AbhishekGhosh1998 commited on
Commit
525945f
·
verified ·
1 Parent(s): 2bbfb21

granularity results granularity_next_b200/tierb_models/attempts/KIMI_K2_L2 @ 2026-09-21T07:52:54Z 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/KIMI_K2_L2/.completed/echo__tierb__gbs8 +1 -0
  2. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/.completed/moonep__tierbprobe__gbs8 +1 -0
  3. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/.completed/ultraep__tierb__gbs8 +1 -0
  4. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/STATUS.txt +5 -2
  5. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/hf_watcher.log +3 -0
  6. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/manifest.txt +19 -0
  7. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/pip_freeze.txt +317 -0
  8. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/provenance.txt +90 -0
  9. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep1/mem_samples.csv +224 -0
  10. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep1/run.env +105 -0
  11. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep1/train.log +935 -0
  12. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep2/mem_samples.csv +120 -0
  13. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep2/run.env +105 -0
  14. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep2/train.log +854 -0
  15. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/results.tsv +2 -0
  16. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/steps.tsv +3 -0
  17. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/study.log +159 -0
  18. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/tierb_timing_capacities.env +1 -0
  19. granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/tierb_timing_capacities.tsv +2 -0
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/.completed/echo__tierb__gbs8 ADDED
@@ -0,0 +1 @@
 
 
1
+ 2026-09-21T07:47:09+00:00
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/.completed/moonep__tierbprobe__gbs8 ADDED
@@ -0,0 +1 @@
 
 
1
+ 2026-09-21T07:43:07+00:00
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/.completed/ultraep__tierb__gbs8 ADDED
@@ -0,0 +1 @@
 
 
1
+ 2026-09-21T07:50:42+00:00
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/STATUS.txt CHANGED
@@ -1,10 +1,13 @@
1
  study_root: /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2
2
  gbs: 8
3
- updated: 2026-09-21T07:40:26+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/KIMI_K2_L2
2
  gbs: 8
3
+ updated: 2026-09-21T07:50:45+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/KIMI_K2_L2/hf_watcher.log CHANGED
@@ -4,3 +4,6 @@ Watching /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2 -> HF
4
  uploaded: /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2 -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2
5
  https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
6
  [2026-09-21T07:42:52+00:00] backup pass
 
 
 
 
4
  uploaded: /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2 -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2
5
  https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
6
  [2026-09-21T07:42:52+00:00] backup pass
7
+ uploaded: /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2 -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2
8
+ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
9
+ [2026-09-21T07:52:54+00:00] backup pass
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/manifest.txt ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ timestamp: 2026-09-21T07:50:59+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/KIMI_K2_L2/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/KIMI_K2_L2/moonep/tierb/provenance.txt ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ==============================================================
2
+ PROVENANCE: moonep__tierb__gbs8
3
+ ==============================================================
4
+ captured_at : 2026-09-21T07:50:45+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/KIMI_K2_L2/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, 29
21
+ 1, 1965 MHz, 1965 MHz, 1000.00 W, 35
22
+ 2, 120 MHz, 1965 MHz, 1000.00 W, 35
23
+ 3, 120 MHz, 1965 MHz, 1000.00 W, 30
24
+ 4, 1965 MHz, 1965 MHz, 1000.00 W, 30
25
+ 5, 120 MHz, 1965 MHz, 1000.00 W, 34
26
+ 6, 157 MHz, 1965 MHz, 1000.00 W, 34
27
+ 7, 120 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/KIMI_K2_L2
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/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__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, 1
25
+ 0, 1
26
+ 1, 1
27
+ 2, 1
28
+ 3, 1
29
+ 4, 1
30
+ 5, 1
31
+ 6, 3
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, 1
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, 4
69
+ 4, 629
70
+ 5, 629
71
+ 6, 4
72
+ 7, 629
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, 35085
90
+ 1, 35085
91
+ 2, 35085
92
+ 3, 35085
93
+ 4, 35085
94
+ 5, 35085
95
+ 6, 35085
96
+ 7, 35085
97
+ 0, 49119
98
+ 1, 49137
99
+ 2, 49145
100
+ 3, 49159
101
+ 4, 49163
102
+ 5, 49169
103
+ 6, 49153
104
+ 7, 49149
105
+ 0, 58717
106
+ 1, 58721
107
+ 2, 58711
108
+ 3, 58745
109
+ 4, 58759
110
+ 5, 58757
111
+ 6, 58737
112
+ 7, 58729
113
+ 0, 59207
114
+ 1, 59209
115
+ 2, 59207
116
+ 3, 59207
117
+ 4, 59207
118
+ 5, 59207
119
+ 6, 59207
120
+ 7, 59207
121
+ 0, 59207
122
+ 1, 59209
123
+ 2, 59207
124
+ 3, 59207
125
+ 4, 59207
126
+ 5, 59207
127
+ 6, 59207
128
+ 7, 59207
129
+ 0, 59207
130
+ 1, 59209
131
+ 2, 59207
132
+ 3, 59207
133
+ 4, 59207
134
+ 5, 59207
135
+ 6, 59207
136
+ 7, 59207
137
+ 0, 64093
138
+ 1, 64091
139
+ 2, 64093
140
+ 3, 64093
141
+ 4, 64093
142
+ 5, 64091
143
+ 6, 64107
144
+ 7, 64091
145
+ 0, 64739
146
+ 1, 64739
147
+ 2, 64739
148
+ 3, 64739
149
+ 4, 64739
150
+ 5, 64739
151
+ 6, 64753
152
+ 7, 64739
153
+ 0, 64741
154
+ 1, 64743
155
+ 2, 64741
156
+ 3, 64741
157
+ 4, 64741
158
+ 5, 64741
159
+ 6, 64755
160
+ 7, 64741
161
+ 0, 67145
162
+ 1, 67119
163
+ 2, 67093
164
+ 3, 67109
165
+ 4, 67091
166
+ 5, 67093
167
+ 6, 67123
168
+ 7, 67087
169
+ 0, 67923
170
+ 1, 67955
171
+ 2, 67963
172
+ 3, 67937
173
+ 4, 67941
174
+ 5, 67949
175
+ 6, 67945
176
+ 7, 67981
177
+ 0, 77117
178
+ 1, 77083
179
+ 2, 77091
180
+ 3, 77233
181
+ 4, 77273
182
+ 5, 77233
183
+ 6, 77287
184
+ 7, 77231
185
+ 0, 86365
186
+ 1, 86369
187
+ 2, 86365
188
+ 3, 86477
189
+ 4, 86477
190
+ 5, 86477
191
+ 6, 86491
192
+ 7, 86477
193
+ 0, 85719
194
+ 1, 85723
195
+ 2, 85719
196
+ 3, 84951
197
+ 4, 85511
198
+ 5, 85831
199
+ 6, 85845
200
+ 7, 85831
201
+ 0, 82771
202
+ 1, 83377
203
+ 2, 83033
204
+ 3, 82883
205
+ 4, 83145
206
+ 5, 83145
207
+ 6, 83215
208
+ 7, 83449
209
+ 0, 388
210
+ 1, 69261
211
+ 2, 69257
212
+ 3, 388
213
+ 4, 69369
214
+ 5, 324
215
+ 6, 68
216
+ 7, 69369
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/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep1/run.env ADDED
@@ -0,0 +1,105 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=moonep
2
+ config=KIMI_K2_L2
3
+ family=tierB
4
+ source_model=Kimi-K2
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=2
14
+ hidden=7168
15
+ ffn_hidden=18432
16
+ moe_ffn_hidden=2048
17
+ num_experts=384
18
+ topk=8
19
+ experts_per_rank=48
20
+ expert_params=44040192
21
+ heads=64
22
+ kv_groups=64
23
+ kv_channels=64
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=1
35
+ moe_layers=1
36
+ rotary_base=50000
37
+ rotary_percent=1.0
38
+ rotary_scaling=32
39
+ mscale=1.0
40
+ mscale_all_dim=1.0
41
+ qkv_bias=0
42
+ router_balancing=seq_aux_loss
43
+ aux_loss_coeff=0
44
+ router_pre_softmax=0
45
+ router_fusion=0
46
+ router_dtype=fp32
47
+ router_groups=1
48
+ router_group_topk=1
49
+ router_score_function=sigmoid
50
+ router_expert_bias=1
51
+ router_bias_update_rate=1e-3
52
+ router_topk_scaling_factor=2.827
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=682
76
+ slot_policy=matched
77
+ slots_per_rank=48
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/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep1/train.log ADDED
@@ -0,0 +1,935 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [07:51:08] WARN: moonep_backend_async_grad_reduce lacks --moe-router-fusion; KIMI_K2_L2 runs unfused (timing only, routing identical)
2
+ ==============================================================
3
+ KIMI_K2_L2 / moonep (slots/rank=48 policy=matched)
4
+ E=384 K=8 H'=2048 m=48
5
+ L=2 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/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep1/memory_snapshots/snapshot.pickle
8
+ repo=/workspace/Megatron-MoE-Echo @ 2e68c8ad
9
+ ==============================================================
10
+ W0921 07:51:09.881000 530811 torch/distributed/run.py:766]
11
+ W0921 07:51:09.881000 530811 torch/distributed/run.py:766] *****************************************
12
+ W0921 07:51:09.881000 530811 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 07:51:09.881000 530811 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
+ 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
19
+ 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
20
+ WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer
21
+ Number of virtual stages per pipeline stage: None
22
+ accumulate and all-reduce gradients in fp32 for bfloat16 data type.
23
+ using torch.bfloat16 for parameters ...
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
+ ------------------------ arguments ------------------------
27
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
28
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
29
+ account_for_embedding_in_pipeline_split ......... False
30
+ account_for_loss_in_pipeline_split .............. False
31
+ accumulate_allreduce_grads_in_fp32 .............. True
32
+ adam_beta1 ...................................... 0.9
33
+ adam_beta2 ...................................... 0.999
34
+ adam_eps ........................................ 1e-08
35
+ add_bias_linear ................................. False
36
+ add_position_embedding .......................... False
37
+ add_qkv_bias .................................... False
38
+ adlr_autoresume ................................. False
39
+ adlr_autoresume_interval ........................ 1000
40
+ align_grad_reduce ............................... True
41
+ align_param_gather .............................. False
42
+ app_tag_run_name ................................ None
43
+ app_tag_run_version ............................. 0.0.0
44
+ apply_layernorm_1p .............................. False
45
+ apply_query_key_layer_scaling ................... False
46
+ apply_residual_connection_post_layernorm ........ False
47
+ apply_rope_fusion ............................... False
48
+ async_save ...................................... None
49
+ async_tensor_model_parallel_allreduce ........... True
50
+ attention_backend ............................... AttnBackend.auto
51
+ attention_dropout ............................... 0.0
52
+ attention_softmax_in_fp32 ....................... False
53
+ auto_detect_ckpt_format ......................... False
54
+ barrier_with_L1_time ............................ True
55
+ bert_binary_head ................................ True
56
+ bert_embedder_type .............................. megatron
57
+ bert_load ....................................... None
58
+ bf16 ............................................ True
59
+ bias_dropout_fusion ............................. True
60
+ bias_gelu_fusion ................................ False
61
+ bias_swiglu_fusion .............................. True
62
+ biencoder_projection_dim ........................ 0
63
+ biencoder_shared_query_context_model ............ False
64
+ block_data_path ................................. None
65
+ calc_ft_timeouts ................................ False
66
+ calculate_per_token_loss ........................ False
67
+ check_for_large_grads ........................... False
68
+ check_for_nan_in_loss_and_grad .................. False
69
+ check_for_spiky_loss ............................ False
70
+ check_weight_hash_across_dp_replicas_interval ... None
71
+ ckpt_assume_constant_structure .................. False
72
+ ckpt_convert_format ............................. None
73
+ ckpt_convert_save ............................... None
74
+ ckpt_convert_update_legacy_dist_opt_format ...... False
75
+ ckpt_format ..................................... torch_dist
76
+ ckpt_fully_parallel_load ........................ False
77
+ ckpt_fully_parallel_save ........................ True
78
+ ckpt_fully_parallel_save_deprecated ............. False
79
+ ckpt_step ....................................... None
80
+ classes_fraction ................................ 1.0
81
+ clip_grad ....................................... 1.0
82
+ clone_scatter_output_in_embedding ............... True
83
+ config_logger_dir ...............................
84
+ consumed_train_samples .......................... 0
85
+ consumed_valid_samples .......................... 0
86
+ context_parallel_size ........................... 1
87
+ cp_comm_type .................................... ['p2p']
88
+ create_attention_mask_in_dataloader ............. True
89
+ cross_entropy_fusion_impl ....................... native
90
+ cross_entropy_loss_fusion ....................... False
91
+ cuda_graph_scope ................................ full_iteration
92
+ cuda_graph_warmup_steps ......................... 3
93
+ data_args_path .................................. None
94
+ data_cache_path ................................. None
95
+ data_parallel_random_init ....................... False
96
+ data_parallel_sharding_strategy ................. no_shard
97
+ data_parallel_size .............................. 8
98
+ data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document']
99
+ data_per_class_fraction ......................... 1.0
100
+ data_sharding ................................... True
101
+ dataloader_type ................................. single
102
+ ddp_average_in_collective ....................... False
103
+ ddp_bucket_size ................................. None
104
+ ddp_num_buckets ................................. None
105
+ ddp_pad_buckets_for_high_nccl_busbw ............. False
106
+ decoder_first_pipeline_num_layers ............... None
107
+ decoder_last_pipeline_num_layers ................ None
108
+ decoder_num_layers .............................. None
109
+ decoder_seq_length .............................. None
110
+ decoupled_lr .................................... None
111
+ decoupled_min_lr ................................ None
112
+ decrease_batch_size_if_needed ................... False
113
+ defer_embedding_wgrad_compute ................... False
114
+ delay_wgrad_compute ............................. False
115
+ deprecated_use_mcore_models ..................... True
116
+ deterministic_mode .............................. False
117
+ dino_bottleneck_size ............................ 256
118
+ dino_freeze_last_layer .......................... 1
119
+ dino_head_hidden_size ........................... 2048
120
+ dino_local_crops_number ......................... 10
121
+ dino_local_img_size ............................. 96
122
+ dino_norm_last_layer ............................ False
123
+ dino_teacher_temp ............................... 0.07
124
+ dino_warmup_teacher_temp ........................ 0.04
125
+ dino_warmup_teacher_temp_epochs ................. 30
126
+ disable_bf16_reduced_precision_matmul ........... False
127
+ disable_mamba_mem_eff_path ...................... False
128
+ disable_straggler_on_startup .................... False
129
+ dist_ckpt_format_deprecated ..................... None
130
+ dist_ckpt_strictness ............................ assume_ok_unexpected
131
+ distribute_saved_activations .................... False
132
+ distributed_backend ............................. nccl
133
+ distributed_timeout_minutes ..................... 10
134
+ embedding_init_method_std ....................... None
135
+ embedding_path .................................. None
136
+ empty_unused_memory_level ....................... 0
137
+ enable_cuda_graph ............................... True
138
+ enable_experimental ............................. False
139
+ enable_ft_package ............................... False
140
+ enable_gloo_process_groups ...................... True
141
+ enable_msc ...................................... True
142
+ enable_one_logger ............................... True
143
+ encoder_num_layers .............................. 2
144
+ encoder_seq_length .............................. 4096
145
+ end_weight_decay ................................ 0.1
146
+ eod_mask_loss ................................... False
147
+ error_injection_rate ............................ 0
148
+ error_injection_type ............................ transient_error
149
+ eval_interval ................................... 1000000
150
+ eval_iters ...................................... 0
151
+ evidence_data_path .............................. None
152
+ exit_duration_in_mins ........................... None
153
+ exit_interval ................................... None
154
+ exit_on_missing_checkpoint ...................... False
155
+ exit_signal_handler ............................. False
156
+ exp_avg_dtype ................................... torch.float32
157
+ exp_avg_sq_dtype ................................ torch.float32
158
+ expert_model_parallel_size ...................... 8
159
+ expert_tensor_parallel_size ..................... 1
160
+ export_force_local_attention .................... False
161
+ export_kd_cfg ................................... None
162
+ export_kd_teacher_ckpt_format ................... None
163
+ export_kd_teacher_load .......................... None
164
+ export_kv_cache_quant ........................... False
165
+ export_legacy_megatron .......................... False
166
+ export_model_type ............................... GPTModel
167
+ export_moe_apply_probs_on_input ................. False
168
+ export_qk_l2_norm ............................... False
169
+ export_quant_cfg ................................ None
170
+ export_real_quant_cfg ........................... None
171
+ export_te_mcore_model ........................... False
172
+ external_cuda_graph ............................. False
173
+ ffn_hidden_size ................................. 18432
174
+ finetune ........................................ False
175
+ finetune_data_split ............................. train
176
+ finetune_hf_dataset ............................. None
177
+ first_last_layers_bf16 .......................... False
178
+ flash_decode .................................... False
179
+ fp16 ............................................ False
180
+ fp16_lm_cross_entropy ........................... False
181
+ fp32_residual_connection ........................ False
182
+ fp8 ............................................. e4m3
183
+ fp8_amax_compute_algo ........................... most_recent
184
+ fp8_amax_history_len ............................ 1
185
+ fp8_interval .................................... 1
186
+ fp8_margin ...................................... 0
187
+ fp8_param_gather ................................ False
188
+ fp8_recipe ...................................... mxfp8
189
+ fp8_wgrad ....................................... True
190
+ fsdp_double_buffer .............................. False
191
+ global_batch_size ............................... 8
192
+ grad_reduce_in_bf16 ............................. False
193
+ gradient_accumulation_fusion .................... True
194
+ gradient_reduce_div_fusion ...................... True
195
+ group_query_attention ........................... False
196
+ head_lr_mult .................................... 1.0
197
+ heterogeneous_layers_config_encoded_json ........ None
198
+ heterogeneous_layers_config_path ................ None
199
+ hidden_dropout .................................. 0.0
200
+ hidden_size ..................................... 7168
201
+ hierarchical_context_parallel_sizes ............. None
202
+ high_priority_stream_groups ..................... []
203
+ hybrid_attention_ratio .......................... 0.0
204
+ hybrid_mlp_ratio ................................ 0.0
205
+ hybrid_override_pattern ......................... None
206
+ hysteresis ...................................... 2
207
+ ict_head_size ................................... None
208
+ ict_load ........................................ None
209
+ img_h ........................................... 224
210
+ img_w ........................................... 224
211
+ indexer_batch_size .............................. 128
212
+ indexer_log_interval ............................ 1000
213
+ inference_batch_times_seqlen_threshold .......... -1
214
+ inference_dynamic_batching ...................... False
215
+ inference_dynamic_batching_buffer_guaranteed_fraction 0.2
216
+ inference_dynamic_batching_buffer_overflow_factor None
217
+ inference_dynamic_batching_buffer_size_gb ....... 40.0
218
+ inference_dynamic_batching_chunk_size ........... 256
219
+ inference_dynamic_batching_max_requests_override None
220
+ inference_dynamic_batching_max_tokens_override .. None
221
+ inference_dynamic_batching_num_cuda_graphs ...... 16
222
+ inference_max_batch_size ........................ 8
223
+ inference_max_seq_length ........................ 2560
224
+ inference_rng_tracker ........................... False
225
+ init_method_std ................................. 0.02
226
+ init_method_xavier_uniform ...................... False
227
+ init_model_with_meta_device ..................... False
228
+ initial_loss_scale .............................. 4294967296
229
+ inprocess_active_world_size ..................... 8
230
+ inprocess_barrier_timeout ....................... 120
231
+ inprocess_completion_timeout .................... 120
232
+ inprocess_empty_cuda_cache ...................... False
233
+ inprocess_granularity ........................... node
234
+ inprocess_hard_timeout .......................... 90
235
+ inprocess_heartbeat_interval .................... 30
236
+ inprocess_heartbeat_timeout ..................... 60
237
+ inprocess_last_call_wait ........................ 1
238
+ inprocess_max_iterations ........................ None
239
+ inprocess_monitor_process_interval .............. 1.0
240
+ inprocess_monitor_thread_interval ............... 1.0
241
+ inprocess_progress_watchdog_interval ............ 1.0
242
+ inprocess_restart ............................... False
243
+ inprocess_soft_timeout .......................... 60
244
+ inprocess_termination_grace_time ................ 1
245
+ is_hybrid_model ................................. False
246
+ iter_per_epoch .................................. 1250
247
+ iterations_to_skip .............................. []
248
+ keep_fp8_transpose_cache_when_using_custom_fsdp . False
249
+ kitchen_config_file ............................. None
250
+ kitchen_recipe_number ........................... None
251
+ kv_channels ..................................... 64
252
+ kv_lora_rank .................................... 512
253
+ lazy_mpu_init ................................... None
254
+ load ............................................ None
255
+ load_main_params_from_ckpt ...................... None
256
+ load_model_opt_format ........................... False
257
+ local_rank ...................................... 0
258
+ log_energy ...................................... False
259
+ log_interval .................................... 1
260
+ log_loss_scale_to_tensorboard ................... True
261
+ log_memory_to_tensorboard ....................... False
262
+ log_num_zeros_in_grad ........................... False
263
+ log_params_norm ................................. False
264
+ log_progress .................................... False
265
+ log_straggler ................................... False
266
+ log_throughput .................................. True
267
+ log_timers_to_tensorboard ....................... False
268
+ log_validation_ppl_to_tensorboard ............... False
269
+ log_world_size_to_tensorboard ................... False
270
+ logging_level ................................... None
271
+ loss_scale ...................................... None
272
+ loss_scale_window ............................... 1000
273
+ lr .............................................. 0.0001
274
+ lr_decay_iters .................................. 8
275
+ lr_decay_samples ................................ None
276
+ lr_decay_style .................................. cosine
277
+ lr_warmup_fraction .............................. None
278
+ lr_warmup_init .................................. 0.0
279
+ lr_warmup_iters ................................. 5
280
+ lr_warmup_samples ............................... 0
281
+ lr_wsd_decay_iters .............................. None
282
+ lr_wsd_decay_samples ............................ None
283
+ lr_wsd_decay_style .............................. exponential
284
+ main_grads_dtype ................................ torch.float32
285
+ main_params_dtype ............................... torch.float32
286
+ make_vocab_size_divisible_by .................... 128
287
+ mamba_head_dim .................................. 64
288
+ mamba_num_groups ................................ 8
289
+ mamba_num_heads ................................. None
290
+ mamba_state_dim ................................. 128
291
+ manual_gc ....................................... False
292
+ manual_gc_eval .................................. True
293
+ manual_gc_interval .............................. 0
294
+ mask_factor ..................................... 1.0
295
+ mask_prob ....................................... 0.15
296
+ mask_type ....................................... random
297
+ masked_softmax_fusion ........................... False
298
+ max_position_embeddings ......................... 4096
299
+ max_tokens_to_oom ............................... 12000
300
+ memory_snapshot_path ............................ snapshot.pickle
301
+ merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt
302
+ micro_batch_size ................................ 1
303
+ microbatch_group_size_per_vp_stage .............. None
304
+ mid_level_dataset_surplus ....................... 0.005
305
+ min_loss_scale .................................. 1.0
306
+ min_lr .......................................... 1e-05
307
+ mlp_chunks_for_prefill .......................... 1
308
+ mmap_bin_files .................................. True
309
+ mock_data ....................................... False
310
+ moe_apply_probs_on_input ........................ False
311
+ moe_aux_loss_coeff .............................. 0.0
312
+ moe_deepep_num_sms .............................. 20
313
+ moe_echo_enable_random_offloading ............... False
314
+ moe_echo_expert_dispatch_overlap ................ False
315
+ moe_echo_expert_dispatcher_type ................. hybridep
316
+ moe_echo_recompute_expert_dispatch .............. False
317
+ moe_enable_deepep ............................... False
318
+ moe_enable_echo ................................. False
319
+ moe_enable_hybridep ............................. False
320
+ moe_enable_moonep ............................... True
321
+ moe_expert_capacity_factor ...................... None
322
+ moe_extended_tp ................................. False
323
+ moe_ffn_hidden_size ............................. 2048
324
+ moe_grouped_gemm ................................ True
325
+ moe_hybridep_num_sms ............................ 16
326
+ moe_input_jitter_eps ............................ None
327
+ moe_layer_freq .................................. [0, 1]
328
+ moe_layer_recompute ............................. False
329
+ moe_moonep_async_grad_reduce .................... True
330
+ moe_moonep_num_sms .............................. None
331
+ moe_num_echo_experts ............................ None
332
+ moe_pad_expert_input_to_capacity ................ False
333
+ moe_per_layer_logging ........................... False
334
+ moe_permute_fusion .............................. False
335
+ moe_received_token_capacity ..................... None
336
+ moe_router_bias_update_rate ..................... 0.001
337
+ moe_router_dtype ................................ fp32
338
+ moe_router_enable_expert_bias ................... True
339
+ moe_router_force_load_balancing ................. False
340
+ moe_router_group_topk ........................... 1
341
+ moe_router_load_balancing_type .................. seq_aux_loss
342
+ moe_router_num_groups ........................... 1
343
+ moe_router_padding_for_fp8 ...................... False
344
+ moe_router_pre_softmax .......................... False
345
+ moe_router_score_function ....................... sigmoid
346
+ moe_router_topk ................................. 8
347
+ moe_router_topk_scaling_factor .................. 2.827
348
+ moe_shared_expert_intermediate_size ............. 2048
349
+ moe_shared_expert_overlap ....................... False
350
+ moe_token_dispatcher_type ....................... flex
351
+ moe_token_drop_policy ........................... probs
352
+ moe_upcycling_granularity ....................... 1
353
+ moe_use_device_initiated_grouped_gemm ........... True
354
+ moe_use_legacy_grouped_gemm ..................... False
355
+ moe_use_upcycling ............................... False
356
+ moe_z_loss_coeff ................................ None
357
+ mrope_section ................................... None
358
+ mscale .......................................... 1.0
359
+ mscale_all_dim .................................. 1.0
360
+ mtp_loss_scaling_factor ......................... 0.1
361
+ mtp_num_layers .................................. None
362
+ multi_latent_attention .......................... True
363
+ nccl_all_reduce_for_prefill ..................... False
364
+ nccl_communicator_config_path ................... None
365
+ nccl_ub ......................................... False
366
+ no_load_optim ................................... None
367
+ no_load_rng ..................................... None
368
+ no_persist_layer_norm ........................... False
369
+ no_rope_freq .................................... None
370
+ no_save_optim ................................... None
371
+ no_save_rng ..................................... None
372
+ non_persistent_ckpt_type ........................ None
373
+ non_persistent_global_ckpt_dir .................. None
374
+ non_persistent_local_ckpt_algo .................. fully_parallel
375
+ non_persistent_local_ckpt_dir ................... None
376
+ non_persistent_save_interval .................... None
377
+ norm_epsilon .................................... 1e-06
378
+ normalization ................................... RMSNorm
379
+ num_attention_heads ............................. 64
380
+ num_channels .................................... 3
381
+ num_classes ..................................... 1000
382
+ num_dataset_builder_threads ..................... 1
383
+ num_distributed_optimizer_instances ............. 1
384
+ num_experts ..................................... 384
385
+ num_layers ...................................... 2
386
+ num_layers_at_end_in_bf16 ....................... 1
387
+ num_layers_at_start_in_bf16 ..................... 1
388
+ num_layers_per_virtual_pipeline_stage ........... None
389
+ num_query_groups ................................ 1
390
+ num_virtual_stages_per_pipeline_rank ............ None
391
+ num_workers ..................................... 2
392
+ object_storage_cache_path ....................... None
393
+ one_logger_async ................................ False
394
+ one_logger_project .............................. megatron-lm
395
+ one_logger_run_name ............................. None
396
+ onnx_safe ....................................... None
397
+ openai_gelu ..................................... False
398
+ optimizer ....................................... adam
399
+ optimizer_cpu_offload ........................... False
400
+ optimizer_offload_fraction ...................... 1.0
401
+ output_bert_embeddings .......................... False
402
+ overlap_cpu_optimizer_d2h_h2d ................... False
403
+ overlap_grad_reduce ............................. True
404
+ overlap_moe_expert_parallel_comm ................ False
405
+ overlap_p2p_comm ................................ False
406
+ overlap_p2p_comm_warmup_flush ................... False
407
+ overlap_param_gather ............................ True
408
+ overlap_param_gather_with_optimizer_step ........ False
409
+ override_opt_param_scheduler .................... False
410
+ params_dtype .................................... torch.bfloat16
411
+ patch_dim ....................................... 16
412
+ per_split_data_args_path ........................ None
413
+ perform_initialization .......................... True
414
+ pin_cpu_grads ................................... True
415
+ pin_cpu_params .................................. True
416
+ pipeline_model_parallel_comm_backend ............ None
417
+ pipeline_model_parallel_layout .................. None
418
+ pipeline_model_parallel_size .................... 1
419
+ position_embedding_type ......................... rope
420
+ pretrained_checkpoint ........................... None
421
+ profile ......................................... False
422
+ profile_ranks ................................... [0]
423
+ profile_step_end ................................ 12
424
+ profile_step_start .............................. 10
425
+ q_lora_rank ..................................... 1536
426
+ qk_head_dim ..................................... 128
427
+ qk_l2_norm ...................................... False
428
+ qk_layernorm .................................... True
429
+ qk_pos_emb_head_dim ............................. 64
430
+ query_in_block_prob ............................. 0.1
431
+ rampup_batch_size ............................... None
432
+ rank ............................................ 0
433
+ recompute_granularity ........................... None
434
+ recompute_method ................................ None
435
+ recompute_modules ............................... None
436
+ recompute_num_layers ............................ None
437
+ record_memory_history ........................... False
438
+ relative_attention_max_distance ................. 128
439
+ relative_attention_num_buckets .................. 32
440
+ replication ..................................... False
441
+ replication_factor .............................. 2
442
+ replication_jump ................................ None
443
+ rerun_mode ...................................... validate_results
444
+ reset_attention_mask ............................ False
445
+ reset_position_ids .............................. False
446
+ result_rejected_tracker_filename ................ None
447
+ retriever_report_topk_accuracies ................ []
448
+ retriever_score_scaling ......................... False
449
+ retriever_seq_length ............................ 256
450
+ retro_add_retriever ............................. False
451
+ retro_attention_gate ............................ 1
452
+ retro_cyclic_train_iters ........................ None
453
+ retro_encoder_attention_dropout ................. 0.1
454
+ retro_encoder_hidden_dropout .................... 0.1
455
+ retro_encoder_layers ............................ 2
456
+ retro_num_neighbors ............................. 2
457
+ retro_num_retrieved_chunks ...................... 2
458
+ retro_project_dir ............................... None
459
+ retro_verify_neighbor_count ..................... True
460
+ reuse_grad_buf_for_mxfp8_param_ag ............... False
461
+ rope_scaling_factor ............................. 8.0
462
+ rope_type ....................................... None
463
+ rotary_base ..................................... 50000
464
+ rotary_interleaved .............................. False
465
+ rotary_percent .................................. 1.0
466
+ rotary_scaling_factor ........................... 32.0
467
+ rotary_seq_len_interpolation_factor ............. None
468
+ run_workload_inspector_server ................... False
469
+ sample_rate ..................................... 1.0
470
+ save ............................................ None
471
+ save_interval ................................... None
472
+ save_retain_interval ............................ None
473
+ scatter_gather_tensors_in_pipeline .............. True
474
+ seed ............................................ 1234
475
+ seq_length ...................................... 4096
476
+ sequence_parallel ............................... False
477
+ sft ............................................. False
478
+ sft_tokenizer_prompt_format ..................... nemotron-h-aligned
479
+ sgd_momentum .................................... 0.9
480
+ short_seq_prob .................................. 0.1
481
+ skip_train ...................................... False
482
+ skipped_train_samples ........................... 0
483
+ spec ............................................ None
484
+ split ........................................... 100,0,0
485
+ squared_relu .................................... False
486
+ start_weight_decay .............................. 0.1
487
+ straggler_ctrlr_port ............................ 65535
488
+ straggler_minmax_count .......................... 1
489
+ suggested_communication_unit_size ............... None
490
+ swiglu .......................................... True
491
+ swin_backbone_type .............................. tiny
492
+ symmetric_ar_type ............................... None
493
+ te_rng_tracker .................................. True
494
+ tensor_model_parallel_size ...................... 1
495
+ tensorboard_dir ................................. None
496
+ tensorboard_log_interval ........................ 1
497
+ tensorboard_queue_size .......................... 1000
498
+ test_data_path .................................. None
499
+ test_mode ....................................... False
500
+ tiktoken_num_special_tokens ..................... 1000
501
+ tiktoken_pattern ................................ None
502
+ tiktoken_special_tokens ......................... None
503
+ timing_log_level ................................ 0
504
+ timing_log_option ............................... minmax
505
+ titles_data_path ................................ None
506
+ tokenizer_model ................................. None
507
+ tokenizer_type .................................. GPT2BPETokenizer
508
+ torch_fsdp2_reshard_after_forward ............... True
509
+ tp_comm_bootstrap_backend ....................... nccl
510
+ tp_comm_bulk_dgrad .............................. True
511
+ tp_comm_bulk_wgrad .............................. True
512
+ tp_comm_overlap ................................. False
513
+ tp_comm_overlap_ag .............................. True
514
+ tp_comm_overlap_cfg ............................. None
515
+ tp_comm_overlap_rs .............................. True
516
+ tp_comm_overlap_rs_dgrad ........................ False
517
+ tp_comm_split_ag ................................ True
518
+ tp_comm_split_rs ................................ True
519
+ train_data_path ................................. None
520
+ train_iters ..................................... 12
521
+ train_samples ................................... None
522
+ train_sync_interval ............................. None
523
+ transformer_impl ................................ transformer_engine
524
+ transformer_pipeline_model_parallel_size ........ 1
525
+ untie_embeddings_and_output_weights ............. True
526
+ use_checkpoint_args ............................. False
527
+ use_checkpoint_opt_param_scheduler .............. False
528
+ use_cpu_initialization .......................... None
529
+ use_custom_fsdp ................................. False
530
+ use_dist_ckpt ................................... True
531
+ use_dist_ckpt_deprecated ........................ False
532
+ use_distributed_optimizer ....................... True
533
+ use_flash_attn .................................. False
534
+ use_legacy_models ............................... False
535
+ use_mp_args_from_checkpoint_args ................ False
536
+ use_one_sent_docs ............................... False
537
+ use_persistent_ckpt_worker ...................... False
538
+ use_precision_aware_optimizer ................... False
539
+ use_pytorch_profiler ............................ False
540
+ use_ring_exchange_p2p ........................... False
541
+ use_rope_scaling ................................ False
542
+ use_rotary_position_embeddings .................. False
543
+ use_sharp ....................................... False
544
+ use_tokenizer_model_from_checkpoint_args ........ True
545
+ use_torch_fsdp2 ................................. False
546
+ use_torch_optimizer_for_cpu_offload ............. False
547
+ use_tp_pp_dp_mapping ............................ False
548
+ v_head_dim ...................................... 128
549
+ valid_data_path ................................. None
550
+ variable_seq_lengths ............................ False
551
+ virtual_pipeline_model_parallel_size ............ None
552
+ vision_backbone_type ............................ vit
553
+ vision_pretraining .............................. False
554
+ vision_pretraining_type ......................... classify
555
+ vocab_extra_ids ................................. 0
556
+ vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json
557
+ vocab_size ...................................... None
558
+ wandb_exp_name ..................................
559
+ wandb_project ...................................
560
+ wandb_save_dir ..................................
561
+ weight_decay .................................... 0.1
562
+ weight_decay_incr_style ......................... constant
563
+ wgrad_deferral_limit ............................ 0
564
+ world_size ...................................... 8
565
+ yaml_cfg ........................................ None
566
+ -------------------- end of arguments ---------------------
567
+ INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1
568
+ > building GPT2BPETokenizer tokenizer ...
569
+ > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304)
570
+ WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS
571
+ > initializing torch distributed ...
572
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
573
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
574
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
575
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
576
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
577
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
578
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
579
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
580
+ [Gloo] Rank 0 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 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
583
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
584
+ [Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
585
+ [Gloo] Rank [Gloo] Rank 4 is connected to 73 peer ranks. is connected to Expected number of connected peer ranks is : 77 peer ranks.
586
+ Expected number of connected peer ranks is : 7
587
+ [Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
588
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
589
+ [Gloo] Rank 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. Expected number of connected peer ranks is : 7
591
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7[Gloo] Rank
592
+ 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
593
+ [Gloo] Rank 6 is connected to [Gloo] Rank 75 peer ranks. is connected to Expected number of connected peer ranks is : 77 peer ranks.
594
+ Expected number of connected peer ranks is : 7
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 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
601
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
602
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
603
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
604
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
605
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
606
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
607
+ [rank3]:[W921 07:51:26.889347288 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.
608
+ [rank6]:[W921 07:51:26.889357851 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.
609
+ [rank7]:[W921 07:51:26.889411270 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.
610
+ [rank1]:[W921 07:51:26.892900024 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
+ [rank2]:[W921 07:51:26.893004294 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.
612
+ [rank5]:[W921 07:51:26.893156617 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.
613
+ [rank4]:[W921 07:51:26.897641184 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.
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 07:51:27.046041478 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: 3.847 seconds
623
+ time to initialize megatron (seconds): 7.419
624
+ [after megatron is initialized] datetime: 2026-09-21 07:51:33
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): 3480525824
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: 9
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.1.mlp.shared_experts.linear_fc1.weight
650
+ module.decoder.layers.1.mlp.shared_experts.linear_fc2.weight
651
+ Params for bucket 3 (45163520 elements, 45163520 padded size):
652
+ module.decoder.layers.1.self_attention.linear_kv_down_proj.weight
653
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.weight
654
+ module.decoder.layers.1.self_attention.linear_q_up_proj.weight
655
+ module.decoder.layers.1.mlp.router.weight
656
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.layer_norm_weight
657
+ module.decoder.layers.1.pre_mlp_layernorm.weight
658
+ module.decoder.layers.1.self_attention.linear_q_up_proj.layer_norm_weight
659
+ module.decoder.layers.1.self_attention.linear_q_down_proj.weight
660
+ Params for bucket 4 (58720256 elements, 58720256 padded size):
661
+ module.decoder.layers.1.self_attention.linear_proj.weight
662
+ Params for bucket 5 (132127744 elements, 132127744 padded size):
663
+ module.decoder.layers.1.input_layernorm.weight
664
+ module.decoder.layers.0.mlp.linear_fc2.weight
665
+ Params for bucket 6 (264241152 elements, 264241152 padded size):
666
+ module.decoder.layers.0.mlp.linear_fc1.weight
667
+ Params for bucket 7 (42411008 elements, 42411008 padded size):
668
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.weight
669
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.layer_norm_weight
670
+ module.decoder.layers.0.self_attention.linear_q_up_proj.layer_norm_weight
671
+ module.decoder.layers.0.mlp.linear_fc1.layer_norm_weight
672
+ module.decoder.layers.0.self_attention.linear_q_up_proj.weight
673
+ module.decoder.layers.0.self_attention.linear_kv_down_proj.weight
674
+ module.decoder.layers.0.self_attention.linear_q_down_proj.weight
675
+ Params for bucket 8 (58720256 elements, 58720256 padded size):
676
+ module.decoder.layers.0.self_attention.linear_proj.weight
677
+ Params for bucket 9 (360586240 elements, 360586240 padded size):
678
+ module.decoder.layers.0.input_layernorm.weight
679
+ module.embedding.word_embeddings.weight
680
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 40
681
+ Params for bucket 1 (44040192 elements, 44040192 padded size):
682
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight0
683
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight1
684
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight2
685
+ Params for bucket 2 (44040192 elements, 44040192 padded size):
686
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight3
687
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight4
688
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight5
689
+ Params for bucket 3 (44040192 elements, 44040192 padded size):
690
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight6
691
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight7
692
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight8
693
+ Params for bucket 4 (44040192 elements, 44040192 padded size):
694
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight9
695
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight10
696
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight11
697
+ Params for bucket 5 (44040192 elements, 44040192 padded size):
698
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight12
699
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight13
700
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight14
701
+ Params for bucket 6 (44040192 elements, 44040192 padded size):
702
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight15
703
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight16
704
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight17
705
+ Params for bucket 7 (44040192 elements, 44040192 padded size):
706
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight18
707
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight19
708
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight20
709
+ Params for bucket 8 (44040192 elements, 44040192 padded size):
710
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight21
711
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight22
712
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight23
713
+ Params for bucket 9 (44040192 elements, 44040192 padded size):
714
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight24
715
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight26
716
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight25
717
+ Params for bucket 10 (44040192 elements, 44040192 padded size):
718
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight27
719
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight29
720
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight28
721
+ Params for bucket 11 (44040192 elements, 44040192 padded size):
722
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight30
723
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight32
724
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight31
725
+ Params for bucket 12 (44040192 elements, 44040192 padded size):
726
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight33
727
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight35
728
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight34
729
+ Params for bucket 13 (44040192 elements, 44040192 padded size):
730
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight36
731
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight38
732
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight37
733
+ Params for bucket 14 (44040192 elements, 44040192 padded size):
734
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight39
735
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight41
736
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight40
737
+ Params for bucket 15 (44040192 elements, 44040192 padded size):
738
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight42
739
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight44
740
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight43
741
+ Params for bucket 16 (44040192 elements, 44040192 padded size):
742
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight45
743
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight46
744
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight47
745
+ Params for bucket 17 (58720256 elements, 58720256 padded size):
746
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight0
747
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight1
748
+ Params for bucket 18 (58720256 elements, 58720256 padded size):
749
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight2
750
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight3
751
+ Params for bucket 19 (58720256 elements, 58720256 padded size):
752
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight4
753
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight5
754
+ Params for bucket 20 (58720256 elements, 58720256 padded size):
755
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight6
756
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight7
757
+ Params for bucket 21 (58720256 elements, 58720256 padded size):
758
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight8
759
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight9
760
+ Params for bucket 22 (58720256 elements, 58720256 padded size):
761
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight10
762
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight11
763
+ Params for bucket 23 (58720256 elements, 58720256 padded size):
764
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight12
765
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight13
766
+ Params for bucket 24 (58720256 elements, 58720256 padded size):
767
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight14
768
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight15
769
+ Params for bucket 25 (58720256 elements, 58720256 padded size):
770
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight16
771
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight17
772
+ Params for bucket 26 (58720256 elements, 58720256 padded size):
773
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight18
774
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight19
775
+ Params for bucket 27 (58720256 elements, 58720256 padded size):
776
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight20
777
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight21
778
+ Params for bucket 28 (58720256 elements, 58720256 padded size):
779
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight22
780
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight23
781
+ Params for bucket 29 (58720256 elements, 58720256 padded size):
782
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight24
783
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight25
784
+ Params for bucket 30 (58720256 elements, 58720256 padded size):
785
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight26
786
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight27
787
+ Params for bucket 31 (58720256 elements, 58720256 padded size):
788
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight28
789
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight29
790
+ Params for bucket 32 (58720256 elements, 58720256 padded size):
791
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight30
792
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight31
793
+ Params for bucket 33 (58720256 elements, 58720256 padded size):
794
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight32
795
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight33
796
+ Params for bucket 34 (58720256 elements, 58720256 padded size):
797
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight34
798
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight35
799
+ Params for bucket 35 (58720256 elements, 58720256 padded size):
800
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight36
801
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight37
802
+ Params for bucket 36 (58720256 elements, 58720256 padded size):
803
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight38
804
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight39
805
+ Params for bucket 37 (58720256 elements, 58720256 padded size):
806
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight40
807
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight41
808
+ Params for bucket 38 (58720256 elements, 58720256 padded size):
809
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight42
810
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight43
811
+ Params for bucket 39 (58720256 elements, 58720256 padded size):
812
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight44
813
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight45
814
+ Params for bucket 40 (58720256 elements, 58720256 padded size):
815
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight46
816
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight47
817
+ 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 0x7a3ac1e82e70>, config_logger_dir='')
818
+ INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine
819
+ [after model, optimizer, and learning rate scheduler are built] datetime: 2026-09-21 07:51:35
820
+ > building train, validation, and test datasets ...
821
+ > datasets target sizes (minimum size):
822
+ train: 96
823
+ validation: 0
824
+ test: 0
825
+ INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None]
826
+ > building train, validation, and test datasets for GPT ...
827
+ 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 0x7a381c792bd0>, 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)
828
+ INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx
829
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths
830
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers
831
+ INFO:megatron.core.datasets.indexed_dataset: Extract the document indices
832
+ INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029
833
+ INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350
834
+ INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices
835
+ INFO:megatron.core.datasets.gpt_dataset: Load the document index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-document_index.npy
836
+ INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-sample_index.npy
837
+ INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-shuffle_index.npy
838
+ INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073
839
+ > finished creating GPT datasets ...
840
+ [after dataloaders are built] datetime: 2026-09-21 07:51:35
841
+ done with setup ...
842
+ (min, max) time across ranks (ms):
843
+ model-and-optimizer-setup ......................: (2613.30, 2620.95)
844
+ train/valid/test-data-iterators-setup ..........: (64.82, 75.64)
845
+ training ...
846
+ Overwriting rerun_state_machine.current_iteration from -1 to 0...
847
+ [before the start of training step] datetime: 2026-09-21 07:51:35
848
+ INFO:moonep.api:MoonEP comm buffer sizes:
849
+ config : R=8, S=4096, H=7168, E=384, K=8, NvS=44960, NvS_padded=45056
850
+ sm_count : num_sms=32, num_sms_dedup=148
851
+ hidden_buffer_size : 616.00 MiB
852
+ meta_buffer_size : 2.00 MiB
853
+ local_temp_buffer_size : 1.29 MiB
854
+ total_size : 619.29 MiB
855
+ [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, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 384, 385, 386, 387, 388, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404, 405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419, 420, 421, 422, 423, 424, 425, 426, 427, 428, 429, 430, 431] sizes [384, 384, 256, 128, 384, 1152, 1664, 1152, 1664, 640, 896, 640, 256, 896, 256, 1280, 384, 512, 1664, 1536, 768, 384, 512, 256, 384, 1152, 640, 256, 256, 384, 384, 640, 640, 2048, 384, 640, 384, 256, 768, 1408, 512, 256, 128, 256, 128, 640, 256, 384, 3456, 640, 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, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[348, 357, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [332, 309, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [314, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [76, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [230, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [208, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [76, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
856
+ [2026-09-21 07:52:01] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 25280.1 | throughput per GPU (TFLOP/s/GPU): 1.4 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.227924E+01 | loss scale: 1.0 | grad norm: 12.616 | number of skipped iterations: 0 | number of nan iterations: 0 |Number of parameters in transformer block in billions: 17.55
857
+
858
+ Number of parameters in embedding layers in billions: 0.72
859
+ Total number of parameters in billions: 18.28
860
+ Number of parameters in most loaded shard in billions: 18.2753
861
+ Theoretical memory footprints: weight and optimizer=130715.23 MB
862
+ [Rank 0] (after 1 iterations) memory (MB) | allocated: 45359.1982421875 | max allocated: 45359.19873046875 | reserved: 45790.0 | max reserved: 45790.0
863
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
864
+ warnings.warn(
865
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
866
+ warnings.warn(
867
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
868
+ warnings.warn(
869
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
870
+ warnings.warn(
871
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
872
+ warnings.warn(
873
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
874
+ warnings.warn(
875
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
876
+ warnings.warn(
877
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
878
+ warnings.warn(
879
+ [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, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 384, 385, 386, 387, 388, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404, 405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419, 420, 421, 422, 423, 424, 425, 426, 427, 428, 429, 430, 431] sizes [384, 384, 384, 128, 512, 1152, 1792, 1024, 1536, 768, 768, 640, 256, 512, 256, 1408, 512, 384, 2048, 1280, 768, 512, 512, 256, 512, 1152, 896, 128, 256, 384, 640, 512, 640, 2048, 512, 768, 512, 256, 896, 1536, 640, 128, 128, 256, 128, 768, 256, 512, 3456, 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, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[348, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [357, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [332, 309, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [230, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [76, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [314, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [76, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
880
+ [2026-09-21 07:52:01] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 135.0 | throughput per GPU (TFLOP/s/GPU): 262.1 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.227087E+01 | loss scale: 1.0 | grad norm: 12.533 | number of skipped iterations: 0 | number of nan iterations: 0 |
881
+ [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, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 384, 385, 386, 387, 388, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404, 405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419, 420, 421, 422, 423, 424, 425, 426, 427, 428, 429, 430, 431] sizes [256, 512, 256, 256, 512, 896, 1920, 1024, 2176, 896, 1536, 640, 256, 640, 384, 640, 512, 640, 1408, 896, 768, 512, 512, 256, 512, 1152, 768, 256, 384, 384, 640, 512, 512, 1408, 512, 896, 512, 256, 768, 1024, 640, 256, 256, 384, 128, 384, 256, 768, 3456, 384, 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, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[313, 309, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [208, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [338, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [230, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [306, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [76, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [332, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
882
+ [2026-09-21 07:52:01] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 110.9 | throughput per GPU (TFLOP/s/GPU): 318.9 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 1.088362E+01 | loss scale: 1.0 | grad norm: 11.029 | number of skipped iterations: 0 | number of nan iterations: 0 |
883
+ Capture CUDA graph for training!!!
884
+ CUDA graph capture done!!!
885
+ [2026-09-21 07:52:01] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 553.4 | throughput per GPU (TFLOP/s/GPU): 63.9 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 1.448777E+01 | loss scale: 1.0 | grad norm: 72.970 | number of skipped iterations: 0 | number of nan iterations: 0 |
886
+ [2026-09-21 07:52:02] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 108.0 | throughput per GPU (TFLOP/s/GPU): 327.7 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.808960E+01 | loss scale: 1.0 | grad norm: 108.333 | number of skipped iterations: 0 | number of nan iterations: 0 |
887
+ [2026-09-21 07:52:02] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 108.4 | throughput per GPU (TFLOP/s/GPU): 326.3 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 1.490298E+01 | loss scale: 1.0 | grad norm: 83.291 | number of skipped iterations: 0 | number of nan iterations: 0 |
888
+ [2026-09-21 07:52:02] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 109.6 | throughput per GPU (TFLOP/s/GPU): 322.7 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 1.142665E+01 | loss scale: 1.0 | grad norm: 26.340 | number of skipped iterations: 0 | number of nan iterations: 0 |
889
+ [2026-09-21 07:52:02] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 108.4 | throughput per GPU (TFLOP/s/GPU): 326.4 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.165925E+01 | loss scale: 1.0 | grad norm: 12.676 | number of skipped iterations: 0 | number of nan iterations: 0 |
890
+ [2026-09-21 07:52:02] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 109.0 | throughput per GPU (TFLOP/s/GPU): 324.6 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.132117E+01 | loss scale: 1.0 | grad norm: 10.523 | number of skipped iterations: 0 | number of nan iterations: 0 |
891
+ [2026-09-21 07:52:02] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 108.6 | throughput per GPU (TFLOP/s/GPU): 325.7 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.097885E+01 | loss scale: 1.0 | grad norm: 9.172 | number of skipped iterations: 0 | number of nan iterations: 0 |
892
+ [2026-09-21 07:52:02] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 108.5 | throughput per GPU (TFLOP/s/GPU): 326.2 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.072919E+01 | loss scale: 1.0 | grad norm: 9.283 | number of skipped iterations: 0 | number of nan iterations: 0 |
893
+ [2026-09-21 07:52:02] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 107.7 | throughput per GPU (TFLOP/s/GPU): 328.3 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.047170E+01 | loss scale: 1.0 | grad norm: 9.277 | number of skipped iterations: 0 | number of nan iterations: 0 |
894
+ [after training is done] datetime: 2026-09-21 07:52:02
895
+ [rank0]:[W921 07:52:04.067116303 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())
896
+ [rank3]:[W921 07:52:04.885154135 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())
897
+ [rank4]:[W921 07:52:05.371335095 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())
898
+ [rank1]:[W921 07:52:05.377021465 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())
899
+ [rank5]:[W921 07:52:05.588258200 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())
900
+ [rank7]:[W921 07:52:05.710176106 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())
901
+ [rank2]:[W921 07:52:05.842622315 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())
902
+ [rank6]:[W921 07:52:05.910178863 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())
903
+ W0921 07:52:10.200000 530811 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers
904
+ W0921 07:52:10.201000 530811 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 531002 closing signal SIGTERM
905
+ W0921 07:52:10.201000 530811 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 531003 closing signal SIGTERM
906
+ W0921 07:52:10.201000 530811 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 531005 closing signal SIGTERM
907
+ W0921 07:52:10.202000 530811 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 531008 closing signal SIGTERM
908
+ Traceback (most recent call last):
909
+ File "/usr/local/bin/torchrun", line 8, in <module>
910
+ sys.exit(main())
911
+ ^^^^^^
912
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
913
+ return f(*args, **kwargs)
914
+ ^^^^^^^^^^^^^^^^^^
915
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 892, in main
916
+ run(args)
917
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 883, in run
918
+ elastic_launch(
919
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
920
+ return launch_agent(self._config, self._entrypoint, list(args))
921
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
922
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 261, in launch_agent
923
+ result = agent.run()
924
+ ^^^^^^^^^^^
925
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper
926
+ result = f(*args, **kwargs)
927
+ ^^^^^^^^^^^^^^^^^^
928
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 711, in run
929
+ result = self._invoke_run(role)
930
+ ^^^^^^^^^^^^^^^^^^^^^^
931
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 870, in _invoke_run
932
+ time.sleep(monitor_interval)
933
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 84, in _terminate_process_handler
934
+ raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval)
935
+ torch.distributed.elastic.multiprocessing.api.SignalException: Process 530811 got signal: 15
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep2/mem_samples.csv ADDED
@@ -0,0 +1,120 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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, 1
25
+ 0, 1
26
+ 1, 1
27
+ 2, 1
28
+ 3, 1
29
+ 4, 1
30
+ 5, 1
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, 1
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, 4
69
+ 4, 4
70
+ 5, 629
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, 2527
90
+ 1, 2527
91
+ 2, 2527
92
+ 3, 2527
93
+ 4, 2527
94
+ 5, 2527
95
+ 6, 2527
96
+ 7, 2527
97
+ 0, 48843
98
+ 1, 48861
99
+ 2, 48861
100
+ 3, 48861
101
+ 4, 48861
102
+ 5, 48861
103
+ 6, 48861
104
+ 7, 48861
105
+ 0, 57941
106
+ 1, 57943
107
+ 2, 57941
108
+ 3, 57943
109
+ 4, 57941
110
+ 5, 58559
111
+ 6, 57941
112
+ 7, 57941
113
+ 0, 59207
114
+ 1, 59209
115
+ 2, 59207
116
+ 3, 59207
117
+ 4, 59207
118
+ 5, 59207
119
+ 6, 59207
120
+ 7, 59207
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep2/run.env ADDED
@@ -0,0 +1,105 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=moonep
2
+ config=KIMI_K2_L2
3
+ family=tierB
4
+ source_model=Kimi-K2
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=2
14
+ hidden=7168
15
+ ffn_hidden=18432
16
+ moe_ffn_hidden=2048
17
+ num_experts=384
18
+ topk=8
19
+ experts_per_rank=48
20
+ expert_params=44040192
21
+ heads=64
22
+ kv_groups=64
23
+ kv_channels=64
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=1
35
+ moe_layers=1
36
+ rotary_base=50000
37
+ rotary_percent=1.0
38
+ rotary_scaling=32
39
+ mscale=1.0
40
+ mscale_all_dim=1.0
41
+ qkv_bias=0
42
+ router_balancing=seq_aux_loss
43
+ aux_loss_coeff=0
44
+ router_pre_softmax=0
45
+ router_fusion=0
46
+ router_dtype=fp32
47
+ router_groups=1
48
+ router_group_topk=1
49
+ router_score_function=sigmoid
50
+ router_expert_bias=1
51
+ router_bias_update_rate=1e-3
52
+ router_topk_scaling_factor=2.827
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=682
76
+ slot_policy=matched
77
+ slots_per_rank=48
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/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep2/train.log ADDED
@@ -0,0 +1,854 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [07:52:21] WARN: moonep_backend_async_grad_reduce lacks --moe-router-fusion; KIMI_K2_L2 runs unfused (timing only, routing identical)
2
+ ==============================================================
3
+ KIMI_K2_L2 / moonep (slots/rank=48 policy=matched)
4
+ E=384 K=8 H'=2048 m=48
5
+ L=2 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/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep2/memory_snapshots/snapshot.pickle
8
+ repo=/workspace/Megatron-MoE-Echo @ 2e68c8ad
9
+ ==============================================================
10
+ W0921 07:52:23.044000 533518 torch/distributed/run.py:766]
11
+ W0921 07:52:23.044000 533518 torch/distributed/run.py:766] *****************************************
12
+ W0921 07:52:23.044000 533518 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 07:52:23.044000 533518 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
+ 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
17
+ WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer
18
+ Number of virtual stages per pipeline stage: None
19
+ accumulate and all-reduce gradients in fp32 for bfloat16 data type.
20
+ using torch.bfloat16 for parameters ...
21
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
22
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
23
+ ------------------------ arguments ------------------------
24
+ account_for_embedding_in_pipeline_split ......... False
25
+ account_for_loss_in_pipeline_split .............. False
26
+ accumulate_allreduce_grads_in_fp32 .............. True
27
+ adam_beta1 ...................................... 0.9
28
+ adam_beta2 ...................................... 0.999
29
+ adam_eps ........................................ 1e-08
30
+ add_bias_linear ................................. False
31
+ add_position_embedding .......................... False
32
+ add_qkv_bias .................................... False
33
+ adlr_autoresume ................................. False
34
+ adlr_autoresume_interval ........................ 1000
35
+ align_grad_reduce ............................... True
36
+ align_param_gather .............................. False
37
+ app_tag_run_name ................................ None
38
+ app_tag_run_version ............................. 0.0.0
39
+ apply_layernorm_1p .............................. False
40
+ apply_query_key_layer_scaling ................... False
41
+ apply_residual_connection_post_layernorm ........ False
42
+ apply_rope_fusion ............................... False
43
+ async_save ...................................... None
44
+ async_tensor_model_parallel_allreduce ........... True
45
+ attention_backend ............................... AttnBackend.auto
46
+ attention_dropout ............................... 0.0
47
+ attention_softmax_in_fp32 ....................... False
48
+ auto_detect_ckpt_format ......................... False
49
+ barrier_with_L1_time ............................ True
50
+ bert_binary_head ................................ True
51
+ bert_embedder_type .............................. megatron
52
+ bert_load ....................................... None
53
+ bf16 ............................................ True
54
+ bias_dropout_fusion ............................. True
55
+ bias_gelu_fusion ................................ False
56
+ bias_swiglu_fusion .............................. True
57
+ biencoder_projection_dim ........................ 0
58
+ biencoder_shared_query_context_model ............ False
59
+ block_data_path ................................. None
60
+ calc_ft_timeouts ................................ False
61
+ calculate_per_token_loss ........................ False
62
+ check_for_large_grads ........................... False
63
+ check_for_nan_in_loss_and_grad .................. False
64
+ check_for_spiky_loss ............................ False
65
+ check_weight_hash_across_dp_replicas_interval ... None
66
+ ckpt_assume_constant_structure .................. False
67
+ ckpt_convert_format ............................. None
68
+ ckpt_convert_save ............................... None
69
+ ckpt_convert_update_legacy_dist_opt_format ...... False
70
+ ckpt_format ..................................... torch_dist
71
+ ckpt_fully_parallel_load ........................ False
72
+ ckpt_fully_parallel_save ........................ True
73
+ ckpt_fully_parallel_save_deprecated ............. False
74
+ ckpt_step ....................................... None
75
+ classes_fraction ................................ 1.0
76
+ clip_grad ....................................... 1.0
77
+ clone_scatter_output_in_embedding ............... True
78
+ config_logger_dir ...............................
79
+ consumed_train_samples .......................... 0
80
+ consumed_valid_samples .......................... 0
81
+ context_parallel_size ........................... 1
82
+ cp_comm_type .................................... ['p2p']
83
+ create_attention_mask_in_dataloader ............. True
84
+ cross_entropy_fusion_impl ....................... native
85
+ cross_entropy_loss_fusion ....................... False
86
+ cuda_graph_scope ................................ full_iteration
87
+ cuda_graph_warmup_steps ......................... 3
88
+ data_args_path .................................. None
89
+ data_cache_path ................................. None
90
+ data_parallel_random_init ....................... False
91
+ data_parallel_sharding_strategy ................. no_shard
92
+ data_parallel_size .............................. 8
93
+ data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document']
94
+ data_per_class_fraction ......................... 1.0
95
+ data_sharding ................................... True
96
+ dataloader_type ................................. single
97
+ ddp_average_in_collective ....................... False
98
+ ddp_bucket_size ................................. None
99
+ ddp_num_buckets ................................. None
100
+ ddp_pad_buckets_for_high_nccl_busbw ............. False
101
+ decoder_first_pipeline_num_layers ............... None
102
+ decoder_last_pipeline_num_layers ................ None
103
+ decoder_num_layers .............................. None
104
+ decoder_seq_length .............................. None
105
+ decoupled_lr .................................... None
106
+ decoupled_min_lr ................................ None
107
+ decrease_batch_size_if_needed ................... False
108
+ defer_embedding_wgrad_compute ................... False
109
+ delay_wgrad_compute ............................. False
110
+ deprecated_use_mcore_models ..................... True
111
+ deterministic_mode .............................. False
112
+ dino_bottleneck_size ............................ 256
113
+ dino_freeze_last_layer .......................... 1
114
+ dino_head_hidden_size ........................... 2048
115
+ dino_local_crops_number ......................... 10
116
+ dino_local_img_size ............................. 96
117
+ dino_norm_last_layer ............................ False
118
+ dino_teacher_temp ............................... 0.07
119
+ dino_warmup_teacher_temp ........................ 0.04
120
+ dino_warmup_teacher_temp_epochs ................. 30
121
+ disable_bf16_reduced_precision_matmul ........... False
122
+ disable_mamba_mem_eff_path ...................... False
123
+ disable_straggler_on_startup .................... False
124
+ dist_ckpt_format_deprecated ..................... None
125
+ dist_ckpt_strictness ............................ assume_ok_unexpected
126
+ distribute_saved_activations .................... False
127
+ distributed_backend ............................. nccl
128
+ distributed_timeout_minutes ..................... 10
129
+ embedding_init_method_std ....................... None
130
+ embedding_path .................................. None
131
+ empty_unused_memory_level ....................... 0
132
+ enable_cuda_graph ............................... True
133
+ enable_experimental ............................. False
134
+ enable_ft_package ............................... False
135
+ enable_gloo_process_groups ...................... True
136
+ enable_msc ...................................... True
137
+ enable_one_logger ............................... True
138
+ encoder_num_layers .............................. 2
139
+ encoder_seq_length .............................. 4096
140
+ end_weight_decay ................................ 0.1
141
+ eod_mask_loss ................................... False
142
+ error_injection_rate ............................ 0
143
+ error_injection_type ............................ transient_error
144
+ eval_interval ................................... 1000000
145
+ eval_iters ...................................... 0
146
+ evidence_data_path .............................. None
147
+ exit_duration_in_mins ........................... None
148
+ exit_interval ................................... None
149
+ exit_on_missing_checkpoint ...................... False
150
+ exit_signal_handler ............................. False
151
+ exp_avg_dtype ................................... torch.float32
152
+ exp_avg_sq_dtype ................................ torch.float32
153
+ expert_model_parallel_size ...................... 8
154
+ expert_tensor_parallel_size ..................... 1
155
+ export_force_local_attention .................... False
156
+ export_kd_cfg ................................... None
157
+ export_kd_teacher_ckpt_format ................... None
158
+ export_kd_teacher_load .......................... None
159
+ export_kv_cache_quant ........................... False
160
+ export_legacy_megatron .......................... False
161
+ export_model_type ............................... GPTModel
162
+ export_moe_apply_probs_on_input ................. False
163
+ export_qk_l2_norm ............................... False
164
+ export_quant_cfg ................................ None
165
+ export_real_quant_cfg ........................... None
166
+ export_te_mcore_model ........................... False
167
+ external_cuda_graph ............................. False
168
+ ffn_hidden_size ................................. 18432
169
+ finetune ........................................ False
170
+ finetune_data_split ............................. train
171
+ finetune_hf_dataset ............................. None
172
+ first_last_layers_bf16 .......................... False
173
+ flash_decode .................................... False
174
+ fp16 ............................................ False
175
+ fp16_lm_cross_entropy ........................... False
176
+ fp32_residual_connection ........................ False
177
+ fp8 ............................................. e4m3
178
+ fp8_amax_compute_algo ........................... most_recent
179
+ fp8_amax_history_len ............................ 1
180
+ fp8_interval .................................... 1
181
+ fp8_margin ...................................... 0
182
+ fp8_param_gather ................................ False
183
+ fp8_recipe ...................................... mxfp8
184
+ fp8_wgrad ....................................... True
185
+ fsdp_double_buffer .............................. False
186
+ global_batch_size ............................... 8
187
+ grad_reduce_in_bf16 ............................. False
188
+ gradient_accumulation_fusion .................... True
189
+ gradient_reduce_div_fusion ...................... True
190
+ group_query_attention ........................... False
191
+ head_lr_mult .................................... 1.0
192
+ heterogeneous_layers_config_encoded_json ........ None
193
+ heterogeneous_layers_config_path ................ None
194
+ hidden_dropout .................................. 0.0
195
+ hidden_size ..................................... 7168
196
+ hierarchical_context_parallel_sizes ............. None
197
+ high_priority_stream_groups ..................... []
198
+ hybrid_attention_ratio .......................... 0.0
199
+ hybrid_mlp_ratio ................................ 0.0
200
+ hybrid_override_pattern ......................... None
201
+ hysteresis ...................................... 2
202
+ ict_head_size ................................... None
203
+ ict_load ........................................ None
204
+ img_h ........................................... 224
205
+ img_w ........................................... 224
206
+ indexer_batch_size .............................. 128
207
+ indexer_log_interval ............................ 1000
208
+ inference_batch_times_seqlen_threshold .......... -1
209
+ inference_dynamic_batching ...................... False
210
+ inference_dynamic_batching_buffer_guaranteed_fraction 0.2
211
+ inference_dynamic_batching_buffer_overflow_factor None
212
+ inference_dynamic_batching_buffer_size_gb ....... 40.0
213
+ inference_dynamic_batching_chunk_size ........... 256
214
+ inference_dynamic_batching_max_requests_override None
215
+ inference_dynamic_batching_max_tokens_override .. None
216
+ inference_dynamic_batching_num_cuda_graphs ...... 16
217
+ inference_max_batch_size ........................ 8
218
+ inference_max_seq_length ........................ 2560
219
+ inference_rng_tracker ........................... False
220
+ init_method_std ................................. 0.02
221
+ init_method_xavier_uniform ...................... False
222
+ init_model_with_meta_device ..................... False
223
+ initial_loss_scale .............................. 4294967296
224
+ inprocess_active_world_size ..................... 8
225
+ inprocess_barrier_timeout ....................... 120
226
+ inprocess_completion_timeout .................... 120
227
+ inprocess_empty_cuda_cache ...................... False
228
+ inprocess_granularity ........................... node
229
+ inprocess_hard_timeout .......................... 90
230
+ inprocess_heartbeat_interval .................... 30
231
+ inprocess_heartbeat_timeout ..................... 60
232
+ inprocess_last_call_wait ........................ 1
233
+ inprocess_max_iterations ........................ None
234
+ inprocess_monitor_process_interval .............. 1.0
235
+ inprocess_monitor_thread_interval ............... 1.0
236
+ inprocess_progress_watchdog_interval ............ 1.0
237
+ inprocess_restart ............................... False
238
+ inprocess_soft_timeout .......................... 60
239
+ inprocess_termination_grace_time ................ 1
240
+ is_hybrid_model ................................. False
241
+ iter_per_epoch .................................. 1250
242
+ iterations_to_skip .............................. []
243
+ keep_fp8_transpose_cache_when_using_custom_fsdp . False
244
+ kitchen_config_file ............................. None
245
+ kitchen_recipe_number ........................... None
246
+ kv_channels ..................................... 64
247
+ kv_lora_rank .................................... 512
248
+ lazy_mpu_init ................................... None
249
+ load ............................................ None
250
+ load_main_params_from_ckpt ...................... None
251
+ load_model_opt_format ........................... False
252
+ local_rank ...................................... 0
253
+ log_energy ...................................... False
254
+ log_interval .................................... 1
255
+ log_loss_scale_to_tensorboard ................... True
256
+ log_memory_to_tensorboard ....................... False
257
+ log_num_zeros_in_grad ........................... False
258
+ log_params_norm ................................. False
259
+ log_progress .................................... False
260
+ log_straggler ................................... False
261
+ log_throughput .................................. True
262
+ log_timers_to_tensorboard ....................... False
263
+ log_validation_ppl_to_tensorboard ............... False
264
+ log_world_size_to_tensorboard ................... False
265
+ logging_level ................................... None
266
+ loss_scale ...................................... None
267
+ loss_scale_window ............................... 1000
268
+ lr .............................................. 0.0001
269
+ lr_decay_iters .................................. 8
270
+ lr_decay_samples ................................ None
271
+ lr_decay_style .................................. cosine
272
+ lr_warmup_fraction .............................. None
273
+ lr_warmup_init .................................. 0.0
274
+ lr_warmup_iters ................................. 5
275
+ lr_warmup_samples ............................... 0
276
+ lr_wsd_decay_iters .............................. None
277
+ lr_wsd_decay_samples ............................ None
278
+ lr_wsd_decay_style .............................. exponential
279
+ main_grads_dtype ................................ torch.float32
280
+ main_params_dtype ............................... torch.float32
281
+ make_vocab_size_divisible_by .................... 128
282
+ mamba_head_dim .................................. 64
283
+ mamba_num_groups ................................ 8
284
+ mamba_num_heads ................................. None
285
+ mamba_state_dim ................................. 128
286
+ manual_gc ....................................... False
287
+ manual_gc_eval .................................. True
288
+ manual_gc_interval .............................. 0
289
+ mask_factor ..................................... 1.0
290
+ mask_prob ....................................... 0.15
291
+ mask_type ....................................... random
292
+ masked_softmax_fusion ........................... False
293
+ max_position_embeddings ......................... 4096
294
+ max_tokens_to_oom ............................... 12000
295
+ memory_snapshot_path ............................ snapshot.pickle
296
+ merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt
297
+ micro_batch_size ................................ 1
298
+ microbatch_group_size_per_vp_stage .............. None
299
+ mid_level_dataset_surplus ....................... 0.005
300
+ min_loss_scale .................................. 1.0
301
+ min_lr .......................................... 1e-05
302
+ mlp_chunks_for_prefill .......................... 1
303
+ mmap_bin_files .................................. True
304
+ mock_data ....................................... False
305
+ moe_apply_probs_on_input ........................ False
306
+ moe_aux_loss_coeff .............................. 0.0
307
+ moe_deepep_num_sms .............................. 20
308
+ moe_echo_enable_random_offloading ............... False
309
+ moe_echo_expert_dispatch_overlap ................ False
310
+ moe_echo_expert_dispatcher_type ................. hybridep
311
+ moe_echo_recompute_expert_dispatch .............. False
312
+ moe_enable_deepep ............................... False
313
+ moe_enable_echo ................................. False
314
+ moe_enable_hybridep ............................. False
315
+ moe_enable_moonep ............................... True
316
+ moe_expert_capacity_factor ...................... None
317
+ moe_extended_tp ................................. False
318
+ moe_ffn_hidden_size ............................. 2048
319
+ moe_grouped_gemm ................................ True
320
+ moe_hybridep_num_sms ............................ 16
321
+ moe_input_jitter_eps ............................ None
322
+ moe_layer_freq .................................. [0, 1]
323
+ moe_layer_recompute ............................. False
324
+ moe_moonep_async_grad_reduce .................... True
325
+ moe_moonep_num_sms .............................. None
326
+ moe_num_echo_experts ............................ None
327
+ moe_pad_expert_input_to_capacity ................ False
328
+ moe_per_layer_logging ........................... False
329
+ moe_permute_fusion .............................. False
330
+ moe_received_token_capacity ..................... None
331
+ moe_router_bias_update_rate ..................... 0.001
332
+ moe_router_dtype ................................ fp32
333
+ moe_router_enable_expert_bias ................... True
334
+ moe_router_force_load_balancing ................. False
335
+ moe_router_group_topk ........................... 1
336
+ moe_router_load_balancing_type .................. seq_aux_loss
337
+ moe_router_num_groups ........................... 1
338
+ moe_router_padding_for_fp8 ...................... False
339
+ moe_router_pre_softmax .......................... False
340
+ moe_router_score_function ....................... sigmoid
341
+ moe_router_topk ................................. 8
342
+ moe_router_topk_scaling_factor .................. 2.827
343
+ moe_shared_expert_intermediate_size ............. 2048
344
+ moe_shared_expert_overlap ....................... False
345
+ moe_token_dispatcher_type ....................... flex
346
+ moe_token_drop_policy ........................... probs
347
+ moe_upcycling_granularity ....................... 1
348
+ moe_use_device_initiated_grouped_gemm ........... True
349
+ moe_use_legacy_grouped_gemm ..................... False
350
+ moe_use_upcycling ............................... False
351
+ moe_z_loss_coeff ................................ None
352
+ mrope_section ................................... None
353
+ mscale .......................................... 1.0
354
+ mscale_all_dim .................................. 1.0
355
+ mtp_loss_scaling_factor ......................... 0.1
356
+ mtp_num_layers .................................. None
357
+ multi_latent_attention .......................... True
358
+ nccl_all_reduce_for_prefill ..................... False
359
+ nccl_communicator_config_path ................... None
360
+ nccl_ub ......................................... False
361
+ no_load_optim ................................... None
362
+ no_load_rng ..................................... None
363
+ no_persist_layer_norm ........................... False
364
+ no_rope_freq .................................... None
365
+ no_save_optim ................................... None
366
+ no_save_rng ..................................... None
367
+ non_persistent_ckpt_type ........................ None
368
+ non_persistent_global_ckpt_dir .................. None
369
+ non_persistent_local_ckpt_algo .................. fully_parallel
370
+ non_persistent_local_ckpt_dir ................... None
371
+ non_persistent_save_interval .................... None
372
+ norm_epsilon .................................... 1e-06
373
+ normalization ................................... RMSNorm
374
+ num_attention_heads ............................. 64
375
+ num_channels .................................... 3
376
+ num_classes ..................................... 1000
377
+ num_dataset_builder_threads ..................... 1
378
+ num_distributed_optimizer_instances ............. 1
379
+ num_experts ..................................... 384
380
+ num_layers ...................................... 2
381
+ num_layers_at_end_in_bf16 ....................... 1
382
+ num_layers_at_start_in_bf16 ..................... 1
383
+ num_layers_per_virtual_pipeline_stage ........... None
384
+ num_query_groups ................................ 1
385
+ num_virtual_stages_per_pipeline_rank ............ None
386
+ num_workers ..................................... 2
387
+ object_storage_cache_path ....................... None
388
+ one_logger_async ................................ False
389
+ one_logger_project .............................. megatron-lm
390
+ one_logger_run_name ............................. None
391
+ onnx_safe ....................................... None
392
+ openai_gelu ..................................... False
393
+ optimizer ....................................... adam
394
+ optimizer_cpu_offload ........................... False
395
+ optimizer_offload_fraction ...................... 1.0
396
+ output_bert_embeddings .......................... False
397
+ overlap_cpu_optimizer_d2h_h2d ................... False
398
+ overlap_grad_reduce ............................. True
399
+ overlap_moe_expert_parallel_comm ................ False
400
+ overlap_p2p_comm ................................ False
401
+ overlap_p2p_comm_warmup_flush ................... False
402
+ overlap_param_gather ............................ True
403
+ overlap_param_gather_with_optimizer_step ........ False
404
+ override_opt_param_scheduler .................... False
405
+ params_dtype .................................... torch.bfloat16
406
+ patch_dim ....................................... 16
407
+ per_split_data_args_path ........................ None
408
+ perform_initialization .......................... True
409
+ pin_cpu_grads ................................... True
410
+ pin_cpu_params .................................. True
411
+ pipeline_model_parallel_comm_backend ............ None
412
+ pipeline_model_parallel_layout .................. None
413
+ pipeline_model_parallel_size .................... 1
414
+ position_embedding_type ......................... rope
415
+ pretrained_checkpoint ........................... None
416
+ profile ......................................... False
417
+ profile_ranks ................................... [0]
418
+ profile_step_end ................................ 12
419
+ profile_step_start .............................. 10
420
+ q_lora_rank ..................................... 1536
421
+ qk_head_dim ..................................... 128
422
+ qk_l2_norm ...................................... False
423
+ qk_layernorm .................................... True
424
+ qk_pos_emb_head_dim ............................. 64
425
+ query_in_block_prob ............................. 0.1
426
+ rampup_batch_size ............................... None
427
+ rank ............................................ 0
428
+ recompute_granularity ........................... None
429
+ recompute_method ................................ None
430
+ recompute_modules ............................... None
431
+ recompute_num_layers ............................ None
432
+ record_memory_history ........................... False
433
+ relative_attention_max_distance ................. 128
434
+ relative_attention_num_buckets .................. 32
435
+ replication ..................................... False
436
+ replication_factor .............................. 2
437
+ replication_jump ................................ None
438
+ rerun_mode ...................................... validate_results
439
+ reset_attention_mask ............................ False
440
+ reset_position_ids .............................. False
441
+ result_rejected_tracker_filename ................ None
442
+ retriever_report_topk_accuracies ................ []
443
+ retriever_score_scaling ......................... False
444
+ retriever_seq_length ............................ 256
445
+ retro_add_retriever ............................. False
446
+ retro_attention_gate ............................ 1
447
+ retro_cyclic_train_iters ........................ None
448
+ retro_encoder_attention_dropout ................. 0.1
449
+ retro_encoder_hidden_dropout .................... 0.1
450
+ retro_encoder_layers ............................ 2
451
+ retro_num_neighbors ............................. 2
452
+ retro_num_retrieved_chunks ...................... 2
453
+ retro_project_dir ............................... None
454
+ retro_verify_neighbor_count ..................... True
455
+ reuse_grad_buf_for_mxfp8_param_ag ............... False
456
+ rope_scaling_factor ............................. 8.0
457
+ rope_type ....................................... None
458
+ rotary_base ..................................... 50000
459
+ rotary_interleaved .............................. False
460
+ rotary_percent .................................. 1.0
461
+ rotary_scaling_factor ........................... 32.0
462
+ rotary_seq_len_interpolation_factor ............. None
463
+ run_workload_inspector_server ................... False
464
+ sample_rate ..................................... 1.0
465
+ save ............................................ None
466
+ save_interval ................................... None
467
+ save_retain_interval ............................ None
468
+ scatter_gather_tensors_in_pipeline .............. True
469
+ seed ............................................ 1234
470
+ seq_length ...................................... 4096
471
+ sequence_parallel ............................... False
472
+ sft ............................................. False
473
+ sft_tokenizer_prompt_format ..................... nemotron-h-aligned
474
+ sgd_momentum .................................... 0.9
475
+ short_seq_prob .................................. 0.1
476
+ skip_train ...................................... False
477
+ skipped_train_samples ........................... 0
478
+ spec ............................................ None
479
+ split ........................................... 100,0,0
480
+ squared_relu .................................... False
481
+ start_weight_decay .............................. 0.1
482
+ straggler_ctrlr_port ............................ 65535
483
+ straggler_minmax_count .......................... 1
484
+ suggested_communication_unit_size ............... None
485
+ swiglu .......................................... True
486
+ swin_backbone_type .............................. tiny
487
+ symmetric_ar_type ............................... None
488
+ te_rng_tracker .................................. True
489
+ tensor_model_parallel_size ...................... 1
490
+ tensorboard_dir ................................. None
491
+ tensorboard_log_interval ........................ 1
492
+ tensorboard_queue_size .......................... 1000
493
+ test_data_path .................................. None
494
+ test_mode ....................................... False
495
+ tiktoken_num_special_tokens ..................... 1000
496
+ tiktoken_pattern ................................ None
497
+ tiktoken_special_tokens ......................... None
498
+ timing_log_level ................................ 0
499
+ timing_log_option ............................... minmax
500
+ titles_data_path ................................ None
501
+ tokenizer_model ................................. None
502
+ tokenizer_type .................................. GPT2BPETokenizer
503
+ torch_fsdp2_reshard_after_forward ............... True
504
+ tp_comm_bootstrap_backend ....................... nccl
505
+ tp_comm_bulk_dgrad .............................. True
506
+ tp_comm_bulk_wgrad .............................. True
507
+ tp_comm_overlap ................................. False
508
+ tp_comm_overlap_ag .............................. True
509
+ tp_comm_overlap_cfg ............................. None
510
+ tp_comm_overlap_rs .............................. True
511
+ tp_comm_overlap_rs_dgrad ........................ False
512
+ tp_comm_split_ag ................................ True
513
+ tp_comm_split_rs ................................ True
514
+ train_data_path ................................. None
515
+ train_iters ..................................... 12
516
+ train_samples ................................... None
517
+ train_sync_interval ............................. None
518
+ transformer_impl ................................ transformer_engine
519
+ transformer_pipeline_model_parallel_size ........ 1
520
+ untie_embeddings_and_output_weights ............. True
521
+ use_checkpoint_args ............................. False
522
+ use_checkpoint_opt_param_scheduler .............. False
523
+ use_cpu_initialization .......................... None
524
+ use_custom_fsdp ................................. False
525
+ use_dist_ckpt ................................... True
526
+ use_dist_ckpt_deprecated ........................ False
527
+ use_distributed_optimizer ....................... True
528
+ use_flash_attn .................................. False
529
+ use_legacy_models ............................... False
530
+ use_mp_args_from_checkpoint_args ................ False
531
+ use_one_sent_docs ............................... False
532
+ use_persistent_ckpt_worker ...................... False
533
+ use_precision_aware_optimizer ................... False
534
+ use_pytorch_profiler ............................ False
535
+ use_ring_exchange_p2p ........................... False
536
+ use_rope_scaling ................................ False
537
+ use_rotary_position_embeddings .................. False
538
+ use_sharp ....................................... False
539
+ use_tokenizer_model_from_checkpoint_args ........ True
540
+ use_torch_fsdp2 ................................. False
541
+ use_torch_optimizer_for_cpu_offload ............. False
542
+ use_tp_pp_dp_mapping ............................ False
543
+ v_head_dim ...................................... 128
544
+ valid_data_path ................................. None
545
+ variable_seq_lengths ............................ False
546
+ virtual_pipeline_model_parallel_size ............ None
547
+ vision_backbone_type ............................ vit
548
+ vision_pretraining .............................. False
549
+ vision_pretraining_type ......................... classify
550
+ vocab_extra_ids ................................. 0
551
+ vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json
552
+ vocab_size ...................................... None
553
+ wandb_exp_name ..................................
554
+ wandb_project ...................................
555
+ wandb_save_dir ..................................
556
+ weight_decay .................................... 0.1
557
+ weight_decay_incr_style ......................... constant
558
+ wgrad_deferral_limit ............................ 0
559
+ world_size ...................................... 8
560
+ yaml_cfg ........................................ None
561
+ -------------------- end of arguments ---------------------
562
+ INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1
563
+ > building GPT2BPETokenizer tokenizer ...
564
+ > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304)
565
+ WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS
566
+ > initializing torch distributed ...
567
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
568
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
569
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
570
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
571
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
572
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
573
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
574
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
575
+ /workspace/Megatron-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 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
582
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
583
+ [Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
584
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
585
+ [Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
586
+ [Gloo] Rank [Gloo] Rank 7 is connected to 67 is connected to peer ranks. 7Expected number of connected peer ranks is : peer ranks. 7Expected number of connected peer ranks is : 7
587
+
588
+ [Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
589
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
590
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
591
+ [Gloo] Rank 2 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 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
594
+ [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 :
595
+ 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 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
+ [rank3]:[W921 07:52:40.529101085 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.
608
+ [rank1]:[W921 07:52:40.529105283 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.
609
+ [rank7]:[W921 07:52:40.529105168 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.
610
+ [rank5]:[W921 07:52:40.529131675 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 5] using GPU 5 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
611
+ [rank4]:[W921 07:52:40.534117045 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.
612
+ [rank6]:[W921 07:52:40.535462095 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 6] using GPU 6 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
613
+ [rank2]:[W921 07:52:40.535462979 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 2] using GPU 2 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
614
+ > 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 07:52:40.705851123 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: 4.020 seconds
623
+ time to initialize megatron (seconds): 7.810
624
+ [after megatron is initialized] datetime: 2026-09-21 07:52:47
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): 3480525824
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: 9
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.1.mlp.shared_experts.linear_fc2.weight
650
+ module.decoder.layers.1.mlp.shared_experts.linear_fc1.weight
651
+ Params for bucket 3 (45163520 elements, 45163520 padded size):
652
+ module.decoder.layers.1.pre_mlp_layernorm.weight
653
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.weight
654
+ module.decoder.layers.1.mlp.router.weight
655
+ module.decoder.layers.1.self_attention.linear_kv_down_proj.weight
656
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.layer_norm_weight
657
+ module.decoder.layers.1.self_attention.linear_q_up_proj.weight
658
+ module.decoder.layers.1.self_attention.linear_q_up_proj.layer_norm_weight
659
+ module.decoder.layers.1.self_attention.linear_q_down_proj.weight
660
+ Params for bucket 4 (58720256 elements, 58720256 padded size):
661
+ module.decoder.layers.1.self_attention.linear_proj.weight
662
+ Params for bucket 5 (132127744 elements, 132127744 padded size):
663
+ module.decoder.layers.1.input_layernorm.weight
664
+ module.decoder.layers.0.mlp.linear_fc2.weight
665
+ Params for bucket 6 (264241152 elements, 264241152 padded size):
666
+ module.decoder.layers.0.mlp.linear_fc1.weight
667
+ Params for bucket 7 (42411008 elements, 42411008 padded size):
668
+ module.decoder.layers.0.mlp.linear_fc1.layer_norm_weight
669
+ module.decoder.layers.0.self_attention.linear_q_up_proj.weight
670
+ module.decoder.layers.0.self_attention.linear_kv_down_proj.weight
671
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.weight
672
+ module.decoder.layers.0.self_attention.linear_q_down_proj.weight
673
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.layer_norm_weight
674
+ module.decoder.layers.0.self_attention.linear_q_up_proj.layer_norm_weight
675
+ Params for bucket 8 (58720256 elements, 58720256 padded size):
676
+ module.decoder.layers.0.self_attention.linear_proj.weight
677
+ Params for bucket 9 (360586240 elements, 360586240 padded size):
678
+ module.decoder.layers.0.input_layernorm.weight
679
+ module.embedding.word_embeddings.weight
680
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 40
681
+ Params for bucket 1 (44040192 elements, 44040192 padded size):
682
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight0
683
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight1
684
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight2
685
+ Params for bucket 2 (44040192 elements, 44040192 padded size):
686
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight3
687
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight4
688
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight5
689
+ Params for bucket 3 (44040192 elements, 44040192 padded size):
690
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight6
691
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight7
692
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight8
693
+ Params for bucket 4 (44040192 elements, 44040192 padded size):
694
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight9
695
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight10
696
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight11
697
+ Params for bucket 5 (44040192 elements, 44040192 padded size):
698
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight12
699
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight13
700
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight14
701
+ Params for bucket 6 (44040192 elements, 44040192 padded size):
702
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight15
703
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight16
704
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight17
705
+ Params for bucket 7 (44040192 elements, 44040192 padded size):
706
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight18
707
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight19
708
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight20
709
+ Params for bucket 8 (44040192 elements, 44040192 padded size):
710
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight21
711
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight23
712
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight22
713
+ Params for bucket 9 (44040192 elements, 44040192 padded size):
714
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight24
715
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight25
716
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight26
717
+ Params for bucket 10 (44040192 elements, 44040192 padded size):
718
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight27
719
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight28
720
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight29
721
+ Params for bucket 11 (44040192 elements, 44040192 padded size):
722
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight30
723
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight31
724
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight32
725
+ Params for bucket 12 (44040192 elements, 44040192 padded size):
726
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight33
727
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight34
728
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight35
729
+ Params for bucket 13 (44040192 elements, 44040192 padded size):
730
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight36
731
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight37
732
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight38
733
+ Params for bucket 14 (44040192 elements, 44040192 padded size):
734
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight39
735
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight40
736
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight41
737
+ Params for bucket 15 (44040192 elements, 44040192 padded size):
738
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight42
739
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight44
740
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight43
741
+ Params for bucket 16 (44040192 elements, 44040192 padded size):
742
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight45
743
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight47
744
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight46
745
+ Params for bucket 17 (58720256 elements, 58720256 padded size):
746
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight0
747
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight1
748
+ Params for bucket 18 (58720256 elements, 58720256 padded size):
749
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight2
750
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight3
751
+ Params for bucket 19 (58720256 elements, 58720256 padded size):
752
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight4
753
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight5
754
+ Params for bucket 20 (58720256 elements, 58720256 padded size):
755
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight6
756
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight7
757
+ Params for bucket 21 (58720256 elements, 58720256 padded size):
758
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight8
759
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight9
760
+ Params for bucket 22 (58720256 elements, 58720256 padded size):
761
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight10
762
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight11
763
+ Params for bucket 23 (58720256 elements, 58720256 padded size):
764
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight12
765
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight13
766
+ Params for bucket 24 (58720256 elements, 58720256 padded size):
767
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight14
768
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight15
769
+ Params for bucket 25 (58720256 elements, 58720256 padded size):
770
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight16
771
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight17
772
+ Params for bucket 26 (58720256 elements, 58720256 padded size):
773
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight18
774
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight19
775
+ Params for bucket 27 (58720256 elements, 58720256 padded size):
776
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight20
777
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight21
778
+ Params for bucket 28 (58720256 elements, 58720256 padded size):
779
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight22
780
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight23
781
+ Params for bucket 29 (58720256 elements, 58720256 padded size):
782
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight24
783
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight25
784
+ Params for bucket 30 (58720256 elements, 58720256 padded size):
785
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight26
786
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight27
787
+ Params for bucket 31 (58720256 elements, 58720256 padded size):
788
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight28
789
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight29
790
+ Params for bucket 32 (58720256 elements, 58720256 padded size):
791
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight30
792
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight31
793
+ Params for bucket 33 (58720256 elements, 58720256 padded size):
794
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight32
795
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight33
796
+ Params for bucket 34 (58720256 elements, 58720256 padded size):
797
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight34
798
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight35
799
+ Params for bucket 35 (58720256 elements, 58720256 padded size):
800
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight36
801
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight37
802
+ Params for bucket 36 (58720256 elements, 58720256 padded size):
803
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight38
804
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight39
805
+ Params for bucket 37 (58720256 elements, 58720256 padded size):
806
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight40
807
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight41
808
+ Params for bucket 38 (58720256 elements, 58720256 padded size):
809
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight42
810
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight43
811
+ Params for bucket 39 (58720256 elements, 58720256 padded size):
812
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight44
813
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight45
814
+ Params for bucket 40 (58720256 elements, 58720256 padded size):
815
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight46
816
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight47
817
+ 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 0x7149aa33af90>, config_logger_dir='')
818
+ INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine
819
+ [after model, optimizer, and learning rate scheduler are built] datetime: 2026-09-21 07:52:48
820
+ > building train, validation, and test datasets ...
821
+ > datasets target sizes (minimum size):
822
+ train: 96
823
+ validation: 0
824
+ test: 0
825
+ INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None]
826
+ > building train, validation, and test datasets for GPT ...
827
+ 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 0x714703e89b80>, 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)
828
+ INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx
829
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths
830
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers
831
+ INFO:megatron.core.datasets.indexed_dataset: Extract the document indices
832
+ INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029
833
+ INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350
834
+ INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices
835
+ INFO:megatron.core.datasets.gpt_dataset: Load the document index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-document_index.npy
836
+ INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-sample_index.npy
837
+ INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-shuffle_index.npy
838
+ INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073
839
+ > finished creating GPT datasets ...
840
+ [after dataloaders are built] datetime: 2026-09-21 07:52:48
841
+ done with setup ...
842
+ (min, max) time across ranks (ms):
843
+ model-and-optimizer-setup ......................: (1886.49, 1888.35)
844
+ train/valid/test-data-iterators-setup ..........: (67.94, 75.42)
845
+ training ...
846
+ Overwriting rerun_state_machine.current_iteration from -1 to 0...
847
+ [before the start of training step] datetime: 2026-09-21 07:52:49
848
+ INFO:moonep.api:MoonEP comm buffer sizes:
849
+ config : R=8, S=4096, H=7168, E=384, K=8, NvS=44960, NvS_padded=45056
850
+ sm_count : num_sms=32, num_sms_dedup=148
851
+ hidden_buffer_size : 616.00 MiB
852
+ meta_buffer_size : 2.00 MiB
853
+ local_temp_buffer_size : 1.29 MiB
854
+ total_size : 619.29 MiB
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/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
+ KIMI_K2_L2 moonep matched 48 4096 8 none none 1 OK 108.60 107.70 86491 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep1
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/steps.tsv CHANGED
@@ -1,3 +1,6 @@
1
  status campaign phase gbs started ended result_dir
2
  OK echo tierbprobe 8 2026-09-21T07:32:50+00:00 2026-09-21T07:36:51+00:00 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/echo/tierbprobe
3
  OK ultraep tierbprobe 8 2026-09-21T07:36:53+00:00 2026-09-21T07:40:24+00:00 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/ultraep/tierbprobe
 
 
 
 
1
  status campaign phase gbs started ended result_dir
2
  OK echo tierbprobe 8 2026-09-21T07:32:50+00:00 2026-09-21T07:36:51+00:00 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/echo/tierbprobe
3
  OK ultraep tierbprobe 8 2026-09-21T07:36:53+00:00 2026-09-21T07:40:24+00:00 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/ultraep/tierbprobe
4
+ OK moonep tierbprobe 8 2026-09-21T07:40:26+00:00 2026-09-21T07:43:07+00:00 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierbprobe
5
+ OK echo tierb 8 2026-09-21T07:43:09+00:00 2026-09-21T07:47:09+00:00 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/echo/tierb
6
+ OK ultraep tierb 8 2026-09-21T07:47:12+00:00 2026-09-21T07:50:42+00:00 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/ultraep/tierb
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/study.log CHANGED
@@ -210,3 +210,162 @@ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
210
  [guard] tearing down process group (outcome=DONE)
211
  --------------------------------------------------------------------------
212
  [07:41:54] CELL KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__forcebal__async (slots=48)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
210
  [guard] tearing down process group (outcome=DONE)
211
  --------------------------------------------------------------------------
212
  [07:41:54] CELL KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__forcebal__async (slots=48)
213
+ [guard] iter 12/12 (70s)
214
+ [guard] tearing down process group (outcome=DONE)
215
+ --------------------------------------------------------------------------
216
+
217
+ ==============================================================
218
+ SWEEP COMPLETE ok=2 failed=0
219
+ ==============================================================
220
+ config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir
221
+ KIMI_K2_L2 moonep matched 48 4096 8 none none 1 OK 121.60 120.60 86157 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierbprobe/tierbprobe_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep1
222
+ KIMI_K2_L2 moonep matched 48 4096 8 none none+forcebal 1 OK 120.39 119.90 86157 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierbprobe/tierbprobe_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__forcebal__async/rep1
223
+
224
+ Results: /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierbprobe/tierbprobe_gbs8/results.tsv
225
+ Roll up: python3 /workspace/blackwell_run_scripts/granularity/tools/collect.py /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierbprobe/tierbprobe_gbs8
226
+ ==============================================================
227
+ 2 probe runs -> /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierbprobe/tierbprobe_gbs8/probe_summary.tsv
228
+ run arm cap rawExpI preRankI postRankI drop% reqCap
229
+ --------------------------------------------------------------------------------------------------------------------
230
+ KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone_ natural none 8.1829 1.2305 1.0260
231
+ KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone_ forced none 1.1328 1.0110 1.0033
232
+
233
+ ECHO note: echo_logged_pre_rank_metric_valid=0 and echo_logged_load_reduction_valid=0 are intentional.
234
+ The branch's built-in fields mix source-rank and destination-rank axes; this tool reconstructs both globally.
235
+
236
+ TIER B PROBE -- what these logs contain, per backend
237
+ ----------------------------------------------------
238
+ Capacity matrix:
239
+ ECHO: start C=2.00; on OOM descend 1.75, 1.50, 1.25. Stop at
240
+ the highest fitting zero-drop capacity. If a fitting capacity drops,
241
+ smaller values cannot help and the model is blocked.
242
+ UltraEP: measure required_capacity at C=1.25, round it UP to the next
243
+ 0.01, then validate that capacity once.
244
+ MoonEP: natural + forced, no capacity factor.
245
+
246
+ The forced arm runs once at the selected/high-headroom capacity. It validates
247
+ instrumentation and is not part of capacity selection.
248
+
249
+ Each backend instruments a different thing:
250
+
251
+ echo local raw-router histograms and local post-reroute placement.
252
+ reduce_probe.py aligns and sums all eight source ranks before
253
+ calculating raw_expert_*, pre_home_rank_*,
254
+ post_physical_rank_* and valid_max_rank_reduction_*.
255
+
256
+ ultraep global pre/post rank vectors plus explicit drop accounting:
257
+ drop_rate_pct and required_capacity -- the capacity factor the run
258
+ WOULD have needed. The vectors are duplicated identically on all
259
+ logging ranks and are deduplicated, not summed.
260
+
261
+ moonep local raw-router histograms, actual payload rows per physical
262
+ destination, and the bytes-copy side. The reducer globally sums
263
+ each (iteration, layer) snapshot first. ECHO reports no comparable
264
+ byte counters, so MoonEP bytes are a self-comparison across models,
265
+ never a cross-backend ranking.
266
+
267
+ Do NOT quote ECHO's logged pre_echo imbalance_ratio or
268
+ max_load_reduction_pct. The branch mixes source-rank and destination-rank
269
+ axes; the reducer marks both invalid and reconstructs the comparison globally.
270
+
271
+ Two boundary cases:
272
+ raw_expert_imax_max == E/K
273
+ -> at least one global snapshot had one expert in every
274
+ token's top-K set. Inspect the MEAN before generalising.
275
+ required_capacity or min-safe-capacity at the configured capacity
276
+ -> clipped by the buffer; it means "at or over", not
277
+ "fits comfortably".
278
+
279
+ tools/reduce_probe.py has already written probe_summary.tsv and .json here.
280
+
281
+ Campaign results: /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierbprobe
282
+ Roll up: python3 /workspace/blackwell_run_scripts/granularity/tools/collect.py /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierbprobe --compare
283
+ Already logged in to Hugging Face as: AbhishekGhosh1998
284
+ uploaded: /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierbprobe -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierbprobe
285
+ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
286
+
287
+ === STAGE 4 -- Tier B timing (probe-selected zero-drop capacities)
288
+
289
+ === STEP echo__tierb__gbs8
290
+ ==============================================================
291
+ CAMPAIGN: echo backend=echo phase=tierb
292
+ Megatron: /workspace/Megatron-MoE-Echo @ bracing-regions-through-kernels
293
+ GBS=8 seq=4096 repeats=3 nsys=1
294
+ results -> /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/echo/tierb
295
+ ==============================================================
296
+ [07:43:23] KIMI_K2_L2 echo: raising floor 1.1171875 -> 1.1328125 (max over clean runs)
297
+ [07:43:23] KIMI_K2_L2: full-CG capacity candidates: 1.17578125 1.15625 1.1328125
298
+ --------------------------------------------------------------------------
299
+ [07:43:23] CELL KIMI_K2_L2__echo__matched__seq4096__gbs8__cap1p17578125 (slots=48)
300
+ [guard] iter 12/12 (70s)
301
+ [guard] tearing down process group (outcome=DONE)
302
+ [guard] iter 12/12 (70s)
303
+ [guard] tearing down process group (outcome=DONE)
304
+ [guard] iter 12/12 (70s)
305
+ [guard] tearing down process group (outcome=DONE)
306
+ --------------------------------------------------------------------------
307
+
308
+ ==============================================================
309
+ SWEEP COMPLETE ok=3 failed=0
310
+ ==============================================================
311
+ config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir
312
+ KIMI_K2_L2 echo matched 48 4096 8 1.17578125 none 1 OK 119.40 118.30 132219 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/echo/tierb/tierb_gbs8/KIMI_K2_L2__echo__matched__seq4096__gbs8__cap1p17578125/rep1
313
+ KIMI_K2_L2 echo matched 48 4096 8 1.17578125 none 2 OK 119.57 118.30 132219 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/echo/tierb/tierb_gbs8/KIMI_K2_L2__echo__matched__seq4096__gbs8__cap1p17578125/rep2
314
+ KIMI_K2_L2 echo matched 48 4096 8 1.17578125 none 3 OK 119.60 118.30 132219 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/echo/tierb/tierb_gbs8/KIMI_K2_L2__echo__matched__seq4096__gbs8__cap1p17578125/rep3
315
+
316
+ Results: /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/echo/tierb/tierb_gbs8/results.tsv
317
+ Roll up: python3 /workspace/blackwell_run_scripts/granularity/tools/collect.py /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/echo/tierb/tierb_gbs8
318
+ ==============================================================
319
+
320
+ Campaign results: /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/echo/tierb
321
+ Roll up: python3 /workspace/blackwell_run_scripts/granularity/tools/collect.py /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/echo/tierb --compare
322
+ Already logged in to Hugging Face as: AbhishekGhosh1998
323
+ uploaded: /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/echo/tierb -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/echo/tierb
324
+ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
325
+
326
+ === STEP ultraep__tierb__gbs8
327
+ ==============================================================
328
+ CAMPAIGN: ultraep backend=ultraep phase=tierb
329
+ Megatron: /workspace/Megatron-UltraEP @ ultraep-bracing-regions-through-kernels
330
+ GBS=8 seq=4096 repeats=3 nsys=1
331
+ results -> /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/ultraep/tierb
332
+ ==============================================================
333
+ --------------------------------------------------------------------------
334
+ [07:47:26] CELL KIMI_K2_L2__ultraep__matched__seq4096__gbs8__cap1p17578125 (slots=48)
335
+ [guard] iter 12/12 (60s)
336
+ [guard] tearing down process group (outcome=DONE)
337
+ [guard] iter 12/12 (60s)
338
+ [guard] tearing down process group (outcome=DONE)
339
+ [guard] iter 12/12 (60s)
340
+ [guard] tearing down process group (outcome=DONE)
341
+ --------------------------------------------------------------------------
342
+
343
+ ==============================================================
344
+ SWEEP COMPLETE ok=3 failed=0
345
+ ==============================================================
346
+ config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir
347
+ KIMI_K2_L2 ultraep matched 48 4096 8 1.17578125 none 1 OK 110.06 108.20 90375 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/ultraep/tierb/tierb_gbs8/KIMI_K2_L2__ultraep__matched__seq4096__gbs8__cap1p17578125/rep1
348
+ KIMI_K2_L2 ultraep matched 48 4096 8 1.17578125 none 2 OK 110.08 108.20 90375 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/ultraep/tierb/tierb_gbs8/KIMI_K2_L2__ultraep__matched__seq4096__gbs8__cap1p17578125/rep2
349
+ KIMI_K2_L2 ultraep matched 48 4096 8 1.17578125 none 3 OK 110.08 108.20 90375 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/ultraep/tierb/tierb_gbs8/KIMI_K2_L2__ultraep__matched__seq4096__gbs8__cap1p17578125/rep3
350
+
351
+ Results: /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/ultraep/tierb/tierb_gbs8/results.tsv
352
+ Roll up: python3 /workspace/blackwell_run_scripts/granularity/tools/collect.py /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/ultraep/tierb/tierb_gbs8
353
+ ==============================================================
354
+
355
+ Campaign results: /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/ultraep/tierb
356
+ Roll up: python3 /workspace/blackwell_run_scripts/granularity/tools/collect.py /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/ultraep/tierb --compare
357
+ Already logged in to Hugging Face as: AbhishekGhosh1998
358
+ uploaded: /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/ultraep/tierb -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/ultraep/tierb
359
+ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
360
+
361
+ === STEP moonep__tierb__gbs8
362
+ ==============================================================
363
+ CAMPAIGN: moonep backend=moonep phase=tierb
364
+ Megatron: /workspace/Megatron-MoE-Echo @ moonep_backend_async_grad_reduce
365
+ GBS=8 seq=4096 repeats=3 nsys=1
366
+ results -> /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb
367
+ ==============================================================
368
+ --------------------------------------------------------------------------
369
+ [07:51:00] CELL KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async (slots=48)
370
+ [guard] iter 12/12 (70s)
371
+ [guard] tearing down process group (outcome=DONE)
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/tierb_timing_capacities.env ADDED
@@ -0,0 +1 @@
 
 
1
+ TIERB_TIMING_CAP_KIMI_K2_L2=1.17578125
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/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
+ KIMI_K2_L2 1.17578125 1.1328125 OK /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/echo/tierb/tierb_gbs8/KIMI_K2_L2__echo__matched__seq4096__gbs8__cap1p17578125/rep3