AbhishekGhosh1998 commited on
Commit
25086c2
·
verified ·
1 Parent(s): a829f7f

granularity results moonep/tierb @ 2026-08-26T18:50:10Z from sp-aiinfra-interactive-aghosh-b5xzj-8xb200-moe-cg-0

Browse files
Files changed (22) hide show
  1. moonep/tierb/manifest.txt +19 -0
  2. moonep/tierb/pip_freeze.txt +317 -0
  3. moonep/tierb/provenance.txt +90 -0
  4. moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep1/mem_samples.csv +200 -0
  5. moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep1/run.env +61 -0
  6. moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep1/train.log +1425 -0
  7. moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep2/mem_samples.csv +248 -0
  8. moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep2/run.env +61 -0
  9. moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep2/train.log +1426 -0
  10. moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep3/mem_samples.csv +232 -0
  11. moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep3/run.env +61 -0
  12. moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep3/train.log +1426 -0
  13. moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep1/mem_samples.csv +200 -0
  14. moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep1/run.env +61 -0
  15. moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep1/train.log +1049 -0
  16. moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep2/mem_samples.csv +192 -0
  17. moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep2/run.env +61 -0
  18. moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep2/train.log +1049 -0
  19. moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep3/mem_samples.csv +192 -0
  20. moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep3/run.env +61 -0
  21. moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep3/train.log +1049 -0
  22. moonep/tierb/tierb_gbs8/results.tsv +7 -0
moonep/tierb/manifest.txt ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ timestamp: 2026-08-26T18:42:15+00:00
2
+ host: sp-aiinfra-interactive-aghosh-b5xzj-8xb200-moe-cg-0
3
+ gpu: NVIDIA B200 x8
4
+ driver: 580.95.05
5
+ study_scripts_branch: granularity-moonep
6
+ study_scripts_commit: ef3305e
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_bracing_regions_through_kernels
18
+ megatron_branch_actual: moonep_backend_bracing_regions_through_kernels
19
+ megatron_commit: 1aac8bdd
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==1.28.0
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@1aac8bdd0626b21390dfc5d98ef05c2400fc10ac#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@c27d35da15f25c112218795894504a65c16f4879#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==0.7.1
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
moonep/tierb/provenance.txt ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ==============================================================
2
+ PROVENANCE: moonep__tierb__gbs8
3
+ ==============================================================
4
+ captured_at : 2026-08-26T18:41:59+00:00
5
+ hostname : sp-aiinfra-interactive-aghosh-b5xzj-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/tierb_study_20260826_144644/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, 127 MHz, 1965 MHz, 1000.00 W, 29
21
+ 1, 142 MHz, 1965 MHz, 1000.00 W, 35
22
+ 2, 172 MHz, 1965 MHz, 1000.00 W, 34
23
+ 3, 217 MHz, 1965 MHz, 1000.00 W, 30
24
+ 4, 195 MHz, 1965 MHz, 1000.00 W, 30
25
+ 5, 180 MHz, 1965 MHz, 1000.00 W, 35
26
+ 6, 382 MHz, 1965 MHz, 1000.00 W, 34
27
+ 7, 232 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: 1.28.0 [/usr/local/lib/python3.12/dist-packages/huggingface_hub/__init__.py]
60
+
61
+ ---- repositories --------------------------------------------
62
+ Megatron-MoE-Echo moonep_backend_bracing_regions_through_kernels 1aac8bdd0626
63
+ Megatron-UltraEP ultraep-bracing-regions-through-kernels 03d57ef7d476
64
+ UltraEP main 94cab099b44f
65
+ DeepEP HEAD 34152ae28f80 (+3 untracked: build output)
66
+ MoonEP megatron-integration c27d35da15f2
67
+ blackwell_run_scripts granularity-moonep ef3305ebfabd
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=tierb_study_20260826_144644
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 Aug 26 14:37 /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 Aug 26 14:09 /usr/local/lib/python3.12/dist-packages/nvidia/nvshmem/lib/libnvshmem_host.so.3
moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep1/mem_samples.csv ADDED
@@ -0,0 +1,200 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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, 0
10
+ 1, 0
11
+ 2, 0
12
+ 3, 0
13
+ 4, 0
14
+ 5, 0
15
+ 6, 0
16
+ 7, 0
17
+ 0, 1
18
+ 1, 1
19
+ 2, 1
20
+ 3, 1
21
+ 4, 1
22
+ 5, 1
23
+ 6, 1
24
+ 7, 1
25
+ 0, 4
26
+ 1, 4
27
+ 2, 4
28
+ 3, 4
29
+ 4, 1
30
+ 5, 1
31
+ 6, 1
32
+ 7, 1
33
+ 0, 4
34
+ 1, 4
35
+ 2, 4
36
+ 3, 4
37
+ 4, 4
38
+ 5, 4
39
+ 6, 4
40
+ 7, 4
41
+ 0, 4
42
+ 1, 629
43
+ 2, 629
44
+ 3, 47
45
+ 4, 47
46
+ 5, 629
47
+ 6, 629
48
+ 7, 629
49
+ 0, 935
50
+ 1, 937
51
+ 2, 957
52
+ 3, 961
53
+ 4, 939
54
+ 5, 937
55
+ 6, 959
56
+ 7, 935
57
+ 0, 1955
58
+ 1, 1955
59
+ 2, 1955
60
+ 3, 1955
61
+ 4, 1955
62
+ 5, 1955
63
+ 6, 1955
64
+ 7, 1955
65
+ 0, 3845
66
+ 1, 3849
67
+ 2, 3857
68
+ 3, 3855
69
+ 4, 3841
70
+ 5, 3851
71
+ 6, 3863
72
+ 7, 3865
73
+ 0, 35381
74
+ 1, 34973
75
+ 2, 34787
76
+ 3, 34891
77
+ 4, 35233
78
+ 5, 34987
79
+ 6, 34997
80
+ 7, 34985
81
+ 0, 35933
82
+ 1, 36057
83
+ 2, 35927
84
+ 3, 35987
85
+ 4, 35939
86
+ 5, 35987
87
+ 6, 35911
88
+ 7, 35993
89
+ 0, 46237
90
+ 1, 46363
91
+ 2, 46221
92
+ 3, 46295
93
+ 4, 46245
94
+ 5, 46293
95
+ 6, 46277
96
+ 7, 46367
97
+ 0, 46291
98
+ 1, 46411
99
+ 2, 46285
100
+ 3, 46343
101
+ 4, 46295
102
+ 5, 46341
103
+ 6, 46333
104
+ 7, 46411
105
+ 0, 46293
106
+ 1, 46415
107
+ 2, 46289
108
+ 3, 46349
109
+ 4, 46301
110
+ 5, 46345
111
+ 6, 46337
112
+ 7, 46415
113
+ 0, 116083
114
+ 1, 116067
115
+ 2, 116079
116
+ 3, 116063
117
+ 4, 116091
118
+ 5, 116061
119
+ 6, 116063
120
+ 7, 116067
121
+ 0, 139411
122
+ 1, 139403
123
+ 2, 139405
124
+ 3, 139401
125
+ 4, 139417
126
+ 5, 139397
127
+ 6, 140175
128
+ 7, 139403
129
+ 0, 139801
130
+ 1, 139793
131
+ 2, 139795
132
+ 3, 139789
133
+ 4, 139805
134
+ 5, 139787
135
+ 6, 140565
136
+ 7, 139793
137
+ 0, 141851
138
+ 1, 141843
139
+ 2, 141845
140
+ 3, 141839
141
+ 4, 141855
142
+ 5, 141837
143
+ 6, 142615
144
+ 7, 141843
145
+ 0, 142699
146
+ 1, 142739
147
+ 2, 142679
148
+ 3, 142715
149
+ 4, 142743
150
+ 5, 142693
151
+ 6, 143477
152
+ 7, 142727
153
+ 0, 147373
154
+ 1, 147365
155
+ 2, 147367
156
+ 3, 147361
157
+ 4, 147377
158
+ 5, 147359
159
+ 6, 146089
160
+ 7, 147365
161
+ 0, 165249
162
+ 1, 165239
163
+ 2, 165243
164
+ 3, 165237
165
+ 4, 165253
166
+ 5, 165233
167
+ 6, 165227
168
+ 7, 165239
169
+ 0, 165291
170
+ 1, 165281
171
+ 2, 165285
172
+ 3, 165279
173
+ 4, 165295
174
+ 5, 165275
175
+ 6, 165269
176
+ 7, 165281
177
+ 0, 165291
178
+ 1, 165281
179
+ 2, 165285
180
+ 3, 165279
181
+ 4, 165295
182
+ 5, 165275
183
+ 6, 165269
184
+ 7, 165281
185
+ 0, 165291
186
+ 1, 165281
187
+ 2, 165285
188
+ 3, 165279
189
+ 4, 165295
190
+ 5, 165275
191
+ 6, 165269
192
+ 7, 165281
193
+ 0, 11177
194
+ 1, 9578
195
+ 2, 9578
196
+ 3, 9578
197
+ 4, 9578
198
+ 5, 9578
199
+ 6, 9898
200
+ 7, 165279
moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep1/run.env ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=moonep
2
+ config=DSV2_LITE_L24
3
+ family=tierB
4
+ repo=/workspace/Megatron-MoE-Echo
5
+ branch=moonep_backend_bracing_regions_through_kernels
6
+ commit=1aac8bdd
7
+ num_layers=24
8
+ hidden=2048
9
+ ffn_hidden=10944
10
+ moe_ffn_hidden=1408
11
+ num_experts=64
12
+ topk=6
13
+ experts_per_rank=8
14
+ expert_params=8650752
15
+ heads=16
16
+ kv_groups=0
17
+ kv_channels=128
18
+ qk_layernorm=1
19
+ shared_ffn=2816
20
+ shared_expert_overlap=0
21
+ mla=1
22
+ dense_first_layers=1
23
+ moe_layers=23
24
+ rotary_base=10000
25
+ rotary_scaling=40
26
+ qkv_bias=0
27
+ router_balancing=seq_aux_loss
28
+ aux_loss_coeff=1e-3
29
+ router_pre_softmax=1
30
+ router_fusion=0
31
+ router_dtype=fp32
32
+ device_initiated_grouped_gemm=1
33
+ norm_epsilon=1e-6
34
+ seq_length=8192
35
+ micro_batch_size=1
36
+ global_batch_size=8
37
+ num_microbatches=1
38
+ assignments_per_rank=49152
39
+ mean_tokens_per_expert=6144
40
+ slot_policy=matched
41
+ slots_per_rank=8
42
+ capacity_mode=none
43
+ capacity=none
44
+ cuda_graph=1
45
+ probe_mode=0
46
+ async_grad_reduce=0
47
+ branch_role=timing
48
+ memory_dump_mode=none
49
+ memory_snapshot_path=
50
+ cuda_graph_warmup_steps=3
51
+ recompute=none
52
+ force_balance=0
53
+ train_iters=12
54
+ ep_size=8
55
+ echo_dispatch_overlap=1
56
+ echo_dispatch_recompute=0
57
+ ultraep_quota_min_tokens=
58
+ ultraep_weight_sync_join=
59
+ profile_enable=0
60
+ kernel_trace=0
61
+ fp8_recipe=mxfp8
moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep1/train.log ADDED
@@ -0,0 +1,1425 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ==============================================================
2
+ DSV2_LITE_L24 / moonep (slots/rank=8 policy=matched)
3
+ E=64 K=6 H'=1408 m=8
4
+ L=24 seq=8192 gbs=8 (1 microbatches)
5
+ cap_mode=none cap=none cg=1 recompute=none
6
+ memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep1/memory_snapshots/snapshot.pickle
7
+ repo=/workspace/Megatron-MoE-Echo @ 1aac8bdd
8
+ ==============================================================
9
+ W0826 18:42:17.103000 547536 torch/distributed/run.py:766]
10
+ W0826 18:42:17.103000 547536 torch/distributed/run.py:766] *****************************************
11
+ W0826 18:42:17.103000 547536 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
12
+ W0826 18:42:17.103000 547536 torch/distributed/run.py:766] *****************************************
13
+ 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
14
+ WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer
15
+ Number of virtual stages per pipeline stage: None
16
+ accumulate and all-reduce gradients in fp32 for bfloat16 data type.
17
+ using torch.bfloat16 for parameters ...
18
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
19
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
20
+ ------------------------ arguments ------------------------
21
+ account_for_embedding_in_pipeline_split ......... False
22
+ account_for_loss_in_pipeline_split .............. False
23
+ accumulate_allreduce_grads_in_fp32 .............. True
24
+ adam_beta1 ...................................... 0.9
25
+ adam_beta2 ...................................... 0.999
26
+ adam_eps ........................................ 1e-08
27
+ add_bias_linear ................................. False
28
+ add_position_embedding .......................... False
29
+ add_qkv_bias .................................... False
30
+ adlr_autoresume ................................. False
31
+ adlr_autoresume_interval ........................ 1000
32
+ align_grad_reduce ............................... True
33
+ align_param_gather .............................. False
34
+ app_tag_run_name ................................ None
35
+ app_tag_run_version ............................. 0.0.0
36
+ apply_layernorm_1p .............................. False
37
+ apply_query_key_layer_scaling ................... False
38
+ apply_residual_connection_post_layernorm ........ False
39
+ apply_rope_fusion ............................... False
40
+ async_save ...................................... None
41
+ async_tensor_model_parallel_allreduce ........... True
42
+ attention_backend ............................... AttnBackend.auto
43
+ attention_dropout ............................... 0.0
44
+ attention_softmax_in_fp32 ....................... False
45
+ auto_detect_ckpt_format ......................... False
46
+ barrier_with_L1_time ............................ True
47
+ bert_binary_head ................................ True
48
+ bert_embedder_type .............................. megatron
49
+ bert_load ....................................... None
50
+ bf16 ............................................ True
51
+ bias_dropout_fusion ............................. True
52
+ bias_gelu_fusion ................................ False
53
+ bias_swiglu_fusion .............................. True
54
+ biencoder_projection_dim ........................ 0
55
+ biencoder_shared_query_context_model ............ False
56
+ block_data_path ................................. None
57
+ calc_ft_timeouts ................................ False
58
+ calculate_per_token_loss ........................ False
59
+ check_for_large_grads ........................... False
60
+ check_for_nan_in_loss_and_grad .................. False
61
+ check_for_spiky_loss ............................ False
62
+ check_weight_hash_across_dp_replicas_interval ... None
63
+ ckpt_assume_constant_structure .................. False
64
+ ckpt_convert_format ............................. None
65
+ ckpt_convert_save ............................... None
66
+ ckpt_convert_update_legacy_dist_opt_format ...... False
67
+ ckpt_format ..................................... torch_dist
68
+ ckpt_fully_parallel_load ........................ False
69
+ ckpt_fully_parallel_save ........................ True
70
+ ckpt_fully_parallel_save_deprecated ............. False
71
+ ckpt_step ....................................... None
72
+ classes_fraction ................................ 1.0
73
+ clip_grad ....................................... 1.0
74
+ clone_scatter_output_in_embedding ............... True
75
+ config_logger_dir ...............................
76
+ consumed_train_samples .......................... 0
77
+ consumed_valid_samples .......................... 0
78
+ context_parallel_size ........................... 1
79
+ cp_comm_type .................................... ['p2p']
80
+ create_attention_mask_in_dataloader ............. True
81
+ cross_entropy_fusion_impl ....................... native
82
+ cross_entropy_loss_fusion ....................... False
83
+ cuda_graph_scope ................................ full_iteration
84
+ cuda_graph_warmup_steps ......................... 3
85
+ data_args_path .................................. None
86
+ data_cache_path ................................. None
87
+ data_parallel_random_init ....................... False
88
+ data_parallel_sharding_strategy ................. no_shard
89
+ data_parallel_size .............................. 8
90
+ data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document']
91
+ data_per_class_fraction ......................... 1.0
92
+ data_sharding ................................... True
93
+ dataloader_type ................................. single
94
+ ddp_average_in_collective ....................... False
95
+ ddp_bucket_size ................................. None
96
+ ddp_num_buckets ................................. None
97
+ ddp_pad_buckets_for_high_nccl_busbw ............. False
98
+ decoder_first_pipeline_num_layers ............... None
99
+ decoder_last_pipeline_num_layers ................ None
100
+ decoder_num_layers .............................. None
101
+ decoder_seq_length .............................. None
102
+ decoupled_lr .................................... None
103
+ decoupled_min_lr ................................ None
104
+ decrease_batch_size_if_needed ................... False
105
+ defer_embedding_wgrad_compute ................... False
106
+ delay_wgrad_compute ............................. False
107
+ deprecated_use_mcore_models ..................... True
108
+ deterministic_mode .............................. False
109
+ dino_bottleneck_size ............................ 256
110
+ dino_freeze_last_layer .......................... 1
111
+ dino_head_hidden_size ........................... 2048
112
+ dino_local_crops_number ......................... 10
113
+ dino_local_img_size ............................. 96
114
+ dino_norm_last_layer ............................ False
115
+ dino_teacher_temp ............................... 0.07
116
+ dino_warmup_teacher_temp ........................ 0.04
117
+ dino_warmup_teacher_temp_epochs ................. 30
118
+ disable_bf16_reduced_precision_matmul ........... False
119
+ disable_mamba_mem_eff_path ...................... False
120
+ disable_straggler_on_startup .................... False
121
+ dist_ckpt_format_deprecated ..................... None
122
+ dist_ckpt_strictness ............................ assume_ok_unexpected
123
+ distribute_saved_activations .................... False
124
+ distributed_backend ............................. nccl
125
+ distributed_timeout_minutes ..................... 10
126
+ embedding_init_method_std ....................... None
127
+ embedding_path .................................. None
128
+ empty_unused_memory_level ....................... 0
129
+ enable_cuda_graph ............................... True
130
+ enable_experimental ............................. False
131
+ enable_ft_package ............................... False
132
+ enable_gloo_process_groups ...................... True
133
+ enable_msc ...................................... True
134
+ enable_one_logger ............................... True
135
+ encoder_num_layers .............................. 24
136
+ encoder_seq_length .............................. 8192
137
+ end_weight_decay ................................ 0.1
138
+ eod_mask_loss ................................... False
139
+ error_injection_rate ............................ 0
140
+ error_injection_type ............................ transient_error
141
+ eval_interval ................................... 1000000
142
+ eval_iters ...................................... 0
143
+ evidence_data_path .............................. None
144
+ exit_duration_in_mins ........................... None
145
+ exit_interval ................................... None
146
+ exit_on_missing_checkpoint ...................... False
147
+ exit_signal_handler ............................. False
148
+ exp_avg_dtype ................................... torch.float32
149
+ exp_avg_sq_dtype ................................ torch.float32
150
+ expert_model_parallel_size ...................... 8
151
+ expert_tensor_parallel_size ..................... 1
152
+ export_force_local_attention .................... False
153
+ export_kd_cfg ................................... None
154
+ export_kd_teacher_ckpt_format ................... None
155
+ export_kd_teacher_load .......................... None
156
+ export_kv_cache_quant ........................... False
157
+ export_legacy_megatron .......................... False
158
+ export_model_type ............................... GPTModel
159
+ export_moe_apply_probs_on_input ................. False
160
+ export_qk_l2_norm ............................... False
161
+ export_quant_cfg ................................ None
162
+ export_real_quant_cfg ........................... None
163
+ export_te_mcore_model ........................... False
164
+ external_cuda_graph ............................. False
165
+ ffn_hidden_size ................................. 10944
166
+ finetune ........................................ False
167
+ finetune_data_split ............................. train
168
+ finetune_hf_dataset ............................. None
169
+ first_last_layers_bf16 .......................... False
170
+ flash_decode .................................... False
171
+ fp16 ............................................ False
172
+ fp16_lm_cross_entropy ........................... False
173
+ fp32_residual_connection ........................ False
174
+ fp8 ............................................. e4m3
175
+ fp8_amax_compute_algo ........................... most_recent
176
+ fp8_amax_history_len ............................ 1
177
+ fp8_interval .................................... 1
178
+ fp8_margin ...................................... 0
179
+ fp8_param_gather ................................ False
180
+ fp8_recipe ...................................... mxfp8
181
+ fp8_wgrad ....................................... True
182
+ fsdp_double_buffer .............................. False
183
+ global_batch_size ............................... 8
184
+ grad_reduce_in_bf16 ............................. False
185
+ gradient_accumulation_fusion .................... True
186
+ gradient_reduce_div_fusion ...................... True
187
+ group_query_attention ........................... False
188
+ head_lr_mult .................................... 1.0
189
+ heterogeneous_layers_config_encoded_json ........ None
190
+ heterogeneous_layers_config_path ................ None
191
+ hidden_dropout .................................. 0.0
192
+ hidden_size ..................................... 2048
193
+ hierarchical_context_parallel_sizes ............. None
194
+ high_priority_stream_groups ..................... []
195
+ hybrid_attention_ratio .......................... 0.0
196
+ hybrid_mlp_ratio ................................ 0.0
197
+ hybrid_override_pattern ......................... None
198
+ hysteresis ...................................... 2
199
+ ict_head_size ................................... None
200
+ ict_load ........................................ None
201
+ img_h ........................................... 224
202
+ img_w ........................................... 224
203
+ indexer_batch_size .............................. 128
204
+ indexer_log_interval ............................ 1000
205
+ inference_batch_times_seqlen_threshold .......... -1
206
+ inference_dynamic_batching ...................... False
207
+ inference_dynamic_batching_buffer_guaranteed_fraction 0.2
208
+ inference_dynamic_batching_buffer_overflow_factor None
209
+ inference_dynamic_batching_buffer_size_gb ....... 40.0
210
+ inference_dynamic_batching_chunk_size ........... 256
211
+ inference_dynamic_batching_max_requests_override None
212
+ inference_dynamic_batching_max_tokens_override .. None
213
+ inference_dynamic_batching_num_cuda_graphs ...... 16
214
+ inference_max_batch_size ........................ 8
215
+ inference_max_seq_length ........................ 2560
216
+ inference_rng_tracker ........................... False
217
+ init_method_std ................................. 0.01
218
+ init_method_xavier_uniform ...................... False
219
+ init_model_with_meta_device ..................... False
220
+ initial_loss_scale .............................. 4294967296
221
+ inprocess_active_world_size ..................... 8
222
+ inprocess_barrier_timeout ....................... 120
223
+ inprocess_completion_timeout .................... 120
224
+ inprocess_empty_cuda_cache ...................... False
225
+ inprocess_granularity ........................... node
226
+ inprocess_hard_timeout .......................... 90
227
+ inprocess_heartbeat_interval .................... 30
228
+ inprocess_heartbeat_timeout ..................... 60
229
+ inprocess_last_call_wait ........................ 1
230
+ inprocess_max_iterations ........................ None
231
+ inprocess_monitor_process_interval .............. 1.0
232
+ inprocess_monitor_thread_interval ............... 1.0
233
+ inprocess_progress_watchdog_interval ............ 1.0
234
+ inprocess_restart ............................... False
235
+ inprocess_soft_timeout .......................... 60
236
+ inprocess_termination_grace_time ................ 1
237
+ is_hybrid_model ................................. False
238
+ iter_per_epoch .................................. 1250
239
+ iterations_to_skip .............................. []
240
+ keep_fp8_transpose_cache_when_using_custom_fsdp . False
241
+ kitchen_config_file ............................. None
242
+ kitchen_recipe_number ........................... None
243
+ kv_channels ..................................... 128
244
+ kv_lora_rank .................................... 512
245
+ lazy_mpu_init ................................... None
246
+ load ............................................ None
247
+ load_main_params_from_ckpt ...................... None
248
+ load_model_opt_format ........................... False
249
+ local_rank ...................................... 0
250
+ log_energy ...................................... False
251
+ log_interval .................................... 1
252
+ log_loss_scale_to_tensorboard ................... True
253
+ log_memory_to_tensorboard ....................... False
254
+ log_num_zeros_in_grad ........................... False
255
+ log_params_norm ................................. False
256
+ log_progress .................................... False
257
+ log_straggler ................................... False
258
+ log_throughput .................................. True
259
+ log_timers_to_tensorboard ....................... False
260
+ log_validation_ppl_to_tensorboard ............... False
261
+ log_world_size_to_tensorboard ................... False
262
+ logging_level ................................... None
263
+ loss_scale ...................................... None
264
+ loss_scale_window ............................... 1000
265
+ lr .............................................. 0.0001
266
+ lr_decay_iters .................................. 8
267
+ lr_decay_samples ................................ None
268
+ lr_decay_style .................................. cosine
269
+ lr_warmup_fraction .............................. None
270
+ lr_warmup_init .................................. 0.0
271
+ lr_warmup_iters ................................. 5
272
+ lr_warmup_samples ............................... 0
273
+ lr_wsd_decay_iters .............................. None
274
+ lr_wsd_decay_samples ............................ None
275
+ lr_wsd_decay_style .............................. exponential
276
+ main_grads_dtype ................................ torch.float32
277
+ main_params_dtype ............................... torch.float32
278
+ make_vocab_size_divisible_by .................... 128
279
+ mamba_head_dim .................................. 64
280
+ mamba_num_groups ................................ 8
281
+ mamba_num_heads ................................. None
282
+ mamba_state_dim ................................. 128
283
+ manual_gc ....................................... False
284
+ manual_gc_eval .................................. True
285
+ manual_gc_interval .............................. 0
286
+ mask_factor ..................................... 1.0
287
+ mask_prob ....................................... 0.15
288
+ mask_type ....................................... random
289
+ masked_softmax_fusion ........................... False
290
+ max_position_embeddings ......................... 32768
291
+ max_tokens_to_oom ............................... 12000
292
+ memory_snapshot_path ............................ snapshot.pickle
293
+ merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt
294
+ micro_batch_size ................................ 1
295
+ microbatch_group_size_per_vp_stage .............. None
296
+ mid_level_dataset_surplus ....................... 0.005
297
+ min_loss_scale .................................. 1.0
298
+ min_lr .......................................... 1e-05
299
+ mlp_chunks_for_prefill .......................... 1
300
+ mmap_bin_files .................................. True
301
+ mock_data ....................................... False
302
+ moe_apply_probs_on_input ........................ False
303
+ moe_aux_loss_coeff .............................. 0.001
304
+ moe_deepep_num_sms .............................. 20
305
+ moe_echo_enable_random_offloading ............... False
306
+ moe_echo_expert_dispatch_overlap ................ False
307
+ moe_echo_expert_dispatcher_type ................. hybridep
308
+ moe_echo_recompute_expert_dispatch .............. False
309
+ moe_enable_deepep ............................... False
310
+ moe_enable_echo ................................. False
311
+ moe_enable_hybridep ............................. False
312
+ moe_enable_moonep ............................... True
313
+ moe_expert_capacity_factor ...................... None
314
+ moe_extended_tp ................................. False
315
+ moe_ffn_hidden_size ............................. 1408
316
+ moe_grouped_gemm ................................ True
317
+ moe_hybridep_num_sms ............................ 16
318
+ moe_input_jitter_eps ............................ None
319
+ moe_layer_freq .................................. [0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
320
+ moe_layer_recompute ............................. False
321
+ moe_moonep_num_sms .............................. None
322
+ moe_num_echo_experts ............................ None
323
+ moe_pad_expert_input_to_capacity ................ False
324
+ moe_per_layer_logging ........................... False
325
+ moe_permute_fusion .............................. False
326
+ moe_received_token_capacity ..................... None
327
+ moe_router_bias_update_rate ..................... 0.001
328
+ moe_router_dtype ................................ fp32
329
+ moe_router_enable_expert_bias ................... False
330
+ moe_router_force_load_balancing ................. False
331
+ moe_router_group_topk ........................... None
332
+ moe_router_load_balancing_type .................. seq_aux_loss
333
+ moe_router_num_groups ........................... None
334
+ moe_router_padding_for_fp8 ...................... False
335
+ moe_router_pre_softmax .......................... True
336
+ moe_router_score_function ....................... softmax
337
+ moe_router_topk ................................. 6
338
+ moe_router_topk_scaling_factor .................. None
339
+ moe_shared_expert_intermediate_size ............. 2816
340
+ moe_shared_expert_overlap ....................... False
341
+ moe_token_dispatcher_type ....................... flex
342
+ moe_token_drop_policy ........................... probs
343
+ moe_upcycling_granularity ....................... 1
344
+ moe_use_device_initiated_grouped_gemm ........... True
345
+ moe_use_legacy_grouped_gemm ..................... False
346
+ moe_use_upcycling ............................... False
347
+ moe_z_loss_coeff ................................ None
348
+ mrope_section ................................... None
349
+ mscale .......................................... 0.707
350
+ mscale_all_dim .................................. 0.707
351
+ mtp_loss_scaling_factor ......................... 0.1
352
+ mtp_num_layers .................................. None
353
+ multi_latent_attention .......................... True
354
+ nccl_all_reduce_for_prefill ..................... False
355
+ nccl_communicator_config_path ................... None
356
+ nccl_ub ......................................... False
357
+ no_load_optim ................................... None
358
+ no_load_rng ..................................... None
359
+ no_persist_layer_norm ........................... False
360
+ no_rope_freq .................................... None
361
+ no_save_optim ................................... None
362
+ no_save_rng ..................................... None
363
+ non_persistent_ckpt_type ........................ None
364
+ non_persistent_global_ckpt_dir .................. None
365
+ non_persistent_local_ckpt_algo .................. fully_parallel
366
+ non_persistent_local_ckpt_dir ................... None
367
+ non_persistent_save_interval .................... None
368
+ norm_epsilon .................................... 1e-06
369
+ normalization ................................... RMSNorm
370
+ num_attention_heads ............................. 16
371
+ num_channels .................................... 3
372
+ num_classes ..................................... 1000
373
+ num_dataset_builder_threads ..................... 1
374
+ num_distributed_optimizer_instances ............. 1
375
+ num_experts ..................................... 64
376
+ num_layers ...................................... 24
377
+ num_layers_at_end_in_bf16 ....................... 1
378
+ num_layers_at_start_in_bf16 ..................... 1
379
+ num_layers_per_virtual_pipeline_stage ........... None
380
+ num_query_groups ................................ 1
381
+ num_virtual_stages_per_pipeline_rank ............ None
382
+ num_workers ..................................... 2
383
+ object_storage_cache_path ....................... None
384
+ one_logger_async ................................ False
385
+ one_logger_project .............................. megatron-lm
386
+ one_logger_run_name ............................. None
387
+ onnx_safe ....................................... None
388
+ openai_gelu ..................................... False
389
+ optimizer ....................................... adam
390
+ optimizer_cpu_offload ........................... False
391
+ optimizer_offload_fraction ...................... 1.0
392
+ output_bert_embeddings .......................... False
393
+ overlap_cpu_optimizer_d2h_h2d ................... False
394
+ overlap_grad_reduce ............................. True
395
+ overlap_moe_expert_parallel_comm ................ False
396
+ overlap_p2p_comm ................................ False
397
+ overlap_p2p_comm_warmup_flush ................... False
398
+ overlap_param_gather ............................ True
399
+ overlap_param_gather_with_optimizer_step ........ False
400
+ override_opt_param_scheduler .................... False
401
+ params_dtype .................................... torch.bfloat16
402
+ patch_dim ....................................... 16
403
+ per_split_data_args_path ........................ None
404
+ perform_initialization .......................... True
405
+ pin_cpu_grads ................................... True
406
+ pin_cpu_params .................................. True
407
+ pipeline_model_parallel_comm_backend ............ None
408
+ pipeline_model_parallel_layout .................. None
409
+ pipeline_model_parallel_size .................... 1
410
+ position_embedding_type ......................... rope
411
+ pretrained_checkpoint ........................... None
412
+ profile ......................................... False
413
+ profile_ranks ................................... [0]
414
+ profile_step_end ................................ 12
415
+ profile_step_start .............................. 10
416
+ q_lora_rank ..................................... None
417
+ qk_head_dim ..................................... 128
418
+ qk_l2_norm ...................................... False
419
+ qk_layernorm .................................... True
420
+ qk_pos_emb_head_dim ............................. 64
421
+ query_in_block_prob ............................. 0.1
422
+ rampup_batch_size ............................... None
423
+ rank ............................................ 0
424
+ recompute_granularity ........................... None
425
+ recompute_method ................................ None
426
+ recompute_modules ............................... None
427
+ recompute_num_layers ............................ None
428
+ record_memory_history ........................... False
429
+ relative_attention_max_distance ................. 128
430
+ relative_attention_num_buckets .................. 32
431
+ replication ..................................... False
432
+ replication_factor .............................. 2
433
+ replication_jump ................................ None
434
+ rerun_mode ...................................... validate_results
435
+ reset_attention_mask ............................ False
436
+ reset_position_ids .............................. False
437
+ result_rejected_tracker_filename ................ None
438
+ retriever_report_topk_accuracies ................ []
439
+ retriever_score_scaling ......................... False
440
+ retriever_seq_length ............................ 256
441
+ retro_add_retriever ............................. False
442
+ retro_attention_gate ............................ 1
443
+ retro_cyclic_train_iters ........................ None
444
+ retro_encoder_attention_dropout ................. 0.1
445
+ retro_encoder_hidden_dropout .................... 0.1
446
+ retro_encoder_layers ............................ 2
447
+ retro_num_neighbors ............................. 2
448
+ retro_num_retrieved_chunks ...................... 2
449
+ retro_project_dir ............................... None
450
+ retro_verify_neighbor_count ..................... True
451
+ reuse_grad_buf_for_mxfp8_param_ag ............... False
452
+ rope_scaling_factor ............................. 8.0
453
+ rope_type ....................................... None
454
+ rotary_base ..................................... 10000
455
+ rotary_interleaved .............................. False
456
+ rotary_percent .................................. 1.0
457
+ rotary_scaling_factor ........................... 40.0
458
+ rotary_seq_len_interpolation_factor ............. None
459
+ run_workload_inspector_server ................... False
460
+ sample_rate ..................................... 1.0
461
+ save ............................................ None
462
+ save_interval ................................... None
463
+ save_retain_interval ............................ None
464
+ scatter_gather_tensors_in_pipeline .............. True
465
+ seed ............................................ 1234
466
+ seq_length ...................................... 8192
467
+ sequence_parallel ............................... False
468
+ sft ............................................. False
469
+ sft_tokenizer_prompt_format ..................... nemotron-h-aligned
470
+ sgd_momentum .................................... 0.9
471
+ short_seq_prob .................................. 0.1
472
+ skip_train ...................................... False
473
+ skipped_train_samples ........................... 0
474
+ spec ............................................ None
475
+ split ........................................... 100,0,0
476
+ squared_relu .................................... False
477
+ start_weight_decay .............................. 0.1
478
+ straggler_ctrlr_port ............................ 65535
479
+ straggler_minmax_count .......................... 1
480
+ suggested_communication_unit_size ............... None
481
+ swiglu .......................................... True
482
+ swin_backbone_type .............................. tiny
483
+ symmetric_ar_type ............................... None
484
+ te_rng_tracker .................................. True
485
+ tensor_model_parallel_size ...................... 1
486
+ tensorboard_dir ................................. None
487
+ tensorboard_log_interval ........................ 1
488
+ tensorboard_queue_size .......................... 1000
489
+ test_data_path .................................. None
490
+ test_mode ....................................... False
491
+ tiktoken_num_special_tokens ..................... 1000
492
+ tiktoken_pattern ................................ None
493
+ tiktoken_special_tokens ......................... None
494
+ timing_log_level ................................ 0
495
+ timing_log_option ............................... minmax
496
+ titles_data_path ................................ None
497
+ tokenizer_model ................................. None
498
+ tokenizer_type .................................. GPT2BPETokenizer
499
+ torch_fsdp2_reshard_after_forward ............... True
500
+ tp_comm_bootstrap_backend ....................... nccl
501
+ tp_comm_bulk_dgrad .............................. True
502
+ tp_comm_bulk_wgrad .............................. True
503
+ tp_comm_overlap ................................. False
504
+ tp_comm_overlap_ag .............................. True
505
+ tp_comm_overlap_cfg ............................. None
506
+ tp_comm_overlap_rs .............................. True
507
+ tp_comm_overlap_rs_dgrad ........................ False
508
+ tp_comm_split_ag ................................ True
509
+ tp_comm_split_rs ................................ True
510
+ train_data_path ................................. None
511
+ train_iters ..................................... 12
512
+ train_samples ................................... None
513
+ train_sync_interval ............................. None
514
+ transformer_impl ................................ transformer_engine
515
+ transformer_pipeline_model_parallel_size ........ 1
516
+ untie_embeddings_and_output_weights ............. True
517
+ use_checkpoint_args ............................. False
518
+ use_checkpoint_opt_param_scheduler .............. False
519
+ use_cpu_initialization .......................... None
520
+ use_custom_fsdp ................................. False
521
+ use_dist_ckpt ................................... True
522
+ use_dist_ckpt_deprecated ........................ False
523
+ use_distributed_optimizer ....................... True
524
+ use_flash_attn .................................. False
525
+ use_legacy_models ............................... False
526
+ use_mp_args_from_checkpoint_args ................ False
527
+ use_one_sent_docs ............................... False
528
+ use_persistent_ckpt_worker ...................... False
529
+ use_precision_aware_optimizer ................... False
530
+ use_pytorch_profiler ............................ False
531
+ use_ring_exchange_p2p ........................... False
532
+ use_rope_scaling ................................ False
533
+ use_rotary_position_embeddings .................. False
534
+ use_sharp ....................................... False
535
+ use_tokenizer_model_from_checkpoint_args ........ True
536
+ use_torch_fsdp2 ................................. False
537
+ use_torch_optimizer_for_cpu_offload ............. False
538
+ use_tp_pp_dp_mapping ............................ False
539
+ v_head_dim ...................................... 128
540
+ valid_data_path ................................. None
541
+ variable_seq_lengths ............................ False
542
+ virtual_pipeline_model_parallel_size ............ None
543
+ vision_backbone_type ............................ vit
544
+ vision_pretraining .............................. False
545
+ vision_pretraining_type ......................... classify
546
+ vocab_extra_ids ................................. 0
547
+ vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json
548
+ vocab_size ...................................... None
549
+ wandb_exp_name ..................................
550
+ wandb_project ...................................
551
+ wandb_save_dir ..................................
552
+ weight_decay .................................... 0.1
553
+ weight_decay_incr_style ......................... constant
554
+ wgrad_deferral_limit ............................ 0
555
+ world_size ...................................... 8
556
+ yaml_cfg ........................................ None
557
+ -------------------- end of arguments ---------------------
558
+ INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1
559
+ > building GPT2BPETokenizer tokenizer ...
560
+ > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304)
561
+ WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS
562
+ > initializing torch distributed ...
563
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
564
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
565
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
566
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
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
+ 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
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
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
579
+ [Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
580
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
581
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
582
+ [Gloo] Rank [Gloo] Rank [Gloo] Rank 154 is connected to is connected to is connected to 777 peer ranks. peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : Expected number of connected peer ranks is : 777
583
+
584
+
585
+ [Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
586
+ [Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
587
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
588
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
589
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
590
+ [Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
591
+ [Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
592
+ [Gloo] Rank [Gloo] Rank 7 is connected to 76 peer ranks. is connected to Expected number of connected peer ranks is : 77 peer ranks. Expected number of connected peer ranks is :
593
+ 7
594
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
595
+ > initialized tensor model parallel with size 1
596
+ > initialized pipeline model parallel with size 1
597
+ > setting random seeds to 1234 ...
598
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
599
+ [Gloo] Rank [Gloo] Rank 00 is connected to is connected to 00 peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : 00
600
+
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
+ [rank1]:[W826 18:42:35.288242748 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.
606
+ [rank4]:[W826 18:42:35.297981227 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.
607
+ [rank6]:[W826 18:42:35.298013433 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.
608
+ [rank3]:[W826 18:42:35.298020269 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.
609
+ [rank5]:[W826 18:42:35.298026316 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 5] using GPU 5 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
610
+ [rank2]:[W826 18:42:35.298074134 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.
611
+ [rank7]:[W826 18:42:35.303180024 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
612
+ > compiling dataset index builder ...
613
+ make: Entering directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
614
+ make: Nothing to be done for 'default'.
615
+ make: Leaving directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
616
+ >>> done with dataset index builder. Compilation time: 0.036 seconds
617
+ WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations.
618
+ > compiling and loading fused kernels ...
619
+ [rank0]:[W826 18:42:35.457842335 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.
620
+ >>> done with compiling and loading fused kernels. Compilation time: 3.828 seconds
621
+ time to initialize megatron (seconds): 7.265
622
+ [after megatron is initialized] datetime: 2026-08-26 18:42:41
623
+ building GPT model ...
624
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
625
+ warnings.warn(
626
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
627
+ warnings.warn(
628
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
629
+ warnings.warn(
630
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
631
+ warnings.warn(
632
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
633
+ warnings.warn(
634
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
635
+ warnings.warn(
636
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
637
+ warnings.warn(
638
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
639
+ warnings.warn(
640
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
641
+ warnings.warn(
642
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
643
+ warnings.warn(
644
+ /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.
645
+ warnings.warn(
646
+ /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.
647
+ warnings.warn(
648
+ /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.
649
+ warnings.warn(
650
+ /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.
651
+ warnings.warn(
652
+ /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.
653
+ warnings.warn(
654
+ /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.
655
+ warnings.warn(
656
+ > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 2596386816
657
+ 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)
658
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 20
659
+ Params for bucket 1 (103022592 elements, 103022592 padded size):
660
+ module.output_layer.weight
661
+ Params for bucket 2 (48503296 elements, 48503296 padded size):
662
+ module.decoder.final_layernorm.weight
663
+ module.decoder.layers.23.mlp.shared_experts.linear_fc1.weight
664
+ module.decoder.layers.23.pre_mlp_layernorm.weight
665
+ module.decoder.layers.23.self_attention.linear_kv_up_proj.layer_norm_weight
666
+ module.decoder.layers.23.self_attention.linear_kv_down_proj.weight
667
+ module.decoder.layers.23.self_attention.linear_q_proj.weight
668
+ module.decoder.layers.23.self_attention.linear_proj.weight
669
+ module.decoder.layers.23.input_layernorm.weight
670
+ module.decoder.layers.22.mlp.shared_experts.linear_fc2.weight
671
+ module.decoder.layers.22.mlp.shared_experts.linear_fc1.weight
672
+ module.decoder.layers.23.mlp.shared_experts.linear_fc2.weight
673
+ module.decoder.layers.23.mlp.router.weight
674
+ module.decoder.layers.23.self_attention.linear_kv_up_proj.weight
675
+ Params for bucket 3 (40901632 elements, 40901632 padded size):
676
+ module.decoder.layers.22.mlp.router.weight
677
+ module.decoder.layers.22.self_attention.linear_kv_up_proj.weight
678
+ module.decoder.layers.22.self_attention.linear_proj.weight
679
+ module.decoder.layers.22.input_layernorm.weight
680
+ module.decoder.layers.21.mlp.shared_experts.linear_fc2.weight
681
+ module.decoder.layers.21.mlp.shared_experts.linear_fc1.weight
682
+ module.decoder.layers.21.mlp.router.weight
683
+ module.decoder.layers.21.self_attention.linear_kv_up_proj.weight
684
+ module.decoder.layers.21.self_attention.linear_q_proj.weight
685
+ module.decoder.layers.22.pre_mlp_layernorm.weight
686
+ module.decoder.layers.22.self_attention.linear_kv_down_proj.weight
687
+ module.decoder.layers.22.self_attention.linear_kv_up_proj.layer_norm_weight
688
+ module.decoder.layers.22.self_attention.linear_q_proj.weight
689
+ module.decoder.layers.21.pre_mlp_layernorm.weight
690
+ module.decoder.layers.21.self_attention.linear_kv_up_proj.layer_norm_weight
691
+ module.decoder.layers.21.self_attention.linear_kv_down_proj.weight
692
+ Params for bucket 4 (41163264 elements, 41163264 padded size):
693
+ module.decoder.layers.21.self_attention.linear_proj.weight
694
+ module.decoder.layers.21.input_layernorm.weight
695
+ module.decoder.layers.20.mlp.shared_experts.linear_fc1.weight
696
+ module.decoder.layers.20.mlp.shared_experts.linear_fc2.weight
697
+ module.decoder.layers.20.pre_mlp_layernorm.weight
698
+ module.decoder.layers.20.self_attention.linear_kv_up_proj.layer_norm_weight
699
+ module.decoder.layers.20.self_attention.linear_kv_down_proj.weight
700
+ module.decoder.layers.20.mlp.router.weight
701
+ module.decoder.layers.20.self_attention.linear_kv_up_proj.weight
702
+ module.decoder.layers.20.self_attention.linear_q_proj.weight
703
+ module.decoder.layers.20.self_attention.linear_proj.weight
704
+ module.decoder.layers.20.input_layernorm.weight
705
+ module.decoder.layers.19.mlp.shared_experts.linear_fc2.weight
706
+ Params for bucket 5 (42734080 elements, 42734080 padded size):
707
+ module.decoder.layers.19.pre_mlp_layernorm.weight
708
+ module.decoder.layers.19.self_attention.linear_kv_up_proj.layer_norm_weight
709
+ module.decoder.layers.19.self_attention.linear_kv_down_proj.weight
710
+ module.decoder.layers.19.mlp.shared_experts.linear_fc1.weight
711
+ module.decoder.layers.19.mlp.router.weight
712
+ module.decoder.layers.19.self_attention.linear_kv_up_proj.weight
713
+ module.decoder.layers.19.self_attention.linear_q_proj.weight
714
+ module.decoder.layers.19.self_attention.linear_proj.weight
715
+ module.decoder.layers.19.input_layernorm.weight
716
+ module.decoder.layers.18.mlp.shared_experts.linear_fc2.weight
717
+ module.decoder.layers.18.mlp.shared_experts.linear_fc1.weight
718
+ Params for bucket 6 (40901632 elements, 40901632 padded size):
719
+ module.decoder.layers.18.self_attention.linear_kv_up_proj.weight
720
+ module.decoder.layers.18.pre_mlp_layernorm.weight
721
+ module.decoder.layers.18.self_attention.linear_q_proj.weight
722
+ module.decoder.layers.18.self_attention.linear_proj.weight
723
+ module.decoder.layers.18.input_layernorm.weight
724
+ module.decoder.layers.17.mlp.shared_experts.linear_fc2.weight
725
+ module.decoder.layers.17.mlp.shared_experts.linear_fc1.weight
726
+ module.decoder.layers.17.pre_mlp_layernorm.weight
727
+ module.decoder.layers.17.self_attention.linear_kv_up_proj.layer_norm_weight
728
+ module.decoder.layers.17.self_attention.linear_kv_down_proj.weight
729
+ module.decoder.layers.18.mlp.router.weight
730
+ module.decoder.layers.18.self_attention.linear_kv_up_proj.layer_norm_weight
731
+ module.decoder.layers.18.self_attention.linear_kv_down_proj.weight
732
+ module.decoder.layers.17.mlp.router.weight
733
+ module.decoder.layers.17.self_attention.linear_kv_up_proj.weight
734
+ module.decoder.layers.17.self_attention.linear_q_proj.weight
735
+ Params for bucket 7 (41163264 elements, 41163264 padded size):
736
+ module.decoder.layers.16.pre_mlp_layernorm.weight
737
+ module.decoder.layers.16.self_attention.linear_kv_up_proj.layer_norm_weight
738
+ module.decoder.layers.16.self_attention.linear_kv_down_proj.weight
739
+ module.decoder.layers.16.self_attention.linear_q_proj.weight
740
+ module.decoder.layers.16.self_attention.linear_proj.weight
741
+ module.decoder.layers.16.input_layernorm.weight
742
+ module.decoder.layers.15.mlp.shared_experts.linear_fc2.weight
743
+ module.decoder.layers.17.self_attention.linear_proj.weight
744
+ module.decoder.layers.16.mlp.shared_experts.linear_fc2.weight
745
+ module.decoder.layers.16.mlp.shared_experts.linear_fc1.weight
746
+ module.decoder.layers.17.input_layernorm.weight
747
+ module.decoder.layers.16.mlp.router.weight
748
+ module.decoder.layers.16.self_attention.linear_kv_up_proj.weight
749
+ Params for bucket 8 (42734080 elements, 42734080 padded size):
750
+ module.decoder.layers.15.mlp.shared_experts.linear_fc1.weight
751
+ module.decoder.layers.15.self_attention.linear_kv_up_proj.layer_norm_weight
752
+ module.decoder.layers.15.pre_mlp_layernorm.weight
753
+ module.decoder.layers.15.mlp.router.weight
754
+ module.decoder.layers.15.self_attention.linear_kv_up_proj.weight
755
+ module.decoder.layers.15.self_attention.linear_kv_down_proj.weight
756
+ module.decoder.layers.15.self_attention.linear_q_proj.weight
757
+ module.decoder.layers.15.self_attention.linear_proj.weight
758
+ module.decoder.layers.15.input_layernorm.weight
759
+ module.decoder.layers.14.mlp.shared_experts.linear_fc2.weight
760
+ module.decoder.layers.14.mlp.shared_experts.linear_fc1.weight
761
+ Params for bucket 9 (40901632 elements, 40901632 padded size):
762
+ module.decoder.layers.14.mlp.router.weight
763
+ module.decoder.layers.14.self_attention.linear_kv_up_proj.weight
764
+ module.decoder.layers.14.self_attention.linear_kv_down_proj.weight
765
+ module.decoder.layers.14.self_attention.linear_q_proj.weight
766
+ module.decoder.layers.14.self_attention.linear_proj.weight
767
+ module.decoder.layers.14.input_layernorm.weight
768
+ module.decoder.layers.13.mlp.shared_experts.linear_fc2.weight
769
+ module.decoder.layers.13.mlp.shared_experts.linear_fc1.weight
770
+ module.decoder.layers.13.pre_mlp_layernorm.weight
771
+ module.decoder.layers.13.self_attention.linear_kv_up_proj.layer_norm_weight
772
+ module.decoder.layers.14.self_attention.linear_kv_up_proj.layer_norm_weight
773
+ module.decoder.layers.14.pre_mlp_layernorm.weight
774
+ module.decoder.layers.13.mlp.router.weight
775
+ module.decoder.layers.13.self_attention.linear_kv_up_proj.weight
776
+ module.decoder.layers.13.self_attention.linear_kv_down_proj.weight
777
+ module.decoder.layers.13.self_attention.linear_q_proj.weight
778
+ Params for bucket 10 (41163264 elements, 41163264 padded size):
779
+ module.decoder.layers.13.input_layernorm.weight
780
+ module.decoder.layers.12.mlp.shared_experts.linear_fc2.weight
781
+ module.decoder.layers.12.mlp.router.weight
782
+ module.decoder.layers.12.self_attention.linear_kv_up_proj.weight
783
+ module.decoder.layers.11.mlp.shared_experts.linear_fc2.weight
784
+ module.decoder.layers.13.self_attention.linear_proj.weight
785
+ module.decoder.layers.12.mlp.shared_experts.linear_fc1.weight
786
+ module.decoder.layers.12.pre_mlp_layernorm.weight
787
+ module.decoder.layers.12.self_attention.linear_kv_up_proj.layer_norm_weight
788
+ module.decoder.layers.12.self_attention.linear_kv_down_proj.weight
789
+ module.decoder.layers.12.self_attention.linear_q_proj.weight
790
+ module.decoder.layers.12.self_attention.linear_proj.weight
791
+ module.decoder.layers.12.input_layernorm.weight
792
+ Params for bucket 11 (42734080 elements, 42734080 padded size):
793
+ module.decoder.layers.11.mlp.router.weight
794
+ module.decoder.layers.11.self_attention.linear_kv_up_proj.weight
795
+ module.decoder.layers.11.self_attention.linear_proj.weight
796
+ module.decoder.layers.11.input_layernorm.weight
797
+ module.decoder.layers.10.mlp.shared_experts.linear_fc2.weight
798
+ module.decoder.layers.10.mlp.shared_experts.linear_fc1.weight
799
+ module.decoder.layers.11.mlp.shared_experts.linear_fc1.weight
800
+ module.decoder.layers.11.pre_mlp_layernorm.weight
801
+ module.decoder.layers.11.self_attention.linear_kv_up_proj.layer_norm_weight
802
+ module.decoder.layers.11.self_attention.linear_kv_down_proj.weight
803
+ module.decoder.layers.11.self_attention.linear_q_proj.weight
804
+ Params for bucket 12 (40901632 elements, 40901632 padded size):
805
+ module.decoder.layers.10.pre_mlp_layernorm.weight
806
+ module.decoder.layers.10.self_attention.linear_kv_down_proj.weight
807
+ module.decoder.layers.10.self_attention.linear_kv_up_proj.layer_norm_weight
808
+ module.decoder.layers.10.self_attention.linear_q_proj.weight
809
+ module.decoder.layers.9.pre_mlp_layernorm.weight
810
+ module.decoder.layers.9.self_attention.linear_kv_up_proj.layer_norm_weight
811
+ module.decoder.layers.9.self_attention.linear_kv_down_proj.weight
812
+ module.decoder.layers.9.self_attention.linear_q_proj.weight
813
+ module.decoder.layers.10.mlp.router.weight
814
+ module.decoder.layers.10.self_attention.linear_kv_up_proj.weight
815
+ module.decoder.layers.10.self_attention.linear_proj.weight
816
+ module.decoder.layers.10.input_layernorm.weight
817
+ module.decoder.layers.9.mlp.shared_experts.linear_fc2.weight
818
+ module.decoder.layers.9.mlp.shared_experts.linear_fc1.weight
819
+ module.decoder.layers.9.mlp.router.weight
820
+ module.decoder.layers.9.self_attention.linear_kv_up_proj.weight
821
+ Params for bucket 13 (41163264 elements, 41163264 padded size):
822
+ module.decoder.layers.8.mlp.router.weight
823
+ module.decoder.layers.8.self_attention.linear_kv_up_proj.weight
824
+ module.decoder.layers.8.self_attention.linear_kv_down_proj.weight
825
+ module.decoder.layers.9.self_attention.linear_proj.weight
826
+ module.decoder.layers.8.mlp.shared_experts.linear_fc1.weight
827
+ module.decoder.layers.8.mlp.shared_experts.linear_fc2.weight
828
+ module.decoder.layers.9.input_layernorm.weight
829
+ module.decoder.layers.8.pre_mlp_layernorm.weight
830
+ module.decoder.layers.8.self_attention.linear_kv_up_proj.layer_norm_weight
831
+ module.decoder.layers.8.self_attention.linear_q_proj.weight
832
+ module.decoder.layers.8.self_attention.linear_proj.weight
833
+ module.decoder.layers.8.input_layernorm.weight
834
+ module.decoder.layers.7.mlp.shared_experts.linear_fc2.weight
835
+ Params for bucket 14 (42734080 elements, 42734080 padded size):
836
+ module.decoder.layers.7.mlp.router.weight
837
+ module.decoder.layers.7.self_attention.linear_kv_up_proj.weight
838
+ module.decoder.layers.7.self_attention.linear_q_proj.weight
839
+ module.decoder.layers.7.self_attention.linear_proj.weight
840
+ module.decoder.layers.7.input_layernorm.weight
841
+ module.decoder.layers.6.mlp.shared_experts.linear_fc2.weight
842
+ module.decoder.layers.6.mlp.shared_experts.linear_fc1.weight
843
+ module.decoder.layers.7.mlp.shared_experts.linear_fc1.weight
844
+ module.decoder.layers.7.pre_mlp_layernorm.weight
845
+ module.decoder.layers.7.self_attention.linear_kv_up_proj.layer_norm_weight
846
+ module.decoder.layers.7.self_attention.linear_kv_down_proj.weight
847
+ Params for bucket 15 (40901632 elements, 40901632 padded size):
848
+ module.decoder.layers.6.self_attention.linear_kv_up_proj.layer_norm_weight
849
+ module.decoder.layers.6.self_attention.linear_kv_down_proj.weight
850
+ module.decoder.layers.6.pre_mlp_layernorm.weight
851
+ module.decoder.layers.6.self_attention.linear_q_proj.weight
852
+ module.decoder.layers.6.self_attention.linear_proj.weight
853
+ module.decoder.layers.6.input_layernorm.weight
854
+ module.decoder.layers.5.mlp.shared_experts.linear_fc2.weight
855
+ module.decoder.layers.5.mlp.shared_experts.linear_fc1.weight
856
+ module.decoder.layers.5.pre_mlp_layernorm.weight
857
+ module.decoder.layers.5.self_attention.linear_kv_up_proj.layer_norm_weight
858
+ module.decoder.layers.5.self_attention.linear_kv_down_proj.weight
859
+ module.decoder.layers.5.self_attention.linear_q_proj.weight
860
+ module.decoder.layers.6.mlp.router.weight
861
+ module.decoder.layers.6.self_attention.linear_kv_up_proj.weight
862
+ module.decoder.layers.5.mlp.router.weight
863
+ module.decoder.layers.5.self_attention.linear_kv_up_proj.weight
864
+ Params for bucket 16 (41163264 elements, 41163264 padded size):
865
+ module.decoder.layers.5.self_attention.linear_proj.weight
866
+ module.decoder.layers.4.mlp.router.weight
867
+ module.decoder.layers.4.mlp.shared_experts.linear_fc2.weight
868
+ module.decoder.layers.4.mlp.shared_experts.linear_fc1.weight
869
+ module.decoder.layers.5.input_layernorm.weight
870
+ module.decoder.layers.4.self_attention.linear_kv_up_proj.weight
871
+ module.decoder.layers.4.pre_mlp_layernorm.weight
872
+ module.decoder.layers.4.self_attention.linear_kv_up_proj.layer_norm_weight
873
+ module.decoder.layers.4.self_attention.linear_kv_down_proj.weight
874
+ module.decoder.layers.4.self_attention.linear_q_proj.weight
875
+ module.decoder.layers.4.self_attention.linear_proj.weight
876
+ module.decoder.layers.4.input_layernorm.weight
877
+ module.decoder.layers.3.mlp.shared_experts.linear_fc2.weight
878
+ Params for bucket 17 (42734080 elements, 42734080 padded size):
879
+ module.decoder.layers.3.pre_mlp_layernorm.weight
880
+ module.decoder.layers.3.self_attention.linear_kv_up_proj.weight
881
+ module.decoder.layers.3.mlp.shared_experts.linear_fc1.weight
882
+ module.decoder.layers.3.self_attention.linear_kv_up_proj.layer_norm_weight
883
+ module.decoder.layers.3.mlp.router.weight
884
+ module.decoder.layers.3.self_attention.linear_kv_down_proj.weight
885
+ module.decoder.layers.3.self_attention.linear_q_proj.weight
886
+ module.decoder.layers.3.self_attention.linear_proj.weight
887
+ module.decoder.layers.3.input_layernorm.weight
888
+ module.decoder.layers.2.mlp.shared_experts.linear_fc2.weight
889
+ module.decoder.layers.2.mlp.shared_experts.linear_fc1.weight
890
+ Params for bucket 18 (40901632 elements, 40901632 padded size):
891
+ module.decoder.layers.2.pre_mlp_layernorm.weight
892
+ module.decoder.layers.2.self_attention.linear_kv_down_proj.weight
893
+ module.decoder.layers.2.self_attention.linear_kv_up_proj.layer_norm_weight
894
+ module.decoder.layers.1.mlp.shared_experts.linear_fc2.weight
895
+ module.decoder.layers.1.mlp.shared_experts.linear_fc1.weight
896
+ module.decoder.layers.1.pre_mlp_layernorm.weight
897
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.layer_norm_weight
898
+ module.decoder.layers.1.self_attention.linear_q_proj.weight
899
+ module.decoder.layers.2.mlp.router.weight
900
+ module.decoder.layers.2.self_attention.linear_kv_up_proj.weight
901
+ module.decoder.layers.2.self_attention.linear_q_proj.weight
902
+ module.decoder.layers.2.self_attention.linear_proj.weight
903
+ module.decoder.layers.2.input_layernorm.weight
904
+ module.decoder.layers.1.mlp.router.weight
905
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.weight
906
+ module.decoder.layers.1.self_attention.linear_kv_down_proj.weight
907
+ Params for bucket 19 (71436288 elements, 71436288 padded size):
908
+ module.decoder.layers.1.self_attention.linear_proj.weight
909
+ module.decoder.layers.0.mlp.linear_fc2.weight
910
+ module.decoder.layers.1.input_layernorm.weight
911
+ module.decoder.layers.0.mlp.linear_fc1.weight
912
+ Params for bucket 20 (116789760 elements, 116789760 padded size):
913
+ module.decoder.layers.0.mlp.linear_fc1.layer_norm_weight
914
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.weight
915
+ module.decoder.layers.0.self_attention.linear_q_proj.weight
916
+ module.decoder.layers.0.self_attention.linear_kv_down_proj.weight
917
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.layer_norm_weight
918
+ module.decoder.layers.0.self_attention.linear_proj.weight
919
+ module.decoder.layers.0.input_layernorm.weight
920
+ module.embedding.word_embeddings.weight
921
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 40
922
+ Params for bucket 1 (40370176 elements, 40370176 padded size):
923
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight7
924
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight5
925
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight3
926
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight1
927
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight7
928
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight5
929
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight6
930
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight4
931
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight2
932
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight0
933
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight6
934
+ Params for bucket 2 (40370176 elements, 40370176 padded size):
935
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight1
936
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight3
937
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight7
938
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight5
939
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight4
940
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight2
941
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight0
942
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight6
943
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight4
944
+ Params for bucket 3 (40370176 elements, 40370176 padded size):
945
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight3
946
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight1
947
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight6
948
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight4
949
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight2
950
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight0
951
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight7
952
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight5
953
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight3
954
+ Params for bucket 4 (40370176 elements, 40370176 padded size):
955
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight2
956
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight7
957
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight0
958
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight5
959
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight3
960
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight1
961
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight6
962
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight1
963
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight4
964
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight2
965
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight0
966
+ Params for bucket 5 (40370176 elements, 40370176 padded size):
967
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight4
968
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight6
969
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight2
970
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight7
971
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight3
972
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight5
973
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight1
974
+ Params for bucket 6 (40370176 elements, 40370176 padded size):
975
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight0
976
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight4
977
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight6
978
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight2
979
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight0
980
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight6
981
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight7
982
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight5
983
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight3
984
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight1
985
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight7
986
+ Params for bucket 7 (40370176 elements, 40370176 padded size):
987
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight4
988
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight2
989
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight0
990
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight7
991
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight5
992
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight1
993
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight3
994
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight6
995
+ Params for bucket 8 (40370176 elements, 40370176 padded size):
996
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight5
997
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight3
998
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight1
999
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight6
1000
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight4
1001
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight2
1002
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight4
1003
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight0
1004
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight7
1005
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight5
1006
+ Params for bucket 9 (40370176 elements, 40370176 padded size):
1007
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight7
1008
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight0
1009
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight2
1010
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight5
1011
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight3
1012
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight3
1013
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight1
1014
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight6
1015
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight4
1016
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight2
1017
+ Params for bucket 10 (40370176 elements, 40370176 padded size):
1018
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight1
1019
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight6
1020
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight4
1021
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight2
1022
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight5
1023
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight7
1024
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight0
1025
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight3
1026
+ Params for bucket 11 (40370176 elements, 40370176 padded size):
1027
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight6
1028
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight0
1029
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight4
1030
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight2
1031
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight0
1032
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight1
1033
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight5
1034
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight7
1035
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight3
1036
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight1
1037
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight7
1038
+ Params for bucket 12 (40370176 elements, 40370176 padded size):
1039
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight6
1040
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight4
1041
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight2
1042
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight0
1043
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight5
1044
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight3
1045
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight1
1046
+ Params for bucket 13 (40370176 elements, 40370176 padded size):
1047
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight7
1048
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight5
1049
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight3
1050
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight1
1051
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight6
1052
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight6
1053
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight4
1054
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight2
1055
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight0
1056
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight7
1057
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight5
1058
+ Params for bucket 14 (40370176 elements, 40370176 padded size):
1059
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight4
1060
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight2
1061
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight0
1062
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight6
1063
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight4
1064
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight3
1065
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight1
1066
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight7
1067
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight5
1068
+ Params for bucket 15 (40370176 elements, 40370176 padded size):
1069
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight7
1070
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight0
1071
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight2
1072
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight5
1073
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight3
1074
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight3
1075
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight1
1076
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight6
1077
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight4
1078
+ Params for bucket 16 (40370176 elements, 40370176 padded size):
1079
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight1
1080
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight7
1081
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight5
1082
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight3
1083
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight1
1084
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight2
1085
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight0
1086
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight6
1087
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight4
1088
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight2
1089
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight0
1090
+ Params for bucket 17 (40370176 elements, 40370176 padded size):
1091
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight6
1092
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight3
1093
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight1
1094
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight7
1095
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight4
1096
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight5
1097
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight2
1098
+ Params for bucket 18 (40370176 elements, 40370176 padded size):
1099
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight4
1100
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight6
1101
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight2
1102
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight0
1103
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight6
1104
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight0
1105
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight7
1106
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight5
1107
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight3
1108
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight1
1109
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight7
1110
+ Params for bucket 19 (40370176 elements, 40370176 padded size):
1111
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight3
1112
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight1
1113
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight7
1114
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight5
1115
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight4
1116
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight2
1117
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight0
1118
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight6
1119
+ Params for bucket 20 (40370176 elements, 40370176 padded size):
1120
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight5
1121
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight3
1122
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight1
1123
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight6
1124
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight4
1125
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight4
1126
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight2
1127
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight0
1128
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight7
1129
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight5
1130
+ Params for bucket 21 (40370176 elements, 40370176 padded size):
1131
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight0
1132
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight2
1133
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight6
1134
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight4
1135
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight2
1136
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight3
1137
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight7
1138
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight1
1139
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight5
1140
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight3
1141
+ Params for bucket 22 (40370176 elements, 40370176 padded size):
1142
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight7
1143
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight0
1144
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight5
1145
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight3
1146
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight1
1147
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight6
1148
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight4
1149
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight2
1150
+ Params for bucket 23 (40370176 elements, 40370176 padded size):
1151
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight1
1152
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight7
1153
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight5
1154
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight3
1155
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight1
1156
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight0
1157
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight6
1158
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight4
1159
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight2
1160
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight0
1161
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight7
1162
+ Params for bucket 24 (40370176 elements, 40370176 padded size):
1163
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight6
1164
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight2
1165
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight4
1166
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight0
1167
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight3
1168
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight5
1169
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight1
1170
+ Params for bucket 25 (40370176 elements, 40370176 padded size):
1171
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight7
1172
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight3
1173
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight5
1174
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight1
1175
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight6
1176
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight5
1177
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight6
1178
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight4
1179
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight2
1180
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight0
1181
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight7
1182
+ Params for bucket 26 (40370176 elements, 40370176 padded size):
1183
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight1
1184
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight3
1185
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight6
1186
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight4
1187
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight4
1188
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight2
1189
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight0
1190
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight7
1191
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight5
1192
+ Params for bucket 27 (40370176 elements, 40370176 padded size):
1193
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight7
1194
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight0
1195
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight2
1196
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight5
1197
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight3
1198
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight3
1199
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight1
1200
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight6
1201
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight4
1202
+ Params for bucket 28 (40370176 elements, 40370176 padded size):
1203
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight7
1204
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight1
1205
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight5
1206
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight3
1207
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight1
1208
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight2
1209
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight0
1210
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight6
1211
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight4
1212
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight2
1213
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight0
1214
+ Params for bucket 29 (40370176 elements, 40370176 padded size):
1215
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight7
1216
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight5
1217
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight3
1218
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight1
1219
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight4
1220
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight6
1221
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight2
1222
+ Params for bucket 30 (40370176 elements, 40370176 padded size):
1223
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight4
1224
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight6
1225
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight2
1226
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight0
1227
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight6
1228
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight0
1229
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight5
1230
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight7
1231
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight3
1232
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight1
1233
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight7
1234
+ Params for bucket 31 (40370176 elements, 40370176 padded size):
1235
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight4
1236
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight2
1237
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight0
1238
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight6
1239
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight5
1240
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight1
1241
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight3
1242
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight7
1243
+ Params for bucket 32 (40370176 elements, 40370176 padded size):
1244
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight4
1245
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight2
1246
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight0
1247
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight7
1248
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight5
1249
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight5
1250
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight1
1251
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight3
1252
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight6
1253
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight4
1254
+ Params for bucket 33 (40370176 elements, 40370176 padded size):
1255
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight3
1256
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight1
1257
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight6
1258
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight4
1259
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight2
1260
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight2
1261
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight7
1262
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight0
1263
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight5
1264
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight3
1265
+ Params for bucket 34 (40370176 elements, 40370176 padded size):
1266
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight5
1267
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight7
1268
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight0
1269
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight3
1270
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight1
1271
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight6
1272
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight4
1273
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight2
1274
+ Params for bucket 35 (40370176 elements, 40370176 padded size):
1275
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight1
1276
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight7
1277
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight5
1278
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight3
1279
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight1
1280
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight0
1281
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight6
1282
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight4
1283
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight2
1284
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight0
1285
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight7
1286
+ Params for bucket 36 (40370176 elements, 40370176 padded size):
1287
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight6
1288
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight5
1289
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight3
1290
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight1
1291
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight2
1292
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight4
1293
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight0
1294
+ Params for bucket 37 (40370176 elements, 40370176 padded size):
1295
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight7
1296
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight3
1297
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight5
1298
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight1
1299
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight6
1300
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight6
1301
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight4
1302
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight2
1303
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight0
1304
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight7
1305
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight5
1306
+ Params for bucket 38 (40370176 elements, 40370176 padded size):
1307
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight4
1308
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight0
1309
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight2
1310
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight7
1311
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight5
1312
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight3
1313
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight1
1314
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight6
1315
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight4
1316
+ Params for bucket 39 (40370176 elements, 40370176 padded size):
1317
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight3
1318
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight1
1319
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight6
1320
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight4
1321
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight0
1322
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight7
1323
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight2
1324
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight5
1325
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight3
1326
+ Params for bucket 40 (17301504 elements, 17301504 padded size):
1327
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight2
1328
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight0
1329
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight1
1330
+ 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 0x7040d59bf920>, config_logger_dir='')
1331
+ INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine
1332
+ [after model, optimizer, and learning rate scheduler are built] datetime: 2026-08-26 18:42:44
1333
+ > building train, validation, and test datasets ...
1334
+ > datasets target sizes (minimum size):
1335
+ train: 96
1336
+ validation: 0
1337
+ test: 0
1338
+ INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None]
1339
+ > building train, validation, and test datasets for GPT ...
1340
+ INFO:megatron.core.datasets.blended_megatron_dataset_builder:Building GPTDataset splits with sizes=(96, 0, 0) and config=GPTDatasetConfig(random_seed=1234, sequence_length=8192, blend=(['/workspace/data/wikitext/wikitext103_gpt2_text_document'], None), blend_per_split=None, split='100,0,0', split_matrix=[(0, 1.0), None, None], num_dataset_builder_threads=1, path_to_cache=None, mmap_bin_files=True, mock=False, tokenizer=<megatron.training.tokenizer.tokenizer._GPT2BPETokenizer object at 0x7040d8dd8ef0>, 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)
1341
+ INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx
1342
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths
1343
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers
1344
+ INFO:megatron.core.datasets.indexed_dataset: Extract the document indices
1345
+ INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029
1346
+ INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350
1347
+ INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices
1348
+ INFO:megatron.core.datasets.gpt_dataset: Load the document index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-document_index.npy
1349
+ INFO:megatron.core.datasets.gpt_dataset: Load the sample index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-sample_index.npy
1350
+ INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-shuffle_index.npy
1351
+ INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 14536
1352
+ > finished creating GPT datasets ...
1353
+ [after dataloaders are built] datetime: 2026-08-26 18:42:44
1354
+ done with setup ...
1355
+ (min, max) time across ranks (ms):
1356
+ model-and-optimizer-setup ......................: (3401.60, 3428.73)
1357
+ train/valid/test-data-iterators-setup ..........: (61.01, 72.34)
1358
+ training ...
1359
+ Overwriting rerun_state_machine.current_iteration from -1 to 0...
1360
+ [before the start of training step] datetime: 2026-08-26 18:42:44
1361
+ INFO:moonep.api:MoonEP comm buffer sizes:
1362
+ config : R=8, S=8192, H=2048, E=64, K=6, NvS=51184, NvS_padded=51200
1363
+ sm_count : num_sms=32, num_sms_dedup=148
1364
+ hidden_buffer_size : 200.00 MiB
1365
+ meta_buffer_size : 2.00 MiB
1366
+ local_temp_buffer_size : 2.01 MiB
1367
+ total_size : 204.01 MiB
1368
+ [2026-08-26 18:43:06] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 21715.0 | throughput per GPU (TFLOP/s/GPU): 5.9 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.092206E+01 | loss scale: 1.0 | grad norm: 30.650 | number of skipped iterations: 0 | number of nan iterations: 0 |
1369
+ Number of parameters in transformer block in billions: 13.53
1370
+ Number of parameters in embedding layers in billions: 0.21
1371
+ Total number of parameters in billions: 13.74
1372
+ Number of parameters in most loaded shard in billions: 13.7356
1373
+ Theoretical memory footprints: weight and optimizer=98244.96 MB
1374
+ [Rank 0] (after 1 iterations) memory (MB) | allocated: 66144.1328125 | max allocated: 124949.08837890625 | reserved: 127502.0 | max reserved: 127502.0
1375
+ [2026-08-26 18:43:07] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 726.9 | throughput per GPU (TFLOP/s/GPU): 175.5 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.091986E+01 | loss scale: 1.0 | grad norm: 30.565 | number of skipped iterations: 0 | number of nan iterations: 0 |
1376
+ [2026-08-26 18:43:08] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 722.0 | throughput per GPU (TFLOP/s/GPU): 176.7 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 9.956084E+00 | loss scale: 1.0 | grad norm: 42.533 | number of skipped iterations: 0 | number of nan iterations: 0 |
1377
+ Capture CUDA graph for training!!!
1378
+ CUDA graph capture done!!!
1379
+ [2026-08-26 18:43:10] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 1968.3 | throughput per GPU (TFLOP/s/GPU): 64.8 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 9.478545E+00 | loss scale: 1.0 | grad norm: 7.152 | number of skipped iterations: 0 | number of nan iterations: 0 |
1380
+ [2026-08-26 18:43:10] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 703.1 | throughput per GPU (TFLOP/s/GPU): 181.5 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.091896E+01 | loss scale: 1.0 | grad norm: 452.732 | number of skipped iterations: 0 | number of nan iterations: 0 |
1381
+ [2026-08-26 18:43:11] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 701.5 | throughput per GPU (TFLOP/s/GPU): 181.9 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 9.375167E+00 | loss scale: 1.0 | grad norm: 4.992 | number of skipped iterations: 0 | number of nan iterations: 0 |
1382
+ [2026-08-26 18:43:12] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 699.1 | throughput per GPU (TFLOP/s/GPU): 182.5 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 8.982457E+00 | loss scale: 1.0 | grad norm: 3.717 | number of skipped iterations: 0 | number of nan iterations: 0 |
1383
+ [2026-08-26 18:43:12] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 700.3 | throughput per GPU (TFLOP/s/GPU): 182.2 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.787875E+00 | loss scale: 1.0 | grad norm: 4.876 | number of skipped iterations: 0 | number of nan iterations: 0 |
1384
+ [2026-08-26 18:43:13] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 702.6 | throughput per GPU (TFLOP/s/GPU): 181.6 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.620937E+00 | loss scale: 1.0 | grad norm: 2.990 | number of skipped iterations: 0 | number of nan iterations: 0 |
1385
+ [2026-08-26 18:43:14] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 702.6 | throughput per GPU (TFLOP/s/GPU): 181.6 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.680684E+00 | loss scale: 1.0 | grad norm: 194.973 | number of skipped iterations: 0 | number of nan iterations: 0 |
1386
+ [2026-08-26 18:43:15] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 702.3 | throughput per GPU (TFLOP/s/GPU): 181.7 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.583418E+00 | loss scale: 1.0 | grad norm: 24.333 | number of skipped iterations: 0 | number of nan iterations: 0 |
1387
+ [2026-08-26 18:43:15] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 698.6 | throughput per GPU (TFLOP/s/GPU): 182.6 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.611391E+00 | loss scale: 1.0 | grad norm: 75.148 | number of skipped iterations: 0 | number of nan iterations: 0 |
1388
+ [after training is done] datetime: 2026-08-26 18:43:15
1389
+ W0826 18:43:15.817000 547536 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers
1390
+ W0826 18:43:15.818000 547536 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 547624 closing signal SIGTERM
1391
+ W0826 18:43:15.819000 547536 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 547625 closing signal SIGTERM
1392
+ W0826 18:43:15.820000 547536 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 547626 closing signal SIGTERM
1393
+ W0826 18:43:15.824000 547536 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 547627 closing signal SIGTERM
1394
+ W0826 18:43:15.828000 547536 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 547628 closing signal SIGTERM
1395
+ W0826 18:43:15.833000 547536 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 547629 closing signal SIGTERM
1396
+ W0826 18:43:15.843000 547536 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 547630 closing signal SIGTERM
1397
+ W0826 18:43:15.847000 547536 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 547631 closing signal SIGTERM
1398
+ Traceback (most recent call last):
1399
+ File "/usr/local/bin/torchrun", line 8, in <module>
1400
+ sys.exit(main())
1401
+ ^^^^^^
1402
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
1403
+ return f(*args, **kwargs)
1404
+ ^^^^^^^^^^^^^^^^^^
1405
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 892, in main
1406
+ run(args)
1407
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 883, in run
1408
+ elastic_launch(
1409
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
1410
+ return launch_agent(self._config, self._entrypoint, list(args))
1411
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1412
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 261, in launch_agent
1413
+ result = agent.run()
1414
+ ^^^^^^^^^^^
1415
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper
1416
+ result = f(*args, **kwargs)
1417
+ ^^^^^^^^^^^^^^^^^^
1418
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 711, in run
1419
+ result = self._invoke_run(role)
1420
+ ^^^^^^^^^^^^^^^^^^^^^^
1421
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 870, in _invoke_run
1422
+ time.sleep(monitor_interval)
1423
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 84, in _terminate_process_handler
1424
+ raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval)
1425
+ torch.distributed.elastic.multiprocessing.api.SignalException: Process 547536 got signal: 15
moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep2/mem_samples.csv ADDED
@@ -0,0 +1,248 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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, 0
10
+ 1, 0
11
+ 2, 0
12
+ 3, 0
13
+ 4, 0
14
+ 5, 0
15
+ 6, 0
16
+ 7, 0
17
+ 0, 2
18
+ 1, 1
19
+ 2, 1
20
+ 3, 1
21
+ 4, 1
22
+ 5, 1
23
+ 6, 1
24
+ 7, 1
25
+ 0, 4
26
+ 1, 4
27
+ 2, 4
28
+ 3, 4
29
+ 4, 1
30
+ 5, 1
31
+ 6, 1
32
+ 7, 1
33
+ 0, 4
34
+ 1, 4
35
+ 2, 4
36
+ 3, 4
37
+ 4, 4
38
+ 5, 4
39
+ 6, 4
40
+ 7, 4
41
+ 0, 4
42
+ 1, 4
43
+ 2, 4
44
+ 3, 4
45
+ 4, 4
46
+ 5, 4
47
+ 6, 4
48
+ 7, 4
49
+ 0, 821
50
+ 1, 821
51
+ 2, 821
52
+ 3, 821
53
+ 4, 821
54
+ 5, 821
55
+ 6, 821
56
+ 7, 821
57
+ 0, 1955
58
+ 1, 1955
59
+ 2, 1955
60
+ 3, 1955
61
+ 4, 1955
62
+ 5, 1955
63
+ 6, 1955
64
+ 7, 1955
65
+ 0, 3863
66
+ 1, 3893
67
+ 2, 3869
68
+ 3, 3849
69
+ 4, 3881
70
+ 5, 3905
71
+ 6, 3875
72
+ 7, 3865
73
+ 0, 35381
74
+ 1, 35505
75
+ 2, 35375
76
+ 3, 35435
77
+ 4, 35387
78
+ 5, 35435
79
+ 6, 35423
80
+ 7, 35505
81
+ 0, 41335
82
+ 1, 41457
83
+ 2, 41331
84
+ 3, 41389
85
+ 4, 41341
86
+ 5, 41387
87
+ 6, 41377
88
+ 7, 41457
89
+ 0, 46289
90
+ 1, 46409
91
+ 2, 46283
92
+ 3, 46341
93
+ 4, 46293
94
+ 5, 46339
95
+ 6, 46331
96
+ 7, 46409
97
+ 0, 46291
98
+ 1, 46411
99
+ 2, 46285
100
+ 3, 46343
101
+ 4, 46295
102
+ 5, 46341
103
+ 6, 46333
104
+ 7, 46411
105
+ 0, 46305
106
+ 1, 46415
107
+ 2, 46289
108
+ 3, 46375
109
+ 4, 46327
110
+ 5, 46371
111
+ 6, 46365
112
+ 7, 46415
113
+ 0, 116085
114
+ 1, 116067
115
+ 2, 116079
116
+ 3, 116063
117
+ 4, 116091
118
+ 5, 116061
119
+ 6, 116063
120
+ 7, 116067
121
+ 0, 139411
122
+ 1, 139403
123
+ 2, 139405
124
+ 3, 139401
125
+ 4, 139417
126
+ 5, 139397
127
+ 6, 140175
128
+ 7, 139403
129
+ 0, 139801
130
+ 1, 139793
131
+ 2, 139795
132
+ 3, 139789
133
+ 4, 139805
134
+ 5, 139787
135
+ 6, 140565
136
+ 7, 139793
137
+ 0, 141851
138
+ 1, 141843
139
+ 2, 141845
140
+ 3, 141839
141
+ 4, 141855
142
+ 5, 141837
143
+ 6, 142615
144
+ 7, 141843
145
+ 0, 142499
146
+ 1, 142491
147
+ 2, 142493
148
+ 3, 142487
149
+ 4, 142503
150
+ 5, 142485
151
+ 6, 143263
152
+ 7, 142491
153
+ 0, 143137
154
+ 1, 143181
155
+ 2, 143193
156
+ 3, 143183
157
+ 4, 143157
158
+ 5, 143181
159
+ 6, 143923
160
+ 7, 143143
161
+ 0, 108209
162
+ 1, 116419
163
+ 2, 114363
164
+ 3, 116417
165
+ 4, 112325
166
+ 5, 116415
167
+ 6, 114351
168
+ 7, 110257
169
+ 0, 165291
170
+ 1, 165281
171
+ 2, 165285
172
+ 3, 165279
173
+ 4, 165295
174
+ 5, 165275
175
+ 6, 165269
176
+ 7, 165281
177
+ 0, 165291
178
+ 1, 165281
179
+ 2, 165285
180
+ 3, 165279
181
+ 4, 165295
182
+ 5, 165275
183
+ 6, 165269
184
+ 7, 165281
185
+ 0, 165291
186
+ 1, 165281
187
+ 2, 165285
188
+ 3, 165279
189
+ 4, 165295
190
+ 5, 165275
191
+ 6, 165269
192
+ 7, 165281
193
+ 0, 165089
194
+ 1, 165079
195
+ 2, 165083
196
+ 3, 165077
197
+ 4, 165093
198
+ 5, 165073
199
+ 6, 165067
200
+ 7, 165079
201
+ 0, 165089
202
+ 1, 165079
203
+ 2, 165083
204
+ 3, 165077
205
+ 4, 165093
206
+ 5, 165073
207
+ 6, 165067
208
+ 7, 165079
209
+ 0, 165089
210
+ 1, 165079
211
+ 2, 165083
212
+ 3, 165077
213
+ 4, 165093
214
+ 5, 165073
215
+ 6, 165067
216
+ 7, 165079
217
+ 0, 165089
218
+ 1, 165079
219
+ 2, 165083
220
+ 3, 165077
221
+ 4, 165093
222
+ 5, 165073
223
+ 6, 165067
224
+ 7, 165079
225
+ 0, 165089
226
+ 1, 165079
227
+ 2, 165083
228
+ 3, 165077
229
+ 4, 165093
230
+ 5, 165073
231
+ 6, 165067
232
+ 7, 165079
233
+ 0, 10973
234
+ 1, 9376
235
+ 2, 9376
236
+ 3, 9376
237
+ 4, 9376
238
+ 5, 9376
239
+ 6, 9696
240
+ 7, 165077
241
+ 0, 0
242
+ 1, 0
243
+ 2, 0
244
+ 3, 0
245
+ 4, 0
246
+ 5, 0
247
+ 6, 0
248
+ 7, 0
moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep2/run.env ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=moonep
2
+ config=DSV2_LITE_L24
3
+ family=tierB
4
+ repo=/workspace/Megatron-MoE-Echo
5
+ branch=moonep_backend_bracing_regions_through_kernels
6
+ commit=1aac8bdd
7
+ num_layers=24
8
+ hidden=2048
9
+ ffn_hidden=10944
10
+ moe_ffn_hidden=1408
11
+ num_experts=64
12
+ topk=6
13
+ experts_per_rank=8
14
+ expert_params=8650752
15
+ heads=16
16
+ kv_groups=0
17
+ kv_channels=128
18
+ qk_layernorm=1
19
+ shared_ffn=2816
20
+ shared_expert_overlap=0
21
+ mla=1
22
+ dense_first_layers=1
23
+ moe_layers=23
24
+ rotary_base=10000
25
+ rotary_scaling=40
26
+ qkv_bias=0
27
+ router_balancing=seq_aux_loss
28
+ aux_loss_coeff=1e-3
29
+ router_pre_softmax=1
30
+ router_fusion=0
31
+ router_dtype=fp32
32
+ device_initiated_grouped_gemm=1
33
+ norm_epsilon=1e-6
34
+ seq_length=8192
35
+ micro_batch_size=1
36
+ global_batch_size=8
37
+ num_microbatches=1
38
+ assignments_per_rank=49152
39
+ mean_tokens_per_expert=6144
40
+ slot_policy=matched
41
+ slots_per_rank=8
42
+ capacity_mode=none
43
+ capacity=none
44
+ cuda_graph=1
45
+ probe_mode=0
46
+ async_grad_reduce=0
47
+ branch_role=timing
48
+ memory_dump_mode=none
49
+ memory_snapshot_path=
50
+ cuda_graph_warmup_steps=3
51
+ recompute=none
52
+ force_balance=0
53
+ train_iters=12
54
+ ep_size=8
55
+ echo_dispatch_overlap=1
56
+ echo_dispatch_recompute=0
57
+ ultraep_quota_min_tokens=
58
+ ultraep_weight_sync_join=
59
+ profile_enable=0
60
+ kernel_trace=0
61
+ fp8_recipe=mxfp8
moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep2/train.log ADDED
@@ -0,0 +1,1426 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ==============================================================
2
+ DSV2_LITE_L24 / moonep (slots/rank=8 policy=matched)
3
+ E=64 K=6 H'=1408 m=8
4
+ L=24 seq=8192 gbs=8 (1 microbatches)
5
+ cap_mode=none cap=none cg=1 recompute=none
6
+ memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep2/memory_snapshots/snapshot.pickle
7
+ repo=/workspace/Megatron-MoE-Echo @ 1aac8bdd
8
+ ==============================================================
9
+ W0826 18:43:23.336000 549952 torch/distributed/run.py:766]
10
+ W0826 18:43:23.336000 549952 torch/distributed/run.py:766] *****************************************
11
+ W0826 18:43:23.336000 549952 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
12
+ W0826 18:43:23.336000 549952 torch/distributed/run.py:766] *****************************************
13
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
14
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
15
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
16
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
17
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
18
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
19
+ 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
20
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
21
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
22
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
23
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
24
+ using world size: 8, data-parallel size: 8, context-parallel size: 1, hierarchical context-parallel sizes: None, tensor-model-parallel size: 1, pipeline-model-parallel size: 1
25
+ WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer
26
+ Number of virtual stages per pipeline stage: None
27
+ accumulate and all-reduce gradients in fp32 for bfloat16 data type.
28
+ using torch.bfloat16 for parameters ...
29
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
30
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
31
+ ------------------------ arguments ------------------------
32
+ account_for_embedding_in_pipeline_split ......... False
33
+ account_for_loss_in_pipeline_split .............. False
34
+ accumulate_allreduce_grads_in_fp32 .............. True
35
+ adam_beta1 ...................................... 0.9
36
+ adam_beta2 ...................................... 0.999
37
+ adam_eps ........................................ 1e-08
38
+ add_bias_linear ................................. False
39
+ add_position_embedding .......................... False
40
+ add_qkv_bias .................................... False
41
+ adlr_autoresume ................................. False
42
+ adlr_autoresume_interval ........................ 1000
43
+ align_grad_reduce ............................... True
44
+ align_param_gather .............................. False
45
+ app_tag_run_name ................................ None
46
+ app_tag_run_version ............................. 0.0.0
47
+ apply_layernorm_1p .............................. False
48
+ apply_query_key_layer_scaling ................... False
49
+ apply_residual_connection_post_layernorm ........ False
50
+ apply_rope_fusion ............................... False
51
+ async_save ...................................... None
52
+ async_tensor_model_parallel_allreduce ........... True
53
+ attention_backend ............................... AttnBackend.auto
54
+ attention_dropout ............................... 0.0
55
+ attention_softmax_in_fp32 ....................... False
56
+ auto_detect_ckpt_format ......................... False
57
+ barrier_with_L1_time ............................ True
58
+ bert_binary_head ................................ True
59
+ bert_embedder_type .............................. megatron
60
+ bert_load ....................................... None
61
+ bf16 ............................................ True
62
+ bias_dropout_fusion ............................. True
63
+ bias_gelu_fusion ................................ False
64
+ bias_swiglu_fusion .............................. True
65
+ biencoder_projection_dim ........................ 0
66
+ biencoder_shared_query_context_model ............ False
67
+ block_data_path ................................. None
68
+ calc_ft_timeouts ................................ False
69
+ calculate_per_token_loss ........................ False
70
+ check_for_large_grads ........................... False
71
+ check_for_nan_in_loss_and_grad .................. False
72
+ check_for_spiky_loss ............................ False
73
+ check_weight_hash_across_dp_replicas_interval ... None
74
+ ckpt_assume_constant_structure .................. False
75
+ ckpt_convert_format ............................. None
76
+ ckpt_convert_save ............................... None
77
+ ckpt_convert_update_legacy_dist_opt_format ...... False
78
+ ckpt_format ..................................... torch_dist
79
+ ckpt_fully_parallel_load ........................ False
80
+ ckpt_fully_parallel_save ........................ True
81
+ ckpt_fully_parallel_save_deprecated ............. False
82
+ ckpt_step ....................................... None
83
+ classes_fraction ................................ 1.0
84
+ clip_grad ....................................... 1.0
85
+ clone_scatter_output_in_embedding ............... True
86
+ config_logger_dir ...............................
87
+ consumed_train_samples .......................... 0
88
+ consumed_valid_samples .......................... 0
89
+ context_parallel_size ........................... 1
90
+ cp_comm_type .................................... ['p2p']
91
+ create_attention_mask_in_dataloader ............. True
92
+ cross_entropy_fusion_impl ....................... native
93
+ cross_entropy_loss_fusion ....................... False
94
+ cuda_graph_scope ................................ full_iteration
95
+ cuda_graph_warmup_steps ......................... 3
96
+ data_args_path .................................. None
97
+ data_cache_path ................................. None
98
+ data_parallel_random_init ....................... False
99
+ data_parallel_sharding_strategy ................. no_shard
100
+ data_parallel_size .............................. 8
101
+ data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document']
102
+ data_per_class_fraction ......................... 1.0
103
+ data_sharding ................................... True
104
+ dataloader_type ................................. single
105
+ ddp_average_in_collective ....................... False
106
+ ddp_bucket_size ................................. None
107
+ ddp_num_buckets ................................. None
108
+ ddp_pad_buckets_for_high_nccl_busbw ............. False
109
+ decoder_first_pipeline_num_layers ............... None
110
+ decoder_last_pipeline_num_layers ................ None
111
+ decoder_num_layers .............................. None
112
+ decoder_seq_length .............................. None
113
+ decoupled_lr .................................... None
114
+ decoupled_min_lr ................................ None
115
+ decrease_batch_size_if_needed ................... False
116
+ defer_embedding_wgrad_compute ................... False
117
+ delay_wgrad_compute ............................. False
118
+ deprecated_use_mcore_models ..................... True
119
+ deterministic_mode .............................. False
120
+ dino_bottleneck_size ............................ 256
121
+ dino_freeze_last_layer .......................... 1
122
+ dino_head_hidden_size ........................... 2048
123
+ dino_local_crops_number ......................... 10
124
+ dino_local_img_size ............................. 96
125
+ dino_norm_last_layer ............................ False
126
+ dino_teacher_temp ............................... 0.07
127
+ dino_warmup_teacher_temp ........................ 0.04
128
+ dino_warmup_teacher_temp_epochs ................. 30
129
+ disable_bf16_reduced_precision_matmul ........... False
130
+ disable_mamba_mem_eff_path ...................... False
131
+ disable_straggler_on_startup .................... False
132
+ dist_ckpt_format_deprecated ..................... None
133
+ dist_ckpt_strictness ............................ assume_ok_unexpected
134
+ distribute_saved_activations .................... False
135
+ distributed_backend ............................. nccl
136
+ distributed_timeout_minutes ..................... 10
137
+ embedding_init_method_std ....................... None
138
+ embedding_path .................................. None
139
+ empty_unused_memory_level ....................... 0
140
+ enable_cuda_graph ............................... True
141
+ enable_experimental ............................. False
142
+ enable_ft_package ............................... False
143
+ enable_gloo_process_groups ...................... True
144
+ enable_msc ...................................... True
145
+ enable_one_logger ............................... True
146
+ encoder_num_layers .............................. 24
147
+ encoder_seq_length .............................. 8192
148
+ end_weight_decay ................................ 0.1
149
+ eod_mask_loss ................................... False
150
+ error_injection_rate ............................ 0
151
+ error_injection_type ............................ transient_error
152
+ eval_interval ................................... 1000000
153
+ eval_iters ...................................... 0
154
+ evidence_data_path .............................. None
155
+ exit_duration_in_mins ........................... None
156
+ exit_interval ................................... None
157
+ exit_on_missing_checkpoint ...................... False
158
+ exit_signal_handler ............................. False
159
+ exp_avg_dtype ................................... torch.float32
160
+ exp_avg_sq_dtype ................................ torch.float32
161
+ expert_model_parallel_size ...................... 8
162
+ expert_tensor_parallel_size ..................... 1
163
+ export_force_local_attention .................... False
164
+ export_kd_cfg ................................... None
165
+ export_kd_teacher_ckpt_format ................... None
166
+ export_kd_teacher_load .......................... None
167
+ export_kv_cache_quant ........................... False
168
+ export_legacy_megatron .......................... False
169
+ export_model_type ............................... GPTModel
170
+ export_moe_apply_probs_on_input ................. False
171
+ export_qk_l2_norm ............................... False
172
+ export_quant_cfg ................................ None
173
+ export_real_quant_cfg ........................... None
174
+ export_te_mcore_model ........................... False
175
+ external_cuda_graph ............................. False
176
+ ffn_hidden_size ................................. 10944
177
+ finetune ........................................ False
178
+ finetune_data_split ............................. train
179
+ finetune_hf_dataset ............................. None
180
+ first_last_layers_bf16 .......................... False
181
+ flash_decode .................................... False
182
+ fp16 ............................................ False
183
+ fp16_lm_cross_entropy ........................... False
184
+ fp32_residual_connection ........................ False
185
+ fp8 ............................................. e4m3
186
+ fp8_amax_compute_algo ........................... most_recent
187
+ fp8_amax_history_len ............................ 1
188
+ fp8_interval .................................... 1
189
+ fp8_margin ...................................... 0
190
+ fp8_param_gather ................................ False
191
+ fp8_recipe ...................................... mxfp8
192
+ fp8_wgrad ....................................... True
193
+ fsdp_double_buffer .............................. False
194
+ global_batch_size ............................... 8
195
+ grad_reduce_in_bf16 ............................. False
196
+ gradient_accumulation_fusion .................... True
197
+ gradient_reduce_div_fusion ...................... True
198
+ group_query_attention ........................... False
199
+ head_lr_mult .................................... 1.0
200
+ heterogeneous_layers_config_encoded_json ........ None
201
+ heterogeneous_layers_config_path ................ None
202
+ hidden_dropout .................................. 0.0
203
+ hidden_size ..................................... 2048
204
+ hierarchical_context_parallel_sizes ............. None
205
+ high_priority_stream_groups ..................... []
206
+ hybrid_attention_ratio .......................... 0.0
207
+ hybrid_mlp_ratio ................................ 0.0
208
+ hybrid_override_pattern ......................... None
209
+ hysteresis ...................................... 2
210
+ ict_head_size ................................... None
211
+ ict_load ........................................ None
212
+ img_h ........................................... 224
213
+ img_w ........................................... 224
214
+ indexer_batch_size .............................. 128
215
+ indexer_log_interval ............................ 1000
216
+ inference_batch_times_seqlen_threshold .......... -1
217
+ inference_dynamic_batching ...................... False
218
+ inference_dynamic_batching_buffer_guaranteed_fraction 0.2
219
+ inference_dynamic_batching_buffer_overflow_factor None
220
+ inference_dynamic_batching_buffer_size_gb ....... 40.0
221
+ inference_dynamic_batching_chunk_size ........... 256
222
+ inference_dynamic_batching_max_requests_override None
223
+ inference_dynamic_batching_max_tokens_override .. None
224
+ inference_dynamic_batching_num_cuda_graphs ...... 16
225
+ inference_max_batch_size ........................ 8
226
+ inference_max_seq_length ........................ 2560
227
+ inference_rng_tracker ........................... False
228
+ init_method_std ................................. 0.01
229
+ init_method_xavier_uniform ...................... False
230
+ init_model_with_meta_device ..................... False
231
+ initial_loss_scale .............................. 4294967296
232
+ inprocess_active_world_size ..................... 8
233
+ inprocess_barrier_timeout ....................... 120
234
+ inprocess_completion_timeout .................... 120
235
+ inprocess_empty_cuda_cache ...................... False
236
+ inprocess_granularity ........................... node
237
+ inprocess_hard_timeout .......................... 90
238
+ inprocess_heartbeat_interval .................... 30
239
+ inprocess_heartbeat_timeout ..................... 60
240
+ inprocess_last_call_wait ........................ 1
241
+ inprocess_max_iterations ........................ None
242
+ inprocess_monitor_process_interval .............. 1.0
243
+ inprocess_monitor_thread_interval ............... 1.0
244
+ inprocess_progress_watchdog_interval ............ 1.0
245
+ inprocess_restart ............................... False
246
+ inprocess_soft_timeout .......................... 60
247
+ inprocess_termination_grace_time ................ 1
248
+ is_hybrid_model ................................. False
249
+ iter_per_epoch .................................. 1250
250
+ iterations_to_skip .............................. []
251
+ keep_fp8_transpose_cache_when_using_custom_fsdp . False
252
+ kitchen_config_file ............................. None
253
+ kitchen_recipe_number ........................... None
254
+ kv_channels ..................................... 128
255
+ kv_lora_rank .................................... 512
256
+ lazy_mpu_init ................................... None
257
+ load ............................................ None
258
+ load_main_params_from_ckpt ...................... None
259
+ load_model_opt_format ........................... False
260
+ local_rank ...................................... 0
261
+ log_energy ...................................... False
262
+ log_interval .................................... 1
263
+ log_loss_scale_to_tensorboard ................... True
264
+ log_memory_to_tensorboard ....................... False
265
+ log_num_zeros_in_grad ........................... False
266
+ log_params_norm ................................. False
267
+ log_progress .................................... False
268
+ log_straggler ................................... False
269
+ log_throughput .................................. True
270
+ log_timers_to_tensorboard ....................... False
271
+ log_validation_ppl_to_tensorboard ............... False
272
+ log_world_size_to_tensorboard ................... False
273
+ logging_level ................................... None
274
+ loss_scale ...................................... None
275
+ loss_scale_window ............................... 1000
276
+ lr .............................................. 0.0001
277
+ lr_decay_iters .................................. 8
278
+ lr_decay_samples ................................ None
279
+ lr_decay_style .................................. cosine
280
+ lr_warmup_fraction .............................. None
281
+ lr_warmup_init .................................. 0.0
282
+ lr_warmup_iters ................................. 5
283
+ lr_warmup_samples ............................... 0
284
+ lr_wsd_decay_iters .............................. None
285
+ lr_wsd_decay_samples ............................ None
286
+ lr_wsd_decay_style .............................. exponential
287
+ main_grads_dtype ................................ torch.float32
288
+ main_params_dtype ............................... torch.float32
289
+ make_vocab_size_divisible_by .................... 128
290
+ mamba_head_dim .................................. 64
291
+ mamba_num_groups ................................ 8
292
+ mamba_num_heads ................................. None
293
+ mamba_state_dim ................................. 128
294
+ manual_gc ....................................... False
295
+ manual_gc_eval .................................. True
296
+ manual_gc_interval .............................. 0
297
+ mask_factor ..................................... 1.0
298
+ mask_prob ....................................... 0.15
299
+ mask_type ....................................... random
300
+ masked_softmax_fusion ........................... False
301
+ max_position_embeddings ......................... 32768
302
+ max_tokens_to_oom ............................... 12000
303
+ memory_snapshot_path ............................ snapshot.pickle
304
+ merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt
305
+ micro_batch_size ................................ 1
306
+ microbatch_group_size_per_vp_stage .............. None
307
+ mid_level_dataset_surplus ....................... 0.005
308
+ min_loss_scale .................................. 1.0
309
+ min_lr .......................................... 1e-05
310
+ mlp_chunks_for_prefill .......................... 1
311
+ mmap_bin_files .................................. True
312
+ mock_data ....................................... False
313
+ moe_apply_probs_on_input ........................ False
314
+ moe_aux_loss_coeff .............................. 0.001
315
+ moe_deepep_num_sms .............................. 20
316
+ moe_echo_enable_random_offloading ............... False
317
+ moe_echo_expert_dispatch_overlap ................ False
318
+ moe_echo_expert_dispatcher_type ................. hybridep
319
+ moe_echo_recompute_expert_dispatch .............. False
320
+ moe_enable_deepep ............................... False
321
+ moe_enable_echo ................................. False
322
+ moe_enable_hybridep ............................. False
323
+ moe_enable_moonep ............................... True
324
+ moe_expert_capacity_factor ...................... None
325
+ moe_extended_tp ................................. False
326
+ moe_ffn_hidden_size ............................. 1408
327
+ moe_grouped_gemm ................................ True
328
+ moe_hybridep_num_sms ............................ 16
329
+ moe_input_jitter_eps ............................ None
330
+ moe_layer_freq .................................. [0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
331
+ moe_layer_recompute ............................. False
332
+ moe_moonep_num_sms .............................. None
333
+ moe_num_echo_experts ............................ None
334
+ moe_pad_expert_input_to_capacity ................ False
335
+ moe_per_layer_logging ........................... False
336
+ moe_permute_fusion .............................. False
337
+ moe_received_token_capacity ..................... None
338
+ moe_router_bias_update_rate ..................... 0.001
339
+ moe_router_dtype ................................ fp32
340
+ moe_router_enable_expert_bias ................... False
341
+ moe_router_force_load_balancing ................. False
342
+ moe_router_group_topk ........................... None
343
+ moe_router_load_balancing_type .................. seq_aux_loss
344
+ moe_router_num_groups ........................... None
345
+ moe_router_padding_for_fp8 ...................... False
346
+ moe_router_pre_softmax .......................... True
347
+ moe_router_score_function ....................... softmax
348
+ moe_router_topk ................................. 6
349
+ moe_router_topk_scaling_factor .................. None
350
+ moe_shared_expert_intermediate_size ............. 2816
351
+ moe_shared_expert_overlap ....................... False
352
+ moe_token_dispatcher_type ....................... flex
353
+ moe_token_drop_policy ........................... probs
354
+ moe_upcycling_granularity ....................... 1
355
+ moe_use_device_initiated_grouped_gemm ........... True
356
+ moe_use_legacy_grouped_gemm ..................... False
357
+ moe_use_upcycling ............................... False
358
+ moe_z_loss_coeff ................................ None
359
+ mrope_section ................................... None
360
+ mscale .......................................... 0.707
361
+ mscale_all_dim .................................. 0.707
362
+ mtp_loss_scaling_factor ......................... 0.1
363
+ mtp_num_layers .................................. None
364
+ multi_latent_attention .......................... True
365
+ nccl_all_reduce_for_prefill ..................... False
366
+ nccl_communicator_config_path ................... None
367
+ nccl_ub ......................................... False
368
+ no_load_optim ................................... None
369
+ no_load_rng ..................................... None
370
+ no_persist_layer_norm ........................... False
371
+ no_rope_freq .................................... None
372
+ no_save_optim ................................... None
373
+ no_save_rng ..................................... None
374
+ non_persistent_ckpt_type ........................ None
375
+ non_persistent_global_ckpt_dir .................. None
376
+ non_persistent_local_ckpt_algo .................. fully_parallel
377
+ non_persistent_local_ckpt_dir ................... None
378
+ non_persistent_save_interval .................... None
379
+ norm_epsilon .................................... 1e-06
380
+ normalization ................................... RMSNorm
381
+ num_attention_heads ............................. 16
382
+ num_channels .................................... 3
383
+ num_classes ..................................... 1000
384
+ num_dataset_builder_threads ..................... 1
385
+ num_distributed_optimizer_instances ............. 1
386
+ num_experts ..................................... 64
387
+ num_layers ...................................... 24
388
+ num_layers_at_end_in_bf16 ....................... 1
389
+ num_layers_at_start_in_bf16 ..................... 1
390
+ num_layers_per_virtual_pipeline_stage ........... None
391
+ num_query_groups ................................ 1
392
+ num_virtual_stages_per_pipeline_rank ............ None
393
+ num_workers ..................................... 2
394
+ object_storage_cache_path ....................... None
395
+ one_logger_async ................................ False
396
+ one_logger_project .............................. megatron-lm
397
+ one_logger_run_name ............................. None
398
+ onnx_safe ....................................... None
399
+ openai_gelu ..................................... False
400
+ optimizer ....................................... adam
401
+ optimizer_cpu_offload ........................... False
402
+ optimizer_offload_fraction ...................... 1.0
403
+ output_bert_embeddings .......................... False
404
+ overlap_cpu_optimizer_d2h_h2d ................... False
405
+ overlap_grad_reduce ............................. True
406
+ overlap_moe_expert_parallel_comm ................ False
407
+ overlap_p2p_comm ................................ False
408
+ overlap_p2p_comm_warmup_flush ................... False
409
+ overlap_param_gather ............................ True
410
+ overlap_param_gather_with_optimizer_step ........ False
411
+ override_opt_param_scheduler .................... False
412
+ params_dtype .................................... torch.bfloat16
413
+ patch_dim ....................................... 16
414
+ per_split_data_args_path ........................ None
415
+ perform_initialization .......................... True
416
+ pin_cpu_grads ................................... True
417
+ pin_cpu_params .................................. True
418
+ pipeline_model_parallel_comm_backend ............ None
419
+ pipeline_model_parallel_layout .................. None
420
+ pipeline_model_parallel_size .................... 1
421
+ position_embedding_type ......................... rope
422
+ pretrained_checkpoint ........................... None
423
+ profile ......................................... False
424
+ profile_ranks ................................... [0]
425
+ profile_step_end ................................ 12
426
+ profile_step_start .............................. 10
427
+ q_lora_rank ..................................... None
428
+ qk_head_dim ..................................... 128
429
+ qk_l2_norm ...................................... False
430
+ qk_layernorm .................................... True
431
+ qk_pos_emb_head_dim ............................. 64
432
+ query_in_block_prob ............................. 0.1
433
+ rampup_batch_size ............................... None
434
+ rank ............................................ 0
435
+ recompute_granularity ........................... None
436
+ recompute_method ................................ None
437
+ recompute_modules ............................... None
438
+ recompute_num_layers ............................ None
439
+ record_memory_history ........................... False
440
+ relative_attention_max_distance ................. 128
441
+ relative_attention_num_buckets .................. 32
442
+ replication ..................................... False
443
+ replication_factor .............................. 2
444
+ replication_jump ................................ None
445
+ rerun_mode ...................................... validate_results
446
+ reset_attention_mask ............................ False
447
+ reset_position_ids .............................. False
448
+ result_rejected_tracker_filename ................ None
449
+ retriever_report_topk_accuracies ................ []
450
+ retriever_score_scaling ......................... False
451
+ retriever_seq_length ............................ 256
452
+ retro_add_retriever ............................. False
453
+ retro_attention_gate ............................ 1
454
+ retro_cyclic_train_iters ........................ None
455
+ retro_encoder_attention_dropout ................. 0.1
456
+ retro_encoder_hidden_dropout .................... 0.1
457
+ retro_encoder_layers ............................ 2
458
+ retro_num_neighbors ............................. 2
459
+ retro_num_retrieved_chunks ...................... 2
460
+ retro_project_dir ............................... None
461
+ retro_verify_neighbor_count ..................... True
462
+ reuse_grad_buf_for_mxfp8_param_ag ............... False
463
+ rope_scaling_factor ............................. 8.0
464
+ rope_type ....................................... None
465
+ rotary_base ..................................... 10000
466
+ rotary_interleaved .............................. False
467
+ rotary_percent .................................. 1.0
468
+ rotary_scaling_factor ........................... 40.0
469
+ rotary_seq_len_interpolation_factor ............. None
470
+ run_workload_inspector_server ................... False
471
+ sample_rate ..................................... 1.0
472
+ save ............................................ None
473
+ save_interval ................................... None
474
+ save_retain_interval ............................ None
475
+ scatter_gather_tensors_in_pipeline .............. True
476
+ seed ............................................ 1234
477
+ seq_length ...................................... 8192
478
+ sequence_parallel ............................... False
479
+ sft ............................................. False
480
+ sft_tokenizer_prompt_format ..................... nemotron-h-aligned
481
+ sgd_momentum .................................... 0.9
482
+ short_seq_prob .................................. 0.1
483
+ skip_train ...................................... False
484
+ skipped_train_samples ........................... 0
485
+ spec ............................................ None
486
+ split ........................................... 100,0,0
487
+ squared_relu .................................... False
488
+ start_weight_decay .............................. 0.1
489
+ straggler_ctrlr_port ............................ 65535
490
+ straggler_minmax_count .......................... 1
491
+ suggested_communication_unit_size ............... None
492
+ swiglu .......................................... True
493
+ swin_backbone_type .............................. tiny
494
+ symmetric_ar_type ............................... None
495
+ te_rng_tracker .................................. True
496
+ tensor_model_parallel_size ...................... 1
497
+ tensorboard_dir ................................. None
498
+ tensorboard_log_interval ........................ 1
499
+ tensorboard_queue_size .......................... 1000
500
+ test_data_path .................................. None
501
+ test_mode ....................................... False
502
+ tiktoken_num_special_tokens ..................... 1000
503
+ tiktoken_pattern ................................ None
504
+ tiktoken_special_tokens ......................... None
505
+ timing_log_level ................................ 0
506
+ timing_log_option ............................... minmax
507
+ titles_data_path ................................ None
508
+ tokenizer_model ................................. None
509
+ tokenizer_type .................................. GPT2BPETokenizer
510
+ torch_fsdp2_reshard_after_forward ............... True
511
+ tp_comm_bootstrap_backend ....................... nccl
512
+ tp_comm_bulk_dgrad .............................. True
513
+ tp_comm_bulk_wgrad .............................. True
514
+ tp_comm_overlap ................................. False
515
+ tp_comm_overlap_ag .............................. True
516
+ tp_comm_overlap_cfg ............................. None
517
+ tp_comm_overlap_rs .............................. True
518
+ tp_comm_overlap_rs_dgrad ........................ False
519
+ tp_comm_split_ag ................................ True
520
+ tp_comm_split_rs ................................ True
521
+ train_data_path ................................. None
522
+ train_iters ..................................... 12
523
+ train_samples ................................... None
524
+ train_sync_interval ............................. None
525
+ transformer_impl ................................ transformer_engine
526
+ transformer_pipeline_model_parallel_size ........ 1
527
+ untie_embeddings_and_output_weights ............. True
528
+ use_checkpoint_args ............................. False
529
+ use_checkpoint_opt_param_scheduler .............. False
530
+ use_cpu_initialization .......................... None
531
+ use_custom_fsdp ................................. False
532
+ use_dist_ckpt ................................... True
533
+ use_dist_ckpt_deprecated ........................ False
534
+ use_distributed_optimizer ....................... True
535
+ use_flash_attn .................................. False
536
+ use_legacy_models ............................... False
537
+ use_mp_args_from_checkpoint_args ................ False
538
+ use_one_sent_docs ............................... False
539
+ use_persistent_ckpt_worker ...................... False
540
+ use_precision_aware_optimizer ................... False
541
+ use_pytorch_profiler ............................ False
542
+ use_ring_exchange_p2p ........................... False
543
+ use_rope_scaling ................................ False
544
+ use_rotary_position_embeddings .................. False
545
+ use_sharp ....................................... False
546
+ use_tokenizer_model_from_checkpoint_args ........ True
547
+ use_torch_fsdp2 ................................. False
548
+ use_torch_optimizer_for_cpu_offload ............. False
549
+ use_tp_pp_dp_mapping ............................ False
550
+ v_head_dim ...................................... 128
551
+ valid_data_path ................................. None
552
+ variable_seq_lengths ............................ False
553
+ virtual_pipeline_model_parallel_size ............ None
554
+ vision_backbone_type ............................ vit
555
+ vision_pretraining .............................. False
556
+ vision_pretraining_type ......................... classify
557
+ vocab_extra_ids ................................. 0
558
+ vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json
559
+ vocab_size ...................................... None
560
+ wandb_exp_name ..................................
561
+ wandb_project ...................................
562
+ wandb_save_dir ..................................
563
+ weight_decay .................................... 0.1
564
+ weight_decay_incr_style ......................... constant
565
+ wgrad_deferral_limit ............................ 0
566
+ world_size ...................................... 8
567
+ yaml_cfg ........................................ None
568
+ -------------------- end of arguments ---------------------
569
+ INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1
570
+ > building GPT2BPETokenizer tokenizer ...
571
+ > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304)
572
+ WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS
573
+ > initializing torch distributed ...
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
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
579
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
580
+ [Gloo] Rank 6 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 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
585
+ [Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
586
+ [Gloo] Rank [Gloo] Rank [Gloo] Rank [Gloo] Rank [Gloo] Rank 1035[Gloo] Rank 4 is connected to is connected to is connected to is connected to 2 is connected to 7777 is connected to 7 peer ranks. peer ranks. peer ranks. peer ranks. 7 peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : Expected number of connected peer ranks is : Expected number of connected peer ranks is : peer ranks. Expected number of connected peer ranks is : 7777Expected number of connected peer ranks is : 7
587
+
588
+
589
+
590
+
591
+ 7
592
+ [Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7[Gloo] Rank
593
+ 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
594
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
595
+ > initialized tensor model parallel with size 1
596
+ > initialized pipeline model parallel with size 1
597
+ > setting random seeds to 1234 ...
598
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
599
+ [Gloo] Rank [Gloo] Rank 00 is connected to is connected to 00 peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : 00
600
+
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
+ [rank6]:[W826 18:43:41.785327486 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.
606
+ [rank3]:[W826 18:43:41.787236207 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.
607
+ [rank5]:[W826 18:43:41.787265951 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.
608
+ [rank2]:[W826 18:43:41.787293880 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 2] using GPU 2 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
609
+ [rank4]:[W826 18:43:41.787345992 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.
610
+ [rank1]:[W826 18:43:41.787389017 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
+ [rank7]:[W826 18:43:41.787599790 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
612
+ > compiling dataset index builder ...
613
+ make: Entering directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
614
+ make: Nothing to be done for 'default'.
615
+ make: Leaving directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
616
+ >>> done with dataset index builder. Compilation time: 0.033 seconds
617
+ WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations.
618
+ > compiling and loading fused kernels ...
619
+ [rank0]:[W826 18:43:41.946017516 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.
620
+ >>> done with compiling and loading fused kernels. Compilation time: 3.859 seconds
621
+ time to initialize megatron (seconds): 7.114
622
+ [after megatron is initialized] datetime: 2026-08-26 18:43:47
623
+ building GPT model ...
624
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
625
+ warnings.warn(
626
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
627
+ warnings.warn(
628
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
629
+ warnings.warn(
630
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
631
+ warnings.warn(
632
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
633
+ warnings.warn(
634
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
635
+ warnings.warn(
636
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
637
+ warnings.warn(
638
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
639
+ warnings.warn(
640
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
641
+ warnings.warn(
642
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
643
+ warnings.warn(
644
+ /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.
645
+ warnings.warn(
646
+ /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.
647
+ warnings.warn(
648
+ /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.
649
+ warnings.warn(
650
+ /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.
651
+ warnings.warn(
652
+ /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.
653
+ warnings.warn(
654
+ /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.
655
+ warnings.warn(
656
+ > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 2596386816
657
+ 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)
658
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 20
659
+ Params for bucket 1 (103022592 elements, 103022592 padded size):
660
+ module.output_layer.weight
661
+ Params for bucket 2 (48503296 elements, 48503296 padded size):
662
+ module.decoder.final_layernorm.weight
663
+ module.decoder.layers.23.mlp.shared_experts.linear_fc1.weight
664
+ module.decoder.layers.23.mlp.router.weight
665
+ module.decoder.layers.23.mlp.shared_experts.linear_fc2.weight
666
+ module.decoder.layers.23.self_attention.linear_kv_up_proj.weight
667
+ module.decoder.layers.23.self_attention.linear_q_proj.weight
668
+ module.decoder.layers.23.self_attention.linear_proj.weight
669
+ module.decoder.layers.23.pre_mlp_layernorm.weight
670
+ module.decoder.layers.23.self_attention.linear_kv_up_proj.layer_norm_weight
671
+ module.decoder.layers.23.self_attention.linear_kv_down_proj.weight
672
+ module.decoder.layers.23.input_layernorm.weight
673
+ module.decoder.layers.22.mlp.shared_experts.linear_fc2.weight
674
+ module.decoder.layers.22.mlp.shared_experts.linear_fc1.weight
675
+ Params for bucket 3 (40901632 elements, 40901632 padded size):
676
+ module.decoder.layers.22.mlp.router.weight
677
+ module.decoder.layers.22.self_attention.linear_kv_down_proj.weight
678
+ module.decoder.layers.22.self_attention.linear_kv_up_proj.layer_norm_weight
679
+ module.decoder.layers.22.self_attention.linear_q_proj.weight
680
+ module.decoder.layers.22.self_attention.linear_proj.weight
681
+ module.decoder.layers.22.input_layernorm.weight
682
+ module.decoder.layers.21.mlp.shared_experts.linear_fc2.weight
683
+ module.decoder.layers.21.mlp.shared_experts.linear_fc1.weight
684
+ module.decoder.layers.21.pre_mlp_layernorm.weight
685
+ module.decoder.layers.21.self_attention.linear_kv_up_proj.layer_norm_weight
686
+ module.decoder.layers.21.self_attention.linear_kv_down_proj.weight
687
+ module.decoder.layers.21.self_attention.linear_q_proj.weight
688
+ module.decoder.layers.22.self_attention.linear_kv_up_proj.weight
689
+ module.decoder.layers.22.pre_mlp_layernorm.weight
690
+ module.decoder.layers.21.mlp.router.weight
691
+ module.decoder.layers.21.self_attention.linear_kv_up_proj.weight
692
+ Params for bucket 4 (41163264 elements, 41163264 padded size):
693
+ module.decoder.layers.20.mlp.shared_experts.linear_fc1.weight
694
+ module.decoder.layers.20.pre_mlp_layernorm.weight
695
+ module.decoder.layers.20.self_attention.linear_kv_up_proj.layer_norm_weight
696
+ module.decoder.layers.20.self_attention.linear_kv_down_proj.weight
697
+ module.decoder.layers.20.self_attention.linear_q_proj.weight
698
+ module.decoder.layers.20.self_attention.linear_proj.weight
699
+ module.decoder.layers.21.self_attention.linear_proj.weight
700
+ module.decoder.layers.20.mlp.router.weight
701
+ module.decoder.layers.21.input_layernorm.weight
702
+ module.decoder.layers.20.mlp.shared_experts.linear_fc2.weight
703
+ module.decoder.layers.20.self_attention.linear_kv_up_proj.weight
704
+ module.decoder.layers.20.input_layernorm.weight
705
+ module.decoder.layers.19.mlp.shared_experts.linear_fc2.weight
706
+ Params for bucket 5 (42734080 elements, 42734080 padded size):
707
+ module.decoder.layers.19.pre_mlp_layernorm.weight
708
+ module.decoder.layers.19.self_attention.linear_kv_up_proj.layer_norm_weight
709
+ module.decoder.layers.19.self_attention.linear_kv_down_proj.weight
710
+ module.decoder.layers.19.mlp.shared_experts.linear_fc1.weight
711
+ module.decoder.layers.19.mlp.router.weight
712
+ module.decoder.layers.19.self_attention.linear_kv_up_proj.weight
713
+ module.decoder.layers.19.self_attention.linear_q_proj.weight
714
+ module.decoder.layers.19.self_attention.linear_proj.weight
715
+ module.decoder.layers.19.input_layernorm.weight
716
+ module.decoder.layers.18.mlp.shared_experts.linear_fc2.weight
717
+ module.decoder.layers.18.mlp.shared_experts.linear_fc1.weight
718
+ Params for bucket 6 (40901632 elements, 40901632 padded size):
719
+ module.decoder.layers.18.mlp.router.weight
720
+ module.decoder.layers.18.self_attention.linear_kv_up_proj.weight
721
+ module.decoder.layers.17.pre_mlp_layernorm.weight
722
+ module.decoder.layers.17.self_attention.linear_kv_up_proj.layer_norm_weight
723
+ module.decoder.layers.17.self_attention.linear_kv_down_proj.weight
724
+ module.decoder.layers.17.self_attention.linear_q_proj.weight
725
+ module.decoder.layers.18.self_attention.linear_kv_up_proj.layer_norm_weight
726
+ module.decoder.layers.18.self_attention.linear_kv_down_proj.weight
727
+ module.decoder.layers.18.pre_mlp_layernorm.weight
728
+ module.decoder.layers.18.self_attention.linear_q_proj.weight
729
+ module.decoder.layers.18.self_attention.linear_proj.weight
730
+ module.decoder.layers.18.input_layernorm.weight
731
+ module.decoder.layers.17.mlp.shared_experts.linear_fc2.weight
732
+ module.decoder.layers.17.mlp.shared_experts.linear_fc1.weight
733
+ module.decoder.layers.17.mlp.router.weight
734
+ module.decoder.layers.17.self_attention.linear_kv_up_proj.weight
735
+ Params for bucket 7 (41163264 elements, 41163264 padded size):
736
+ module.decoder.layers.16.pre_mlp_layernorm.weight
737
+ module.decoder.layers.16.self_attention.linear_kv_up_proj.layer_norm_weight
738
+ module.decoder.layers.16.self_attention.linear_kv_down_proj.weight
739
+ module.decoder.layers.16.self_attention.linear_q_proj.weight
740
+ module.decoder.layers.16.self_attention.linear_proj.weight
741
+ module.decoder.layers.16.input_layernorm.weight
742
+ module.decoder.layers.15.mlp.shared_experts.linear_fc2.weight
743
+ module.decoder.layers.17.self_attention.linear_proj.weight
744
+ module.decoder.layers.17.input_layernorm.weight
745
+ module.decoder.layers.16.mlp.shared_experts.linear_fc1.weight
746
+ module.decoder.layers.16.mlp.shared_experts.linear_fc2.weight
747
+ module.decoder.layers.16.mlp.router.weight
748
+ module.decoder.layers.16.self_attention.linear_kv_up_proj.weight
749
+ Params for bucket 8 (42734080 elements, 42734080 padded size):
750
+ module.decoder.layers.15.mlp.shared_experts.linear_fc1.weight
751
+ module.decoder.layers.15.mlp.router.weight
752
+ module.decoder.layers.15.self_attention.linear_kv_up_proj.weight
753
+ module.decoder.layers.15.self_attention.linear_q_proj.weight
754
+ module.decoder.layers.15.self_attention.linear_proj.weight
755
+ module.decoder.layers.15.input_layernorm.weight
756
+ module.decoder.layers.14.mlp.shared_experts.linear_fc2.weight
757
+ module.decoder.layers.14.mlp.shared_experts.linear_fc1.weight
758
+ module.decoder.layers.15.self_attention.linear_kv_up_proj.layer_norm_weight
759
+ module.decoder.layers.15.pre_mlp_layernorm.weight
760
+ module.decoder.layers.15.self_attention.linear_kv_down_proj.weight
761
+ Params for bucket 9 (40901632 elements, 40901632 padded size):
762
+ module.decoder.layers.14.mlp.router.weight
763
+ module.decoder.layers.14.self_attention.linear_kv_up_proj.weight
764
+ module.decoder.layers.14.self_attention.linear_q_proj.weight
765
+ module.decoder.layers.14.self_attention.linear_proj.weight
766
+ module.decoder.layers.14.input_layernorm.weight
767
+ module.decoder.layers.13.mlp.shared_experts.linear_fc2.weight
768
+ module.decoder.layers.13.mlp.shared_experts.linear_fc1.weight
769
+ module.decoder.layers.13.mlp.router.weight
770
+ module.decoder.layers.13.self_attention.linear_kv_up_proj.weight
771
+ module.decoder.layers.13.self_attention.linear_q_proj.weight
772
+ module.decoder.layers.14.self_attention.linear_kv_down_proj.weight
773
+ module.decoder.layers.14.self_attention.linear_kv_up_proj.layer_norm_weight
774
+ module.decoder.layers.14.pre_mlp_layernorm.weight
775
+ module.decoder.layers.13.pre_mlp_layernorm.weight
776
+ module.decoder.layers.13.self_attention.linear_kv_up_proj.layer_norm_weight
777
+ module.decoder.layers.13.self_attention.linear_kv_down_proj.weight
778
+ Params for bucket 10 (41163264 elements, 41163264 padded size):
779
+ module.decoder.layers.13.self_attention.linear_proj.weight
780
+ module.decoder.layers.12.mlp.router.weight
781
+ module.decoder.layers.13.input_layernorm.weight
782
+ module.decoder.layers.12.mlp.shared_experts.linear_fc2.weight
783
+ module.decoder.layers.12.mlp.shared_experts.linear_fc1.weight
784
+ module.decoder.layers.12.self_attention.linear_kv_up_proj.weight
785
+ module.decoder.layers.12.self_attention.linear_q_proj.weight
786
+ module.decoder.layers.12.self_attention.linear_proj.weight
787
+ module.decoder.layers.12.input_layernorm.weight
788
+ module.decoder.layers.11.mlp.shared_experts.linear_fc2.weight
789
+ module.decoder.layers.12.pre_mlp_layernorm.weight
790
+ module.decoder.layers.12.self_attention.linear_kv_up_proj.layer_norm_weight
791
+ module.decoder.layers.12.self_attention.linear_kv_down_proj.weight
792
+ Params for bucket 11 (42734080 elements, 42734080 padded size):
793
+ module.decoder.layers.11.mlp.shared_experts.linear_fc1.weight
794
+ module.decoder.layers.11.pre_mlp_layernorm.weight
795
+ module.decoder.layers.11.self_attention.linear_kv_up_proj.layer_norm_weight
796
+ module.decoder.layers.11.self_attention.linear_kv_down_proj.weight
797
+ module.decoder.layers.11.self_attention.linear_q_proj.weight
798
+ module.decoder.layers.11.self_attention.linear_proj.weight
799
+ module.decoder.layers.11.input_layernorm.weight
800
+ module.decoder.layers.10.mlp.shared_experts.linear_fc2.weight
801
+ module.decoder.layers.10.mlp.shared_experts.linear_fc1.weight
802
+ module.decoder.layers.11.mlp.router.weight
803
+ module.decoder.layers.11.self_attention.linear_kv_up_proj.weight
804
+ Params for bucket 12 (40901632 elements, 40901632 padded size):
805
+ module.decoder.layers.10.self_attention.linear_kv_up_proj.layer_norm_weight
806
+ module.decoder.layers.10.pre_mlp_layernorm.weight
807
+ module.decoder.layers.9.pre_mlp_layernorm.weight
808
+ module.decoder.layers.9.self_attention.linear_kv_up_proj.layer_norm_weight
809
+ module.decoder.layers.9.self_attention.linear_kv_down_proj.weight
810
+ module.decoder.layers.9.self_attention.linear_q_proj.weight
811
+ module.decoder.layers.10.mlp.router.weight
812
+ module.decoder.layers.10.self_attention.linear_kv_down_proj.weight
813
+ module.decoder.layers.10.self_attention.linear_kv_up_proj.weight
814
+ module.decoder.layers.10.self_attention.linear_q_proj.weight
815
+ module.decoder.layers.10.self_attention.linear_proj.weight
816
+ module.decoder.layers.10.input_layernorm.weight
817
+ module.decoder.layers.9.mlp.shared_experts.linear_fc2.weight
818
+ module.decoder.layers.9.mlp.shared_experts.linear_fc1.weight
819
+ module.decoder.layers.9.mlp.router.weight
820
+ module.decoder.layers.9.self_attention.linear_kv_up_proj.weight
821
+ Params for bucket 13 (41163264 elements, 41163264 padded size):
822
+ module.decoder.layers.8.mlp.router.weight
823
+ module.decoder.layers.8.self_attention.linear_kv_up_proj.weight
824
+ module.decoder.layers.7.mlp.shared_experts.linear_fc2.weight
825
+ module.decoder.layers.9.self_attention.linear_proj.weight
826
+ module.decoder.layers.8.mlp.shared_experts.linear_fc2.weight
827
+ module.decoder.layers.9.input_layernorm.weight
828
+ module.decoder.layers.8.mlp.shared_experts.linear_fc1.weight
829
+ module.decoder.layers.8.pre_mlp_layernorm.weight
830
+ module.decoder.layers.8.self_attention.linear_kv_up_proj.layer_norm_weight
831
+ module.decoder.layers.8.self_attention.linear_kv_down_proj.weight
832
+ module.decoder.layers.8.self_attention.linear_q_proj.weight
833
+ module.decoder.layers.8.self_attention.linear_proj.weight
834
+ module.decoder.layers.8.input_layernorm.weight
835
+ Params for bucket 14 (42734080 elements, 42734080 padded size):
836
+ module.decoder.layers.7.mlp.router.weight
837
+ module.decoder.layers.7.self_attention.linear_kv_up_proj.weight
838
+ module.decoder.layers.6.mlp.shared_experts.linear_fc1.weight
839
+ module.decoder.layers.7.mlp.shared_experts.linear_fc1.weight
840
+ module.decoder.layers.7.pre_mlp_layernorm.weight
841
+ module.decoder.layers.7.self_attention.linear_kv_up_proj.layer_norm_weight
842
+ module.decoder.layers.7.self_attention.linear_kv_down_proj.weight
843
+ module.decoder.layers.7.self_attention.linear_q_proj.weight
844
+ module.decoder.layers.7.self_attention.linear_proj.weight
845
+ module.decoder.layers.7.input_layernorm.weight
846
+ module.decoder.layers.6.mlp.shared_experts.linear_fc2.weight
847
+ Params for bucket 15 (40901632 elements, 40901632 padded size):
848
+ module.decoder.layers.6.mlp.router.weight
849
+ module.decoder.layers.6.self_attention.linear_kv_down_proj.weight
850
+ module.decoder.layers.6.self_attention.linear_kv_up_proj.weight
851
+ module.decoder.layers.5.pre_mlp_layernorm.weight
852
+ module.decoder.layers.5.self_attention.linear_kv_up_proj.layer_norm_weight
853
+ module.decoder.layers.5.self_attention.linear_kv_down_proj.weight
854
+ module.decoder.layers.6.self_attention.linear_kv_up_proj.layer_norm_weight
855
+ module.decoder.layers.6.pre_mlp_layernorm.weight
856
+ module.decoder.layers.6.self_attention.linear_q_proj.weight
857
+ module.decoder.layers.6.self_attention.linear_proj.weight
858
+ module.decoder.layers.6.input_layernorm.weight
859
+ module.decoder.layers.5.mlp.shared_experts.linear_fc2.weight
860
+ module.decoder.layers.5.mlp.shared_experts.linear_fc1.weight
861
+ module.decoder.layers.5.mlp.router.weight
862
+ module.decoder.layers.5.self_attention.linear_kv_up_proj.weight
863
+ module.decoder.layers.5.self_attention.linear_q_proj.weight
864
+ Params for bucket 16 (41163264 elements, 41163264 padded size):
865
+ module.decoder.layers.4.pre_mlp_layernorm.weight
866
+ module.decoder.layers.4.self_attention.linear_kv_up_proj.layer_norm_weight
867
+ module.decoder.layers.4.self_attention.linear_kv_down_proj.weight
868
+ module.decoder.layers.4.self_attention.linear_q_proj.weight
869
+ module.decoder.layers.4.self_attention.linear_proj.weight
870
+ module.decoder.layers.4.input_layernorm.weight
871
+ module.decoder.layers.3.mlp.shared_experts.linear_fc2.weight
872
+ module.decoder.layers.5.self_attention.linear_proj.weight
873
+ module.decoder.layers.4.mlp.shared_experts.linear_fc2.weight
874
+ module.decoder.layers.4.mlp.shared_experts.linear_fc1.weight
875
+ module.decoder.layers.4.mlp.router.weight
876
+ module.decoder.layers.5.input_layernorm.weight
877
+ module.decoder.layers.4.self_attention.linear_kv_up_proj.weight
878
+ Params for bucket 17 (42734080 elements, 42734080 padded size):
879
+ module.decoder.layers.3.mlp.shared_experts.linear_fc1.weight
880
+ module.decoder.layers.3.self_attention.linear_kv_up_proj.layer_norm_weight
881
+ module.decoder.layers.3.pre_mlp_layernorm.weight
882
+ module.decoder.layers.3.self_attention.linear_kv_down_proj.weight
883
+ module.decoder.layers.3.self_attention.linear_q_proj.weight
884
+ module.decoder.layers.3.self_attention.linear_proj.weight
885
+ module.decoder.layers.3.input_layernorm.weight
886
+ module.decoder.layers.2.mlp.shared_experts.linear_fc2.weight
887
+ module.decoder.layers.2.mlp.shared_experts.linear_fc1.weight
888
+ module.decoder.layers.3.self_attention.linear_kv_up_proj.weight
889
+ module.decoder.layers.3.mlp.router.weight
890
+ Params for bucket 18 (40901632 elements, 40901632 padded size):
891
+ module.decoder.layers.2.mlp.router.weight
892
+ module.decoder.layers.2.self_attention.linear_kv_up_proj.weight
893
+ module.decoder.layers.1.mlp.shared_experts.linear_fc2.weight
894
+ module.decoder.layers.1.mlp.shared_experts.linear_fc1.weight
895
+ module.decoder.layers.1.pre_mlp_layernorm.weight
896
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.layer_norm_weight
897
+ module.decoder.layers.2.pre_mlp_layernorm.weight
898
+ module.decoder.layers.2.self_attention.linear_kv_down_proj.weight
899
+ module.decoder.layers.2.self_attention.linear_kv_up_proj.layer_norm_weight
900
+ module.decoder.layers.2.self_attention.linear_q_proj.weight
901
+ module.decoder.layers.2.self_attention.linear_proj.weight
902
+ module.decoder.layers.2.input_layernorm.weight
903
+ module.decoder.layers.1.mlp.router.weight
904
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.weight
905
+ module.decoder.layers.1.self_attention.linear_kv_down_proj.weight
906
+ module.decoder.layers.1.self_attention.linear_q_proj.weight
907
+ Params for bucket 19 (71436288 elements, 71436288 padded size):
908
+ module.decoder.layers.1.self_attention.linear_proj.weight
909
+ module.decoder.layers.1.input_layernorm.weight
910
+ module.decoder.layers.0.mlp.linear_fc1.weight
911
+ module.decoder.layers.0.mlp.linear_fc2.weight
912
+ Params for bucket 20 (116789760 elements, 116789760 padded size):
913
+ module.decoder.layers.0.self_attention.linear_kv_down_proj.weight
914
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.weight
915
+ module.embedding.word_embeddings.weight
916
+ module.decoder.layers.0.mlp.linear_fc1.layer_norm_weight
917
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.layer_norm_weight
918
+ module.decoder.layers.0.self_attention.linear_q_proj.weight
919
+ module.decoder.layers.0.self_attention.linear_proj.weight
920
+ module.decoder.layers.0.input_layernorm.weight
921
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 40
922
+ Params for bucket 1 (40370176 elements, 40370176 padded size):
923
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight7
924
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight5
925
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight3
926
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight1
927
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight6
928
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight4
929
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight6
930
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight2
931
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight0
932
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight7
933
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight5
934
+ Params for bucket 2 (40370176 elements, 40370176 padded size):
935
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight4
936
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight0
937
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight2
938
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight6
939
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight4
940
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight1
941
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight3
942
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight7
943
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight5
944
+ Params for bucket 3 (40370176 elements, 40370176 padded size):
945
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight0
946
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight2
947
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight7
948
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight5
949
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight3
950
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight3
951
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight1
952
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight6
953
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight4
954
+ Params for bucket 4 (40370176 elements, 40370176 padded size):
955
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight6
956
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight1
957
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight4
958
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight2
959
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight0
960
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight2
961
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight0
962
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight7
963
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight5
964
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight3
965
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight1
966
+ Params for bucket 5 (40370176 elements, 40370176 padded size):
967
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight7
968
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight5
969
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight3
970
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight1
971
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight6
972
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight4
973
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight2
974
+ Params for bucket 6 (40370176 elements, 40370176 padded size):
975
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight6
976
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight4
977
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight2
978
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight0
979
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight7
980
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight0
981
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight7
982
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight5
983
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight3
984
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight1
985
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight6
986
+ Params for bucket 7 (40370176 elements, 40370176 padded size):
987
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight5
988
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight1
989
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight3
990
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight6
991
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight4
992
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight2
993
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight0
994
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight7
995
+ Params for bucket 8 (40370176 elements, 40370176 padded size):
996
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight4
997
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight2
998
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight0
999
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight7
1000
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight5
1001
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight5
1002
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight3
1003
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight1
1004
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight6
1005
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight4
1006
+ Params for bucket 9 (40370176 elements, 40370176 padded size):
1007
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight3
1008
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight7
1009
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight1
1010
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight5
1011
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight3
1012
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight2
1013
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight0
1014
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight6
1015
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight4
1016
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight2
1017
+ Params for bucket 10 (40370176 elements, 40370176 padded size):
1018
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight1
1019
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight5
1020
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight7
1021
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight3
1022
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight6
1023
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight0
1024
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight4
1025
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight2
1026
+ Params for bucket 11 (40370176 elements, 40370176 padded size):
1027
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight1
1028
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight7
1029
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight5
1030
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight3
1031
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight1
1032
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight0
1033
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight6
1034
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight4
1035
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight2
1036
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight0
1037
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight7
1038
+ Params for bucket 12 (40370176 elements, 40370176 padded size):
1039
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight6
1040
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight2
1041
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight4
1042
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight0
1043
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight5
1044
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight3
1045
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight1
1046
+ Params for bucket 13 (40370176 elements, 40370176 padded size):
1047
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight7
1048
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight3
1049
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight5
1050
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight1
1051
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight6
1052
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight6
1053
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight4
1054
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight2
1055
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight0
1056
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight7
1057
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight5
1058
+ Params for bucket 14 (40370176 elements, 40370176 padded size):
1059
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight4
1060
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight0
1061
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight2
1062
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight6
1063
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight4
1064
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight3
1065
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight1
1066
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight7
1067
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight5
1068
+ Params for bucket 15 (40370176 elements, 40370176 padded size):
1069
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight7
1070
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight2
1071
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight0
1072
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight5
1073
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight3
1074
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight3
1075
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight1
1076
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight6
1077
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight4
1078
+ Params for bucket 16 (40370176 elements, 40370176 padded size):
1079
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight1
1080
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight6
1081
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight4
1082
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight2
1083
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight0
1084
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight2
1085
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight7
1086
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight0
1087
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight5
1088
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight3
1089
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight1
1090
+ Params for bucket 17 (40370176 elements, 40370176 padded size):
1091
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight7
1092
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight3
1093
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight5
1094
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight1
1095
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight4
1096
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight6
1097
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight2
1098
+ Params for bucket 18 (40370176 elements, 40370176 padded size):
1099
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight6
1100
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight4
1101
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight2
1102
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight0
1103
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight7
1104
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight0
1105
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight5
1106
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight7
1107
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight3
1108
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight1
1109
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight6
1110
+ Params for bucket 19 (40370176 elements, 40370176 padded size):
1111
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight5
1112
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight3
1113
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight1
1114
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight6
1115
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight2
1116
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight4
1117
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight0
1118
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight7
1119
+ Params for bucket 20 (40370176 elements, 40370176 padded size):
1120
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight4
1121
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight2
1122
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight0
1123
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight7
1124
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight5
1125
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight5
1126
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight1
1127
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight3
1128
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight6
1129
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight4
1130
+ Params for bucket 21 (40370176 elements, 40370176 padded size):
1131
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight3
1132
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight1
1133
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight7
1134
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight5
1135
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight3
1136
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight2
1137
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight0
1138
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight6
1139
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight4
1140
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight2
1141
+ Params for bucket 22 (40370176 elements, 40370176 padded size):
1142
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight1
1143
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight6
1144
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight4
1145
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight2
1146
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight5
1147
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight0
1148
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight7
1149
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight3
1150
+ Params for bucket 23 (40370176 elements, 40370176 padded size):
1151
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight0
1152
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight7
1153
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight5
1154
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight3
1155
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight1
1156
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight1
1157
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight6
1158
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight4
1159
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight2
1160
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight0
1161
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight7
1162
+ Params for bucket 24 (40370176 elements, 40370176 padded size):
1163
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight6
1164
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight4
1165
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight2
1166
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight0
1167
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight5
1168
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight3
1169
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight1
1170
+ Params for bucket 25 (40370176 elements, 40370176 padded size):
1171
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight7
1172
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight5
1173
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight3
1174
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight1
1175
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight7
1176
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight5
1177
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight4
1178
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight6
1179
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight2
1180
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight0
1181
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight6
1182
+ Params for bucket 26 (40370176 elements, 40370176 padded size):
1183
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight3
1184
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight1
1185
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight6
1186
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight4
1187
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight4
1188
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight0
1189
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight2
1190
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight7
1191
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight5
1192
+ Params for bucket 27 (40370176 elements, 40370176 padded size):
1193
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight7
1194
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight2
1195
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight0
1196
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight5
1197
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight3
1198
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight3
1199
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight1
1200
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight6
1201
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight4
1202
+ Params for bucket 28 (40370176 elements, 40370176 padded size):
1203
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight1
1204
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight6
1205
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight4
1206
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight2
1207
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight0
1208
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight2
1209
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight7
1210
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight0
1211
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight5
1212
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight3
1213
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight1
1214
+ Params for bucket 29 (40370176 elements, 40370176 padded size):
1215
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight7
1216
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight3
1217
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight5
1218
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight1
1219
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight4
1220
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight6
1221
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight2
1222
+ Params for bucket 30 (40370176 elements, 40370176 padded size):
1223
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight5
1224
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight7
1225
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight3
1226
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight1
1227
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight7
1228
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight0
1229
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight6
1230
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight4
1231
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight2
1232
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight0
1233
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight6
1234
+ Params for bucket 31 (40370176 elements, 40370176 padded size):
1235
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight5
1236
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight1
1237
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight3
1238
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight7
1239
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight2
1240
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight4
1241
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight0
1242
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight6
1243
+ Params for bucket 32 (40370176 elements, 40370176 padded size):
1244
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight5
1245
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight3
1246
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight1
1247
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight7
1248
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight5
1249
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight4
1250
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight2
1251
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight0
1252
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight6
1253
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight4
1254
+ Params for bucket 33 (40370176 elements, 40370176 padded size):
1255
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight3
1256
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight7
1257
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight1
1258
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight5
1259
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight3
1260
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight0
1261
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight2
1262
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight6
1263
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight4
1264
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight2
1265
+ Params for bucket 34 (40370176 elements, 40370176 padded size):
1266
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight1
1267
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight6
1268
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight4
1269
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight2
1270
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight5
1271
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight7
1272
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight0
1273
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight3
1274
+ Params for bucket 35 (40370176 elements, 40370176 padded size):
1275
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight6
1276
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight0
1277
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight4
1278
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight2
1279
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight0
1280
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight1
1281
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight5
1282
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight7
1283
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight3
1284
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight1
1285
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight7
1286
+ Params for bucket 36 (40370176 elements, 40370176 padded size):
1287
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight6
1288
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight4
1289
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight2
1290
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight0
1291
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight5
1292
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight3
1293
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight1
1294
+ Params for bucket 37 (40370176 elements, 40370176 padded size):
1295
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight6
1296
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight4
1297
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight2
1298
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight0
1299
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight7
1300
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight5
1301
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight7
1302
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight5
1303
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight3
1304
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight1
1305
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight6
1306
+ Params for bucket 38 (40370176 elements, 40370176 padded size):
1307
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight3
1308
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight1
1309
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight6
1310
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight4
1311
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight4
1312
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight0
1313
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight2
1314
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight7
1315
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight5
1316
+ Params for bucket 39 (40370176 elements, 40370176 padded size):
1317
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight0
1318
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight2
1319
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight7
1320
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight5
1321
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight3
1322
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight3
1323
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight1
1324
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight6
1325
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight4
1326
+ Params for bucket 40 (17301504 elements, 17301504 padded size):
1327
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight2
1328
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight1
1329
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight0
1330
+ 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 0x748284166090>, config_logger_dir='')
1331
+ INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine
1332
+ [after model, optimizer, and learning rate scheduler are built] datetime: 2026-08-26 18:43:51
1333
+ > building train, validation, and test datasets ...
1334
+ > datasets target sizes (minimum size):
1335
+ train: 96
1336
+ validation: 0
1337
+ test: 0
1338
+ INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None]
1339
+ > building train, validation, and test datasets for GPT ...
1340
+ INFO:megatron.core.datasets.blended_megatron_dataset_builder:Building GPTDataset splits with sizes=(96, 0, 0) and config=GPTDatasetConfig(random_seed=1234, sequence_length=8192, blend=(['/workspace/data/wikitext/wikitext103_gpt2_text_document'], None), blend_per_split=None, split='100,0,0', split_matrix=[(0, 1.0), None, None], num_dataset_builder_threads=1, path_to_cache=None, mmap_bin_files=True, mock=False, tokenizer=<megatron.training.tokenizer.tokenizer._GPT2BPETokenizer object at 0x74828553b200>, 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)
1341
+ INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx
1342
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths
1343
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers
1344
+ INFO:megatron.core.datasets.indexed_dataset: Extract the document indices
1345
+ INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029
1346
+ INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350
1347
+ INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices
1348
+ INFO:megatron.core.datasets.gpt_dataset: Load the document index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-document_index.npy
1349
+ INFO:megatron.core.datasets.gpt_dataset: Load the sample index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-sample_index.npy
1350
+ INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-shuffle_index.npy
1351
+ INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 14536
1352
+ > finished creating GPT datasets ...
1353
+ [after dataloaders are built] datetime: 2026-08-26 18:43:51
1354
+ done with setup ...
1355
+ (min, max) time across ranks (ms):
1356
+ model-and-optimizer-setup ......................: (3246.94, 3277.11)
1357
+ train/valid/test-data-iterators-setup ..........: (59.73, 73.68)
1358
+ training ...
1359
+ Overwriting rerun_state_machine.current_iteration from -1 to 0...
1360
+ [before the start of training step] datetime: 2026-08-26 18:43:51
1361
+ INFO:moonep.api:MoonEP comm buffer sizes:
1362
+ config : R=8, S=8192, H=2048, E=64, K=6, NvS=51184, NvS_padded=51200
1363
+ sm_count : num_sms=32, num_sms_dedup=148
1364
+ hidden_buffer_size : 200.00 MiB
1365
+ meta_buffer_size : 2.00 MiB
1366
+ local_temp_buffer_size : 2.01 MiB
1367
+ total_size : 204.01 MiB
1368
+ Number of parameters in transformer block in billions: 13.53 [2026-08-26 18:44:13] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 21927.2 | throughput per GPU (TFLOP/s/GPU): 5.8 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.092206E+01 | loss scale: 1.0 | grad norm: 30.650 | number of skipped iterations: 0 | number of nan iterations: 0 |
1369
+
1370
+ Number of parameters in embedding layers in billions: 0.21
1371
+ Total number of parameters in billions: 13.74
1372
+ Number of parameters in most loaded shard in billions: 13.7356
1373
+ Theoretical memory footprints: weight and optimizer=98244.96 MB
1374
+ [Rank 0] (after 1 iterations) memory (MB) | allocated: 66144.1884765625 | max allocated: 124949.08837890625 | reserved: 127502.0 | max reserved: 127502.0
1375
+ [2026-08-26 18:44:13] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 724.9 | throughput per GPU (TFLOP/s/GPU): 176.0 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.091986E+01 | loss scale: 1.0 | grad norm: 30.565 | number of skipped iterations: 0 | number of nan iterations: 0 |
1376
+ [2026-08-26 18:44:14] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 721.6 | throughput per GPU (TFLOP/s/GPU): 176.8 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 9.956084E+00 | loss scale: 1.0 | grad norm: 42.533 | number of skipped iterations: 0 | number of nan iterations: 0 |
1377
+ Capture CUDA graph for training!!!
1378
+ CUDA graph capture done!!!
1379
+ [2026-08-26 18:44:16] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 1861.4 | throughput per GPU (TFLOP/s/GPU): 68.5 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 9.478545E+00 | loss scale: 1.0 | grad norm: 7.152 | number of skipped iterations: 0 | number of nan iterations: 0 |
1380
+ [2026-08-26 18:44:17] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 701.3 | throughput per GPU (TFLOP/s/GPU): 181.9 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.091896E+01 | loss scale: 1.0 | grad norm: 452.732 | number of skipped iterations: 0 | number of nan iterations: 0 |
1381
+ [2026-08-26 18:44:17] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 703.0 | throughput per GPU (TFLOP/s/GPU): 181.5 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 9.375167E+00 | loss scale: 1.0 | grad norm: 4.992 | number of skipped iterations: 0 | number of nan iterations: 0 |
1382
+ [2026-08-26 18:44:18] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 700.3 | throughput per GPU (TFLOP/s/GPU): 182.2 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 8.982457E+00 | loss scale: 1.0 | grad norm: 3.717 | number of skipped iterations: 0 | number of nan iterations: 0 |
1383
+ [2026-08-26 18:44:19] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 700.7 | throughput per GPU (TFLOP/s/GPU): 182.1 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.787875E+00 | loss scale: 1.0 | grad norm: 4.876 | number of skipped iterations: 0 | number of nan iterations: 0 |
1384
+ [2026-08-26 18:44:20] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 706.3 | throughput per GPU (TFLOP/s/GPU): 180.6 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.620937E+00 | loss scale: 1.0 | grad norm: 2.990 | number of skipped iterations: 0 | number of nan iterations: 0 |
1385
+ [2026-08-26 18:44:20] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 701.1 | throughput per GPU (TFLOP/s/GPU): 182.0 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.680684E+00 | loss scale: 1.0 | grad norm: 194.973 | number of skipped iterations: 0 | number of nan iterations: 0 |
1386
+ [2026-08-26 18:44:21] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 701.4 | throughput per GPU (TFLOP/s/GPU): 181.9 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.583418E+00 | loss scale: 1.0 | grad norm: 24.333 | number of skipped iterations: 0 | number of nan iterations: 0 |
1387
+ [2026-08-26 18:44:22] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 702.6 | throughput per GPU (TFLOP/s/GPU): 181.6 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.611391E+00 | loss scale: 1.0 | grad norm: 75.148 | number of skipped iterations: 0 | number of nan iterations: 0 |
1388
+ [after training is done] datetime: 2026-08-26 18:44:22
1389
+ [rank0]:[W826 18:44:23.872992215 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())
1390
+ W0826 18:44:32.023000 549952 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers
1391
+ W0826 18:44:32.024000 549952 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 550040 closing signal SIGTERM
1392
+ W0826 18:44:32.024000 549952 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 550041 closing signal SIGTERM
1393
+ W0826 18:44:32.027000 549952 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 550042 closing signal SIGTERM
1394
+ W0826 18:44:32.032000 549952 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 550043 closing signal SIGTERM
1395
+ W0826 18:44:32.034000 549952 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 550044 closing signal SIGTERM
1396
+ W0826 18:44:32.036000 549952 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 550045 closing signal SIGTERM
1397
+ W0826 18:44:32.037000 549952 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 550046 closing signal SIGTERM
1398
+ W0826 18:44:32.039000 549952 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 550047 closing signal SIGTERM
1399
+ Traceback (most recent call last):
1400
+ File "/usr/local/bin/torchrun", line 8, in <module>
1401
+ sys.exit(main())
1402
+ ^^^^^^
1403
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
1404
+ return f(*args, **kwargs)
1405
+ ^^^^^^^^^^^^^^^^^^
1406
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 892, in main
1407
+ run(args)
1408
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 883, in run
1409
+ elastic_launch(
1410
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
1411
+ return launch_agent(self._config, self._entrypoint, list(args))
1412
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1413
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 261, in launch_agent
1414
+ result = agent.run()
1415
+ ^^^^^^^^^^^
1416
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper
1417
+ result = f(*args, **kwargs)
1418
+ ^^^^^^^^^^^^^^^^^^
1419
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 711, in run
1420
+ result = self._invoke_run(role)
1421
+ ^^^^^^^^^^^^^^^^^^^^^^
1422
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 870, in _invoke_run
1423
+ time.sleep(monitor_interval)
1424
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 84, in _terminate_process_handler
1425
+ raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval)
1426
+ torch.distributed.elastic.multiprocessing.api.SignalException: Process 549952 got signal: 15
moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep3/mem_samples.csv ADDED
@@ -0,0 +1,232 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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, 0
10
+ 1, 0
11
+ 2, 0
12
+ 3, 0
13
+ 4, 0
14
+ 5, 0
15
+ 6, 0
16
+ 7, 0
17
+ 0, 1
18
+ 1, 1
19
+ 2, 1
20
+ 3, 1
21
+ 4, 1
22
+ 5, 1
23
+ 6, 1
24
+ 7, 1
25
+ 0, 4
26
+ 1, 4
27
+ 2, 4
28
+ 3, 4
29
+ 4, 1
30
+ 5, 1
31
+ 6, 1
32
+ 7, 1
33
+ 0, 4
34
+ 1, 4
35
+ 2, 4
36
+ 3, 4
37
+ 4, 4
38
+ 5, 4
39
+ 6, 4
40
+ 7, 4
41
+ 0, 4
42
+ 1, 26
43
+ 2, 27
44
+ 3, 27
45
+ 4, 629
46
+ 5, 27
47
+ 6, 27
48
+ 7, 629
49
+ 0, 867
50
+ 1, 885
51
+ 2, 869
52
+ 3, 879
53
+ 4, 875
54
+ 5, 877
55
+ 6, 883
56
+ 7, 885
57
+ 0, 1955
58
+ 1, 1955
59
+ 2, 1955
60
+ 3, 1955
61
+ 4, 1955
62
+ 5, 1955
63
+ 6, 1955
64
+ 7, 1955
65
+ 0, 3933
66
+ 1, 3949
67
+ 2, 3943
68
+ 3, 3937
69
+ 4, 3947
70
+ 5, 3923
71
+ 6, 3933
72
+ 7, 3915
73
+ 0, 31739
74
+ 1, 31739
75
+ 2, 31739
76
+ 3, 31739
77
+ 4, 31739
78
+ 5, 31739
79
+ 6, 31739
80
+ 7, 31739
81
+ 0, 35749
82
+ 1, 35869
83
+ 2, 35755
84
+ 3, 35807
85
+ 4, 35763
86
+ 5, 35835
87
+ 6, 35785
88
+ 7, 35919
89
+ 0, 45853
90
+ 1, 45961
91
+ 2, 45809
92
+ 3, 45881
93
+ 4, 45801
94
+ 5, 45871
95
+ 6, 45847
96
+ 7, 45945
97
+ 0, 46291
98
+ 1, 46411
99
+ 2, 46285
100
+ 3, 46343
101
+ 4, 46295
102
+ 5, 46341
103
+ 6, 46333
104
+ 7, 46411
105
+ 0, 46293
106
+ 1, 46415
107
+ 2, 46289
108
+ 3, 46349
109
+ 4, 46295
110
+ 5, 46345
111
+ 6, 46337
112
+ 7, 46415
113
+ 0, 116057
114
+ 1, 116047
115
+ 2, 116059
116
+ 3, 116045
117
+ 4, 116063
118
+ 5, 116043
119
+ 6, 116037
120
+ 7, 116049
121
+ 0, 139411
122
+ 1, 139403
123
+ 2, 139405
124
+ 3, 139401
125
+ 4, 139417
126
+ 5, 139397
127
+ 6, 140175
128
+ 7, 139403
129
+ 0, 139801
130
+ 1, 139745
131
+ 2, 139687
132
+ 3, 139769
133
+ 4, 139729
134
+ 5, 139757
135
+ 6, 140501
136
+ 7, 139753
137
+ 0, 139801
138
+ 1, 139793
139
+ 2, 139795
140
+ 3, 139789
141
+ 4, 139805
142
+ 5, 139787
143
+ 6, 140565
144
+ 7, 139793
145
+ 0, 142115
146
+ 1, 142107
147
+ 2, 142109
148
+ 3, 142103
149
+ 4, 142119
150
+ 5, 142101
151
+ 6, 142879
152
+ 7, 142107
153
+ 0, 143203
154
+ 1, 143179
155
+ 2, 143157
156
+ 3, 143193
157
+ 4, 143165
158
+ 5, 143177
159
+ 6, 143973
160
+ 7, 143149
161
+ 0, 108211
162
+ 1, 114361
163
+ 2, 104097
164
+ 3, 104093
165
+ 4, 104109
166
+ 5, 104089
167
+ 6, 104083
168
+ 7, 104095
169
+ 0, 165291
170
+ 1, 165281
171
+ 2, 165285
172
+ 3, 165279
173
+ 4, 165295
174
+ 5, 165275
175
+ 6, 165269
176
+ 7, 165281
177
+ 0, 165291
178
+ 1, 165281
179
+ 2, 165285
180
+ 3, 165279
181
+ 4, 165295
182
+ 5, 165275
183
+ 6, 165269
184
+ 7, 165281
185
+ 0, 165291
186
+ 1, 165281
187
+ 2, 165285
188
+ 3, 165279
189
+ 4, 165295
190
+ 5, 165275
191
+ 6, 165269
192
+ 7, 165281
193
+ 0, 165089
194
+ 1, 165079
195
+ 2, 165083
196
+ 3, 165077
197
+ 4, 165093
198
+ 5, 165073
199
+ 6, 165067
200
+ 7, 165079
201
+ 0, 165089
202
+ 1, 165079
203
+ 2, 165083
204
+ 3, 165077
205
+ 4, 165093
206
+ 5, 165073
207
+ 6, 165067
208
+ 7, 165079
209
+ 0, 165089
210
+ 1, 165079
211
+ 2, 165083
212
+ 3, 165077
213
+ 4, 165093
214
+ 5, 165073
215
+ 6, 165067
216
+ 7, 165079
217
+ 0, 165089
218
+ 1, 165079
219
+ 2, 165083
220
+ 3, 165077
221
+ 4, 165093
222
+ 5, 165073
223
+ 6, 165067
224
+ 7, 165079
225
+ 0, 10975
226
+ 1, 9376
227
+ 2, 9376
228
+ 3, 9376
229
+ 4, 9376
230
+ 5, 9376
231
+ 6, 9696
232
+ 7, 165077
moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep3/run.env ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=moonep
2
+ config=DSV2_LITE_L24
3
+ family=tierB
4
+ repo=/workspace/Megatron-MoE-Echo
5
+ branch=moonep_backend_bracing_regions_through_kernels
6
+ commit=1aac8bdd
7
+ num_layers=24
8
+ hidden=2048
9
+ ffn_hidden=10944
10
+ moe_ffn_hidden=1408
11
+ num_experts=64
12
+ topk=6
13
+ experts_per_rank=8
14
+ expert_params=8650752
15
+ heads=16
16
+ kv_groups=0
17
+ kv_channels=128
18
+ qk_layernorm=1
19
+ shared_ffn=2816
20
+ shared_expert_overlap=0
21
+ mla=1
22
+ dense_first_layers=1
23
+ moe_layers=23
24
+ rotary_base=10000
25
+ rotary_scaling=40
26
+ qkv_bias=0
27
+ router_balancing=seq_aux_loss
28
+ aux_loss_coeff=1e-3
29
+ router_pre_softmax=1
30
+ router_fusion=0
31
+ router_dtype=fp32
32
+ device_initiated_grouped_gemm=1
33
+ norm_epsilon=1e-6
34
+ seq_length=8192
35
+ micro_batch_size=1
36
+ global_batch_size=8
37
+ num_microbatches=1
38
+ assignments_per_rank=49152
39
+ mean_tokens_per_expert=6144
40
+ slot_policy=matched
41
+ slots_per_rank=8
42
+ capacity_mode=none
43
+ capacity=none
44
+ cuda_graph=1
45
+ probe_mode=0
46
+ async_grad_reduce=0
47
+ branch_role=timing
48
+ memory_dump_mode=none
49
+ memory_snapshot_path=
50
+ cuda_graph_warmup_steps=3
51
+ recompute=none
52
+ force_balance=0
53
+ train_iters=12
54
+ ep_size=8
55
+ echo_dispatch_overlap=1
56
+ echo_dispatch_recompute=0
57
+ ultraep_quota_min_tokens=
58
+ ultraep_weight_sync_join=
59
+ profile_enable=0
60
+ kernel_trace=0
61
+ fp8_recipe=mxfp8
moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep3/train.log ADDED
@@ -0,0 +1,1426 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ==============================================================
2
+ DSV2_LITE_L24 / moonep (slots/rank=8 policy=matched)
3
+ E=64 K=6 H'=1408 m=8
4
+ L=24 seq=8192 gbs=8 (1 microbatches)
5
+ cap_mode=none cap=none cg=1 recompute=none
6
+ memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep3/memory_snapshots/snapshot.pickle
7
+ repo=/workspace/Megatron-MoE-Echo @ 1aac8bdd
8
+ ==============================================================
9
+ W0826 18:44:39.574000 552370 torch/distributed/run.py:766]
10
+ W0826 18:44:39.574000 552370 torch/distributed/run.py:766] *****************************************
11
+ W0826 18:44:39.574000 552370 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
12
+ W0826 18:44:39.574000 552370 torch/distributed/run.py:766] *****************************************
13
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
14
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
15
+ 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
16
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
17
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
18
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
19
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
20
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
21
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
22
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
23
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
24
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
25
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
26
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
27
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
28
+ 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
29
+ WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer
30
+ Number of virtual stages per pipeline stage: None
31
+ accumulate and all-reduce gradients in fp32 for bfloat16 data type.
32
+ using torch.bfloat16 for parameters ...
33
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
34
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
35
+ ------------------------ arguments ------------------------
36
+ account_for_embedding_in_pipeline_split ......... False
37
+ account_for_loss_in_pipeline_split .............. False
38
+ accumulate_allreduce_grads_in_fp32 .............. True
39
+ adam_beta1 ...................................... 0.9
40
+ adam_beta2 ...................................... 0.999
41
+ adam_eps ........................................ 1e-08
42
+ add_bias_linear ................................. False
43
+ add_position_embedding .......................... False
44
+ add_qkv_bias .................................... False
45
+ adlr_autoresume ................................. False
46
+ adlr_autoresume_interval ........................ 1000
47
+ align_grad_reduce ............................... True
48
+ align_param_gather .............................. False
49
+ app_tag_run_name ................................ None
50
+ app_tag_run_version ............................. 0.0.0
51
+ apply_layernorm_1p .............................. False
52
+ apply_query_key_layer_scaling ................... False
53
+ apply_residual_connection_post_layernorm ........ False
54
+ apply_rope_fusion ............................... False
55
+ async_save ...................................... None
56
+ async_tensor_model_parallel_allreduce ........... True
57
+ attention_backend ............................... AttnBackend.auto
58
+ attention_dropout ............................... 0.0
59
+ attention_softmax_in_fp32 ....................... False
60
+ auto_detect_ckpt_format ......................... False
61
+ barrier_with_L1_time ............................ True
62
+ bert_binary_head ................................ True
63
+ bert_embedder_type .............................. megatron
64
+ bert_load ....................................... None
65
+ bf16 ............................................ True
66
+ bias_dropout_fusion ............................. True
67
+ bias_gelu_fusion ................................ False
68
+ bias_swiglu_fusion .............................. True
69
+ biencoder_projection_dim ........................ 0
70
+ biencoder_shared_query_context_model ............ False
71
+ block_data_path ................................. None
72
+ calc_ft_timeouts ................................ False
73
+ calculate_per_token_loss ........................ False
74
+ check_for_large_grads ........................... False
75
+ check_for_nan_in_loss_and_grad .................. False
76
+ check_for_spiky_loss ............................ False
77
+ check_weight_hash_across_dp_replicas_interval ... None
78
+ ckpt_assume_constant_structure .................. False
79
+ ckpt_convert_format ............................. None
80
+ ckpt_convert_save ............................... None
81
+ ckpt_convert_update_legacy_dist_opt_format ...... False
82
+ ckpt_format ..................................... torch_dist
83
+ ckpt_fully_parallel_load ........................ False
84
+ ckpt_fully_parallel_save ........................ True
85
+ ckpt_fully_parallel_save_deprecated ............. False
86
+ ckpt_step ....................................... None
87
+ classes_fraction ................................ 1.0
88
+ clip_grad ....................................... 1.0
89
+ clone_scatter_output_in_embedding ............... True
90
+ config_logger_dir ...............................
91
+ consumed_train_samples .......................... 0
92
+ consumed_valid_samples .......................... 0
93
+ context_parallel_size ........................... 1
94
+ cp_comm_type .................................... ['p2p']
95
+ create_attention_mask_in_dataloader ............. True
96
+ cross_entropy_fusion_impl ....................... native
97
+ cross_entropy_loss_fusion ....................... False
98
+ cuda_graph_scope ................................ full_iteration
99
+ cuda_graph_warmup_steps ......................... 3
100
+ data_args_path .................................. None
101
+ data_cache_path ................................. None
102
+ data_parallel_random_init ....................... False
103
+ data_parallel_sharding_strategy ................. no_shard
104
+ data_parallel_size .............................. 8
105
+ data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document']
106
+ data_per_class_fraction ......................... 1.0
107
+ data_sharding ................................... True
108
+ dataloader_type ................................. single
109
+ ddp_average_in_collective ....................... False
110
+ ddp_bucket_size ................................. None
111
+ ddp_num_buckets ................................. None
112
+ ddp_pad_buckets_for_high_nccl_busbw ............. False
113
+ decoder_first_pipeline_num_layers ............... None
114
+ decoder_last_pipeline_num_layers ................ None
115
+ decoder_num_layers .............................. None
116
+ decoder_seq_length .............................. None
117
+ decoupled_lr .................................... None
118
+ decoupled_min_lr ................................ None
119
+ decrease_batch_size_if_needed ................... False
120
+ defer_embedding_wgrad_compute ................... False
121
+ delay_wgrad_compute ............................. False
122
+ deprecated_use_mcore_models ..................... True
123
+ deterministic_mode .............................. False
124
+ dino_bottleneck_size ............................ 256
125
+ dino_freeze_last_layer .......................... 1
126
+ dino_head_hidden_size ........................... 2048
127
+ dino_local_crops_number ......................... 10
128
+ dino_local_img_size ............................. 96
129
+ dino_norm_last_layer ............................ False
130
+ dino_teacher_temp ............................... 0.07
131
+ dino_warmup_teacher_temp ........................ 0.04
132
+ dino_warmup_teacher_temp_epochs ................. 30
133
+ disable_bf16_reduced_precision_matmul ........... False
134
+ disable_mamba_mem_eff_path ...................... False
135
+ disable_straggler_on_startup .................... False
136
+ dist_ckpt_format_deprecated ..................... None
137
+ dist_ckpt_strictness ............................ assume_ok_unexpected
138
+ distribute_saved_activations .................... False
139
+ distributed_backend ............................. nccl
140
+ distributed_timeout_minutes ..................... 10
141
+ embedding_init_method_std ....................... None
142
+ embedding_path .................................. None
143
+ empty_unused_memory_level ....................... 0
144
+ enable_cuda_graph ............................... True
145
+ enable_experimental ............................. False
146
+ enable_ft_package ............................... False
147
+ enable_gloo_process_groups ...................... True
148
+ enable_msc ...................................... True
149
+ enable_one_logger ............................... True
150
+ encoder_num_layers .............................. 24
151
+ encoder_seq_length .............................. 8192
152
+ end_weight_decay ................................ 0.1
153
+ eod_mask_loss ................................... False
154
+ error_injection_rate ............................ 0
155
+ error_injection_type ............................ transient_error
156
+ eval_interval ................................... 1000000
157
+ eval_iters ...................................... 0
158
+ evidence_data_path .............................. None
159
+ exit_duration_in_mins ........................... None
160
+ exit_interval ................................... None
161
+ exit_on_missing_checkpoint ...................... False
162
+ exit_signal_handler ............................. False
163
+ exp_avg_dtype ................................... torch.float32
164
+ exp_avg_sq_dtype ................................ torch.float32
165
+ expert_model_parallel_size ...................... 8
166
+ expert_tensor_parallel_size ..................... 1
167
+ export_force_local_attention .................... False
168
+ export_kd_cfg ................................... None
169
+ export_kd_teacher_ckpt_format ................... None
170
+ export_kd_teacher_load .......................... None
171
+ export_kv_cache_quant ........................... False
172
+ export_legacy_megatron .......................... False
173
+ export_model_type ............................... GPTModel
174
+ export_moe_apply_probs_on_input ................. False
175
+ export_qk_l2_norm ............................... False
176
+ export_quant_cfg ................................ None
177
+ export_real_quant_cfg ........................... None
178
+ export_te_mcore_model ........................... False
179
+ external_cuda_graph ............................. False
180
+ ffn_hidden_size ................................. 10944
181
+ finetune ........................................ False
182
+ finetune_data_split ............................. train
183
+ finetune_hf_dataset ............................. None
184
+ first_last_layers_bf16 .......................... False
185
+ flash_decode .................................... False
186
+ fp16 ............................................ False
187
+ fp16_lm_cross_entropy ........................... False
188
+ fp32_residual_connection ........................ False
189
+ fp8 ............................................. e4m3
190
+ fp8_amax_compute_algo ........................... most_recent
191
+ fp8_amax_history_len ............................ 1
192
+ fp8_interval .................................... 1
193
+ fp8_margin ...................................... 0
194
+ fp8_param_gather ................................ False
195
+ fp8_recipe ...................................... mxfp8
196
+ fp8_wgrad ....................................... True
197
+ fsdp_double_buffer .............................. False
198
+ global_batch_size ............................... 8
199
+ grad_reduce_in_bf16 ............................. False
200
+ gradient_accumulation_fusion .................... True
201
+ gradient_reduce_div_fusion ...................... True
202
+ group_query_attention ........................... False
203
+ head_lr_mult .................................... 1.0
204
+ heterogeneous_layers_config_encoded_json ........ None
205
+ heterogeneous_layers_config_path ................ None
206
+ hidden_dropout .................................. 0.0
207
+ hidden_size ..................................... 2048
208
+ hierarchical_context_parallel_sizes ............. None
209
+ high_priority_stream_groups ..................... []
210
+ hybrid_attention_ratio .......................... 0.0
211
+ hybrid_mlp_ratio ................................ 0.0
212
+ hybrid_override_pattern ......................... None
213
+ hysteresis ...................................... 2
214
+ ict_head_size ................................... None
215
+ ict_load ........................................ None
216
+ img_h ........................................... 224
217
+ img_w ........................................... 224
218
+ indexer_batch_size .............................. 128
219
+ indexer_log_interval ............................ 1000
220
+ inference_batch_times_seqlen_threshold .......... -1
221
+ inference_dynamic_batching ...................... False
222
+ inference_dynamic_batching_buffer_guaranteed_fraction 0.2
223
+ inference_dynamic_batching_buffer_overflow_factor None
224
+ inference_dynamic_batching_buffer_size_gb ....... 40.0
225
+ inference_dynamic_batching_chunk_size ........... 256
226
+ inference_dynamic_batching_max_requests_override None
227
+ inference_dynamic_batching_max_tokens_override .. None
228
+ inference_dynamic_batching_num_cuda_graphs ...... 16
229
+ inference_max_batch_size ........................ 8
230
+ inference_max_seq_length ........................ 2560
231
+ inference_rng_tracker ........................... False
232
+ init_method_std ................................. 0.01
233
+ init_method_xavier_uniform ...................... False
234
+ init_model_with_meta_device ..................... False
235
+ initial_loss_scale .............................. 4294967296
236
+ inprocess_active_world_size ..................... 8
237
+ inprocess_barrier_timeout ....................... 120
238
+ inprocess_completion_timeout .................... 120
239
+ inprocess_empty_cuda_cache ...................... False
240
+ inprocess_granularity ........................... node
241
+ inprocess_hard_timeout .......................... 90
242
+ inprocess_heartbeat_interval .................... 30
243
+ inprocess_heartbeat_timeout ..................... 60
244
+ inprocess_last_call_wait ........................ 1
245
+ inprocess_max_iterations ........................ None
246
+ inprocess_monitor_process_interval .............. 1.0
247
+ inprocess_monitor_thread_interval ............... 1.0
248
+ inprocess_progress_watchdog_interval ............ 1.0
249
+ inprocess_restart ............................... False
250
+ inprocess_soft_timeout .......................... 60
251
+ inprocess_termination_grace_time ................ 1
252
+ is_hybrid_model ................................. False
253
+ iter_per_epoch .................................. 1250
254
+ iterations_to_skip .............................. []
255
+ keep_fp8_transpose_cache_when_using_custom_fsdp . False
256
+ kitchen_config_file ............................. None
257
+ kitchen_recipe_number ........................... None
258
+ kv_channels ..................................... 128
259
+ kv_lora_rank .................................... 512
260
+ lazy_mpu_init ................................... None
261
+ load ............................................ None
262
+ load_main_params_from_ckpt ...................... None
263
+ load_model_opt_format ........................... False
264
+ local_rank ...................................... 0
265
+ log_energy ...................................... False
266
+ log_interval .................................... 1
267
+ log_loss_scale_to_tensorboard ................... True
268
+ log_memory_to_tensorboard ....................... False
269
+ log_num_zeros_in_grad ........................... False
270
+ log_params_norm ................................. False
271
+ log_progress .................................... False
272
+ log_straggler ................................... False
273
+ log_throughput .................................. True
274
+ log_timers_to_tensorboard ....................... False
275
+ log_validation_ppl_to_tensorboard ............... False
276
+ log_world_size_to_tensorboard ................... False
277
+ logging_level ................................... None
278
+ loss_scale ...................................... None
279
+ loss_scale_window ............................... 1000
280
+ lr .............................................. 0.0001
281
+ lr_decay_iters .................................. 8
282
+ lr_decay_samples ................................ None
283
+ lr_decay_style .................................. cosine
284
+ lr_warmup_fraction .............................. None
285
+ lr_warmup_init .................................. 0.0
286
+ lr_warmup_iters ................................. 5
287
+ lr_warmup_samples ............................... 0
288
+ lr_wsd_decay_iters .............................. None
289
+ lr_wsd_decay_samples ............................ None
290
+ lr_wsd_decay_style .............................. exponential
291
+ main_grads_dtype ................................ torch.float32
292
+ main_params_dtype ............................... torch.float32
293
+ make_vocab_size_divisible_by .................... 128
294
+ mamba_head_dim .................................. 64
295
+ mamba_num_groups ................................ 8
296
+ mamba_num_heads ................................. None
297
+ mamba_state_dim ................................. 128
298
+ manual_gc ....................................... False
299
+ manual_gc_eval .................................. True
300
+ manual_gc_interval .............................. 0
301
+ mask_factor ..................................... 1.0
302
+ mask_prob ....................................... 0.15
303
+ mask_type ....................................... random
304
+ masked_softmax_fusion ........................... False
305
+ max_position_embeddings ......................... 32768
306
+ max_tokens_to_oom ............................... 12000
307
+ memory_snapshot_path ............................ snapshot.pickle
308
+ merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt
309
+ micro_batch_size ................................ 1
310
+ microbatch_group_size_per_vp_stage .............. None
311
+ mid_level_dataset_surplus ....................... 0.005
312
+ min_loss_scale .................................. 1.0
313
+ min_lr .......................................... 1e-05
314
+ mlp_chunks_for_prefill .......................... 1
315
+ mmap_bin_files .................................. True
316
+ mock_data ....................................... False
317
+ moe_apply_probs_on_input ........................ False
318
+ moe_aux_loss_coeff .............................. 0.001
319
+ moe_deepep_num_sms .............................. 20
320
+ moe_echo_enable_random_offloading ............... False
321
+ moe_echo_expert_dispatch_overlap ................ False
322
+ moe_echo_expert_dispatcher_type ................. hybridep
323
+ moe_echo_recompute_expert_dispatch .............. False
324
+ moe_enable_deepep ............................... False
325
+ moe_enable_echo ................................. False
326
+ moe_enable_hybridep ............................. False
327
+ moe_enable_moonep ............................... True
328
+ moe_expert_capacity_factor ...................... None
329
+ moe_extended_tp ................................. False
330
+ moe_ffn_hidden_size ............................. 1408
331
+ moe_grouped_gemm ................................ True
332
+ moe_hybridep_num_sms ............................ 16
333
+ moe_input_jitter_eps ............................ None
334
+ moe_layer_freq .................................. [0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
335
+ moe_layer_recompute ............................. False
336
+ moe_moonep_num_sms .............................. None
337
+ moe_num_echo_experts ............................ None
338
+ moe_pad_expert_input_to_capacity ................ False
339
+ moe_per_layer_logging ........................... False
340
+ moe_permute_fusion .............................. False
341
+ moe_received_token_capacity ..................... None
342
+ moe_router_bias_update_rate ..................... 0.001
343
+ moe_router_dtype ................................ fp32
344
+ moe_router_enable_expert_bias ................... False
345
+ moe_router_force_load_balancing ................. False
346
+ moe_router_group_topk ........................... None
347
+ moe_router_load_balancing_type .................. seq_aux_loss
348
+ moe_router_num_groups ........................... None
349
+ moe_router_padding_for_fp8 ...................... False
350
+ moe_router_pre_softmax .......................... True
351
+ moe_router_score_function ....................... softmax
352
+ moe_router_topk ................................. 6
353
+ moe_router_topk_scaling_factor .................. None
354
+ moe_shared_expert_intermediate_size ............. 2816
355
+ moe_shared_expert_overlap ....................... False
356
+ moe_token_dispatcher_type ....................... flex
357
+ moe_token_drop_policy ........................... probs
358
+ moe_upcycling_granularity ....................... 1
359
+ moe_use_device_initiated_grouped_gemm ........... True
360
+ moe_use_legacy_grouped_gemm ..................... False
361
+ moe_use_upcycling ............................... False
362
+ moe_z_loss_coeff ................................ None
363
+ mrope_section ................................... None
364
+ mscale .......................................... 0.707
365
+ mscale_all_dim .................................. 0.707
366
+ mtp_loss_scaling_factor ......................... 0.1
367
+ mtp_num_layers .................................. None
368
+ multi_latent_attention .......................... True
369
+ nccl_all_reduce_for_prefill ..................... False
370
+ nccl_communicator_config_path ................... None
371
+ nccl_ub ......................................... False
372
+ no_load_optim ................................... None
373
+ no_load_rng ..................................... None
374
+ no_persist_layer_norm ........................... False
375
+ no_rope_freq .................................... None
376
+ no_save_optim ................................... None
377
+ no_save_rng ..................................... None
378
+ non_persistent_ckpt_type ........................ None
379
+ non_persistent_global_ckpt_dir .................. None
380
+ non_persistent_local_ckpt_algo .................. fully_parallel
381
+ non_persistent_local_ckpt_dir ................... None
382
+ non_persistent_save_interval .................... None
383
+ norm_epsilon .................................... 1e-06
384
+ normalization ................................... RMSNorm
385
+ num_attention_heads ............................. 16
386
+ num_channels .................................... 3
387
+ num_classes ..................................... 1000
388
+ num_dataset_builder_threads ..................... 1
389
+ num_distributed_optimizer_instances ............. 1
390
+ num_experts ..................................... 64
391
+ num_layers ...................................... 24
392
+ num_layers_at_end_in_bf16 ....................... 1
393
+ num_layers_at_start_in_bf16 ..................... 1
394
+ num_layers_per_virtual_pipeline_stage ........... None
395
+ num_query_groups ................................ 1
396
+ num_virtual_stages_per_pipeline_rank ............ None
397
+ num_workers ..................................... 2
398
+ object_storage_cache_path ....................... None
399
+ one_logger_async ................................ False
400
+ one_logger_project .............................. megatron-lm
401
+ one_logger_run_name ............................. None
402
+ onnx_safe ....................................... None
403
+ openai_gelu ..................................... False
404
+ optimizer ....................................... adam
405
+ optimizer_cpu_offload ........................... False
406
+ optimizer_offload_fraction ...................... 1.0
407
+ output_bert_embeddings .......................... False
408
+ overlap_cpu_optimizer_d2h_h2d ................... False
409
+ overlap_grad_reduce ............................. True
410
+ overlap_moe_expert_parallel_comm ................ False
411
+ overlap_p2p_comm ................................ False
412
+ overlap_p2p_comm_warmup_flush ................... False
413
+ overlap_param_gather ............................ True
414
+ overlap_param_gather_with_optimizer_step ........ False
415
+ override_opt_param_scheduler .................... False
416
+ params_dtype .................................... torch.bfloat16
417
+ patch_dim ....................................... 16
418
+ per_split_data_args_path ........................ None
419
+ perform_initialization .......................... True
420
+ pin_cpu_grads ................................... True
421
+ pin_cpu_params .................................. True
422
+ pipeline_model_parallel_comm_backend ............ None
423
+ pipeline_model_parallel_layout .................. None
424
+ pipeline_model_parallel_size .................... 1
425
+ position_embedding_type ......................... rope
426
+ pretrained_checkpoint ........................... None
427
+ profile ......................................... False
428
+ profile_ranks ................................... [0]
429
+ profile_step_end ................................ 12
430
+ profile_step_start .............................. 10
431
+ q_lora_rank ..................................... None
432
+ qk_head_dim ..................................... 128
433
+ qk_l2_norm ...................................... False
434
+ qk_layernorm .................................... True
435
+ qk_pos_emb_head_dim ............................. 64
436
+ query_in_block_prob ............................. 0.1
437
+ rampup_batch_size ............................... None
438
+ rank ............................................ 0
439
+ recompute_granularity ........................... None
440
+ recompute_method ................................ None
441
+ recompute_modules ............................... None
442
+ recompute_num_layers ............................ None
443
+ record_memory_history ........................... False
444
+ relative_attention_max_distance ................. 128
445
+ relative_attention_num_buckets .................. 32
446
+ replication ..................................... False
447
+ replication_factor .............................. 2
448
+ replication_jump ................................ None
449
+ rerun_mode ...................................... validate_results
450
+ reset_attention_mask ............................ False
451
+ reset_position_ids .............................. False
452
+ result_rejected_tracker_filename ................ None
453
+ retriever_report_topk_accuracies ................ []
454
+ retriever_score_scaling ......................... False
455
+ retriever_seq_length ............................ 256
456
+ retro_add_retriever ............................. False
457
+ retro_attention_gate ............................ 1
458
+ retro_cyclic_train_iters ........................ None
459
+ retro_encoder_attention_dropout ................. 0.1
460
+ retro_encoder_hidden_dropout .................... 0.1
461
+ retro_encoder_layers ............................ 2
462
+ retro_num_neighbors ............................. 2
463
+ retro_num_retrieved_chunks ...................... 2
464
+ retro_project_dir ............................... None
465
+ retro_verify_neighbor_count ..................... True
466
+ reuse_grad_buf_for_mxfp8_param_ag ............... False
467
+ rope_scaling_factor ............................. 8.0
468
+ rope_type ....................................... None
469
+ rotary_base ..................................... 10000
470
+ rotary_interleaved .............................. False
471
+ rotary_percent .................................. 1.0
472
+ rotary_scaling_factor ........................... 40.0
473
+ rotary_seq_len_interpolation_factor ............. None
474
+ run_workload_inspector_server ................... False
475
+ sample_rate ..................................... 1.0
476
+ save ............................................ None
477
+ save_interval ................................... None
478
+ save_retain_interval ............................ None
479
+ scatter_gather_tensors_in_pipeline .............. True
480
+ seed ............................................ 1234
481
+ seq_length ...................................... 8192
482
+ sequence_parallel ............................... False
483
+ sft ............................................. False
484
+ sft_tokenizer_prompt_format ..................... nemotron-h-aligned
485
+ sgd_momentum .................................... 0.9
486
+ short_seq_prob .................................. 0.1
487
+ skip_train ...................................... False
488
+ skipped_train_samples ........................... 0
489
+ spec ............................................ None
490
+ split ........................................... 100,0,0
491
+ squared_relu .................................... False
492
+ start_weight_decay .............................. 0.1
493
+ straggler_ctrlr_port ............................ 65535
494
+ straggler_minmax_count .......................... 1
495
+ suggested_communication_unit_size ............... None
496
+ swiglu .......................................... True
497
+ swin_backbone_type .............................. tiny
498
+ symmetric_ar_type ............................... None
499
+ te_rng_tracker .................................. True
500
+ tensor_model_parallel_size ...................... 1
501
+ tensorboard_dir ................................. None
502
+ tensorboard_log_interval ........................ 1
503
+ tensorboard_queue_size .......................... 1000
504
+ test_data_path .................................. None
505
+ test_mode ....................................... False
506
+ tiktoken_num_special_tokens ..................... 1000
507
+ tiktoken_pattern ................................ None
508
+ tiktoken_special_tokens ......................... None
509
+ timing_log_level ................................ 0
510
+ timing_log_option ............................... minmax
511
+ titles_data_path ................................ None
512
+ tokenizer_model ................................. None
513
+ tokenizer_type .................................. GPT2BPETokenizer
514
+ torch_fsdp2_reshard_after_forward ............... True
515
+ tp_comm_bootstrap_backend ....................... nccl
516
+ tp_comm_bulk_dgrad .............................. True
517
+ tp_comm_bulk_wgrad .............................. True
518
+ tp_comm_overlap ................................. False
519
+ tp_comm_overlap_ag .............................. True
520
+ tp_comm_overlap_cfg ............................. None
521
+ tp_comm_overlap_rs .............................. True
522
+ tp_comm_overlap_rs_dgrad ........................ False
523
+ tp_comm_split_ag ................................ True
524
+ tp_comm_split_rs ................................ True
525
+ train_data_path ................................. None
526
+ train_iters ..................................... 12
527
+ train_samples ................................... None
528
+ train_sync_interval ............................. None
529
+ transformer_impl ................................ transformer_engine
530
+ transformer_pipeline_model_parallel_size ........ 1
531
+ untie_embeddings_and_output_weights ............. True
532
+ use_checkpoint_args ............................. False
533
+ use_checkpoint_opt_param_scheduler .............. False
534
+ use_cpu_initialization .......................... None
535
+ use_custom_fsdp ................................. False
536
+ use_dist_ckpt ................................... True
537
+ use_dist_ckpt_deprecated ........................ False
538
+ use_distributed_optimizer ....................... True
539
+ use_flash_attn .................................. False
540
+ use_legacy_models ............................... False
541
+ use_mp_args_from_checkpoint_args ................ False
542
+ use_one_sent_docs ............................... False
543
+ use_persistent_ckpt_worker ...................... False
544
+ use_precision_aware_optimizer ................... False
545
+ use_pytorch_profiler ............................ False
546
+ use_ring_exchange_p2p ........................... False
547
+ use_rope_scaling ................................ False
548
+ use_rotary_position_embeddings .................. False
549
+ use_sharp ....................................... False
550
+ use_tokenizer_model_from_checkpoint_args ........ True
551
+ use_torch_fsdp2 ................................. False
552
+ use_torch_optimizer_for_cpu_offload ............. False
553
+ use_tp_pp_dp_mapping ............................ False
554
+ v_head_dim ...................................... 128
555
+ valid_data_path ................................. None
556
+ variable_seq_lengths ............................ False
557
+ virtual_pipeline_model_parallel_size ............ None
558
+ vision_backbone_type ............................ vit
559
+ vision_pretraining .............................. False
560
+ vision_pretraining_type ......................... classify
561
+ vocab_extra_ids ................................. 0
562
+ vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json
563
+ vocab_size ...................................... None
564
+ wandb_exp_name ..................................
565
+ wandb_project ...................................
566
+ wandb_save_dir ..................................
567
+ weight_decay .................................... 0.1
568
+ weight_decay_incr_style ......................... constant
569
+ wgrad_deferral_limit ............................ 0
570
+ world_size ...................................... 8
571
+ yaml_cfg ........................................ None
572
+ -------------------- end of arguments ---------------------
573
+ INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1
574
+ > building GPT2BPETokenizer tokenizer ...
575
+ > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304)
576
+ WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS
577
+ > initializing torch distributed ...
578
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
579
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
580
+ [Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
581
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
582
+ [Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
583
+ [Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
584
+ [Gloo] Rank 7 is connected to 7 peer ranks. [Gloo] Rank Expected number of connected peer ranks is : 76
585
+ is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
586
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
587
+ [Gloo] Rank [Gloo] Rank 7 is connected to 17 is connected to peer ranks. 7Expected number of connected peer ranks is : peer ranks. 7Expected number of connected peer ranks is :
588
+ 7
589
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
590
+ [Gloo] Rank 4 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
592
+ [Gloo] Rank [Gloo] Rank 5 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 :
593
+ 7
594
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
595
+ > initialized tensor model parallel with size 1
596
+ > initialized pipeline model parallel with size 1
597
+ > setting random seeds to 1234 ...
598
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
599
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
600
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
601
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
602
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
603
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
604
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
605
+ [rank2]:[W826 18:44:57.809376370 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.
606
+ [rank5]:[W826 18:44:57.809579890 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.
607
+ [rank7]:[W826 18:44:57.810964509 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
608
+ [rank6]:[W826 18:44:57.811816213 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
+ [rank4]:[W826 18:44:57.811849938 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.
610
+ [rank3]:[W826 18:44:57.811927304 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.
611
+ [rank1]:[W826 18:44:57.819100826 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.
612
+ > compiling dataset index builder ...
613
+ make: Entering directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
614
+ make: Nothing to be done for 'default'.
615
+ make: Leaving directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
616
+ >>> done with dataset index builder. Compilation time: 0.035 seconds
617
+ WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations.
618
+ > compiling and loading fused kernels ...
619
+ [rank0]:[W826 18:44:57.975079572 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.
620
+ >>> done with compiling and loading fused kernels. Compilation time: 3.591 seconds
621
+ time to initialize megatron (seconds): 7.084
622
+ [after megatron is initialized] datetime: 2026-08-26 18:45:03
623
+ building GPT model ...
624
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
625
+ warnings.warn(
626
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
627
+ warnings.warn(
628
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
629
+ warnings.warn(
630
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
631
+ warnings.warn(
632
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
633
+ warnings.warn(
634
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
635
+ warnings.warn(
636
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
637
+ warnings.warn(
638
+ /workspace/Megatron-MoE-Echo/megatron/core/transformer/transformer_config.py:1531: UserWarning: MLA config max_position_embeddings is overridden by customer input, please use the original_max_position_embeddings instead for YaRN!max_position_embeddings will be deprecated in the future.Assigned original_max_position_embeddings to max_position_embeddings if not set,and assigned max_position_embeddings back to the original value.
639
+ warnings.warn(
640
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
641
+ warnings.warn(
642
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
643
+ warnings.warn(
644
+ /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.
645
+ warnings.warn(
646
+ /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.
647
+ warnings.warn(
648
+ /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.
649
+ warnings.warn(
650
+ /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.
651
+ warnings.warn(
652
+ /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.
653
+ warnings.warn(
654
+ /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.
655
+ warnings.warn(
656
+ > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 2596386816
657
+ 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)
658
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 20
659
+ Params for bucket 1 (103022592 elements, 103022592 padded size):
660
+ module.output_layer.weight
661
+ Params for bucket 2 (48503296 elements, 48503296 padded size):
662
+ module.decoder.layers.23.pre_mlp_layernorm.weight
663
+ module.decoder.layers.23.mlp.shared_experts.linear_fc1.weight
664
+ module.decoder.layers.23.self_attention.linear_kv_up_proj.layer_norm_weight
665
+ module.decoder.layers.23.self_attention.linear_kv_down_proj.weight
666
+ module.decoder.layers.23.self_attention.linear_q_proj.weight
667
+ module.decoder.layers.23.input_layernorm.weight
668
+ module.decoder.layers.22.mlp.shared_experts.linear_fc2.weight
669
+ module.decoder.layers.22.mlp.shared_experts.linear_fc1.weight
670
+ module.decoder.final_layernorm.weight
671
+ module.decoder.layers.23.mlp.shared_experts.linear_fc2.weight
672
+ module.decoder.layers.23.mlp.router.weight
673
+ module.decoder.layers.23.self_attention.linear_kv_up_proj.weight
674
+ module.decoder.layers.23.self_attention.linear_proj.weight
675
+ Params for bucket 3 (40901632 elements, 40901632 padded size):
676
+ module.decoder.layers.22.mlp.router.weight
677
+ module.decoder.layers.22.self_attention.linear_kv_up_proj.weight
678
+ module.decoder.layers.22.self_attention.linear_q_proj.weight
679
+ module.decoder.layers.22.self_attention.linear_proj.weight
680
+ module.decoder.layers.22.input_layernorm.weight
681
+ module.decoder.layers.21.mlp.shared_experts.linear_fc2.weight
682
+ module.decoder.layers.21.mlp.shared_experts.linear_fc1.weight
683
+ module.decoder.layers.21.pre_mlp_layernorm.weight
684
+ module.decoder.layers.21.self_attention.linear_kv_up_proj.layer_norm_weight
685
+ module.decoder.layers.21.self_attention.linear_kv_down_proj.weight
686
+ module.decoder.layers.21.self_attention.linear_q_proj.weight
687
+ module.decoder.layers.22.self_attention.linear_kv_down_proj.weight
688
+ module.decoder.layers.22.self_attention.linear_kv_up_proj.layer_norm_weight
689
+ module.decoder.layers.22.pre_mlp_layernorm.weight
690
+ module.decoder.layers.21.mlp.router.weight
691
+ module.decoder.layers.21.self_attention.linear_kv_up_proj.weight
692
+ Params for bucket 4 (41163264 elements, 41163264 padded size):
693
+ module.decoder.layers.21.self_attention.linear_proj.weight
694
+ module.decoder.layers.21.input_layernorm.weight
695
+ module.decoder.layers.20.mlp.shared_experts.linear_fc1.weight
696
+ module.decoder.layers.20.mlp.shared_experts.linear_fc2.weight
697
+ module.decoder.layers.20.pre_mlp_layernorm.weight
698
+ module.decoder.layers.20.self_attention.linear_kv_up_proj.layer_norm_weight
699
+ module.decoder.layers.20.self_attention.linear_kv_down_proj.weight
700
+ module.decoder.layers.20.self_attention.linear_q_proj.weight
701
+ module.decoder.layers.20.input_layernorm.weight
702
+ module.decoder.layers.19.mlp.shared_experts.linear_fc2.weight
703
+ module.decoder.layers.20.mlp.router.weight
704
+ module.decoder.layers.20.self_attention.linear_kv_up_proj.weight
705
+ module.decoder.layers.20.self_attention.linear_proj.weight
706
+ Params for bucket 5 (42734080 elements, 42734080 padded size):
707
+ module.decoder.layers.19.mlp.shared_experts.linear_fc1.weight
708
+ module.decoder.layers.19.mlp.router.weight
709
+ module.decoder.layers.19.self_attention.linear_kv_up_proj.weight
710
+ module.decoder.layers.19.self_attention.linear_kv_down_proj.weight
711
+ module.decoder.layers.19.self_attention.linear_kv_up_proj.layer_norm_weight
712
+ module.decoder.layers.19.pre_mlp_layernorm.weight
713
+ module.decoder.layers.19.self_attention.linear_q_proj.weight
714
+ module.decoder.layers.19.self_attention.linear_proj.weight
715
+ module.decoder.layers.19.input_layernorm.weight
716
+ module.decoder.layers.18.mlp.shared_experts.linear_fc2.weight
717
+ module.decoder.layers.18.mlp.shared_experts.linear_fc1.weight
718
+ Params for bucket 6 (40901632 elements, 40901632 padded size):
719
+ module.decoder.layers.18.mlp.router.weight
720
+ module.decoder.layers.18.self_attention.linear_kv_up_proj.weight
721
+ module.decoder.layers.17.mlp.router.weight
722
+ module.decoder.layers.17.self_attention.linear_kv_up_proj.weight
723
+ module.decoder.layers.17.self_attention.linear_q_proj.weight
724
+ module.decoder.layers.18.pre_mlp_layernorm.weight
725
+ module.decoder.layers.18.self_attention.linear_kv_up_proj.layer_norm_weight
726
+ module.decoder.layers.18.self_attention.linear_kv_down_proj.weight
727
+ module.decoder.layers.18.self_attention.linear_q_proj.weight
728
+ module.decoder.layers.18.self_attention.linear_proj.weight
729
+ module.decoder.layers.18.input_layernorm.weight
730
+ module.decoder.layers.17.mlp.shared_experts.linear_fc2.weight
731
+ module.decoder.layers.17.mlp.shared_experts.linear_fc1.weight
732
+ module.decoder.layers.17.pre_mlp_layernorm.weight
733
+ module.decoder.layers.17.self_attention.linear_kv_up_proj.layer_norm_weight
734
+ module.decoder.layers.17.self_attention.linear_kv_down_proj.weight
735
+ Params for bucket 7 (41163264 elements, 41163264 padded size):
736
+ module.decoder.layers.17.self_attention.linear_proj.weight
737
+ module.decoder.layers.17.input_layernorm.weight
738
+ module.decoder.layers.16.mlp.shared_experts.linear_fc1.weight
739
+ module.decoder.layers.16.mlp.shared_experts.linear_fc2.weight
740
+ module.decoder.layers.16.mlp.router.weight
741
+ module.decoder.layers.16.self_attention.linear_kv_up_proj.weight
742
+ module.decoder.layers.16.pre_mlp_layernorm.weight
743
+ module.decoder.layers.16.self_attention.linear_kv_up_proj.layer_norm_weight
744
+ module.decoder.layers.16.self_attention.linear_kv_down_proj.weight
745
+ module.decoder.layers.16.self_attention.linear_q_proj.weight
746
+ module.decoder.layers.16.self_attention.linear_proj.weight
747
+ module.decoder.layers.16.input_layernorm.weight
748
+ module.decoder.layers.15.mlp.shared_experts.linear_fc2.weight
749
+ Params for bucket 8 (42734080 elements, 42734080 padded size):
750
+ module.decoder.layers.15.self_attention.linear_kv_up_proj.weight
751
+ module.decoder.layers.15.pre_mlp_layernorm.weight
752
+ module.decoder.layers.15.self_attention.linear_q_proj.weight
753
+ module.decoder.layers.15.self_attention.linear_proj.weight
754
+ module.decoder.layers.15.input_layernorm.weight
755
+ module.decoder.layers.14.mlp.shared_experts.linear_fc2.weight
756
+ module.decoder.layers.14.mlp.shared_experts.linear_fc1.weight
757
+ module.decoder.layers.15.mlp.shared_experts.linear_fc1.weight
758
+ module.decoder.layers.15.self_attention.linear_kv_up_proj.layer_norm_weight
759
+ module.decoder.layers.15.mlp.router.weight
760
+ module.decoder.layers.15.self_attention.linear_kv_down_proj.weight
761
+ Params for bucket 9 (40901632 elements, 40901632 padded size):
762
+ module.decoder.layers.14.mlp.router.weight
763
+ module.decoder.layers.14.self_attention.linear_kv_up_proj.weight
764
+ module.decoder.layers.13.pre_mlp_layernorm.weight
765
+ module.decoder.layers.13.self_attention.linear_kv_up_proj.layer_norm_weight
766
+ module.decoder.layers.13.self_attention.linear_kv_down_proj.weight
767
+ module.decoder.layers.13.self_attention.linear_q_proj.weight
768
+ module.decoder.layers.14.pre_mlp_layernorm.weight
769
+ module.decoder.layers.14.self_attention.linear_kv_up_proj.layer_norm_weight
770
+ module.decoder.layers.14.self_attention.linear_kv_down_proj.weight
771
+ module.decoder.layers.14.self_attention.linear_q_proj.weight
772
+ module.decoder.layers.14.self_attention.linear_proj.weight
773
+ module.decoder.layers.14.input_layernorm.weight
774
+ module.decoder.layers.13.mlp.shared_experts.linear_fc2.weight
775
+ module.decoder.layers.13.mlp.shared_experts.linear_fc1.weight
776
+ module.decoder.layers.13.mlp.router.weight
777
+ module.decoder.layers.13.self_attention.linear_kv_up_proj.weight
778
+ Params for bucket 10 (41163264 elements, 41163264 padded size):
779
+ module.decoder.layers.12.pre_mlp_layernorm.weight
780
+ module.decoder.layers.12.self_attention.linear_kv_up_proj.layer_norm_weight
781
+ module.decoder.layers.12.self_attention.linear_kv_down_proj.weight
782
+ module.decoder.layers.12.self_attention.linear_q_proj.weight
783
+ module.decoder.layers.12.input_layernorm.weight
784
+ module.decoder.layers.11.mlp.shared_experts.linear_fc2.weight
785
+ module.decoder.layers.13.self_attention.linear_proj.weight
786
+ module.decoder.layers.12.mlp.shared_experts.linear_fc2.weight
787
+ module.decoder.layers.12.mlp.router.weight
788
+ module.decoder.layers.13.input_layernorm.weight
789
+ module.decoder.layers.12.mlp.shared_experts.linear_fc1.weight
790
+ module.decoder.layers.12.self_attention.linear_kv_up_proj.weight
791
+ module.decoder.layers.12.self_attention.linear_proj.weight
792
+ Params for bucket 11 (42734080 elements, 42734080 padded size):
793
+ module.decoder.layers.11.mlp.router.weight
794
+ module.decoder.layers.11.self_attention.linear_kv_up_proj.weight
795
+ module.decoder.layers.11.self_attention.linear_q_proj.weight
796
+ module.decoder.layers.11.mlp.shared_experts.linear_fc1.weight
797
+ module.decoder.layers.11.pre_mlp_layernorm.weight
798
+ module.decoder.layers.11.self_attention.linear_kv_up_proj.layer_norm_weight
799
+ module.decoder.layers.11.self_attention.linear_kv_down_proj.weight
800
+ module.decoder.layers.11.self_attention.linear_proj.weight
801
+ module.decoder.layers.11.input_layernorm.weight
802
+ module.decoder.layers.10.mlp.shared_experts.linear_fc2.weight
803
+ module.decoder.layers.10.mlp.shared_experts.linear_fc1.weight
804
+ Params for bucket 12 (40901632 elements, 40901632 padded size):
805
+ module.decoder.layers.10.mlp.router.weight
806
+ module.decoder.layers.10.self_attention.linear_kv_up_proj.weight
807
+ module.decoder.layers.9.pre_mlp_layernorm.weight
808
+ module.decoder.layers.9.self_attention.linear_kv_up_proj.layer_norm_weight
809
+ module.decoder.layers.9.self_attention.linear_kv_down_proj.weight
810
+ module.decoder.layers.9.self_attention.linear_q_proj.weight
811
+ module.decoder.layers.10.pre_mlp_layernorm.weight
812
+ module.decoder.layers.10.self_attention.linear_kv_down_proj.weight
813
+ module.decoder.layers.10.self_attention.linear_kv_up_proj.layer_norm_weight
814
+ module.decoder.layers.10.self_attention.linear_q_proj.weight
815
+ module.decoder.layers.10.self_attention.linear_proj.weight
816
+ module.decoder.layers.10.input_layernorm.weight
817
+ module.decoder.layers.9.mlp.shared_experts.linear_fc2.weight
818
+ module.decoder.layers.9.mlp.shared_experts.linear_fc1.weight
819
+ module.decoder.layers.9.mlp.router.weight
820
+ module.decoder.layers.9.self_attention.linear_kv_up_proj.weight
821
+ Params for bucket 13 (41163264 elements, 41163264 padded size):
822
+ module.decoder.layers.9.self_attention.linear_proj.weight
823
+ module.decoder.layers.9.input_layernorm.weight
824
+ module.decoder.layers.8.mlp.router.weight
825
+ module.decoder.layers.8.mlp.shared_experts.linear_fc2.weight
826
+ module.decoder.layers.8.mlp.shared_experts.linear_fc1.weight
827
+ module.decoder.layers.8.self_attention.linear_kv_up_proj.weight
828
+ module.decoder.layers.8.self_attention.linear_proj.weight
829
+ module.decoder.layers.8.input_layernorm.weight
830
+ module.decoder.layers.8.pre_mlp_layernorm.weight
831
+ module.decoder.layers.8.self_attention.linear_kv_up_proj.layer_norm_weight
832
+ module.decoder.layers.8.self_attention.linear_kv_down_proj.weight
833
+ module.decoder.layers.8.self_attention.linear_q_proj.weight
834
+ module.decoder.layers.7.mlp.shared_experts.linear_fc2.weight
835
+ Params for bucket 14 (42734080 elements, 42734080 padded size):
836
+ module.decoder.layers.7.mlp.shared_experts.linear_fc1.weight
837
+ module.decoder.layers.7.self_attention.linear_kv_up_proj.layer_norm_weight
838
+ module.decoder.layers.7.pre_mlp_layernorm.weight
839
+ module.decoder.layers.7.self_attention.linear_kv_down_proj.weight
840
+ module.decoder.layers.7.input_layernorm.weight
841
+ module.decoder.layers.6.mlp.shared_experts.linear_fc2.weight
842
+ module.decoder.layers.6.mlp.shared_experts.linear_fc1.weight
843
+ module.decoder.layers.7.mlp.router.weight
844
+ module.decoder.layers.7.self_attention.linear_kv_up_proj.weight
845
+ module.decoder.layers.7.self_attention.linear_q_proj.weight
846
+ module.decoder.layers.7.self_attention.linear_proj.weight
847
+ Params for bucket 15 (40901632 elements, 40901632 padded size):
848
+ module.decoder.layers.6.pre_mlp_layernorm.weight
849
+ module.decoder.layers.6.self_attention.linear_kv_down_proj.weight
850
+ module.decoder.layers.6.self_attention.linear_kv_up_proj.layer_norm_weight
851
+ module.decoder.layers.5.pre_mlp_layernorm.weight
852
+ module.decoder.layers.5.self_attention.linear_kv_up_proj.layer_norm_weight
853
+ module.decoder.layers.5.self_attention.linear_kv_down_proj.weight
854
+ module.decoder.layers.6.mlp.router.weight
855
+ module.decoder.layers.6.self_attention.linear_kv_up_proj.weight
856
+ module.decoder.layers.6.self_attention.linear_q_proj.weight
857
+ module.decoder.layers.6.self_attention.linear_proj.weight
858
+ module.decoder.layers.6.input_layernorm.weight
859
+ module.decoder.layers.5.mlp.shared_experts.linear_fc2.weight
860
+ module.decoder.layers.5.mlp.shared_experts.linear_fc1.weight
861
+ module.decoder.layers.5.mlp.router.weight
862
+ module.decoder.layers.5.self_attention.linear_kv_up_proj.weight
863
+ module.decoder.layers.5.self_attention.linear_q_proj.weight
864
+ Params for bucket 16 (41163264 elements, 41163264 padded size):
865
+ module.decoder.layers.4.pre_mlp_layernorm.weight
866
+ module.decoder.layers.4.self_attention.linear_kv_up_proj.layer_norm_weight
867
+ module.decoder.layers.4.self_attention.linear_kv_down_proj.weight
868
+ module.decoder.layers.4.self_attention.linear_q_proj.weight
869
+ module.decoder.layers.4.self_attention.linear_proj.weight
870
+ module.decoder.layers.4.input_layernorm.weight
871
+ module.decoder.layers.3.mlp.shared_experts.linear_fc2.weight
872
+ module.decoder.layers.5.self_attention.linear_proj.weight
873
+ module.decoder.layers.4.mlp.router.weight
874
+ module.decoder.layers.4.mlp.shared_experts.linear_fc2.weight
875
+ module.decoder.layers.5.input_layernorm.weight
876
+ module.decoder.layers.4.mlp.shared_experts.linear_fc1.weight
877
+ module.decoder.layers.4.self_attention.linear_kv_up_proj.weight
878
+ Params for bucket 17 (42734080 elements, 42734080 padded size):
879
+ module.decoder.layers.3.mlp.shared_experts.linear_fc1.weight
880
+ module.decoder.layers.3.mlp.router.weight
881
+ module.decoder.layers.3.self_attention.linear_kv_up_proj.weight
882
+ module.decoder.layers.3.self_attention.linear_q_proj.weight
883
+ module.decoder.layers.3.self_attention.linear_proj.weight
884
+ module.decoder.layers.3.input_layernorm.weight
885
+ module.decoder.layers.2.mlp.shared_experts.linear_fc2.weight
886
+ module.decoder.layers.2.mlp.shared_experts.linear_fc1.weight
887
+ module.decoder.layers.3.self_attention.linear_kv_up_proj.layer_norm_weight
888
+ module.decoder.layers.3.pre_mlp_layernorm.weight
889
+ module.decoder.layers.3.self_attention.linear_kv_down_proj.weight
890
+ Params for bucket 18 (40901632 elements, 40901632 padded size):
891
+ module.decoder.layers.2.mlp.router.weight
892
+ module.decoder.layers.2.self_attention.linear_kv_down_proj.weight
893
+ module.decoder.layers.2.self_attention.linear_kv_up_proj.weight
894
+ module.decoder.layers.2.self_attention.linear_q_proj.weight
895
+ module.decoder.layers.2.self_attention.linear_proj.weight
896
+ module.decoder.layers.2.input_layernorm.weight
897
+ module.decoder.layers.1.mlp.shared_experts.linear_fc2.weight
898
+ module.decoder.layers.1.mlp.router.weight
899
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.weight
900
+ module.decoder.layers.2.self_attention.linear_kv_up_proj.layer_norm_weight
901
+ module.decoder.layers.2.pre_mlp_layernorm.weight
902
+ module.decoder.layers.1.mlp.shared_experts.linear_fc1.weight
903
+ module.decoder.layers.1.pre_mlp_layernorm.weight
904
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.layer_norm_weight
905
+ module.decoder.layers.1.self_attention.linear_kv_down_proj.weight
906
+ module.decoder.layers.1.self_attention.linear_q_proj.weight
907
+ Params for bucket 19 (71436288 elements, 71436288 padded size):
908
+ module.decoder.layers.1.self_attention.linear_proj.weight
909
+ module.decoder.layers.1.input_layernorm.weight
910
+ module.decoder.layers.0.mlp.linear_fc1.weight
911
+ module.decoder.layers.0.mlp.linear_fc2.weight
912
+ Params for bucket 20 (116789760 elements, 116789760 padded size):
913
+ module.decoder.layers.0.mlp.linear_fc1.layer_norm_weight
914
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.weight
915
+ module.decoder.layers.0.self_attention.linear_kv_down_proj.weight
916
+ module.decoder.layers.0.self_attention.linear_proj.weight
917
+ module.decoder.layers.0.input_layernorm.weight
918
+ module.embedding.word_embeddings.weight
919
+ module.decoder.layers.0.self_attention.linear_q_proj.weight
920
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.layer_norm_weight
921
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 40
922
+ Params for bucket 1 (40370176 elements, 40370176 padded size):
923
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight7
924
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight5
925
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight3
926
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight1
927
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight6
928
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight6
929
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight4
930
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight2
931
+ module.decoder.layers.23.mlp.experts.linear_fc2.weight0
932
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight7
933
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight5
934
+ Params for bucket 2 (40370176 elements, 40370176 padded size):
935
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight4
936
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight2
937
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight0
938
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight7
939
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight5
940
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight3
941
+ module.decoder.layers.23.mlp.experts.linear_fc1.weight1
942
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight6
943
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight4
944
+ Params for bucket 3 (40370176 elements, 40370176 padded size):
945
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight3
946
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight1
947
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight6
948
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight4
949
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight0
950
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight7
951
+ module.decoder.layers.22.mlp.experts.linear_fc2.weight2
952
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight5
953
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight3
954
+ Params for bucket 4 (40370176 elements, 40370176 padded size):
955
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight2
956
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight0
957
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight7
958
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight5
959
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight3
960
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight1
961
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight6
962
+ module.decoder.layers.22.mlp.experts.linear_fc1.weight1
963
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight4
964
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight2
965
+ module.decoder.layers.21.mlp.experts.linear_fc2.weight0
966
+ Params for bucket 5 (40370176 elements, 40370176 padded size):
967
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight6
968
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight4
969
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight2
970
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight7
971
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight3
972
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight5
973
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight1
974
+ Params for bucket 6 (40370176 elements, 40370176 padded size):
975
+ module.decoder.layers.21.mlp.experts.linear_fc1.weight0
976
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight5
977
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight7
978
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight3
979
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight1
980
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight6
981
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight4
982
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight6
983
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight2
984
+ module.decoder.layers.20.mlp.experts.linear_fc2.weight0
985
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight7
986
+ Params for bucket 7 (40370176 elements, 40370176 padded size):
987
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight2
988
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight4
989
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight0
990
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight7
991
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight5
992
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight1
993
+ module.decoder.layers.20.mlp.experts.linear_fc1.weight3
994
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight6
995
+ Params for bucket 8 (40370176 elements, 40370176 padded size):
996
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight5
997
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight1
998
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight3
999
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight6
1000
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight4
1001
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight2
1002
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight4
1003
+ module.decoder.layers.19.mlp.experts.linear_fc2.weight0
1004
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight7
1005
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight5
1006
+ Params for bucket 9 (40370176 elements, 40370176 padded size):
1007
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight2
1008
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight0
1009
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight7
1010
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight5
1011
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight3
1012
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight3
1013
+ module.decoder.layers.19.mlp.experts.linear_fc1.weight1
1014
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight6
1015
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight4
1016
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight2
1017
+ Params for bucket 10 (40370176 elements, 40370176 padded size):
1018
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight1
1019
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight6
1020
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight4
1021
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight2
1022
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight7
1023
+ module.decoder.layers.18.mlp.experts.linear_fc2.weight0
1024
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight5
1025
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight3
1026
+ Params for bucket 11 (40370176 elements, 40370176 padded size):
1027
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight0
1028
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight7
1029
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight5
1030
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight3
1031
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight1
1032
+ module.decoder.layers.18.mlp.experts.linear_fc1.weight1
1033
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight6
1034
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight4
1035
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight2
1036
+ module.decoder.layers.17.mlp.experts.linear_fc2.weight0
1037
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight7
1038
+ Params for bucket 12 (40370176 elements, 40370176 padded size):
1039
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight6
1040
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight4
1041
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight2
1042
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight0
1043
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight3
1044
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight5
1045
+ module.decoder.layers.17.mlp.experts.linear_fc1.weight1
1046
+ Params for bucket 13 (40370176 elements, 40370176 padded size):
1047
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight6
1048
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight4
1049
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight2
1050
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight0
1051
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight7
1052
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight5
1053
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight7
1054
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight5
1055
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight3
1056
+ module.decoder.layers.16.mlp.experts.linear_fc2.weight1
1057
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight6
1058
+ Params for bucket 14 (40370176 elements, 40370176 padded size):
1059
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight3
1060
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight1
1061
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight7
1062
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight5
1063
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight4
1064
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight2
1065
+ module.decoder.layers.16.mlp.experts.linear_fc1.weight0
1066
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight6
1067
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight4
1068
+ Params for bucket 15 (40370176 elements, 40370176 padded size):
1069
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight3
1070
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight1
1071
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight6
1072
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight4
1073
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight7
1074
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight0
1075
+ module.decoder.layers.15.mlp.experts.linear_fc2.weight2
1076
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight5
1077
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight3
1078
+ Params for bucket 16 (40370176 elements, 40370176 padded size):
1079
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight2
1080
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight0
1081
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight6
1082
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight4
1083
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight2
1084
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight0
1085
+ module.decoder.layers.15.mlp.experts.linear_fc1.weight1
1086
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight7
1087
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight5
1088
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight3
1089
+ module.decoder.layers.14.mlp.experts.linear_fc2.weight1
1090
+ Params for bucket 17 (40370176 elements, 40370176 padded size):
1091
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight7
1092
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight3
1093
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight5
1094
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight1
1095
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight6
1096
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight4
1097
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight2
1098
+ Params for bucket 18 (40370176 elements, 40370176 padded size):
1099
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight7
1100
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight5
1101
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight3
1102
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight1
1103
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight7
1104
+ module.decoder.layers.14.mlp.experts.linear_fc1.weight0
1105
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight6
1106
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight4
1107
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight2
1108
+ module.decoder.layers.13.mlp.experts.linear_fc2.weight0
1109
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight6
1110
+ Params for bucket 19 (40370176 elements, 40370176 padded size):
1111
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight5
1112
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight3
1113
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight1
1114
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight7
1115
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight4
1116
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight2
1117
+ module.decoder.layers.13.mlp.experts.linear_fc1.weight0
1118
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight6
1119
+ Params for bucket 20 (40370176 elements, 40370176 padded size):
1120
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight5
1121
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight3
1122
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight1
1123
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight6
1124
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight4
1125
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight4
1126
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight2
1127
+ module.decoder.layers.12.mlp.experts.linear_fc2.weight0
1128
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight7
1129
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight5
1130
+ Params for bucket 21 (40370176 elements, 40370176 padded size):
1131
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight0
1132
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight7
1133
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight2
1134
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight5
1135
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight3
1136
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight3
1137
+ module.decoder.layers.12.mlp.experts.linear_fc1.weight1
1138
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight6
1139
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight4
1140
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight2
1141
+ Params for bucket 22 (40370176 elements, 40370176 padded size):
1142
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight1
1143
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight6
1144
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight4
1145
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight2
1146
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight5
1147
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight0
1148
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight7
1149
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight3
1150
+ Params for bucket 23 (40370176 elements, 40370176 padded size):
1151
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight0
1152
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight6
1153
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight4
1154
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight2
1155
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight0
1156
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight7
1157
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight1
1158
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight7
1159
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight5
1160
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight3
1161
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight1
1162
+ Params for bucket 24 (40370176 elements, 40370176 padded size):
1163
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight3
1164
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight5
1165
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight1
1166
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight6
1167
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight4
1168
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight2
1169
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight0
1170
+ Params for bucket 25 (40370176 elements, 40370176 padded size):
1171
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight7
1172
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight5
1173
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight3
1174
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight1
1175
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight7
1176
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight5
1177
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight6
1178
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight4
1179
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight2
1180
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight0
1181
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight6
1182
+ Params for bucket 26 (40370176 elements, 40370176 padded size):
1183
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight1
1184
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight3
1185
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight7
1186
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight5
1187
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight4
1188
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight2
1189
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight0
1190
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight6
1191
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight4
1192
+ Params for bucket 27 (40370176 elements, 40370176 padded size):
1193
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight3
1194
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight1
1195
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight6
1196
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight4
1197
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight0
1198
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight7
1199
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight2
1200
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight5
1201
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight3
1202
+ Params for bucket 28 (40370176 elements, 40370176 padded size):
1203
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight2
1204
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight0
1205
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight6
1206
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight4
1207
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight2
1208
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight0
1209
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight1
1210
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight7
1211
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight5
1212
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight3
1213
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight1
1214
+ Params for bucket 29 (40370176 elements, 40370176 padded size):
1215
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight4
1216
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight6
1217
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight2
1218
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight7
1219
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight5
1220
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight3
1221
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight1
1222
+ Params for bucket 30 (40370176 elements, 40370176 padded size):
1223
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight0
1224
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight5
1225
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight7
1226
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight3
1227
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight1
1228
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight6
1229
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight6
1230
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight4
1231
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight2
1232
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight0
1233
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight7
1234
+ Params for bucket 31 (40370176 elements, 40370176 padded size):
1235
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight4
1236
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight2
1237
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight0
1238
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight7
1239
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight5
1240
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight1
1241
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight3
1242
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight6
1243
+ Params for bucket 32 (40370176 elements, 40370176 padded size):
1244
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight5
1245
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight3
1246
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight1
1247
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight6
1248
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight4
1249
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight4
1250
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight2
1251
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight0
1252
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight7
1253
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight5
1254
+ Params for bucket 33 (40370176 elements, 40370176 padded size):
1255
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight0
1256
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight7
1257
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight2
1258
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight5
1259
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight3
1260
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight3
1261
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight1
1262
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight6
1263
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight4
1264
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight2
1265
+ Params for bucket 34 (40370176 elements, 40370176 padded size):
1266
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight1
1267
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight6
1268
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight4
1269
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight2
1270
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight0
1271
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight5
1272
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight7
1273
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight3
1274
+ Params for bucket 35 (40370176 elements, 40370176 padded size):
1275
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight0
1276
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight6
1277
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight4
1278
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight2
1279
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight0
1280
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight7
1281
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight1
1282
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight7
1283
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight5
1284
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight3
1285
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight1
1286
+ Params for bucket 36 (40370176 elements, 40370176 padded size):
1287
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight5
1288
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight3
1289
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight1
1290
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight6
1291
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight2
1292
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight4
1293
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight0
1294
+ Params for bucket 37 (40370176 elements, 40370176 padded size):
1295
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight4
1296
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight6
1297
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight2
1298
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight0
1299
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight6
1300
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight7
1301
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight5
1302
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight3
1303
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight1
1304
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight7
1305
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight5
1306
+ Params for bucket 38 (40370176 elements, 40370176 padded size):
1307
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight4
1308
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight0
1309
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight2
1310
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight7
1311
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight5
1312
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight1
1313
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight3
1314
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight6
1315
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight4
1316
+ Params for bucket 39 (40370176 elements, 40370176 padded size):
1317
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight3
1318
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight7
1319
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight1
1320
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight5
1321
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight3
1322
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight0
1323
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight2
1324
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight6
1325
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight4
1326
+ Params for bucket 40 (17301504 elements, 17301504 padded size):
1327
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight2
1328
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight1
1329
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight0
1330
+ 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 0x71365922f7d0>, config_logger_dir='')
1331
+ INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine
1332
+ [after model, optimizer, and learning rate scheduler are built] datetime: 2026-08-26 18:45:07
1333
+ > building train, validation, and test datasets ...
1334
+ > datasets target sizes (minimum size):
1335
+ train: 96
1336
+ validation: 0
1337
+ test: 0
1338
+ INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None]
1339
+ > building train, validation, and test datasets for GPT ...
1340
+ INFO:megatron.core.datasets.blended_megatron_dataset_builder:Building GPTDataset splits with sizes=(96, 0, 0) and config=GPTDatasetConfig(random_seed=1234, sequence_length=8192, blend=(['/workspace/data/wikitext/wikitext103_gpt2_text_document'], None), blend_per_split=None, split='100,0,0', split_matrix=[(0, 1.0), None, None], num_dataset_builder_threads=1, path_to_cache=None, mmap_bin_files=True, mock=False, tokenizer=<megatron.training.tokenizer.tokenizer._GPT2BPETokenizer object at 0x7133b2f45bb0>, 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)
1341
+ INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx
1342
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths
1343
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers
1344
+ INFO:megatron.core.datasets.indexed_dataset: Extract the document indices
1345
+ INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029
1346
+ INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350
1347
+ INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices
1348
+ INFO:megatron.core.datasets.gpt_dataset: Load the document index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-document_index.npy
1349
+ INFO:megatron.core.datasets.gpt_dataset: Load the sample index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-sample_index.npy
1350
+ INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-shuffle_index.npy
1351
+ INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 14536
1352
+ > finished creating GPT datasets ...
1353
+ [after dataloaders are built] datetime: 2026-08-26 18:45:07
1354
+ done with setup ...
1355
+ (min, max) time across ranks (ms):
1356
+ model-and-optimizer-setup ......................: (3802.70, 3820.41)
1357
+ train/valid/test-data-iterators-setup ..........: (65.14, 75.88)
1358
+ training ...
1359
+ Overwriting rerun_state_machine.current_iteration from -1 to 0...
1360
+ [before the start of training step] datetime: 2026-08-26 18:45:07
1361
+ INFO:moonep.api:MoonEP comm buffer sizes:
1362
+ config : R=8, S=8192, H=2048, E=64, K=6, NvS=51184, NvS_padded=51200
1363
+ sm_count : num_sms=32, num_sms_dedup=148
1364
+ hidden_buffer_size : 200.00 MiB
1365
+ meta_buffer_size : 2.00 MiB
1366
+ local_temp_buffer_size : 2.01 MiB
1367
+ total_size : 204.01 MiB
1368
+ [2026-08-26 18:45:30] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 22681.7 | throughput per GPU (TFLOP/s/GPU): 5.6 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.092206E+01 | loss scale: 1.0 | grad norm: 30.650 | number of skipped iterations: 0 | number of nan iterations: 0 |
1369
+ Number of parameters in transformer block in billions: 13.53
1370
+ Number of parameters in embedding layers in billions: 0.21
1371
+ Total number of parameters in billions: 13.74
1372
+ Number of parameters in most loaded shard in billions: 13.7356
1373
+ Theoretical memory footprints: weight and optimizer=98244.96 MB
1374
+ [Rank 0] (after 1 iterations) memory (MB) | allocated: 66144.1328125 | max allocated: 124949.08837890625 | reserved: 127502.0 | max reserved: 127502.0
1375
+ [2026-08-26 18:45:31] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 726.3 | throughput per GPU (TFLOP/s/GPU): 175.7 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.091986E+01 | loss scale: 1.0 | grad norm: 30.565 | number of skipped iterations: 0 | number of nan iterations: 0 |
1376
+ [2026-08-26 18:45:31] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 720.9 | throughput per GPU (TFLOP/s/GPU): 177.0 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 9.956084E+00 | loss scale: 1.0 | grad norm: 42.533 | number of skipped iterations: 0 | number of nan iterations: 0 |
1377
+ Capture CUDA graph for training!!!
1378
+ CUDA graph capture done!!!
1379
+ [2026-08-26 18:45:33] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 1912.6 | throughput per GPU (TFLOP/s/GPU): 66.7 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 9.478545E+00 | loss scale: 1.0 | grad norm: 7.152 | number of skipped iterations: 0 | number of nan iterations: 0 |
1380
+ [2026-08-26 18:45:34] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 699.2 | throughput per GPU (TFLOP/s/GPU): 182.5 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.091896E+01 | loss scale: 1.0 | grad norm: 452.732 | number of skipped iterations: 0 | number of nan iterations: 0 |
1381
+ [2026-08-26 18:45:35] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 704.9 | throughput per GPU (TFLOP/s/GPU): 181.0 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 9.375167E+00 | loss scale: 1.0 | grad norm: 4.992 | number of skipped iterations: 0 | number of nan iterations: 0 |
1382
+ [2026-08-26 18:45:35] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 699.4 | throughput per GPU (TFLOP/s/GPU): 182.4 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 8.982457E+00 | loss scale: 1.0 | grad norm: 3.717 | number of skipped iterations: 0 | number of nan iterations: 0 |
1383
+ [2026-08-26 18:45:36] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 699.8 | throughput per GPU (TFLOP/s/GPU): 182.3 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.787875E+00 | loss scale: 1.0 | grad norm: 4.876 | number of skipped iterations: 0 | number of nan iterations: 0 |
1384
+ [2026-08-26 18:45:37] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 700.3 | throughput per GPU (TFLOP/s/GPU): 182.2 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.620937E+00 | loss scale: 1.0 | grad norm: 2.990 | number of skipped iterations: 0 | number of nan iterations: 0 |
1385
+ [2026-08-26 18:45:37] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 701.0 | throughput per GPU (TFLOP/s/GPU): 182.0 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.680684E+00 | loss scale: 1.0 | grad norm: 194.973 | number of skipped iterations: 0 | number of nan iterations: 0 |
1386
+ [2026-08-26 18:45:38] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 701.6 | throughput per GPU (TFLOP/s/GPU): 181.8 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.583418E+00 | loss scale: 1.0 | grad norm: 24.333 | number of skipped iterations: 0 | number of nan iterations: 0 |
1387
+ [2026-08-26 18:45:39] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 698.4 | throughput per GPU (TFLOP/s/GPU): 182.7 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.611391E+00 | loss scale: 1.0 | grad norm: 75.148 | number of skipped iterations: 0 | number of nan iterations: 0 |
1388
+ [after training is done] datetime: 2026-08-26 18:45:39
1389
+ [rank0]:[W826 18:45:40.743637444 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())
1390
+ W0826 18:45:48.261000 552370 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers
1391
+ W0826 18:45:48.261000 552370 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 552458 closing signal SIGTERM
1392
+ W0826 18:45:48.262000 552370 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 552459 closing signal SIGTERM
1393
+ W0826 18:45:48.264000 552370 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 552460 closing signal SIGTERM
1394
+ W0826 18:45:48.270000 552370 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 552461 closing signal SIGTERM
1395
+ W0826 18:45:48.272000 552370 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 552462 closing signal SIGTERM
1396
+ W0826 18:45:48.276000 552370 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 552463 closing signal SIGTERM
1397
+ W0826 18:45:48.278000 552370 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 552464 closing signal SIGTERM
1398
+ W0826 18:45:48.281000 552370 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 552465 closing signal SIGTERM
1399
+ Traceback (most recent call last):
1400
+ File "/usr/local/bin/torchrun", line 8, in <module>
1401
+ sys.exit(main())
1402
+ ^^^^^^
1403
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
1404
+ return f(*args, **kwargs)
1405
+ ^^^^^^^^^^^^^^^^^^
1406
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 892, in main
1407
+ run(args)
1408
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 883, in run
1409
+ elastic_launch(
1410
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
1411
+ return launch_agent(self._config, self._entrypoint, list(args))
1412
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1413
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 261, in launch_agent
1414
+ result = agent.run()
1415
+ ^^^^^^^^^^^
1416
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper
1417
+ result = f(*args, **kwargs)
1418
+ ^^^^^^^^^^^^^^^^^^
1419
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 711, in run
1420
+ result = self._invoke_run(role)
1421
+ ^^^^^^^^^^^^^^^^^^^^^^
1422
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 870, in _invoke_run
1423
+ time.sleep(monitor_interval)
1424
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 84, in _terminate_process_handler
1425
+ raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval)
1426
+ torch.distributed.elastic.multiprocessing.api.SignalException: Process 552370 got signal: 15
moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep1/mem_samples.csv ADDED
@@ -0,0 +1,200 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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, 0
10
+ 1, 0
11
+ 2, 0
12
+ 3, 0
13
+ 4, 0
14
+ 5, 0
15
+ 6, 0
16
+ 7, 0
17
+ 0, 1
18
+ 1, 1
19
+ 2, 1
20
+ 3, 1
21
+ 4, 1
22
+ 5, 1
23
+ 6, 1
24
+ 7, 1
25
+ 0, 4
26
+ 1, 4
27
+ 2, 4
28
+ 3, 4
29
+ 4, 1
30
+ 5, 1
31
+ 6, 1
32
+ 7, 1
33
+ 0, 4
34
+ 1, 4
35
+ 2, 4
36
+ 3, 4
37
+ 4, 4
38
+ 5, 4
39
+ 6, 1
40
+ 7, 1
41
+ 0, 4
42
+ 1, 4
43
+ 2, 4
44
+ 3, 4
45
+ 4, 4
46
+ 5, 4
47
+ 6, 4
48
+ 7, 4
49
+ 0, 4
50
+ 1, 41
51
+ 2, 45
52
+ 3, 35
53
+ 4, 629
54
+ 5, 629
55
+ 6, 39
56
+ 7, 45
57
+ 0, 1053
58
+ 1, 1055
59
+ 2, 1055
60
+ 3, 1067
61
+ 4, 1067
62
+ 5, 1067
63
+ 6, 1059
64
+ 7, 1039
65
+ 0, 1803
66
+ 1, 1803
67
+ 2, 1803
68
+ 3, 1803
69
+ 4, 1803
70
+ 5, 1803
71
+ 6, 1803
72
+ 7, 1803
73
+ 0, 6767
74
+ 1, 6769
75
+ 2, 6759
76
+ 3, 6747
77
+ 4, 6763
78
+ 5, 6733
79
+ 6, 6747
80
+ 7, 6765
81
+ 0, 58521
82
+ 1, 57875
83
+ 2, 57881
84
+ 3, 58617
85
+ 4, 58605
86
+ 5, 58589
87
+ 6, 58565
88
+ 7, 58589
89
+ 0, 59109
90
+ 1, 58541
91
+ 2, 58541
92
+ 3, 59241
93
+ 4, 59229
94
+ 5, 59241
95
+ 6, 59239
96
+ 7, 59241
97
+ 0, 59737
98
+ 1, 59169
99
+ 2, 59169
100
+ 3, 59869
101
+ 4, 59869
102
+ 5, 59869
103
+ 6, 59869
104
+ 7, 59869
105
+ 0, 60439
106
+ 1, 59869
107
+ 2, 59867
108
+ 3, 60569
109
+ 4, 60569
110
+ 5, 60569
111
+ 6, 60569
112
+ 7, 60569
113
+ 0, 84165
114
+ 1, 84581
115
+ 2, 84205
116
+ 3, 84407
117
+ 4, 84295
118
+ 5, 85833
119
+ 6, 84707
120
+ 7, 84519
121
+ 0, 118055
122
+ 1, 118069
123
+ 2, 118071
124
+ 3, 118101
125
+ 4, 118077
126
+ 5, 118077
127
+ 6, 118073
128
+ 7, 118055
129
+ 0, 118939
130
+ 1, 118981
131
+ 2, 118963
132
+ 3, 118977
133
+ 4, 118977
134
+ 5, 118977
135
+ 6, 118977
136
+ 7, 118977
137
+ 0, 118939
138
+ 1, 118983
139
+ 2, 118965
140
+ 3, 118979
141
+ 4, 118979
142
+ 5, 118979
143
+ 6, 118979
144
+ 7, 118979
145
+ 0, 119599
146
+ 1, 119641
147
+ 2, 119623
148
+ 3, 119637
149
+ 4, 119637
150
+ 5, 119637
151
+ 6, 119637
152
+ 7, 119637
153
+ 0, 128787
154
+ 1, 128497
155
+ 2, 128531
156
+ 3, 128659
157
+ 4, 128647
158
+ 5, 128635
159
+ 6, 128645
160
+ 7, 128631
161
+ 0, 142739
162
+ 1, 142553
163
+ 2, 142535
164
+ 3, 142671
165
+ 4, 142671
166
+ 5, 142671
167
+ 6, 142671
168
+ 7, 142671
169
+ 0, 142739
170
+ 1, 142553
171
+ 2, 142535
172
+ 3, 142671
173
+ 4, 142671
174
+ 5, 142671
175
+ 6, 142671
176
+ 7, 142671
177
+ 0, 142499
178
+ 1, 142313
179
+ 2, 142295
180
+ 3, 142431
181
+ 4, 142431
182
+ 5, 142431
183
+ 6, 142431
184
+ 7, 142431
185
+ 0, 15364
186
+ 1, 142311
187
+ 2, 142293
188
+ 3, 142429
189
+ 4, 142429
190
+ 5, 142429
191
+ 6, 142429
192
+ 7, 142429
193
+ 0, 13441
194
+ 1, 13441
195
+ 2, 13441
196
+ 3, 13441
197
+ 4, 13761
198
+ 5, 142429
199
+ 6, 15044
200
+ 7, 13444
moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep1/run.env ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=moonep
2
+ config=QWEN2_57B
3
+ family=tierB
4
+ repo=/workspace/Megatron-MoE-Echo
5
+ branch=moonep_backend_bracing_regions_through_kernels
6
+ commit=1aac8bdd
7
+ num_layers=10
8
+ hidden=3584
9
+ ffn_hidden=2560
10
+ moe_ffn_hidden=2560
11
+ num_experts=64
12
+ topk=8
13
+ experts_per_rank=8
14
+ expert_params=27525120
15
+ heads=28
16
+ kv_groups=4
17
+ kv_channels=
18
+ qk_layernorm=0
19
+ shared_ffn=20480
20
+ shared_expert_overlap=0
21
+ mla=0
22
+ dense_first_layers=0
23
+ moe_layers=10
24
+ rotary_base=1000000
25
+ rotary_scaling=
26
+ qkv_bias=1
27
+ router_balancing=aux_loss
28
+ aux_loss_coeff=1e-2
29
+ router_pre_softmax=1
30
+ router_fusion=0
31
+ router_dtype=fp32
32
+ device_initiated_grouped_gemm=1
33
+ norm_epsilon=1e-6
34
+ seq_length=4096
35
+ micro_batch_size=1
36
+ global_batch_size=8
37
+ num_microbatches=1
38
+ assignments_per_rank=32768
39
+ mean_tokens_per_expert=4096
40
+ slot_policy=matched
41
+ slots_per_rank=8
42
+ capacity_mode=none
43
+ capacity=none
44
+ cuda_graph=1
45
+ probe_mode=0
46
+ async_grad_reduce=0
47
+ branch_role=timing
48
+ memory_dump_mode=none
49
+ memory_snapshot_path=
50
+ cuda_graph_warmup_steps=3
51
+ recompute=none
52
+ force_balance=0
53
+ train_iters=12
54
+ ep_size=8
55
+ echo_dispatch_overlap=1
56
+ echo_dispatch_recompute=0
57
+ ultraep_quota_min_tokens=
58
+ ultraep_weight_sync_join=
59
+ profile_enable=0
60
+ kernel_trace=0
61
+ fp8_recipe=mxfp8
moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep1/train.log ADDED
@@ -0,0 +1,1049 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ==============================================================
2
+ QWEN2_57B / moonep (slots/rank=8 policy=matched)
3
+ E=64 K=8 H'=2560 m=8
4
+ L=10 seq=4096 gbs=8 (1 microbatches)
5
+ cap_mode=none cap=none cg=1 recompute=none
6
+ memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep1/memory_snapshots/snapshot.pickle
7
+ repo=/workspace/Megatron-MoE-Echo @ 1aac8bdd
8
+ ==============================================================
9
+ W0826 18:45:54.778000 554845 torch/distributed/run.py:766]
10
+ W0826 18:45:54.778000 554845 torch/distributed/run.py:766] *****************************************
11
+ W0826 18:45:54.778000 554845 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
12
+ W0826 18:45:54.778000 554845 torch/distributed/run.py:766] *****************************************
13
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
14
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
15
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
16
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
17
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
18
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
19
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
20
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
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
+ 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
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
+ 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
27
+ WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer
28
+ Number of virtual stages per pipeline stage: None
29
+ accumulate and all-reduce gradients in fp32 for bfloat16 data type.
30
+ using torch.bfloat16 for parameters ...
31
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
32
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
33
+ ------------------------ arguments ------------------------
34
+ account_for_embedding_in_pipeline_split ......... False
35
+ account_for_loss_in_pipeline_split .............. False
36
+ accumulate_allreduce_grads_in_fp32 .............. True
37
+ adam_beta1 ...................................... 0.9
38
+ adam_beta2 ...................................... 0.999
39
+ adam_eps ........................................ 1e-08
40
+ add_bias_linear ................................. False
41
+ add_position_embedding .......................... False
42
+ add_qkv_bias .................................... True
43
+ adlr_autoresume ................................. False
44
+ adlr_autoresume_interval ........................ 1000
45
+ align_grad_reduce ............................... True
46
+ align_param_gather .............................. False
47
+ app_tag_run_name ................................ None
48
+ app_tag_run_version ............................. 0.0.0
49
+ apply_layernorm_1p .............................. False
50
+ apply_query_key_layer_scaling ................... False
51
+ apply_residual_connection_post_layernorm ........ False
52
+ apply_rope_fusion ............................... True
53
+ async_save ...................................... None
54
+ async_tensor_model_parallel_allreduce ........... True
55
+ attention_backend ............................... AttnBackend.auto
56
+ attention_dropout ............................... 0.0
57
+ attention_softmax_in_fp32 ....................... False
58
+ auto_detect_ckpt_format ......................... False
59
+ barrier_with_L1_time ............................ True
60
+ bert_binary_head ................................ True
61
+ bert_embedder_type .............................. megatron
62
+ bert_load ....................................... None
63
+ bf16 ............................................ True
64
+ bias_dropout_fusion ............................. True
65
+ bias_gelu_fusion ................................ False
66
+ bias_swiglu_fusion .............................. True
67
+ biencoder_projection_dim ........................ 0
68
+ biencoder_shared_query_context_model ............ False
69
+ block_data_path ................................. None
70
+ calc_ft_timeouts ................................ False
71
+ calculate_per_token_loss ........................ False
72
+ check_for_large_grads ........................... False
73
+ check_for_nan_in_loss_and_grad .................. False
74
+ check_for_spiky_loss ............................ False
75
+ check_weight_hash_across_dp_replicas_interval ... None
76
+ ckpt_assume_constant_structure .................. False
77
+ ckpt_convert_format ............................. None
78
+ ckpt_convert_save ............................... None
79
+ ckpt_convert_update_legacy_dist_opt_format ...... False
80
+ ckpt_format ..................................... torch_dist
81
+ ckpt_fully_parallel_load ........................ False
82
+ ckpt_fully_parallel_save ........................ True
83
+ ckpt_fully_parallel_save_deprecated ............. False
84
+ ckpt_step ....................................... None
85
+ classes_fraction ................................ 1.0
86
+ clip_grad ....................................... 1.0
87
+ clone_scatter_output_in_embedding ............... True
88
+ config_logger_dir ...............................
89
+ consumed_train_samples .......................... 0
90
+ consumed_valid_samples .......................... 0
91
+ context_parallel_size ........................... 1
92
+ cp_comm_type .................................... ['p2p']
93
+ create_attention_mask_in_dataloader ............. True
94
+ cross_entropy_fusion_impl ....................... native
95
+ cross_entropy_loss_fusion ....................... False
96
+ cuda_graph_scope ................................ full_iteration
97
+ cuda_graph_warmup_steps ......................... 3
98
+ data_args_path .................................. None
99
+ data_cache_path ................................. None
100
+ data_parallel_random_init ....................... False
101
+ data_parallel_sharding_strategy ................. no_shard
102
+ data_parallel_size .............................. 8
103
+ data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document']
104
+ data_per_class_fraction ......................... 1.0
105
+ data_sharding ................................... True
106
+ dataloader_type ................................. single
107
+ ddp_average_in_collective ....................... False
108
+ ddp_bucket_size ................................. None
109
+ ddp_num_buckets ................................. None
110
+ ddp_pad_buckets_for_high_nccl_busbw ............. False
111
+ decoder_first_pipeline_num_layers ............... None
112
+ decoder_last_pipeline_num_layers ................ None
113
+ decoder_num_layers .............................. None
114
+ decoder_seq_length .............................. None
115
+ decoupled_lr .................................... None
116
+ decoupled_min_lr ................................ None
117
+ decrease_batch_size_if_needed ................... False
118
+ defer_embedding_wgrad_compute ................... False
119
+ delay_wgrad_compute ............................. False
120
+ deprecated_use_mcore_models ..................... True
121
+ deterministic_mode .............................. False
122
+ dino_bottleneck_size ............................ 256
123
+ dino_freeze_last_layer .......................... 1
124
+ dino_head_hidden_size ........................... 2048
125
+ dino_local_crops_number ......................... 10
126
+ dino_local_img_size ............................. 96
127
+ dino_norm_last_layer ............................ False
128
+ dino_teacher_temp ............................... 0.07
129
+ dino_warmup_teacher_temp ........................ 0.04
130
+ dino_warmup_teacher_temp_epochs ................. 30
131
+ disable_bf16_reduced_precision_matmul ........... False
132
+ disable_mamba_mem_eff_path ...................... False
133
+ disable_straggler_on_startup .................... False
134
+ dist_ckpt_format_deprecated ..................... None
135
+ dist_ckpt_strictness ............................ assume_ok_unexpected
136
+ distribute_saved_activations .................... False
137
+ distributed_backend ............................. nccl
138
+ distributed_timeout_minutes ..................... 10
139
+ embedding_init_method_std ....................... None
140
+ embedding_path .................................. None
141
+ empty_unused_memory_level ....................... 0
142
+ enable_cuda_graph ............................... True
143
+ enable_experimental ............................. False
144
+ enable_ft_package ............................... False
145
+ enable_gloo_process_groups ...................... True
146
+ enable_msc ...................................... True
147
+ enable_one_logger ............................... True
148
+ encoder_num_layers .............................. 10
149
+ encoder_seq_length .............................. 4096
150
+ end_weight_decay ................................ 0.1
151
+ eod_mask_loss ................................... False
152
+ error_injection_rate ............................ 0
153
+ error_injection_type ............................ transient_error
154
+ eval_interval ................................... 1000000
155
+ eval_iters ...................................... 0
156
+ evidence_data_path .............................. None
157
+ exit_duration_in_mins ........................... None
158
+ exit_interval ................................... None
159
+ exit_on_missing_checkpoint ...................... False
160
+ exit_signal_handler ............................. False
161
+ exp_avg_dtype ................................... torch.float32
162
+ exp_avg_sq_dtype ................................ torch.float32
163
+ expert_model_parallel_size ...................... 8
164
+ expert_tensor_parallel_size ..................... 1
165
+ export_force_local_attention .................... False
166
+ export_kd_cfg ................................... None
167
+ export_kd_teacher_ckpt_format ................... None
168
+ export_kd_teacher_load .......................... None
169
+ export_kv_cache_quant ........................... False
170
+ export_legacy_megatron .......................... False
171
+ export_model_type ............................... GPTModel
172
+ export_moe_apply_probs_on_input ................. False
173
+ export_qk_l2_norm ............................... False
174
+ export_quant_cfg ................................ None
175
+ export_real_quant_cfg ........................... None
176
+ export_te_mcore_model ........................... False
177
+ external_cuda_graph ............................. False
178
+ ffn_hidden_size ................................. 2560
179
+ finetune ........................................ False
180
+ finetune_data_split ............................. train
181
+ finetune_hf_dataset ............................. None
182
+ first_last_layers_bf16 .......................... False
183
+ flash_decode .................................... False
184
+ fp16 ............................................ False
185
+ fp16_lm_cross_entropy ........................... False
186
+ fp32_residual_connection ........................ False
187
+ fp8 ............................................. e4m3
188
+ fp8_amax_compute_algo ........................... most_recent
189
+ fp8_amax_history_len ............................ 1
190
+ fp8_interval .................................... 1
191
+ fp8_margin ...................................... 0
192
+ fp8_param_gather ................................ False
193
+ fp8_recipe ...................................... mxfp8
194
+ fp8_wgrad ....................................... True
195
+ fsdp_double_buffer .............................. False
196
+ global_batch_size ............................... 8
197
+ grad_reduce_in_bf16 ............................. False
198
+ gradient_accumulation_fusion .................... True
199
+ gradient_reduce_div_fusion ...................... True
200
+ group_query_attention ........................... True
201
+ head_lr_mult .................................... 1.0
202
+ heterogeneous_layers_config_encoded_json ........ None
203
+ heterogeneous_layers_config_path ................ None
204
+ hidden_dropout .................................. 0.0
205
+ hidden_size ..................................... 3584
206
+ hierarchical_context_parallel_sizes ............. None
207
+ high_priority_stream_groups ..................... []
208
+ hybrid_attention_ratio .......................... 0.0
209
+ hybrid_mlp_ratio ................................ 0.0
210
+ hybrid_override_pattern ......................... None
211
+ hysteresis ...................................... 2
212
+ ict_head_size ................................... None
213
+ ict_load ........................................ None
214
+ img_h ........................................... 224
215
+ img_w ........................................... 224
216
+ indexer_batch_size .............................. 128
217
+ indexer_log_interval ............................ 1000
218
+ inference_batch_times_seqlen_threshold .......... -1
219
+ inference_dynamic_batching ...................... False
220
+ inference_dynamic_batching_buffer_guaranteed_fraction 0.2
221
+ inference_dynamic_batching_buffer_overflow_factor None
222
+ inference_dynamic_batching_buffer_size_gb ....... 40.0
223
+ inference_dynamic_batching_chunk_size ........... 256
224
+ inference_dynamic_batching_max_requests_override None
225
+ inference_dynamic_batching_max_tokens_override .. None
226
+ inference_dynamic_batching_num_cuda_graphs ...... 16
227
+ inference_max_batch_size ........................ 8
228
+ inference_max_seq_length ........................ 2560
229
+ inference_rng_tracker ........................... False
230
+ init_method_std ................................. 0.01
231
+ init_method_xavier_uniform ...................... False
232
+ init_model_with_meta_device ..................... False
233
+ initial_loss_scale .............................. 4294967296
234
+ inprocess_active_world_size ..................... 8
235
+ inprocess_barrier_timeout ....................... 120
236
+ inprocess_completion_timeout .................... 120
237
+ inprocess_empty_cuda_cache ...................... False
238
+ inprocess_granularity ........................... node
239
+ inprocess_hard_timeout .......................... 90
240
+ inprocess_heartbeat_interval .................... 30
241
+ inprocess_heartbeat_timeout ..................... 60
242
+ inprocess_last_call_wait ........................ 1
243
+ inprocess_max_iterations ........................ None
244
+ inprocess_monitor_process_interval .............. 1.0
245
+ inprocess_monitor_thread_interval ............... 1.0
246
+ inprocess_progress_watchdog_interval ............ 1.0
247
+ inprocess_restart ............................... False
248
+ inprocess_soft_timeout .......................... 60
249
+ inprocess_termination_grace_time ................ 1
250
+ is_hybrid_model ................................. False
251
+ iter_per_epoch .................................. 1250
252
+ iterations_to_skip .............................. []
253
+ keep_fp8_transpose_cache_when_using_custom_fsdp . False
254
+ kitchen_config_file ............................. None
255
+ kitchen_recipe_number ........................... None
256
+ kv_channels ..................................... 128
257
+ kv_lora_rank .................................... 32
258
+ lazy_mpu_init ................................... None
259
+ load ............................................ None
260
+ load_main_params_from_ckpt ...................... None
261
+ load_model_opt_format ........................... False
262
+ local_rank ...................................... 0
263
+ log_energy ...................................... False
264
+ log_interval .................................... 1
265
+ log_loss_scale_to_tensorboard ................... True
266
+ log_memory_to_tensorboard ....................... False
267
+ log_num_zeros_in_grad ........................... False
268
+ log_params_norm ................................. False
269
+ log_progress .................................... False
270
+ log_straggler ................................... False
271
+ log_throughput .................................. True
272
+ log_timers_to_tensorboard ....................... False
273
+ log_validation_ppl_to_tensorboard ............... False
274
+ log_world_size_to_tensorboard ................... False
275
+ logging_level ................................... None
276
+ loss_scale ...................................... None
277
+ loss_scale_window ............................... 1000
278
+ lr .............................................. 0.0001
279
+ lr_decay_iters .................................. 8
280
+ lr_decay_samples ................................ None
281
+ lr_decay_style .................................. cosine
282
+ lr_warmup_fraction .............................. None
283
+ lr_warmup_init .................................. 0.0
284
+ lr_warmup_iters ................................. 5
285
+ lr_warmup_samples ............................... 0
286
+ lr_wsd_decay_iters .............................. None
287
+ lr_wsd_decay_samples ............................ None
288
+ lr_wsd_decay_style .............................. exponential
289
+ main_grads_dtype ................................ torch.float32
290
+ main_params_dtype ............................... torch.float32
291
+ make_vocab_size_divisible_by .................... 128
292
+ mamba_head_dim .................................. 64
293
+ mamba_num_groups ................................ 8
294
+ mamba_num_heads ................................. None
295
+ mamba_state_dim ................................. 128
296
+ manual_gc ....................................... False
297
+ manual_gc_eval .................................. True
298
+ manual_gc_interval .............................. 0
299
+ mask_factor ..................................... 1.0
300
+ mask_prob ....................................... 0.15
301
+ mask_type ....................................... random
302
+ masked_softmax_fusion ........................... False
303
+ max_position_embeddings ......................... 32768
304
+ max_tokens_to_oom ............................... 12000
305
+ memory_snapshot_path ............................ snapshot.pickle
306
+ merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt
307
+ micro_batch_size ................................ 1
308
+ microbatch_group_size_per_vp_stage .............. None
309
+ mid_level_dataset_surplus ....................... 0.005
310
+ min_loss_scale .................................. 1.0
311
+ min_lr .......................................... 1e-05
312
+ mlp_chunks_for_prefill .......................... 1
313
+ mmap_bin_files .................................. True
314
+ mock_data ....................................... False
315
+ moe_apply_probs_on_input ........................ False
316
+ moe_aux_loss_coeff .............................. 0.01
317
+ moe_deepep_num_sms .............................. 20
318
+ moe_echo_enable_random_offloading ............... False
319
+ moe_echo_expert_dispatch_overlap ................ False
320
+ moe_echo_expert_dispatcher_type ................. hybridep
321
+ moe_echo_recompute_expert_dispatch .............. False
322
+ moe_enable_deepep ............................... False
323
+ moe_enable_echo ................................. False
324
+ moe_enable_hybridep ............................. False
325
+ moe_enable_moonep ............................... True
326
+ moe_expert_capacity_factor ...................... None
327
+ moe_extended_tp ................................. False
328
+ moe_ffn_hidden_size ............................. 2560
329
+ moe_grouped_gemm ................................ True
330
+ moe_hybridep_num_sms ............................ 16
331
+ moe_input_jitter_eps ............................ None
332
+ moe_layer_freq .................................. 1
333
+ moe_layer_recompute ............................. False
334
+ moe_moonep_num_sms .............................. None
335
+ moe_num_echo_experts ............................ None
336
+ moe_pad_expert_input_to_capacity ................ False
337
+ moe_per_layer_logging ........................... False
338
+ moe_permute_fusion .............................. False
339
+ moe_received_token_capacity ..................... None
340
+ moe_router_bias_update_rate ..................... 0.001
341
+ moe_router_dtype ................................ fp32
342
+ moe_router_enable_expert_bias ................... False
343
+ moe_router_force_load_balancing ................. False
344
+ moe_router_group_topk ........................... None
345
+ moe_router_load_balancing_type .................. aux_loss
346
+ moe_router_num_groups ........................... None
347
+ moe_router_padding_for_fp8 ...................... False
348
+ moe_router_pre_softmax .......................... True
349
+ moe_router_score_function ....................... softmax
350
+ moe_router_topk ................................. 8
351
+ moe_router_topk_scaling_factor .................. None
352
+ moe_shared_expert_intermediate_size ............. 20480
353
+ moe_shared_expert_overlap ....................... False
354
+ moe_token_dispatcher_type ....................... flex
355
+ moe_token_drop_policy ........................... probs
356
+ moe_upcycling_granularity ....................... 1
357
+ moe_use_device_initiated_grouped_gemm ........... True
358
+ moe_use_legacy_grouped_gemm ..................... False
359
+ moe_use_upcycling ............................... False
360
+ moe_z_loss_coeff ................................ None
361
+ mrope_section ................................... None
362
+ mscale .......................................... 1.0
363
+ mscale_all_dim .................................. 0.0
364
+ mtp_loss_scaling_factor ......................... 0.1
365
+ mtp_num_layers .................................. None
366
+ multi_latent_attention .......................... False
367
+ nccl_all_reduce_for_prefill ..................... False
368
+ nccl_communicator_config_path ................... None
369
+ nccl_ub ......................................... False
370
+ no_load_optim ................................... None
371
+ no_load_rng ..................................... None
372
+ no_persist_layer_norm ........................... False
373
+ no_rope_freq .................................... None
374
+ no_save_optim ................................... None
375
+ no_save_rng ..................................... None
376
+ non_persistent_ckpt_type ........................ None
377
+ non_persistent_global_ckpt_dir .................. None
378
+ non_persistent_local_ckpt_algo .................. fully_parallel
379
+ non_persistent_local_ckpt_dir ................... None
380
+ non_persistent_save_interval .................... None
381
+ norm_epsilon .................................... 1e-06
382
+ normalization ................................... RMSNorm
383
+ num_attention_heads ............................. 28
384
+ num_channels .................................... 3
385
+ num_classes ..................................... 1000
386
+ num_dataset_builder_threads ..................... 1
387
+ num_distributed_optimizer_instances ............. 1
388
+ num_experts ..................................... 64
389
+ num_layers ...................................... 10
390
+ num_layers_at_end_in_bf16 ....................... 1
391
+ num_layers_at_start_in_bf16 ..................... 1
392
+ num_layers_per_virtual_pipeline_stage ........... None
393
+ num_query_groups ................................ 4
394
+ num_virtual_stages_per_pipeline_rank ............ None
395
+ num_workers ..................................... 2
396
+ object_storage_cache_path ....................... None
397
+ one_logger_async ................................ False
398
+ one_logger_project .............................. megatron-lm
399
+ one_logger_run_name ............................. None
400
+ onnx_safe ....................................... None
401
+ openai_gelu ..................................... False
402
+ optimizer ....................................... adam
403
+ optimizer_cpu_offload ........................... False
404
+ optimizer_offload_fraction ...................... 1.0
405
+ output_bert_embeddings .......................... False
406
+ overlap_cpu_optimizer_d2h_h2d ................... False
407
+ overlap_grad_reduce ............................. True
408
+ overlap_moe_expert_parallel_comm ................ False
409
+ overlap_p2p_comm ................................ False
410
+ overlap_p2p_comm_warmup_flush ................... False
411
+ overlap_param_gather ............................ True
412
+ overlap_param_gather_with_optimizer_step ........ False
413
+ override_opt_param_scheduler .................... False
414
+ params_dtype .................................... torch.bfloat16
415
+ patch_dim ....................................... 16
416
+ per_split_data_args_path ........................ None
417
+ perform_initialization .......................... True
418
+ pin_cpu_grads ................................... True
419
+ pin_cpu_params .................................. True
420
+ pipeline_model_parallel_comm_backend ............ None
421
+ pipeline_model_parallel_layout .................. None
422
+ pipeline_model_parallel_size .................... 1
423
+ position_embedding_type ......................... rope
424
+ pretrained_checkpoint ........................... None
425
+ profile ......................................... False
426
+ profile_ranks ................................... [0]
427
+ profile_step_end ................................ 12
428
+ profile_step_start .............................. 10
429
+ q_lora_rank ..................................... None
430
+ qk_head_dim ..................................... 128
431
+ qk_l2_norm ...................................... False
432
+ qk_layernorm .................................... False
433
+ qk_pos_emb_head_dim ............................. 64
434
+ query_in_block_prob ............................. 0.1
435
+ rampup_batch_size ............................... None
436
+ rank ............................................ 0
437
+ recompute_granularity ........................... None
438
+ recompute_method ................................ None
439
+ recompute_modules ............................... None
440
+ recompute_num_layers ............................ None
441
+ record_memory_history ........................... False
442
+ relative_attention_max_distance ................. 128
443
+ relative_attention_num_buckets .................. 32
444
+ replication ..................................... False
445
+ replication_factor .............................. 2
446
+ replication_jump ................................ None
447
+ rerun_mode ...................................... validate_results
448
+ reset_attention_mask ............................ False
449
+ reset_position_ids .............................. False
450
+ result_rejected_tracker_filename ................ None
451
+ retriever_report_topk_accuracies ................ []
452
+ retriever_score_scaling ......................... False
453
+ retriever_seq_length ............................ 256
454
+ retro_add_retriever ............................. False
455
+ retro_attention_gate ............................ 1
456
+ retro_cyclic_train_iters ........................ None
457
+ retro_encoder_attention_dropout ................. 0.1
458
+ retro_encoder_hidden_dropout .................... 0.1
459
+ retro_encoder_layers ............................ 2
460
+ retro_num_neighbors ............................. 2
461
+ retro_num_retrieved_chunks ...................... 2
462
+ retro_project_dir ............................... None
463
+ retro_verify_neighbor_count ..................... True
464
+ reuse_grad_buf_for_mxfp8_param_ag ............... False
465
+ rope_scaling_factor ............................. 8.0
466
+ rope_type ....................................... None
467
+ rotary_base ..................................... 1000000
468
+ rotary_interleaved .............................. False
469
+ rotary_percent .................................. 1.0
470
+ rotary_scaling_factor ........................... 1.0
471
+ rotary_seq_len_interpolation_factor ............. None
472
+ run_workload_inspector_server ................... False
473
+ sample_rate ..................................... 1.0
474
+ save ............................................ None
475
+ save_interval ................................... None
476
+ save_retain_interval ............................ None
477
+ scatter_gather_tensors_in_pipeline .............. True
478
+ seed ............................................ 1234
479
+ seq_length ...................................... 4096
480
+ sequence_parallel ............................... False
481
+ sft ............................................. False
482
+ sft_tokenizer_prompt_format ..................... nemotron-h-aligned
483
+ sgd_momentum .................................... 0.9
484
+ short_seq_prob .................................. 0.1
485
+ skip_train ...................................... False
486
+ skipped_train_samples ........................... 0
487
+ spec ............................................ None
488
+ split ........................................... 100,0,0
489
+ squared_relu .................................... False
490
+ start_weight_decay .............................. 0.1
491
+ straggler_ctrlr_port ............................ 65535
492
+ straggler_minmax_count .......................... 1
493
+ suggested_communication_unit_size ............... None
494
+ swiglu .......................................... True
495
+ swin_backbone_type .............................. tiny
496
+ symmetric_ar_type ............................... None
497
+ te_rng_tracker .................................. True
498
+ tensor_model_parallel_size ...................... 1
499
+ tensorboard_dir ................................. None
500
+ tensorboard_log_interval ........................ 1
501
+ tensorboard_queue_size .......................... 1000
502
+ test_data_path .................................. None
503
+ test_mode ....................................... False
504
+ tiktoken_num_special_tokens ..................... 1000
505
+ tiktoken_pattern ................................ None
506
+ tiktoken_special_tokens ......................... None
507
+ timing_log_level ................................ 0
508
+ timing_log_option ............................... minmax
509
+ titles_data_path ................................ None
510
+ tokenizer_model ................................. None
511
+ tokenizer_type .................................. GPT2BPETokenizer
512
+ torch_fsdp2_reshard_after_forward ............... True
513
+ tp_comm_bootstrap_backend ....................... nccl
514
+ tp_comm_bulk_dgrad .............................. True
515
+ tp_comm_bulk_wgrad .............................. True
516
+ tp_comm_overlap ................................. False
517
+ tp_comm_overlap_ag .............................. True
518
+ tp_comm_overlap_cfg ............................. None
519
+ tp_comm_overlap_rs .............................. True
520
+ tp_comm_overlap_rs_dgrad ........................ False
521
+ tp_comm_split_ag ................................ True
522
+ tp_comm_split_rs ................................ True
523
+ train_data_path ................................. None
524
+ train_iters ..................................... 12
525
+ train_samples ................................... None
526
+ train_sync_interval ............................. None
527
+ transformer_impl ................................ transformer_engine
528
+ transformer_pipeline_model_parallel_size ........ 1
529
+ untie_embeddings_and_output_weights ............. True
530
+ use_checkpoint_args ............................. False
531
+ use_checkpoint_opt_param_scheduler .............. False
532
+ use_cpu_initialization .......................... None
533
+ use_custom_fsdp ................................. False
534
+ use_dist_ckpt ................................... True
535
+ use_dist_ckpt_deprecated ........................ False
536
+ use_distributed_optimizer ....................... True
537
+ use_flash_attn .................................. False
538
+ use_legacy_models ............................... False
539
+ use_mp_args_from_checkpoint_args ................ False
540
+ use_one_sent_docs ............................... False
541
+ use_persistent_ckpt_worker ...................... False
542
+ use_precision_aware_optimizer ................... False
543
+ use_pytorch_profiler ............................ False
544
+ use_ring_exchange_p2p ........................... False
545
+ use_rope_scaling ................................ False
546
+ use_rotary_position_embeddings .................. False
547
+ use_sharp ....................................... False
548
+ use_tokenizer_model_from_checkpoint_args ........ True
549
+ use_torch_fsdp2 ................................. False
550
+ use_torch_optimizer_for_cpu_offload ............. False
551
+ use_tp_pp_dp_mapping ............................ False
552
+ v_head_dim ...................................... 128
553
+ valid_data_path ................................. None
554
+ variable_seq_lengths ............................ False
555
+ virtual_pipeline_model_parallel_size ............ None
556
+ vision_backbone_type ............................ vit
557
+ vision_pretraining .............................. False
558
+ vision_pretraining_type ......................... classify
559
+ vocab_extra_ids ................................. 0
560
+ vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json
561
+ vocab_size ...................................... None
562
+ wandb_exp_name ..................................
563
+ wandb_project ...................................
564
+ wandb_save_dir ..................................
565
+ weight_decay .................................... 0.1
566
+ weight_decay_incr_style ......................... constant
567
+ wgrad_deferral_limit ............................ 0
568
+ world_size ...................................... 8
569
+ yaml_cfg ........................................ None
570
+ -------------------- end of arguments ---------------------
571
+ INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1
572
+ > building GPT2BPETokenizer tokenizer ...
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
+ > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304)
576
+ WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS
577
+ > initializing torch distributed ...
578
+ [Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
579
+ [Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
580
+ [Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
581
+ [Gloo] Rank [Gloo] Rank [Gloo] Rank [Gloo] Rank [Gloo] Rank 0214 is connected to 3 is connected to is connected to is connected to 7 is connected to 777 peer ranks. 7 peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : 77
582
+ Expected number of connected peer ranks is :
583
+ 7
584
+ peer ranks. Expected number of connected peer ranks is : 7
585
+ peer ranks. Expected number of connected peer ranks is : 7
586
+ [Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
587
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
588
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
589
+ [Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
590
+ [Gloo] Rank 6 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
592
+ [Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
593
+ [Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
594
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
595
+ > initialized tensor model parallel with size 1
596
+ > initialized pipeline model parallel with size 1
597
+ > setting random seeds to 1234 ...
598
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
599
+ [Gloo] Rank [Gloo] Rank [Gloo] Rank 000 is connected to is connected to is connected to 000 peer ranks. peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : Expected number of connected peer ranks is : 00
600
+ 0
601
+
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
+ [rank2]:[W826 18:46:13.881817328 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.
606
+ [rank6]:[W826 18:46:13.883220521 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.
607
+ [rank5]:[W826 18:46:13.883336352 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.
608
+ [rank7]:[W826 18:46:13.883388600 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.
609
+ [rank1]:[W826 18:46:13.883415032 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
610
+ [rank3]:[W826 18:46:13.884963900 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.
611
+ [rank4]:[W826 18:46:13.884970073 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
+ > compiling dataset index builder ...
613
+ make: Entering directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
614
+ make: Nothing to be done for 'default'.
615
+ make: Leaving directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
616
+ >>> done with dataset index builder. Compilation time: 0.033 seconds
617
+ WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations.
618
+ > compiling and loading fused kernels ...
619
+ [rank0]:[W826 18:46:14.047087888 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.
620
+ >>> done with compiling and loading fused kernels. Compilation time: 3.939 seconds
621
+ time to initialize megatron (seconds): 7.686
622
+ [after megatron is initialized] datetime: 2026-08-26 18:46:20
623
+ building GPT model ...
624
+ /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.
625
+ warnings.warn(
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
+ > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 5060614656
641
+ 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)
642
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 22
643
+ Params for bucket 1 (180289536 elements, 180289536 padded size):
644
+ module.output_layer.weight
645
+ Params for bucket 2 (73403904 elements, 73403904 padded size):
646
+ module.decoder.final_layernorm.weight
647
+ module.decoder.layers.9.mlp.shared_experts.linear_fc2.weight
648
+ Params for bucket 3 (146800640 elements, 146800640 padded size):
649
+ module.decoder.layers.9.mlp.shared_experts.linear_fc1.weight
650
+ Params for bucket 4 (103001600 elements, 103001600 padded size):
651
+ module.decoder.layers.9.mlp.router.weight
652
+ module.decoder.layers.9.self_attention.linear_qkv.weight
653
+ module.decoder.layers.9.pre_mlp_layernorm.weight
654
+ module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight
655
+ module.decoder.layers.9.self_attention.linear_qkv.bias
656
+ module.decoder.layers.9.self_attention.linear_proj.weight
657
+ module.decoder.layers.8.mlp.shared_experts.linear_fc2.weight
658
+ Params for bucket 5 (146800640 elements, 146800640 padded size):
659
+ module.decoder.layers.8.mlp.shared_experts.linear_fc1.weight
660
+ Params for bucket 6 (103001600 elements, 103001600 padded size):
661
+ module.decoder.layers.8.mlp.router.weight
662
+ module.decoder.layers.8.self_attention.linear_qkv.weight
663
+ module.decoder.layers.8.self_attention.linear_proj.weight
664
+ module.decoder.layers.7.mlp.shared_experts.linear_fc2.weight
665
+ module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight
666
+ module.decoder.layers.8.pre_mlp_layernorm.weight
667
+ module.decoder.layers.8.self_attention.linear_qkv.bias
668
+ Params for bucket 7 (146800640 elements, 146800640 padded size):
669
+ module.decoder.layers.7.mlp.shared_experts.linear_fc1.weight
670
+ Params for bucket 8 (103001600 elements, 103001600 padded size):
671
+ module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight
672
+ module.decoder.layers.7.self_attention.linear_qkv.bias
673
+ module.decoder.layers.7.pre_mlp_layernorm.weight
674
+ module.decoder.layers.7.self_attention.linear_proj.weight
675
+ module.decoder.layers.6.mlp.shared_experts.linear_fc2.weight
676
+ module.decoder.layers.7.mlp.router.weight
677
+ module.decoder.layers.7.self_attention.linear_qkv.weight
678
+ Params for bucket 9 (146800640 elements, 146800640 padded size):
679
+ module.decoder.layers.6.mlp.shared_experts.linear_fc1.weight
680
+ Params for bucket 10 (103001600 elements, 103001600 padded size):
681
+ module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight
682
+ module.decoder.layers.6.self_attention.linear_qkv.bias
683
+ module.decoder.layers.6.pre_mlp_layernorm.weight
684
+ module.decoder.layers.6.mlp.router.weight
685
+ module.decoder.layers.6.self_attention.linear_qkv.weight
686
+ module.decoder.layers.6.self_attention.linear_proj.weight
687
+ module.decoder.layers.5.mlp.shared_experts.linear_fc2.weight
688
+ Params for bucket 11 (146800640 elements, 146800640 padded size):
689
+ module.decoder.layers.5.mlp.shared_experts.linear_fc1.weight
690
+ Params for bucket 12 (103001600 elements, 103001600 padded size):
691
+ module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight
692
+ module.decoder.layers.5.self_attention.linear_qkv.bias
693
+ module.decoder.layers.5.pre_mlp_layernorm.weight
694
+ module.decoder.layers.5.self_attention.linear_proj.weight
695
+ module.decoder.layers.4.mlp.shared_experts.linear_fc2.weight
696
+ module.decoder.layers.5.mlp.router.weight
697
+ module.decoder.layers.5.self_attention.linear_qkv.weight
698
+ Params for bucket 13 (146800640 elements, 146800640 padded size):
699
+ module.decoder.layers.4.mlp.shared_experts.linear_fc1.weight
700
+ Params for bucket 14 (103001600 elements, 103001600 padded size):
701
+ module.decoder.layers.4.mlp.router.weight
702
+ module.decoder.layers.4.self_attention.linear_qkv.weight
703
+ module.decoder.layers.4.self_attention.linear_proj.weight
704
+ module.decoder.layers.3.mlp.shared_experts.linear_fc2.weight
705
+ module.decoder.layers.4.self_attention.linear_qkv.bias
706
+ module.decoder.layers.4.self_attention.linear_qkv.layer_norm_weight
707
+ module.decoder.layers.4.pre_mlp_layernorm.weight
708
+ Params for bucket 15 (146800640 elements, 146800640 padded size):
709
+ module.decoder.layers.3.mlp.shared_experts.linear_fc1.weight
710
+ Params for bucket 16 (103001600 elements, 103001600 padded size):
711
+ module.decoder.layers.3.pre_mlp_layernorm.weight
712
+ module.decoder.layers.3.self_attention.linear_qkv.bias
713
+ module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight
714
+ module.decoder.layers.2.mlp.shared_experts.linear_fc2.weight
715
+ module.decoder.layers.3.mlp.router.weight
716
+ module.decoder.layers.3.self_attention.linear_qkv.weight
717
+ module.decoder.layers.3.self_attention.linear_proj.weight
718
+ Params for bucket 17 (146800640 elements, 146800640 padded size):
719
+ module.decoder.layers.2.mlp.shared_experts.linear_fc1.weight
720
+ Params for bucket 18 (103001600 elements, 103001600 padded size):
721
+ module.decoder.layers.2.self_attention.linear_qkv.bias
722
+ module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight
723
+ module.decoder.layers.2.pre_mlp_layernorm.weight
724
+ module.decoder.layers.2.mlp.router.weight
725
+ module.decoder.layers.2.self_attention.linear_qkv.weight
726
+ module.decoder.layers.2.self_attention.linear_proj.weight
727
+ module.decoder.layers.1.mlp.shared_experts.linear_fc2.weight
728
+ Params for bucket 19 (146800640 elements, 146800640 padded size):
729
+ module.decoder.layers.1.mlp.shared_experts.linear_fc1.weight
730
+ Params for bucket 20 (103001600 elements, 103001600 padded size):
731
+ module.decoder.layers.1.mlp.router.weight
732
+ module.decoder.layers.1.self_attention.linear_qkv.weight
733
+ module.decoder.layers.0.mlp.shared_experts.linear_fc2.weight
734
+ module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight
735
+ module.decoder.layers.1.self_attention.linear_qkv.bias
736
+ module.decoder.layers.1.pre_mlp_layernorm.weight
737
+ module.decoder.layers.1.self_attention.linear_proj.weight
738
+ Params for bucket 21 (146800640 elements, 146800640 padded size):
739
+ module.decoder.layers.0.mlp.shared_experts.linear_fc1.weight
740
+ Params for bucket 22 (209890816 elements, 209890816 padded size):
741
+ module.decoder.layers.0.mlp.router.weight
742
+ module.decoder.layers.0.pre_mlp_layernorm.weight
743
+ module.decoder.layers.0.self_attention.linear_qkv.layer_norm_weight
744
+ module.decoder.layers.0.self_attention.linear_proj.weight
745
+ module.embedding.word_embeddings.weight
746
+ module.decoder.layers.0.self_attention.linear_qkv.weight
747
+ module.decoder.layers.0.self_attention.linear_qkv.bias
748
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 44
749
+ Params for bucket 1 (45875200 elements, 45875200 padded size):
750
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight4
751
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight6
752
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight7
753
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight5
754
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight3
755
+ Params for bucket 2 (45875200 elements, 45875200 padded size):
756
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight2
757
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight1
758
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight0
759
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight7
760
+ Params for bucket 3 (55050240 elements, 55050240 padded size):
761
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight6
762
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight4
763
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight5
764
+ Params for bucket 4 (55050240 elements, 55050240 padded size):
765
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight3
766
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight1
767
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight2
768
+ Params for bucket 5 (45875200 elements, 45875200 padded size):
769
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight0
770
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight6
771
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight7
772
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight5
773
+ Params for bucket 6 (45875200 elements, 45875200 padded size):
774
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight4
775
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight0
776
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight2
777
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight3
778
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight1
779
+ Params for bucket 7 (55050240 elements, 55050240 padded size):
780
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight7
781
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight5
782
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight6
783
+ Params for bucket 8 (55050240 elements, 55050240 padded size):
784
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight4
785
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight2
786
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight3
787
+ Params for bucket 9 (45875200 elements, 45875200 padded size):
788
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight1
789
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight0
790
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight7
791
+ Params for bucket 10 (45875200 elements, 45875200 padded size):
792
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight6
793
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight4
794
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight2
795
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight5
796
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight3
797
+ Params for bucket 11 (55050240 elements, 55050240 padded size):
798
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight1
799
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight7
800
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight6
801
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight0
802
+ Params for bucket 12 (55050240 elements, 55050240 padded size):
803
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight5
804
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight4
805
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight3
806
+ Params for bucket 13 (55050240 elements, 55050240 padded size):
807
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight2
808
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight0
809
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight1
810
+ Params for bucket 14 (45875200 elements, 45875200 padded size):
811
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight4
812
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight6
813
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight7
814
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight5
815
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight3
816
+ Params for bucket 15 (45875200 elements, 45875200 padded size):
817
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight2
818
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight0
819
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight1
820
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight7
821
+ Params for bucket 16 (55050240 elements, 55050240 padded size):
822
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight6
823
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight4
824
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight5
825
+ Params for bucket 17 (55050240 elements, 55050240 padded size):
826
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight3
827
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight1
828
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight2
829
+ Params for bucket 18 (45875200 elements, 45875200 padded size):
830
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight0
831
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight6
832
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight7
833
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight5
834
+ Params for bucket 19 (45875200 elements, 45875200 padded size):
835
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight3
836
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight1
837
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight4
838
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight0
839
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight2
840
+ Params for bucket 20 (55050240 elements, 55050240 padded size):
841
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight7
842
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight6
843
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight5
844
+ Params for bucket 21 (55050240 elements, 55050240 padded size):
845
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight4
846
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight3
847
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight2
848
+ Params for bucket 22 (45875200 elements, 45875200 padded size):
849
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight1
850
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight0
851
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight7
852
+ Params for bucket 23 (45875200 elements, 45875200 padded size):
853
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight6
854
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight4
855
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight2
856
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight5
857
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight3
858
+ Params for bucket 24 (55050240 elements, 55050240 padded size):
859
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight1
860
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight7
861
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight6
862
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight0
863
+ Params for bucket 25 (55050240 elements, 55050240 padded size):
864
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight5
865
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight4
866
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight3
867
+ Params for bucket 26 (55050240 elements, 55050240 padded size):
868
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight2
869
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight1
870
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight0
871
+ Params for bucket 27 (45875200 elements, 45875200 padded size):
872
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight7
873
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight3
874
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight5
875
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight4
876
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight6
877
+ Params for bucket 28 (45875200 elements, 45875200 padded size):
878
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight2
879
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight0
880
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight1
881
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight7
882
+ Params for bucket 29 (55050240 elements, 55050240 padded size):
883
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight6
884
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight4
885
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight5
886
+ Params for bucket 30 (55050240 elements, 55050240 padded size):
887
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight3
888
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight1
889
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight2
890
+ Params for bucket 31 (45875200 elements, 45875200 padded size):
891
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight0
892
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight5
893
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight6
894
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight7
895
+ Params for bucket 32 (45875200 elements, 45875200 padded size):
896
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight4
897
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight2
898
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight0
899
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight1
900
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight3
901
+ Params for bucket 33 (55050240 elements, 55050240 padded size):
902
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight7
903
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight5
904
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight6
905
+ Params for bucket 34 (55050240 elements, 55050240 padded size):
906
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight4
907
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight2
908
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight3
909
+ Params for bucket 35 (45875200 elements, 45875200 padded size):
910
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight1
911
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight0
912
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight7
913
+ Params for bucket 36 (45875200 elements, 45875200 padded size):
914
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight5
915
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight3
916
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight6
917
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight2
918
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight4
919
+ Params for bucket 37 (55050240 elements, 55050240 padded size):
920
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight1
921
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight7
922
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight0
923
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight6
924
+ Params for bucket 38 (55050240 elements, 55050240 padded size):
925
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight5
926
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight4
927
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight3
928
+ Params for bucket 39 (55050240 elements, 55050240 padded size):
929
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight2
930
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight1
931
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight0
932
+ Params for bucket 40 (45875200 elements, 45875200 padded size):
933
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight6
934
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight4
935
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight7
936
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight5
937
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight3
938
+ Params for bucket 41 (45875200 elements, 45875200 padded size):
939
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight2
940
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight0
941
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight1
942
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight7
943
+ Params for bucket 42 (55050240 elements, 55050240 padded size):
944
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight6
945
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight4
946
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight5
947
+ Params for bucket 43 (55050240 elements, 55050240 padded size):
948
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight3
949
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight1
950
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight2
951
+ Params for bucket 44 (18350080 elements, 18350080 padded size):
952
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight0
953
+ INFO:megatron.core.optimizer:Setting up optimizer with config OptimizerConfig(optimizer='adam', lr=0.0001, min_lr=1e-05, decoupled_lr=None, decoupled_min_lr=None, weight_decay=0.1, fp8_recipe='mxfp8', fp16=False, bf16=True, reuse_grad_buf_for_mxfp8_param_ag=False, params_dtype=torch.bfloat16, use_precision_aware_optimizer=False, store_param_remainders=True, main_grads_dtype=torch.float32, main_params_dtype=torch.float32, exp_avg_dtype=torch.float32, exp_avg_sq_dtype=torch.float32, loss_scale=None, initial_loss_scale=4294967296, min_loss_scale=1.0, loss_scale_window=1000, hysteresis=2, adam_beta1=0.9, adam_beta2=0.999, adam_eps=1e-08, sgd_momentum=0.9, use_distributed_optimizer=True, overlap_param_gather_with_optimizer_step=False, optimizer_cpu_offload=False, optimizer_offload_fraction=1.0, use_torch_optimizer_for_cpu_offload=False, overlap_cpu_optimizer_d2h_h2d=False, pin_cpu_grads=True, pin_cpu_params=True, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=<megatron.core.timers.Timers object at 0x78cc431a7470>, config_logger_dir='')
954
+ INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine
955
+ [after model, optimizer, and learning rate scheduler are built] datetime: 2026-08-26 18:46:22
956
+ > building train, validation, and test datasets ...
957
+ > datasets target sizes (minimum size):
958
+ train: 96
959
+ validation: 0
960
+ test: 0
961
+ INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None]
962
+ > building train, validation, and test datasets for GPT ...
963
+ INFO:megatron.core.datasets.blended_megatron_dataset_builder:Building GPTDataset splits with sizes=(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 0x78cc436b2db0>, mid_level_dataset_surplus=0.005, reset_position_ids=False, reset_attention_mask=False, eod_mask_loss=False, create_attention_mask=True, drop_last_partial_validation_sequence=True, add_extra_token_to_sequence=True, object_storage_cache_path=None)
964
+ INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx
965
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths
966
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers
967
+ INFO:megatron.core.datasets.indexed_dataset: Extract the document indices
968
+ INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029
969
+ INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350
970
+ INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices
971
+ INFO:megatron.core.datasets.gpt_dataset: Load the document index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-document_index.npy
972
+ INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-sample_index.npy
973
+ INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-shuffle_index.npy
974
+ INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073
975
+ > finished creating GPT datasets ...
976
+ [after dataloaders are built] datetime: 2026-08-26 18:46:22
977
+ done with setup ...
978
+ (min, max) time across ranks (ms):
979
+ model-and-optimizer-setup ......................: (2707.16, 2714.18)
980
+ train/valid/test-data-iterators-setup ..........: (61.58, 67.12)
981
+ training ...
982
+ Overwriting rerun_state_machine.current_iteration from -1 to 0...
983
+ [before the start of training step] datetime: 2026-08-26 18:46:22
984
+ INFO:moonep.api:MoonEP comm buffer sizes:
985
+ config : R=8, S=4096, H=3584, E=64, K=8, NvS=34800, NvS_padded=34816
986
+ sm_count : num_sms=32, num_sms_dedup=148
987
+ hidden_buffer_size : 238.00 MiB
988
+ meta_buffer_size : 2.00 MiB
989
+ local_temp_buffer_size : 1.26 MiB
990
+ total_size : 241.26 MiB
991
+ [2026-08-26 18:46:45] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 22845.5 | throughput per GPU (TFLOP/s/GPU): 5.4 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.097850E+01 | loss scale: 1.0 | grad norm: 29.821 | number of skipped iterations: 0 | number of nan iterations: 0 |
992
+ Number of parameters in transformer block in billions: 20.11
993
+ Number of parameters in embedding layers in billions: 0.36
994
+ Total number of parameters in billions: 20.47
995
+ Number of parameters in most loaded shard in billions: 20.4724
996
+ Theoretical memory footprints: weight and optimizer=146430.14 MB
997
+ [Rank 0] (after 1 iterations) memory (MB) | allocated: 105000.43212890625 | max allocated: 105000.4326171875 | reserved: 108632.0 | max reserved: 108632.0
998
+ [2026-08-26 18:46:46] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 350.8 | throughput per GPU (TFLOP/s/GPU): 352.0 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.097848E+01 | loss scale: 1.0 | grad norm: 30.569 | number of skipped iterations: 0 | number of nan iterations: 0 |
999
+ [2026-08-26 18:46:46] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 329.9 | throughput per GPU (TFLOP/s/GPU): 374.3 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 9.533927E+00 | loss scale: 1.0 | grad norm: 32.044 | number of skipped iterations: 0 | number of nan iterations: 0 |
1000
+ Capture CUDA graph for training!!!
1001
+ CUDA graph capture done!!!
1002
+ [2026-08-26 18:46:47] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 900.3 | throughput per GPU (TFLOP/s/GPU): 137.2 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 1.111380E+01 | loss scale: 1.0 | grad norm: 107.612 | number of skipped iterations: 0 | number of nan iterations: 0 |
1003
+ [2026-08-26 18:46:47] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 326.4 | throughput per GPU (TFLOP/s/GPU): 378.4 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 9.944556E+00 | loss scale: 1.0 | grad norm: 23.032 | number of skipped iterations: 0 | number of nan iterations: 0 |
1004
+ [2026-08-26 18:46:48] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 323.4 | throughput per GPU (TFLOP/s/GPU): 381.9 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 1.046731E+01 | loss scale: 1.0 | grad norm: 8.613 | number of skipped iterations: 0 | number of nan iterations: 0 |
1005
+ [2026-08-26 18:46:48] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 327.4 | throughput per GPU (TFLOP/s/GPU): 377.2 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 9.390327E+00 | loss scale: 1.0 | grad norm: 4.976 | number of skipped iterations: 0 | number of nan iterations: 0 |
1006
+ [2026-08-26 18:46:48] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 322.1 | throughput per GPU (TFLOP/s/GPU): 383.4 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.896387E+00 | loss scale: 1.0 | grad norm: 4.626 | number of skipped iterations: 0 | number of nan iterations: 0 |
1007
+ [2026-08-26 18:46:49] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 326.3 | throughput per GPU (TFLOP/s/GPU): 378.5 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.671444E+00 | loss scale: 1.0 | grad norm: 4.295 | number of skipped iterations: 0 | number of nan iterations: 0 |
1008
+ [2026-08-26 18:46:49] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 326.4 | throughput per GPU (TFLOP/s/GPU): 378.3 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.603101E+00 | loss scale: 1.0 | grad norm: 4.395 | number of skipped iterations: 0 | number of nan iterations: 0 |
1009
+ [2026-08-26 18:46:49] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 324.9 | throughput per GPU (TFLOP/s/GPU): 380.1 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.543194E+00 | loss scale: 1.0 | grad norm: 4.301 | number of skipped iterations: 0 | number of nan iterations: 0 |
1010
+ [2026-08-26 18:46:50] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 325.5 | throughput per GPU (TFLOP/s/GPU): 379.3 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.469308E+00 | loss scale: 1.0 | grad norm: 4.148 | number of skipped iterations: 0 | number of nan iterations: 0 |
1011
+ [after training is done] datetime: 2026-08-26 18:46:50
1012
+ [rank0]:[W826 18:46:51.398588069 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
1013
+ W0826 18:46:53.453000 554845 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers
1014
+ W0826 18:46:53.453000 554845 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 554939 closing signal SIGTERM
1015
+ W0826 18:46:53.454000 554845 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 554940 closing signal SIGTERM
1016
+ W0826 18:46:53.457000 554845 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 554941 closing signal SIGTERM
1017
+ W0826 18:46:53.459000 554845 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 554942 closing signal SIGTERM
1018
+ W0826 18:46:53.464000 554845 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 554943 closing signal SIGTERM
1019
+ W0826 18:46:53.467000 554845 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 554944 closing signal SIGTERM
1020
+ W0826 18:46:53.468000 554845 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 554945 closing signal SIGTERM
1021
+ W0826 18:46:53.471000 554845 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 554946 closing signal SIGTERM
1022
+ Traceback (most recent call last):
1023
+ File "/usr/local/bin/torchrun", line 8, in <module>
1024
+ sys.exit(main())
1025
+ ^^^^^^
1026
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
1027
+ return f(*args, **kwargs)
1028
+ ^^^^^^^^^^^^^^^^^^
1029
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 892, in main
1030
+ run(args)
1031
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 883, in run
1032
+ elastic_launch(
1033
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
1034
+ return launch_agent(self._config, self._entrypoint, list(args))
1035
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1036
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 261, in launch_agent
1037
+ result = agent.run()
1038
+ ^^^^^^^^^^^
1039
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper
1040
+ result = f(*args, **kwargs)
1041
+ ^^^^^^^^^^^^^^^^^^
1042
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 711, in run
1043
+ result = self._invoke_run(role)
1044
+ ^^^^^^^^^^^^^^^^^^^^^^
1045
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 870, in _invoke_run
1046
+ time.sleep(monitor_interval)
1047
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 84, in _terminate_process_handler
1048
+ raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval)
1049
+ torch.distributed.elastic.multiprocessing.api.SignalException: Process 554845 got signal: 15
moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep2/mem_samples.csv ADDED
@@ -0,0 +1,192 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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, 0
10
+ 1, 0
11
+ 2, 0
12
+ 3, 0
13
+ 4, 0
14
+ 5, 0
15
+ 6, 0
16
+ 7, 0
17
+ 0, 1
18
+ 1, 1
19
+ 2, 1
20
+ 3, 1
21
+ 4, 1
22
+ 5, 1
23
+ 6, 1
24
+ 7, 1
25
+ 0, 4
26
+ 1, 4
27
+ 2, 4
28
+ 3, 4
29
+ 4, 1
30
+ 5, 1
31
+ 6, 1
32
+ 7, 1
33
+ 0, 4
34
+ 1, 4
35
+ 2, 4
36
+ 3, 4
37
+ 4, 4
38
+ 5, 4
39
+ 6, 4
40
+ 7, 1
41
+ 0, 4
42
+ 1, 4
43
+ 2, 4
44
+ 3, 4
45
+ 4, 4
46
+ 5, 4
47
+ 6, 4
48
+ 7, 4
49
+ 0, 821
50
+ 1, 821
51
+ 2, 821
52
+ 3, 821
53
+ 4, 821
54
+ 5, 821
55
+ 6, 821
56
+ 7, 821
57
+ 0, 1803
58
+ 1, 1803
59
+ 2, 1803
60
+ 3, 1803
61
+ 4, 1803
62
+ 5, 1803
63
+ 6, 1803
64
+ 7, 1803
65
+ 0, 6705
66
+ 1, 6711
67
+ 2, 6703
68
+ 3, 6713
69
+ 4, 6699
70
+ 5, 6705
71
+ 6, 6699
72
+ 7, 6707
73
+ 0, 58469
74
+ 1, 57935
75
+ 2, 57947
76
+ 3, 58617
77
+ 4, 58653
78
+ 5, 58633
79
+ 6, 58617
80
+ 7, 58611
81
+ 0, 59497
82
+ 1, 58929
83
+ 2, 58929
84
+ 3, 59629
85
+ 4, 59629
86
+ 5, 59629
87
+ 6, 59629
88
+ 7, 59629
89
+ 0, 59737
90
+ 1, 59169
91
+ 2, 59169
92
+ 3, 59869
93
+ 4, 59869
94
+ 5, 59869
95
+ 6, 59869
96
+ 7, 59869
97
+ 0, 60937
98
+ 1, 60371
99
+ 2, 60383
100
+ 3, 61489
101
+ 4, 61415
102
+ 5, 60569
103
+ 6, 61379
104
+ 7, 61471
105
+ 0, 89207
106
+ 1, 88505
107
+ 2, 88991
108
+ 3, 90929
109
+ 4, 89211
110
+ 5, 89155
111
+ 6, 89155
112
+ 7, 90741
113
+ 0, 117903
114
+ 1, 117945
115
+ 2, 117927
116
+ 3, 117941
117
+ 4, 117941
118
+ 5, 117941
119
+ 6, 117941
120
+ 7, 117941
121
+ 0, 118939
122
+ 1, 118981
123
+ 2, 118963
124
+ 3, 118977
125
+ 4, 118977
126
+ 5, 118977
127
+ 6, 118977
128
+ 7, 118977
129
+ 0, 118939
130
+ 1, 118983
131
+ 2, 118965
132
+ 3, 118979
133
+ 4, 118979
134
+ 5, 118979
135
+ 6, 118979
136
+ 7, 118979
137
+ 0, 119599
138
+ 1, 119635
139
+ 2, 119623
140
+ 3, 119637
141
+ 4, 119637
142
+ 5, 119637
143
+ 6, 119637
144
+ 7, 119633
145
+ 0, 128715
146
+ 1, 128565
147
+ 2, 128519
148
+ 3, 128707
149
+ 4, 128621
150
+ 5, 128665
151
+ 6, 128613
152
+ 7, 128671
153
+ 0, 142739
154
+ 1, 142553
155
+ 2, 142535
156
+ 3, 142671
157
+ 4, 142671
158
+ 5, 142671
159
+ 6, 142671
160
+ 7, 142671
161
+ 0, 142739
162
+ 1, 142553
163
+ 2, 142535
164
+ 3, 142671
165
+ 4, 142671
166
+ 5, 142671
167
+ 6, 142671
168
+ 7, 142671
169
+ 0, 142499
170
+ 1, 142313
171
+ 2, 142295
172
+ 3, 142431
173
+ 4, 142431
174
+ 5, 142431
175
+ 6, 142431
176
+ 7, 142431
177
+ 0, 142499
178
+ 1, 142313
179
+ 2, 142295
180
+ 3, 142431
181
+ 4, 142431
182
+ 5, 142431
183
+ 6, 142431
184
+ 7, 142431
185
+ 0, 13764
186
+ 1, 142311
187
+ 2, 15044
188
+ 3, 13444
189
+ 4, 13444
190
+ 5, 13444
191
+ 6, 13444
192
+ 7, 13444
moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep2/run.env ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=moonep
2
+ config=QWEN2_57B
3
+ family=tierB
4
+ repo=/workspace/Megatron-MoE-Echo
5
+ branch=moonep_backend_bracing_regions_through_kernels
6
+ commit=1aac8bdd
7
+ num_layers=10
8
+ hidden=3584
9
+ ffn_hidden=2560
10
+ moe_ffn_hidden=2560
11
+ num_experts=64
12
+ topk=8
13
+ experts_per_rank=8
14
+ expert_params=27525120
15
+ heads=28
16
+ kv_groups=4
17
+ kv_channels=
18
+ qk_layernorm=0
19
+ shared_ffn=20480
20
+ shared_expert_overlap=0
21
+ mla=0
22
+ dense_first_layers=0
23
+ moe_layers=10
24
+ rotary_base=1000000
25
+ rotary_scaling=
26
+ qkv_bias=1
27
+ router_balancing=aux_loss
28
+ aux_loss_coeff=1e-2
29
+ router_pre_softmax=1
30
+ router_fusion=0
31
+ router_dtype=fp32
32
+ device_initiated_grouped_gemm=1
33
+ norm_epsilon=1e-6
34
+ seq_length=4096
35
+ micro_batch_size=1
36
+ global_batch_size=8
37
+ num_microbatches=1
38
+ assignments_per_rank=32768
39
+ mean_tokens_per_expert=4096
40
+ slot_policy=matched
41
+ slots_per_rank=8
42
+ capacity_mode=none
43
+ capacity=none
44
+ cuda_graph=1
45
+ probe_mode=0
46
+ async_grad_reduce=0
47
+ branch_role=timing
48
+ memory_dump_mode=none
49
+ memory_snapshot_path=
50
+ cuda_graph_warmup_steps=3
51
+ recompute=none
52
+ force_balance=0
53
+ train_iters=12
54
+ ep_size=8
55
+ echo_dispatch_overlap=1
56
+ echo_dispatch_recompute=0
57
+ ultraep_quota_min_tokens=
58
+ ultraep_weight_sync_join=
59
+ profile_enable=0
60
+ kernel_trace=0
61
+ fp8_recipe=mxfp8
moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep2/train.log ADDED
@@ -0,0 +1,1049 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ==============================================================
2
+ QWEN2_57B / moonep (slots/rank=8 policy=matched)
3
+ E=64 K=8 H'=2560 m=8
4
+ L=10 seq=4096 gbs=8 (1 microbatches)
5
+ cap_mode=none cap=none cg=1 recompute=none
6
+ memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep2/memory_snapshots/snapshot.pickle
7
+ repo=/workspace/Megatron-MoE-Echo @ 1aac8bdd
8
+ ==============================================================
9
+ W0826 18:47:00.967000 557198 torch/distributed/run.py:766]
10
+ W0826 18:47:00.967000 557198 torch/distributed/run.py:766] *****************************************
11
+ W0826 18:47:00.967000 557198 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
12
+ W0826 18:47:00.967000 557198 torch/distributed/run.py:766] *****************************************
13
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
14
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
15
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
16
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
17
+ 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
18
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
19
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
20
+ 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
21
+ WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer
22
+ Number of virtual stages per pipeline stage: None
23
+ accumulate and all-reduce gradients in fp32 for bfloat16 data type.
24
+ using torch.bfloat16 for parameters ...
25
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
26
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
27
+ ------------------------ arguments ------------------------
28
+ account_for_embedding_in_pipeline_split ......... False
29
+ account_for_loss_in_pipeline_split .............. False
30
+ accumulate_allreduce_grads_in_fp32 .............. True
31
+ adam_beta1 ...................................... 0.9
32
+ adam_beta2 ...................................... 0.999
33
+ adam_eps ........................................ 1e-08
34
+ add_bias_linear ................................. False
35
+ add_position_embedding .......................... False
36
+ add_qkv_bias .................................... True
37
+ adlr_autoresume ................................. False
38
+ adlr_autoresume_interval ........................ 1000
39
+ align_grad_reduce ............................... True
40
+ align_param_gather .............................. False
41
+ app_tag_run_name ................................ None
42
+ app_tag_run_version ............................. 0.0.0
43
+ apply_layernorm_1p .............................. False
44
+ apply_query_key_layer_scaling ................... False
45
+ apply_residual_connection_post_layernorm ........ False
46
+ apply_rope_fusion ............................... True
47
+ async_save ...................................... None
48
+ async_tensor_model_parallel_allreduce ........... True
49
+ attention_backend ............................... AttnBackend.auto
50
+ attention_dropout ............................... 0.0
51
+ attention_softmax_in_fp32 ....................... False
52
+ auto_detect_ckpt_format ......................... False
53
+ barrier_with_L1_time ............................ True
54
+ bert_binary_head ................................ True
55
+ bert_embedder_type .............................. megatron
56
+ bert_load ....................................... None
57
+ bf16 ............................................ True
58
+ bias_dropout_fusion ............................. True
59
+ bias_gelu_fusion ................................ False
60
+ bias_swiglu_fusion .............................. True
61
+ biencoder_projection_dim ........................ 0
62
+ biencoder_shared_query_context_model ............ False
63
+ block_data_path ................................. None
64
+ calc_ft_timeouts ................................ False
65
+ calculate_per_token_loss ........................ False
66
+ check_for_large_grads ........................... False
67
+ check_for_nan_in_loss_and_grad .................. False
68
+ check_for_spiky_loss ............................ False
69
+ check_weight_hash_across_dp_replicas_interval ... None
70
+ ckpt_assume_constant_structure .................. False
71
+ ckpt_convert_format ............................. None
72
+ ckpt_convert_save ............................... None
73
+ ckpt_convert_update_legacy_dist_opt_format ...... False
74
+ ckpt_format ..................................... torch_dist
75
+ ckpt_fully_parallel_load ........................ False
76
+ ckpt_fully_parallel_save ........................ True
77
+ ckpt_fully_parallel_save_deprecated ............. False
78
+ ckpt_step ....................................... None
79
+ classes_fraction ................................ 1.0
80
+ clip_grad ....................................... 1.0
81
+ clone_scatter_output_in_embedding ............... True
82
+ config_logger_dir ...............................
83
+ consumed_train_samples .......................... 0
84
+ consumed_valid_samples .......................... 0
85
+ context_parallel_size ........................... 1
86
+ cp_comm_type .................................... ['p2p']
87
+ create_attention_mask_in_dataloader ............. True
88
+ cross_entropy_fusion_impl ....................... native
89
+ cross_entropy_loss_fusion ....................... False
90
+ cuda_graph_scope ................................ full_iteration
91
+ cuda_graph_warmup_steps ......................... 3
92
+ data_args_path .................................. None
93
+ data_cache_path ................................. None
94
+ data_parallel_random_init ....................... False
95
+ data_parallel_sharding_strategy ................. no_shard
96
+ data_parallel_size .............................. 8
97
+ data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document']
98
+ data_per_class_fraction ......................... 1.0
99
+ data_sharding ................................... True
100
+ dataloader_type ................................. single
101
+ ddp_average_in_collective ....................... False
102
+ ddp_bucket_size ................................. None
103
+ ddp_num_buckets ................................. None
104
+ ddp_pad_buckets_for_high_nccl_busbw ............. False
105
+ decoder_first_pipeline_num_layers ............... None
106
+ decoder_last_pipeline_num_layers ................ None
107
+ decoder_num_layers .............................. None
108
+ decoder_seq_length .............................. None
109
+ decoupled_lr .................................... None
110
+ decoupled_min_lr ................................ None
111
+ decrease_batch_size_if_needed ................... False
112
+ defer_embedding_wgrad_compute ................... False
113
+ delay_wgrad_compute ............................. False
114
+ deprecated_use_mcore_models ..................... True
115
+ deterministic_mode .............................. False
116
+ dino_bottleneck_size ............................ 256
117
+ dino_freeze_last_layer .......................... 1
118
+ dino_head_hidden_size ........................... 2048
119
+ dino_local_crops_number ......................... 10
120
+ dino_local_img_size ............................. 96
121
+ dino_norm_last_layer ............................ False
122
+ dino_teacher_temp ............................... 0.07
123
+ dino_warmup_teacher_temp ........................ 0.04
124
+ dino_warmup_teacher_temp_epochs ................. 30
125
+ disable_bf16_reduced_precision_matmul ........... False
126
+ disable_mamba_mem_eff_path ...................... False
127
+ disable_straggler_on_startup .................... False
128
+ dist_ckpt_format_deprecated ..................... None
129
+ dist_ckpt_strictness ............................ assume_ok_unexpected
130
+ distribute_saved_activations .................... False
131
+ distributed_backend ............................. nccl
132
+ distributed_timeout_minutes ..................... 10
133
+ embedding_init_method_std ....................... None
134
+ embedding_path .................................. None
135
+ empty_unused_memory_level ....................... 0
136
+ enable_cuda_graph ............................... True
137
+ enable_experimental ............................. False
138
+ enable_ft_package ............................... False
139
+ enable_gloo_process_groups ...................... True
140
+ enable_msc ...................................... True
141
+ enable_one_logger ............................... True
142
+ encoder_num_layers .............................. 10
143
+ encoder_seq_length .............................. 4096
144
+ end_weight_decay ................................ 0.1
145
+ eod_mask_loss ................................... False
146
+ error_injection_rate ............................ 0
147
+ error_injection_type ............................ transient_error
148
+ eval_interval ................................... 1000000
149
+ eval_iters ...................................... 0
150
+ evidence_data_path .............................. None
151
+ exit_duration_in_mins ........................... None
152
+ exit_interval ................................... None
153
+ exit_on_missing_checkpoint ...................... False
154
+ exit_signal_handler ............................. False
155
+ exp_avg_dtype ................................... torch.float32
156
+ exp_avg_sq_dtype ................................ torch.float32
157
+ expert_model_parallel_size ...................... 8
158
+ expert_tensor_parallel_size ..................... 1
159
+ export_force_local_attention .................... False
160
+ export_kd_cfg ................................... None
161
+ export_kd_teacher_ckpt_format ................... None
162
+ export_kd_teacher_load .......................... None
163
+ export_kv_cache_quant ........................... False
164
+ export_legacy_megatron .......................... False
165
+ export_model_type ............................... GPTModel
166
+ export_moe_apply_probs_on_input ................. False
167
+ export_qk_l2_norm ............................... False
168
+ export_quant_cfg ................................ None
169
+ export_real_quant_cfg ........................... None
170
+ export_te_mcore_model ........................... False
171
+ external_cuda_graph ............................. False
172
+ ffn_hidden_size ................................. 2560
173
+ finetune ........................................ False
174
+ finetune_data_split ............................. train
175
+ finetune_hf_dataset ............................. None
176
+ first_last_layers_bf16 .......................... False
177
+ flash_decode .................................... False
178
+ fp16 ............................................ False
179
+ fp16_lm_cross_entropy ........................... False
180
+ fp32_residual_connection ........................ False
181
+ fp8 ............................................. e4m3
182
+ fp8_amax_compute_algo ........................... most_recent
183
+ fp8_amax_history_len ............................ 1
184
+ fp8_interval .................................... 1
185
+ fp8_margin ...................................... 0
186
+ fp8_param_gather ................................ False
187
+ fp8_recipe ...................................... mxfp8
188
+ fp8_wgrad ....................................... True
189
+ fsdp_double_buffer .............................. False
190
+ global_batch_size ............................... 8
191
+ grad_reduce_in_bf16 ............................. False
192
+ gradient_accumulation_fusion .................... True
193
+ gradient_reduce_div_fusion ...................... True
194
+ group_query_attention ........................... True
195
+ head_lr_mult .................................... 1.0
196
+ heterogeneous_layers_config_encoded_json ........ None
197
+ heterogeneous_layers_config_path ................ None
198
+ hidden_dropout .................................. 0.0
199
+ hidden_size ..................................... 3584
200
+ hierarchical_context_parallel_sizes ............. None
201
+ high_priority_stream_groups ..................... []
202
+ hybrid_attention_ratio .......................... 0.0
203
+ hybrid_mlp_ratio ................................ 0.0
204
+ hybrid_override_pattern ......................... None
205
+ hysteresis ...................................... 2
206
+ ict_head_size ................................... None
207
+ ict_load ........................................ None
208
+ img_h ........................................... 224
209
+ img_w ........................................... 224
210
+ indexer_batch_size .............................. 128
211
+ indexer_log_interval ............................ 1000
212
+ inference_batch_times_seqlen_threshold .......... -1
213
+ inference_dynamic_batching ...................... False
214
+ inference_dynamic_batching_buffer_guaranteed_fraction 0.2
215
+ inference_dynamic_batching_buffer_overflow_factor None
216
+ inference_dynamic_batching_buffer_size_gb ....... 40.0
217
+ inference_dynamic_batching_chunk_size ........... 256
218
+ inference_dynamic_batching_max_requests_override None
219
+ inference_dynamic_batching_max_tokens_override .. None
220
+ inference_dynamic_batching_num_cuda_graphs ...... 16
221
+ inference_max_batch_size ........................ 8
222
+ inference_max_seq_length ........................ 2560
223
+ inference_rng_tracker ........................... False
224
+ init_method_std ................................. 0.01
225
+ init_method_xavier_uniform ...................... False
226
+ init_model_with_meta_device ..................... False
227
+ initial_loss_scale .............................. 4294967296
228
+ inprocess_active_world_size ..................... 8
229
+ inprocess_barrier_timeout ....................... 120
230
+ inprocess_completion_timeout .................... 120
231
+ inprocess_empty_cuda_cache ...................... False
232
+ inprocess_granularity ........................... node
233
+ inprocess_hard_timeout .......................... 90
234
+ inprocess_heartbeat_interval .................... 30
235
+ inprocess_heartbeat_timeout ..................... 60
236
+ inprocess_last_call_wait ........................ 1
237
+ inprocess_max_iterations ........................ None
238
+ inprocess_monitor_process_interval .............. 1.0
239
+ inprocess_monitor_thread_interval ............... 1.0
240
+ inprocess_progress_watchdog_interval ............ 1.0
241
+ inprocess_restart ............................... False
242
+ inprocess_soft_timeout .......................... 60
243
+ inprocess_termination_grace_time ................ 1
244
+ is_hybrid_model ................................. False
245
+ iter_per_epoch .................................. 1250
246
+ iterations_to_skip .............................. []
247
+ keep_fp8_transpose_cache_when_using_custom_fsdp . False
248
+ kitchen_config_file ............................. None
249
+ kitchen_recipe_number ........................... None
250
+ kv_channels ..................................... 128
251
+ kv_lora_rank .................................... 32
252
+ lazy_mpu_init ................................... None
253
+ load ............................................ None
254
+ load_main_params_from_ckpt ...................... None
255
+ load_model_opt_format ........................... False
256
+ local_rank ...................................... 0
257
+ log_energy ...................................... False
258
+ log_interval .................................... 1
259
+ log_loss_scale_to_tensorboard ................... True
260
+ log_memory_to_tensorboard ....................... False
261
+ log_num_zeros_in_grad ........................... False
262
+ log_params_norm ................................. False
263
+ log_progress .................................... False
264
+ log_straggler ................................... False
265
+ log_throughput .................................. True
266
+ log_timers_to_tensorboard ....................... False
267
+ log_validation_ppl_to_tensorboard ............... False
268
+ log_world_size_to_tensorboard ................... False
269
+ logging_level ................................... None
270
+ loss_scale ...................................... None
271
+ loss_scale_window ............................... 1000
272
+ lr .............................................. 0.0001
273
+ lr_decay_iters .................................. 8
274
+ lr_decay_samples ................................ None
275
+ lr_decay_style .................................. cosine
276
+ lr_warmup_fraction .............................. None
277
+ lr_warmup_init .................................. 0.0
278
+ lr_warmup_iters ................................. 5
279
+ lr_warmup_samples ............................... 0
280
+ lr_wsd_decay_iters .............................. None
281
+ lr_wsd_decay_samples ............................ None
282
+ lr_wsd_decay_style .............................. exponential
283
+ main_grads_dtype ................................ torch.float32
284
+ main_params_dtype ............................... torch.float32
285
+ make_vocab_size_divisible_by .................... 128
286
+ mamba_head_dim .................................. 64
287
+ mamba_num_groups ................................ 8
288
+ mamba_num_heads ................................. None
289
+ mamba_state_dim ................................. 128
290
+ manual_gc ....................................... False
291
+ manual_gc_eval .................................. True
292
+ manual_gc_interval .............................. 0
293
+ mask_factor ..................................... 1.0
294
+ mask_prob ....................................... 0.15
295
+ mask_type ....................................... random
296
+ masked_softmax_fusion ........................... False
297
+ max_position_embeddings ......................... 32768
298
+ max_tokens_to_oom ............................... 12000
299
+ memory_snapshot_path ............................ snapshot.pickle
300
+ merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt
301
+ micro_batch_size ................................ 1
302
+ microbatch_group_size_per_vp_stage .............. None
303
+ mid_level_dataset_surplus ....................... 0.005
304
+ min_loss_scale .................................. 1.0
305
+ min_lr .......................................... 1e-05
306
+ mlp_chunks_for_prefill .......................... 1
307
+ mmap_bin_files .................................. True
308
+ mock_data ....................................... False
309
+ moe_apply_probs_on_input ........................ False
310
+ moe_aux_loss_coeff .............................. 0.01
311
+ moe_deepep_num_sms .............................. 20
312
+ moe_echo_enable_random_offloading ............... False
313
+ moe_echo_expert_dispatch_overlap ................ False
314
+ moe_echo_expert_dispatcher_type ................. hybridep
315
+ moe_echo_recompute_expert_dispatch .............. False
316
+ moe_enable_deepep ............................... False
317
+ moe_enable_echo ................................. False
318
+ moe_enable_hybridep ............................. False
319
+ moe_enable_moonep ............................... True
320
+ moe_expert_capacity_factor ...................... None
321
+ moe_extended_tp ................................. False
322
+ moe_ffn_hidden_size ............................. 2560
323
+ moe_grouped_gemm ................................ True
324
+ moe_hybridep_num_sms ............................ 16
325
+ moe_input_jitter_eps ............................ None
326
+ moe_layer_freq .................................. 1
327
+ moe_layer_recompute ............................. False
328
+ moe_moonep_num_sms .............................. None
329
+ moe_num_echo_experts ............................ None
330
+ moe_pad_expert_input_to_capacity ................ False
331
+ moe_per_layer_logging ........................... False
332
+ moe_permute_fusion .............................. False
333
+ moe_received_token_capacity ..................... None
334
+ moe_router_bias_update_rate ..................... 0.001
335
+ moe_router_dtype ................................ fp32
336
+ moe_router_enable_expert_bias ................... False
337
+ moe_router_force_load_balancing ................. False
338
+ moe_router_group_topk ........................... None
339
+ moe_router_load_balancing_type .................. aux_loss
340
+ moe_router_num_groups ........................... None
341
+ moe_router_padding_for_fp8 ...................... False
342
+ moe_router_pre_softmax .......................... True
343
+ moe_router_score_function ....................... softmax
344
+ moe_router_topk ................................. 8
345
+ moe_router_topk_scaling_factor .................. None
346
+ moe_shared_expert_intermediate_size ............. 20480
347
+ moe_shared_expert_overlap ....................... False
348
+ moe_token_dispatcher_type ....................... flex
349
+ moe_token_drop_policy ........................... probs
350
+ moe_upcycling_granularity ....................... 1
351
+ moe_use_device_initiated_grouped_gemm ........... True
352
+ moe_use_legacy_grouped_gemm ..................... False
353
+ moe_use_upcycling ............................... False
354
+ moe_z_loss_coeff ................................ None
355
+ mrope_section ................................... None
356
+ mscale .......................................... 1.0
357
+ mscale_all_dim .................................. 0.0
358
+ mtp_loss_scaling_factor ......................... 0.1
359
+ mtp_num_layers .................................. None
360
+ multi_latent_attention .......................... False
361
+ nccl_all_reduce_for_prefill ..................... False
362
+ nccl_communicator_config_path ................... None
363
+ nccl_ub ......................................... False
364
+ no_load_optim ................................... None
365
+ no_load_rng ..................................... None
366
+ no_persist_layer_norm ........................... False
367
+ no_rope_freq .................................... None
368
+ no_save_optim ................................... None
369
+ no_save_rng ..................................... None
370
+ non_persistent_ckpt_type ........................ None
371
+ non_persistent_global_ckpt_dir .................. None
372
+ non_persistent_local_ckpt_algo .................. fully_parallel
373
+ non_persistent_local_ckpt_dir ................... None
374
+ non_persistent_save_interval .................... None
375
+ norm_epsilon .................................... 1e-06
376
+ normalization ................................... RMSNorm
377
+ num_attention_heads ............................. 28
378
+ num_channels .................................... 3
379
+ num_classes ..................................... 1000
380
+ num_dataset_builder_threads ..................... 1
381
+ num_distributed_optimizer_instances ............. 1
382
+ num_experts ..................................... 64
383
+ num_layers ...................................... 10
384
+ num_layers_at_end_in_bf16 ....................... 1
385
+ num_layers_at_start_in_bf16 ..................... 1
386
+ num_layers_per_virtual_pipeline_stage ........... None
387
+ num_query_groups ................................ 4
388
+ num_virtual_stages_per_pipeline_rank ............ None
389
+ num_workers ..................................... 2
390
+ object_storage_cache_path ....................... None
391
+ one_logger_async ................................ False
392
+ one_logger_project .............................. megatron-lm
393
+ one_logger_run_name ............................. None
394
+ onnx_safe ....................................... None
395
+ openai_gelu ..................................... False
396
+ optimizer ....................................... adam
397
+ optimizer_cpu_offload ........................... False
398
+ optimizer_offload_fraction ...................... 1.0
399
+ output_bert_embeddings .......................... False
400
+ overlap_cpu_optimizer_d2h_h2d ................... False
401
+ overlap_grad_reduce ............................. True
402
+ overlap_moe_expert_parallel_comm ................ False
403
+ overlap_p2p_comm ................................ False
404
+ overlap_p2p_comm_warmup_flush ................... False
405
+ overlap_param_gather ............................ True
406
+ overlap_param_gather_with_optimizer_step ........ False
407
+ override_opt_param_scheduler .................... False
408
+ params_dtype .................................... torch.bfloat16
409
+ patch_dim ....................................... 16
410
+ per_split_data_args_path ........................ None
411
+ perform_initialization .......................... True
412
+ pin_cpu_grads ................................... True
413
+ pin_cpu_params .................................. True
414
+ pipeline_model_parallel_comm_backend ............ None
415
+ pipeline_model_parallel_layout .................. None
416
+ pipeline_model_parallel_size .................... 1
417
+ position_embedding_type ......................... rope
418
+ pretrained_checkpoint ........................... None
419
+ profile ......................................... False
420
+ profile_ranks ................................... [0]
421
+ profile_step_end ................................ 12
422
+ profile_step_start .............................. 10
423
+ q_lora_rank ..................................... None
424
+ qk_head_dim ..................................... 128
425
+ qk_l2_norm ...................................... False
426
+ qk_layernorm .................................... False
427
+ qk_pos_emb_head_dim ............................. 64
428
+ query_in_block_prob ............................. 0.1
429
+ rampup_batch_size ............................... None
430
+ rank ............................................ 0
431
+ recompute_granularity ........................... None
432
+ recompute_method ................................ None
433
+ recompute_modules ............................... None
434
+ recompute_num_layers ............................ None
435
+ record_memory_history ........................... False
436
+ relative_attention_max_distance ................. 128
437
+ relative_attention_num_buckets .................. 32
438
+ replication ..................................... False
439
+ replication_factor .............................. 2
440
+ replication_jump ................................ None
441
+ rerun_mode ...................................... validate_results
442
+ reset_attention_mask ............................ False
443
+ reset_position_ids .............................. False
444
+ result_rejected_tracker_filename ................ None
445
+ retriever_report_topk_accuracies ................ []
446
+ retriever_score_scaling ......................... False
447
+ retriever_seq_length ............................ 256
448
+ retro_add_retriever ............................. False
449
+ retro_attention_gate ............................ 1
450
+ retro_cyclic_train_iters ........................ None
451
+ retro_encoder_attention_dropout ................. 0.1
452
+ retro_encoder_hidden_dropout .................... 0.1
453
+ retro_encoder_layers ............................ 2
454
+ retro_num_neighbors ............................. 2
455
+ retro_num_retrieved_chunks ...................... 2
456
+ retro_project_dir ............................... None
457
+ retro_verify_neighbor_count ..................... True
458
+ reuse_grad_buf_for_mxfp8_param_ag ............... False
459
+ rope_scaling_factor ............................. 8.0
460
+ rope_type ....................................... None
461
+ rotary_base ..................................... 1000000
462
+ rotary_interleaved .............................. False
463
+ rotary_percent .................................. 1.0
464
+ rotary_scaling_factor ........................... 1.0
465
+ rotary_seq_len_interpolation_factor ............. None
466
+ run_workload_inspector_server ................... False
467
+ sample_rate ..................................... 1.0
468
+ save ............................................ None
469
+ save_interval ................................... None
470
+ save_retain_interval ............................ None
471
+ scatter_gather_tensors_in_pipeline .............. True
472
+ seed ............................................ 1234
473
+ seq_length ...................................... 4096
474
+ sequence_parallel ............................... False
475
+ sft ............................................. False
476
+ sft_tokenizer_prompt_format ..................... nemotron-h-aligned
477
+ sgd_momentum .................................... 0.9
478
+ short_seq_prob .................................. 0.1
479
+ skip_train ...................................... False
480
+ skipped_train_samples ........................... 0
481
+ spec ............................................ None
482
+ split ........................................... 100,0,0
483
+ squared_relu .................................... False
484
+ start_weight_decay .............................. 0.1
485
+ straggler_ctrlr_port ............................ 65535
486
+ straggler_minmax_count .......................... 1
487
+ suggested_communication_unit_size ............... None
488
+ swiglu .......................................... True
489
+ swin_backbone_type .............................. tiny
490
+ symmetric_ar_type ............................... None
491
+ te_rng_tracker .................................. True
492
+ tensor_model_parallel_size ...................... 1
493
+ tensorboard_dir ................................. None
494
+ tensorboard_log_interval ........................ 1
495
+ tensorboard_queue_size .......................... 1000
496
+ test_data_path .................................. None
497
+ test_mode ....................................... False
498
+ tiktoken_num_special_tokens ..................... 1000
499
+ tiktoken_pattern ................................ None
500
+ tiktoken_special_tokens ......................... None
501
+ timing_log_level ................................ 0
502
+ timing_log_option ............................... minmax
503
+ titles_data_path ................................ None
504
+ tokenizer_model ................................. None
505
+ tokenizer_type .................................. GPT2BPETokenizer
506
+ torch_fsdp2_reshard_after_forward ............... True
507
+ tp_comm_bootstrap_backend ....................... nccl
508
+ tp_comm_bulk_dgrad .............................. True
509
+ tp_comm_bulk_wgrad .............................. True
510
+ tp_comm_overlap ................................. False
511
+ tp_comm_overlap_ag .............................. True
512
+ tp_comm_overlap_cfg ............................. None
513
+ tp_comm_overlap_rs .............................. True
514
+ tp_comm_overlap_rs_dgrad ........................ False
515
+ tp_comm_split_ag ................................ True
516
+ tp_comm_split_rs ................................ True
517
+ train_data_path ................................. None
518
+ train_iters ..................................... 12
519
+ train_samples ................................... None
520
+ train_sync_interval ............................. None
521
+ transformer_impl ................................ transformer_engine
522
+ transformer_pipeline_model_parallel_size ........ 1
523
+ untie_embeddings_and_output_weights ............. True
524
+ use_checkpoint_args ............................. False
525
+ use_checkpoint_opt_param_scheduler .............. False
526
+ use_cpu_initialization .......................... None
527
+ use_custom_fsdp ................................. False
528
+ use_dist_ckpt ................................... True
529
+ use_dist_ckpt_deprecated ........................ False
530
+ use_distributed_optimizer ....................... True
531
+ use_flash_attn .................................. False
532
+ use_legacy_models ............................... False
533
+ use_mp_args_from_checkpoint_args ................ False
534
+ use_one_sent_docs ............................... False
535
+ use_persistent_ckpt_worker ...................... False
536
+ use_precision_aware_optimizer ................... False
537
+ use_pytorch_profiler ............................ False
538
+ use_ring_exchange_p2p ........................... False
539
+ use_rope_scaling ................................ False
540
+ use_rotary_position_embeddings .................. False
541
+ use_sharp ....................................... False
542
+ use_tokenizer_model_from_checkpoint_args ........ True
543
+ use_torch_fsdp2 ................................. False
544
+ use_torch_optimizer_for_cpu_offload ............. False
545
+ use_tp_pp_dp_mapping ............................ False
546
+ v_head_dim ...................................... 128
547
+ valid_data_path ................................. None
548
+ variable_seq_lengths ............................ False
549
+ virtual_pipeline_model_parallel_size ............ None
550
+ vision_backbone_type ............................ vit
551
+ vision_pretraining .............................. False
552
+ vision_pretraining_type ......................... classify
553
+ vocab_extra_ids ................................. 0
554
+ vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json
555
+ vocab_size ...................................... None
556
+ wandb_exp_name ..................................
557
+ wandb_project ...................................
558
+ wandb_save_dir ..................................
559
+ weight_decay .................................... 0.1
560
+ weight_decay_incr_style ......................... constant
561
+ wgrad_deferral_limit ............................ 0
562
+ world_size ...................................... 8
563
+ yaml_cfg ........................................ None
564
+ -------------------- end of arguments ---------------------
565
+ INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1
566
+ > building GPT2BPETokenizer tokenizer ...
567
+ > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304)
568
+ WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS
569
+ > initializing torch distributed ...
570
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
571
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
572
+ /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
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
579
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
580
+ [Gloo] Rank 6[Gloo] Rank is connected to 17 is connected to peer ranks. 7Expected number of connected peer ranks is : peer ranks. 7Expected number of connected peer ranks is : 7
581
+
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 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
585
+ [Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
586
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
587
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
588
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
589
+ [Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
590
+ [Gloo] Rank [Gloo] Rank 4 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 :
591
+ 7
592
+ [Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
593
+ [Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
594
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
595
+ > initialized tensor model parallel with size 1
596
+ > initialized pipeline model parallel with size 1
597
+ > setting random seeds to 1234 ...
598
+ [Gloo] Rank [Gloo] Rank 00 is connected to is connected to 00 peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : 00
599
+
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
+ [rank2]:[W826 18:47:19.794623131 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.
606
+ [rank7]:[W826 18:47:19.796537974 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.
607
+ [rank4]:[W826 18:47:19.796582622 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 4] using GPU 4 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
608
+ [rank5]:[W826 18:47:19.796620326 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.
609
+ [rank3]:[W826 18:47:19.796928559 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
610
+ [rank1]:[W826 18:47:19.797013198 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
+ [rank6]:[W826 18:47:19.797013410 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.
612
+ > compiling dataset index builder ...
613
+ make: Entering directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
614
+ make: Nothing to be done for 'default'.
615
+ make: Leaving directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
616
+ >>> done with dataset index builder. Compilation time: 0.034 seconds
617
+ WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations.
618
+ > compiling and loading fused kernels ...
619
+ [rank0]:[W826 18:47:19.972139520 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.
620
+ >>> done with compiling and loading fused kernels. Compilation time: 3.637 seconds
621
+ time to initialize megatron (seconds): 6.986
622
+ [after megatron is initialized] datetime: 2026-08-26 18:47:25
623
+ building GPT model ...
624
+ /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.
625
+ warnings.warn(
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
+ > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 5060614656
641
+ 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)
642
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 22
643
+ Params for bucket 1 (180289536 elements, 180289536 padded size):
644
+ module.output_layer.weight
645
+ Params for bucket 2 (73403904 elements, 73403904 padded size):
646
+ module.decoder.final_layernorm.weight
647
+ module.decoder.layers.9.mlp.shared_experts.linear_fc2.weight
648
+ Params for bucket 3 (146800640 elements, 146800640 padded size):
649
+ module.decoder.layers.9.mlp.shared_experts.linear_fc1.weight
650
+ Params for bucket 4 (103001600 elements, 103001600 padded size):
651
+ module.decoder.layers.9.mlp.router.weight
652
+ module.decoder.layers.9.self_attention.linear_qkv.weight
653
+ module.decoder.layers.9.pre_mlp_layernorm.weight
654
+ module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight
655
+ module.decoder.layers.9.self_attention.linear_qkv.bias
656
+ module.decoder.layers.9.self_attention.linear_proj.weight
657
+ module.decoder.layers.8.mlp.shared_experts.linear_fc2.weight
658
+ Params for bucket 5 (146800640 elements, 146800640 padded size):
659
+ module.decoder.layers.8.mlp.shared_experts.linear_fc1.weight
660
+ Params for bucket 6 (103001600 elements, 103001600 padded size):
661
+ module.decoder.layers.8.self_attention.linear_qkv.bias
662
+ module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight
663
+ module.decoder.layers.8.pre_mlp_layernorm.weight
664
+ module.decoder.layers.8.self_attention.linear_proj.weight
665
+ module.decoder.layers.7.mlp.shared_experts.linear_fc2.weight
666
+ module.decoder.layers.8.mlp.router.weight
667
+ module.decoder.layers.8.self_attention.linear_qkv.weight
668
+ Params for bucket 7 (146800640 elements, 146800640 padded size):
669
+ module.decoder.layers.7.mlp.shared_experts.linear_fc1.weight
670
+ Params for bucket 8 (103001600 elements, 103001600 padded size):
671
+ module.decoder.layers.7.self_attention.linear_qkv.bias
672
+ module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight
673
+ module.decoder.layers.7.pre_mlp_layernorm.weight
674
+ module.decoder.layers.7.mlp.router.weight
675
+ module.decoder.layers.7.self_attention.linear_qkv.weight
676
+ module.decoder.layers.7.self_attention.linear_proj.weight
677
+ module.decoder.layers.6.mlp.shared_experts.linear_fc2.weight
678
+ Params for bucket 9 (146800640 elements, 146800640 padded size):
679
+ module.decoder.layers.6.mlp.shared_experts.linear_fc1.weight
680
+ Params for bucket 10 (103001600 elements, 103001600 padded size):
681
+ module.decoder.layers.6.self_attention.linear_qkv.bias
682
+ module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight
683
+ module.decoder.layers.6.pre_mlp_layernorm.weight
684
+ module.decoder.layers.6.self_attention.linear_proj.weight
685
+ module.decoder.layers.5.mlp.shared_experts.linear_fc2.weight
686
+ module.decoder.layers.6.mlp.router.weight
687
+ module.decoder.layers.6.self_attention.linear_qkv.weight
688
+ Params for bucket 11 (146800640 elements, 146800640 padded size):
689
+ module.decoder.layers.5.mlp.shared_experts.linear_fc1.weight
690
+ Params for bucket 12 (103001600 elements, 103001600 padded size):
691
+ module.decoder.layers.5.mlp.router.weight
692
+ module.decoder.layers.5.self_attention.linear_qkv.weight
693
+ module.decoder.layers.5.pre_mlp_layernorm.weight
694
+ module.decoder.layers.5.self_attention.linear_qkv.bias
695
+ module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight
696
+ module.decoder.layers.5.self_attention.linear_proj.weight
697
+ module.decoder.layers.4.mlp.shared_experts.linear_fc2.weight
698
+ Params for bucket 13 (146800640 elements, 146800640 padded size):
699
+ module.decoder.layers.4.mlp.shared_experts.linear_fc1.weight
700
+ Params for bucket 14 (103001600 elements, 103001600 padded size):
701
+ module.decoder.layers.4.mlp.router.weight
702
+ module.decoder.layers.4.self_attention.linear_qkv.weight
703
+ module.decoder.layers.4.pre_mlp_layernorm.weight
704
+ module.decoder.layers.4.self_attention.linear_qkv.bias
705
+ module.decoder.layers.4.self_attention.linear_qkv.layer_norm_weight
706
+ module.decoder.layers.4.self_attention.linear_proj.weight
707
+ module.decoder.layers.3.mlp.shared_experts.linear_fc2.weight
708
+ Params for bucket 15 (146800640 elements, 146800640 padded size):
709
+ module.decoder.layers.3.mlp.shared_experts.linear_fc1.weight
710
+ Params for bucket 16 (103001600 elements, 103001600 padded size):
711
+ module.decoder.layers.3.pre_mlp_layernorm.weight
712
+ module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight
713
+ module.decoder.layers.3.self_attention.linear_qkv.bias
714
+ module.decoder.layers.3.mlp.router.weight
715
+ module.decoder.layers.3.self_attention.linear_qkv.weight
716
+ module.decoder.layers.3.self_attention.linear_proj.weight
717
+ module.decoder.layers.2.mlp.shared_experts.linear_fc2.weight
718
+ Params for bucket 17 (146800640 elements, 146800640 padded size):
719
+ module.decoder.layers.2.mlp.shared_experts.linear_fc1.weight
720
+ Params for bucket 18 (103001600 elements, 103001600 padded size):
721
+ module.decoder.layers.2.mlp.router.weight
722
+ module.decoder.layers.2.self_attention.linear_qkv.weight
723
+ module.decoder.layers.2.self_attention.linear_proj.weight
724
+ module.decoder.layers.2.self_attention.linear_qkv.bias
725
+ module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight
726
+ module.decoder.layers.2.pre_mlp_layernorm.weight
727
+ module.decoder.layers.1.mlp.shared_experts.linear_fc2.weight
728
+ Params for bucket 19 (146800640 elements, 146800640 padded size):
729
+ module.decoder.layers.1.mlp.shared_experts.linear_fc1.weight
730
+ Params for bucket 20 (103001600 elements, 103001600 padded size):
731
+ module.decoder.layers.1.mlp.router.weight
732
+ module.decoder.layers.1.self_attention.linear_qkv.weight
733
+ module.decoder.layers.1.self_attention.linear_proj.weight
734
+ module.decoder.layers.0.mlp.shared_experts.linear_fc2.weight
735
+ module.decoder.layers.1.pre_mlp_layernorm.weight
736
+ module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight
737
+ module.decoder.layers.1.self_attention.linear_qkv.bias
738
+ Params for bucket 21 (146800640 elements, 146800640 padded size):
739
+ module.decoder.layers.0.mlp.shared_experts.linear_fc1.weight
740
+ Params for bucket 22 (209890816 elements, 209890816 padded size):
741
+ module.decoder.layers.0.mlp.router.weight
742
+ module.decoder.layers.0.self_attention.linear_qkv.weight
743
+ module.decoder.layers.0.self_attention.linear_qkv.bias
744
+ module.embedding.word_embeddings.weight
745
+ module.decoder.layers.0.self_attention.linear_qkv.layer_norm_weight
746
+ module.decoder.layers.0.pre_mlp_layernorm.weight
747
+ module.decoder.layers.0.self_attention.linear_proj.weight
748
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 44
749
+ Params for bucket 1 (45875200 elements, 45875200 padded size):
750
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight7
751
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight3
752
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight5
753
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight6
754
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight4
755
+ Params for bucket 2 (45875200 elements, 45875200 padded size):
756
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight2
757
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight0
758
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight7
759
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight1
760
+ Params for bucket 3 (55050240 elements, 55050240 padded size):
761
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight6
762
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight4
763
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight5
764
+ Params for bucket 4 (55050240 elements, 55050240 padded size):
765
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight3
766
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight1
767
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight2
768
+ Params for bucket 5 (45875200 elements, 45875200 padded size):
769
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight0
770
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight6
771
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight7
772
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight5
773
+ Params for bucket 6 (45875200 elements, 45875200 padded size):
774
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight3
775
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight1
776
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight4
777
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight0
778
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight2
779
+ Params for bucket 7 (55050240 elements, 55050240 padded size):
780
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight7
781
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight5
782
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight6
783
+ Params for bucket 8 (55050240 elements, 55050240 padded size):
784
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight4
785
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight2
786
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight3
787
+ Params for bucket 9 (45875200 elements, 45875200 padded size):
788
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight1
789
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight7
790
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight0
791
+ Params for bucket 10 (45875200 elements, 45875200 padded size):
792
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight6
793
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight2
794
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight4
795
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight5
796
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight3
797
+ Params for bucket 11 (55050240 elements, 55050240 padded size):
798
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight1
799
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight7
800
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight0
801
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight6
802
+ Params for bucket 12 (55050240 elements, 55050240 padded size):
803
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight5
804
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight4
805
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight3
806
+ Params for bucket 13 (55050240 elements, 55050240 padded size):
807
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight2
808
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight1
809
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight0
810
+ Params for bucket 14 (45875200 elements, 45875200 padded size):
811
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight7
812
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight3
813
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight5
814
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight4
815
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight6
816
+ Params for bucket 15 (45875200 elements, 45875200 padded size):
817
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight2
818
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight0
819
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight1
820
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight7
821
+ Params for bucket 16 (55050240 elements, 55050240 padded size):
822
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight6
823
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight4
824
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight5
825
+ Params for bucket 17 (55050240 elements, 55050240 padded size):
826
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight3
827
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight1
828
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight2
829
+ Params for bucket 18 (45875200 elements, 45875200 padded size):
830
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight0
831
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight5
832
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight6
833
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight7
834
+ Params for bucket 19 (45875200 elements, 45875200 padded size):
835
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight4
836
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight2
837
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight0
838
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight1
839
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight3
840
+ Params for bucket 20 (55050240 elements, 55050240 padded size):
841
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight7
842
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight6
843
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight5
844
+ Params for bucket 21 (55050240 elements, 55050240 padded size):
845
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight4
846
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight3
847
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight2
848
+ Params for bucket 22 (45875200 elements, 45875200 padded size):
849
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight1
850
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight0
851
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight7
852
+ Params for bucket 23 (45875200 elements, 45875200 padded size):
853
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight3
854
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight5
855
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight6
856
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight4
857
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight2
858
+ Params for bucket 24 (55050240 elements, 55050240 padded size):
859
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight1
860
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight0
861
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight6
862
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight7
863
+ Params for bucket 25 (55050240 elements, 55050240 padded size):
864
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight5
865
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight3
866
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight4
867
+ Params for bucket 26 (55050240 elements, 55050240 padded size):
868
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight2
869
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight0
870
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight1
871
+ Params for bucket 27 (45875200 elements, 45875200 padded size):
872
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight7
873
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight3
874
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight5
875
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight4
876
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight6
877
+ Params for bucket 28 (45875200 elements, 45875200 padded size):
878
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight2
879
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight0
880
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight1
881
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight7
882
+ Params for bucket 29 (55050240 elements, 55050240 padded size):
883
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight6
884
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight5
885
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight4
886
+ Params for bucket 30 (55050240 elements, 55050240 padded size):
887
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight3
888
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight2
889
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight1
890
+ Params for bucket 31 (45875200 elements, 45875200 padded size):
891
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight0
892
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight5
893
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight7
894
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight6
895
+ Params for bucket 32 (45875200 elements, 45875200 padded size):
896
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight4
897
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight2
898
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight0
899
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight1
900
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight3
901
+ Params for bucket 33 (55050240 elements, 55050240 padded size):
902
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight7
903
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight5
904
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight6
905
+ Params for bucket 34 (55050240 elements, 55050240 padded size):
906
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight4
907
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight2
908
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight3
909
+ Params for bucket 35 (45875200 elements, 45875200 padded size):
910
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight1
911
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight7
912
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight0
913
+ Params for bucket 36 (45875200 elements, 45875200 padded size):
914
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight5
915
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight3
916
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight6
917
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight4
918
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight2
919
+ Params for bucket 37 (55050240 elements, 55050240 padded size):
920
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight1
921
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight7
922
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight6
923
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight0
924
+ Params for bucket 38 (55050240 elements, 55050240 padded size):
925
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight5
926
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight4
927
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight3
928
+ Params for bucket 39 (55050240 elements, 55050240 padded size):
929
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight2
930
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight1
931
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight0
932
+ Params for bucket 40 (45875200 elements, 45875200 padded size):
933
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight7
934
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight3
935
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight5
936
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight6
937
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight4
938
+ Params for bucket 41 (45875200 elements, 45875200 padded size):
939
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight2
940
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight0
941
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight1
942
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight7
943
+ Params for bucket 42 (55050240 elements, 55050240 padded size):
944
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight6
945
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight5
946
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight4
947
+ Params for bucket 43 (55050240 elements, 55050240 padded size):
948
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight3
949
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight2
950
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight1
951
+ Params for bucket 44 (18350080 elements, 18350080 padded size):
952
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight0
953
+ INFO:megatron.core.optimizer:Setting up optimizer with config OptimizerConfig(optimizer='adam', lr=0.0001, min_lr=1e-05, decoupled_lr=None, decoupled_min_lr=None, weight_decay=0.1, fp8_recipe='mxfp8', fp16=False, bf16=True, reuse_grad_buf_for_mxfp8_param_ag=False, params_dtype=torch.bfloat16, use_precision_aware_optimizer=False, store_param_remainders=True, main_grads_dtype=torch.float32, main_params_dtype=torch.float32, exp_avg_dtype=torch.float32, exp_avg_sq_dtype=torch.float32, loss_scale=None, initial_loss_scale=4294967296, min_loss_scale=1.0, loss_scale_window=1000, hysteresis=2, adam_beta1=0.9, adam_beta2=0.999, adam_eps=1e-08, sgd_momentum=0.9, use_distributed_optimizer=True, overlap_param_gather_with_optimizer_step=False, optimizer_cpu_offload=False, optimizer_offload_fraction=1.0, use_torch_optimizer_for_cpu_offload=False, overlap_cpu_optimizer_d2h_h2d=False, pin_cpu_grads=True, pin_cpu_params=True, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=<megatron.core.timers.Timers object at 0x797d62a63290>, config_logger_dir='')
954
+ INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine
955
+ [after model, optimizer, and learning rate scheduler are built] datetime: 2026-08-26 18:47:28
956
+ > building train, validation, and test datasets ...
957
+ > datasets target sizes (minimum size):
958
+ train: 96
959
+ validation: 0
960
+ test: 0
961
+ INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None]
962
+ > building train, validation, and test datasets for GPT ...
963
+ INFO:megatron.core.datasets.blended_megatron_dataset_builder:Building GPTDataset splits with sizes=(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 0x797d6310adb0>, mid_level_dataset_surplus=0.005, reset_position_ids=False, reset_attention_mask=False, eod_mask_loss=False, create_attention_mask=True, drop_last_partial_validation_sequence=True, add_extra_token_to_sequence=True, object_storage_cache_path=None)
964
+ INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx
965
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths
966
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers
967
+ INFO:megatron.core.datasets.indexed_dataset: Extract the document indices
968
+ INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029
969
+ INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350
970
+ INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices
971
+ INFO:megatron.core.datasets.gpt_dataset: Load the document index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-document_index.npy
972
+ INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-sample_index.npy
973
+ INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-shuffle_index.npy
974
+ INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073
975
+ > finished creating GPT datasets ...
976
+ [after dataloaders are built] datetime: 2026-08-26 18:47:28
977
+ done with setup ...
978
+ (min, max) time across ranks (ms):
979
+ model-and-optimizer-setup ......................: (2534.13, 2539.73)
980
+ train/valid/test-data-iterators-setup ..........: (61.58, 75.18)
981
+ training ...
982
+ Overwriting rerun_state_machine.current_iteration from -1 to 0...
983
+ [before the start of training step] datetime: 2026-08-26 18:47:28
984
+ INFO:moonep.api:MoonEP comm buffer sizes:
985
+ config : R=8, S=4096, H=3584, E=64, K=8, NvS=34800, NvS_padded=34816
986
+ sm_count : num_sms=32, num_sms_dedup=148
987
+ hidden_buffer_size : 238.00 MiB
988
+ meta_buffer_size : 2.00 MiB
989
+ local_temp_buffer_size : 1.26 MiB
990
+ total_size : 241.26 MiB
991
+ [2026-08-26 18:47:51] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 22730.5 | throughput per GPU (TFLOP/s/GPU): 5.4 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.097850E+01 | loss scale: 1.0 | grad norm: 29.821 | number of skipped iterations: 0 | number of nan iterations: 0 |
992
+ Number of parameters in transformer block in billions: 20.11
993
+ Number of parameters in embedding layers in billions: 0.36
994
+ Total number of parameters in billions: 20.47
995
+ Number of parameters in most loaded shard in billions: 20.4724
996
+ Theoretical memory footprints: weight and optimizer=146430.14 MB
997
+ [Rank 0] (after 1 iterations) memory (MB) | allocated: 105000.43212890625 | max allocated: 105000.4326171875 | reserved: 108632.0 | max reserved: 108632.0
998
+ [2026-08-26 18:47:51] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 361.8 | throughput per GPU (TFLOP/s/GPU): 341.3 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.097848E+01 | loss scale: 1.0 | grad norm: 30.565 | number of skipped iterations: 0 | number of nan iterations: 0 |
999
+ [2026-08-26 18:47:51] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 326.5 | throughput per GPU (TFLOP/s/GPU): 378.2 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 9.532912E+00 | loss scale: 1.0 | grad norm: 31.962 | number of skipped iterations: 0 | number of nan iterations: 0 |
1000
+ Capture CUDA graph for training!!!
1001
+ CUDA graph capture done!!!
1002
+ [2026-08-26 18:47:52] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 924.6 | throughput per GPU (TFLOP/s/GPU): 133.6 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 1.111276E+01 | loss scale: 1.0 | grad norm: 107.584 | number of skipped iterations: 0 | number of nan iterations: 0 |
1003
+ [2026-08-26 18:47:53] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 325.6 | throughput per GPU (TFLOP/s/GPU): 379.3 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 9.938559E+00 | loss scale: 1.0 | grad norm: 22.707 | number of skipped iterations: 0 | number of nan iterations: 0 |
1004
+ [2026-08-26 18:47:53] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 325.4 | throughput per GPU (TFLOP/s/GPU): 379.5 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 1.046246E+01 | loss scale: 1.0 | grad norm: 8.596 | number of skipped iterations: 0 | number of nan iterations: 0 |
1005
+ [2026-08-26 18:47:53] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 325.9 | throughput per GPU (TFLOP/s/GPU): 378.9 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 9.384308E+00 | loss scale: 1.0 | grad norm: 4.989 | number of skipped iterations: 0 | number of nan iterations: 0 |
1006
+ [2026-08-26 18:47:54] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 324.2 | throughput per GPU (TFLOP/s/GPU): 380.9 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.891587E+00 | loss scale: 1.0 | grad norm: 4.679 | number of skipped iterations: 0 | number of nan iterations: 0 |
1007
+ [2026-08-26 18:47:54] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 324.8 | throughput per GPU (TFLOP/s/GPU): 380.2 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.666572E+00 | loss scale: 1.0 | grad norm: 4.308 | number of skipped iterations: 0 | number of nan iterations: 0 |
1008
+ [2026-08-26 18:47:54] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 325.9 | throughput per GPU (TFLOP/s/GPU): 378.9 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.597811E+00 | loss scale: 1.0 | grad norm: 4.389 | number of skipped iterations: 0 | number of nan iterations: 0 |
1009
+ [2026-08-26 18:47:55] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 322.1 | throughput per GPU (TFLOP/s/GPU): 383.3 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.538117E+00 | loss scale: 1.0 | grad norm: 4.292 | number of skipped iterations: 0 | number of nan iterations: 0 |
1010
+ [2026-08-26 18:47:55] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 326.8 | throughput per GPU (TFLOP/s/GPU): 377.9 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.464961E+00 | loss scale: 1.0 | grad norm: 4.135 | number of skipped iterations: 0 | number of nan iterations: 0 |
1011
+ [after training is done] datetime: 2026-08-26 18:47:55
1012
+ [rank0]:[W826 18:47:56.866983430 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
1013
+ W0826 18:47:59.643000 557198 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers
1014
+ W0826 18:47:59.644000 557198 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 557286 closing signal SIGTERM
1015
+ W0826 18:47:59.645000 557198 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 557287 closing signal SIGTERM
1016
+ W0826 18:47:59.648000 557198 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 557288 closing signal SIGTERM
1017
+ W0826 18:47:59.651000 557198 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 557289 closing signal SIGTERM
1018
+ W0826 18:47:59.652000 557198 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 557290 closing signal SIGTERM
1019
+ W0826 18:47:59.655000 557198 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 557291 closing signal SIGTERM
1020
+ W0826 18:47:59.658000 557198 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 557292 closing signal SIGTERM
1021
+ W0826 18:47:59.659000 557198 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 557293 closing signal SIGTERM
1022
+ Traceback (most recent call last):
1023
+ File "/usr/local/bin/torchrun", line 8, in <module>
1024
+ sys.exit(main())
1025
+ ^^^^^^
1026
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
1027
+ return f(*args, **kwargs)
1028
+ ^^^^^^^^^^^^^^^^^^
1029
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 892, in main
1030
+ run(args)
1031
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 883, in run
1032
+ elastic_launch(
1033
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
1034
+ return launch_agent(self._config, self._entrypoint, list(args))
1035
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1036
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 261, in launch_agent
1037
+ result = agent.run()
1038
+ ^^^^^^^^^^^
1039
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper
1040
+ result = f(*args, **kwargs)
1041
+ ^^^^^^^^^^^^^^^^^^
1042
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 711, in run
1043
+ result = self._invoke_run(role)
1044
+ ^^^^^^^^^^^^^^^^^^^^^^
1045
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 870, in _invoke_run
1046
+ time.sleep(monitor_interval)
1047
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 84, in _terminate_process_handler
1048
+ raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval)
1049
+ torch.distributed.elastic.multiprocessing.api.SignalException: Process 557198 got signal: 15
moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep3/mem_samples.csv ADDED
@@ -0,0 +1,192 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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, 0
10
+ 1, 0
11
+ 2, 0
12
+ 3, 0
13
+ 4, 0
14
+ 5, 0
15
+ 6, 0
16
+ 7, 0
17
+ 0, 1
18
+ 1, 1
19
+ 2, 1
20
+ 3, 1
21
+ 4, 1
22
+ 5, 1
23
+ 6, 1
24
+ 7, 1
25
+ 0, 4
26
+ 1, 4
27
+ 2, 4
28
+ 3, 4
29
+ 4, 1
30
+ 5, 1
31
+ 6, 1
32
+ 7, 1
33
+ 0, 4
34
+ 1, 4
35
+ 2, 4
36
+ 3, 4
37
+ 4, 4
38
+ 5, 4
39
+ 6, 4
40
+ 7, 4
41
+ 0, 781
42
+ 1, 737
43
+ 2, 735
44
+ 3, 733
45
+ 4, 737
46
+ 5, 757
47
+ 6, 741
48
+ 7, 771
49
+ 0, 1537
50
+ 1, 1537
51
+ 2, 1537
52
+ 3, 1537
53
+ 4, 1537
54
+ 5, 1537
55
+ 6, 1537
56
+ 7, 1537
57
+ 0, 1863
58
+ 1, 1863
59
+ 2, 1863
60
+ 3, 1863
61
+ 4, 1863
62
+ 5, 1863
63
+ 6, 1863
64
+ 7, 1863
65
+ 0, 6895
66
+ 1, 6895
67
+ 2, 6895
68
+ 3, 6895
69
+ 4, 6895
70
+ 5, 6895
71
+ 6, 6895
72
+ 7, 6895
73
+ 0, 58437
74
+ 1, 57853
75
+ 2, 57933
76
+ 3, 58587
77
+ 4, 58607
78
+ 5, 58621
79
+ 6, 58619
80
+ 7, 58591
81
+ 0, 59113
82
+ 1, 58545
83
+ 2, 58545
84
+ 3, 59245
85
+ 4, 59245
86
+ 5, 59245
87
+ 6, 59245
88
+ 7, 59245
89
+ 0, 59737
90
+ 1, 59169
91
+ 2, 59169
92
+ 3, 59869
93
+ 4, 59869
94
+ 5, 59869
95
+ 6, 59869
96
+ 7, 59869
97
+ 0, 60439
98
+ 1, 59869
99
+ 2, 59869
100
+ 3, 60569
101
+ 4, 60569
102
+ 5, 60569
103
+ 6, 60569
104
+ 7, 60569
105
+ 0, 85061
106
+ 1, 86591
107
+ 2, 86983
108
+ 3, 87641
109
+ 4, 90313
110
+ 5, 86297
111
+ 6, 92149
112
+ 7, 88519
113
+ 0, 117993
114
+ 1, 118045
115
+ 2, 118017
116
+ 3, 118013
117
+ 4, 118017
118
+ 5, 118029
119
+ 6, 118027
120
+ 7, 118029
121
+ 0, 118939
122
+ 1, 118981
123
+ 2, 118963
124
+ 3, 118977
125
+ 4, 118977
126
+ 5, 118977
127
+ 6, 118977
128
+ 7, 118977
129
+ 0, 118939
130
+ 1, 118983
131
+ 2, 118965
132
+ 3, 118979
133
+ 4, 118979
134
+ 5, 118979
135
+ 6, 118979
136
+ 7, 118979
137
+ 0, 119215
138
+ 1, 119257
139
+ 2, 119239
140
+ 3, 119253
141
+ 4, 119253
142
+ 5, 119253
143
+ 6, 119253
144
+ 7, 119253
145
+ 0, 128667
146
+ 1, 128457
147
+ 2, 128437
148
+ 3, 128615
149
+ 4, 128583
150
+ 5, 128585
151
+ 6, 128595
152
+ 7, 128611
153
+ 0, 142739
154
+ 1, 142553
155
+ 2, 142535
156
+ 3, 142671
157
+ 4, 142671
158
+ 5, 142671
159
+ 6, 142671
160
+ 7, 142671
161
+ 0, 142739
162
+ 1, 142553
163
+ 2, 142535
164
+ 3, 142671
165
+ 4, 142671
166
+ 5, 142671
167
+ 6, 142671
168
+ 7, 142671
169
+ 0, 142499
170
+ 1, 142313
171
+ 2, 142295
172
+ 3, 142431
173
+ 4, 142431
174
+ 5, 142431
175
+ 6, 142431
176
+ 7, 142431
177
+ 0, 142499
178
+ 1, 142313
179
+ 2, 142295
180
+ 3, 142431
181
+ 4, 142431
182
+ 5, 142431
183
+ 6, 142431
184
+ 7, 142431
185
+ 0, 13444
186
+ 1, 13444
187
+ 2, 13444
188
+ 3, 13764
189
+ 4, 142429
190
+ 5, 15044
191
+ 6, 13444
192
+ 7, 13444
moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep3/run.env ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=moonep
2
+ config=QWEN2_57B
3
+ family=tierB
4
+ repo=/workspace/Megatron-MoE-Echo
5
+ branch=moonep_backend_bracing_regions_through_kernels
6
+ commit=1aac8bdd
7
+ num_layers=10
8
+ hidden=3584
9
+ ffn_hidden=2560
10
+ moe_ffn_hidden=2560
11
+ num_experts=64
12
+ topk=8
13
+ experts_per_rank=8
14
+ expert_params=27525120
15
+ heads=28
16
+ kv_groups=4
17
+ kv_channels=
18
+ qk_layernorm=0
19
+ shared_ffn=20480
20
+ shared_expert_overlap=0
21
+ mla=0
22
+ dense_first_layers=0
23
+ moe_layers=10
24
+ rotary_base=1000000
25
+ rotary_scaling=
26
+ qkv_bias=1
27
+ router_balancing=aux_loss
28
+ aux_loss_coeff=1e-2
29
+ router_pre_softmax=1
30
+ router_fusion=0
31
+ router_dtype=fp32
32
+ device_initiated_grouped_gemm=1
33
+ norm_epsilon=1e-6
34
+ seq_length=4096
35
+ micro_batch_size=1
36
+ global_batch_size=8
37
+ num_microbatches=1
38
+ assignments_per_rank=32768
39
+ mean_tokens_per_expert=4096
40
+ slot_policy=matched
41
+ slots_per_rank=8
42
+ capacity_mode=none
43
+ capacity=none
44
+ cuda_graph=1
45
+ probe_mode=0
46
+ async_grad_reduce=0
47
+ branch_role=timing
48
+ memory_dump_mode=none
49
+ memory_snapshot_path=
50
+ cuda_graph_warmup_steps=3
51
+ recompute=none
52
+ force_balance=0
53
+ train_iters=12
54
+ ep_size=8
55
+ echo_dispatch_overlap=1
56
+ echo_dispatch_recompute=0
57
+ ultraep_quota_min_tokens=
58
+ ultraep_weight_sync_join=
59
+ profile_enable=0
60
+ kernel_trace=0
61
+ fp8_recipe=mxfp8
moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep3/train.log ADDED
@@ -0,0 +1,1049 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ==============================================================
2
+ QWEN2_57B / moonep (slots/rank=8 policy=matched)
3
+ E=64 K=8 H'=2560 m=8
4
+ L=10 seq=4096 gbs=8 (1 microbatches)
5
+ cap_mode=none cap=none cg=1 recompute=none
6
+ memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep3/memory_snapshots/snapshot.pickle
7
+ repo=/workspace/Megatron-MoE-Echo @ 1aac8bdd
8
+ ==============================================================
9
+ W0826 18:48:07.155000 559545 torch/distributed/run.py:766]
10
+ W0826 18:48:07.155000 559545 torch/distributed/run.py:766] *****************************************
11
+ W0826 18:48:07.155000 559545 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
12
+ W0826 18:48:07.155000 559545 torch/distributed/run.py:766] *****************************************
13
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
14
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
15
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
16
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
17
+ 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
18
+ WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer
19
+ Number of virtual stages per pipeline stage: None
20
+ accumulate and all-reduce gradients in fp32 for bfloat16 data type.
21
+ using torch.bfloat16 for parameters ...
22
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
23
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
24
+ ------------------------ arguments ------------------------
25
+ account_for_embedding_in_pipeline_split ......... False
26
+ account_for_loss_in_pipeline_split .............. False
27
+ accumulate_allreduce_grads_in_fp32 .............. True
28
+ adam_beta1 ...................................... 0.9
29
+ adam_beta2 ...................................... 0.999
30
+ adam_eps ........................................ 1e-08
31
+ add_bias_linear ................................. False
32
+ add_position_embedding .......................... False
33
+ add_qkv_bias .................................... True
34
+ adlr_autoresume ................................. False
35
+ adlr_autoresume_interval ........................ 1000
36
+ align_grad_reduce ............................... True
37
+ align_param_gather .............................. False
38
+ app_tag_run_name ................................ None
39
+ app_tag_run_version ............................. 0.0.0
40
+ apply_layernorm_1p .............................. False
41
+ apply_query_key_layer_scaling ................... False
42
+ apply_residual_connection_post_layernorm ........ False
43
+ apply_rope_fusion ............................... True
44
+ async_save ...................................... None
45
+ async_tensor_model_parallel_allreduce ........... True
46
+ attention_backend ............................... AttnBackend.auto
47
+ attention_dropout ............................... 0.0
48
+ attention_softmax_in_fp32 ....................... False
49
+ auto_detect_ckpt_format ......................... False
50
+ barrier_with_L1_time ............................ True
51
+ bert_binary_head ................................ True
52
+ bert_embedder_type .............................. megatron
53
+ bert_load ....................................... None
54
+ bf16 ............................................ True
55
+ bias_dropout_fusion ............................. True
56
+ bias_gelu_fusion ................................ False
57
+ bias_swiglu_fusion .............................. True
58
+ biencoder_projection_dim ........................ 0
59
+ biencoder_shared_query_context_model ............ False
60
+ block_data_path ................................. None
61
+ calc_ft_timeouts ................................ False
62
+ calculate_per_token_loss ........................ False
63
+ check_for_large_grads ........................... False
64
+ check_for_nan_in_loss_and_grad .................. False
65
+ check_for_spiky_loss ............................ False
66
+ check_weight_hash_across_dp_replicas_interval ... None
67
+ ckpt_assume_constant_structure .................. False
68
+ ckpt_convert_format ............................. None
69
+ ckpt_convert_save ............................... None
70
+ ckpt_convert_update_legacy_dist_opt_format ...... False
71
+ ckpt_format ..................................... torch_dist
72
+ ckpt_fully_parallel_load ........................ False
73
+ ckpt_fully_parallel_save ........................ True
74
+ ckpt_fully_parallel_save_deprecated ............. False
75
+ ckpt_step ....................................... None
76
+ classes_fraction ................................ 1.0
77
+ clip_grad ....................................... 1.0
78
+ clone_scatter_output_in_embedding ............... True
79
+ config_logger_dir ...............................
80
+ consumed_train_samples .......................... 0
81
+ consumed_valid_samples .......................... 0
82
+ context_parallel_size ........................... 1
83
+ cp_comm_type .................................... ['p2p']
84
+ create_attention_mask_in_dataloader ............. True
85
+ cross_entropy_fusion_impl ....................... native
86
+ cross_entropy_loss_fusion ....................... False
87
+ cuda_graph_scope ................................ full_iteration
88
+ cuda_graph_warmup_steps ......................... 3
89
+ data_args_path .................................. None
90
+ data_cache_path ................................. None
91
+ data_parallel_random_init ....................... False
92
+ data_parallel_sharding_strategy ................. no_shard
93
+ data_parallel_size .............................. 8
94
+ data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document']
95
+ data_per_class_fraction ......................... 1.0
96
+ data_sharding ................................... True
97
+ dataloader_type ................................. single
98
+ ddp_average_in_collective ....................... False
99
+ ddp_bucket_size ................................. None
100
+ ddp_num_buckets ................................. None
101
+ ddp_pad_buckets_for_high_nccl_busbw ............. False
102
+ decoder_first_pipeline_num_layers ............... None
103
+ decoder_last_pipeline_num_layers ................ None
104
+ decoder_num_layers .............................. None
105
+ decoder_seq_length .............................. None
106
+ decoupled_lr .................................... None
107
+ decoupled_min_lr ................................ None
108
+ decrease_batch_size_if_needed ................... False
109
+ defer_embedding_wgrad_compute ................... False
110
+ delay_wgrad_compute ............................. False
111
+ deprecated_use_mcore_models ..................... True
112
+ deterministic_mode .............................. False
113
+ dino_bottleneck_size ............................ 256
114
+ dino_freeze_last_layer .......................... 1
115
+ dino_head_hidden_size ........................... 2048
116
+ dino_local_crops_number ......................... 10
117
+ dino_local_img_size ............................. 96
118
+ dino_norm_last_layer ............................ False
119
+ dino_teacher_temp ............................... 0.07
120
+ dino_warmup_teacher_temp ........................ 0.04
121
+ dino_warmup_teacher_temp_epochs ................. 30
122
+ disable_bf16_reduced_precision_matmul ........... False
123
+ disable_mamba_mem_eff_path ...................... False
124
+ disable_straggler_on_startup .................... False
125
+ dist_ckpt_format_deprecated ..................... None
126
+ dist_ckpt_strictness ............................ assume_ok_unexpected
127
+ distribute_saved_activations .................... False
128
+ distributed_backend ............................. nccl
129
+ distributed_timeout_minutes ..................... 10
130
+ embedding_init_method_std ....................... None
131
+ embedding_path .................................. None
132
+ empty_unused_memory_level ....................... 0
133
+ enable_cuda_graph ............................... True
134
+ enable_experimental ............................. False
135
+ enable_ft_package ............................... False
136
+ enable_gloo_process_groups ...................... True
137
+ enable_msc ...................................... True
138
+ enable_one_logger ............................... True
139
+ encoder_num_layers .............................. 10
140
+ encoder_seq_length .............................. 4096
141
+ end_weight_decay ................................ 0.1
142
+ eod_mask_loss ................................... False
143
+ error_injection_rate ............................ 0
144
+ error_injection_type ............................ transient_error
145
+ eval_interval ................................... 1000000
146
+ eval_iters ...................................... 0
147
+ evidence_data_path .............................. None
148
+ exit_duration_in_mins ........................... None
149
+ exit_interval ................................... None
150
+ exit_on_missing_checkpoint ...................... False
151
+ exit_signal_handler ............................. False
152
+ exp_avg_dtype ................................... torch.float32
153
+ exp_avg_sq_dtype ................................ torch.float32
154
+ expert_model_parallel_size ...................... 8
155
+ expert_tensor_parallel_size ..................... 1
156
+ export_force_local_attention .................... False
157
+ export_kd_cfg ................................... None
158
+ export_kd_teacher_ckpt_format ................... None
159
+ export_kd_teacher_load .......................... None
160
+ export_kv_cache_quant ........................... False
161
+ export_legacy_megatron .......................... False
162
+ export_model_type ............................... GPTModel
163
+ export_moe_apply_probs_on_input ................. False
164
+ export_qk_l2_norm ............................... False
165
+ export_quant_cfg ................................ None
166
+ export_real_quant_cfg ........................... None
167
+ export_te_mcore_model ........................... False
168
+ external_cuda_graph ............................. False
169
+ ffn_hidden_size ................................. 2560
170
+ finetune ........................................ False
171
+ finetune_data_split ............................. train
172
+ finetune_hf_dataset ............................. None
173
+ first_last_layers_bf16 .......................... False
174
+ flash_decode .................................... False
175
+ fp16 ............................................ False
176
+ fp16_lm_cross_entropy ........................... False
177
+ fp32_residual_connection ........................ False
178
+ fp8 ............................................. e4m3
179
+ fp8_amax_compute_algo ........................... most_recent
180
+ fp8_amax_history_len ............................ 1
181
+ fp8_interval .................................... 1
182
+ fp8_margin ...................................... 0
183
+ fp8_param_gather ................................ False
184
+ fp8_recipe ...................................... mxfp8
185
+ fp8_wgrad ....................................... True
186
+ fsdp_double_buffer .............................. False
187
+ global_batch_size ............................... 8
188
+ grad_reduce_in_bf16 ............................. False
189
+ gradient_accumulation_fusion .................... True
190
+ gradient_reduce_div_fusion ...................... True
191
+ group_query_attention ........................... True
192
+ head_lr_mult .................................... 1.0
193
+ heterogeneous_layers_config_encoded_json ........ None
194
+ heterogeneous_layers_config_path ................ None
195
+ hidden_dropout .................................. 0.0
196
+ hidden_size ..................................... 3584
197
+ hierarchical_context_parallel_sizes ............. None
198
+ high_priority_stream_groups ..................... []
199
+ hybrid_attention_ratio .......................... 0.0
200
+ hybrid_mlp_ratio ................................ 0.0
201
+ hybrid_override_pattern ......................... None
202
+ hysteresis ...................................... 2
203
+ ict_head_size ................................... None
204
+ ict_load ........................................ None
205
+ img_h ........................................... 224
206
+ img_w ........................................... 224
207
+ indexer_batch_size .............................. 128
208
+ indexer_log_interval ............................ 1000
209
+ inference_batch_times_seqlen_threshold .......... -1
210
+ inference_dynamic_batching ...................... False
211
+ inference_dynamic_batching_buffer_guaranteed_fraction 0.2
212
+ inference_dynamic_batching_buffer_overflow_factor None
213
+ inference_dynamic_batching_buffer_size_gb ....... 40.0
214
+ inference_dynamic_batching_chunk_size ........... 256
215
+ inference_dynamic_batching_max_requests_override None
216
+ inference_dynamic_batching_max_tokens_override .. None
217
+ inference_dynamic_batching_num_cuda_graphs ...... 16
218
+ inference_max_batch_size ........................ 8
219
+ inference_max_seq_length ........................ 2560
220
+ inference_rng_tracker ........................... False
221
+ init_method_std ................................. 0.01
222
+ init_method_xavier_uniform ...................... False
223
+ init_model_with_meta_device ..................... False
224
+ initial_loss_scale .............................. 4294967296
225
+ inprocess_active_world_size ..................... 8
226
+ inprocess_barrier_timeout ....................... 120
227
+ inprocess_completion_timeout .................... 120
228
+ inprocess_empty_cuda_cache ...................... False
229
+ inprocess_granularity ........................... node
230
+ inprocess_hard_timeout .......................... 90
231
+ inprocess_heartbeat_interval .................... 30
232
+ inprocess_heartbeat_timeout ..................... 60
233
+ inprocess_last_call_wait ........................ 1
234
+ inprocess_max_iterations ........................ None
235
+ inprocess_monitor_process_interval .............. 1.0
236
+ inprocess_monitor_thread_interval ............... 1.0
237
+ inprocess_progress_watchdog_interval ............ 1.0
238
+ inprocess_restart ............................... False
239
+ inprocess_soft_timeout .......................... 60
240
+ inprocess_termination_grace_time ................ 1
241
+ is_hybrid_model ................................. False
242
+ iter_per_epoch .................................. 1250
243
+ iterations_to_skip .............................. []
244
+ keep_fp8_transpose_cache_when_using_custom_fsdp . False
245
+ kitchen_config_file ............................. None
246
+ kitchen_recipe_number ........................... None
247
+ kv_channels ..................................... 128
248
+ kv_lora_rank .................................... 32
249
+ lazy_mpu_init ................................... None
250
+ load ............................................ None
251
+ load_main_params_from_ckpt ...................... None
252
+ load_model_opt_format ........................... False
253
+ local_rank ...................................... 0
254
+ log_energy ...................................... False
255
+ log_interval .................................... 1
256
+ log_loss_scale_to_tensorboard ................... True
257
+ log_memory_to_tensorboard ....................... False
258
+ log_num_zeros_in_grad ........................... False
259
+ log_params_norm ................................. False
260
+ log_progress .................................... False
261
+ log_straggler ................................... False
262
+ log_throughput .................................. True
263
+ log_timers_to_tensorboard ....................... False
264
+ log_validation_ppl_to_tensorboard ............... False
265
+ log_world_size_to_tensorboard ................... False
266
+ logging_level ................................... None
267
+ loss_scale ...................................... None
268
+ loss_scale_window ............................... 1000
269
+ lr .............................................. 0.0001
270
+ lr_decay_iters .................................. 8
271
+ lr_decay_samples ................................ None
272
+ lr_decay_style .................................. cosine
273
+ lr_warmup_fraction .............................. None
274
+ lr_warmup_init .................................. 0.0
275
+ lr_warmup_iters ................................. 5
276
+ lr_warmup_samples ............................... 0
277
+ lr_wsd_decay_iters .............................. None
278
+ lr_wsd_decay_samples ............................ None
279
+ lr_wsd_decay_style .............................. exponential
280
+ main_grads_dtype ................................ torch.float32
281
+ main_params_dtype ............................... torch.float32
282
+ make_vocab_size_divisible_by .................... 128
283
+ mamba_head_dim .................................. 64
284
+ mamba_num_groups ................................ 8
285
+ mamba_num_heads ................................. None
286
+ mamba_state_dim ................................. 128
287
+ manual_gc ....................................... False
288
+ manual_gc_eval .................................. True
289
+ manual_gc_interval .............................. 0
290
+ mask_factor ..................................... 1.0
291
+ mask_prob ....................................... 0.15
292
+ mask_type ....................................... random
293
+ masked_softmax_fusion ........................... False
294
+ max_position_embeddings ......................... 32768
295
+ max_tokens_to_oom ............................... 12000
296
+ memory_snapshot_path ............................ snapshot.pickle
297
+ merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt
298
+ micro_batch_size ................................ 1
299
+ microbatch_group_size_per_vp_stage .............. None
300
+ mid_level_dataset_surplus ....................... 0.005
301
+ min_loss_scale .................................. 1.0
302
+ min_lr .......................................... 1e-05
303
+ mlp_chunks_for_prefill .......................... 1
304
+ mmap_bin_files .................................. True
305
+ mock_data ....................................... False
306
+ moe_apply_probs_on_input ........................ False
307
+ moe_aux_loss_coeff .............................. 0.01
308
+ moe_deepep_num_sms .............................. 20
309
+ moe_echo_enable_random_offloading ............... False
310
+ moe_echo_expert_dispatch_overlap ................ False
311
+ moe_echo_expert_dispatcher_type ................. hybridep
312
+ moe_echo_recompute_expert_dispatch .............. False
313
+ moe_enable_deepep ............................... False
314
+ moe_enable_echo ................................. False
315
+ moe_enable_hybridep ............................. False
316
+ moe_enable_moonep ............................... True
317
+ moe_expert_capacity_factor ...................... None
318
+ moe_extended_tp ................................. False
319
+ moe_ffn_hidden_size ............................. 2560
320
+ moe_grouped_gemm ................................ True
321
+ moe_hybridep_num_sms ............................ 16
322
+ moe_input_jitter_eps ............................ None
323
+ moe_layer_freq .................................. 1
324
+ moe_layer_recompute ............................. False
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 ................... False
334
+ moe_router_force_load_balancing ................. False
335
+ moe_router_group_topk ........................... None
336
+ moe_router_load_balancing_type .................. aux_loss
337
+ moe_router_num_groups ........................... None
338
+ moe_router_padding_for_fp8 ...................... False
339
+ moe_router_pre_softmax .......................... True
340
+ moe_router_score_function ....................... softmax
341
+ moe_router_topk ................................. 8
342
+ moe_router_topk_scaling_factor .................. None
343
+ moe_shared_expert_intermediate_size ............. 20480
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 .................................. 0.0
355
+ mtp_loss_scaling_factor ......................... 0.1
356
+ mtp_num_layers .................................. None
357
+ multi_latent_attention .......................... False
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 ............................. 28
375
+ num_channels .................................... 3
376
+ num_classes ..................................... 1000
377
+ num_dataset_builder_threads ..................... 1
378
+ num_distributed_optimizer_instances ............. 1
379
+ num_experts ..................................... 64
380
+ num_layers ...................................... 10
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 ................................ 4
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 ..................................... None
421
+ qk_head_dim ..................................... 128
422
+ qk_l2_norm ...................................... False
423
+ qk_layernorm .................................... False
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 ..................................... 1000000
459
+ rotary_interleaved .............................. False
460
+ rotary_percent .................................. 1.0
461
+ rotary_scaling_factor ........................... 1.0
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
+ 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
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
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
579
+ [Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
580
+ [Gloo] Rank 2 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 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
583
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
584
+ [Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
585
+ [Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
586
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
587
+ [Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
588
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
589
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
590
+ [Gloo] Rank [Gloo] Rank 5 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 : [Gloo] Rank
591
+ 74
592
+ is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
593
+ [Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
594
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
595
+ > initialized tensor model parallel with size 1
596
+ > initialized pipeline model parallel with size 1
597
+ > setting random seeds to 1234 ...
598
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
599
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
600
+ [Gloo] Rank 0 is connected to 0[Gloo] Rank peer ranks. 0Expected number of connected peer ranks is : is connected to 00
601
+ 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
+ [rank5]:[W826 18:48:25.596908457 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.
606
+ [rank7]:[W826 18:48:25.599649064 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.
607
+ [rank2]:[W826 18:48:25.599650679 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.
608
+ [rank6]:[W826 18:48:25.599784749 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
+ [rank1]:[W826 18:48:25.599793508 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
610
+ [rank4]:[W826 18:48:25.601148546 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.
611
+ [rank3]:[W826 18:48:25.602093144 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
612
+ > compiling dataset index builder ...
613
+ make: Entering directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
614
+ make: Nothing to be done for 'default'.
615
+ make: Leaving directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
616
+ >>> done with dataset index builder. Compilation time: 0.036 seconds
617
+ WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations.
618
+ > compiling and loading fused kernels ...
619
+ [rank0]:[W826 18:48:25.758806697 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.
620
+ >>> done with compiling and loading fused kernels. Compilation time: 3.491 seconds
621
+ time to initialize megatron (seconds): 7.233
622
+ [after megatron is initialized] datetime: 2026-08-26 18:48:31
623
+ building GPT model ...
624
+ /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.
625
+ warnings.warn(
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
+ > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 5060614656
641
+ 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)
642
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 22
643
+ Params for bucket 1 (180289536 elements, 180289536 padded size):
644
+ module.output_layer.weight
645
+ Params for bucket 2 (73403904 elements, 73403904 padded size):
646
+ module.decoder.final_layernorm.weight
647
+ module.decoder.layers.9.mlp.shared_experts.linear_fc2.weight
648
+ Params for bucket 3 (146800640 elements, 146800640 padded size):
649
+ module.decoder.layers.9.mlp.shared_experts.linear_fc1.weight
650
+ Params for bucket 4 (103001600 elements, 103001600 padded size):
651
+ module.decoder.layers.9.mlp.router.weight
652
+ module.decoder.layers.9.self_attention.linear_qkv.weight
653
+ module.decoder.layers.9.self_attention.linear_proj.weight
654
+ module.decoder.layers.8.mlp.shared_experts.linear_fc2.weight
655
+ module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight
656
+ module.decoder.layers.9.pre_mlp_layernorm.weight
657
+ module.decoder.layers.9.self_attention.linear_qkv.bias
658
+ Params for bucket 5 (146800640 elements, 146800640 padded size):
659
+ module.decoder.layers.8.mlp.shared_experts.linear_fc1.weight
660
+ Params for bucket 6 (103001600 elements, 103001600 padded size):
661
+ module.decoder.layers.8.pre_mlp_layernorm.weight
662
+ module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight
663
+ module.decoder.layers.8.self_attention.linear_qkv.bias
664
+ module.decoder.layers.8.self_attention.linear_proj.weight
665
+ module.decoder.layers.8.mlp.router.weight
666
+ module.decoder.layers.8.self_attention.linear_qkv.weight
667
+ module.decoder.layers.7.mlp.shared_experts.linear_fc2.weight
668
+ Params for bucket 7 (146800640 elements, 146800640 padded size):
669
+ module.decoder.layers.7.mlp.shared_experts.linear_fc1.weight
670
+ Params for bucket 8 (103001600 elements, 103001600 padded size):
671
+ module.decoder.layers.7.pre_mlp_layernorm.weight
672
+ module.decoder.layers.7.self_attention.linear_qkv.bias
673
+ module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight
674
+ module.decoder.layers.7.self_attention.linear_proj.weight
675
+ module.decoder.layers.6.mlp.shared_experts.linear_fc2.weight
676
+ module.decoder.layers.7.mlp.router.weight
677
+ module.decoder.layers.7.self_attention.linear_qkv.weight
678
+ Params for bucket 9 (146800640 elements, 146800640 padded size):
679
+ module.decoder.layers.6.mlp.shared_experts.linear_fc1.weight
680
+ Params for bucket 10 (103001600 elements, 103001600 padded size):
681
+ module.decoder.layers.6.mlp.router.weight
682
+ module.decoder.layers.6.self_attention.linear_qkv.weight
683
+ module.decoder.layers.6.self_attention.linear_proj.weight
684
+ module.decoder.layers.5.mlp.shared_experts.linear_fc2.weight
685
+ module.decoder.layers.6.self_attention.linear_qkv.bias
686
+ module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight
687
+ module.decoder.layers.6.pre_mlp_layernorm.weight
688
+ Params for bucket 11 (146800640 elements, 146800640 padded size):
689
+ module.decoder.layers.5.mlp.shared_experts.linear_fc1.weight
690
+ Params for bucket 12 (103001600 elements, 103001600 padded size):
691
+ module.decoder.layers.5.mlp.router.weight
692
+ module.decoder.layers.5.self_attention.linear_qkv.weight
693
+ module.decoder.layers.5.self_attention.linear_proj.weight
694
+ module.decoder.layers.4.mlp.shared_experts.linear_fc2.weight
695
+ module.decoder.layers.5.self_attention.linear_qkv.bias
696
+ module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight
697
+ module.decoder.layers.5.pre_mlp_layernorm.weight
698
+ Params for bucket 13 (146800640 elements, 146800640 padded size):
699
+ module.decoder.layers.4.mlp.shared_experts.linear_fc1.weight
700
+ Params for bucket 14 (103001600 elements, 103001600 padded size):
701
+ module.decoder.layers.4.mlp.router.weight
702
+ module.decoder.layers.4.self_attention.linear_qkv.layer_norm_weight
703
+ module.decoder.layers.4.pre_mlp_layernorm.weight
704
+ module.decoder.layers.4.self_attention.linear_qkv.bias
705
+ module.decoder.layers.4.self_attention.linear_qkv.weight
706
+ module.decoder.layers.4.self_attention.linear_proj.weight
707
+ module.decoder.layers.3.mlp.shared_experts.linear_fc2.weight
708
+ Params for bucket 15 (146800640 elements, 146800640 padded size):
709
+ module.decoder.layers.3.mlp.shared_experts.linear_fc1.weight
710
+ Params for bucket 16 (103001600 elements, 103001600 padded size):
711
+ module.decoder.layers.3.self_attention.linear_qkv.bias
712
+ module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight
713
+ module.decoder.layers.3.pre_mlp_layernorm.weight
714
+ module.decoder.layers.3.mlp.router.weight
715
+ module.decoder.layers.3.self_attention.linear_qkv.weight
716
+ module.decoder.layers.3.self_attention.linear_proj.weight
717
+ module.decoder.layers.2.mlp.shared_experts.linear_fc2.weight
718
+ Params for bucket 17 (146800640 elements, 146800640 padded size):
719
+ module.decoder.layers.2.mlp.shared_experts.linear_fc1.weight
720
+ Params for bucket 18 (103001600 elements, 103001600 padded size):
721
+ module.decoder.layers.2.mlp.router.weight
722
+ module.decoder.layers.2.self_attention.linear_qkv.weight
723
+ module.decoder.layers.2.pre_mlp_layernorm.weight
724
+ module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight
725
+ module.decoder.layers.2.self_attention.linear_qkv.bias
726
+ module.decoder.layers.2.self_attention.linear_proj.weight
727
+ module.decoder.layers.1.mlp.shared_experts.linear_fc2.weight
728
+ Params for bucket 19 (146800640 elements, 146800640 padded size):
729
+ module.decoder.layers.1.mlp.shared_experts.linear_fc1.weight
730
+ Params for bucket 20 (103001600 elements, 103001600 padded size):
731
+ module.decoder.layers.1.mlp.router.weight
732
+ module.decoder.layers.1.self_attention.linear_qkv.weight
733
+ module.decoder.layers.1.self_attention.linear_proj.weight
734
+ module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight
735
+ module.decoder.layers.1.pre_mlp_layernorm.weight
736
+ module.decoder.layers.1.self_attention.linear_qkv.bias
737
+ module.decoder.layers.0.mlp.shared_experts.linear_fc2.weight
738
+ Params for bucket 21 (146800640 elements, 146800640 padded size):
739
+ module.decoder.layers.0.mlp.shared_experts.linear_fc1.weight
740
+ Params for bucket 22 (209890816 elements, 209890816 padded size):
741
+ module.decoder.layers.0.self_attention.linear_qkv.weight
742
+ module.decoder.layers.0.mlp.router.weight
743
+ module.decoder.layers.0.self_attention.linear_qkv.layer_norm_weight
744
+ module.decoder.layers.0.self_attention.linear_qkv.bias
745
+ module.decoder.layers.0.pre_mlp_layernorm.weight
746
+ module.decoder.layers.0.self_attention.linear_proj.weight
747
+ module.embedding.word_embeddings.weight
748
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 44
749
+ Params for bucket 1 (45875200 elements, 45875200 padded size):
750
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight4
751
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight6
752
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight7
753
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight5
754
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight3
755
+ Params for bucket 2 (45875200 elements, 45875200 padded size):
756
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight2
757
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight1
758
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight0
759
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight7
760
+ Params for bucket 3 (55050240 elements, 55050240 padded size):
761
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight6
762
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight5
763
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight4
764
+ Params for bucket 4 (55050240 elements, 55050240 padded size):
765
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight3
766
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight2
767
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight1
768
+ Params for bucket 5 (45875200 elements, 45875200 padded size):
769
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight0
770
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight6
771
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight7
772
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight5
773
+ Params for bucket 6 (45875200 elements, 45875200 padded size):
774
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight3
775
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight1
776
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight4
777
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight0
778
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight2
779
+ Params for bucket 7 (55050240 elements, 55050240 padded size):
780
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight7
781
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight6
782
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight5
783
+ Params for bucket 8 (55050240 elements, 55050240 padded size):
784
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight4
785
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight3
786
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight2
787
+ Params for bucket 9 (45875200 elements, 45875200 padded size):
788
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight1
789
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight0
790
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight7
791
+ Params for bucket 10 (45875200 elements, 45875200 padded size):
792
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight6
793
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight4
794
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight2
795
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight5
796
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight3
797
+ Params for bucket 11 (55050240 elements, 55050240 padded size):
798
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight1
799
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight0
800
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight7
801
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight6
802
+ Params for bucket 12 (55050240 elements, 55050240 padded size):
803
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight5
804
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight3
805
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight4
806
+ Params for bucket 13 (55050240 elements, 55050240 padded size):
807
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight2
808
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight0
809
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight1
810
+ Params for bucket 14 (45875200 elements, 45875200 padded size):
811
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight6
812
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight4
813
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight7
814
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight5
815
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight3
816
+ Params for bucket 15 (45875200 elements, 45875200 padded size):
817
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight2
818
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight0
819
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight1
820
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight7
821
+ Params for bucket 16 (55050240 elements, 55050240 padded size):
822
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight6
823
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight4
824
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight5
825
+ Params for bucket 17 (55050240 elements, 55050240 padded size):
826
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight3
827
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight1
828
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight2
829
+ Params for bucket 18 (45875200 elements, 45875200 padded size):
830
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight0
831
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight6
832
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight7
833
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight5
834
+ Params for bucket 19 (45875200 elements, 45875200 padded size):
835
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight4
836
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight0
837
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight2
838
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight3
839
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight1
840
+ Params for bucket 20 (55050240 elements, 55050240 padded size):
841
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight7
842
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight5
843
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight6
844
+ Params for bucket 21 (55050240 elements, 55050240 padded size):
845
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight4
846
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight3
847
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight2
848
+ Params for bucket 22 (45875200 elements, 45875200 padded size):
849
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight1
850
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight0
851
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight7
852
+ Params for bucket 23 (45875200 elements, 45875200 padded size):
853
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight6
854
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight2
855
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight4
856
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight5
857
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight3
858
+ Params for bucket 24 (55050240 elements, 55050240 padded size):
859
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight1
860
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight7
861
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight6
862
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight0
863
+ Params for bucket 25 (55050240 elements, 55050240 padded size):
864
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight5
865
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight3
866
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight4
867
+ Params for bucket 26 (55050240 elements, 55050240 padded size):
868
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight2
869
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight0
870
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight1
871
+ Params for bucket 27 (45875200 elements, 45875200 padded size):
872
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight7
873
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight5
874
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight3
875
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight4
876
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight6
877
+ Params for bucket 28 (45875200 elements, 45875200 padded size):
878
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight2
879
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight1
880
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight7
881
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight0
882
+ Params for bucket 29 (55050240 elements, 55050240 padded size):
883
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight6
884
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight5
885
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight4
886
+ Params for bucket 30 (55050240 elements, 55050240 padded size):
887
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight3
888
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight2
889
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight1
890
+ Params for bucket 31 (45875200 elements, 45875200 padded size):
891
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight0
892
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight5
893
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight6
894
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight7
895
+ Params for bucket 32 (45875200 elements, 45875200 padded size):
896
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight4
897
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight2
898
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight0
899
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight3
900
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight1
901
+ Params for bucket 33 (55050240 elements, 55050240 padded size):
902
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight7
903
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight5
904
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight6
905
+ Params for bucket 34 (55050240 elements, 55050240 padded size):
906
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight4
907
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight2
908
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight3
909
+ Params for bucket 35 (45875200 elements, 45875200 padded size):
910
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight1
911
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight7
912
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight0
913
+ Params for bucket 36 (45875200 elements, 45875200 padded size):
914
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight3
915
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight5
916
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight6
917
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight4
918
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight2
919
+ Params for bucket 37 (55050240 elements, 55050240 padded size):
920
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight1
921
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight0
922
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight6
923
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight7
924
+ Params for bucket 38 (55050240 elements, 55050240 padded size):
925
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight5
926
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight3
927
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight4
928
+ Params for bucket 39 (55050240 elements, 55050240 padded size):
929
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight2
930
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight1
931
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight0
932
+ Params for bucket 40 (45875200 elements, 45875200 padded size):
933
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight7
934
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight5
935
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight3
936
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight6
937
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight4
938
+ Params for bucket 41 (45875200 elements, 45875200 padded size):
939
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight2
940
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight1
941
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight7
942
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight0
943
+ Params for bucket 42 (55050240 elements, 55050240 padded size):
944
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight6
945
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight4
946
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight5
947
+ Params for bucket 43 (55050240 elements, 55050240 padded size):
948
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight3
949
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight1
950
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight2
951
+ Params for bucket 44 (18350080 elements, 18350080 padded size):
952
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight0
953
+ INFO:megatron.core.optimizer:Setting up optimizer with config OptimizerConfig(optimizer='adam', lr=0.0001, min_lr=1e-05, decoupled_lr=None, decoupled_min_lr=None, weight_decay=0.1, fp8_recipe='mxfp8', fp16=False, bf16=True, reuse_grad_buf_for_mxfp8_param_ag=False, params_dtype=torch.bfloat16, use_precision_aware_optimizer=False, store_param_remainders=True, main_grads_dtype=torch.float32, main_params_dtype=torch.float32, exp_avg_dtype=torch.float32, exp_avg_sq_dtype=torch.float32, loss_scale=None, initial_loss_scale=4294967296, min_loss_scale=1.0, loss_scale_window=1000, hysteresis=2, adam_beta1=0.9, adam_beta2=0.999, adam_eps=1e-08, sgd_momentum=0.9, use_distributed_optimizer=True, overlap_param_gather_with_optimizer_step=False, optimizer_cpu_offload=False, optimizer_offload_fraction=1.0, use_torch_optimizer_for_cpu_offload=False, overlap_cpu_optimizer_d2h_h2d=False, pin_cpu_grads=True, pin_cpu_params=True, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=<megatron.core.timers.Timers object at 0x7dc3aee1f560>, config_logger_dir='')
954
+ INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine
955
+ [after model, optimizer, and learning rate scheduler are built] datetime: 2026-08-26 18:48:33
956
+ > building train, validation, and test datasets ...
957
+ > datasets target sizes (minimum size):
958
+ train: 96
959
+ validation: 0
960
+ test: 0
961
+ INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None]
962
+ > building train, validation, and test datasets for GPT ...
963
+ INFO:megatron.core.datasets.blended_megatron_dataset_builder:Building GPTDataset splits with sizes=(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 0x7dc10a5d5430>, mid_level_dataset_surplus=0.005, reset_position_ids=False, reset_attention_mask=False, eod_mask_loss=False, create_attention_mask=True, drop_last_partial_validation_sequence=True, add_extra_token_to_sequence=True, object_storage_cache_path=None)
964
+ INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx
965
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths
966
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers
967
+ INFO:megatron.core.datasets.indexed_dataset: Extract the document indices
968
+ INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029
969
+ INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350
970
+ INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices
971
+ INFO:megatron.core.datasets.gpt_dataset: Load the document index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-document_index.npy
972
+ INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-sample_index.npy
973
+ INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-shuffle_index.npy
974
+ INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073
975
+ > finished creating GPT datasets ...
976
+ [after dataloaders are built] datetime: 2026-08-26 18:48:33
977
+ done with setup ...
978
+ (min, max) time across ranks (ms):
979
+ model-and-optimizer-setup ......................: (2452.04, 2457.61)
980
+ train/valid/test-data-iterators-setup ..........: (59.47, 65.80)
981
+ training ...
982
+ Overwriting rerun_state_machine.current_iteration from -1 to 0...
983
+ [before the start of training step] datetime: 2026-08-26 18:48:33
984
+ INFO:moonep.api:MoonEP comm buffer sizes:
985
+ config : R=8, S=4096, H=3584, E=64, K=8, NvS=34800, NvS_padded=34816
986
+ sm_count : num_sms=32, num_sms_dedup=148
987
+ hidden_buffer_size : 238.00 MiB
988
+ meta_buffer_size : 2.00 MiB
989
+ local_temp_buffer_size : 1.26 MiB
990
+ total_size : 241.26 MiB
991
+ [2026-08-26 18:48:56] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 22694.3 | throughput per GPU (TFLOP/s/GPU): 5.4 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.097850E+01 | loss scale: 1.0 | grad norm: 29.826 | number of skipped iterations: 0 | number of nan iterations: 0 |
992
+ Number of parameters in transformer block in billions: 20.11
993
+ Number of parameters in embedding layers in billions: 0.36
994
+ Total number of parameters in billions: 20.47
995
+ Number of parameters in most loaded shard in billions: 20.4724
996
+ Theoretical memory footprints: weight and optimizer=146430.14 MB
997
+ [Rank 0] (after 1 iterations) memory (MB) | allocated: 105000.43212890625 | max allocated: 105000.4326171875 | reserved: 108632.0 | max reserved: 108632.0
998
+ [2026-08-26 18:48:57] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 359.5 | throughput per GPU (TFLOP/s/GPU): 343.5 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.097848E+01 | loss scale: 1.0 | grad norm: 30.561 | number of skipped iterations: 0 | number of nan iterations: 0 |
999
+ [2026-08-26 18:48:57] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 329.4 | throughput per GPU (TFLOP/s/GPU): 374.9 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 9.533901E+00 | loss scale: 1.0 | grad norm: 32.037 | number of skipped iterations: 0 | number of nan iterations: 0 |
1000
+ Capture CUDA graph for training!!!
1001
+ CUDA graph capture done!!!
1002
+ [2026-08-26 18:48:58] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 895.6 | throughput per GPU (TFLOP/s/GPU): 137.9 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 1.110981E+01 | loss scale: 1.0 | grad norm: 107.507 | number of skipped iterations: 0 | number of nan iterations: 0 |
1003
+ [2026-08-26 18:48:58] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 325.9 | throughput per GPU (TFLOP/s/GPU): 378.9 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 9.943975E+00 | loss scale: 1.0 | grad norm: 23.010 | number of skipped iterations: 0 | number of nan iterations: 0 |
1004
+ [2026-08-26 18:48:58] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 319.7 | throughput per GPU (TFLOP/s/GPU): 386.3 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 1.046578E+01 | loss scale: 1.0 | grad norm: 8.600 | number of skipped iterations: 0 | number of nan iterations: 0 |
1005
+ [2026-08-26 18:48:59] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 323.3 | throughput per GPU (TFLOP/s/GPU): 381.9 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 9.389576E+00 | loss scale: 1.0 | grad norm: 4.970 | number of skipped iterations: 0 | number of nan iterations: 0 |
1006
+ [2026-08-26 18:48:59] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 325.8 | throughput per GPU (TFLOP/s/GPU): 379.0 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.895209E+00 | loss scale: 1.0 | grad norm: 4.617 | number of skipped iterations: 0 | number of nan iterations: 0 |
1007
+ [2026-08-26 18:48:59] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 328.6 | throughput per GPU (TFLOP/s/GPU): 375.8 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.670596E+00 | loss scale: 1.0 | grad norm: 4.294 | number of skipped iterations: 0 | number of nan iterations: 0 |
1008
+ [2026-08-26 18:49:00] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 320.9 | throughput per GPU (TFLOP/s/GPU): 384.8 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.601568E+00 | loss scale: 1.0 | grad norm: 4.395 | number of skipped iterations: 0 | number of nan iterations: 0 |
1009
+ [2026-08-26 18:49:00] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 326.2 | throughput per GPU (TFLOP/s/GPU): 378.6 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.541773E+00 | loss scale: 1.0 | grad norm: 4.297 | number of skipped iterations: 0 | number of nan iterations: 0 |
1010
+ [2026-08-26 18:49:00] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 325.5 | throughput per GPU (TFLOP/s/GPU): 379.4 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 8.468069E+00 | loss scale: 1.0 | grad norm: 4.144 | number of skipped iterations: 0 | number of nan iterations: 0 |
1011
+ [after training is done] datetime: 2026-08-26 18:49:00
1012
+ [rank0]:[W826 18:49:02.300158672 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
1013
+ W0826 18:49:05.832000 559545 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers
1014
+ W0826 18:49:05.833000 559545 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 559633 closing signal SIGTERM
1015
+ W0826 18:49:05.833000 559545 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 559634 closing signal SIGTERM
1016
+ W0826 18:49:05.839000 559545 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 559635 closing signal SIGTERM
1017
+ W0826 18:49:05.841000 559545 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 559636 closing signal SIGTERM
1018
+ W0826 18:49:05.843000 559545 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 559637 closing signal SIGTERM
1019
+ W0826 18:49:05.845000 559545 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 559638 closing signal SIGTERM
1020
+ W0826 18:49:05.847000 559545 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 559639 closing signal SIGTERM
1021
+ W0826 18:49:05.850000 559545 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 559640 closing signal SIGTERM
1022
+ Traceback (most recent call last):
1023
+ File "/usr/local/bin/torchrun", line 8, in <module>
1024
+ sys.exit(main())
1025
+ ^^^^^^
1026
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
1027
+ return f(*args, **kwargs)
1028
+ ^^^^^^^^^^^^^^^^^^
1029
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 892, in main
1030
+ run(args)
1031
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 883, in run
1032
+ elastic_launch(
1033
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
1034
+ return launch_agent(self._config, self._entrypoint, list(args))
1035
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1036
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 261, in launch_agent
1037
+ result = agent.run()
1038
+ ^^^^^^^^^^^
1039
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper
1040
+ result = f(*args, **kwargs)
1041
+ ^^^^^^^^^^^^^^^^^^
1042
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 711, in run
1043
+ result = self._invoke_run(role)
1044
+ ^^^^^^^^^^^^^^^^^^^^^^
1045
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 870, in _invoke_run
1046
+ time.sleep(monitor_interval)
1047
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 84, in _terminate_process_handler
1048
+ raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval)
1049
+ torch.distributed.elastic.multiprocessing.api.SignalException: Process 559545 got signal: 15
moonep/tierb/tierb_gbs8/results.tsv ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir
2
+ DSV2_LITE_L24 moonep matched 8 8192 8 none none 1 OK 701.00 698.60 165295 /workspace/tierb_study_20260826_144644/moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep1
3
+ DSV2_LITE_L24 moonep matched 8 8192 8 none none 2 OK 701.60 698.60 165295 /workspace/tierb_study_20260826_144644/moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep2
4
+ DSV2_LITE_L24 moonep matched 8 8192 8 none none 3 OK 701.32 698.40 165295 /workspace/tierb_study_20260826_144644/moonep/tierb/tierb_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep3
5
+ QWEN2_57B moonep matched 8 4096 8 none none 1 OK 325.14 322.10 142739 /workspace/tierb_study_20260826_144644/moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep1
6
+ QWEN2_57B moonep matched 8 4096 8 none none 2 OK 325.08 322.10 142739 /workspace/tierb_study_20260826_144644/moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep2
7
+ QWEN2_57B moonep matched 8 4096 8 none none 3 OK 324.81 319.70 142739 /workspace/tierb_study_20260826_144644/moonep/tierb/tierb_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep3