AbhishekGhosh1998 commited on
Commit
1bb110e
·
verified ·
1 Parent(s): a17cffb

granularity results tierb_study_20260826_144644 @ 2026-08-26T16:02:45Z from sp-aiinfra-interactive-aghosh-b5xzj-8xb200-moe-cg-0

Browse files
tierb_study_20260826_144644/.completed/moonep__tierbprobe__gbs8 ADDED
@@ -0,0 +1 @@
 
 
1
+ 2026-08-26T16:00:40+00:00
tierb_study_20260826_144644/.completed/ultraep__tierbprobe__gbs8 ADDED
@@ -0,0 +1 @@
 
 
1
+ 2026-08-26T15:53:23+00:00
tierb_study_20260826_144644/STATUS.txt CHANGED
@@ -1,10 +1,12 @@
1
  study_root: /workspace/tierb_study_20260826_144644
2
  gbs: 8
3
- updated: 2026-08-26T15:43:26+00:00
4
- current: ultraep__tierbprobe__gbs8
5
 
6
  completed steps:
7
  echo__tierbprobe__gbs8
 
 
8
 
9
  failed steps:
10
  FAIL ultraep tierbprobe 8 2026-08-26T14:57:57+00:00 2026-08-26T15:05:21+00:00 /workspace/tierb_study_20260826_144644/ultraep/tierbprobe
 
1
  study_root: /workspace/tierb_study_20260826_144644
2
  gbs: 8
3
+ updated: 2026-08-26T16:00:42+00:00
4
+ current: echo__tierb__gbs8
5
 
6
  completed steps:
7
  echo__tierbprobe__gbs8
8
+ moonep__tierbprobe__gbs8
9
+ ultraep__tierbprobe__gbs8
10
 
11
  failed steps:
12
  FAIL ultraep tierbprobe 8 2026-08-26T14:57:57+00:00 2026-08-26T15:05:21+00:00 /workspace/tierb_study_20260826_144644/ultraep/tierbprobe
tierb_study_20260826_144644/echo/tierb/manifest.txt ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ timestamp: 2026-08-26T16:00:57+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-echo
6
+ study_scripts_commit: d11391f
7
+ campaign: echo
8
+ backend: echo
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: bracing-regions-through-kernels
18
+ megatron_branch_actual: bracing-regions-through-kernels
19
+ megatron_commit: 9a236793
tierb_study_20260826_144644/echo/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@9a23679390d7ab1fe462bc61fa31d90d393db49c#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
tierb_study_20260826_144644/echo/tierb/provenance.txt ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ==============================================================
2
+ PROVENANCE: echo__tierb__gbs8
3
+ ==============================================================
4
+ captured_at : 2026-08-26T16:00:42+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/echo/tierb --nsys
9
+
10
+ ---- GPU / driver --------------------------------------------
11
+ 0, NVIDIA B200, 10.0, 183359 MiB, 580.95.05
12
+ 1, NVIDIA B200, 10.0, 183359 MiB, 580.95.05
13
+ 2, NVIDIA B200, 10.0, 183359 MiB, 580.95.05
14
+ 3, NVIDIA B200, 10.0, 183359 MiB, 580.95.05
15
+ 4, NVIDIA B200, 10.0, 183359 MiB, 580.95.05
16
+ 5, NVIDIA B200, 10.0, 183359 MiB, 580.95.05
17
+ 6, NVIDIA B200, 10.0, 183359 MiB, 580.95.05
18
+ 7, NVIDIA B200, 10.0, 183359 MiB, 580.95.05
19
+ persistence/clocks:
20
+ 0, 120 MHz, 1965 MHz, 1000.00 W, 29
21
+ 1, 120 MHz, 1965 MHz, 1000.00 W, 34
22
+ 2, 120 MHz, 1965 MHz, 1000.00 W, 34
23
+ 3, 120 MHz, 1965 MHz, 1000.00 W, 30
24
+ 4, 120 MHz, 1965 MHz, 1000.00 W, 30
25
+ 5, 120 MHz, 1965 MHz, 1000.00 W, 34
26
+ 6, 127 MHz, 1965 MHz, 1000.00 W, 34
27
+ 7, 120 MHz, 1965 MHz, 1000.00 W, 30
28
+
29
+ NVLink topology:
30
+ GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7 NIC0 NIC1 NIC2 NIC3 NIC4 NIC5 NIC6 NIC7 NIC8 CPU Affinity NUMA Affinity GPU NUMA ID
31
+ GPU0 X NV18 NV18 NV18 NV18 NV18 NV18 NV18 PIX NODE SYS SYS SYS SYS SYS SYS NODE 0-27 0 N/A
32
+ GPU1 NV18 X NV18 NV18 NV18 NV18 NV18 NV18 NODE PIX SYS SYS SYS SYS SYS SYS NODE 0-27 0 N/A
33
+ GPU2 NV18 NV18 X NV18 NV18 NV18 NV18 NV18 SYS SYS PIX NODE SYS SYS SYS SYS SYS 28-55 1 N/A
34
+ GPU3 NV18 NV18 NV18 X NV18 NV18 NV18 NV18 SYS SYS NODE PIX SYS SYS SYS SYS SYS 28-55 1 N/A
35
+ GPU4 NV18 NV18 NV18 NV18 X NV18 NV18 NV18 SYS SYS SYS SYS PIX NODE SYS SYS SYS 56-83 2 N/A
36
+ GPU5 NV18 NV18 NV18 NV18 NV18 X NV18 NV18 SYS SYS SYS SYS NODE PIX SYS SYS SYS 56-83 2 N/A
37
+ GPU6 NV18 NV18 NV18 NV18 NV18 NV18 X NV18 SYS SYS SYS SYS SYS SYS PIX NODE SYS 84-111 3 N/A
38
+ GPU7 NV18 NV18 NV18 NV18 NV18 NV18 NV18 X SYS SYS SYS SYS SYS SYS NODE PIX SYS 84-111 3 N/A
39
+ NIC0 PIX NODE SYS SYS SYS SYS SYS SYS X NODE SYS SYS SYS SYS SYS SYS NODE
40
+ NIC1 NODE PIX SYS SYS SYS SYS SYS SYS NODE X SYS SYS SYS SYS SYS SYS NODE
41
+ NIC2 SYS SYS PIX NODE SYS SYS SYS SYS SYS SYS X NODE SYS SYS SYS SYS SYS
42
+ NIC3 SYS SYS NODE PIX SYS SYS SYS SYS SYS SYS NODE X SYS SYS SYS SYS SYS
43
+ NIC4 SYS SYS SYS SYS PIX NODE SYS SYS SYS SYS SYS SYS X NODE SYS SYS SYS
44
+ unavailable
45
+
46
+ ---- toolchain -----------------------------------------------
47
+ nvcc : release 12.9, V12.9.86
48
+ gcc : gcc (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
49
+ python : 3.12.3
50
+ torch : 2.8.0a0+5228986c39.nv25.06 cuda=12.9 nccl=2.27.3
51
+ torch arch : ['sm_75', 'sm_80', 'sm_86', 'sm_90', 'sm_100', 'sm_120', 'compute_120']
52
+ transformer_engine: 2.10.0.dev0+10f6128d [/usr/local/lib/python3.12/dist-packages/transformer_engine/__init__.py]
53
+ deep_ep : 1.2.1+34152ae [/workspace/DeepEP/deep_ep/__init__.py]
54
+ ultra_ep : 1.0.0 [/workspace/UltraEP/ultra_ep/__init__.py]
55
+ moonep : 0.0.1 [/workspace/MoonEP/moonep/__init__.py]
56
+ megatron : 0.14.0rc5 [None]
57
+ flash_attn : 2.7.4.post1 [/usr/local/lib/python3.12/dist-packages/flash_attn/__init__.py]
58
+ numpy : 1.26.4 [/usr/local/lib/python3.12/dist-packages/numpy/__init__.py]
59
+ huggingface_hub: 1.28.0 [/usr/local/lib/python3.12/dist-packages/huggingface_hub/__init__.py]
60
+
61
+ ---- repositories --------------------------------------------
62
+ Megatron-MoE-Echo bracing-regions-through-kernels 9a23679390d7
63
+ Megatron-UltraEP ultraep-token-drop-analysis eaf829e7ad21
64
+ UltraEP main 94cab099b44f
65
+ DeepEP HEAD 34152ae28f80 (+3 untracked: build output)
66
+ MoonEP megatron-integration c27d35da15f2
67
+ blackwell_run_scripts granularity-echo d11391fc9103
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
tierb_study_20260826_144644/echo/tierb/tierb_gbs8/QWEN3_235B__echo__matched__seq8192__gbs8__cap2p00/rep1/mem_samples.csv ADDED
@@ -0,0 +1,256 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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, 4
44
+ 3, 629
45
+ 4, 4
46
+ 5, 4
47
+ 6, 4
48
+ 7, 37
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, 1975
58
+ 1, 1975
59
+ 2, 1975
60
+ 3, 1975
61
+ 4, 1975
62
+ 5, 1975
63
+ 6, 1975
64
+ 7, 1975
65
+ 0, 47803
66
+ 1, 47803
67
+ 2, 47803
68
+ 3, 47803
69
+ 4, 47515
70
+ 5, 47803
71
+ 6, 47803
72
+ 7, 47803
73
+ 0, 49057
74
+ 1, 49057
75
+ 2, 49057
76
+ 3, 49057
77
+ 4, 48769
78
+ 5, 49057
79
+ 6, 49057
80
+ 7, 49057
81
+ 0, 49929
82
+ 1, 49929
83
+ 2, 49929
84
+ 3, 49931
85
+ 4, 49641
86
+ 5, 49929
87
+ 6, 49929
88
+ 7, 49929
89
+ 0, 53699
90
+ 1, 53699
91
+ 2, 53699
92
+ 3, 53699
93
+ 4, 53411
94
+ 5, 53699
95
+ 6, 53699
96
+ 7, 53699
97
+ 0, 54107
98
+ 1, 54107
99
+ 2, 54107
100
+ 3, 54107
101
+ 4, 53819
102
+ 5, 54107
103
+ 6, 54107
104
+ 7, 54107
105
+ 0, 54433
106
+ 1, 54431
107
+ 2, 54429
108
+ 3, 54451
109
+ 4, 54143
110
+ 5, 54433
111
+ 6, 54421
112
+ 7, 54451
113
+ 0, 55545
114
+ 1, 55545
115
+ 2, 55545
116
+ 3, 55545
117
+ 4, 55257
118
+ 5, 55545
119
+ 6, 55545
120
+ 7, 55545
121
+ 0, 56569
122
+ 1, 56569
123
+ 2, 56569
124
+ 3, 56569
125
+ 4, 56281
126
+ 5, 56569
127
+ 6, 56569
128
+ 7, 56569
129
+ 0, 59801
130
+ 1, 59801
131
+ 2, 59801
132
+ 3, 59801
133
+ 4, 59513
134
+ 5, 59801
135
+ 6, 59801
136
+ 7, 59801
137
+ 0, 59801
138
+ 1, 59801
139
+ 2, 59801
140
+ 3, 59801
141
+ 4, 59513
142
+ 5, 59801
143
+ 6, 59801
144
+ 7, 59801
145
+ 0, 105853
146
+ 1, 105925
147
+ 2, 105925
148
+ 3, 105899
149
+ 4, 105885
150
+ 5, 105911
151
+ 6, 105901
152
+ 7, 105907
153
+ 0, 106307
154
+ 1, 106347
155
+ 2, 106347
156
+ 3, 106347
157
+ 4, 106327
158
+ 5, 106347
159
+ 6, 106347
160
+ 7, 106347
161
+ 0, 106307
162
+ 1, 106347
163
+ 2, 106347
164
+ 3, 106347
165
+ 4, 106327
166
+ 5, 106347
167
+ 6, 106347
168
+ 7, 106347
169
+ 0, 106307
170
+ 1, 106347
171
+ 2, 106347
172
+ 3, 106347
173
+ 4, 106327
174
+ 5, 106347
175
+ 6, 106347
176
+ 7, 106347
177
+ 0, 106307
178
+ 1, 106347
179
+ 2, 106347
180
+ 3, 106347
181
+ 4, 106327
182
+ 5, 106347
183
+ 6, 106347
184
+ 7, 106347
185
+ 0, 107657
186
+ 1, 107713
187
+ 2, 107697
188
+ 3, 107697
189
+ 4, 107677
190
+ 5, 107697
191
+ 6, 107713
192
+ 7, 107713
193
+ 0, 111393
194
+ 1, 111433
195
+ 2, 111433
196
+ 3, 111433
197
+ 4, 111413
198
+ 5, 111433
199
+ 6, 111433
200
+ 7, 111433
201
+ 0, 112059
202
+ 1, 112079
203
+ 2, 112151
204
+ 3, 112125
205
+ 4, 112069
206
+ 5, 112151
207
+ 6, 112159
208
+ 7, 112139
209
+ 0, 130219
210
+ 1, 130923
211
+ 2, 131083
212
+ 3, 131667
213
+ 4, 133127
214
+ 5, 94717
215
+ 6, 130967
216
+ 7, 122835
217
+ 0, 155643
218
+ 1, 155809
219
+ 2, 155809
220
+ 3, 155809
221
+ 4, 155423
222
+ 5, 155809
223
+ 6, 155809
224
+ 7, 155809
225
+ 0, 155643
226
+ 1, 155809
227
+ 2, 155809
228
+ 3, 155809
229
+ 4, 155423
230
+ 5, 155809
231
+ 6, 155809
232
+ 7, 155809
233
+ 0, 154397
234
+ 1, 154563
235
+ 2, 154563
236
+ 3, 154563
237
+ 4, 154177
238
+ 5, 154563
239
+ 6, 154563
240
+ 7, 154563
241
+ 0, 3363
242
+ 1, 3366
243
+ 2, 3622
244
+ 3, 154561
245
+ 4, 4646
246
+ 5, 3366
247
+ 6, 3366
248
+ 7, 3366
249
+ 0, 0
250
+ 1, 0
251
+ 2, 0
252
+ 3, 0
253
+ 4, 0
254
+ 5, 0
255
+ 6, 0
256
+ 7, 0
tierb_study_20260826_144644/echo/tierb/tierb_gbs8/QWEN3_235B__echo__matched__seq8192__gbs8__cap2p00/rep1/run.env ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=echo
2
+ config=QWEN3_235B
3
+ family=tierB
4
+ repo=/workspace/Megatron-MoE-Echo
5
+ branch=bracing-regions-through-kernels
6
+ commit=9a236793
7
+ num_layers=12
8
+ hidden=4096
9
+ ffn_hidden=12288
10
+ moe_ffn_hidden=1536
11
+ num_experts=128
12
+ topk=8
13
+ experts_per_rank=16
14
+ expert_params=18874368
15
+ heads=64
16
+ kv_groups=4
17
+ kv_channels=128
18
+ qk_layernorm=1
19
+ shared_ffn=0
20
+ shared_expert_overlap=0
21
+ mla=0
22
+ dense_first_layers=0
23
+ moe_layers=12
24
+ rotary_base=1000000
25
+ rotary_scaling=
26
+ qkv_bias=0
27
+ router_balancing=aux_loss
28
+ aux_loss_coeff=1e-3
29
+ router_pre_softmax=0
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=65536
39
+ mean_tokens_per_expert=4096
40
+ slot_policy=matched
41
+ slots_per_rank=16
42
+ capacity_mode=static
43
+ capacity=2.00
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
tierb_study_20260826_144644/echo/tierb/tierb_gbs8/QWEN3_235B__echo__matched__seq8192__gbs8__cap2p00/rep1/train.log ADDED
The diff for this file is too large to render. See raw diff
 
tierb_study_20260826_144644/echo/tierb/tierb_gbs8/QWEN3_235B__echo__matched__seq8192__gbs8__cap2p00/rep2/mem_samples.csv ADDED
@@ -0,0 +1,80 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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, 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, 1537
58
+ 1, 1537
59
+ 2, 1537
60
+ 3, 1537
61
+ 4, 1537
62
+ 5, 1537
63
+ 6, 1537
64
+ 7, 1537
65
+ 0, 2551
66
+ 1, 2551
67
+ 2, 2551
68
+ 3, 2551
69
+ 4, 2551
70
+ 5, 2551
71
+ 6, 2551
72
+ 7, 2551
73
+ 0, 48171
74
+ 1, 48147
75
+ 2, 48155
76
+ 3, 48165
77
+ 4, 47865
78
+ 5, 48127
79
+ 6, 48143
80
+ 7, 48171
tierb_study_20260826_144644/echo/tierb/tierb_gbs8/QWEN3_235B__echo__matched__seq8192__gbs8__cap2p00/rep2/run.env ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=echo
2
+ config=QWEN3_235B
3
+ family=tierB
4
+ repo=/workspace/Megatron-MoE-Echo
5
+ branch=bracing-regions-through-kernels
6
+ commit=9a236793
7
+ num_layers=12
8
+ hidden=4096
9
+ ffn_hidden=12288
10
+ moe_ffn_hidden=1536
11
+ num_experts=128
12
+ topk=8
13
+ experts_per_rank=16
14
+ expert_params=18874368
15
+ heads=64
16
+ kv_groups=4
17
+ kv_channels=128
18
+ qk_layernorm=1
19
+ shared_ffn=0
20
+ shared_expert_overlap=0
21
+ mla=0
22
+ dense_first_layers=0
23
+ moe_layers=12
24
+ rotary_base=1000000
25
+ rotary_scaling=
26
+ qkv_bias=0
27
+ router_balancing=aux_loss
28
+ aux_loss_coeff=1e-3
29
+ router_pre_softmax=0
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=65536
39
+ mean_tokens_per_expert=4096
40
+ slot_policy=matched
41
+ slots_per_rank=16
42
+ capacity_mode=static
43
+ capacity=2.00
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
tierb_study_20260826_144644/echo/tierb/tierb_gbs8/QWEN3_235B__echo__matched__seq8192__gbs8__cap2p00/rep2/train.log ADDED
@@ -0,0 +1,1234 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [16:02:18] WARN: bracing-regions-through-kernels lacks --moe-router-fusion; QWEN3_235B runs unfused (timing only, routing identical)
2
+ ==============================================================
3
+ QWEN3_235B / echo (slots/rank=16 policy=matched)
4
+ E=128 K=8 H'=1536 m=16
5
+ L=12 seq=8192 gbs=8 (1 microbatches)
6
+ cap_mode=static cap=2.00 cg=1 recompute=none
7
+ memory_dump=none snapshots=/workspace/tierb_study_20260826_144644/echo/tierb/tierb_gbs8/QWEN3_235B__echo__matched__seq8192__gbs8__cap2p00/rep2/memory_snapshots/snapshot.pickle
8
+ repo=/workspace/Megatron-MoE-Echo @ 9a236793
9
+ ==============================================================
10
+ W0826 16:02:19.875000 172070 torch/distributed/run.py:766]
11
+ W0826 16:02:19.875000 172070 torch/distributed/run.py:766] *****************************************
12
+ W0826 16:02:19.875000 172070 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
13
+ W0826 16:02:19.875000 172070 torch/distributed/run.py:766] *****************************************
14
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
15
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
16
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
17
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
18
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
19
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
20
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
21
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
22
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
23
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
24
+ 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
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
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
28
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
29
+ using world size: 8, data-parallel size: 8, context-parallel size: 1, hierarchical context-parallel sizes: None, tensor-model-parallel size: 1, pipeline-model-parallel size: 1
30
+ WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer
31
+ Number of virtual stages per pipeline stage: None
32
+ accumulate and all-reduce gradients in fp32 for bfloat16 data type.
33
+ using torch.bfloat16 for parameters ...
34
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
35
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
36
+ ------------------------ arguments ------------------------
37
+ account_for_embedding_in_pipeline_split ......... False
38
+ account_for_loss_in_pipeline_split .............. False
39
+ accumulate_allreduce_grads_in_fp32 .............. True
40
+ adam_beta1 ...................................... 0.9
41
+ adam_beta2 ...................................... 0.999
42
+ adam_eps ........................................ 1e-08
43
+ add_bias_linear ................................. False
44
+ add_position_embedding .......................... False
45
+ add_qkv_bias .................................... False
46
+ adlr_autoresume ................................. False
47
+ adlr_autoresume_interval ........................ 1000
48
+ align_grad_reduce ............................... True
49
+ align_param_gather .............................. False
50
+ app_tag_run_name ................................ None
51
+ app_tag_run_version ............................. 0.0.0
52
+ apply_layernorm_1p .............................. False
53
+ apply_query_key_layer_scaling ................... False
54
+ apply_residual_connection_post_layernorm ........ False
55
+ apply_rope_fusion ............................... True
56
+ async_save ...................................... None
57
+ async_tensor_model_parallel_allreduce ........... True
58
+ attention_backend ............................... AttnBackend.auto
59
+ attention_dropout ............................... 0.0
60
+ attention_softmax_in_fp32 ....................... False
61
+ auto_detect_ckpt_format ......................... False
62
+ barrier_with_L1_time ............................ True
63
+ bert_binary_head ................................ True
64
+ bert_embedder_type .............................. megatron
65
+ bert_load ....................................... None
66
+ bf16 ............................................ True
67
+ bias_dropout_fusion ............................. True
68
+ bias_gelu_fusion ................................ False
69
+ bias_swiglu_fusion .............................. True
70
+ biencoder_projection_dim ........................ 0
71
+ biencoder_shared_query_context_model ............ False
72
+ block_data_path ................................. None
73
+ calc_ft_timeouts ................................ False
74
+ calculate_per_token_loss ........................ False
75
+ check_for_large_grads ........................... False
76
+ check_for_nan_in_loss_and_grad .................. False
77
+ check_for_spiky_loss ............................ False
78
+ check_weight_hash_across_dp_replicas_interval ... None
79
+ ckpt_assume_constant_structure .................. False
80
+ ckpt_convert_format ............................. None
81
+ ckpt_convert_save ............................... None
82
+ ckpt_convert_update_legacy_dist_opt_format ...... False
83
+ ckpt_format ..................................... torch_dist
84
+ ckpt_fully_parallel_load ........................ False
85
+ ckpt_fully_parallel_save ........................ True
86
+ ckpt_fully_parallel_save_deprecated ............. False
87
+ ckpt_step ....................................... None
88
+ classes_fraction ................................ 1.0
89
+ clip_grad ....................................... 1.0
90
+ clone_scatter_output_in_embedding ............... True
91
+ config_logger_dir ...............................
92
+ consumed_train_samples .......................... 0
93
+ consumed_valid_samples .......................... 0
94
+ context_parallel_size ........................... 1
95
+ cp_comm_type .................................... ['p2p']
96
+ create_attention_mask_in_dataloader ............. True
97
+ cross_entropy_fusion_impl ....................... native
98
+ cross_entropy_loss_fusion ....................... False
99
+ cuda_graph_scope ................................ full_iteration
100
+ cuda_graph_warmup_steps ......................... 3
101
+ data_args_path .................................. None
102
+ data_cache_path ................................. None
103
+ data_parallel_random_init ....................... False
104
+ data_parallel_sharding_strategy ................. no_shard
105
+ data_parallel_size .............................. 8
106
+ data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document']
107
+ data_per_class_fraction ......................... 1.0
108
+ data_sharding ................................... True
109
+ dataloader_type ................................. single
110
+ ddp_average_in_collective ....................... False
111
+ ddp_bucket_size ................................. None
112
+ ddp_num_buckets ................................. None
113
+ ddp_pad_buckets_for_high_nccl_busbw ............. False
114
+ decoder_first_pipeline_num_layers ............... None
115
+ decoder_last_pipeline_num_layers ................ None
116
+ decoder_num_layers .............................. None
117
+ decoder_seq_length .............................. None
118
+ decoupled_lr .................................... None
119
+ decoupled_min_lr ................................ None
120
+ decrease_batch_size_if_needed ................... False
121
+ defer_embedding_wgrad_compute ................... False
122
+ delay_wgrad_compute ............................. False
123
+ deprecated_use_mcore_models ..................... True
124
+ deterministic_mode .............................. False
125
+ dino_bottleneck_size ............................ 256
126
+ dino_freeze_last_layer .......................... 1
127
+ dino_head_hidden_size ........................... 2048
128
+ dino_local_crops_number ......................... 10
129
+ dino_local_img_size ............................. 96
130
+ dino_norm_last_layer ............................ False
131
+ dino_teacher_temp ............................... 0.07
132
+ dino_warmup_teacher_temp ........................ 0.04
133
+ dino_warmup_teacher_temp_epochs ................. 30
134
+ disable_bf16_reduced_precision_matmul ........... False
135
+ disable_mamba_mem_eff_path ...................... False
136
+ disable_straggler_on_startup .................... False
137
+ dist_ckpt_format_deprecated ..................... None
138
+ dist_ckpt_strictness ............................ assume_ok_unexpected
139
+ distribute_saved_activations .................... False
140
+ distributed_backend ............................. nccl
141
+ distributed_timeout_minutes ..................... 10
142
+ embedding_init_method_std ....................... None
143
+ embedding_path .................................. None
144
+ empty_unused_memory_level ....................... 0
145
+ enable_cuda_graph ............................... True
146
+ enable_experimental ............................. False
147
+ enable_ft_package ............................... False
148
+ enable_gloo_process_groups ...................... True
149
+ enable_msc ...................................... True
150
+ enable_one_logger ............................... True
151
+ encoder_num_layers .............................. 12
152
+ encoder_seq_length .............................. 8192
153
+ end_weight_decay ................................ 0.1
154
+ eod_mask_loss ................................... False
155
+ error_injection_rate ............................ 0
156
+ error_injection_type ............................ transient_error
157
+ eval_interval ................................... 1000000
158
+ eval_iters ...................................... 0
159
+ evidence_data_path .............................. None
160
+ exit_duration_in_mins ........................... None
161
+ exit_interval ................................... None
162
+ exit_on_missing_checkpoint ...................... False
163
+ exit_signal_handler ............................. False
164
+ exp_avg_dtype ................................... torch.float32
165
+ exp_avg_sq_dtype ................................ torch.float32
166
+ expert_model_parallel_size ...................... 8
167
+ expert_tensor_parallel_size ..................... 1
168
+ export_force_local_attention .................... False
169
+ export_kd_cfg ................................... None
170
+ export_kd_teacher_ckpt_format ................... None
171
+ export_kd_teacher_load .......................... None
172
+ export_kv_cache_quant ........................... False
173
+ export_legacy_megatron .......................... False
174
+ export_model_type ............................... GPTModel
175
+ export_moe_apply_probs_on_input ................. False
176
+ export_qk_l2_norm ............................... False
177
+ export_quant_cfg ................................ None
178
+ export_real_quant_cfg ........................... None
179
+ export_te_mcore_model ........................... False
180
+ external_cuda_graph ............................. False
181
+ ffn_hidden_size ................................. 12288
182
+ finetune ........................................ False
183
+ finetune_data_split ............................. train
184
+ finetune_hf_dataset ............................. None
185
+ first_last_layers_bf16 .......................... False
186
+ flash_decode .................................... False
187
+ fp16 ............................................ False
188
+ fp16_lm_cross_entropy ........................... False
189
+ fp32_residual_connection ........................ False
190
+ fp8 ............................................. e4m3
191
+ fp8_amax_compute_algo ........................... most_recent
192
+ fp8_amax_history_len ............................ 1
193
+ fp8_interval .................................... 1
194
+ fp8_margin ...................................... 0
195
+ fp8_param_gather ................................ False
196
+ fp8_recipe ...................................... mxfp8
197
+ fp8_wgrad ....................................... True
198
+ fsdp_double_buffer .............................. False
199
+ global_batch_size ............................... 8
200
+ grad_reduce_in_bf16 ............................. False
201
+ gradient_accumulation_fusion .................... True
202
+ gradient_reduce_div_fusion ...................... True
203
+ group_query_attention ........................... True
204
+ head_lr_mult .................................... 1.0
205
+ heterogeneous_layers_config_encoded_json ........ None
206
+ heterogeneous_layers_config_path ................ None
207
+ hidden_dropout .................................. 0.0
208
+ hidden_size ..................................... 4096
209
+ hierarchical_context_parallel_sizes ............. None
210
+ high_priority_stream_groups ..................... []
211
+ hybrid_attention_ratio .......................... 0.0
212
+ hybrid_mlp_ratio ................................ 0.0
213
+ hybrid_override_pattern ......................... None
214
+ hysteresis ...................................... 2
215
+ ict_head_size ................................... None
216
+ ict_load ........................................ None
217
+ img_h ........................................... 224
218
+ img_w ........................................... 224
219
+ indexer_batch_size .............................. 128
220
+ indexer_log_interval ............................ 1000
221
+ inference_batch_times_seqlen_threshold .......... -1
222
+ inference_dynamic_batching ...................... False
223
+ inference_dynamic_batching_buffer_guaranteed_fraction 0.2
224
+ inference_dynamic_batching_buffer_overflow_factor None
225
+ inference_dynamic_batching_buffer_size_gb ....... 40.0
226
+ inference_dynamic_batching_chunk_size ........... 256
227
+ inference_dynamic_batching_max_requests_override None
228
+ inference_dynamic_batching_max_tokens_override .. None
229
+ inference_dynamic_batching_num_cuda_graphs ...... 16
230
+ inference_max_batch_size ........................ 8
231
+ inference_max_seq_length ........................ 2560
232
+ inference_rng_tracker ........................... False
233
+ init_method_std ................................. 0.01
234
+ init_method_xavier_uniform ...................... False
235
+ init_model_with_meta_device ..................... False
236
+ initial_loss_scale .............................. 4294967296
237
+ inprocess_active_world_size ..................... 8
238
+ inprocess_barrier_timeout ....................... 120
239
+ inprocess_completion_timeout .................... 120
240
+ inprocess_empty_cuda_cache ...................... False
241
+ inprocess_granularity ........................... node
242
+ inprocess_hard_timeout .......................... 90
243
+ inprocess_heartbeat_interval .................... 30
244
+ inprocess_heartbeat_timeout ..................... 60
245
+ inprocess_last_call_wait ........................ 1
246
+ inprocess_max_iterations ........................ None
247
+ inprocess_monitor_process_interval .............. 1.0
248
+ inprocess_monitor_thread_interval ............... 1.0
249
+ inprocess_progress_watchdog_interval ............ 1.0
250
+ inprocess_restart ............................... False
251
+ inprocess_soft_timeout .......................... 60
252
+ inprocess_termination_grace_time ................ 1
253
+ is_hybrid_model ................................. False
254
+ iter_per_epoch .................................. 1250
255
+ iterations_to_skip .............................. []
256
+ keep_fp8_transpose_cache_when_using_custom_fsdp . False
257
+ kitchen_config_file ............................. None
258
+ kitchen_recipe_number ........................... None
259
+ kv_channels ..................................... 128
260
+ kv_lora_rank .................................... 32
261
+ lazy_mpu_init ................................... None
262
+ load ............................................ None
263
+ load_main_params_from_ckpt ...................... None
264
+ load_model_opt_format ........................... False
265
+ local_rank ...................................... 0
266
+ log_energy ...................................... False
267
+ log_interval .................................... 1
268
+ log_loss_scale_to_tensorboard ................... True
269
+ log_memory_to_tensorboard ....................... False
270
+ log_num_zeros_in_grad ........................... False
271
+ log_params_norm ................................. False
272
+ log_progress .................................... False
273
+ log_straggler ................................... False
274
+ log_throughput .................................. True
275
+ log_timers_to_tensorboard ....................... False
276
+ log_validation_ppl_to_tensorboard ............... False
277
+ log_world_size_to_tensorboard ................... False
278
+ logging_level ................................... None
279
+ loss_scale ...................................... None
280
+ loss_scale_window ............................... 1000
281
+ lr .............................................. 0.0001
282
+ lr_decay_iters .................................. 8
283
+ lr_decay_samples ................................ None
284
+ lr_decay_style .................................. cosine
285
+ lr_warmup_fraction .............................. None
286
+ lr_warmup_init .................................. 0.0
287
+ lr_warmup_iters ................................. 5
288
+ lr_warmup_samples ............................... 0
289
+ lr_wsd_decay_iters .............................. None
290
+ lr_wsd_decay_samples ............................ None
291
+ lr_wsd_decay_style .............................. exponential
292
+ main_grads_dtype ................................ torch.float32
293
+ main_params_dtype ............................... torch.float32
294
+ make_vocab_size_divisible_by .................... 128
295
+ mamba_head_dim .................................. 64
296
+ mamba_num_groups ................................ 8
297
+ mamba_num_heads ................................. None
298
+ mamba_state_dim ................................. 128
299
+ manual_gc ....................................... False
300
+ manual_gc_eval .................................. True
301
+ manual_gc_interval .............................. 0
302
+ mask_factor ..................................... 1.0
303
+ mask_prob ....................................... 0.15
304
+ mask_type ....................................... random
305
+ masked_softmax_fusion ........................... False
306
+ max_position_embeddings ......................... 32768
307
+ max_tokens_to_oom ............................... 12000
308
+ memory_snapshot_path ............................ snapshot.pickle
309
+ merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt
310
+ micro_batch_size ................................ 1
311
+ microbatch_group_size_per_vp_stage .............. None
312
+ mid_level_dataset_surplus ....................... 0.005
313
+ min_loss_scale .................................. 1.0
314
+ min_lr .......................................... 1e-05
315
+ mlp_chunks_for_prefill .......................... 1
316
+ mmap_bin_files .................................. True
317
+ mock_data ....................................... False
318
+ moe_apply_probs_on_input ........................ False
319
+ moe_aux_loss_coeff .............................. 0.001
320
+ moe_deepep_num_sms .............................. 20
321
+ moe_echo_enable_random_offloading ............... False
322
+ moe_echo_expert_dispatch_overlap ................ True
323
+ moe_echo_expert_dispatcher_type ................. hybridep
324
+ moe_echo_recompute_expert_dispatch .............. False
325
+ moe_enable_deepep ............................... False
326
+ moe_enable_echo ................................. True
327
+ moe_enable_hybridep ............................. True
328
+ moe_expert_capacity_factor ...................... None
329
+ moe_extended_tp ................................. False
330
+ moe_ffn_hidden_size ............................. 1536
331
+ moe_grouped_gemm ................................ True
332
+ moe_hybridep_num_sms ............................ 16
333
+ moe_input_jitter_eps ............................ None
334
+ moe_layer_freq .................................. 1
335
+ moe_layer_recompute ............................. False
336
+ moe_num_echo_experts ............................ 128
337
+ moe_pad_expert_input_to_capacity ................ False
338
+ moe_per_layer_logging ........................... False
339
+ moe_permute_fusion .............................. False
340
+ moe_received_token_capacity ..................... 2.0
341
+ moe_router_bias_update_rate ..................... 0.001
342
+ moe_router_dtype ................................ fp32
343
+ moe_router_enable_expert_bias ................... False
344
+ moe_router_force_load_balancing ................. False
345
+ moe_router_group_topk ........................... None
346
+ moe_router_load_balancing_type .................. aux_loss
347
+ moe_router_num_groups ........................... None
348
+ moe_router_padding_for_fp8 ...................... False
349
+ moe_router_pre_softmax .......................... False
350
+ moe_router_score_function ....................... softmax
351
+ moe_router_topk ................................. 8
352
+ moe_router_topk_scaling_factor .................. None
353
+ moe_shared_expert_intermediate_size ............. None
354
+ moe_shared_expert_overlap ....................... False
355
+ moe_token_dispatcher_type ....................... flex
356
+ moe_token_drop_policy ........................... probs
357
+ moe_upcycling_granularity ....................... 1
358
+ moe_use_device_initiated_grouped_gemm ........... True
359
+ moe_use_legacy_grouped_gemm ..................... False
360
+ moe_use_upcycling ............................... False
361
+ moe_z_loss_coeff ................................ None
362
+ mrope_section ................................... None
363
+ mscale .......................................... 1.0
364
+ mscale_all_dim .................................. 0.0
365
+ mtp_loss_scaling_factor ......................... 0.1
366
+ mtp_num_layers .................................. None
367
+ multi_latent_attention .......................... False
368
+ nccl_all_reduce_for_prefill ..................... False
369
+ nccl_communicator_config_path ................... None
370
+ nccl_ub ......................................... False
371
+ no_load_optim ................................... None
372
+ no_load_rng ..................................... None
373
+ no_persist_layer_norm ........................... False
374
+ no_rope_freq .................................... None
375
+ no_save_optim ................................... None
376
+ no_save_rng ..................................... None
377
+ non_persistent_ckpt_type ........................ None
378
+ non_persistent_global_ckpt_dir .................. None
379
+ non_persistent_local_ckpt_algo .................. fully_parallel
380
+ non_persistent_local_ckpt_dir ................... None
381
+ non_persistent_save_interval .................... None
382
+ norm_epsilon .................................... 1e-06
383
+ normalization ................................... RMSNorm
384
+ num_attention_heads ............................. 64
385
+ num_channels .................................... 3
386
+ num_classes ..................................... 1000
387
+ num_dataset_builder_threads ..................... 1
388
+ num_distributed_optimizer_instances ............. 1
389
+ num_experts ..................................... 128
390
+ num_layers ...................................... 12
391
+ num_layers_at_end_in_bf16 ....................... 1
392
+ num_layers_at_start_in_bf16 ..................... 1
393
+ num_layers_per_virtual_pipeline_stage ........... None
394
+ num_query_groups ................................ 4
395
+ num_virtual_stages_per_pipeline_rank ............ None
396
+ num_workers ..................................... 2
397
+ object_storage_cache_path ....................... None
398
+ one_logger_async ................................ False
399
+ one_logger_project .............................. megatron-lm
400
+ one_logger_run_name ............................. None
401
+ onnx_safe ....................................... None
402
+ openai_gelu ..................................... False
403
+ optimizer ....................................... adam
404
+ optimizer_cpu_offload ........................... False
405
+ optimizer_offload_fraction ...................... 1.0
406
+ output_bert_embeddings .......................... False
407
+ overlap_cpu_optimizer_d2h_h2d ................... False
408
+ overlap_grad_reduce ............................. True
409
+ overlap_moe_expert_parallel_comm ................ False
410
+ overlap_p2p_comm ................................ False
411
+ overlap_p2p_comm_warmup_flush ................... False
412
+ overlap_param_gather ............................ True
413
+ overlap_param_gather_with_optimizer_step ........ False
414
+ override_opt_param_scheduler .................... False
415
+ params_dtype .................................... torch.bfloat16
416
+ patch_dim ....................................... 16
417
+ per_split_data_args_path ........................ None
418
+ perform_initialization .......................... True
419
+ pin_cpu_grads ................................... True
420
+ pin_cpu_params .................................. True
421
+ pipeline_model_parallel_comm_backend ............ None
422
+ pipeline_model_parallel_layout .................. None
423
+ pipeline_model_parallel_size .................... 1
424
+ position_embedding_type ......................... rope
425
+ pretrained_checkpoint ........................... None
426
+ profile ......................................... False
427
+ profile_ranks ................................... [0]
428
+ profile_step_end ................................ 12
429
+ profile_step_start .............................. 10
430
+ q_lora_rank ..................................... None
431
+ qk_head_dim ..................................... 128
432
+ qk_l2_norm ...................................... False
433
+ qk_layernorm .................................... True
434
+ qk_pos_emb_head_dim ............................. 64
435
+ query_in_block_prob ............................. 0.1
436
+ rampup_batch_size ............................... None
437
+ rank ............................................ 0
438
+ recompute_granularity ........................... None
439
+ recompute_method ................................ None
440
+ recompute_modules ............................... None
441
+ recompute_num_layers ............................ None
442
+ record_memory_history ........................... False
443
+ relative_attention_max_distance ................. 128
444
+ relative_attention_num_buckets .................. 32
445
+ replication ..................................... False
446
+ replication_factor .............................. 2
447
+ replication_jump ................................ None
448
+ rerun_mode ...................................... validate_results
449
+ reset_attention_mask ............................ False
450
+ reset_position_ids .............................. False
451
+ result_rejected_tracker_filename ................ None
452
+ retriever_report_topk_accuracies ................ []
453
+ retriever_score_scaling ......................... False
454
+ retriever_seq_length ............................ 256
455
+ retro_add_retriever ............................. False
456
+ retro_attention_gate ............................ 1
457
+ retro_cyclic_train_iters ........................ None
458
+ retro_encoder_attention_dropout ................. 0.1
459
+ retro_encoder_hidden_dropout .................... 0.1
460
+ retro_encoder_layers ............................ 2
461
+ retro_num_neighbors ............................. 2
462
+ retro_num_retrieved_chunks ...................... 2
463
+ retro_project_dir ............................... None
464
+ retro_verify_neighbor_count ..................... True
465
+ reuse_grad_buf_for_mxfp8_param_ag ............... False
466
+ rope_scaling_factor ............................. 8.0
467
+ rope_type ....................................... None
468
+ rotary_base ..................................... 1000000
469
+ rotary_interleaved .............................. False
470
+ rotary_percent .................................. 1.0
471
+ rotary_scaling_factor ........................... 1.0
472
+ rotary_seq_len_interpolation_factor ............. None
473
+ run_workload_inspector_server ................... False
474
+ sample_rate ..................................... 1.0
475
+ save ............................................ None
476
+ save_interval ................................... None
477
+ save_retain_interval ............................ None
478
+ scatter_gather_tensors_in_pipeline .............. True
479
+ seed ............................................ 1234
480
+ seq_length ...................................... 8192
481
+ sequence_parallel ............................... False
482
+ sft ............................................. False
483
+ sft_tokenizer_prompt_format ..................... nemotron-h-aligned
484
+ sgd_momentum .................................... 0.9
485
+ short_seq_prob .................................. 0.1
486
+ skip_train ...................................... False
487
+ skipped_train_samples ........................... 0
488
+ spec ............................................ None
489
+ split ........................................... 100,0,0
490
+ squared_relu .................................... False
491
+ start_weight_decay .............................. 0.1
492
+ straggler_ctrlr_port ............................ 65535
493
+ straggler_minmax_count .......................... 1
494
+ suggested_communication_unit_size ............... None
495
+ swiglu .......................................... True
496
+ swin_backbone_type .............................. tiny
497
+ symmetric_ar_type ............................... None
498
+ te_rng_tracker .................................. True
499
+ tensor_model_parallel_size ...................... 1
500
+ tensorboard_dir ................................. None
501
+ tensorboard_log_interval ........................ 1
502
+ tensorboard_queue_size .......................... 1000
503
+ test_data_path .................................. None
504
+ test_mode ....................................... False
505
+ tiktoken_num_special_tokens ..................... 1000
506
+ tiktoken_pattern ................................ None
507
+ tiktoken_special_tokens ......................... None
508
+ timing_log_level ................................ 0
509
+ timing_log_option ............................... minmax
510
+ titles_data_path ................................ None
511
+ tokenizer_model ................................. None
512
+ tokenizer_type .................................. GPT2BPETokenizer
513
+ torch_fsdp2_reshard_after_forward ............... True
514
+ tp_comm_bootstrap_backend ....................... nccl
515
+ tp_comm_bulk_dgrad .............................. True
516
+ tp_comm_bulk_wgrad .............................. True
517
+ tp_comm_overlap ................................. False
518
+ tp_comm_overlap_ag .............................. True
519
+ tp_comm_overlap_cfg ............................. None
520
+ tp_comm_overlap_rs .............................. True
521
+ tp_comm_overlap_rs_dgrad ........................ False
522
+ tp_comm_split_ag ................................ True
523
+ tp_comm_split_rs ................................ True
524
+ train_data_path ................................. None
525
+ train_iters ..................................... 12
526
+ train_samples ................................... None
527
+ train_sync_interval ............................. None
528
+ transformer_impl ................................ transformer_engine
529
+ transformer_pipeline_model_parallel_size ........ 1
530
+ untie_embeddings_and_output_weights ............. True
531
+ use_checkpoint_args ............................. False
532
+ use_checkpoint_opt_param_scheduler .............. False
533
+ use_cpu_initialization .......................... None
534
+ use_custom_fsdp ................................. False
535
+ use_dist_ckpt ................................... True
536
+ use_dist_ckpt_deprecated ........................ False
537
+ use_distributed_optimizer ....................... True
538
+ use_flash_attn .................................. False
539
+ use_legacy_models ............................... False
540
+ use_mp_args_from_checkpoint_args ................ False
541
+ use_one_sent_docs ............................... False
542
+ use_persistent_ckpt_worker ...................... False
543
+ use_precision_aware_optimizer ................... False
544
+ use_pytorch_profiler ............................ False
545
+ use_ring_exchange_p2p ........................... False
546
+ use_rope_scaling ................................ False
547
+ use_rotary_position_embeddings .................. False
548
+ use_sharp ....................................... False
549
+ use_tokenizer_model_from_checkpoint_args ........ True
550
+ use_torch_fsdp2 ................................. False
551
+ use_torch_optimizer_for_cpu_offload ............. False
552
+ use_tp_pp_dp_mapping ............................ False
553
+ v_head_dim ...................................... 128
554
+ valid_data_path ................................. None
555
+ variable_seq_lengths ............................ False
556
+ virtual_pipeline_model_parallel_size ............ None
557
+ vision_backbone_type ............................ vit
558
+ vision_pretraining .............................. False
559
+ vision_pretraining_type ......................... classify
560
+ vocab_extra_ids ................................. 0
561
+ vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json
562
+ vocab_size ...................................... None
563
+ wandb_exp_name ..................................
564
+ wandb_project ...................................
565
+ wandb_save_dir ..................................
566
+ weight_decay .................................... 0.1
567
+ weight_decay_incr_style ......................... constant
568
+ wgrad_deferral_limit ............................ 0
569
+ world_size ...................................... 8
570
+ yaml_cfg ........................................ None
571
+ -------------------- end of arguments ---------------------
572
+ INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1
573
+ > building GPT2BPETokenizer tokenizer ...
574
+ > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304)
575
+ WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS
576
+ > initializing torch distributed ...
577
+ [Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
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 [Gloo] Rank 1 is connected to 72 peer ranks. is connected to Expected number of connected peer ranks is : 77 peer ranks.
581
+ 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 7 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 0 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 1 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 [Gloo] Rank 5 is connected to 47 is connected to peer ranks. 7Expected number of connected peer ranks is : peer ranks. 7Expected number of connected peer ranks is :
590
+ 7
591
+ [Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
592
+ [Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
593
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
594
+ > initialized tensor model parallel with size 1
595
+ > initialized pipeline model parallel with size 1
596
+ > setting random seeds to 1234 ...
597
+ [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
598
+
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
+ [rank4]:[W826 16:02:37.029625998 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.
605
+ [rank5]:[W826 16:02:37.030324870 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
+ [rank6]:[W826 16:02:37.032559023 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
+ [rank2]:[W826 16:02:37.036812469 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
+ [rank7]:[W826 16:02:37.036826843 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
+ [rank3]:[W826 16:02:37.036831730 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 16:02:37.036861610 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
+ > compiling dataset index builder ...
612
+ make: Entering directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
613
+ make: Nothing to be done for 'default'.
614
+ make: Leaving directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
615
+ >>> done with dataset index builder. Compilation time: 0.036 seconds
616
+ WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations.
617
+ > compiling and loading fused kernels ...
618
+ [rank0]:[W826 16:02:38.212575228 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.
619
+ >>> done with compiling and loading fused kernels. Compilation time: 3.598 seconds
620
+ time to initialize megatron (seconds): 6.921
621
+ [after megatron is initialized] datetime: 2026-08-26 16:02:43
622
+ building GPT model ...
623
+ /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.
624
+ warnings.warn(
625
+ /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.
626
+ warnings.warn(
627
+ /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.
628
+ warnings.warn(
629
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
630
+ warnings.warn(
631
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
632
+ warnings.warn(
633
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
634
+ warnings.warn(
635
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
636
+ warnings.warn(
637
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
638
+ warnings.warn(
639
+ > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 4898003968
640
+ 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)
641
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 14
642
+ Params for bucket 1 (206045184 elements, 206045184 padded size):
643
+ module.output_layer.weight
644
+ Params for bucket 2 (71840000 elements, 71840000 padded size):
645
+ module.decoder.layers.11.self_attention.k_layernorm.weight
646
+ module.decoder.layers.11.self_attention.linear_qkv.layer_norm_weight
647
+ module.decoder.layers.11.self_attention.linear_proj.weight
648
+ module.decoder.final_layernorm.weight
649
+ module.decoder.layers.11.pre_mlp_layernorm.weight
650
+ module.decoder.layers.11.self_attention.q_layernorm.weight
651
+ module.decoder.layers.11.mlp.router.weight
652
+ module.decoder.layers.11.self_attention.linear_qkv.weight
653
+ Params for bucket 3 (71835904 elements, 71835904 padded size):
654
+ module.decoder.layers.10.self_attention.q_layernorm.weight
655
+ module.decoder.layers.10.pre_mlp_layernorm.weight
656
+ module.decoder.layers.10.self_attention.linear_qkv.weight
657
+ module.decoder.layers.10.self_attention.linear_qkv.layer_norm_weight
658
+ module.decoder.layers.10.self_attention.linear_proj.weight
659
+ module.decoder.layers.10.mlp.router.weight
660
+ module.decoder.layers.10.self_attention.k_layernorm.weight
661
+ Params for bucket 4 (71835904 elements, 71835904 padded size):
662
+ module.decoder.layers.9.self_attention.q_layernorm.weight
663
+ module.decoder.layers.9.self_attention.linear_qkv.weight
664
+ module.decoder.layers.9.pre_mlp_layernorm.weight
665
+ module.decoder.layers.9.self_attention.linear_proj.weight
666
+ module.decoder.layers.9.mlp.router.weight
667
+ module.decoder.layers.9.self_attention.k_layernorm.weight
668
+ module.decoder.layers.9.self_attention.linear_qkv.layer_norm_weight
669
+ Params for bucket 5 (71835904 elements, 71835904 padded size):
670
+ module.decoder.layers.8.mlp.router.weight
671
+ module.decoder.layers.8.self_attention.q_layernorm.weight
672
+ module.decoder.layers.8.self_attention.linear_qkv.weight
673
+ module.decoder.layers.8.self_attention.k_layernorm.weight
674
+ module.decoder.layers.8.pre_mlp_layernorm.weight
675
+ module.decoder.layers.8.self_attention.linear_qkv.layer_norm_weight
676
+ module.decoder.layers.8.self_attention.linear_proj.weight
677
+ Params for bucket 6 (71835904 elements, 71835904 padded size):
678
+ module.decoder.layers.7.mlp.router.weight
679
+ module.decoder.layers.7.pre_mlp_layernorm.weight
680
+ module.decoder.layers.7.self_attention.linear_qkv.weight
681
+ module.decoder.layers.7.self_attention.q_layernorm.weight
682
+ module.decoder.layers.7.self_attention.k_layernorm.weight
683
+ module.decoder.layers.7.self_attention.linear_qkv.layer_norm_weight
684
+ module.decoder.layers.7.self_attention.linear_proj.weight
685
+ Params for bucket 7 (71835904 elements, 71835904 padded size):
686
+ module.decoder.layers.6.pre_mlp_layernorm.weight
687
+ module.decoder.layers.6.self_attention.linear_qkv.weight
688
+ module.decoder.layers.6.self_attention.q_layernorm.weight
689
+ module.decoder.layers.6.self_attention.linear_proj.weight
690
+ module.decoder.layers.6.mlp.router.weight
691
+ module.decoder.layers.6.self_attention.k_layernorm.weight
692
+ module.decoder.layers.6.self_attention.linear_qkv.layer_norm_weight
693
+ Params for bucket 8 (71835904 elements, 71835904 padded size):
694
+ module.decoder.layers.5.pre_mlp_layernorm.weight
695
+ module.decoder.layers.5.self_attention.q_layernorm.weight
696
+ module.decoder.layers.5.self_attention.linear_qkv.layer_norm_weight
697
+ module.decoder.layers.5.mlp.router.weight
698
+ module.decoder.layers.5.self_attention.k_layernorm.weight
699
+ module.decoder.layers.5.self_attention.linear_qkv.weight
700
+ module.decoder.layers.5.self_attention.linear_proj.weight
701
+ Params for bucket 9 (71835904 elements, 71835904 padded size):
702
+ module.decoder.layers.4.self_attention.q_layernorm.weight
703
+ module.decoder.layers.4.pre_mlp_layernorm.weight
704
+ module.decoder.layers.4.self_attention.linear_qkv.weight
705
+ module.decoder.layers.4.mlp.router.weight
706
+ module.decoder.layers.4.self_attention.k_layernorm.weight
707
+ module.decoder.layers.4.self_attention.linear_qkv.layer_norm_weight
708
+ module.decoder.layers.4.self_attention.linear_proj.weight
709
+ Params for bucket 10 (71835904 elements, 71835904 padded size):
710
+ module.decoder.layers.3.self_attention.q_layernorm.weight
711
+ module.decoder.layers.3.pre_mlp_layernorm.weight
712
+ module.decoder.layers.3.self_attention.linear_qkv.weight
713
+ module.decoder.layers.3.self_attention.linear_proj.weight
714
+ module.decoder.layers.3.mlp.router.weight
715
+ module.decoder.layers.3.self_attention.k_layernorm.weight
716
+ module.decoder.layers.3.self_attention.linear_qkv.layer_norm_weight
717
+ Params for bucket 11 (71835904 elements, 71835904 padded size):
718
+ module.decoder.layers.2.mlp.router.weight
719
+ module.decoder.layers.2.self_attention.linear_qkv.weight
720
+ module.decoder.layers.2.pre_mlp_layernorm.weight
721
+ module.decoder.layers.2.self_attention.q_layernorm.weight
722
+ module.decoder.layers.2.self_attention.k_layernorm.weight
723
+ module.decoder.layers.2.self_attention.linear_qkv.layer_norm_weight
724
+ module.decoder.layers.2.self_attention.linear_proj.weight
725
+ Params for bucket 12 (71835904 elements, 71835904 padded size):
726
+ module.decoder.layers.1.self_attention.q_layernorm.weight
727
+ module.decoder.layers.1.self_attention.k_layernorm.weight
728
+ module.decoder.layers.1.self_attention.linear_qkv.layer_norm_weight
729
+ module.decoder.layers.1.self_attention.linear_proj.weight
730
+ module.decoder.layers.1.mlp.router.weight
731
+ module.decoder.layers.1.pre_mlp_layernorm.weight
732
+ module.decoder.layers.1.self_attention.linear_qkv.weight
733
+ Params for bucket 13 (71835904 elements, 71835904 padded size):
734
+ module.decoder.layers.0.pre_mlp_layernorm.weight
735
+ module.decoder.layers.0.self_attention.linear_qkv.layer_norm_weight
736
+ module.decoder.layers.0.self_attention.linear_proj.weight
737
+ module.decoder.layers.0.mlp.router.weight
738
+ module.decoder.layers.0.self_attention.linear_qkv.weight
739
+ module.decoder.layers.0.self_attention.q_layernorm.weight
740
+ module.decoder.layers.0.self_attention.k_layernorm.weight
741
+ Params for bucket 14 (206045184 elements, 206045184 padded size):
742
+ module.embedding.word_embeddings.weight
743
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 76
744
+ Params for bucket 1 (44040192 elements, 44040192 padded size):
745
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight14
746
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight12
747
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight10
748
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight15
749
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight11
750
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight13
751
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight9
752
+ Params for bucket 2 (44040192 elements, 44040192 padded size):
753
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight8
754
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight4
755
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight6
756
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight2
757
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight7
758
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight5
759
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight3
760
+ Params for bucket 3 (50331648 elements, 50331648 padded size):
761
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight13
762
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight15
763
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight0
764
+ module.decoder.layers.11.mlp.experts.linear_fc2.weight1
765
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight14
766
+ Params for bucket 4 (50331648 elements, 50331648 padded size):
767
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight12
768
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight11
769
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight10
770
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight9
771
+ Params for bucket 5 (50331648 elements, 50331648 padded size):
772
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight8
773
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight7
774
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight5
775
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight6
776
+ Params for bucket 6 (50331648 elements, 50331648 padded size):
777
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight4
778
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight1
779
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight2
780
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight3
781
+ Params for bucket 7 (44040192 elements, 44040192 padded size):
782
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight13
783
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight15
784
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight11
785
+ module.decoder.layers.11.mlp.experts.linear_fc1.weight0
786
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight14
787
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight12
788
+ Params for bucket 8 (44040192 elements, 44040192 padded size):
789
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight7
790
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight9
791
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight5
792
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight10
793
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight6
794
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight8
795
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight4
796
+ Params for bucket 9 (50331648 elements, 50331648 padded size):
797
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight3
798
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight1
799
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight14
800
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight2
801
+ module.decoder.layers.10.mlp.experts.linear_fc2.weight0
802
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight15
803
+ Params for bucket 10 (50331648 elements, 50331648 padded size):
804
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight13
805
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight10
806
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight11
807
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight12
808
+ Params for bucket 11 (50331648 elements, 50331648 padded size):
809
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight9
810
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight6
811
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight7
812
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight8
813
+ Params for bucket 12 (50331648 elements, 50331648 padded size):
814
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight5
815
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight3
816
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight4
817
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight2
818
+ Params for bucket 13 (44040192 elements, 44040192 padded size):
819
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight0
820
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight13
821
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight15
822
+ module.decoder.layers.10.mlp.experts.linear_fc1.weight1
823
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight14
824
+ Params for bucket 14 (44040192 elements, 44040192 padded size):
825
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight9
826
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight11
827
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight7
828
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight12
829
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight10
830
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight8
831
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight6
832
+ Params for bucket 15 (50331648 elements, 50331648 padded size):
833
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight5
834
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight3
835
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight1
836
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight4
837
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight2
838
+ module.decoder.layers.9.mlp.experts.linear_fc2.weight0
839
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight15
840
+ Params for bucket 16 (50331648 elements, 50331648 padded size):
841
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight14
842
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight11
843
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight13
844
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight12
845
+ Params for bucket 17 (50331648 elements, 50331648 padded size):
846
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight10
847
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight7
848
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight8
849
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight9
850
+ Params for bucket 18 (50331648 elements, 50331648 padded size):
851
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight6
852
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight4
853
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight5
854
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight3
855
+ Params for bucket 19 (44040192 elements, 44040192 padded size):
856
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight2
857
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight1
858
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight15
859
+ module.decoder.layers.9.mlp.experts.linear_fc1.weight0
860
+ Params for bucket 20 (44040192 elements, 44040192 padded size):
861
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight14
862
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight12
863
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight10
864
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight8
865
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight13
866
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight11
867
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight9
868
+ Params for bucket 21 (44040192 elements, 44040192 padded size):
869
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight6
870
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight4
871
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight2
872
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight7
873
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight3
874
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight5
875
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight1
876
+ Params for bucket 22 (44040192 elements, 44040192 padded size):
877
+ module.decoder.layers.8.mlp.experts.linear_fc2.weight0
878
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight15
879
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight13
880
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight14
881
+ Params for bucket 23 (50331648 elements, 50331648 padded size):
882
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight12
883
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight9
884
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight10
885
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight11
886
+ Params for bucket 24 (50331648 elements, 50331648 padded size):
887
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight8
888
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight6
889
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight5
890
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight7
891
+ Params for bucket 25 (50331648 elements, 50331648 padded size):
892
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight4
893
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight3
894
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight2
895
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight1
896
+ Params for bucket 26 (44040192 elements, 44040192 padded size):
897
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight14
898
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight12
899
+ module.decoder.layers.8.mlp.experts.linear_fc1.weight0
900
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight13
901
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight15
902
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight11
903
+ Params for bucket 27 (44040192 elements, 44040192 padded size):
904
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight10
905
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight6
906
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight8
907
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight4
908
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight7
909
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight9
910
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight5
911
+ Params for bucket 28 (50331648 elements, 50331648 padded size):
912
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight0
913
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight15
914
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight2
915
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight3
916
+ module.decoder.layers.7.mlp.experts.linear_fc2.weight1
917
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight14
918
+ Params for bucket 29 (50331648 elements, 50331648 padded size):
919
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight13
920
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight10
921
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight12
922
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight11
923
+ Params for bucket 30 (50331648 elements, 50331648 padded size):
924
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight9
925
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight6
926
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight7
927
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight8
928
+ Params for bucket 31 (50331648 elements, 50331648 padded size):
929
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight5
930
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight3
931
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight4
932
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight2
933
+ Params for bucket 32 (44040192 elements, 44040192 padded size):
934
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight1
935
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight14
936
+ module.decoder.layers.7.mlp.experts.linear_fc1.weight0
937
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight13
938
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight15
939
+ Params for bucket 33 (44040192 elements, 44040192 padded size):
940
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight12
941
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight10
942
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight8
943
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight6
944
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight9
945
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight11
946
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight7
947
+ Params for bucket 34 (50331648 elements, 50331648 padded size):
948
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight4
949
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight2
950
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight0
951
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight15
952
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight5
953
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight3
954
+ module.decoder.layers.6.mlp.experts.linear_fc2.weight1
955
+ Params for bucket 35 (50331648 elements, 50331648 padded size):
956
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight14
957
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight13
958
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight11
959
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight12
960
+ Params for bucket 36 (50331648 elements, 50331648 padded size):
961
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight10
962
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight9
963
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight8
964
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight7
965
+ Params for bucket 37 (50331648 elements, 50331648 padded size):
966
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight6
967
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight3
968
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight4
969
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight5
970
+ Params for bucket 38 (44040192 elements, 44040192 padded size):
971
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight2
972
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight0
973
+ module.decoder.layers.6.mlp.experts.linear_fc1.weight1
974
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight15
975
+ Params for bucket 39 (44040192 elements, 44040192 padded size):
976
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight11
977
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight13
978
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight9
979
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight14
980
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight10
981
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight12
982
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight8
983
+ Params for bucket 40 (44040192 elements, 44040192 padded size):
984
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight7
985
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight5
986
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight3
987
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight1
988
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight4
989
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight6
990
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight2
991
+ Params for bucket 41 (44040192 elements, 44040192 padded size):
992
+ module.decoder.layers.5.mlp.experts.linear_fc2.weight0
993
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight14
994
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight15
995
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight13
996
+ Params for bucket 42 (50331648 elements, 50331648 padded size):
997
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight12
998
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight10
999
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight11
1000
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight9
1001
+ Params for bucket 43 (50331648 elements, 50331648 padded size):
1002
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight8
1003
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight7
1004
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight5
1005
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight6
1006
+ Params for bucket 44 (50331648 elements, 50331648 padded size):
1007
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight4
1008
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight1
1009
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight2
1010
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight3
1011
+ Params for bucket 45 (44040192 elements, 44040192 padded size):
1012
+ module.decoder.layers.5.mlp.experts.linear_fc1.weight0
1013
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight13
1014
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight15
1015
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight11
1016
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight12
1017
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight14
1018
+ Params for bucket 46 (44040192 elements, 44040192 padded size):
1019
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight7
1020
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight9
1021
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight5
1022
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight10
1023
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight6
1024
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight8
1025
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight4
1026
+ Params for bucket 47 (50331648 elements, 50331648 padded size):
1027
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight3
1028
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight1
1029
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight14
1030
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight0
1031
+ module.decoder.layers.4.mlp.experts.linear_fc2.weight2
1032
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight15
1033
+ Params for bucket 48 (50331648 elements, 50331648 padded size):
1034
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight13
1035
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight10
1036
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight12
1037
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight11
1038
+ Params for bucket 49 (50331648 elements, 50331648 padded size):
1039
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight9
1040
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight6
1041
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight7
1042
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight8
1043
+ Params for bucket 50 (50331648 elements, 50331648 padded size):
1044
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight5
1045
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight3
1046
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight4
1047
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight2
1048
+ Params for bucket 51 (44040192 elements, 44040192 padded size):
1049
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight0
1050
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight14
1051
+ module.decoder.layers.4.mlp.experts.linear_fc1.weight1
1052
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight15
1053
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight13
1054
+ Params for bucket 52 (44040192 elements, 44040192 padded size):
1055
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight12
1056
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight10
1057
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight8
1058
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight6
1059
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight11
1060
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight9
1061
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight7
1062
+ Params for bucket 53 (50331648 elements, 50331648 padded size):
1063
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight4
1064
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight2
1065
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight0
1066
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight5
1067
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight1
1068
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight3
1069
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight15
1070
+ Params for bucket 54 (50331648 elements, 50331648 padded size):
1071
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight14
1072
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight13
1073
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight11
1074
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight12
1075
+ Params for bucket 55 (50331648 elements, 50331648 padded size):
1076
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight10
1077
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight7
1078
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight8
1079
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight9
1080
+ Params for bucket 56 (50331648 elements, 50331648 padded size):
1081
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight6
1082
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight3
1083
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight4
1084
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight5
1085
+ Params for bucket 57 (44040192 elements, 44040192 padded size):
1086
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight2
1087
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight0
1088
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight1
1089
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight15
1090
+ Params for bucket 58 (44040192 elements, 44040192 padded size):
1091
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight13
1092
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight11
1093
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight9
1094
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight14
1095
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight10
1096
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight12
1097
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight8
1098
+ Params for bucket 59 (44040192 elements, 44040192 padded size):
1099
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight7
1100
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight3
1101
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight5
1102
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight1
1103
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight4
1104
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight6
1105
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight2
1106
+ Params for bucket 60 (44040192 elements, 44040192 padded size):
1107
+ module.decoder.layers.2.mlp.experts.linear_fc2.weight0
1108
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight13
1109
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight14
1110
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight15
1111
+ Params for bucket 61 (50331648 elements, 50331648 padded size):
1112
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight12
1113
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight10
1114
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight11
1115
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight9
1116
+ Params for bucket 62 (50331648 elements, 50331648 padded size):
1117
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight8
1118
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight6
1119
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight7
1120
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight5
1121
+ Params for bucket 63 (50331648 elements, 50331648 padded size):
1122
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight4
1123
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight3
1124
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight1
1125
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight2
1126
+ Params for bucket 64 (44040192 elements, 44040192 padded size):
1127
+ module.decoder.layers.2.mlp.experts.linear_fc1.weight0
1128
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight14
1129
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight12
1130
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight15
1131
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight13
1132
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight11
1133
+ Params for bucket 65 (44040192 elements, 44040192 padded size):
1134
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight10
1135
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight6
1136
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight8
1137
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight4
1138
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight9
1139
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight7
1140
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight5
1141
+ Params for bucket 66 (50331648 elements, 50331648 padded size):
1142
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight15
1143
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight0
1144
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight2
1145
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight3
1146
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight1
1147
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight14
1148
+ Params for bucket 67 (50331648 elements, 50331648 padded size):
1149
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight13
1150
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight12
1151
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight10
1152
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight11
1153
+ Params for bucket 68 (50331648 elements, 50331648 padded size):
1154
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight9
1155
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight6
1156
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight7
1157
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight8
1158
+ Params for bucket 69 (50331648 elements, 50331648 padded size):
1159
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight5
1160
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight3
1161
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight2
1162
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight4
1163
+ Params for bucket 70 (44040192 elements, 44040192 padded size):
1164
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight1
1165
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight14
1166
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight13
1167
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight0
1168
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight15
1169
+ Params for bucket 71 (44040192 elements, 44040192 padded size):
1170
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight12
1171
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight10
1172
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight8
1173
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight6
1174
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight11
1175
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight9
1176
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight7
1177
+ Params for bucket 72 (50331648 elements, 50331648 padded size):
1178
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight4
1179
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight2
1180
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight0
1181
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight5
1182
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight1
1183
+ module.decoder.layers.0.mlp.experts.linear_fc2.weight3
1184
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight15
1185
+ Params for bucket 73 (50331648 elements, 50331648 padded size):
1186
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight14
1187
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight13
1188
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight11
1189
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight12
1190
+ Params for bucket 74 (50331648 elements, 50331648 padded size):
1191
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight10
1192
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight7
1193
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight8
1194
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight9
1195
+ Params for bucket 75 (50331648 elements, 50331648 padded size):
1196
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight6
1197
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight5
1198
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight4
1199
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight3
1200
+ Params for bucket 76 (37748736 elements, 37748736 padded size):
1201
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight2
1202
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight1
1203
+ module.decoder.layers.0.mlp.experts.linear_fc1.weight0
1204
+ 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 0x7eaaba118140>, config_logger_dir='')
1205
+ INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine
1206
+ [after model, optimizer, and learning rate scheduler are built] datetime: 2026-08-26 16:02:44
1207
+ > building train, validation, and test datasets ...
1208
+ > datasets target sizes (minimum size):
1209
+ train: 96
1210
+ validation: 0
1211
+ test: 0
1212
+ INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None]
1213
+ > building train, validation, and test datasets for GPT ...
1214
+ 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 0x7eaabb13e8a0>, 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)
1215
+ INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx
1216
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths
1217
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers
1218
+ INFO:megatron.core.datasets.indexed_dataset: Extract the document indices
1219
+ INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029
1220
+ INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350
1221
+ INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices
1222
+ INFO:megatron.core.datasets.gpt_dataset: Load the document index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-document_index.npy
1223
+ INFO:megatron.core.datasets.gpt_dataset: Load the sample index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-sample_index.npy
1224
+ INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from eadc3756447e39d609e8aa728a91f451-GPTDataset-train-shuffle_index.npy
1225
+ INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 14536
1226
+ > finished creating GPT datasets ...
1227
+ [after dataloaders are built] datetime: 2026-08-26 16:02:44
1228
+ done with setup ...
1229
+ (min, max) time across ranks (ms):
1230
+ model-and-optimizer-setup ......................: (634.18, 646.50)
1231
+ train/valid/test-data-iterators-setup ..........: (63.59, 74.46)
1232
+ training ...
1233
+ Overwriting rerun_state_machine.current_iteration from -1 to 0...
1234
+ [before the start of training step] datetime: 2026-08-26 16:02:44
tierb_study_20260826_144644/echo/tierb/tierb_gbs8/results.tsv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir
2
+ QWEN3_235B echo matched 16 8192 8 2.00 none 1 OK 323.26 322.10 155809 /workspace/tierb_study_20260826_144644/echo/tierb/tierb_gbs8/QWEN3_235B__echo__matched__seq8192__gbs8__cap2p00/rep1
tierb_study_20260826_144644/hf_watcher.log CHANGED
@@ -23,3 +23,6 @@ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
23
  uploaded: /workspace/tierb_study_20260826_144644 -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tierb_study_20260826_144644
24
  https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
25
  [2026-08-26T15:52:40+00:00] backup pass
 
 
 
 
23
  uploaded: /workspace/tierb_study_20260826_144644 -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tierb_study_20260826_144644
24
  https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
25
  [2026-08-26T15:52:40+00:00] backup pass
26
+ uploaded: /workspace/tierb_study_20260826_144644 -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tierb_study_20260826_144644
27
+ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
28
+ [2026-08-26T16:02:44+00:00] backup pass
tierb_study_20260826_144644/steps.tsv CHANGED
@@ -3,3 +3,5 @@ OK echo tierbprobe 8 2026-08-26T14:46:44+00:00 2026-08-26T14:57:54+00:00 /worksp
3
  FAIL ultraep tierbprobe 8 2026-08-26T14:57:57+00:00 2026-08-26T15:05:21+00:00 /workspace/tierb_study_20260826_144644/ultraep/tierbprobe
4
  FAIL moonep tierbprobe 8 2026-08-26T15:05:24+00:00 2026-08-26T15:12:52+00:00 /workspace/tierb_study_20260826_144644/moonep/tierbprobe
5
  OK echo tierbprobe 8 2026-08-26T15:32:33+00:00 2026-08-26T15:43:24+00:00 /workspace/tierb_study_20260826_144644/echo/tierbprobe
 
 
 
3
  FAIL ultraep tierbprobe 8 2026-08-26T14:57:57+00:00 2026-08-26T15:05:21+00:00 /workspace/tierb_study_20260826_144644/ultraep/tierbprobe
4
  FAIL moonep tierbprobe 8 2026-08-26T15:05:24+00:00 2026-08-26T15:12:52+00:00 /workspace/tierb_study_20260826_144644/moonep/tierbprobe
5
  OK echo tierbprobe 8 2026-08-26T15:32:33+00:00 2026-08-26T15:43:24+00:00 /workspace/tierb_study_20260826_144644/echo/tierbprobe
6
+ OK ultraep tierbprobe 8 2026-08-26T15:43:26+00:00 2026-08-26T15:53:23+00:00 /workspace/tierb_study_20260826_144644/ultraep/tierbprobe
7
+ OK moonep tierbprobe 8 2026-08-26T15:53:25+00:00 2026-08-26T16:00:40+00:00 /workspace/tierb_study_20260826_144644/moonep/tierbprobe
tierb_study_20260826_144644/study.log CHANGED
@@ -635,3 +635,221 @@ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
635
  --------------------------------------------------------------------------
636
  [15:52:03] CELL QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p25__forcebal (slots=8)
637
  [15:52:06] waiting for stale ranks to exit ...
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
635
  --------------------------------------------------------------------------
636
  [15:52:03] CELL QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p25__forcebal (slots=8)
637
  [15:52:06] waiting for stale ranks to exit ...
638
+ [guard] iter 12/12 (60s)
639
+ [guard] tearing down process group (outcome=DONE)
640
+ --------------------------------------------------------------------------
641
+
642
+ ==============================================================
643
+ SWEEP COMPLETE ok=8 failed=0
644
+ ==============================================================
645
+ config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir
646
+ QWEN3_235B ultraep matched 16 8192 8 1.25 none 1 OK 340.64 339.20 122751 /workspace/tierb_study_20260826_144644/ultraep/tierbprobe/tierbprobe_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p25/rep1
647
+ QWEN3_235B ultraep matched 16 8192 8 1.25 none+forcebal 1 OK 317.89 317.40 122739 /workspace/tierb_study_20260826_144644/ultraep/tierbprobe/tierbprobe_gbs8/QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p25__forcebal/rep1
648
+ QWEN3_30B ultraep matched 16 4096 8 1.25 none 1 OK 749.31 712.40 117927 /workspace/tierb_study_20260826_144644/ultraep/tierbprobe/tierbprobe_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p25/rep1
649
+ QWEN3_30B ultraep matched 16 4096 8 1.25 none+forcebal 1 OK 737.56 705.60 117867 /workspace/tierb_study_20260826_144644/ultraep/tierbprobe/tierbprobe_gbs8/QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p25__forcebal/rep1
650
+ DSV2_LITE_L24 ultraep matched 8 8192 8 1.25 none 1 OK 639.96 638.80 130023 /workspace/tierb_study_20260826_144644/ultraep/tierbprobe/tierbprobe_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p25/rep1
651
+ DSV2_LITE_L24 ultraep matched 8 8192 8 1.25 none+forcebal 1 OK 625.83 624.60 129971 /workspace/tierb_study_20260826_144644/ultraep/tierbprobe/tierbprobe_gbs8/DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1p25__forcebal/rep1
652
+ QWEN2_57B ultraep matched 8 4096 8 1.25 none 1 OK 238.29 230.10 97675 /workspace/tierb_study_20260826_144644/ultraep/tierbprobe/tierbprobe_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p25/rep1
653
+ QWEN2_57B ultraep matched 8 4096 8 1.25 none+forcebal 1 OK 269.69 212.00 97677 /workspace/tierb_study_20260826_144644/ultraep/tierbprobe/tierbprobe_gbs8/QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p25__forcebal/rep1
654
+
655
+ Results: /workspace/tierb_study_20260826_144644/ultraep/tierbprobe/tierbprobe_gbs8/results.tsv
656
+ Roll up: python3 /workspace/blackwell_run_scripts/granularity/tools/collect.py /workspace/tierb_study_20260826_144644/ultraep/tierbprobe/tierbprobe_gbs8
657
+ ==============================================================
658
+ 8 probe runs -> /workspace/tierb_study_20260826_144644/ultraep/tierbprobe/tierbprobe_gbs8/probe_summary.tsv
659
+ run arm cap rawExpI preRankI postRankI drop% reqCap
660
+ --------------------------------------------------------------------------------------------------------------------
661
+ DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1 natural 1.25 2.1074 1.0100 0.0000 1.0101
662
+ DSV2_LITE_L24__ultraep__matched__seq8192__gbs8__cap1 forced 1.25 1.0061 1.0060 0.0000 1.0100
663
+ QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p25/ natural 1.25 2.1724 1.0100 0.0000 1.0101
664
+ QWEN2_57B__ultraep__matched__seq4096__gbs8__cap1p25_ forced 1.25 1.0074 1.0071 0.0000 1.0100
665
+ QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p25 natural 1.25 2.4475 1.0100 0.0000 1.0101
666
+ QWEN3_235B__ultraep__matched__seq8192__gbs8__cap1p25 forced 1.25 1.0053 1.0053 0.0000 1.0100
667
+ QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p25/ natural 1.25 2.3516 1.0100 0.0000 1.0101
668
+ QWEN3_30B__ultraep__matched__seq4096__gbs8__cap1p25_ forced 1.25 1.0077 1.0073 0.0000 1.0100
669
+
670
+ ECHO note: echo_logged_pre_rank_metric_valid=0 and echo_logged_load_reduction_valid=0 are intentional.
671
+ The branch's built-in fields mix source-rank and destination-rank axes; this tool reconstructs both globally.
672
+
673
+ TIER B PROBE -- what these logs contain, per backend
674
+ ----------------------------------------------------
675
+ Capacity matrix:
676
+ ECHO: start C=2.00; on OOM descend 1.75, 1.50, 1.25. Stop at
677
+ the highest fitting zero-drop capacity. If a fitting capacity drops,
678
+ smaller values cannot help and the model is blocked.
679
+ UltraEP: measure required_capacity at C=1.25, round it UP to the next
680
+ 0.01, then validate that capacity once.
681
+ MoonEP: natural + forced, no capacity factor.
682
+
683
+ The forced arm runs once at the selected/high-headroom capacity. It validates
684
+ instrumentation and is not part of capacity selection.
685
+
686
+ Each backend instruments a different thing:
687
+
688
+ echo local raw-router histograms and local post-reroute placement.
689
+ reduce_probe.py aligns and sums all eight source ranks before
690
+ calculating raw_expert_*, pre_home_rank_*,
691
+ post_physical_rank_* and valid_max_rank_reduction_*.
692
+
693
+ ultraep global pre/post rank vectors plus explicit drop accounting:
694
+ drop_rate_pct and required_capacity -- the capacity factor the run
695
+ WOULD have needed. The vectors are duplicated identically on all
696
+ logging ranks and are deduplicated, not summed.
697
+
698
+ moonep local raw-router histograms, actual payload rows per physical
699
+ destination, and the bytes-copy side. The reducer globally sums
700
+ each (iteration, layer) snapshot first. ECHO reports no comparable
701
+ byte counters, so MoonEP bytes are a self-comparison across models,
702
+ never a cross-backend ranking.
703
+
704
+ Do NOT quote ECHO's logged pre_echo imbalance_ratio or
705
+ max_load_reduction_pct. The branch mixes source-rank and destination-rank
706
+ axes; the reducer marks both invalid and reconstructs the comparison globally.
707
+
708
+ Two boundary cases:
709
+ raw_expert_imax_max == E/K
710
+ -> at least one global snapshot had one expert in every
711
+ token's top-K set. Inspect the MEAN before generalising.
712
+ required_capacity or min-safe-capacity at the configured capacity
713
+ -> clipped by the buffer; it means "at or over", not
714
+ "fits comfortably".
715
+
716
+ tools/reduce_probe.py has already written probe_summary.tsv and .json here.
717
+
718
+ Campaign results: /workspace/tierb_study_20260826_144644/ultraep/tierbprobe
719
+ Roll up: python3 /workspace/blackwell_run_scripts/granularity/tools/collect.py /workspace/tierb_study_20260826_144644/ultraep/tierbprobe --compare
720
+ Already logged in to Hugging Face as: AbhishekGhosh1998
721
+ uploaded: /workspace/tierb_study_20260826_144644/ultraep/tierbprobe -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tierb_study_20260826_144644/ultraep/tierbprobe
722
+ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
723
+
724
+ === STEP moonep__tierbprobe__gbs8
725
+ ==============================================================
726
+ CAMPAIGN: moonep backend=moonep phase=tierbprobe
727
+ Megatron: /workspace/Megatron-MoE-Echo @ moonep_backend_token_routing_analysis
728
+ GBS=8 seq=4096 repeats=3 nsys=1
729
+ results -> /workspace/tierb_study_20260826_144644/moonep/tierbprobe
730
+ ==============================================================
731
+ --------------------------------------------------------------------------
732
+ [15:53:40] CELL QWEN3_235B__moonep__matched__seq8192__gbs8__capnone (slots=16)
733
+ [guard] tearing down process group (outcome=OOM)
734
+ [15:54:41] WARN: QWEN3_235B__moonep__matched__seq8192__gbs8__capnone rep1 -> OOM
735
+ [15:54:41] WARN: QWEN3_235B does not fit moonep (OOM); excluded from moonep Tier B timing
736
+ --------------------------------------------------------------------------
737
+ [15:54:41] CELL QWEN3_30B__moonep__matched__seq4096__gbs8__capnone (slots=16)
738
+ [15:54:41] waiting for stale ranks to exit ...
739
+ [guard] tearing down process group (outcome=OOM)
740
+ [15:55:53] WARN: QWEN3_30B__moonep__matched__seq4096__gbs8__capnone rep1 -> OOM
741
+ [15:55:53] WARN: QWEN3_30B does not fit moonep (OOM); excluded from moonep Tier B timing
742
+ --------------------------------------------------------------------------
743
+ [15:55:53] CELL DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone (slots=8)
744
+ [15:55:53] waiting for stale ranks to exit ...
745
+ [guard] iter 11/12 (60s)
746
+ [guard] iter 12/12 (70s)
747
+ [guard] tearing down process group (outcome=DONE)
748
+ --------------------------------------------------------------------------
749
+ [15:57:15] CELL DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone__forcebal (slots=8)
750
+ [guard] iter 10/12 (60s)
751
+ [guard] iter 12/12 (70s)
752
+ [guard] tearing down process group (outcome=DONE)
753
+ --------------------------------------------------------------------------
754
+ [15:58:27] CELL QWEN2_57B__moonep__matched__seq4096__gbs8__capnone (slots=8)
755
+ [guard] iter 1/12 (50s)
756
+ [guard] iter 12/12 (60s)
757
+ [guard] tearing down process group (outcome=DONE)
758
+ --------------------------------------------------------------------------
759
+ [15:59:33] CELL QWEN2_57B__moonep__matched__seq4096__gbs8__capnone__forcebal (slots=8)
760
+ [guard] iter 12/12 (60s)
761
+ [guard] tearing down process group (outcome=DONE)
762
+ --------------------------------------------------------------------------
763
+
764
+ ==============================================================
765
+ SWEEP COMPLETE ok=4 failed=2
766
+ ==============================================================
767
+ config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir
768
+ QWEN3_235B moonep matched 16 8192 8 none none 1 OOM NA NA 182589 /workspace/tierb_study_20260826_144644/moonep/tierbprobe/tierbprobe_gbs8/QWEN3_235B__moonep__matched__seq8192__gbs8__capnone/rep1
769
+ QWEN3_30B moonep matched 16 4096 8 none none 1 OOM NA NA 182599 /workspace/tierb_study_20260826_144644/moonep/tierbprobe/tierbprobe_gbs8/QWEN3_30B__moonep__matched__seq4096__gbs8__capnone/rep1
770
+ DSV2_LITE_L24 moonep matched 8 8192 8 none none 1 OK 793.84 791.50 157151 /workspace/tierb_study_20260826_144644/moonep/tierbprobe/tierbprobe_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone/rep1
771
+ DSV2_LITE_L24 moonep matched 8 8192 8 none none+forcebal 1 OK 794.89 790.90 157151 /workspace/tierb_study_20260826_144644/moonep/tierbprobe/tierbprobe_gbs8/DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capnone__forcebal/rep1
772
+ QWEN2_57B moonep matched 8 4096 8 none none 1 OK 401.09 360.60 141555 /workspace/tierb_study_20260826_144644/moonep/tierbprobe/tierbprobe_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/rep1
773
+ QWEN2_57B moonep matched 8 4096 8 none none+forcebal 1 OK 406.04 364.50 141555 /workspace/tierb_study_20260826_144644/moonep/tierbprobe/tierbprobe_gbs8/QWEN2_57B__moonep__matched__seq4096__gbs8__capnone__forcebal/rep1
774
+
775
+ Results: /workspace/tierb_study_20260826_144644/moonep/tierbprobe/tierbprobe_gbs8/results.tsv
776
+ Roll up: python3 /workspace/blackwell_run_scripts/granularity/tools/collect.py /workspace/tierb_study_20260826_144644/moonep/tierbprobe/tierbprobe_gbs8
777
+ ==============================================================
778
+ 4 probe runs -> /workspace/tierb_study_20260826_144644/moonep/tierbprobe/tierbprobe_gbs8/probe_summary.tsv
779
+ run arm cap rawExpI preRankI postRankI drop% reqCap
780
+ --------------------------------------------------------------------------------------------------------------------
781
+ DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capno natural none 8.3174* 2.2370 1.1122
782
+ DSV2_LITE_L24__moonep__matched__seq8192__gbs8__capno forced none 1.0286 1.0061 1.0026
783
+ QWEN2_57B__moonep__matched__seq4096__gbs8__capnone/r natural none 7.4086* 2.2416 1.0986
784
+ QWEN2_57B__moonep__matched__seq4096__gbs8__capnone__ forced none 1.0348 1.0073 1.0033
785
+
786
+ * At least one globally aggregated snapshot reached the E/K expert-Imax ceiling.
787
+ Inspect raw_expert_imax_mean and hottest_expert_token_fraction_mean before generalising.
788
+
789
+ ECHO note: echo_logged_pre_rank_metric_valid=0 and echo_logged_load_reduction_valid=0 are intentional.
790
+ The branch's built-in fields mix source-rank and destination-rank axes; this tool reconstructs both globally.
791
+
792
+ TIER B PROBE -- what these logs contain, per backend
793
+ ----------------------------------------------------
794
+ Capacity matrix:
795
+ ECHO: start C=2.00; on OOM descend 1.75, 1.50, 1.25. Stop at
796
+ the highest fitting zero-drop capacity. If a fitting capacity drops,
797
+ smaller values cannot help and the model is blocked.
798
+ UltraEP: measure required_capacity at C=1.25, round it UP to the next
799
+ 0.01, then validate that capacity once.
800
+ MoonEP: natural + forced, no capacity factor.
801
+
802
+ The forced arm runs once at the selected/high-headroom capacity. It validates
803
+ instrumentation and is not part of capacity selection.
804
+
805
+ Each backend instruments a different thing:
806
+
807
+ echo local raw-router histograms and local post-reroute placement.
808
+ reduce_probe.py aligns and sums all eight source ranks before
809
+ calculating raw_expert_*, pre_home_rank_*,
810
+ post_physical_rank_* and valid_max_rank_reduction_*.
811
+
812
+ ultraep global pre/post rank vectors plus explicit drop accounting:
813
+ drop_rate_pct and required_capacity -- the capacity factor the run
814
+ WOULD have needed. The vectors are duplicated identically on all
815
+ logging ranks and are deduplicated, not summed.
816
+
817
+ moonep local raw-router histograms, actual payload rows per physical
818
+ destination, and the bytes-copy side. The reducer globally sums
819
+ each (iteration, layer) snapshot first. ECHO reports no comparable
820
+ byte counters, so MoonEP bytes are a self-comparison across models,
821
+ never a cross-backend ranking.
822
+
823
+ Do NOT quote ECHO's logged pre_echo imbalance_ratio or
824
+ max_load_reduction_pct. The branch mixes source-rank and destination-rank
825
+ axes; the reducer marks both invalid and reconstructs the comparison globally.
826
+
827
+ Two boundary cases:
828
+ raw_expert_imax_max == E/K
829
+ -> at least one global snapshot had one expert in every
830
+ token's top-K set. Inspect the MEAN before generalising.
831
+ required_capacity or min-safe-capacity at the configured capacity
832
+ -> clipped by the buffer; it means "at or over", not
833
+ "fits comfortably".
834
+
835
+ tools/reduce_probe.py has already written probe_summary.tsv and .json here.
836
+
837
+ Campaign results: /workspace/tierb_study_20260826_144644/moonep/tierbprobe
838
+ Roll up: python3 /workspace/blackwell_run_scripts/granularity/tools/collect.py /workspace/tierb_study_20260826_144644/moonep/tierbprobe --compare
839
+ Already logged in to Hugging Face as: AbhishekGhosh1998
840
+ uploaded: /workspace/tierb_study_20260826_144644/moonep/tierbprobe -> MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tierb_study_20260826_144644/moonep/tierbprobe
841
+ https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
842
+
843
+ === STAGE 4 -- Tier B timing (probe-selected zero-drop capacities)
844
+
845
+ === STEP echo__tierb__gbs8
846
+ ==============================================================
847
+ CAMPAIGN: echo backend=echo phase=tierb
848
+ Megatron: /workspace/Megatron-MoE-Echo @ bracing-regions-through-kernels
849
+ GBS=8 seq=4096 repeats=3 nsys=1
850
+ results -> /workspace/tierb_study_20260826_144644/echo/tierb
851
+ ==============================================================
852
+ --------------------------------------------------------------------------
853
+ [16:00:58] CELL QWEN3_235B__echo__matched__seq8192__gbs8__cap2p00 (slots=16)
854
+ [guard] iter 8/12 (70s)
855
+ [guard] iter 12/12 (80s)
tierb_study_20260826_144644/tierb_timing_capacities.env.tmp.168529 ADDED
File without changes
tierb_study_20260826_144644/tierb_timing_capacities.tsv.tmp.168529 ADDED
@@ -0,0 +1 @@
 
 
1
+ model common_timing_capacity echo_probe_required_capacity echo_full_cg_status echo_run_dir