43
+ NIC4 SYS SYS SYS SYS PIX NODE SYS SYS SYS SYS SYS SYS X NODE SYS SYS SYS
44
+ unavailable
45
+
46
+ ---- toolchain -----------------------------------------------
47
+ nvcc : release 12.9, V12.9.86
48
+ gcc : gcc (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
49
+ python : 3.12.3
50
+ torch : 2.8.0a0+5228986c39.nv25.06 cuda=12.9 nccl=2.27.3
51
+ torch arch : ['sm_75', 'sm_80', 'sm_86', 'sm_90', 'sm_100', 'sm_120', 'compute_120']
52
+ transformer_engine: 2.10.0.dev0+10f6128d [/usr/local/lib/python3.12/dist-packages/transformer_engine/__init__.py]
53
+ deep_ep : 1.2.1+34152ae [/workspace/DeepEP/deep_ep/__init__.py]
54
+ ultra_ep : 1.0.0 [/workspace/UltraEP/ultra_ep/__init__.py]
55
+ moonep : 0.0.1 [/workspace/MoonEP/moonep/__init__.py]
56
+ megatron : 0.14.0rc5 [None]
57
+ flash_attn : 2.7.4.post1 [/usr/local/lib/python3.12/dist-packages/flash_attn/__init__.py]
58
+ numpy : 1.26.4 [/usr/local/lib/python3.12/dist-packages/numpy/__init__.py]
59
+ huggingface_hub: 1.32.0 [/usr/local/lib/python3.12/dist-packages/huggingface_hub/__init__.py]
60
+
61
+ ---- repositories --------------------------------------------
62
+ Megatron-MoE-Echo moonep_backend_async_grad_reduce 2e68c8ad6bfb
63
+ Megatron-UltraEP ultraep-bracing-regions-through-kernels 03d57ef7d476
64
+ UltraEP main 94cab099b44f
65
+ DeepEP HEAD 34152ae28f80 (+3 untracked: build output)
66
+ MoonEP megatron-ddp-symmetric-params d8bc6667fad7
67
+ blackwell_run_scripts granularity-moonep d29f6570cdce
68
+
69
+ ---- environment ---------------------------------------------
70
+ CUDA_ARCH_LIST=7.5 8.0 8.6 9.0 10.0 12.0
71
+ CUDA_DRIVER_VERSION=575.57.08
72
+ CUDA_HOME=/usr/local/cuda
73
+ CUDA_MODULE_LOADING=LAZY
74
+ CUDA_VERSION=12.9.1.010
75
+ HF_REMOTE_ROOT=granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4
76
+ NCCL_VERSION=2.27.3
77
+ NVSHMEM_VERSION=3.2.5
78
+ PYTORCH_BUILD_NUMBER=0
79
+ PYTORCH_BUILD_VERSION=2.8.0a0+5228986
80
+ PYTORCH_HOME=/opt/pytorch/pytorch
81
+ PYTORCH_VERSION=2.8.0a0+5228986
82
+ TORCH_ALLOW_TF32_CUBLAS_OVERRIDE=1
83
+ TORCH_CUDA_ARCH_LIST=7.5 8.0 8.6 9.0 10.0 12.0+PTX
84
+ TORCH_NCCL_USE_COMM_NONBLOCKING=0
85
+
86
+ ---- shared library resolution -------------------------------
87
+ libnvshmem_host.so: not in ldconfig
88
+ libnccl.so: /lib/x86_64-linux-gnu/libnccl.so.2
89
+ lrwxrwxrwx 1 root root 20 Sep 21 04:57 /usr/local/lib/python3.12/dist-packages/nvidia/nvshmem/lib/libnvshmem_host.so -> libnvshmem_host.so.3
90
+ -rw-r--r-- 1 root root 179045744 Sep 21 04:54 /usr/local/lib/python3.12/dist-packages/nvidia/nvshmem/lib/libnvshmem_host.so.3
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep1/mem_samples.csv ADDED
@@ -0,0 +1,224 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 0, 0
2
+ 1, 0
3
+ 2, 0
4
+ 3, 0
5
+ 4, 0
6
+ 5, 0
7
+ 6, 0
8
+ 7, 0
9
+ 0, 4
10
+ 1, 4
11
+ 2, 4
12
+ 3, 1
13
+ 4, 1
14
+ 5, 1
15
+ 6, 1
16
+ 7, 1
17
+ 0, 4
18
+ 1, 4
19
+ 2, 4
20
+ 3, 4
21
+ 4, 4
22
+ 5, 4
23
+ 6, 4
24
+ 7, 2
25
+ 0, 0
26
+ 1, 0
27
+ 2, 0
28
+ 3, 0
29
+ 4, 0
30
+ 5, 0
31
+ 6, 0
32
+ 7, 0
33
+ 0, 0
34
+ 1, 0
35
+ 2, 0
36
+ 3, 0
37
+ 4, 0
38
+ 5, 0
39
+ 6, 0
40
+ 7, 0
41
+ 0, 4
42
+ 1, 1
43
+ 2, 1
44
+ 3, 1
45
+ 4, 1
46
+ 5, 1
47
+ 6, 1
48
+ 7, 1
49
+ 0, 4
50
+ 1, 4
51
+ 2, 4
52
+ 3, 4
53
+ 4, 4
54
+ 5, 1
55
+ 6, 1
56
+ 7, 1
57
+ 0, 4
58
+ 1, 4
59
+ 2, 4
60
+ 3, 4
61
+ 4, 4
62
+ 5, 4
63
+ 6, 4
64
+ 7, 4
65
+ 0, 4
66
+ 1, 4
67
+ 2, 4
68
+ 3, 47
69
+ 4, 4
70
+ 5, 4
71
+ 6, 4
72
+ 7, 4
73
+ 0, 821
74
+ 1, 821
75
+ 2, 821
76
+ 3, 821
77
+ 4, 821
78
+ 5, 821
79
+ 6, 821
80
+ 7, 821
81
+ 0, 1927
82
+ 1, 1927
83
+ 2, 1927
84
+ 3, 1927
85
+ 4, 1927
86
+ 5, 1927
87
+ 6, 1927
88
+ 7, 1927
89
+ 0, 37737
90
+ 1, 37755
91
+ 2, 37755
92
+ 3, 37769
93
+ 4, 37759
94
+ 5, 37755
95
+ 6, 37759
96
+ 7, 37757
97
+ 0, 45021
98
+ 1, 45133
99
+ 2, 45133
100
+ 3, 45133
101
+ 4, 45133
102
+ 5, 45133
103
+ 6, 45133
104
+ 7, 45133
105
+ 0, 75427
106
+ 1, 75743
107
+ 2, 75747
108
+ 3, 75741
109
+ 4, 75739
110
+ 5, 75733
111
+ 6, 75725
112
+ 7, 75749
113
+ 0, 76023
114
+ 1, 76329
115
+ 2, 76329
116
+ 3, 76329
117
+ 4, 76329
118
+ 5, 76329
119
+ 6, 76329
120
+ 7, 76329
121
+ 0, 76025
122
+ 1, 76329
123
+ 2, 76329
124
+ 3, 76329
125
+ 4, 76329
126
+ 5, 76329
127
+ 6, 76329
128
+ 7, 76329
129
+ 0, 76223
130
+ 1, 76527
131
+ 2, 76527
132
+ 3, 76527
133
+ 4, 76527
134
+ 5, 76527
135
+ 6, 76527
136
+ 7, 76527
137
+ 0, 76935
138
+ 1, 77239
139
+ 2, 77239
140
+ 3, 77239
141
+ 4, 77239
142
+ 5, 77239
143
+ 6, 77239
144
+ 7, 77239
145
+ 0, 77497
146
+ 1, 77241
147
+ 2, 77241
148
+ 3, 77241
149
+ 4, 77241
150
+ 5, 77241
151
+ 6, 77241
152
+ 7, 77241
153
+ 0, 81817
154
+ 1, 81609
155
+ 2, 81597
156
+ 3, 81583
157
+ 4, 81621
158
+ 5, 81613
159
+ 6, 81651
160
+ 7, 81659
161
+ 0, 82823
162
+ 1, 82545
163
+ 2, 82553
164
+ 3, 82563
165
+ 4, 82567
166
+ 5, 82579
167
+ 6, 82573
168
+ 7, 82557
169
+ 0, 83863
170
+ 1, 83635
171
+ 2, 83633
172
+ 3, 83607
173
+ 4, 83639
174
+ 5, 83653
175
+ 6, 83617
176
+ 7, 83605
177
+ 0, 97849
178
+ 1, 97509
179
+ 2, 97481
180
+ 3, 97505
181
+ 4, 97509
182
+ 5, 97509
183
+ 6, 97509
184
+ 7, 97509
185
+ 0, 97529
186
+ 1, 97509
187
+ 2, 97481
188
+ 3, 97505
189
+ 4, 97509
190
+ 5, 97509
191
+ 6, 97509
192
+ 7, 97247
193
+ 0, 95025
194
+ 1, 95267
195
+ 2, 94977
196
+ 3, 95263
197
+ 4, 95069
198
+ 5, 94685
199
+ 6, 95267
200
+ 7, 94553
201
+ 0, 93297
202
+ 1, 93155
203
+ 2, 93047
204
+ 3, 93583
205
+ 4, 93261
206
+ 5, 92941
207
+ 6, 92941
208
+ 7, 76425
209
+ 0, 68
210
+ 1, 76425
211
+ 2, 388
212
+ 3, 84485
213
+ 4, 324
214
+ 5, 4
215
+ 6, 4
216
+ 7, 4
217
+ 0, 0
218
+ 1, 0
219
+ 2, 0
220
+ 3, 0
221
+ 4, 0
222
+ 5, 0
223
+ 6, 0
224
+ 7, 0
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep1/run.env ADDED
@@ -0,0 +1,105 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=moonep
2
+ config=DEEPSEEK_V3_L4
3
+ family=tierB
4
+ source_model=DeepSeek-V3
5
+ standard_num_layers=61
6
+ base_rung=
7
+ repo=/workspace/Megatron-MoE-Echo
8
+ branch=moonep_backend_async_grad_reduce
9
+ commit=2e68c8ad
10
+ moonep_source_dir=/workspace/MoonEP
11
+ moonep_source_branch=megatron-ddp-symmetric-params
12
+ moonep_source_commit=d8bc666
13
+ num_layers=4
14
+ hidden=7168
15
+ ffn_hidden=18432
16
+ moe_ffn_hidden=2048
17
+ num_experts=256
18
+ topk=8
19
+ experts_per_rank=32
20
+ expert_params=44040192
21
+ heads=128
22
+ kv_groups=128
23
+ kv_channels=128
24
+ qk_layernorm=1
25
+ shared_ffn=2048
26
+ shared_expert_overlap=0
27
+ moe_token_dispatcher=flex
28
+ mla=1
29
+ q_lora_rank=1536
30
+ kv_lora_rank=512
31
+ qk_head_dim=128
32
+ qk_pos_emb_head_dim=64
33
+ v_head_dim=128
34
+ dense_first_layers=3
35
+ moe_layers=1
36
+ rotary_base=10000
37
+ rotary_percent=1.0
38
+ rotary_scaling=40
39
+ mscale=1.0
40
+ mscale_all_dim=1.0
41
+ qkv_bias=0
42
+ router_balancing=seq_aux_loss
43
+ aux_loss_coeff=1e-4
44
+ router_pre_softmax=0
45
+ router_fusion=0
46
+ router_dtype=fp32
47
+ router_groups=8
48
+ router_group_topk=4
49
+ router_score_function=sigmoid
50
+ router_expert_bias=1
51
+ router_bias_update_rate=1e-3
52
+ router_topk_scaling_factor=2.5
53
+ shared_expert_gate=0
54
+ apply_layernorm_1p=0
55
+ attention_output_gate=0
56
+ linear_attention_type=
57
+ linear_attention_freq=
58
+ linear_conv_kernel_dim=
59
+ linear_key_head_dim=
60
+ linear_value_head_dim=
61
+ linear_num_key_heads=
62
+ linear_num_value_heads=
63
+ mtp_num_layers=0
64
+ mtp_loss_scaling_factor=0.1
65
+ no_weight_decay_cond_type=
66
+ max_position_embeddings=4096
67
+ init_method_std=0.02
68
+ device_initiated_grouped_gemm=1
69
+ norm_epsilon=1e-6
70
+ seq_length=4096
71
+ micro_batch_size=1
72
+ global_batch_size=8
73
+ num_microbatches=1
74
+ assignments_per_rank=32768
75
+ mean_tokens_per_expert=1024
76
+ slot_policy=matched
77
+ slots_per_rank=32
78
+ capacity_mode=none
79
+ capacity=none
80
+ cuda_graph=1
81
+ probe_mode=0
82
+ async_grad_reduce=1
83
+ branch_role=timing
84
+ memory_dump_mode=none
85
+ memory_snapshot_path=
86
+ cuda_graph_warmup_steps=3
87
+ recompute=none
88
+ force_balance=0
89
+ train_iters=12
90
+ ep_size=8
91
+ echo_dispatch_overlap=1
92
+ echo_dispatch_recompute=0
93
+ ultraep_quota_min_tokens=
94
+ ultraep_weight_sync_join=
95
+ profile_enable=0
96
+ kernel_trace=0
97
+ fp8_mode=mxfp8
98
+ fp8_recipe=mxfp8
99
+ training_seed=1234
100
+ learning_rate=1e-4
101
+ min_learning_rate=1.0e-5
102
+ lr_warmup_iters=5
103
+ state_dump_enable=0
104
+ state_dump_dir=
105
+ moonep_assert_no_pending_reduce=1
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep1/train.log ADDED
@@ -0,0 +1,901 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [07:00:34] WARN: moonep_backend_async_grad_reduce lacks --moe-router-fusion; DEEPSEEK_V3_L4 runs unfused (timing only, routing identical)
2
+ ==============================================================
3
+ DEEPSEEK_V3_L4 / moonep (slots/rank=32 policy=matched)
4
+ E=256 K=8 H'=2048 m=32
5
+ L=4 seq=4096 gbs=8 (1 microbatches)
6
+ cap_mode=none cap=none cg=1 recompute=none
7
+ memory_dump=none snapshots=/workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep1/memory_snapshots/snapshot.pickle
8
+ repo=/workspace/Megatron-MoE-Echo @ 2e68c8ad
9
+ ==============================================================
10
+ W0921 07:00:35.571000 360201 torch/distributed/run.py:766]
11
+ W0921 07:00:35.571000 360201 torch/distributed/run.py:766] *****************************************
12
+ W0921 07:00:35.571000 360201 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
13
+ W0921 07:00:35.571000 360201 torch/distributed/run.py:766] *****************************************
14
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
15
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
16
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
17
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
18
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
19
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
20
+ /workspace/Megatron-MoE-Echo/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
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
23
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
24
+ 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
25
+ WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer
26
+ Number of virtual stages per pipeline stage: None
27
+ accumulate and all-reduce gradients in fp32 for bfloat16 data type.
28
+ using torch.bfloat16 for parameters ...
29
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
30
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
31
+ ------------------------ arguments ------------------------
32
+ account_for_embedding_in_pipeline_split ......... False
33
+ account_for_loss_in_pipeline_split .............. False
34
+ accumulate_allreduce_grads_in_fp32 .............. True
35
+ adam_beta1 ...................................... 0.9
36
+ adam_beta2 ...................................... 0.999
37
+ adam_eps ........................................ 1e-08
38
+ add_bias_linear ................................. False
39
+ add_position_embedding .......................... False
40
+ add_qkv_bias .................................... False
41
+ adlr_autoresume ................................. False
42
+ adlr_autoresume_interval ........................ 1000
43
+ align_grad_reduce ............................... True
44
+ align_param_gather .............................. False
45
+ app_tag_run_name ................................ None
46
+ app_tag_run_version ............................. 0.0.0
47
+ apply_layernorm_1p .............................. False
48
+ apply_query_key_layer_scaling ................... False
49
+ apply_residual_connection_post_layernorm ........ False
50
+ apply_rope_fusion ............................... False
51
+ async_save ...................................... None
52
+ async_tensor_model_parallel_allreduce ........... True
53
+ attention_backend ............................... AttnBackend.auto
54
+ attention_dropout ............................... 0.0
55
+ attention_softmax_in_fp32 ....................... False
56
+ auto_detect_ckpt_format ......................... False
57
+ barrier_with_L1_time ............................ True
58
+ bert_binary_head ................................ True
59
+ bert_embedder_type .............................. megatron
60
+ bert_load ....................................... None
61
+ bf16 ............................................ True
62
+ bias_dropout_fusion ............................. True
63
+ bias_gelu_fusion ................................ False
64
+ bias_swiglu_fusion .............................. True
65
+ biencoder_projection_dim ........................ 0
66
+ biencoder_shared_query_context_model ............ False
67
+ block_data_path ................................. None
68
+ calc_ft_timeouts ................................ False
69
+ calculate_per_token_loss ........................ False
70
+ check_for_large_grads ........................... False
71
+ check_for_nan_in_loss_and_grad .................. False
72
+ check_for_spiky_loss ............................ False
73
+ check_weight_hash_across_dp_replicas_interval ... None
74
+ ckpt_assume_constant_structure .................. False
75
+ ckpt_convert_format ............................. None
76
+ ckpt_convert_save ............................... None
77
+ ckpt_convert_update_legacy_dist_opt_format ...... False
78
+ ckpt_format ..................................... torch_dist
79
+ ckpt_fully_parallel_load ........................ False
80
+ ckpt_fully_parallel_save ........................ True
81
+ ckpt_fully_parallel_save_deprecated ............. False
82
+ ckpt_step ....................................... None
83
+ classes_fraction ................................ 1.0
84
+ clip_grad ....................................... 1.0
85
+ clone_scatter_output_in_embedding ............... True
86
+ config_logger_dir ...............................
87
+ consumed_train_samples .......................... 0
88
+ consumed_valid_samples .......................... 0
89
+ context_parallel_size ........................... 1
90
+ cp_comm_type .................................... ['p2p']
91
+ create_attention_mask_in_dataloader ............. True
92
+ cross_entropy_fusion_impl ....................... native
93
+ cross_entropy_loss_fusion ....................... False
94
+ cuda_graph_scope ................................ full_iteration
95
+ cuda_graph_warmup_steps ......................... 3
96
+ data_args_path .................................. None
97
+ data_cache_path ................................. None
98
+ data_parallel_random_init ....................... False
99
+ data_parallel_sharding_strategy ................. no_shard
100
+ data_parallel_size .............................. 8
101
+ data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document']
102
+ data_per_class_fraction ......................... 1.0
103
+ data_sharding ................................... True
104
+ dataloader_type ................................. single
105
+ ddp_average_in_collective ....................... False
106
+ ddp_bucket_size ................................. None
107
+ ddp_num_buckets ................................. None
108
+ ddp_pad_buckets_for_high_nccl_busbw ............. False
109
+ decoder_first_pipeline_num_layers ............... None
110
+ decoder_last_pipeline_num_layers ................ None
111
+ decoder_num_layers .............................. None
112
+ decoder_seq_length .............................. None
113
+ decoupled_lr .................................... None
114
+ decoupled_min_lr ................................ None
115
+ decrease_batch_size_if_needed ................... False
116
+ defer_embedding_wgrad_compute ................... False
117
+ delay_wgrad_compute ............................. False
118
+ deprecated_use_mcore_models ..................... True
119
+ deterministic_mode .............................. False
120
+ dino_bottleneck_size ............................ 256
121
+ dino_freeze_last_layer .......................... 1
122
+ dino_head_hidden_size ........................... 2048
123
+ dino_local_crops_number ......................... 10
124
+ dino_local_img_size ............................. 96
125
+ dino_norm_last_layer ............................ False
126
+ dino_teacher_temp ............................... 0.07
127
+ dino_warmup_teacher_temp ........................ 0.04
128
+ dino_warmup_teacher_temp_epochs ................. 30
129
+ disable_bf16_reduced_precision_matmul ........... False
130
+ disable_mamba_mem_eff_path ...................... False
131
+ disable_straggler_on_startup .................... False
132
+ dist_ckpt_format_deprecated ..................... None
133
+ dist_ckpt_strictness ............................ assume_ok_unexpected
134
+ distribute_saved_activations .................... False
135
+ distributed_backend ............................. nccl
136
+ distributed_timeout_minutes ..................... 10
137
+ embedding_init_method_std ....................... None
138
+ embedding_path .................................. None
139
+ empty_unused_memory_level ....................... 0
140
+ enable_cuda_graph ............................... True
141
+ enable_experimental ............................. False
142
+ enable_ft_package ............................... False
143
+ enable_gloo_process_groups ...................... True
144
+ enable_msc ...................................... True
145
+ enable_one_logger ............................... True
146
+ encoder_num_layers .............................. 4
147
+ encoder_seq_length .............................. 4096
148
+ end_weight_decay ................................ 0.1
149
+ eod_mask_loss ................................... False
150
+ error_injection_rate ............................ 0
151
+ error_injection_type ............................ transient_error
152
+ eval_interval ................................... 1000000
153
+ eval_iters ...................................... 0
154
+ evidence_data_path .............................. None
155
+ exit_duration_in_mins ........................... None
156
+ exit_interval ................................... None
157
+ exit_on_missing_checkpoint ...................... False
158
+ exit_signal_handler ............................. False
159
+ exp_avg_dtype ................................... torch.float32
160
+ exp_avg_sq_dtype ................................ torch.float32
161
+ expert_model_parallel_size ...................... 8
162
+ expert_tensor_parallel_size ..................... 1
163
+ export_force_local_attention .................... False
164
+ export_kd_cfg ................................... None
165
+ export_kd_teacher_ckpt_format ................... None
166
+ export_kd_teacher_load .......................... None
167
+ export_kv_cache_quant ........................... False
168
+ export_legacy_megatron .......................... False
169
+ export_model_type ............................... GPTModel
170
+ export_moe_apply_probs_on_input ................. False
171
+ export_qk_l2_norm ............................... False
172
+ export_quant_cfg ................................ None
173
+ export_real_quant_cfg ........................... None
174
+ export_te_mcore_model ........................... False
175
+ external_cuda_graph ............................. False
176
+ ffn_hidden_size ................................. 18432
177
+ finetune ........................................ False
178
+ finetune_data_split ............................. train
179
+ finetune_hf_dataset ............................. None
180
+ first_last_layers_bf16 .......................... False
181
+ flash_decode .................................... False
182
+ fp16 ............................................ False
183
+ fp16_lm_cross_entropy ........................... False
184
+ fp32_residual_connection ........................ False
185
+ fp8 ............................................. e4m3
186
+ fp8_amax_compute_algo ........................... most_recent
187
+ fp8_amax_history_len ............................ 1
188
+ fp8_interval .................................... 1
189
+ fp8_margin ...................................... 0
190
+ fp8_param_gather ................................ False
191
+ fp8_recipe ...................................... mxfp8
192
+ fp8_wgrad ....................................... True
193
+ fsdp_double_buffer .............................. False
194
+ global_batch_size ............................... 8
195
+ grad_reduce_in_bf16 ............................. False
196
+ gradient_accumulation_fusion .................... True
197
+ gradient_reduce_div_fusion ...................... True
198
+ group_query_attention ........................... False
199
+ head_lr_mult .................................... 1.0
200
+ heterogeneous_layers_config_encoded_json ........ None
201
+ heterogeneous_layers_config_path ................ None
202
+ hidden_dropout .................................. 0.0
203
+ hidden_size ..................................... 7168
204
+ hierarchical_context_parallel_sizes ............. None
205
+ high_priority_stream_groups ..................... []
206
+ hybrid_attention_ratio .......................... 0.0
207
+ hybrid_mlp_ratio ................................ 0.0
208
+ hybrid_override_pattern ......................... None
209
+ hysteresis ...................................... 2
210
+ ict_head_size ................................... None
211
+ ict_load ........................................ None
212
+ img_h ........................................... 224
213
+ img_w ........................................... 224
214
+ indexer_batch_size .............................. 128
215
+ indexer_log_interval ............................ 1000
216
+ inference_batch_times_seqlen_threshold .......... -1
217
+ inference_dynamic_batching ...................... False
218
+ inference_dynamic_batching_buffer_guaranteed_fraction 0.2
219
+ inference_dynamic_batching_buffer_overflow_factor None
220
+ inference_dynamic_batching_buffer_size_gb ....... 40.0
221
+ inference_dynamic_batching_chunk_size ........... 256
222
+ inference_dynamic_batching_max_requests_override None
223
+ inference_dynamic_batching_max_tokens_override .. None
224
+ inference_dynamic_batching_num_cuda_graphs ...... 16
225
+ inference_max_batch_size ........................ 8
226
+ inference_max_seq_length ........................ 2560
227
+ inference_rng_tracker ........................... False
228
+ init_method_std ................................. 0.02
229
+ init_method_xavier_uniform ...................... False
230
+ init_model_with_meta_device ..................... False
231
+ initial_loss_scale .............................. 4294967296
232
+ inprocess_active_world_size ..................... 8
233
+ inprocess_barrier_timeout ....................... 120
234
+ inprocess_completion_timeout .................... 120
235
+ inprocess_empty_cuda_cache ...................... False
236
+ inprocess_granularity ........................... node
237
+ inprocess_hard_timeout .......................... 90
238
+ inprocess_heartbeat_interval .................... 30
239
+ inprocess_heartbeat_timeout ..................... 60
240
+ inprocess_last_call_wait ........................ 1
241
+ inprocess_max_iterations ........................ None
242
+ inprocess_monitor_process_interval .............. 1.0
243
+ inprocess_monitor_thread_interval ............... 1.0
244
+ inprocess_progress_watchdog_interval ............ 1.0
245
+ inprocess_restart ............................... False
246
+ inprocess_soft_timeout .......................... 60
247
+ inprocess_termination_grace_time ................ 1
248
+ is_hybrid_model ................................. False
249
+ iter_per_epoch .................................. 1250
250
+ iterations_to_skip .............................. []
251
+ keep_fp8_transpose_cache_when_using_custom_fsdp . False
252
+ kitchen_config_file ............................. None
253
+ kitchen_recipe_number ........................... None
254
+ kv_channels ..................................... 128
255
+ kv_lora_rank .................................... 512
256
+ lazy_mpu_init ................................... None
257
+ load ............................................ None
258
+ load_main_params_from_ckpt ...................... None
259
+ load_model_opt_format ........................... False
260
+ local_rank ...................................... 0
261
+ log_energy ...................................... False
262
+ log_interval .................................... 1
263
+ log_loss_scale_to_tensorboard ................... True
264
+ log_memory_to_tensorboard ....................... False
265
+ log_num_zeros_in_grad ........................... False
266
+ log_params_norm ................................. False
267
+ log_progress .................................... False
268
+ log_straggler ................................... False
269
+ log_throughput .................................. True
270
+ log_timers_to_tensorboard ....................... False
271
+ log_validation_ppl_to_tensorboard ............... False
272
+ log_world_size_to_tensorboard ................... False
273
+ logging_level ................................... None
274
+ loss_scale ...................................... None
275
+ loss_scale_window ............................... 1000
276
+ lr .............................................. 0.0001
277
+ lr_decay_iters .................................. 8
278
+ lr_decay_samples ................................ None
279
+ lr_decay_style .................................. cosine
280
+ lr_warmup_fraction .............................. None
281
+ lr_warmup_init .................................. 0.0
282
+ lr_warmup_iters ................................. 5
283
+ lr_warmup_samples ............................... 0
284
+ lr_wsd_decay_iters .............................. None
285
+ lr_wsd_decay_samples ............................ None
286
+ lr_wsd_decay_style .............................. exponential
287
+ main_grads_dtype ................................ torch.float32
288
+ main_params_dtype ............................... torch.float32
289
+ make_vocab_size_divisible_by .................... 128
290
+ mamba_head_dim .................................. 64
291
+ mamba_num_groups ................................ 8
292
+ mamba_num_heads ................................. None
293
+ mamba_state_dim ................................. 128
294
+ manual_gc ....................................... False
295
+ manual_gc_eval .................................. True
296
+ manual_gc_interval .............................. 0
297
+ mask_factor ..................................... 1.0
298
+ mask_prob ....................................... 0.15
299
+ mask_type ....................................... random
300
+ masked_softmax_fusion ........................... False
301
+ max_position_embeddings ......................... 4096
302
+ max_tokens_to_oom ............................... 12000
303
+ memory_snapshot_path ............................ snapshot.pickle
304
+ merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt
305
+ micro_batch_size ................................ 1
306
+ microbatch_group_size_per_vp_stage .............. None
307
+ mid_level_dataset_surplus ....................... 0.005
308
+ min_loss_scale .................................. 1.0
309
+ min_lr .......................................... 1e-05
310
+ mlp_chunks_for_prefill .......................... 1
311
+ mmap_bin_files .................................. True
312
+ mock_data ....................................... False
313
+ moe_apply_probs_on_input ........................ False
314
+ moe_aux_loss_coeff .............................. 0.0001
315
+ moe_deepep_num_sms .............................. 20
316
+ moe_echo_enable_random_offloading ............... False
317
+ moe_echo_expert_dispatch_overlap ................ False
318
+ moe_echo_expert_dispatcher_type ................. hybridep
319
+ moe_echo_recompute_expert_dispatch .............. False
320
+ moe_enable_deepep ............................... False
321
+ moe_enable_echo ................................. False
322
+ moe_enable_hybridep ............................. False
323
+ moe_enable_moonep ............................... True
324
+ moe_expert_capacity_factor ...................... None
325
+ moe_extended_tp ................................. False
326
+ moe_ffn_hidden_size ............................. 2048
327
+ moe_grouped_gemm ................................ True
328
+ moe_hybridep_num_sms ............................ 16
329
+ moe_input_jitter_eps ............................ None
330
+ moe_layer_freq .................................. [0, 0, 0, 1]
331
+ moe_layer_recompute ............................. False
332
+ moe_moonep_async_grad_reduce .................... True
333
+ moe_moonep_num_sms .............................. None
334
+ moe_num_echo_experts ............................ None
335
+ moe_pad_expert_input_to_capacity ................ False
336
+ moe_per_layer_logging ........................... False
337
+ moe_permute_fusion .............................. False
338
+ moe_received_token_capacity ..................... None
339
+ moe_router_bias_update_rate ..................... 0.001
340
+ moe_router_dtype ................................ fp32
341
+ moe_router_enable_expert_bias ................... True
342
+ moe_router_force_load_balancing ................. False
343
+ moe_router_group_topk ........................... 4
344
+ moe_router_load_balancing_type .................. seq_aux_loss
345
+ moe_router_num_groups ........................... 8
346
+ moe_router_padding_for_fp8 ...................... False
347
+ moe_router_pre_softmax .......................... False
348
+ moe_router_score_function ....................... sigmoid
349
+ moe_router_topk ................................. 8
350
+ moe_router_topk_scaling_factor .................. 2.5
351
+ moe_shared_expert_intermediate_size ............. 2048
352
+ moe_shared_expert_overlap ....................... False
353
+ moe_token_dispatcher_type ....................... flex
354
+ moe_token_drop_policy ........................... probs
355
+ moe_upcycling_granularity ....................... 1
356
+ moe_use_device_initiated_grouped_gemm ........... True
357
+ moe_use_legacy_grouped_gemm ..................... False
358
+ moe_use_upcycling ............................... False
359
+ moe_z_loss_coeff ................................ None
360
+ mrope_section ................................... None
361
+ mscale .......................................... 1.0
362
+ mscale_all_dim .................................. 1.0
363
+ mtp_loss_scaling_factor ......................... 0.1
364
+ mtp_num_layers .................................. None
365
+ multi_latent_attention .......................... True
366
+ nccl_all_reduce_for_prefill ..................... False
367
+ nccl_communicator_config_path ................... None
368
+ nccl_ub ......................................... False
369
+ no_load_optim ................................... None
370
+ no_load_rng ..................................... None
371
+ no_persist_layer_norm ........................... False
372
+ no_rope_freq .................................... None
373
+ no_save_optim ................................... None
374
+ no_save_rng ..................................... None
375
+ non_persistent_ckpt_type ........................ None
376
+ non_persistent_global_ckpt_dir .................. None
377
+ non_persistent_local_ckpt_algo .................. fully_parallel
378
+ non_persistent_local_ckpt_dir ................... None
379
+ non_persistent_save_interval .................... None
380
+ norm_epsilon .................................... 1e-06
381
+ normalization ................................... RMSNorm
382
+ num_attention_heads ............................. 128
383
+ num_channels .................................... 3
384
+ num_classes ..................................... 1000
385
+ num_dataset_builder_threads ..................... 1
386
+ num_distributed_optimizer_instances ............. 1
387
+ num_experts ..................................... 256
388
+ num_layers ...................................... 4
389
+ num_layers_at_end_in_bf16 ....................... 1
390
+ num_layers_at_start_in_bf16 ..................... 1
391
+ num_layers_per_virtual_pipeline_stage ........... None
392
+ num_query_groups ................................ 1
393
+ num_virtual_stages_per_pipeline_rank ............ None
394
+ num_workers ..................................... 2
395
+ object_storage_cache_path ....................... None
396
+ one_logger_async ................................ False
397
+ one_logger_project .............................. megatron-lm
398
+ one_logger_run_name ............................. None
399
+ onnx_safe ....................................... None
400
+ openai_gelu ..................................... False
401
+ optimizer ....................................... adam
402
+ optimizer_cpu_offload ........................... False
403
+ optimizer_offload_fraction ...................... 1.0
404
+ output_bert_embeddings .......................... False
405
+ overlap_cpu_optimizer_d2h_h2d ................... False
406
+ overlap_grad_reduce ............................. True
407
+ overlap_moe_expert_parallel_comm ................ False
408
+ overlap_p2p_comm ................................ False
409
+ overlap_p2p_comm_warmup_flush ................... False
410
+ overlap_param_gather ............................ True
411
+ overlap_param_gather_with_optimizer_step ........ False
412
+ override_opt_param_scheduler .................... False
413
+ params_dtype .................................... torch.bfloat16
414
+ patch_dim ....................................... 16
415
+ per_split_data_args_path ........................ None
416
+ perform_initialization .......................... True
417
+ pin_cpu_grads ................................... True
418
+ pin_cpu_params .................................. True
419
+ pipeline_model_parallel_comm_backend ............ None
420
+ pipeline_model_parallel_layout .................. None
421
+ pipeline_model_parallel_size .................... 1
422
+ position_embedding_type ......................... rope
423
+ pretrained_checkpoint ........................... None
424
+ profile ......................................... False
425
+ profile_ranks ................................... [0]
426
+ profile_step_end ................................ 12
427
+ profile_step_start .............................. 10
428
+ q_lora_rank ..................................... 1536
429
+ qk_head_dim ..................................... 128
430
+ qk_l2_norm ...................................... False
431
+ qk_layernorm .................................... True
432
+ qk_pos_emb_head_dim ............................. 64
433
+ query_in_block_prob ............................. 0.1
434
+ rampup_batch_size ............................... None
435
+ rank ............................................ 0
436
+ recompute_granularity ........................... None
437
+ recompute_method ................................ None
438
+ recompute_modules ............................... None
439
+ recompute_num_layers ............................ None
440
+ record_memory_history ........................... False
441
+ relative_attention_max_distance ................. 128
442
+ relative_attention_num_buckets .................. 32
443
+ replication ..................................... False
444
+ replication_factor .............................. 2
445
+ replication_jump ................................ None
446
+ rerun_mode ...................................... validate_results
447
+ reset_attention_mask ............................ False
448
+ reset_position_ids .............................. False
449
+ result_rejected_tracker_filename ................ None
450
+ retriever_report_topk_accuracies ................ []
451
+ retriever_score_scaling ......................... False
452
+ retriever_seq_length ............................ 256
453
+ retro_add_retriever ............................. False
454
+ retro_attention_gate ............................ 1
455
+ retro_cyclic_train_iters ........................ None
456
+ retro_encoder_attention_dropout ................. 0.1
457
+ retro_encoder_hidden_dropout .................... 0.1
458
+ retro_encoder_layers ............................ 2
459
+ retro_num_neighbors ............................. 2
460
+ retro_num_retrieved_chunks ...................... 2
461
+ retro_project_dir ............................... None
462
+ retro_verify_neighbor_count ..................... True
463
+ reuse_grad_buf_for_mxfp8_param_ag ............... False
464
+ rope_scaling_factor ............................. 8.0
465
+ rope_type ....................................... None
466
+ rotary_base ..................................... 10000
467
+ rotary_interleaved .............................. False
468
+ rotary_percent .................................. 1.0
469
+ rotary_scaling_factor ........................... 40.0
470
+ rotary_seq_len_interpolation_factor ............. None
471
+ run_workload_inspector_server ................... False
472
+ sample_rate ..................................... 1.0
473
+ save ............................................ None
474
+ save_interval ................................... None
475
+ save_retain_interval ............................ None
476
+ scatter_gather_tensors_in_pipeline .............. True
477
+ seed ............................................ 1234
478
+ seq_length ...................................... 4096
479
+ sequence_parallel ............................... False
480
+ sft ............................................. False
481
+ sft_tokenizer_prompt_format ..................... nemotron-h-aligned
482
+ sgd_momentum .................................... 0.9
483
+ short_seq_prob .................................. 0.1
484
+ skip_train ...................................... False
485
+ skipped_train_samples ........................... 0
486
+ spec ............................................ None
487
+ split ........................................... 100,0,0
488
+ squared_relu .................................... False
489
+ start_weight_decay .............................. 0.1
490
+ straggler_ctrlr_port ............................ 65535
491
+ straggler_minmax_count .......................... 1
492
+ suggested_communication_unit_size ............... None
493
+ swiglu .......................................... True
494
+ swin_backbone_type .............................. tiny
495
+ symmetric_ar_type ............................... None
496
+ te_rng_tracker .................................. True
497
+ tensor_model_parallel_size ...................... 1
498
+ tensorboard_dir ................................. None
499
+ tensorboard_log_interval ........................ 1
500
+ tensorboard_queue_size .......................... 1000
501
+ test_data_path .................................. None
502
+ test_mode ....................................... False
503
+ tiktoken_num_special_tokens ..................... 1000
504
+ tiktoken_pattern ................................ None
505
+ tiktoken_special_tokens ......................... None
506
+ timing_log_level ................................ 0
507
+ timing_log_option ............................... minmax
508
+ titles_data_path ................................ None
509
+ tokenizer_model ................................. None
510
+ tokenizer_type .................................. GPT2BPETokenizer
511
+ torch_fsdp2_reshard_after_forward ............... True
512
+ tp_comm_bootstrap_backend ....................... nccl
513
+ tp_comm_bulk_dgrad .............................. True
514
+ tp_comm_bulk_wgrad .............................. True
515
+ tp_comm_overlap ................................. False
516
+ tp_comm_overlap_ag .............................. True
517
+ tp_comm_overlap_cfg ............................. None
518
+ tp_comm_overlap_rs .............................. True
519
+ tp_comm_overlap_rs_dgrad ........................ False
520
+ tp_comm_split_ag ................................ True
521
+ tp_comm_split_rs ................................ True
522
+ train_data_path ................................. None
523
+ train_iters ..................................... 12
524
+ train_samples ................................... None
525
+ train_sync_interval ............................. None
526
+ transformer_impl ................................ transformer_engine
527
+ transformer_pipeline_model_parallel_size ........ 1
528
+ untie_embeddings_and_output_weights ............. True
529
+ use_checkpoint_args ............................. False
530
+ use_checkpoint_opt_param_scheduler .............. False
531
+ use_cpu_initialization .......................... None
532
+ use_custom_fsdp ................................. False
533
+ use_dist_ckpt ................................... True
534
+ use_dist_ckpt_deprecated ........................ False
535
+ use_distributed_optimizer ....................... True
536
+ use_flash_attn .................................. False
537
+ use_legacy_models ............................... False
538
+ use_mp_args_from_checkpoint_args ................ False
539
+ use_one_sent_docs ............................... False
540
+ use_persistent_ckpt_worker ...................... False
541
+ use_precision_aware_optimizer ................... False
542
+ use_pytorch_profiler ............................ False
543
+ use_ring_exchange_p2p ........................... False
544
+ use_rope_scaling ................................ False
545
+ use_rotary_position_embeddings .................. False
546
+ use_sharp ....................................... False
547
+ use_tokenizer_model_from_checkpoint_args ........ True
548
+ use_torch_fsdp2 ................................. False
549
+ use_torch_optimizer_for_cpu_offload ............. False
550
+ use_tp_pp_dp_mapping ............................ False
551
+ v_head_dim ...................................... 128
552
+ valid_data_path ................................. None
553
+ variable_seq_lengths ............................ False
554
+ virtual_pipeline_model_parallel_size ............ None
555
+ vision_backbone_type ............................ vit
556
+ vision_pretraining .............................. False
557
+ vision_pretraining_type ......................... classify
558
+ vocab_extra_ids ................................. 0
559
+ vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json
560
+ vocab_size ...................................... None
561
+ wandb_exp_name ..................................
562
+ wandb_project ...................................
563
+ wandb_save_dir ..................................
564
+ weight_decay .................................... 0.1
565
+ weight_decay_incr_style ......................... constant
566
+ wgrad_deferral_limit ............................ 0
567
+ world_size ...................................... 8
568
+ yaml_cfg ........................................ None
569
+ -------------------- end of arguments ---------------------
570
+ INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1
571
+ > building GPT2BPETokenizer tokenizer ...
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
+ /workspace/Megatron-MoE-Echo/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-MoE-Echo/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
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
581
+ [Gloo] Rank [Gloo] Rank 2 is connected to 17 is connected to peer ranks. [Gloo] Rank 7Expected number of connected peer ranks is : peer ranks. 47Expected number of connected peer ranks is : is connected to
582
+ 77
583
+ peer ranks. Expected number of connected peer ranks is : 7
584
+ [Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
585
+ [Gloo] Rank [Gloo] Rank 6 is connected to 77 is connected to peer ranks. 7Expected number of connected peer ranks is : peer ranks. 7Expected number of connected peer ranks is :
586
+ 7
587
+ [Gloo] Rank 3 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 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
590
+ [Gloo] Rank 1 is connected to 7 peer ranks. [Gloo] Rank Expected number of connected peer ranks is : 75
591
+ is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
592
+ [Gloo] Rank 4 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
595
+ [Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
596
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
597
+ > initialized tensor model parallel with size 1
598
+ > initialized pipeline model parallel with size 1
599
+ > setting random seeds to 1234 ...
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 [Gloo] Rank 00 is connected to is connected to 00 peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : 00
606
+
607
+ [rank4]:[W921 07:00:52.580330149 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.
608
+ [rank2]:[W921 07:00:52.580909327 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.
609
+ [rank1]:[W921 07:00:52.580938519 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.
610
+ [rank6]:[W921 07:00:52.581302862 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.
611
+ [rank7]:[W921 07:00:52.581332662 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.
612
+ [rank3]:[W921 07:00:52.581443502 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.
613
+ [rank5]:[W921 07:00:52.582662342 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.
614
+ > compiling dataset index builder ...
615
+ make: Entering directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
616
+ make: Nothing to be done for 'default'.
617
+ make: Leaving directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
618
+ >>> done with dataset index builder. Compilation time: 0.034 seconds
619
+ WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations.
620
+ > compiling and loading fused kernels ...
621
+ [rank0]:[W921 07:00:52.761131081 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.
622
+ >>> done with compiling and loading fused kernels. Compilation time: 3.955 seconds
623
+ time to initialize megatron (seconds): 7.458
624
+ [after megatron is initialized] datetime: 2026-09-21 07:00:58
625
+ building GPT model ...
626
+ /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.
627
+ warnings.warn(
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
+ > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 4113898496
643
+ 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)
644
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 17
645
+ Params for bucket 1 (360579072 elements, 360579072 padded size):
646
+ module.output_layer.weight
647
+ Params for bucket 2 (44047360 elements, 44047360 padded size):
648
+ module.decoder.final_layernorm.weight
649
+ module.decoder.layers.3.mlp.shared_experts.linear_fc2.weight
650
+ module.decoder.layers.3.mlp.shared_experts.linear_fc1.weight
651
+ Params for bucket 3 (60497408 elements, 60497408 padded size):
652
+ module.decoder.layers.3.self_attention.linear_kv_down_proj.weight
653
+ module.decoder.layers.3.self_attention.linear_kv_up_proj.layer_norm_weight
654
+ module.decoder.layers.3.self_attention.linear_q_up_proj.weight
655
+ module.decoder.layers.3.mlp.router.weight
656
+ module.decoder.layers.3.pre_mlp_layernorm.weight
657
+ module.decoder.layers.3.self_attention.linear_kv_up_proj.weight
658
+ Params for bucket 4 (128452096 elements, 128452096 padded size):
659
+ module.decoder.layers.3.self_attention.linear_q_up_proj.layer_norm_weight
660
+ module.decoder.layers.3.self_attention.linear_proj.weight
661
+ module.decoder.layers.3.self_attention.linear_q_down_proj.weight
662
+ Params for bucket 5 (132127744 elements, 132127744 padded size):
663
+ module.decoder.layers.3.input_layernorm.weight
664
+ module.decoder.layers.2.mlp.linear_fc2.weight
665
+ Params for bucket 6 (264241152 elements, 264241152 padded size):
666
+ module.decoder.layers.2.mlp.linear_fc1.weight
667
+ Params for bucket 7 (58662400 elements, 58662400 padded size):
668
+ module.decoder.layers.2.mlp.linear_fc1.layer_norm_weight
669
+ module.decoder.layers.2.self_attention.linear_kv_up_proj.layer_norm_weight
670
+ module.decoder.layers.2.self_attention.linear_kv_up_proj.weight
671
+ module.decoder.layers.2.self_attention.linear_kv_down_proj.weight
672
+ module.decoder.layers.2.self_attention.linear_q_up_proj.weight
673
+ Params for bucket 8 (128452096 elements, 128452096 padded size):
674
+ module.decoder.layers.2.self_attention.linear_q_up_proj.layer_norm_weight
675
+ module.decoder.layers.2.self_attention.linear_proj.weight
676
+ module.decoder.layers.2.self_attention.linear_q_down_proj.weight
677
+ Params for bucket 9 (132127744 elements, 132127744 padded size):
678
+ module.decoder.layers.2.input_layernorm.weight
679
+ module.decoder.layers.1.mlp.linear_fc2.weight
680
+ Params for bucket 10 (264241152 elements, 264241152 padded size):
681
+ module.decoder.layers.1.mlp.linear_fc1.weight
682
+ Params for bucket 11 (58662400 elements, 58662400 padded size):
683
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.layer_norm_weight
684
+ module.decoder.layers.1.mlp.linear_fc1.layer_norm_weight
685
+ module.decoder.layers.1.self_attention.linear_q_up_proj.weight
686
+ module.decoder.layers.1.self_attention.linear_kv_down_proj.weight
687
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.weight
688
+ Params for bucket 12 (128452096 elements, 128452096 padded size):
689
+ module.decoder.layers.1.self_attention.linear_q_up_proj.layer_norm_weight
690
+ module.decoder.layers.1.self_attention.linear_q_down_proj.weight
691
+ module.decoder.layers.1.self_attention.linear_proj.weight
692
+ Params for bucket 13 (132127744 elements, 132127744 padded size):
693
+ module.decoder.layers.1.input_layernorm.weight
694
+ module.decoder.layers.0.mlp.linear_fc2.weight
695
+ Params for bucket 14 (264241152 elements, 264241152 padded size):
696
+ module.decoder.layers.0.mlp.linear_fc1.weight
697
+ Params for bucket 15 (58662400 elements, 58662400 padded size):
698
+ module.decoder.layers.0.mlp.linear_fc1.layer_norm_weight
699
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.weight
700
+ module.decoder.layers.0.self_attention.linear_kv_down_proj.weight
701
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.layer_norm_weight
702
+ module.decoder.layers.0.self_attention.linear_q_up_proj.weight
703
+ Params for bucket 16 (128452096 elements, 128452096 padded size):
704
+ module.decoder.layers.0.self_attention.linear_q_up_proj.layer_norm_weight
705
+ module.decoder.layers.0.self_attention.linear_proj.weight
706
+ module.decoder.layers.0.self_attention.linear_q_down_proj.weight
707
+ Params for bucket 17 (360586240 elements, 360586240 padded size):
708
+ module.decoder.layers.0.input_layernorm.weight
709
+ module.embedding.word_embeddings.weight
710
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 27
711
+ Params for bucket 1 (44040192 elements, 44040192 padded size):
712
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight0
713
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight1
714
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight2
715
+ Params for bucket 2 (44040192 elements, 44040192 padded size):
716
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight3
717
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight4
718
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight5
719
+ Params for bucket 3 (44040192 elements, 44040192 padded size):
720
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight6
721
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight7
722
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight8
723
+ Params for bucket 4 (44040192 elements, 44040192 padded size):
724
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight9
725
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight10
726
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight11
727
+ Params for bucket 5 (44040192 elements, 44040192 padded size):
728
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight12
729
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight13
730
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight14
731
+ Params for bucket 6 (44040192 elements, 44040192 padded size):
732
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight15
733
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight16
734
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight17
735
+ Params for bucket 7 (44040192 elements, 44040192 padded size):
736
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight18
737
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight19
738
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight20
739
+ Params for bucket 8 (44040192 elements, 44040192 padded size):
740
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight21
741
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight22
742
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight23
743
+ Params for bucket 9 (44040192 elements, 44040192 padded size):
744
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight24
745
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight25
746
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight26
747
+ Params for bucket 10 (44040192 elements, 44040192 padded size):
748
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight27
749
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight29
750
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight28
751
+ Params for bucket 11 (58720256 elements, 58720256 padded size):
752
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight30
753
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight0
754
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight31
755
+ Params for bucket 12 (58720256 elements, 58720256 padded size):
756
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight1
757
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight2
758
+ Params for bucket 13 (58720256 elements, 58720256 padded size):
759
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight3
760
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight4
761
+ Params for bucket 14 (58720256 elements, 58720256 padded size):
762
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight5
763
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight6
764
+ Params for bucket 15 (58720256 elements, 58720256 padded size):
765
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight7
766
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight8
767
+ Params for bucket 16 (58720256 elements, 58720256 padded size):
768
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight9
769
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight10
770
+ Params for bucket 17 (58720256 elements, 58720256 padded size):
771
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight11
772
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight12
773
+ Params for bucket 18 (58720256 elements, 58720256 padded size):
774
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight13
775
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight14
776
+ Params for bucket 19 (58720256 elements, 58720256 padded size):
777
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight15
778
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight16
779
+ Params for bucket 20 (58720256 elements, 58720256 padded size):
780
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight17
781
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight18
782
+ Params for bucket 21 (58720256 elements, 58720256 padded size):
783
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight19
784
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight20
785
+ Params for bucket 22 (58720256 elements, 58720256 padded size):
786
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight21
787
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight22
788
+ Params for bucket 23 (58720256 elements, 58720256 padded size):
789
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight23
790
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight24
791
+ Params for bucket 24 (58720256 elements, 58720256 padded size):
792
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight25
793
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight26
794
+ Params for bucket 25 (58720256 elements, 58720256 padded size):
795
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight27
796
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight28
797
+ Params for bucket 26 (58720256 elements, 58720256 padded size):
798
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight29
799
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight30
800
+ Params for bucket 27 (29360128 elements, 29360128 padded size):
801
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight31
802
+ 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 0x7a8e1f036b10>, config_logger_dir='')
803
+ INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine
804
+ [after model, optimizer, and learning rate scheduler are built] datetime: 2026-09-21 07:01:00
805
+ > building train, validation, and test datasets ...
806
+ > datasets target sizes (minimum size):
807
+ train: 96
808
+ validation: 0
809
+ test: 0
810
+ INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None]
811
+ > building train, validation, and test datasets for GPT ...
812
+ INFO:megatron.core.datasets.blended_megatron_dataset_builder:Building GPTDataset splits with sizes=(96, 0, 0) and config=GPTDatasetConfig(random_seed=1234, sequence_length=4096, blend=(['/workspace/data/wikitext/wikitext103_gpt2_text_document'], None), blend_per_split=None, split='100,0,0', split_matrix=[(0, 1.0), None, None], num_dataset_builder_threads=1, path_to_cache=None, mmap_bin_files=True, mock=False, tokenizer=<megatron.training.tokenizer.tokenizer._GPT2BPETokenizer object at 0x7a8b791ecbc0>, 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)
813
+ INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx
814
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths
815
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers
816
+ INFO:megatron.core.datasets.indexed_dataset: Extract the document indices
817
+ INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029
818
+ INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350
819
+ INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices
820
+ INFO:megatron.core.datasets.gpt_dataset: Load the document index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-document_index.npy
821
+ INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-sample_index.npy
822
+ INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-shuffle_index.npy
823
+ INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073
824
+ > finished creating GPT datasets ...
825
+ [after dataloaders are built] datetime: 2026-09-21 07:01:00
826
+ done with setup ...
827
+ training ...
828
+ Overwriting rerun_state_machine.current_iteration from -1 to 0...
829
+ (min, max) time across ranks (ms):
830
+ model-and-optimizer-setup ......................: (1898.47, 1898.56)
831
+ train/valid/test-data-iterators-setup ..........: (66.96, 71.97)
832
+ [before the start of training step] datetime: 2026-09-21 07:01:01
833
+ INFO:moonep.api:MoonEP comm buffer sizes:
834
+ config : R=8, S=4096, H=7168, E=256, K=8, NvS=40896, NvS_padded=40960
835
+ sm_count : num_sms=32, num_sms_dedup=148
836
+ hidden_buffer_size : 560.00 MiB
837
+ meta_buffer_size : 2.00 MiB
838
+ local_temp_buffer_size : 1.27 MiB
839
+ total_size : 563.27 MiB
840
+ [MoonEP selfcheck 1/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287] sizes [768, 640, 640, 1024, 640, 640, 640, 512, 1280, 640, 896, 640, 1536, 2304, 1536, 896, 1408, 384, 1280, 1664, 1152, 640, 640, 1152, 1280, 1536, 768, 512, 1664, 384, 384, 896, 2432, 1792, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[215, 197, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [194, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [205, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [78, 91, 86, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [203, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [50, 32, 34, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [82, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
841
+ Number of parameters in transformer block in billions: 13.26
842
+ Number of parameters in embedding layers in billions: 0.72
843
+ Total number of parameters in billions: 13.98
844
+ Number of parameters in most loaded shard in billions: 13.9771
845
+ Theoretical memory footprints: weight and optimizer=99972.23 MB
846
+ [2026-09-21 07:01:24] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 23919.2 | throughput per GPU (TFLOP/s/GPU): 3.1 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.226155E+01 | loss scale: 1.0 | grad norm: 15.011 | number of skipped iterations: 0 | number of nan iterations: 0 |
847
+ [Rank 0] (after 1 iterations) memory (MB) | allocated: 46904.38818359375 | max allocated: 57624.96875 | reserved: 60652.0 | max reserved: 60652.0
848
+ [MoonEP selfcheck 2/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287] sizes [768, 768, 768, 640, 512, 640, 768, 512, 1408, 640, 896, 768, 1408, 2560, 1408, 1024, 1280, 512, 1664, 1664, 1152, 640, 640, 1024, 1280, 1408, 640, 512, 1536, 384, 384, 896, 2176, 1664, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[205, 197, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [158, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [151, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [82, 78, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [50, 32, 34, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [91, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [215, 194, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
849
+ [2026-09-21 07:01:25] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 209.4 | throughput per GPU (TFLOP/s/GPU): 355.6 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.227770E+01 | loss scale: 1.0 | grad norm: 14.853 | number of skipped iterations: 0 | number of nan iterations: 0 |
850
+ [MoonEP selfcheck 3/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287] sizes [384, 640, 1152, 512, 384, 768, 896, 256, 896, 512, 1024, 768, 1664, 1920, 1024, 896, 1408, 256, 1152, 1792, 256, 768, 384, 896, 768, 1024, 384, 256, 640, 384, 384, 1024, 5248, 3200, 1280, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[50, 34, 62, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [66, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [151, 158, 137, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [81, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [82, 78, 91, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [239, 252, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [52, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
851
+ [2026-09-21 07:01:25] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 197.9 | throughput per GPU (TFLOP/s/GPU): 376.4 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 1.070648E+01 | loss scale: 1.0 | grad norm: 18.850 | number of skipped iterations: 0 | number of nan iterations: 0 |
852
+ Capture CUDA graph for training!!!
853
+ CUDA graph capture done!!!
854
+ [2026-09-21 07:01:26] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 722.0 | throughput per GPU (TFLOP/s/GPU): 103.1 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 1.557410E+01 | loss scale: 1.0 | grad norm: 97.887 | number of skipped iterations: 0 | number of nan iterations: 0 |
855
+ [2026-09-21 07:01:26] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 193.3 | throughput per GPU (TFLOP/s/GPU): 385.3 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.447998E+01 | loss scale: 1.0 | grad norm: 97.902 | number of skipped iterations: 0 | number of nan iterations: 0 |
856
+ [2026-09-21 07:01:26] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 192.3 | throughput per GPU (TFLOP/s/GPU): 387.3 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 1.293705E+01 | loss scale: 1.0 | grad norm: 16.034 | number of skipped iterations: 0 | number of nan iterations: 0 |
857
+ [2026-09-21 07:01:26] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 192.9 | throughput per GPU (TFLOP/s/GPU): 385.9 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 1.689716E+01 | loss scale: 1.0 | grad norm: 100.878 | number of skipped iterations: 0 | number of nan iterations: 0 |
858
+ [2026-09-21 07:01:26] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 192.4 | throughput per GPU (TFLOP/s/GPU): 387.0 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.098987E+01 | loss scale: 1.0 | grad norm: 18.714 | number of skipped iterations: 0 | number of nan iterations: 0 |
859
+ [2026-09-21 07:01:27] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 192.3 | throughput per GPU (TFLOP/s/GPU): 387.3 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.048015E+01 | loss scale: 1.0 | grad norm: 16.061 | number of skipped iterations: 0 | number of nan iterations: 0 |
860
+ [2026-09-21 07:01:27] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 193.1 | throughput per GPU (TFLOP/s/GPU): 385.6 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.017142E+01 | loss scale: 1.0 | grad norm: 11.421 | number of skipped iterations: 0 | number of nan iterations: 0 |
861
+ [2026-09-21 07:01:27] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 192.4 | throughput per GPU (TFLOP/s/GPU): 387.0 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.971597E+00 | loss scale: 1.0 | grad norm: 11.756 | number of skipped iterations: 0 | number of nan iterations: 0 |
862
+ [2026-09-21 07:01:27] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 193.3 | throughput per GPU (TFLOP/s/GPU): 385.2 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.694415E+00 | loss scale: 1.0 | grad norm: 17.317 | number of skipped iterations: 0 | number of nan iterations: 0 |
863
+ [after training is done] datetime: 2026-09-21 07:01:27
864
+ [rank0]:[W921 07:01:28.770602025 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
865
+ [rank7]:[W921 07:01:29.718959758 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
866
+ [rank4]:[W921 07:01:29.891130677 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
867
+ [rank5]:[W921 07:01:29.891176373 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
868
+ [rank2]:[W921 07:01:29.892298178 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
869
+ [rank1]:[W921 07:01:30.254574517 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
870
+ [rank6]:[W921 07:01:30.274196412 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
871
+ [rank3]:[W921 07:01:30.307623770 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
872
+ W0921 07:01:35.935000 360201 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers
873
+ W0921 07:01:35.935000 360201 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 360394 closing signal SIGTERM
874
+ Traceback (most recent call last):
875
+ File "/usr/local/bin/torchrun", line 8, in <module>
876
+ sys.exit(main())
877
+ ^^^^^^
878
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
879
+ return f(*args, **kwargs)
880
+ ^^^^^^^^^^^^^^^^^^
881
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 892, in main
882
+ run(args)
883
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 883, in run
884
+ elastic_launch(
885
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
886
+ return launch_agent(self._config, self._entrypoint, list(args))
887
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
888
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 261, in launch_agent
889
+ result = agent.run()
890
+ ^^^^^^^^^^^
891
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper
892
+ result = f(*args, **kwargs)
893
+ ^^^^^^^^^^^^^^^^^^
894
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 711, in run
895
+ result = self._invoke_run(role)
896
+ ^^^^^^^^^^^^^^^^^^^^^^
897
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 870, in _invoke_run
898
+ time.sleep(monitor_interval)
899
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 84, in _terminate_process_handler
900
+ raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval)
901
+ torch.distributed.elastic.multiprocessing.api.SignalException: Process 360201 got signal: 15
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep2/mem_samples.csv ADDED
@@ -0,0 +1,152 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 0, 0
2
+ 1, 0
3
+ 2, 0
4
+ 3, 0
5
+ 4, 0
6
+ 5, 0
7
+ 6, 0
8
+ 7, 0
9
+ 0, 4
10
+ 1, 2
11
+ 2, 1
12
+ 3, 1
13
+ 4, 1
14
+ 5, 1
15
+ 6, 1
16
+ 7, 1
17
+ 0, 4
18
+ 1, 4
19
+ 2, 4
20
+ 3, 4
21
+ 4, 4
22
+ 5, 4
23
+ 6, 2
24
+ 7, 1
25
+ 0, 1
26
+ 1, 1
27
+ 2, 1
28
+ 3, 1
29
+ 4, 1
30
+ 5, 4
31
+ 6, 4
32
+ 7, 4
33
+ 0, 0
34
+ 1, 0
35
+ 2, 0
36
+ 3, 0
37
+ 4, 0
38
+ 5, 0
39
+ 6, 0
40
+ 7, 0
41
+ 0, 1
42
+ 1, 1
43
+ 2, 1
44
+ 3, 1
45
+ 4, 1
46
+ 5, 1
47
+ 6, 1
48
+ 7, 1
49
+ 0, 4
50
+ 1, 4
51
+ 2, 4
52
+ 3, 4
53
+ 4, 4
54
+ 5, 1
55
+ 6, 1
56
+ 7, 1
57
+ 0, 4
58
+ 1, 4
59
+ 2, 4
60
+ 3, 4
61
+ 4, 4
62
+ 5, 4
63
+ 6, 4
64
+ 7, 1
65
+ 0, 4
66
+ 1, 4
67
+ 2, 4
68
+ 3, 4
69
+ 4, 4
70
+ 5, 4
71
+ 6, 4
72
+ 7, 4
73
+ 0, 821
74
+ 1, 821
75
+ 2, 821
76
+ 3, 821
77
+ 4, 821
78
+ 5, 821
79
+ 6, 821
80
+ 7, 821
81
+ 0, 1927
82
+ 1, 1927
83
+ 2, 1927
84
+ 3, 1927
85
+ 4, 1927
86
+ 5, 1927
87
+ 6, 1927
88
+ 7, 1927
89
+ 0, 2359
90
+ 1, 1783
91
+ 2, 2359
92
+ 3, 1783
93
+ 4, 2359
94
+ 5, 1783
95
+ 6, 1783
96
+ 7, 2359
97
+ 0, 44613
98
+ 1, 44613
99
+ 2, 44613
100
+ 3, 44613
101
+ 4, 44613
102
+ 5, 44613
103
+ 6, 44613
104
+ 7, 44613
105
+ 0, 54307
106
+ 1, 54611
107
+ 2, 54611
108
+ 3, 54613
109
+ 4, 54611
110
+ 5, 54611
111
+ 6, 54611
112
+ 7, 54611
113
+ 0, 76021
114
+ 1, 76325
115
+ 2, 76325
116
+ 3, 76325
117
+ 4, 76325
118
+ 5, 76325
119
+ 6, 76325
120
+ 7, 76325
121
+ 0, 76023
122
+ 1, 76329
123
+ 2, 76329
124
+ 3, 76329
125
+ 4, 76329
126
+ 5, 76329
127
+ 6, 76329
128
+ 7, 76329
129
+ 0, 76025
130
+ 1, 76329
131
+ 2, 76329
132
+ 3, 76329
133
+ 4, 76329
134
+ 5, 76329
135
+ 6, 76329
136
+ 7, 76329
137
+ 0, 76223
138
+ 1, 76527
139
+ 2, 76527
140
+ 3, 76527
141
+ 4, 76527
142
+ 5, 76527
143
+ 6, 76527
144
+ 7, 76527
145
+ 0, 76935
146
+ 1, 77239
147
+ 2, 77239
148
+ 3, 77239
149
+ 4, 77239
150
+ 5, 77239
151
+ 6, 77239
152
+ 7, 77239
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep2/run.env ADDED
@@ -0,0 +1,105 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=moonep
2
+ config=DEEPSEEK_V3_L4
3
+ family=tierB
4
+ source_model=DeepSeek-V3
5
+ standard_num_layers=61
6
+ base_rung=
7
+ repo=/workspace/Megatron-MoE-Echo
8
+ branch=moonep_backend_async_grad_reduce
9
+ commit=2e68c8ad
10
+ moonep_source_dir=/workspace/MoonEP
11
+ moonep_source_branch=megatron-ddp-symmetric-params
12
+ moonep_source_commit=d8bc666
13
+ num_layers=4
14
+ hidden=7168
15
+ ffn_hidden=18432
16
+ moe_ffn_hidden=2048
17
+ num_experts=256
18
+ topk=8
19
+ experts_per_rank=32
20
+ expert_params=44040192
21
+ heads=128
22
+ kv_groups=128
23
+ kv_channels=128
24
+ qk_layernorm=1
25
+ shared_ffn=2048
26
+ shared_expert_overlap=0
27
+ moe_token_dispatcher=flex
28
+ mla=1
29
+ q_lora_rank=1536
30
+ kv_lora_rank=512
31
+ qk_head_dim=128
32
+ qk_pos_emb_head_dim=64
33
+ v_head_dim=128
34
+ dense_first_layers=3
35
+ moe_layers=1
36
+ rotary_base=10000
37
+ rotary_percent=1.0
38
+ rotary_scaling=40
39
+ mscale=1.0
40
+ mscale_all_dim=1.0
41
+ qkv_bias=0
42
+ router_balancing=seq_aux_loss
43
+ aux_loss_coeff=1e-4
44
+ router_pre_softmax=0
45
+ router_fusion=0
46
+ router_dtype=fp32
47
+ router_groups=8
48
+ router_group_topk=4
49
+ router_score_function=sigmoid
50
+ router_expert_bias=1
51
+ router_bias_update_rate=1e-3
52
+ router_topk_scaling_factor=2.5
53
+ shared_expert_gate=0
54
+ apply_layernorm_1p=0
55
+ attention_output_gate=0
56
+ linear_attention_type=
57
+ linear_attention_freq=
58
+ linear_conv_kernel_dim=
59
+ linear_key_head_dim=
60
+ linear_value_head_dim=
61
+ linear_num_key_heads=
62
+ linear_num_value_heads=
63
+ mtp_num_layers=0
64
+ mtp_loss_scaling_factor=0.1
65
+ no_weight_decay_cond_type=
66
+ max_position_embeddings=4096
67
+ init_method_std=0.02
68
+ device_initiated_grouped_gemm=1
69
+ norm_epsilon=1e-6
70
+ seq_length=4096
71
+ micro_batch_size=1
72
+ global_batch_size=8
73
+ num_microbatches=1
74
+ assignments_per_rank=32768
75
+ mean_tokens_per_expert=1024
76
+ slot_policy=matched
77
+ slots_per_rank=32
78
+ capacity_mode=none
79
+ capacity=none
80
+ cuda_graph=1
81
+ probe_mode=0
82
+ async_grad_reduce=1
83
+ branch_role=timing
84
+ memory_dump_mode=none
85
+ memory_snapshot_path=
86
+ cuda_graph_warmup_steps=3
87
+ recompute=none
88
+ force_balance=0
89
+ train_iters=12
90
+ ep_size=8
91
+ echo_dispatch_overlap=1
92
+ echo_dispatch_recompute=0
93
+ ultraep_quota_min_tokens=
94
+ ultraep_weight_sync_join=
95
+ profile_enable=0
96
+ kernel_trace=0
97
+ fp8_mode=mxfp8
98
+ fp8_recipe=mxfp8
99
+ training_seed=1234
100
+ learning_rate=1e-4
101
+ min_learning_rate=1.0e-5
102
+ lr_warmup_iters=5
103
+ state_dump_enable=0
104
+ state_dump_dir=
105
+ moonep_assert_no_pending_reduce=1
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep2/train.log ADDED
@@ -0,0 +1,840 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [07:01:46] WARN: moonep_backend_async_grad_reduce lacks --moe-router-fusion; DEEPSEEK_V3_L4 runs unfused (timing only, routing identical)
2
+ ==============================================================
3
+ DEEPSEEK_V3_L4 / moonep (slots/rank=32 policy=matched)
4
+ E=256 K=8 H'=2048 m=32
5
+ L=4 seq=4096 gbs=8 (1 microbatches)
6
+ cap_mode=none cap=none cg=1 recompute=none
7
+ memory_dump=none snapshots=/workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep2/memory_snapshots/snapshot.pickle
8
+ repo=/workspace/Megatron-MoE-Echo @ 2e68c8ad
9
+ ==============================================================
10
+ W0921 07:01:48.068000 362905 torch/distributed/run.py:766]
11
+ W0921 07:01:48.068000 362905 torch/distributed/run.py:766] *****************************************
12
+ W0921 07:01:48.068000 362905 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
13
+ W0921 07:01:48.068000 362905 torch/distributed/run.py:766] *****************************************
14
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
15
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
16
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
17
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
18
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
19
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
20
+ /workspace/Megatron-MoE-Echo/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
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
23
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
24
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
25
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
26
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
27
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
28
+ 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
29
+ 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
30
+ WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer
31
+ Number of virtual stages per pipeline stage: None
32
+ accumulate and all-reduce gradients in fp32 for bfloat16 data type.
33
+ using torch.bfloat16 for parameters ...
34
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
35
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
36
+ ------------------------ arguments ------------------------
37
+ account_for_embedding_in_pipeline_split ......... False
38
+ account_for_loss_in_pipeline_split .............. False
39
+ accumulate_allreduce_grads_in_fp32 .............. True
40
+ adam_beta1 ...................................... 0.9
41
+ adam_beta2 ...................................... 0.999
42
+ adam_eps ........................................ 1e-08
43
+ add_bias_linear ................................. False
44
+ add_position_embedding .......................... False
45
+ add_qkv_bias .................................... False
46
+ adlr_autoresume ................................. False
47
+ adlr_autoresume_interval ........................ 1000
48
+ align_grad_reduce ............................... True
49
+ align_param_gather .............................. False
50
+ app_tag_run_name ................................ None
51
+ app_tag_run_version ............................. 0.0.0
52
+ apply_layernorm_1p .............................. False
53
+ apply_query_key_layer_scaling ................... False
54
+ apply_residual_connection_post_layernorm ........ False
55
+ apply_rope_fusion ............................... False
56
+ async_save ...................................... None
57
+ async_tensor_model_parallel_allreduce ........... True
58
+ attention_backend ............................... AttnBackend.auto
59
+ attention_dropout ............................... 0.0
60
+ attention_softmax_in_fp32 ....................... False
61
+ auto_detect_ckpt_format ......................... False
62
+ barrier_with_L1_time ............................ True
63
+ bert_binary_head ................................ True
64
+ bert_embedder_type .............................. megatron
65
+ bert_load ....................................... None
66
+ bf16 ............................................ True
67
+ bias_dropout_fusion ............................. True
68
+ bias_gelu_fusion ................................ False
69
+ bias_swiglu_fusion .............................. True
70
+ biencoder_projection_dim ........................ 0
71
+ biencoder_shared_query_context_model ............ False
72
+ block_data_path ................................. None
73
+ calc_ft_timeouts ................................ False
74
+ calculate_per_token_loss ........................ False
75
+ check_for_large_grads ........................... False
76
+ check_for_nan_in_loss_and_grad .................. False
77
+ check_for_spiky_loss ............................ False
78
+ check_weight_hash_across_dp_replicas_interval ... None
79
+ ckpt_assume_constant_structure .................. False
80
+ ckpt_convert_format ............................. None
81
+ ckpt_convert_save ............................... None
82
+ ckpt_convert_update_legacy_dist_opt_format ...... False
83
+ ckpt_format ..................................... torch_dist
84
+ ckpt_fully_parallel_load ........................ False
85
+ ckpt_fully_parallel_save ........................ True
86
+ ckpt_fully_parallel_save_deprecated ............. False
87
+ ckpt_step ....................................... None
88
+ classes_fraction ................................ 1.0
89
+ clip_grad ....................................... 1.0
90
+ clone_scatter_output_in_embedding ............... True
91
+ config_logger_dir ...............................
92
+ consumed_train_samples .......................... 0
93
+ consumed_valid_samples .......................... 0
94
+ context_parallel_size ........................... 1
95
+ cp_comm_type .................................... ['p2p']
96
+ create_attention_mask_in_dataloader ............. True
97
+ cross_entropy_fusion_impl ....................... native
98
+ cross_entropy_loss_fusion ....................... False
99
+ cuda_graph_scope ................................ full_iteration
100
+ cuda_graph_warmup_steps ......................... 3
101
+ data_args_path .................................. None
102
+ data_cache_path ................................. None
103
+ data_parallel_random_init ....................... False
104
+ data_parallel_sharding_strategy ................. no_shard
105
+ data_parallel_size .............................. 8
106
+ data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document']
107
+ data_per_class_fraction ......................... 1.0
108
+ data_sharding ................................... True
109
+ dataloader_type ................................. single
110
+ ddp_average_in_collective ....................... False
111
+ ddp_bucket_size ................................. None
112
+ ddp_num_buckets ................................. None
113
+ ddp_pad_buckets_for_high_nccl_busbw ............. False
114
+ decoder_first_pipeline_num_layers ............... None
115
+ decoder_last_pipeline_num_layers ................ None
116
+ decoder_num_layers .............................. None
117
+ decoder_seq_length .............................. None
118
+ decoupled_lr .................................... None
119
+ decoupled_min_lr ................................ None
120
+ decrease_batch_size_if_needed ................... False
121
+ defer_embedding_wgrad_compute ................... False
122
+ delay_wgrad_compute ............................. False
123
+ deprecated_use_mcore_models ..................... True
124
+ deterministic_mode .............................. False
125
+ dino_bottleneck_size ............................ 256
126
+ dino_freeze_last_layer .......................... 1
127
+ dino_head_hidden_size ........................... 2048
128
+ dino_local_crops_number ......................... 10
129
+ dino_local_img_size ............................. 96
130
+ dino_norm_last_layer ............................ False
131
+ dino_teacher_temp ............................... 0.07
132
+ dino_warmup_teacher_temp ........................ 0.04
133
+ dino_warmup_teacher_temp_epochs ................. 30
134
+ disable_bf16_reduced_precision_matmul ........... False
135
+ disable_mamba_mem_eff_path ...................... False
136
+ disable_straggler_on_startup .................... False
137
+ dist_ckpt_format_deprecated ..................... None
138
+ dist_ckpt_strictness ............................ assume_ok_unexpected
139
+ distribute_saved_activations .................... False
140
+ distributed_backend ............................. nccl
141
+ distributed_timeout_minutes ..................... 10
142
+ embedding_init_method_std ....................... None
143
+ embedding_path .................................. None
144
+ empty_unused_memory_level ....................... 0
145
+ enable_cuda_graph ............................... True
146
+ enable_experimental ............................. False
147
+ enable_ft_package ............................... False
148
+ enable_gloo_process_groups ...................... True
149
+ enable_msc ...................................... True
150
+ enable_one_logger ............................... True
151
+ encoder_num_layers .............................. 4
152
+ encoder_seq_length .............................. 4096
153
+ end_weight_decay ................................ 0.1
154
+ eod_mask_loss ................................... False
155
+ error_injection_rate ............................ 0
156
+ error_injection_type ............................ transient_error
157
+ eval_interval ................................... 1000000
158
+ eval_iters ...................................... 0
159
+ evidence_data_path .............................. None
160
+ exit_duration_in_mins ........................... None
161
+ exit_interval ................................... None
162
+ exit_on_missing_checkpoint ...................... False
163
+ exit_signal_handler ............................. False
164
+ exp_avg_dtype ................................... torch.float32
165
+ exp_avg_sq_dtype ................................ torch.float32
166
+ expert_model_parallel_size ...................... 8
167
+ expert_tensor_parallel_size ..................... 1
168
+ export_force_local_attention .................... False
169
+ export_kd_cfg ................................... None
170
+ export_kd_teacher_ckpt_format ................... None
171
+ export_kd_teacher_load .......................... None
172
+ export_kv_cache_quant ........................... False
173
+ export_legacy_megatron .......................... False
174
+ export_model_type ............................... GPTModel
175
+ export_moe_apply_probs_on_input ................. False
176
+ export_qk_l2_norm ............................... False
177
+ export_quant_cfg ................................ None
178
+ export_real_quant_cfg ........................... None
179
+ export_te_mcore_model ........................... False
180
+ external_cuda_graph ............................. False
181
+ ffn_hidden_size ................................. 18432
182
+ finetune ........................................ False
183
+ finetune_data_split ............................. train
184
+ finetune_hf_dataset ............................. None
185
+ first_last_layers_bf16 .......................... False
186
+ flash_decode .................................... False
187
+ fp16 ............................................ False
188
+ fp16_lm_cross_entropy ........................... False
189
+ fp32_residual_connection ........................ False
190
+ fp8 ............................................. e4m3
191
+ fp8_amax_compute_algo ........................... most_recent
192
+ fp8_amax_history_len ............................ 1
193
+ fp8_interval .................................... 1
194
+ fp8_margin ...................................... 0
195
+ fp8_param_gather ................................ False
196
+ fp8_recipe ...................................... mxfp8
197
+ fp8_wgrad ....................................... True
198
+ fsdp_double_buffer .............................. False
199
+ global_batch_size ............................... 8
200
+ grad_reduce_in_bf16 ............................. False
201
+ gradient_accumulation_fusion .................... True
202
+ gradient_reduce_div_fusion ...................... True
203
+ group_query_attention ........................... False
204
+ head_lr_mult .................................... 1.0
205
+ heterogeneous_layers_config_encoded_json ........ None
206
+ heterogeneous_layers_config_path ................ None
207
+ hidden_dropout .................................. 0.0
208
+ hidden_size ..................................... 7168
209
+ hierarchical_context_parallel_sizes ............. None
210
+ high_priority_stream_groups ..................... []
211
+ hybrid_attention_ratio .......................... 0.0
212
+ hybrid_mlp_ratio ................................ 0.0
213
+ hybrid_override_pattern ......................... None
214
+ hysteresis ...................................... 2
215
+ ict_head_size ................................... None
216
+ ict_load ........................................ None
217
+ img_h ........................................... 224
218
+ img_w ........................................... 224
219
+ indexer_batch_size .............................. 128
220
+ indexer_log_interval ............................ 1000
221
+ inference_batch_times_seqlen_threshold .......... -1
222
+ inference_dynamic_batching ...................... False
223
+ inference_dynamic_batching_buffer_guaranteed_fraction 0.2
224
+ inference_dynamic_batching_buffer_overflow_factor None
225
+ inference_dynamic_batching_buffer_size_gb ....... 40.0
226
+ inference_dynamic_batching_chunk_size ........... 256
227
+ inference_dynamic_batching_max_requests_override None
228
+ inference_dynamic_batching_max_tokens_override .. None
229
+ inference_dynamic_batching_num_cuda_graphs ...... 16
230
+ inference_max_batch_size ........................ 8
231
+ inference_max_seq_length ........................ 2560
232
+ inference_rng_tracker ........................... False
233
+ init_method_std ................................. 0.02
234
+ init_method_xavier_uniform ...................... False
235
+ init_model_with_meta_device ..................... False
236
+ initial_loss_scale .............................. 4294967296
237
+ inprocess_active_world_size ..................... 8
238
+ inprocess_barrier_timeout ....................... 120
239
+ inprocess_completion_timeout .................... 120
240
+ inprocess_empty_cuda_cache ...................... False
241
+ inprocess_granularity ........................... node
242
+ inprocess_hard_timeout .......................... 90
243
+ inprocess_heartbeat_interval .................... 30
244
+ inprocess_heartbeat_timeout ..................... 60
245
+ inprocess_last_call_wait ........................ 1
246
+ inprocess_max_iterations ........................ None
247
+ inprocess_monitor_process_interval .............. 1.0
248
+ inprocess_monitor_thread_interval ............... 1.0
249
+ inprocess_progress_watchdog_interval ............ 1.0
250
+ inprocess_restart ............................... False
251
+ inprocess_soft_timeout .......................... 60
252
+ inprocess_termination_grace_time ................ 1
253
+ is_hybrid_model ................................. False
254
+ iter_per_epoch .................................. 1250
255
+ iterations_to_skip .............................. []
256
+ keep_fp8_transpose_cache_when_using_custom_fsdp . False
257
+ kitchen_config_file ............................. None
258
+ kitchen_recipe_number ........................... None
259
+ kv_channels ..................................... 128
260
+ kv_lora_rank .................................... 512
261
+ lazy_mpu_init ................................... None
262
+ load ............................................ None
263
+ load_main_params_from_ckpt ...................... None
264
+ load_model_opt_format ........................... False
265
+ local_rank ...................................... 0
266
+ log_energy ...................................... False
267
+ log_interval .................................... 1
268
+ log_loss_scale_to_tensorboard ................... True
269
+ log_memory_to_tensorboard ....................... False
270
+ log_num_zeros_in_grad ........................... False
271
+ log_params_norm ................................. False
272
+ log_progress .................................... False
273
+ log_straggler ................................... False
274
+ log_throughput .................................. True
275
+ log_timers_to_tensorboard ....................... False
276
+ log_validation_ppl_to_tensorboard ............... False
277
+ log_world_size_to_tensorboard ................... False
278
+ logging_level ................................... None
279
+ loss_scale ...................................... None
280
+ loss_scale_window ............................... 1000
281
+ lr .............................................. 0.0001
282
+ lr_decay_iters .................................. 8
283
+ lr_decay_samples ................................ None
284
+ lr_decay_style .................................. cosine
285
+ lr_warmup_fraction .............................. None
286
+ lr_warmup_init .................................. 0.0
287
+ lr_warmup_iters ................................. 5
288
+ lr_warmup_samples ............................... 0
289
+ lr_wsd_decay_iters .............................. None
290
+ lr_wsd_decay_samples ............................ None
291
+ lr_wsd_decay_style .............................. exponential
292
+ main_grads_dtype ................................ torch.float32
293
+ main_params_dtype ............................... torch.float32
294
+ make_vocab_size_divisible_by .................... 128
295
+ mamba_head_dim .................................. 64
296
+ mamba_num_groups ................................ 8
297
+ mamba_num_heads ................................. None
298
+ mamba_state_dim ................................. 128
299
+ manual_gc ....................................... False
300
+ manual_gc_eval .................................. True
301
+ manual_gc_interval .............................. 0
302
+ mask_factor ..................................... 1.0
303
+ mask_prob ....................................... 0.15
304
+ mask_type ....................................... random
305
+ masked_softmax_fusion ........................... False
306
+ max_position_embeddings ......................... 4096
307
+ max_tokens_to_oom ............................... 12000
308
+ memory_snapshot_path ............................ snapshot.pickle
309
+ merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt
310
+ micro_batch_size ................................ 1
311
+ microbatch_group_size_per_vp_stage .............. None
312
+ mid_level_dataset_surplus ....................... 0.005
313
+ min_loss_scale .................................. 1.0
314
+ min_lr .......................................... 1e-05
315
+ mlp_chunks_for_prefill .......................... 1
316
+ mmap_bin_files .................................. True
317
+ mock_data ....................................... False
318
+ moe_apply_probs_on_input ........................ False
319
+ moe_aux_loss_coeff .............................. 0.0001
320
+ moe_deepep_num_sms .............................. 20
321
+ moe_echo_enable_random_offloading ............... False
322
+ moe_echo_expert_dispatch_overlap ................ False
323
+ moe_echo_expert_dispatcher_type ................. hybridep
324
+ moe_echo_recompute_expert_dispatch .............. False
325
+ moe_enable_deepep ............................... False
326
+ moe_enable_echo ................................. False
327
+ moe_enable_hybridep ............................. False
328
+ moe_enable_moonep ............................... True
329
+ moe_expert_capacity_factor ...................... None
330
+ moe_extended_tp ................................. False
331
+ moe_ffn_hidden_size ............................. 2048
332
+ moe_grouped_gemm ................................ True
333
+ moe_hybridep_num_sms ............................ 16
334
+ moe_input_jitter_eps ............................ None
335
+ moe_layer_freq .................................. [0, 0, 0, 1]
336
+ moe_layer_recompute ............................. False
337
+ moe_moonep_async_grad_reduce .................... True
338
+ moe_moonep_num_sms .............................. None
339
+ moe_num_echo_experts ............................ None
340
+ moe_pad_expert_input_to_capacity ................ False
341
+ moe_per_layer_logging ........................... False
342
+ moe_permute_fusion .............................. False
343
+ moe_received_token_capacity ..................... None
344
+ moe_router_bias_update_rate ..................... 0.001
345
+ moe_router_dtype ................................ fp32
346
+ moe_router_enable_expert_bias ................... True
347
+ moe_router_force_load_balancing ................. False
348
+ moe_router_group_topk ........................... 4
349
+ moe_router_load_balancing_type .................. seq_aux_loss
350
+ moe_router_num_groups ........................... 8
351
+ moe_router_padding_for_fp8 ...................... False
352
+ moe_router_pre_softmax .......................... False
353
+ moe_router_score_function ....................... sigmoid
354
+ moe_router_topk ................................. 8
355
+ moe_router_topk_scaling_factor .................. 2.5
356
+ moe_shared_expert_intermediate_size ............. 2048
357
+ moe_shared_expert_overlap ....................... False
358
+ moe_token_dispatcher_type ....................... flex
359
+ moe_token_drop_policy ........................... probs
360
+ moe_upcycling_granularity ....................... 1
361
+ moe_use_device_initiated_grouped_gemm ........... True
362
+ moe_use_legacy_grouped_gemm ..................... False
363
+ moe_use_upcycling ............................... False
364
+ moe_z_loss_coeff ................................ None
365
+ mrope_section ................................... None
366
+ mscale .......................................... 1.0
367
+ mscale_all_dim .................................. 1.0
368
+ mtp_loss_scaling_factor ......................... 0.1
369
+ mtp_num_layers .................................. None
370
+ multi_latent_attention .......................... True
371
+ nccl_all_reduce_for_prefill ..................... False
372
+ nccl_communicator_config_path ................... None
373
+ nccl_ub ......................................... False
374
+ no_load_optim ................................... None
375
+ no_load_rng ..................................... None
376
+ no_persist_layer_norm ........................... False
377
+ no_rope_freq .................................... None
378
+ no_save_optim ................................... None
379
+ no_save_rng ..................................... None
380
+ non_persistent_ckpt_type ........................ None
381
+ non_persistent_global_ckpt_dir .................. None
382
+ non_persistent_local_ckpt_algo .................. fully_parallel
383
+ non_persistent_local_ckpt_dir ................... None
384
+ non_persistent_save_interval .................... None
385
+ norm_epsilon .................................... 1e-06
386
+ normalization ................................... RMSNorm
387
+ num_attention_heads ............................. 128
388
+ num_channels .................................... 3
389
+ num_classes ..................................... 1000
390
+ num_dataset_builder_threads ..................... 1
391
+ num_distributed_optimizer_instances ............. 1
392
+ num_experts ..................................... 256
393
+ num_layers ...................................... 4
394
+ num_layers_at_end_in_bf16 ....................... 1
395
+ num_layers_at_start_in_bf16 ..................... 1
396
+ num_layers_per_virtual_pipeline_stage ........... None
397
+ num_query_groups ................................ 1
398
+ num_virtual_stages_per_pipeline_rank ............ None
399
+ num_workers ..................................... 2
400
+ object_storage_cache_path ....................... None
401
+ one_logger_async ................................ False
402
+ one_logger_project .............................. megatron-lm
403
+ one_logger_run_name ............................. None
404
+ onnx_safe ....................................... None
405
+ openai_gelu ..................................... False
406
+ optimizer ....................................... adam
407
+ optimizer_cpu_offload ........................... False
408
+ optimizer_offload_fraction ...................... 1.0
409
+ output_bert_embeddings .......................... False
410
+ overlap_cpu_optimizer_d2h_h2d ................... False
411
+ overlap_grad_reduce ............................. True
412
+ overlap_moe_expert_parallel_comm ................ False
413
+ overlap_p2p_comm ................................ False
414
+ overlap_p2p_comm_warmup_flush ................... False
415
+ overlap_param_gather ............................ True
416
+ overlap_param_gather_with_optimizer_step ........ False
417
+ override_opt_param_scheduler .................... False
418
+ params_dtype .................................... torch.bfloat16
419
+ patch_dim ....................................... 16
420
+ per_split_data_args_path ........................ None
421
+ perform_initialization .......................... True
422
+ pin_cpu_grads ................................... True
423
+ pin_cpu_params .................................. True
424
+ pipeline_model_parallel_comm_backend ............ None
425
+ pipeline_model_parallel_layout .................. None
426
+ pipeline_model_parallel_size .................... 1
427
+ position_embedding_type ......................... rope
428
+ pretrained_checkpoint ........................... None
429
+ profile ......................................... False
430
+ profile_ranks ................................... [0]
431
+ profile_step_end ................................ 12
432
+ profile_step_start .............................. 10
433
+ q_lora_rank ..................................... 1536
434
+ qk_head_dim ..................................... 128
435
+ qk_l2_norm ...................................... False
436
+ qk_layernorm .................................... True
437
+ qk_pos_emb_head_dim ............................. 64
438
+ query_in_block_prob ............................. 0.1
439
+ rampup_batch_size ............................... None
440
+ rank ............................................ 0
441
+ recompute_granularity ........................... None
442
+ recompute_method ................................ None
443
+ recompute_modules ............................... None
444
+ recompute_num_layers ............................ None
445
+ record_memory_history ........................... False
446
+ relative_attention_max_distance ................. 128
447
+ relative_attention_num_buckets .................. 32
448
+ replication ..................................... False
449
+ replication_factor .............................. 2
450
+ replication_jump ................................ None
451
+ rerun_mode ...................................... validate_results
452
+ reset_attention_mask ............................ False
453
+ reset_position_ids .............................. False
454
+ result_rejected_tracker_filename ................ None
455
+ retriever_report_topk_accuracies ................ []
456
+ retriever_score_scaling ......................... False
457
+ retriever_seq_length ............................ 256
458
+ retro_add_retriever ............................. False
459
+ retro_attention_gate ............................ 1
460
+ retro_cyclic_train_iters ........................ None
461
+ retro_encoder_attention_dropout ................. 0.1
462
+ retro_encoder_hidden_dropout .................... 0.1
463
+ retro_encoder_layers ............................ 2
464
+ retro_num_neighbors ............................. 2
465
+ retro_num_retrieved_chunks ...................... 2
466
+ retro_project_dir ............................... None
467
+ retro_verify_neighbor_count ..................... True
468
+ reuse_grad_buf_for_mxfp8_param_ag ............... False
469
+ rope_scaling_factor ............................. 8.0
470
+ rope_type ....................................... None
471
+ rotary_base ..................................... 10000
472
+ rotary_interleaved .............................. False
473
+ rotary_percent .................................. 1.0
474
+ rotary_scaling_factor ........................... 40.0
475
+ rotary_seq_len_interpolation_factor ............. None
476
+ run_workload_inspector_server ................... False
477
+ sample_rate ..................................... 1.0
478
+ save ............................................ None
479
+ save_interval ................................... None
480
+ save_retain_interval ............................ None
481
+ scatter_gather_tensors_in_pipeline .............. True
482
+ seed ............................................ 1234
483
+ seq_length ...................................... 4096
484
+ sequence_parallel ............................... False
485
+ sft ............................................. False
486
+ sft_tokenizer_prompt_format ..................... nemotron-h-aligned
487
+ sgd_momentum .................................... 0.9
488
+ short_seq_prob .................................. 0.1
489
+ skip_train ...................................... False
490
+ skipped_train_samples ........................... 0
491
+ spec ............................................ None
492
+ split ........................................... 100,0,0
493
+ squared_relu .................................... False
494
+ start_weight_decay .............................. 0.1
495
+ straggler_ctrlr_port ............................ 65535
496
+ straggler_minmax_count .......................... 1
497
+ suggested_communication_unit_size ............... None
498
+ swiglu .......................................... True
499
+ swin_backbone_type .............................. tiny
500
+ symmetric_ar_type ............................... None
501
+ te_rng_tracker .................................. True
502
+ tensor_model_parallel_size ...................... 1
503
+ tensorboard_dir ................................. None
504
+ tensorboard_log_interval ........................ 1
505
+ tensorboard_queue_size .......................... 1000
506
+ test_data_path .................................. None
507
+ test_mode ....................................... False
508
+ tiktoken_num_special_tokens ..................... 1000
509
+ tiktoken_pattern ................................ None
510
+ tiktoken_special_tokens ......................... None
511
+ timing_log_level ................................ 0
512
+ timing_log_option ............................... minmax
513
+ titles_data_path ................................ None
514
+ tokenizer_model ................................. None
515
+ tokenizer_type .................................. GPT2BPETokenizer
516
+ torch_fsdp2_reshard_after_forward ............... True
517
+ tp_comm_bootstrap_backend ....................... nccl
518
+ tp_comm_bulk_dgrad .............................. True
519
+ tp_comm_bulk_wgrad .............................. True
520
+ tp_comm_overlap ................................. False
521
+ tp_comm_overlap_ag .............................. True
522
+ tp_comm_overlap_cfg ............................. None
523
+ tp_comm_overlap_rs .............................. True
524
+ tp_comm_overlap_rs_dgrad ........................ False
525
+ tp_comm_split_ag ................................ True
526
+ tp_comm_split_rs ................................ True
527
+ train_data_path ................................. None
528
+ train_iters ..................................... 12
529
+ train_samples ................................... None
530
+ train_sync_interval ............................. None
531
+ transformer_impl ................................ transformer_engine
532
+ transformer_pipeline_model_parallel_size ........ 1
533
+ untie_embeddings_and_output_weights ............. True
534
+ use_checkpoint_args ............................. False
535
+ use_checkpoint_opt_param_scheduler .............. False
536
+ use_cpu_initialization .......................... None
537
+ use_custom_fsdp ................................. False
538
+ use_dist_ckpt ................................... True
539
+ use_dist_ckpt_deprecated ........................ False
540
+ use_distributed_optimizer ....................... True
541
+ use_flash_attn .................................. False
542
+ use_legacy_models ............................... False
543
+ use_mp_args_from_checkpoint_args ................ False
544
+ use_one_sent_docs ............................... False
545
+ use_persistent_ckpt_worker ...................... False
546
+ use_precision_aware_optimizer ................... False
547
+ use_pytorch_profiler ............................ False
548
+ use_ring_exchange_p2p ........................... False
549
+ use_rope_scaling ................................ False
550
+ use_rotary_position_embeddings .................. False
551
+ use_sharp ....................................... False
552
+ use_tokenizer_model_from_checkpoint_args ........ True
553
+ use_torch_fsdp2 ................................. False
554
+ use_torch_optimizer_for_cpu_offload ............. False
555
+ use_tp_pp_dp_mapping ............................ False
556
+ v_head_dim ...................................... 128
557
+ valid_data_path ................................. None
558
+ variable_seq_lengths ............................ False
559
+ virtual_pipeline_model_parallel_size ............ None
560
+ vision_backbone_type ............................ vit
561
+ vision_pretraining .............................. False
562
+ vision_pretraining_type ......................... classify
563
+ vocab_extra_ids ................................. 0
564
+ vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json
565
+ vocab_size ...................................... None
566
+ wandb_exp_name ..................................
567
+ wandb_project ...................................
568
+ wandb_save_dir ..................................
569
+ weight_decay .................................... 0.1
570
+ weight_decay_incr_style ......................... constant
571
+ wgrad_deferral_limit ............................ 0
572
+ world_size ...................................... 8
573
+ yaml_cfg ........................................ None
574
+ -------------------- end of arguments ---------------------
575
+ INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1
576
+ > building GPT2BPETokenizer tokenizer ...
577
+ > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304)
578
+ WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS
579
+ > initializing torch distributed ...
580
+ [Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
581
+ [Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
582
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
583
+ [Gloo] Rank [Gloo] Rank 3 is connected to 27 is connected to peer ranks. 7Expected number of connected peer ranks is : peer ranks. 7Expected number of connected peer ranks is : 7
584
+
585
+ [Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : [Gloo] Rank 76
586
+ 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 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
589
+ [Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
590
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
591
+ [Gloo] Rank [Gloo] Rank 4 is connected to 37 is connected to peer ranks. 7Expected number of connected peer ranks is : peer ranks. 7Expected number of connected peer ranks is :
592
+ 7
593
+ [Gloo] Rank [Gloo] Rank 65 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
594
+
595
+ [Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
596
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
597
+ > initialized tensor model parallel with size 1
598
+ > initialized pipeline model parallel with size 1
599
+ > setting random seeds to 1234 ...
600
+ [Gloo] Rank [Gloo] Rank 0[Gloo] Rank 0 is connected to 0 is connected to 0 is connected to 0 peer ranks. 0 peer ranks. Expected number of connected peer ranks is : peer ranks. Expected number of connected peer ranks is : 0Expected number of connected peer ranks is : 0
601
+
602
+ 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
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
607
+ [rank7]:[W921 07:02:05.322292508 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.
608
+ [rank4]:[W921 07:02:05.322294433 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.
609
+ [rank5]:[W921 07:02:05.323023122 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]:[W921 07:02:05.330288127 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]:[W921 07:02:05.330312717 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
612
+ [rank2]:[W921 07:02:05.330376113 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
+ [rank6]:[W921 07:02:05.330424754 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.
614
+ > compiling dataset index builder ...
615
+ make: Entering directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
616
+ make: Nothing to be done for 'default'.
617
+ make: Leaving directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
618
+ >>> done with dataset index builder. Compilation time: 0.035 seconds
619
+ WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations.
620
+ > compiling and loading fused kernels ...
621
+ [rank0]:[W921 07:02:05.491284272 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.
622
+ >>> done with compiling and loading fused kernels. Compilation time: 3.845 seconds
623
+ time to initialize megatron (seconds): 7.461
624
+ [after megatron is initialized] datetime: 2026-09-21 07:02:11
625
+ building GPT model ...
626
+ /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.
627
+ warnings.warn(
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
+ > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 4113898496
643
+ 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)
644
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 17
645
+ Params for bucket 1 (360579072 elements, 360579072 padded size):
646
+ module.output_layer.weight
647
+ Params for bucket 2 (44047360 elements, 44047360 padded size):
648
+ module.decoder.final_layernorm.weight
649
+ module.decoder.layers.3.mlp.shared_experts.linear_fc2.weight
650
+ module.decoder.layers.3.mlp.shared_experts.linear_fc1.weight
651
+ Params for bucket 3 (60497408 elements, 60497408 padded size):
652
+ module.decoder.layers.3.mlp.router.weight
653
+ module.decoder.layers.3.self_attention.linear_kv_down_proj.weight
654
+ module.decoder.layers.3.self_attention.linear_kv_up_proj.layer_norm_weight
655
+ module.decoder.layers.3.self_attention.linear_q_up_proj.weight
656
+ module.decoder.layers.3.pre_mlp_layernorm.weight
657
+ module.decoder.layers.3.self_attention.linear_kv_up_proj.weight
658
+ Params for bucket 4 (128452096 elements, 128452096 padded size):
659
+ module.decoder.layers.3.self_attention.linear_q_up_proj.layer_norm_weight
660
+ module.decoder.layers.3.self_attention.linear_proj.weight
661
+ module.decoder.layers.3.self_attention.linear_q_down_proj.weight
662
+ Params for bucket 5 (132127744 elements, 132127744 padded size):
663
+ module.decoder.layers.3.input_layernorm.weight
664
+ module.decoder.layers.2.mlp.linear_fc2.weight
665
+ Params for bucket 6 (264241152 elements, 264241152 padded size):
666
+ module.decoder.layers.2.mlp.linear_fc1.weight
667
+ Params for bucket 7 (58662400 elements, 58662400 padded size):
668
+ module.decoder.layers.2.mlp.linear_fc1.layer_norm_weight
669
+ module.decoder.layers.2.self_attention.linear_kv_up_proj.layer_norm_weight
670
+ module.decoder.layers.2.self_attention.linear_kv_up_proj.weight
671
+ module.decoder.layers.2.self_attention.linear_kv_down_proj.weight
672
+ module.decoder.layers.2.self_attention.linear_q_up_proj.weight
673
+ Params for bucket 8 (128452096 elements, 128452096 padded size):
674
+ module.decoder.layers.2.self_attention.linear_q_up_proj.layer_norm_weight
675
+ module.decoder.layers.2.self_attention.linear_q_down_proj.weight
676
+ module.decoder.layers.2.self_attention.linear_proj.weight
677
+ Params for bucket 9 (132127744 elements, 132127744 padded size):
678
+ module.decoder.layers.2.input_layernorm.weight
679
+ module.decoder.layers.1.mlp.linear_fc2.weight
680
+ Params for bucket 10 (264241152 elements, 264241152 padded size):
681
+ module.decoder.layers.1.mlp.linear_fc1.weight
682
+ Params for bucket 11 (58662400 elements, 58662400 padded size):
683
+ module.decoder.layers.1.mlp.linear_fc1.layer_norm_weight
684
+ module.decoder.layers.1.self_attention.linear_q_up_proj.weight
685
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.layer_norm_weight
686
+ module.decoder.layers.1.self_attention.linear_kv_down_proj.weight
687
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.weight
688
+ Params for bucket 12 (128452096 elements, 128452096 padded size):
689
+ module.decoder.layers.1.self_attention.linear_q_up_proj.layer_norm_weight
690
+ module.decoder.layers.1.self_attention.linear_proj.weight
691
+ module.decoder.layers.1.self_attention.linear_q_down_proj.weight
692
+ Params for bucket 13 (132127744 elements, 132127744 padded size):
693
+ module.decoder.layers.1.input_layernorm.weight
694
+ module.decoder.layers.0.mlp.linear_fc2.weight
695
+ Params for bucket 14 (264241152 elements, 264241152 padded size):
696
+ module.decoder.layers.0.mlp.linear_fc1.weight
697
+ Params for bucket 15 (58662400 elements, 58662400 padded size):
698
+ module.decoder.layers.0.mlp.linear_fc1.layer_norm_weight
699
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.layer_norm_weight
700
+ module.decoder.layers.0.self_attention.linear_q_up_proj.weight
701
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.weight
702
+ module.decoder.layers.0.self_attention.linear_kv_down_proj.weight
703
+ Params for bucket 16 (128452096 elements, 128452096 padded size):
704
+ module.decoder.layers.0.self_attention.linear_q_up_proj.layer_norm_weight
705
+ module.decoder.layers.0.self_attention.linear_proj.weight
706
+ module.decoder.layers.0.self_attention.linear_q_down_proj.weight
707
+ Params for bucket 17 (360586240 elements, 360586240 padded size):
708
+ module.decoder.layers.0.input_layernorm.weight
709
+ module.embedding.word_embeddings.weight
710
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 27
711
+ Params for bucket 1 (44040192 elements, 44040192 padded size):
712
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight0
713
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight1
714
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight2
715
+ Params for bucket 2 (44040192 elements, 44040192 padded size):
716
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight3
717
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight4
718
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight5
719
+ Params for bucket 3 (44040192 elements, 44040192 padded size):
720
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight6
721
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight7
722
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight8
723
+ Params for bucket 4 (44040192 elements, 44040192 padded size):
724
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight9
725
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight10
726
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight11
727
+ Params for bucket 5 (44040192 elements, 44040192 padded size):
728
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight12
729
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight13
730
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight14
731
+ Params for bucket 6 (44040192 elements, 44040192 padded size):
732
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight15
733
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight16
734
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight17
735
+ Params for bucket 7 (44040192 elements, 44040192 padded size):
736
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight18
737
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight19
738
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight20
739
+ Params for bucket 8 (44040192 elements, 44040192 padded size):
740
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight21
741
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight22
742
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight23
743
+ Params for bucket 9 (44040192 elements, 44040192 padded size):
744
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight24
745
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight26
746
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight25
747
+ Params for bucket 10 (44040192 elements, 44040192 padded size):
748
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight27
749
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight29
750
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight28
751
+ Params for bucket 11 (58720256 elements, 58720256 padded size):
752
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight30
753
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight0
754
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight31
755
+ Params for bucket 12 (58720256 elements, 58720256 padded size):
756
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight1
757
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight2
758
+ Params for bucket 13 (58720256 elements, 58720256 padded size):
759
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight3
760
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight4
761
+ Params for bucket 14 (58720256 elements, 58720256 padded size):
762
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight5
763
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight6
764
+ Params for bucket 15 (58720256 elements, 58720256 padded size):
765
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight7
766
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight8
767
+ Params for bucket 16 (58720256 elements, 58720256 padded size):
768
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight9
769
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight10
770
+ Params for bucket 17 (58720256 elements, 58720256 padded size):
771
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight11
772
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight12
773
+ Params for bucket 18 (58720256 elements, 58720256 padded size):
774
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight13
775
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight14
776
+ Params for bucket 19 (58720256 elements, 58720256 padded size):
777
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight15
778
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight16
779
+ Params for bucket 20 (58720256 elements, 58720256 padded size):
780
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight17
781
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight18
782
+ Params for bucket 21 (58720256 elements, 58720256 padded size):
783
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight19
784
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight20
785
+ Params for bucket 22 (58720256 elements, 58720256 padded size):
786
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight21
787
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight22
788
+ Params for bucket 23 (58720256 elements, 58720256 padded size):
789
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight23
790
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight24
791
+ Params for bucket 24 (58720256 elements, 58720256 padded size):
792
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight25
793
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight26
794
+ Params for bucket 25 (58720256 elements, 58720256 padded size):
795
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight27
796
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight28
797
+ Params for bucket 26 (58720256 elements, 58720256 padded size):
798
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight29
799
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight30
800
+ Params for bucket 27 (29360128 elements, 29360128 padded size):
801
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight31
802
+ 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 0x7ec9843d3e60>, config_logger_dir='')
803
+ INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine
804
+ [after model, optimizer, and learning rate scheduler are built] datetime: 2026-09-21 07:02:13
805
+ > building train, validation, and test datasets ...
806
+ > datasets target sizes (minimum size):
807
+ train: 96
808
+ validation: 0
809
+ test: 0
810
+ INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None]
811
+ > building train, validation, and test datasets for GPT ...
812
+ INFO:megatron.core.datasets.blended_megatron_dataset_builder:Building GPTDataset splits with sizes=(96, 0, 0) and config=GPTDatasetConfig(random_seed=1234, sequence_length=4096, blend=(['/workspace/data/wikitext/wikitext103_gpt2_text_document'], None), blend_per_split=None, split='100,0,0', split_matrix=[(0, 1.0), None, None], num_dataset_builder_threads=1, path_to_cache=None, mmap_bin_files=True, mock=False, tokenizer=<megatron.training.tokenizer.tokenizer._GPT2BPETokenizer object at 0x7ec9841eafc0>, 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)
813
+ INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx
814
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths
815
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers
816
+ INFO:megatron.core.datasets.indexed_dataset: Extract the document indices
817
+ INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029
818
+ INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350
819
+ INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices
820
+ INFO:megatron.core.datasets.gpt_dataset: Load the document index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-document_index.npy
821
+ INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-sample_index.npy
822
+ INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-shuffle_index.npy
823
+ INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073
824
+ > finished creating GPT datasets ...
825
+ [after dataloaders are built] datetime: 2026-09-21 07:02:13
826
+ done with setup ...
827
+ training ...
828
+ Overwriting rerun_state_machine.current_iteration from -1 to 0...
829
+ (min, max) time across ranks (ms):
830
+ model-and-optimizer-setup ......................: (2031.71, 2031.82)
831
+ train/valid/test-data-iterators-setup ..........: (70.08, 80.16)
832
+ [before the start of training step] datetime: 2026-09-21 07:02:13
833
+ INFO:moonep.api:MoonEP comm buffer sizes:
834
+ config : R=8, S=4096, H=7168, E=256, K=8, NvS=40896, NvS_padded=40960
835
+ sm_count : num_sms=32, num_sms_dedup=148
836
+ hidden_buffer_size : 560.00 MiB
837
+ meta_buffer_size : 2.00 MiB
838
+ local_temp_buffer_size : 1.27 MiB
839
+ total_size : 563.27 MiB
840
+ [MoonEP selfcheck 1/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287] sizes [768, 640, 640, 1024, 640, 640, 640, 512, 1280, 640, 896, 640, 1536, 2304, 1536, 896, 1408, 384, 1280, 1664, 1152, 640, 640, 1152, 1280, 1536, 768, 512, 1664, 384, 384, 896, 2432, 1792, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[215, 197, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [194, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [205, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [78, 91, 86, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [203, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [50, 32, 34, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [82, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/results.tsv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir
2
+ DEEPSEEK_V3_L4 moonep matched 32 4096 8 none none 1 OK 192.67 192.30 97849 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep1
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/steps.tsv CHANGED
@@ -1,3 +1,6 @@
1
  status campaign phase gbs started ended result_dir
2
  OK echo tierbprobe 8 2026-09-21T06:42:21+00:00 2026-09-21T06:46:19+00:00 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/echo/tierbprobe
3
  OK ultraep tierbprobe 8 2026-09-21T06:46:22+00:00 2026-09-21T06:49:53+00:00 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/ultraep/tierbprobe
 
 
 
 
1
  status campaign phase gbs started ended result_dir
2
  OK echo tierbprobe 8 2026-09-21T06:42:21+00:00 2026-09-21T06:46:19+00:00 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/echo/tierbprobe
3
  OK ultraep tierbprobe 8 2026-09-21T06:46:22+00:00 2026-09-21T06:49:53+00:00 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/ultraep/tierbprobe
4
+ OK moonep tierbprobe 8 2026-09-21T06:49:55+00:00 2026-09-21T06:52:36+00:00 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierbprobe
5
+ OK echo tierb 8 2026-09-21T06:52:38+00:00 2026-09-21T06:56:36+00:00 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/echo/tierb
6
+ OK ultraep tierb 8 2026-09-21T06:56:38+00:00 2026-09-21T07:00:08+00:00 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/ultraep/tierb
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/study.log CHANGED
@@ -212,3 +212,162 @@ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
212
  --------------------------------------------------------------------------
213
  [06:51:22] CELL DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__forcebal__async (slots=32)
214
  [guard] iter 2/12 (60s)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
212
  --------------------------------------------------------------------------
213
  [06:51:22] CELL DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__forcebal__async (slots=32)
214
  [guard] iter 2/12 (60s)
215
+ [guard] iter 12/12 (70s)
216
+ [guard] tearing down process group (outcome=DONE)
217
+ --------------------------------------------------------------------------
218
+
219
+ ==============================================================
220
+ SWEEP COMPLETE ok=2 failed=0
221
+ ==============================================================
222
+ config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir
223
+ DEEPSEEK_V3_L4 moonep matched 32 4096 8 none none 1 OK 209.43 208.50 100169 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierbprobe/tierbprobe_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep1
224
+ DEEPSEEK_V3_L4 moonep matched 32 4096 8 none none+forcebal 1 OK 207.81 207.30 100169 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierbprobe/tierbprobe_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__forcebal__async/rep1
225
+
226
+ Results: /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierbprobe/tierbprobe_gbs8/results.tsv
227
+ Roll up: python3 /workspace/blackwell_run_scripts/granularity/tools/collect.py /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierbprobe/tierbprobe_gbs8
228
+ ==============================================================
229
+ 2 probe runs -> /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierbprobe/tierbprobe_gbs8/probe_summary.tsv
230
+ run arm cap rawExpI preRankI postRankI drop% reqCap
231
+ --------------------------------------------------------------------------------------------------------------------
232
+ DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capn natural none 5.6274 1.3085 1.0516
233
+ DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capn forced none 1.1010 1.0112 1.0051
234
+
235
+ ECHO note: echo_logged_pre_rank_metric_valid=0 and echo_logged_load_reduction_valid=0 are intentional.
236
+ The branch's built-in fields mix source-rank and destination-rank axes; this tool reconstructs both globally.
237
+
238
+ TIER B PROBE -- what these logs contain, per backend
239
+ ----------------------------------------------------
240
+ Capacity matrix:
241
+ ECHO: start C=2.00; on OOM descend 1.75, 1.50, 1.25. Stop at
242
+ the highest fitting zero-drop capacity. If a fitting capacity drops,
243
+ smaller values cannot help and the model is blocked.
244
+ UltraEP: measure required_capacity at C=1.25, round it UP to the next
245
+ 0.01, then validate that capacity once.
246
+ MoonEP: natural + forced, no capacity factor.
247
+
248
+ The forced arm runs once at the selected/high-headroom capacity. It validates
249
+ instrumentation and is not part of capacity selection.
250
+
251
+ Each backend instruments a different thing:
252
+
253
+ echo local raw-router histograms and local post-reroute placement.
254
+ reduce_probe.py aligns and sums all eight source ranks before
255
+ calculating raw_expert_*, pre_home_rank_*,
256
+ post_physical_rank_* and valid_max_rank_reduction_*.
257
+
258
+ ultraep global pre/post rank vectors plus explicit drop accounting:
259
+ drop_rate_pct and required_capacity -- the capacity factor the run
260
+ WOULD have needed. The vectors are duplicated identically on all
261
+ logging ranks and are deduplicated, not summed.
262
+
263
+ moonep local raw-router histograms, actual payload rows per physical
264
+ destination, and the bytes-copy side. The reducer globally sums
265
+ each (iteration, layer) snapshot first. ECHO reports no comparable
266
+ byte counters, so MoonEP bytes are a self-comparison across models,
267
+ never a cross-backend ranking.
268
+
269
+ Do NOT quote ECHO's logged pre_echo imbalance_ratio or
270
+ max_load_reduction_pct. The branch mixes source-rank and destination-rank
271
+ axes; the reducer marks both invalid and reconstructs the comparison globally.
272
+
273
+ Two boundary cases:
274
+ raw_expert_imax_max == E/K
275
+ -> at least one global snapshot had one expert in every
276
+ token's top-K set. Inspect the MEAN before generalising.
277
+ required_capacity or min-safe-capacity at the configured capacity
278
+ -> clipped by the buffer; it means "at or over", not
279
+ "fits comfortably".
280
+
281
+ tools/reduce_probe.py has already written probe_summary.tsv and .json here.
282
+
283
+ Campaign results: /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierbprobe
284
+ Roll up: python3 /workspace/blackwell_run_scripts/granularity/tools/collect.py /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierbprobe --compare
285
+ Already logged in to Hugging Face as: AbhishekGhosh1998
286
+ uploaded: /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierbprobe -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierbprobe
287
+ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
288
+
289
+ === STAGE 4 -- Tier B timing (probe-selected zero-drop capacities)
290
+
291
+ === STEP echo__tierb__gbs8
292
+ ==============================================================
293
+ CAMPAIGN: echo backend=echo phase=tierb
294
+ Megatron: /workspace/Megatron-MoE-Echo @ bracing-regions-through-kernels
295
+ GBS=8 seq=4096 repeats=3 nsys=1
296
+ results -> /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/echo/tierb
297
+ ==============================================================
298
+ [06:52:53] DEEPSEEK_V3_L4: full-CG capacity candidates: 1.13671875 1.10546875 1.08203125
299
+ --------------------------------------------------------------------------
300
+ [06:52:53] CELL DEEPSEEK_V3_L4__echo__matched__seq4096__gbs8__cap1p13671875 (slots=32)
301
+ [guard] iter 12/12 (70s)
302
+ [guard] tearing down process group (outcome=DONE)
303
+ [guard] iter 12/12 (70s)
304
+ [guard] tearing down process group (outcome=DONE)
305
+ [guard] iter 12/12 (70s)
306
+ [guard] tearing down process group (outcome=DONE)
307
+ --------------------------------------------------------------------------
308
+
309
+ ==============================================================
310
+ SWEEP COMPLETE ok=3 failed=0
311
+ ==============================================================
312
+ config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir
313
+ DEEPSEEK_V3_L4 echo matched 32 4096 8 1.13671875 none 1 OK 198.47 197.00 131075 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/echo/tierb/tierb_gbs8/DEEPSEEK_V3_L4__echo__matched__seq4096__gbs8__cap1p13671875/rep1
314
+ DEEPSEEK_V3_L4 echo matched 32 4096 8 1.13671875 none 2 OK 197.80 195.60 131075 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/echo/tierb/tierb_gbs8/DEEPSEEK_V3_L4__echo__matched__seq4096__gbs8__cap1p13671875/rep2
315
+ DEEPSEEK_V3_L4 echo matched 32 4096 8 1.13671875 none 3 OK 197.67 195.60 131075 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/echo/tierb/tierb_gbs8/DEEPSEEK_V3_L4__echo__matched__seq4096__gbs8__cap1p13671875/rep3
316
+
317
+ Results: /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/echo/tierb/tierb_gbs8/results.tsv
318
+ Roll up: python3 /workspace/blackwell_run_scripts/granularity/tools/collect.py /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/echo/tierb/tierb_gbs8
319
+ ==============================================================
320
+
321
+ Campaign results: /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/echo/tierb
322
+ Roll up: python3 /workspace/blackwell_run_scripts/granularity/tools/collect.py /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/echo/tierb --compare
323
+ Already logged in to Hugging Face as: AbhishekGhosh1998
324
+ uploaded: /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/echo/tierb -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/echo/tierb
325
+ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
326
+
327
+ === STEP ultraep__tierb__gbs8
328
+ ==============================================================
329
+ CAMPAIGN: ultraep backend=ultraep phase=tierb
330
+ Megatron: /workspace/Megatron-UltraEP @ ultraep-bracing-regions-through-kernels
331
+ GBS=8 seq=4096 repeats=3 nsys=1
332
+ results -> /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/ultraep/tierb
333
+ ==============================================================
334
+ --------------------------------------------------------------------------
335
+ [06:56:53] CELL DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875 (slots=32)
336
+ [guard] iter 12/12 (60s)
337
+ [guard] tearing down process group (outcome=DONE)
338
+ [guard] iter 12/12 (60s)
339
+ [guard] tearing down process group (outcome=DONE)
340
+ [guard] iter 12/12 (60s)
341
+ [guard] tearing down process group (outcome=DONE)
342
+ --------------------------------------------------------------------------
343
+
344
+ ==============================================================
345
+ SWEEP COMPLETE ok=3 failed=0
346
+ ==============================================================
347
+ config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir
348
+ DEEPSEEK_V3_L4 ultraep matched 32 4096 8 1.13671875 none 1 OK 195.76 194.20 101995 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/ultraep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875/rep1
349
+ DEEPSEEK_V3_L4 ultraep matched 32 4096 8 1.13671875 none 2 OK 195.69 194.00 101995 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/ultraep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875/rep2
350
+ DEEPSEEK_V3_L4 ultraep matched 32 4096 8 1.13671875 none 3 OK 195.66 194.00 101995 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/ultraep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__ultraep__matched__seq4096__gbs8__cap1p13671875/rep3
351
+
352
+ Results: /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/ultraep/tierb/tierb_gbs8/results.tsv
353
+ Roll up: python3 /workspace/blackwell_run_scripts/granularity/tools/collect.py /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/ultraep/tierb/tierb_gbs8
354
+ ==============================================================
355
+
356
+ Campaign results: /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/ultraep/tierb
357
+ Roll up: python3 /workspace/blackwell_run_scripts/granularity/tools/collect.py /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/ultraep/tierb --compare
358
+ Already logged in to Hugging Face as: AbhishekGhosh1998
359
+ uploaded: /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/ultraep/tierb -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/ultraep/tierb
360
+ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
361
+
362
+ === STEP moonep__tierb__gbs8
363
+ ==============================================================
364
+ CAMPAIGN: moonep backend=moonep phase=tierb
365
+ Megatron: /workspace/Megatron-MoE-Echo @ moonep_backend_async_grad_reduce
366
+ GBS=8 seq=4096 repeats=3 nsys=1
367
+ results -> /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb
368
+ ==============================================================
369
+ --------------------------------------------------------------------------
370
+ [07:00:25] CELL DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async (slots=32)
371
+ [guard] iter 3/12 (60s)
372
+ [guard] iter 12/12 (70s)
373
+ [guard] tearing down process group (outcome=DONE)
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/tierb_timing_capacities.env ADDED
@@ -0,0 +1 @@
 
 
1
+ TIERB_TIMING_CAP_DEEPSEEK_V3_L4=1.13671875
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/tierb_timing_capacities.tsv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ model common_timing_capacity echo_probe_required_capacity echo_full_cg_status echo_run_dir
2
+ DEEPSEEK_V3_L4 1.13671875 1.08203125 OK /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/echo/tierb/tierb_gbs8/DEEPSEEK_V3_L4__echo__matched__seq4096__gbs8__cap1p13671875/rep3