AbhishekGhosh1998 commited on
Commit
667bb86
·
verified ·
1 Parent(s): 525945f

granularity results granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb @ 2026-09-21T07:54:40Z from sp-aiinfra-interactive-aghosh-8-mb99k-8xb200-moe-cg-0

Browse files
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep2/mem_samples.csv CHANGED
@@ -118,3 +118,107 @@
118
  5, 59207
119
  6, 59207
120
  7, 59207
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
118
  5, 59207
119
  6, 59207
120
  7, 59207
121
+ 0, 59207
122
+ 1, 59209
123
+ 2, 59207
124
+ 3, 59207
125
+ 4, 59207
126
+ 5, 59207
127
+ 6, 59207
128
+ 7, 59207
129
+ 0, 59207
130
+ 1, 59209
131
+ 2, 59207
132
+ 3, 59207
133
+ 4, 59207
134
+ 5, 59207
135
+ 6, 59207
136
+ 7, 59207
137
+ 0, 59405
138
+ 1, 59407
139
+ 2, 59405
140
+ 3, 59405
141
+ 4, 59405
142
+ 5, 59405
143
+ 6, 59405
144
+ 7, 59405
145
+ 0, 64191
146
+ 1, 64199
147
+ 2, 64213
148
+ 3, 64197
149
+ 4, 64219
150
+ 5, 64195
151
+ 6, 64215
152
+ 7, 64235
153
+ 0, 64741
154
+ 1, 64743
155
+ 2, 64741
156
+ 3, 64741
157
+ 4, 64741
158
+ 5, 64741
159
+ 6, 64755
160
+ 7, 64741
161
+ 0, 67031
162
+ 1, 67011
163
+ 2, 67039
164
+ 3, 67033
165
+ 4, 67035
166
+ 5, 67037
167
+ 6, 67111
168
+ 7, 67039
169
+ 0, 67965
170
+ 1, 67941
171
+ 2, 67943
172
+ 3, 67949
173
+ 4, 67971
174
+ 5, 67937
175
+ 6, 68009
176
+ 7, 67949
177
+ 0, 77099
178
+ 1, 77129
179
+ 2, 77111
180
+ 3, 77203
181
+ 4, 77233
182
+ 5, 77283
183
+ 6, 77303
184
+ 7, 77263
185
+ 0, 86365
186
+ 1, 86369
187
+ 2, 86365
188
+ 3, 86477
189
+ 4, 86477
190
+ 5, 86477
191
+ 6, 86491
192
+ 7, 86477
193
+ 0, 84455
194
+ 1, 84523
195
+ 2, 85719
196
+ 3, 84887
197
+ 4, 84631
198
+ 5, 85831
199
+ 6, 84901
200
+ 7, 84567
201
+ 0, 82575
202
+ 1, 82455
203
+ 2, 83033
204
+ 3, 82819
205
+ 4, 82443
206
+ 5, 83533
207
+ 6, 82897
208
+ 7, 83145
209
+ 0, 4
210
+ 1, 68
211
+ 2, 69603
212
+ 3, 324
213
+ 4, 68
214
+ 5, 69709
215
+ 6, 324
216
+ 7, 4
217
+ 0, 0
218
+ 1, 0
219
+ 2, 0
220
+ 3, 0
221
+ 4, 0
222
+ 5, 0
223
+ 6, 0
224
+ 7, 0
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep2/train.log CHANGED
@@ -852,3 +852,84 @@ INFO:moonep.api:MoonEP comm buffer sizes:
852
  meta_buffer_size : 2.00 MiB
853
  local_temp_buffer_size : 1.29 MiB
854
  total_size : 619.29 MiB
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
852
  meta_buffer_size : 2.00 MiB
853
  local_temp_buffer_size : 1.29 MiB
854
  total_size : 619.29 MiB
855
+ [MoonEP selfcheck 1/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 384, 385, 386, 387, 388, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404, 405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419, 420, 421, 422, 423, 424, 425, 426, 427, 428, 429, 430, 431] sizes [384, 384, 256, 128, 384, 1152, 1664, 1152, 1664, 640, 896, 640, 256, 896, 256, 1280, 384, 512, 1664, 1536, 768, 384, 512, 256, 384, 1152, 640, 256, 256, 384, 384, 640, 640, 2048, 384, 640, 384, 256, 768, 1408, 512, 256, 128, 256, 128, 640, 256, 384, 3456, 640, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[348, 357, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [332, 309, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [314, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [76, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [230, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [208, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [76, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
856
+ [2026-09-21 07:53:14] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 25725.8 | throughput per GPU (TFLOP/s/GPU): 1.4 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.227924E+01 | loss scale: 1.0 | grad norm: 12.616 | number of skipped iterations: 0 | number of nan iterations: 0 |
857
+ Number of parameters in transformer block in billions: 17.55
858
+ Number of parameters in embedding layers in billions: 0.72
859
+ Total number of parameters in billions: 18.28
860
+ Number of parameters in most loaded shard in billions: 18.2753
861
+ Theoretical memory footprints: weight and optimizer=130715.23 MB
862
+ [Rank 0] (after 1 iterations) memory (MB) | allocated: 45359.1982421875 | max allocated: 45359.19873046875 | reserved: 45790.0 | max reserved: 45790.0
863
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
864
+ warnings.warn(
865
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
866
+ warnings.warn(
867
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
868
+ warnings.warn(
869
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
870
+ warnings.warn(
871
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
872
+ warnings.warn(
873
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
874
+ warnings.warn(
875
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
876
+ warnings.warn(
877
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
878
+ warnings.warn(
879
+ [MoonEP selfcheck 2/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 384, 385, 386, 387, 388, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404, 405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419, 420, 421, 422, 423, 424, 425, 426, 427, 428, 429, 430, 431] sizes [384, 384, 384, 128, 512, 1152, 1792, 1024, 1536, 768, 768, 640, 256, 512, 256, 1408, 512, 384, 2048, 1280, 768, 512, 512, 256, 512, 1152, 896, 128, 256, 384, 640, 512, 640, 2048, 512, 768, 512, 256, 896, 1536, 640, 128, 128, 256, 128, 768, 256, 512, 3456, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[348, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [357, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [332, 309, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [230, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [76, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [314, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [76, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
880
+ [2026-09-21 07:53:14] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 131.0 | throughput per GPU (TFLOP/s/GPU): 270.0 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.227087E+01 | loss scale: 1.0 | grad norm: 12.533 | number of skipped iterations: 0 | number of nan iterations: 0 |
881
+ [MoonEP selfcheck 3/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 384, 385, 386, 387, 388, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404, 405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419, 420, 421, 422, 423, 424, 425, 426, 427, 428, 429, 430, 431] sizes [256, 512, 256, 256, 512, 896, 1920, 1024, 2176, 896, 1536, 640, 256, 640, 384, 640, 512, 640, 1408, 896, 768, 512, 512, 256, 512, 1152, 768, 256, 384, 384, 640, 512, 512, 1408, 512, 896, 512, 256, 768, 1024, 640, 256, 256, 384, 128, 384, 256, 768, 3456, 384, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[313, 309, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [208, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [338, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [230, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [306, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [76, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [332, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
882
+ [2026-09-21 07:53:15] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 110.9 | throughput per GPU (TFLOP/s/GPU): 319.1 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 1.088362E+01 | loss scale: 1.0 | grad norm: 11.029 | number of skipped iterations: 0 | number of nan iterations: 0 |
883
+ Capture CUDA graph for training!!!
884
+ CUDA graph capture done!!!
885
+ [2026-09-21 07:53:15] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 485.0 | throughput per GPU (TFLOP/s/GPU): 72.9 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 1.448777E+01 | loss scale: 1.0 | grad norm: 72.970 | number of skipped iterations: 0 | number of nan iterations: 0 |
886
+ [2026-09-21 07:53:15] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 108.5 | throughput per GPU (TFLOP/s/GPU): 326.1 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.808960E+01 | loss scale: 1.0 | grad norm: 108.333 | number of skipped iterations: 0 | number of nan iterations: 0 |
887
+ [2026-09-21 07:53:15] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 108.8 | throughput per GPU (TFLOP/s/GPU): 325.1 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 1.490298E+01 | loss scale: 1.0 | grad norm: 83.291 | number of skipped iterations: 0 | number of nan iterations: 0 |
888
+ [2026-09-21 07:53:15] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 109.4 | throughput per GPU (TFLOP/s/GPU): 323.4 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 1.142665E+01 | loss scale: 1.0 | grad norm: 26.340 | number of skipped iterations: 0 | number of nan iterations: 0 |
889
+ [2026-09-21 07:53:15] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 108.6 | throughput per GPU (TFLOP/s/GPU): 325.9 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.165925E+01 | loss scale: 1.0 | grad norm: 12.676 | number of skipped iterations: 0 | number of nan iterations: 0 |
890
+ [2026-09-21 07:53:16] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 109.8 | throughput per GPU (TFLOP/s/GPU): 322.1 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.132117E+01 | loss scale: 1.0 | grad norm: 10.523 | number of skipped iterations: 0 | number of nan iterations: 0 |
891
+ [2026-09-21 07:53:16] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 108.9 | throughput per GPU (TFLOP/s/GPU): 324.7 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.097885E+01 | loss scale: 1.0 | grad norm: 9.172 | number of skipped iterations: 0 | number of nan iterations: 0 |
892
+ [2026-09-21 07:53:16] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 108.5 | throughput per GPU (TFLOP/s/GPU): 325.9 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.072919E+01 | loss scale: 1.0 | grad norm: 9.283 | number of skipped iterations: 0 | number of nan iterations: 0 |
893
+ [2026-09-21 07:53:16] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 108.1 | throughput per GPU (TFLOP/s/GPU): 327.3 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.047170E+01 | loss scale: 1.0 | grad norm: 9.277 | number of skipped iterations: 0 | number of nan iterations: 0 |
894
+ [after training is done] datetime: 2026-09-21 07:53:16
895
+ [rank0]:[W921 07:53:17.557828365 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
896
+ [rank3]:[W921 07:53:18.397054620 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
897
+ [rank7]:[W921 07:53:18.402746735 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
898
+ [rank1]:[W921 07:53:18.403545243 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
899
+ [rank6]:[W921 07:53:18.403900871 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
900
+ [rank4]:[W921 07:53:18.503929987 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
901
+ [rank5]:[W921 07:53:19.006220053 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
902
+ [rank2]:[W921 07:53:19.118086742 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
903
+ W0921 07:53:23.410000 533518 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers
904
+ W0921 07:53:23.411000 533518 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 533709 closing signal SIGTERM
905
+ W0921 07:53:23.411000 533518 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 533710 closing signal SIGTERM
906
+ W0921 07:53:23.412000 533518 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 533713 closing signal SIGTERM
907
+ W0921 07:53:23.413000 533518 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 533715 closing signal SIGTERM
908
+ Traceback (most recent call last):
909
+ File "/usr/local/bin/torchrun", line 8, in <module>
910
+ sys.exit(main())
911
+ ^^^^^^
912
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
913
+ return f(*args, **kwargs)
914
+ ^^^^^^^^^^^^^^^^^^
915
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 892, in main
916
+ run(args)
917
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 883, in run
918
+ elastic_launch(
919
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
920
+ return launch_agent(self._config, self._entrypoint, list(args))
921
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
922
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 261, in launch_agent
923
+ result = agent.run()
924
+ ^^^^^^^^^^^
925
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper
926
+ result = f(*args, **kwargs)
927
+ ^^^^^^^^^^^^^^^^^^
928
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 711, in run
929
+ result = self._invoke_run(role)
930
+ ^^^^^^^^^^^^^^^^^^^^^^
931
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 870, in _invoke_run
932
+ time.sleep(monitor_interval)
933
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 84, in _terminate_process_handler
934
+ raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval)
935
+ torch.distributed.elastic.multiprocessing.api.SignalException: Process 533518 got signal: 15
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep3/mem_samples.csv ADDED
@@ -0,0 +1,216 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 0, 0
2
+ 1, 0
3
+ 2, 0
4
+ 3, 0
5
+ 4, 0
6
+ 5, 0
7
+ 6, 0
8
+ 7, 0
9
+ 0, 4
10
+ 1, 4
11
+ 2, 1
12
+ 3, 1
13
+ 4, 1
14
+ 5, 1
15
+ 6, 1
16
+ 7, 1
17
+ 0, 4
18
+ 1, 4
19
+ 2, 4
20
+ 3, 4
21
+ 4, 4
22
+ 5, 4
23
+ 6, 4
24
+ 7, 1
25
+ 0, 1
26
+ 1, 1
27
+ 2, 1
28
+ 3, 1
29
+ 4, 1
30
+ 5, 4
31
+ 6, 4
32
+ 7, 4
33
+ 0, 0
34
+ 1, 0
35
+ 2, 0
36
+ 3, 0
37
+ 4, 0
38
+ 5, 0
39
+ 6, 0
40
+ 7, 0
41
+ 0, 1
42
+ 1, 1
43
+ 2, 1
44
+ 3, 1
45
+ 4, 1
46
+ 5, 1
47
+ 6, 1
48
+ 7, 1
49
+ 0, 4
50
+ 1, 4
51
+ 2, 4
52
+ 3, 4
53
+ 4, 1
54
+ 5, 1
55
+ 6, 1
56
+ 7, 1
57
+ 0, 4
58
+ 1, 4
59
+ 2, 4
60
+ 3, 4
61
+ 4, 4
62
+ 5, 4
63
+ 6, 4
64
+ 7, 4
65
+ 0, 4
66
+ 1, 629
67
+ 2, 629
68
+ 3, 629
69
+ 4, 629
70
+ 5, 629
71
+ 6, 629
72
+ 7, 629
73
+ 0, 953
74
+ 1, 955
75
+ 2, 965
76
+ 3, 953
77
+ 4, 961
78
+ 5, 955
79
+ 6, 963
80
+ 7, 963
81
+ 0, 1927
82
+ 1, 1927
83
+ 2, 1927
84
+ 3, 1927
85
+ 4, 1927
86
+ 5, 1927
87
+ 6, 1927
88
+ 7, 1927
89
+ 0, 35183
90
+ 1, 35221
91
+ 2, 35199
92
+ 3, 35225
93
+ 4, 35207
94
+ 5, 35223
95
+ 6, 35215
96
+ 7, 35207
97
+ 0, 49249
98
+ 1, 49249
99
+ 2, 49249
100
+ 3, 49249
101
+ 4, 49249
102
+ 5, 49249
103
+ 6, 49249
104
+ 7, 49249
105
+ 0, 58703
106
+ 1, 58711
107
+ 2, 58703
108
+ 3, 58733
109
+ 4, 58727
110
+ 5, 58709
111
+ 6, 58735
112
+ 7, 58745
113
+ 0, 59207
114
+ 1, 59209
115
+ 2, 59207
116
+ 3, 59207
117
+ 4, 59207
118
+ 5, 59207
119
+ 6, 59207
120
+ 7, 59207
121
+ 0, 59207
122
+ 1, 59209
123
+ 2, 59207
124
+ 3, 59207
125
+ 4, 59207
126
+ 5, 59207
127
+ 6, 59207
128
+ 7, 59207
129
+ 0, 59207
130
+ 1, 59209
131
+ 2, 59207
132
+ 3, 59207
133
+ 4, 59207
134
+ 5, 59207
135
+ 6, 59207
136
+ 7, 59207
137
+ 0, 64233
138
+ 1, 64237
139
+ 2, 64223
140
+ 3, 64219
141
+ 4, 64233
142
+ 5, 64253
143
+ 6, 64235
144
+ 7, 64243
145
+ 0, 64741
146
+ 1, 64743
147
+ 2, 64741
148
+ 3, 64741
149
+ 4, 64741
150
+ 5, 64741
151
+ 6, 64755
152
+ 7, 64741
153
+ 0, 67045
154
+ 1, 67009
155
+ 2, 67045
156
+ 3, 67039
157
+ 4, 67043
158
+ 5, 67069
159
+ 6, 67061
160
+ 7, 67035
161
+ 0, 67995
162
+ 1, 68017
163
+ 2, 68001
164
+ 3, 67999
165
+ 4, 67995
166
+ 5, 68053
167
+ 6, 68021
168
+ 7, 68037
169
+ 0, 77195
170
+ 1, 77191
171
+ 2, 77187
172
+ 3, 77309
173
+ 4, 77305
174
+ 5, 77269
175
+ 6, 77321
176
+ 7, 77301
177
+ 0, 86365
178
+ 1, 86369
179
+ 2, 86365
180
+ 3, 86477
181
+ 4, 86477
182
+ 5, 86477
183
+ 6, 86491
184
+ 7, 86477
185
+ 0, 84519
186
+ 1, 85723
187
+ 2, 84775
188
+ 3, 84567
189
+ 4, 84631
190
+ 5, 85831
191
+ 6, 84901
192
+ 7, 84567
193
+ 0, 82997
194
+ 1, 83425
195
+ 2, 83389
196
+ 3, 83145
197
+ 4, 83107
198
+ 5, 83533
199
+ 6, 82897
200
+ 7, 83533
201
+ 0, 388
202
+ 1, 69261
203
+ 2, 69257
204
+ 3, 324
205
+ 4, 68
206
+ 5, 69369
207
+ 6, 388
208
+ 7, 69369
209
+ 0, 0
210
+ 1, 0
211
+ 2, 0
212
+ 3, 0
213
+ 4, 0
214
+ 5, 0
215
+ 6, 0
216
+ 7, 0
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep3/run.env ADDED
@@ -0,0 +1,105 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=moonep
2
+ config=KIMI_K2_L2
3
+ family=tierB
4
+ source_model=Kimi-K2
5
+ standard_num_layers=61
6
+ base_rung=
7
+ repo=/workspace/Megatron-MoE-Echo
8
+ branch=moonep_backend_async_grad_reduce
9
+ commit=2e68c8ad
10
+ moonep_source_dir=/workspace/MoonEP
11
+ moonep_source_branch=megatron-ddp-symmetric-params
12
+ moonep_source_commit=d8bc666
13
+ num_layers=2
14
+ hidden=7168
15
+ ffn_hidden=18432
16
+ moe_ffn_hidden=2048
17
+ num_experts=384
18
+ topk=8
19
+ experts_per_rank=48
20
+ expert_params=44040192
21
+ heads=64
22
+ kv_groups=64
23
+ kv_channels=64
24
+ qk_layernorm=1
25
+ shared_ffn=2048
26
+ shared_expert_overlap=0
27
+ moe_token_dispatcher=flex
28
+ mla=1
29
+ q_lora_rank=1536
30
+ kv_lora_rank=512
31
+ qk_head_dim=128
32
+ qk_pos_emb_head_dim=64
33
+ v_head_dim=128
34
+ dense_first_layers=1
35
+ moe_layers=1
36
+ rotary_base=50000
37
+ rotary_percent=1.0
38
+ rotary_scaling=32
39
+ mscale=1.0
40
+ mscale_all_dim=1.0
41
+ qkv_bias=0
42
+ router_balancing=seq_aux_loss
43
+ aux_loss_coeff=0
44
+ router_pre_softmax=0
45
+ router_fusion=0
46
+ router_dtype=fp32
47
+ router_groups=1
48
+ router_group_topk=1
49
+ router_score_function=sigmoid
50
+ router_expert_bias=1
51
+ router_bias_update_rate=1e-3
52
+ router_topk_scaling_factor=2.827
53
+ shared_expert_gate=0
54
+ apply_layernorm_1p=0
55
+ attention_output_gate=0
56
+ linear_attention_type=
57
+ linear_attention_freq=
58
+ linear_conv_kernel_dim=
59
+ linear_key_head_dim=
60
+ linear_value_head_dim=
61
+ linear_num_key_heads=
62
+ linear_num_value_heads=
63
+ mtp_num_layers=0
64
+ mtp_loss_scaling_factor=0.1
65
+ no_weight_decay_cond_type=
66
+ max_position_embeddings=4096
67
+ init_method_std=0.02
68
+ device_initiated_grouped_gemm=1
69
+ norm_epsilon=1e-6
70
+ seq_length=4096
71
+ micro_batch_size=1
72
+ global_batch_size=8
73
+ num_microbatches=1
74
+ assignments_per_rank=32768
75
+ mean_tokens_per_expert=682
76
+ slot_policy=matched
77
+ slots_per_rank=48
78
+ capacity_mode=none
79
+ capacity=none
80
+ cuda_graph=1
81
+ probe_mode=0
82
+ async_grad_reduce=1
83
+ branch_role=timing
84
+ memory_dump_mode=none
85
+ memory_snapshot_path=
86
+ cuda_graph_warmup_steps=3
87
+ recompute=none
88
+ force_balance=0
89
+ train_iters=12
90
+ ep_size=8
91
+ echo_dispatch_overlap=1
92
+ echo_dispatch_recompute=0
93
+ ultraep_quota_min_tokens=
94
+ ultraep_weight_sync_join=
95
+ profile_enable=0
96
+ kernel_trace=0
97
+ fp8_mode=mxfp8
98
+ fp8_recipe=mxfp8
99
+ training_seed=1234
100
+ learning_rate=1e-4
101
+ min_learning_rate=1.0e-5
102
+ lr_warmup_iters=5
103
+ state_dump_enable=0
104
+ state_dump_dir=
105
+ moonep_assert_no_pending_reduce=1
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep3/train.log ADDED
@@ -0,0 +1,935 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [07:53:35] WARN: moonep_backend_async_grad_reduce lacks --moe-router-fusion; KIMI_K2_L2 runs unfused (timing only, routing identical)
2
+ ==============================================================
3
+ KIMI_K2_L2 / moonep (slots/rank=48 policy=matched)
4
+ E=384 K=8 H'=2048 m=48
5
+ L=2 seq=4096 gbs=8 (1 microbatches)
6
+ cap_mode=none cap=none cg=1 recompute=none
7
+ memory_dump=none snapshots=/workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep3/memory_snapshots/snapshot.pickle
8
+ repo=/workspace/Megatron-MoE-Echo @ 2e68c8ad
9
+ ==============================================================
10
+ W0921 07:53:36.381000 542044 torch/distributed/run.py:766]
11
+ W0921 07:53:36.381000 542044 torch/distributed/run.py:766] *****************************************
12
+ W0921 07:53:36.381000 542044 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
13
+ W0921 07:53:36.381000 542044 torch/distributed/run.py:766] *****************************************
14
+ 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
15
+ WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer
16
+ Number of virtual stages per pipeline stage: None
17
+ accumulate and all-reduce gradients in fp32 for bfloat16 data type.
18
+ using torch.bfloat16 for parameters ...
19
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
20
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
21
+ ------------------------ arguments ------------------------
22
+ account_for_embedding_in_pipeline_split ......... False
23
+ account_for_loss_in_pipeline_split .............. False
24
+ accumulate_allreduce_grads_in_fp32 .............. True
25
+ adam_beta1 ...................................... 0.9
26
+ adam_beta2 ...................................... 0.999
27
+ adam_eps ........................................ 1e-08
28
+ add_bias_linear ................................. False
29
+ add_position_embedding .......................... False
30
+ add_qkv_bias .................................... False
31
+ adlr_autoresume ................................. False
32
+ adlr_autoresume_interval ........................ 1000
33
+ align_grad_reduce ............................... True
34
+ align_param_gather .............................. False
35
+ app_tag_run_name ................................ None
36
+ app_tag_run_version ............................. 0.0.0
37
+ apply_layernorm_1p .............................. False
38
+ apply_query_key_layer_scaling ................... False
39
+ apply_residual_connection_post_layernorm ........ False
40
+ apply_rope_fusion ............................... False
41
+ async_save ...................................... None
42
+ async_tensor_model_parallel_allreduce ........... True
43
+ attention_backend ............................... AttnBackend.auto
44
+ attention_dropout ............................... 0.0
45
+ attention_softmax_in_fp32 ....................... False
46
+ auto_detect_ckpt_format ......................... False
47
+ barrier_with_L1_time ............................ True
48
+ bert_binary_head ................................ True
49
+ bert_embedder_type .............................. megatron
50
+ bert_load ....................................... None
51
+ bf16 ............................................ True
52
+ bias_dropout_fusion ............................. True
53
+ bias_gelu_fusion ................................ False
54
+ bias_swiglu_fusion .............................. True
55
+ biencoder_projection_dim ........................ 0
56
+ biencoder_shared_query_context_model ............ False
57
+ block_data_path ................................. None
58
+ calc_ft_timeouts ................................ False
59
+ calculate_per_token_loss ........................ False
60
+ check_for_large_grads ........................... False
61
+ check_for_nan_in_loss_and_grad .................. False
62
+ check_for_spiky_loss ............................ False
63
+ check_weight_hash_across_dp_replicas_interval ... None
64
+ ckpt_assume_constant_structure .................. False
65
+ ckpt_convert_format ............................. None
66
+ ckpt_convert_save ............................... None
67
+ ckpt_convert_update_legacy_dist_opt_format ...... False
68
+ ckpt_format ..................................... torch_dist
69
+ ckpt_fully_parallel_load ........................ False
70
+ ckpt_fully_parallel_save ........................ True
71
+ ckpt_fully_parallel_save_deprecated ............. False
72
+ ckpt_step ....................................... None
73
+ classes_fraction ................................ 1.0
74
+ clip_grad ....................................... 1.0
75
+ clone_scatter_output_in_embedding ............... True
76
+ config_logger_dir ...............................
77
+ consumed_train_samples .......................... 0
78
+ consumed_valid_samples .......................... 0
79
+ context_parallel_size ........................... 1
80
+ cp_comm_type .................................... ['p2p']
81
+ create_attention_mask_in_dataloader ............. True
82
+ cross_entropy_fusion_impl ....................... native
83
+ cross_entropy_loss_fusion ....................... False
84
+ cuda_graph_scope ................................ full_iteration
85
+ cuda_graph_warmup_steps ......................... 3
86
+ data_args_path .................................. None
87
+ data_cache_path ................................. None
88
+ data_parallel_random_init ....................... False
89
+ data_parallel_sharding_strategy ................. no_shard
90
+ data_parallel_size .............................. 8
91
+ data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document']
92
+ data_per_class_fraction ......................... 1.0
93
+ data_sharding ................................... True
94
+ dataloader_type ................................. single
95
+ ddp_average_in_collective ....................... False
96
+ ddp_bucket_size ................................. None
97
+ ddp_num_buckets ................................. None
98
+ ddp_pad_buckets_for_high_nccl_busbw ............. False
99
+ decoder_first_pipeline_num_layers ............... None
100
+ decoder_last_pipeline_num_layers ................ None
101
+ decoder_num_layers .............................. None
102
+ decoder_seq_length .............................. None
103
+ decoupled_lr .................................... None
104
+ decoupled_min_lr ................................ None
105
+ decrease_batch_size_if_needed ................... False
106
+ defer_embedding_wgrad_compute ................... False
107
+ delay_wgrad_compute ............................. False
108
+ deprecated_use_mcore_models ..................... True
109
+ deterministic_mode .............................. False
110
+ dino_bottleneck_size ............................ 256
111
+ dino_freeze_last_layer .......................... 1
112
+ dino_head_hidden_size ........................... 2048
113
+ dino_local_crops_number ......................... 10
114
+ dino_local_img_size ............................. 96
115
+ dino_norm_last_layer ............................ False
116
+ dino_teacher_temp ............................... 0.07
117
+ dino_warmup_teacher_temp ........................ 0.04
118
+ dino_warmup_teacher_temp_epochs ................. 30
119
+ disable_bf16_reduced_precision_matmul ........... False
120
+ disable_mamba_mem_eff_path ...................... False
121
+ disable_straggler_on_startup .................... False
122
+ dist_ckpt_format_deprecated ..................... None
123
+ dist_ckpt_strictness ............................ assume_ok_unexpected
124
+ distribute_saved_activations .................... False
125
+ distributed_backend ............................. nccl
126
+ distributed_timeout_minutes ..................... 10
127
+ embedding_init_method_std ....................... None
128
+ embedding_path .................................. None
129
+ empty_unused_memory_level ....................... 0
130
+ enable_cuda_graph ............................... True
131
+ enable_experimental ............................. False
132
+ enable_ft_package ............................... False
133
+ enable_gloo_process_groups ...................... True
134
+ enable_msc ...................................... True
135
+ enable_one_logger ............................... True
136
+ encoder_num_layers .............................. 2
137
+ encoder_seq_length .............................. 4096
138
+ end_weight_decay ................................ 0.1
139
+ eod_mask_loss ................................... False
140
+ error_injection_rate ............................ 0
141
+ error_injection_type ............................ transient_error
142
+ eval_interval ................................... 1000000
143
+ eval_iters ...................................... 0
144
+ evidence_data_path .............................. None
145
+ exit_duration_in_mins ........................... None
146
+ exit_interval ................................... None
147
+ exit_on_missing_checkpoint ...................... False
148
+ exit_signal_handler ............................. False
149
+ exp_avg_dtype ................................... torch.float32
150
+ exp_avg_sq_dtype ................................ torch.float32
151
+ expert_model_parallel_size ...................... 8
152
+ expert_tensor_parallel_size ..................... 1
153
+ export_force_local_attention .................... False
154
+ export_kd_cfg ................................... None
155
+ export_kd_teacher_ckpt_format ................... None
156
+ export_kd_teacher_load .......................... None
157
+ export_kv_cache_quant ........................... False
158
+ export_legacy_megatron .......................... False
159
+ export_model_type ............................... GPTModel
160
+ export_moe_apply_probs_on_input ................. False
161
+ export_qk_l2_norm ............................... False
162
+ export_quant_cfg ................................ None
163
+ export_real_quant_cfg ........................... None
164
+ export_te_mcore_model ........................... False
165
+ external_cuda_graph ............................. False
166
+ ffn_hidden_size ................................. 18432
167
+ finetune ........................................ False
168
+ finetune_data_split ............................. train
169
+ finetune_hf_dataset ............................. None
170
+ first_last_layers_bf16 .......................... False
171
+ flash_decode .................................... False
172
+ fp16 ............................................ False
173
+ fp16_lm_cross_entropy ........................... False
174
+ fp32_residual_connection ........................ False
175
+ fp8 ............................................. e4m3
176
+ fp8_amax_compute_algo ........................... most_recent
177
+ fp8_amax_history_len ............................ 1
178
+ fp8_interval .................................... 1
179
+ fp8_margin ...................................... 0
180
+ fp8_param_gather ................................ False
181
+ fp8_recipe ...................................... mxfp8
182
+ fp8_wgrad ....................................... True
183
+ fsdp_double_buffer .............................. False
184
+ global_batch_size ............................... 8
185
+ grad_reduce_in_bf16 ............................. False
186
+ gradient_accumulation_fusion .................... True
187
+ gradient_reduce_div_fusion ...................... True
188
+ group_query_attention ........................... False
189
+ head_lr_mult .................................... 1.0
190
+ heterogeneous_layers_config_encoded_json ........ None
191
+ heterogeneous_layers_config_path ................ None
192
+ hidden_dropout .................................. 0.0
193
+ hidden_size ..................................... 7168
194
+ hierarchical_context_parallel_sizes ............. None
195
+ high_priority_stream_groups ..................... []
196
+ hybrid_attention_ratio .......................... 0.0
197
+ hybrid_mlp_ratio ................................ 0.0
198
+ hybrid_override_pattern ......................... None
199
+ hysteresis ...................................... 2
200
+ ict_head_size ................................... None
201
+ ict_load ........................................ None
202
+ img_h ........................................... 224
203
+ img_w ........................................... 224
204
+ indexer_batch_size .............................. 128
205
+ indexer_log_interval ............................ 1000
206
+ inference_batch_times_seqlen_threshold .......... -1
207
+ inference_dynamic_batching ...................... False
208
+ inference_dynamic_batching_buffer_guaranteed_fraction 0.2
209
+ inference_dynamic_batching_buffer_overflow_factor None
210
+ inference_dynamic_batching_buffer_size_gb ....... 40.0
211
+ inference_dynamic_batching_chunk_size ........... 256
212
+ inference_dynamic_batching_max_requests_override None
213
+ inference_dynamic_batching_max_tokens_override .. None
214
+ inference_dynamic_batching_num_cuda_graphs ...... 16
215
+ inference_max_batch_size ........................ 8
216
+ inference_max_seq_length ........................ 2560
217
+ inference_rng_tracker ........................... False
218
+ init_method_std ................................. 0.02
219
+ init_method_xavier_uniform ...................... False
220
+ init_model_with_meta_device ..................... False
221
+ initial_loss_scale .............................. 4294967296
222
+ inprocess_active_world_size ..................... 8
223
+ inprocess_barrier_timeout ....................... 120
224
+ inprocess_completion_timeout .................... 120
225
+ inprocess_empty_cuda_cache ...................... False
226
+ inprocess_granularity ........................... node
227
+ inprocess_hard_timeout .......................... 90
228
+ inprocess_heartbeat_interval .................... 30
229
+ inprocess_heartbeat_timeout ..................... 60
230
+ inprocess_last_call_wait ........................ 1
231
+ inprocess_max_iterations ........................ None
232
+ inprocess_monitor_process_interval .............. 1.0
233
+ inprocess_monitor_thread_interval ............... 1.0
234
+ inprocess_progress_watchdog_interval ............ 1.0
235
+ inprocess_restart ............................... False
236
+ inprocess_soft_timeout .......................... 60
237
+ inprocess_termination_grace_time ................ 1
238
+ is_hybrid_model ................................. False
239
+ iter_per_epoch .................................. 1250
240
+ iterations_to_skip .............................. []
241
+ keep_fp8_transpose_cache_when_using_custom_fsdp . False
242
+ kitchen_config_file ............................. None
243
+ kitchen_recipe_number ........................... None
244
+ kv_channels ..................................... 64
245
+ kv_lora_rank .................................... 512
246
+ lazy_mpu_init ................................... None
247
+ load ............................................ None
248
+ load_main_params_from_ckpt ...................... None
249
+ load_model_opt_format ........................... False
250
+ local_rank ...................................... 0
251
+ log_energy ...................................... False
252
+ log_interval .................................... 1
253
+ log_loss_scale_to_tensorboard ................... True
254
+ log_memory_to_tensorboard ....................... False
255
+ log_num_zeros_in_grad ........................... False
256
+ log_params_norm ................................. False
257
+ log_progress .................................... False
258
+ log_straggler ................................... False
259
+ log_throughput .................................. True
260
+ log_timers_to_tensorboard ....................... False
261
+ log_validation_ppl_to_tensorboard ............... False
262
+ log_world_size_to_tensorboard ................... False
263
+ logging_level ................................... None
264
+ loss_scale ...................................... None
265
+ loss_scale_window ............................... 1000
266
+ lr .............................................. 0.0001
267
+ lr_decay_iters .................................. 8
268
+ lr_decay_samples ................................ None
269
+ lr_decay_style .................................. cosine
270
+ lr_warmup_fraction .............................. None
271
+ lr_warmup_init .................................. 0.0
272
+ lr_warmup_iters ................................. 5
273
+ lr_warmup_samples ............................... 0
274
+ lr_wsd_decay_iters .............................. None
275
+ lr_wsd_decay_samples ............................ None
276
+ lr_wsd_decay_style .............................. exponential
277
+ main_grads_dtype ................................ torch.float32
278
+ main_params_dtype ............................... torch.float32
279
+ make_vocab_size_divisible_by .................... 128
280
+ mamba_head_dim .................................. 64
281
+ mamba_num_groups ................................ 8
282
+ mamba_num_heads ................................. None
283
+ mamba_state_dim ................................. 128
284
+ manual_gc ....................................... False
285
+ manual_gc_eval .................................. True
286
+ manual_gc_interval .............................. 0
287
+ mask_factor ..................................... 1.0
288
+ mask_prob ....................................... 0.15
289
+ mask_type ....................................... random
290
+ masked_softmax_fusion ........................... False
291
+ max_position_embeddings ......................... 4096
292
+ max_tokens_to_oom ............................... 12000
293
+ memory_snapshot_path ............................ snapshot.pickle
294
+ merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt
295
+ micro_batch_size ................................ 1
296
+ microbatch_group_size_per_vp_stage .............. None
297
+ mid_level_dataset_surplus ....................... 0.005
298
+ min_loss_scale .................................. 1.0
299
+ min_lr .......................................... 1e-05
300
+ mlp_chunks_for_prefill .......................... 1
301
+ mmap_bin_files .................................. True
302
+ mock_data ....................................... False
303
+ moe_apply_probs_on_input ........................ False
304
+ moe_aux_loss_coeff .............................. 0.0
305
+ moe_deepep_num_sms .............................. 20
306
+ moe_echo_enable_random_offloading ............... False
307
+ moe_echo_expert_dispatch_overlap ................ False
308
+ moe_echo_expert_dispatcher_type ................. hybridep
309
+ moe_echo_recompute_expert_dispatch .............. False
310
+ moe_enable_deepep ............................... False
311
+ moe_enable_echo ................................. False
312
+ moe_enable_hybridep ............................. False
313
+ moe_enable_moonep ............................... True
314
+ moe_expert_capacity_factor ...................... None
315
+ moe_extended_tp ................................. False
316
+ moe_ffn_hidden_size ............................. 2048
317
+ moe_grouped_gemm ................................ True
318
+ moe_hybridep_num_sms ............................ 16
319
+ moe_input_jitter_eps ............................ None
320
+ moe_layer_freq .................................. [0, 1]
321
+ moe_layer_recompute ............................. False
322
+ moe_moonep_async_grad_reduce .................... True
323
+ moe_moonep_num_sms .............................. None
324
+ moe_num_echo_experts ............................ None
325
+ moe_pad_expert_input_to_capacity ................ False
326
+ moe_per_layer_logging ........................... False
327
+ moe_permute_fusion .............................. False
328
+ moe_received_token_capacity ..................... None
329
+ moe_router_bias_update_rate ..................... 0.001
330
+ moe_router_dtype ................................ fp32
331
+ moe_router_enable_expert_bias ................... True
332
+ moe_router_force_load_balancing ................. False
333
+ moe_router_group_topk ........................... 1
334
+ moe_router_load_balancing_type .................. seq_aux_loss
335
+ moe_router_num_groups ........................... 1
336
+ moe_router_padding_for_fp8 ...................... False
337
+ moe_router_pre_softmax .......................... False
338
+ moe_router_score_function ....................... sigmoid
339
+ moe_router_topk ................................. 8
340
+ moe_router_topk_scaling_factor .................. 2.827
341
+ moe_shared_expert_intermediate_size ............. 2048
342
+ moe_shared_expert_overlap ....................... False
343
+ moe_token_dispatcher_type ....................... flex
344
+ moe_token_drop_policy ........................... probs
345
+ moe_upcycling_granularity ....................... 1
346
+ moe_use_device_initiated_grouped_gemm ........... True
347
+ moe_use_legacy_grouped_gemm ..................... False
348
+ moe_use_upcycling ............................... False
349
+ moe_z_loss_coeff ................................ None
350
+ mrope_section ................................... None
351
+ mscale .......................................... 1.0
352
+ mscale_all_dim .................................. 1.0
353
+ mtp_loss_scaling_factor ......................... 0.1
354
+ mtp_num_layers .................................. None
355
+ multi_latent_attention .......................... True
356
+ nccl_all_reduce_for_prefill ..................... False
357
+ nccl_communicator_config_path ................... None
358
+ nccl_ub ......................................... False
359
+ no_load_optim ................................... None
360
+ no_load_rng ..................................... None
361
+ no_persist_layer_norm ........................... False
362
+ no_rope_freq .................................... None
363
+ no_save_optim ................................... None
364
+ no_save_rng ..................................... None
365
+ non_persistent_ckpt_type ........................ None
366
+ non_persistent_global_ckpt_dir .................. None
367
+ non_persistent_local_ckpt_algo .................. fully_parallel
368
+ non_persistent_local_ckpt_dir ................... None
369
+ non_persistent_save_interval .................... None
370
+ norm_epsilon .................................... 1e-06
371
+ normalization ................................... RMSNorm
372
+ num_attention_heads ............................. 64
373
+ num_channels .................................... 3
374
+ num_classes ..................................... 1000
375
+ num_dataset_builder_threads ..................... 1
376
+ num_distributed_optimizer_instances ............. 1
377
+ num_experts ..................................... 384
378
+ num_layers ...................................... 2
379
+ num_layers_at_end_in_bf16 ....................... 1
380
+ num_layers_at_start_in_bf16 ..................... 1
381
+ num_layers_per_virtual_pipeline_stage ........... None
382
+ num_query_groups ................................ 1
383
+ num_virtual_stages_per_pipeline_rank ............ None
384
+ num_workers ..................................... 2
385
+ object_storage_cache_path ....................... None
386
+ one_logger_async ................................ False
387
+ one_logger_project .............................. megatron-lm
388
+ one_logger_run_name ............................. None
389
+ onnx_safe ....................................... None
390
+ openai_gelu ..................................... False
391
+ optimizer ....................................... adam
392
+ optimizer_cpu_offload ........................... False
393
+ optimizer_offload_fraction ...................... 1.0
394
+ output_bert_embeddings .......................... False
395
+ overlap_cpu_optimizer_d2h_h2d ................... False
396
+ overlap_grad_reduce ............................. True
397
+ overlap_moe_expert_parallel_comm ................ False
398
+ overlap_p2p_comm ................................ False
399
+ overlap_p2p_comm_warmup_flush ................... False
400
+ overlap_param_gather ............................ True
401
+ overlap_param_gather_with_optimizer_step ........ False
402
+ override_opt_param_scheduler .................... False
403
+ params_dtype .................................... torch.bfloat16
404
+ patch_dim ....................................... 16
405
+ per_split_data_args_path ........................ None
406
+ perform_initialization .......................... True
407
+ pin_cpu_grads ................................... True
408
+ pin_cpu_params .................................. True
409
+ pipeline_model_parallel_comm_backend ............ None
410
+ pipeline_model_parallel_layout .................. None
411
+ pipeline_model_parallel_size .................... 1
412
+ position_embedding_type ......................... rope
413
+ pretrained_checkpoint ........................... None
414
+ profile ......................................... False
415
+ profile_ranks ................................... [0]
416
+ profile_step_end ................................ 12
417
+ profile_step_start .............................. 10
418
+ q_lora_rank ..................................... 1536
419
+ qk_head_dim ..................................... 128
420
+ qk_l2_norm ...................................... False
421
+ qk_layernorm .................................... True
422
+ qk_pos_emb_head_dim ............................. 64
423
+ query_in_block_prob ............................. 0.1
424
+ rampup_batch_size ............................... None
425
+ rank ............................................ 0
426
+ recompute_granularity ........................... None
427
+ recompute_method ................................ None
428
+ recompute_modules ............................... None
429
+ recompute_num_layers ............................ None
430
+ record_memory_history ........................... False
431
+ relative_attention_max_distance ................. 128
432
+ relative_attention_num_buckets .................. 32
433
+ replication ..................................... False
434
+ replication_factor .............................. 2
435
+ replication_jump ................................ None
436
+ rerun_mode ...................................... validate_results
437
+ reset_attention_mask ............................ False
438
+ reset_position_ids .............................. False
439
+ result_rejected_tracker_filename ................ None
440
+ retriever_report_topk_accuracies ................ []
441
+ retriever_score_scaling ......................... False
442
+ retriever_seq_length ............................ 256
443
+ retro_add_retriever ............................. False
444
+ retro_attention_gate ............................ 1
445
+ retro_cyclic_train_iters ........................ None
446
+ retro_encoder_attention_dropout ................. 0.1
447
+ retro_encoder_hidden_dropout .................... 0.1
448
+ retro_encoder_layers ............................ 2
449
+ retro_num_neighbors ............................. 2
450
+ retro_num_retrieved_chunks ...................... 2
451
+ retro_project_dir ............................... None
452
+ retro_verify_neighbor_count ..................... True
453
+ reuse_grad_buf_for_mxfp8_param_ag ............... False
454
+ rope_scaling_factor ............................. 8.0
455
+ rope_type ....................................... None
456
+ rotary_base ..................................... 50000
457
+ rotary_interleaved .............................. False
458
+ rotary_percent .................................. 1.0
459
+ rotary_scaling_factor ........................... 32.0
460
+ rotary_seq_len_interpolation_factor ............. None
461
+ run_workload_inspector_server ................... False
462
+ sample_rate ..................................... 1.0
463
+ save ............................................ None
464
+ save_interval ................................... None
465
+ save_retain_interval ............................ None
466
+ scatter_gather_tensors_in_pipeline .............. True
467
+ seed ............................................ 1234
468
+ seq_length ...................................... 4096
469
+ sequence_parallel ............................... False
470
+ sft ............................................. False
471
+ sft_tokenizer_prompt_format ..................... nemotron-h-aligned
472
+ sgd_momentum .................................... 0.9
473
+ short_seq_prob .................................. 0.1
474
+ skip_train ...................................... False
475
+ skipped_train_samples ........................... 0
476
+ spec ............................................ None
477
+ split ........................................... 100,0,0
478
+ squared_relu .................................... False
479
+ start_weight_decay .............................. 0.1
480
+ straggler_ctrlr_port ............................ 65535
481
+ straggler_minmax_count .......................... 1
482
+ suggested_communication_unit_size ............... None
483
+ swiglu .......................................... True
484
+ swin_backbone_type .............................. tiny
485
+ symmetric_ar_type ............................... None
486
+ te_rng_tracker .................................. True
487
+ tensor_model_parallel_size ...................... 1
488
+ tensorboard_dir ................................. None
489
+ tensorboard_log_interval ........................ 1
490
+ tensorboard_queue_size .......................... 1000
491
+ test_data_path .................................. None
492
+ test_mode ....................................... False
493
+ tiktoken_num_special_tokens ..................... 1000
494
+ tiktoken_pattern ................................ None
495
+ tiktoken_special_tokens ......................... None
496
+ timing_log_level ................................ 0
497
+ timing_log_option ............................... minmax
498
+ titles_data_path ................................ None
499
+ tokenizer_model ................................. None
500
+ tokenizer_type .................................. GPT2BPETokenizer
501
+ torch_fsdp2_reshard_after_forward ............... True
502
+ tp_comm_bootstrap_backend ....................... nccl
503
+ tp_comm_bulk_dgrad .............................. True
504
+ tp_comm_bulk_wgrad .............................. True
505
+ tp_comm_overlap ................................. False
506
+ tp_comm_overlap_ag .............................. True
507
+ tp_comm_overlap_cfg ............................. None
508
+ tp_comm_overlap_rs .............................. True
509
+ tp_comm_overlap_rs_dgrad ........................ False
510
+ tp_comm_split_ag ................................ True
511
+ tp_comm_split_rs ................................ True
512
+ train_data_path ................................. None
513
+ train_iters ..................................... 12
514
+ train_samples ................................... None
515
+ train_sync_interval ............................. None
516
+ transformer_impl ................................ transformer_engine
517
+ transformer_pipeline_model_parallel_size ........ 1
518
+ untie_embeddings_and_output_weights ............. True
519
+ use_checkpoint_args ............................. False
520
+ use_checkpoint_opt_param_scheduler .............. False
521
+ use_cpu_initialization .......................... None
522
+ use_custom_fsdp ................................. False
523
+ use_dist_ckpt ................................... True
524
+ use_dist_ckpt_deprecated ........................ False
525
+ use_distributed_optimizer ....................... True
526
+ use_flash_attn .................................. False
527
+ use_legacy_models ............................... False
528
+ use_mp_args_from_checkpoint_args ................ False
529
+ use_one_sent_docs ............................... False
530
+ use_persistent_ckpt_worker ...................... False
531
+ use_precision_aware_optimizer ................... False
532
+ use_pytorch_profiler ............................ False
533
+ use_ring_exchange_p2p ........................... False
534
+ use_rope_scaling ................................ False
535
+ use_rotary_position_embeddings .................. False
536
+ use_sharp ....................................... False
537
+ use_tokenizer_model_from_checkpoint_args ........ True
538
+ use_torch_fsdp2 ................................. False
539
+ use_torch_optimizer_for_cpu_offload ............. False
540
+ use_tp_pp_dp_mapping ............................ False
541
+ v_head_dim ...................................... 128
542
+ valid_data_path ................................. None
543
+ variable_seq_lengths ............................ False
544
+ virtual_pipeline_model_parallel_size ............ None
545
+ vision_backbone_type ............................ vit
546
+ vision_pretraining .............................. False
547
+ vision_pretraining_type ......................... classify
548
+ vocab_extra_ids ................................. 0
549
+ vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json
550
+ vocab_size ...................................... None
551
+ wandb_exp_name ..................................
552
+ wandb_project ...................................
553
+ wandb_save_dir ..................................
554
+ weight_decay .................................... 0.1
555
+ weight_decay_incr_style ......................... constant
556
+ wgrad_deferral_limit ............................ 0
557
+ world_size ...................................... 8
558
+ yaml_cfg ........................................ None
559
+ -------------------- end of arguments ---------------------
560
+ INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1
561
+ > building GPT2BPETokenizer tokenizer ...
562
+ > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304)
563
+ WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS
564
+ > initializing torch distributed ...
565
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
566
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
567
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
568
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
569
+ 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
570
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
571
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
572
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
573
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
574
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
575
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
576
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
577
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
578
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
579
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
580
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
581
+ [Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
582
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
583
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
584
+ [Gloo] Rank [Gloo] Rank [Gloo] Rank [Gloo] Rank 4576 is connected to is connected to is connected to is connected to 7777 peer ranks. peer ranks. peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : Expected number of connected peer ranks is : Expected number of connected peer ranks is : 7777
585
+
586
+
587
+
588
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
589
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
590
+ [Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
591
+ [Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
592
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : [Gloo] Rank 7[Gloo] Rank 6
593
+ 5 is connected to is connected to 77 peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : 77
594
+
595
+ [Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
596
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
597
+ > initialized tensor model parallel with size 1
598
+ > initialized pipeline model parallel with size 1
599
+ > setting random seeds to 1234 ...
600
+ [Gloo] Rank [Gloo] Rank 0 is connected to 00 is connected to peer ranks. 0Expected number of connected peer ranks is : peer ranks. 0Expected number of connected peer ranks is :
601
+ 0
602
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
603
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
604
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
605
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
606
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
607
+ [rank3]:[W921 07:53:53.222525889 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
608
+ [rank5]:[W921 07:53:53.222642533 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 5] using GPU 5 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
609
+ [rank1]:[W921 07:53:53.222644058 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
610
+ [rank4]:[W921 07:53:53.222672331 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 4] using GPU 4 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
611
+ [rank2]:[W921 07:53:53.222734423 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 2] using GPU 2 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
612
+ [rank6]:[W921 07:53:53.222782446 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 6] using GPU 6 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
613
+ [rank7]:[W921 07:53:53.223915471 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
614
+ > compiling dataset index builder ...
615
+ make: Entering directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
616
+ make: Nothing to be done for 'default'.
617
+ make: Leaving directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
618
+ >>> done with dataset index builder. Compilation time: 0.033 seconds
619
+ WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations.
620
+ > compiling and loading fused kernels ...
621
+ [rank0]:[W921 07:53:53.383864505 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 0] using GPU 0 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
622
+ >>> done with compiling and loading fused kernels. Compilation time: 3.796 seconds
623
+ time to initialize megatron (seconds): 7.336
624
+ [after megatron is initialized] datetime: 2026-09-21 07:53:59
625
+ building GPT model ...
626
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
627
+ warnings.warn(
628
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
629
+ warnings.warn(
630
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
631
+ warnings.warn(
632
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
633
+ warnings.warn(
634
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
635
+ warnings.warn(
636
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
637
+ warnings.warn(
638
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
639
+ warnings.warn(
640
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
641
+ warnings.warn(
642
+ > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 3480525824
643
+ INFO:megatron.core.distributed.distributed_data_parallel:Setting up DistributedDataParallel with config DistributedDataParallelConfig(grad_reduce_in_fp32=True, overlap_grad_reduce=True, overlap_param_gather=True, align_param_gather=False, use_distributed_optimizer=True, num_distributed_optimizer_instances=1, check_for_nan_in_grad=False, check_for_large_grads=False, bucket_size=40000000, pad_buckets_for_high_nccl_busbw=False, average_in_collective=False, fp8_param_gather=False, reuse_grad_buf_for_mxfp8_param_ag=False, use_custom_fsdp=False, data_parallel_sharding_strategy='no_shard', gradient_reduce_div_fusion=True, suggested_communication_unit_size=None, preserve_fp32_weights=True, keep_fp8_transpose_cache_when_using_custom_fsdp=False, nccl_ub=False, fsdp_double_buffer=False)
644
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 9
645
+ Params for bucket 1 (360579072 elements, 360579072 padded size):
646
+ module.output_layer.weight
647
+ Params for bucket 2 (44047360 elements, 44047360 padded size):
648
+ module.decoder.final_layernorm.weight
649
+ module.decoder.layers.1.mlp.shared_experts.linear_fc2.weight
650
+ module.decoder.layers.1.mlp.shared_experts.linear_fc1.weight
651
+ Params for bucket 3 (45163520 elements, 45163520 padded size):
652
+ module.decoder.layers.1.mlp.router.weight
653
+ module.decoder.layers.1.self_attention.linear_kv_down_proj.weight
654
+ module.decoder.layers.1.pre_mlp_layernorm.weight
655
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.layer_norm_weight
656
+ module.decoder.layers.1.self_attention.linear_q_up_proj.weight
657
+ module.decoder.layers.1.self_attention.linear_q_up_proj.layer_norm_weight
658
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.weight
659
+ module.decoder.layers.1.self_attention.linear_q_down_proj.weight
660
+ Params for bucket 4 (58720256 elements, 58720256 padded size):
661
+ module.decoder.layers.1.self_attention.linear_proj.weight
662
+ Params for bucket 5 (132127744 elements, 132127744 padded size):
663
+ module.decoder.layers.1.input_layernorm.weight
664
+ module.decoder.layers.0.mlp.linear_fc2.weight
665
+ Params for bucket 6 (264241152 elements, 264241152 padded size):
666
+ module.decoder.layers.0.mlp.linear_fc1.weight
667
+ Params for bucket 7 (42411008 elements, 42411008 padded size):
668
+ module.decoder.layers.0.mlp.linear_fc1.layer_norm_weight
669
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.layer_norm_weight
670
+ module.decoder.layers.0.self_attention.linear_kv_down_proj.weight
671
+ module.decoder.layers.0.self_attention.linear_q_up_proj.weight
672
+ module.decoder.layers.0.self_attention.linear_q_up_proj.layer_norm_weight
673
+ module.decoder.layers.0.self_attention.linear_q_down_proj.weight
674
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.weight
675
+ Params for bucket 8 (58720256 elements, 58720256 padded size):
676
+ module.decoder.layers.0.self_attention.linear_proj.weight
677
+ Params for bucket 9 (360586240 elements, 360586240 padded size):
678
+ module.decoder.layers.0.input_layernorm.weight
679
+ module.embedding.word_embeddings.weight
680
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 40
681
+ Params for bucket 1 (44040192 elements, 44040192 padded size):
682
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight0
683
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight1
684
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight2
685
+ Params for bucket 2 (44040192 elements, 44040192 padded size):
686
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight3
687
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight4
688
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight5
689
+ Params for bucket 3 (44040192 elements, 44040192 padded size):
690
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight6
691
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight7
692
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight8
693
+ Params for bucket 4 (44040192 elements, 44040192 padded size):
694
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight9
695
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight11
696
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight10
697
+ Params for bucket 5 (44040192 elements, 44040192 padded size):
698
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight12
699
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight14
700
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight13
701
+ Params for bucket 6 (44040192 elements, 44040192 padded size):
702
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight15
703
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight17
704
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight16
705
+ Params for bucket 7 (44040192 elements, 44040192 padded size):
706
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight18
707
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight20
708
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight19
709
+ Params for bucket 8 (44040192 elements, 44040192 padded size):
710
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight21
711
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight23
712
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight22
713
+ Params for bucket 9 (44040192 elements, 44040192 padded size):
714
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight24
715
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight25
716
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight26
717
+ Params for bucket 10 (44040192 elements, 44040192 padded size):
718
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight27
719
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight28
720
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight29
721
+ Params for bucket 11 (44040192 elements, 44040192 padded size):
722
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight30
723
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight31
724
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight32
725
+ Params for bucket 12 (44040192 elements, 44040192 padded size):
726
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight33
727
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight34
728
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight35
729
+ Params for bucket 13 (44040192 elements, 44040192 padded size):
730
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight36
731
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight37
732
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight38
733
+ Params for bucket 14 (44040192 elements, 44040192 padded size):
734
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight39
735
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight40
736
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight41
737
+ Params for bucket 15 (44040192 elements, 44040192 padded size):
738
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight42
739
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight43
740
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight44
741
+ Params for bucket 16 (44040192 elements, 44040192 padded size):
742
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight45
743
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight46
744
+ module.decoder.layers.1.mlp.experts.linear_fc2.weight47
745
+ Params for bucket 17 (58720256 elements, 58720256 padded size):
746
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight0
747
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight1
748
+ Params for bucket 18 (58720256 elements, 58720256 padded size):
749
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight2
750
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight3
751
+ Params for bucket 19 (58720256 elements, 58720256 padded size):
752
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight4
753
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight5
754
+ Params for bucket 20 (58720256 elements, 58720256 padded size):
755
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight6
756
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight7
757
+ Params for bucket 21 (58720256 elements, 58720256 padded size):
758
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight8
759
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight9
760
+ Params for bucket 22 (58720256 elements, 58720256 padded size):
761
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight10
762
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight11
763
+ Params for bucket 23 (58720256 elements, 58720256 padded size):
764
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight12
765
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight13
766
+ Params for bucket 24 (58720256 elements, 58720256 padded size):
767
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight14
768
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight15
769
+ Params for bucket 25 (58720256 elements, 58720256 padded size):
770
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight16
771
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight17
772
+ Params for bucket 26 (58720256 elements, 58720256 padded size):
773
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight18
774
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight19
775
+ Params for bucket 27 (58720256 elements, 58720256 padded size):
776
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight20
777
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight21
778
+ Params for bucket 28 (58720256 elements, 58720256 padded size):
779
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight22
780
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight23
781
+ Params for bucket 29 (58720256 elements, 58720256 padded size):
782
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight24
783
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight25
784
+ Params for bucket 30 (58720256 elements, 58720256 padded size):
785
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight26
786
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight27
787
+ Params for bucket 31 (58720256 elements, 58720256 padded size):
788
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight28
789
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight29
790
+ Params for bucket 32 (58720256 elements, 58720256 padded size):
791
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight30
792
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight31
793
+ Params for bucket 33 (58720256 elements, 58720256 padded size):
794
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight32
795
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight33
796
+ Params for bucket 34 (58720256 elements, 58720256 padded size):
797
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight34
798
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight35
799
+ Params for bucket 35 (58720256 elements, 58720256 padded size):
800
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight36
801
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight37
802
+ Params for bucket 36 (58720256 elements, 58720256 padded size):
803
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight38
804
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight39
805
+ Params for bucket 37 (58720256 elements, 58720256 padded size):
806
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight40
807
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight41
808
+ Params for bucket 38 (58720256 elements, 58720256 padded size):
809
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight42
810
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight43
811
+ Params for bucket 39 (58720256 elements, 58720256 padded size):
812
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight44
813
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight45
814
+ Params for bucket 40 (58720256 elements, 58720256 padded size):
815
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight46
816
+ module.decoder.layers.1.mlp.experts.linear_fc1.weight47
817
+ INFO:megatron.core.optimizer:Setting up optimizer with config OptimizerConfig(optimizer='adam', lr=0.0001, min_lr=1e-05, decoupled_lr=None, decoupled_min_lr=None, weight_decay=0.1, fp8_recipe='mxfp8', fp16=False, bf16=True, reuse_grad_buf_for_mxfp8_param_ag=False, params_dtype=torch.bfloat16, use_precision_aware_optimizer=False, store_param_remainders=True, main_grads_dtype=torch.float32, main_params_dtype=torch.float32, exp_avg_dtype=torch.float32, exp_avg_sq_dtype=torch.float32, loss_scale=None, initial_loss_scale=4294967296, min_loss_scale=1.0, loss_scale_window=1000, hysteresis=2, adam_beta1=0.9, adam_beta2=0.999, adam_eps=1e-08, sgd_momentum=0.9, use_distributed_optimizer=True, overlap_param_gather_with_optimizer_step=False, optimizer_cpu_offload=False, optimizer_offload_fraction=1.0, use_torch_optimizer_for_cpu_offload=False, overlap_cpu_optimizer_d2h_h2d=False, pin_cpu_grads=True, pin_cpu_params=True, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=<megatron.core.timers.Timers object at 0x71207e83ed50>, config_logger_dir='')
818
+ INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine
819
+ [after model, optimizer, and learning rate scheduler are built] datetime: 2026-09-21 07:54:01
820
+ > building train, validation, and test datasets ...
821
+ > datasets target sizes (minimum size):
822
+ train: 96
823
+ validation: 0
824
+ test: 0
825
+ INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None]
826
+ > building train, validation, and test datasets for GPT ...
827
+ INFO:megatron.core.datasets.blended_megatron_dataset_builder:Building GPTDataset splits with sizes=(96, 0, 0) and config=GPTDatasetConfig(random_seed=1234, sequence_length=4096, blend=(['/workspace/data/wikitext/wikitext103_gpt2_text_document'], None), blend_per_split=None, split='100,0,0', split_matrix=[(0, 1.0), None, None], num_dataset_builder_threads=1, path_to_cache=None, mmap_bin_files=True, mock=False, tokenizer=<megatron.training.tokenizer.tokenizer._GPT2BPETokenizer object at 0x711dd81d0da0>, mid_level_dataset_surplus=0.005, reset_position_ids=False, reset_attention_mask=False, eod_mask_loss=False, create_attention_mask=True, drop_last_partial_validation_sequence=True, add_extra_token_to_sequence=True, object_storage_cache_path=None)
828
+ INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx
829
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths
830
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers
831
+ INFO:megatron.core.datasets.indexed_dataset: Extract the document indices
832
+ INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029
833
+ INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350
834
+ INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices
835
+ INFO:megatron.core.datasets.gpt_dataset: Load the document index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-document_index.npy
836
+ INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-sample_index.npy
837
+ INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-shuffle_index.npy
838
+ INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073
839
+ > finished creating GPT datasets ...
840
+ [after dataloaders are built] datetime: 2026-09-21 07:54:01
841
+ done with setup ...
842
+ (min, max) time across ranks (ms):
843
+ model-and-optimizer-setup ......................: (2273.96, 2282.37)
844
+ train/valid/test-data-iterators-setup ..........: (66.08, 78.02)
845
+ training ...
846
+ Overwriting rerun_state_machine.current_iteration from -1 to 0...
847
+ [before the start of training step] datetime: 2026-09-21 07:54:02
848
+ INFO:moonep.api:MoonEP comm buffer sizes:
849
+ config : R=8, S=4096, H=7168, E=384, K=8, NvS=44960, NvS_padded=45056
850
+ sm_count : num_sms=32, num_sms_dedup=148
851
+ hidden_buffer_size : 616.00 MiB
852
+ meta_buffer_size : 2.00 MiB
853
+ local_temp_buffer_size : 1.29 MiB
854
+ total_size : 619.29 MiB
855
+ [MoonEP selfcheck 1/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 384, 385, 386, 387, 388, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404, 405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419, 420, 421, 422, 423, 424, 425, 426, 427, 428, 429, 430, 431] sizes [384, 384, 256, 128, 384, 1152, 1664, 1152, 1664, 640, 896, 640, 256, 896, 256, 1280, 384, 512, 1664, 1536, 768, 384, 512, 256, 384, 1152, 640, 256, 256, 384, 384, 640, 640, 2048, 384, 640, 384, 256, 768, 1408, 512, 256, 128, 256, 128, 640, 256, 384, 3456, 640, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[348, 357, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [332, 309, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [314, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [76, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [230, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [208, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [76, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
856
+ [2026-09-21 07:54:27] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 25230.2 | throughput per GPU (TFLOP/s/GPU): 1.4 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.227924E+01 | loss scale: 1.0 | grad norm: 12.616 | number of skipped iterations: 0 | number of nan iterations: 0 |Number of parameters in transformer block in billions: 17.55
857
+
858
+ Number of parameters in embedding layers in billions: 0.72
859
+ Total number of parameters in billions: 18.28
860
+ Number of parameters in most loaded shard in billions: 18.2753
861
+ Theoretical memory footprints: weight and optimizer=130715.23 MB
862
+ [Rank 0] (after 1 iterations) memory (MB) | allocated: 45359.1982421875 | max allocated: 45359.19873046875 | reserved: 45790.0 | max reserved: 45790.0
863
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
864
+ warnings.warn(
865
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
866
+ warnings.warn(
867
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
868
+ warnings.warn(
869
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
870
+ warnings.warn(
871
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
872
+ warnings.warn(
873
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
874
+ warnings.warn(
875
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
876
+ warnings.warn(
877
+ /workspace/Megatron-MoE-Echo/megatron/core/distributed/param_and_grad_buffer.py:368: UserWarning: The next bucket's parameter all-gather operation has already been dispatched. This may be caused by a mismatch between the order of parameter registration and forward pass execution, which will hurt the communication-computation overlap performance.
878
+ warnings.warn(
879
+ [MoonEP selfcheck 2/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 384, 385, 386, 387, 388, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404, 405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419, 420, 421, 422, 423, 424, 425, 426, 427, 428, 429, 430, 431] sizes [384, 384, 384, 128, 512, 1152, 1792, 1024, 1536, 768, 768, 640, 256, 512, 256, 1408, 512, 384, 2048, 1280, 768, 512, 512, 256, 512, 1152, 896, 128, 256, 384, 640, 512, 640, 2048, 512, 768, 512, 256, 896, 1536, 640, 128, 128, 256, 128, 768, 256, 512, 3456, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[348, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [357, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [332, 309, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [230, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [76, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [314, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [76, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
880
+ [2026-09-21 07:54:27] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 125.5 | throughput per GPU (TFLOP/s/GPU): 281.9 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.227087E+01 | loss scale: 1.0 | grad norm: 12.533 | number of skipped iterations: 0 | number of nan iterations: 0 |
881
+ [MoonEP selfcheck 3/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 384, 385, 386, 387, 388, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404, 405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419, 420, 421, 422, 423, 424, 425, 426, 427, 428, 429, 430, 431] sizes [256, 512, 256, 256, 512, 896, 1920, 1024, 2176, 896, 1536, 640, 256, 640, 384, 640, 512, 640, 1408, 896, 768, 512, 512, 256, 512, 1152, 768, 256, 384, 384, 640, 512, 512, 1408, 512, 896, 512, 256, 768, 1024, 640, 256, 256, 384, 128, 384, 256, 768, 3456, 384, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[313, 309, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [208, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [338, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [230, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [306, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [76, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [332, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
882
+ [2026-09-21 07:54:27] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 111.0 | throughput per GPU (TFLOP/s/GPU): 318.6 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 1.088362E+01 | loss scale: 1.0 | grad norm: 11.029 | number of skipped iterations: 0 | number of nan iterations: 0 |
883
+ Capture CUDA graph for training!!!
884
+ CUDA graph capture done!!!
885
+ [2026-09-21 07:54:28] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 494.3 | throughput per GPU (TFLOP/s/GPU): 71.6 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 1.448777E+01 | loss scale: 1.0 | grad norm: 72.970 | number of skipped iterations: 0 | number of nan iterations: 0 |
886
+ [2026-09-21 07:54:28] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 108.1 | throughput per GPU (TFLOP/s/GPU): 327.2 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.808960E+01 | loss scale: 1.0 | grad norm: 108.333 | number of skipped iterations: 0 | number of nan iterations: 0 |
887
+ [2026-09-21 07:54:28] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 109.0 | throughput per GPU (TFLOP/s/GPU): 324.5 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 1.490298E+01 | loss scale: 1.0 | grad norm: 83.291 | number of skipped iterations: 0 | number of nan iterations: 0 |
888
+ [2026-09-21 07:54:28] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 109.8 | throughput per GPU (TFLOP/s/GPU): 322.1 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 1.142665E+01 | loss scale: 1.0 | grad norm: 26.340 | number of skipped iterations: 0 | number of nan iterations: 0 |
889
+ [2026-09-21 07:54:28] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 108.8 | throughput per GPU (TFLOP/s/GPU): 325.3 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.165925E+01 | loss scale: 1.0 | grad norm: 12.676 | number of skipped iterations: 0 | number of nan iterations: 0 |
890
+ [2026-09-21 07:54:28] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 109.1 | throughput per GPU (TFLOP/s/GPU): 324.2 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.132117E+01 | loss scale: 1.0 | grad norm: 10.523 | number of skipped iterations: 0 | number of nan iterations: 0 |
891
+ [2026-09-21 07:54:28] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 109.3 | throughput per GPU (TFLOP/s/GPU): 323.6 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.097885E+01 | loss scale: 1.0 | grad norm: 9.172 | number of skipped iterations: 0 | number of nan iterations: 0 |
892
+ [2026-09-21 07:54:28] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 108.6 | throughput per GPU (TFLOP/s/GPU): 325.8 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.072919E+01 | loss scale: 1.0 | grad norm: 9.283 | number of skipped iterations: 0 | number of nan iterations: 0 |
893
+ [2026-09-21 07:54:28] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 108.4 | throughput per GPU (TFLOP/s/GPU): 326.4 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.047170E+01 | loss scale: 1.0 | grad norm: 9.277 | number of skipped iterations: 0 | number of nan iterations: 0 |
894
+ [after training is done] datetime: 2026-09-21 07:54:28
895
+ [rank0]:[W921 07:54:30.955123182 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
896
+ [rank2]:[W921 07:54:31.261639254 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
897
+ [rank3]:[W921 07:54:31.261963895 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
898
+ [rank7]:[W921 07:54:31.262935101 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
899
+ [rank4]:[W921 07:54:31.263176425 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
900
+ [rank6]:[W921 07:54:31.263407998 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
901
+ [rank5]:[W921 07:54:31.524302258 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
902
+ [rank1]:[W921 07:54:32.029826445 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
903
+ W0921 07:54:36.601000 542044 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers
904
+ W0921 07:54:36.601000 542044 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 543623 closing signal SIGTERM
905
+ W0921 07:54:36.602000 542044 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 543624 closing signal SIGTERM
906
+ W0921 07:54:36.602000 542044 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 543629 closing signal SIGTERM
907
+ W0921 07:54:36.602000 542044 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 543631 closing signal SIGTERM
908
+ Traceback (most recent call last):
909
+ File "/usr/local/bin/torchrun", line 8, in <module>
910
+ sys.exit(main())
911
+ ^^^^^^
912
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
913
+ return f(*args, **kwargs)
914
+ ^^^^^^^^^^^^^^^^^^
915
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 892, in main
916
+ run(args)
917
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 883, in run
918
+ elastic_launch(
919
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
920
+ return launch_agent(self._config, self._entrypoint, list(args))
921
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
922
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 261, in launch_agent
923
+ result = agent.run()
924
+ ^^^^^^^^^^^
925
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper
926
+ result = f(*args, **kwargs)
927
+ ^^^^^^^^^^^^^^^^^^
928
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 711, in run
929
+ result = self._invoke_run(role)
930
+ ^^^^^^^^^^^^^^^^^^^^^^
931
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 870, in _invoke_run
932
+ time.sleep(monitor_interval)
933
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 84, in _terminate_process_handler
934
+ raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval)
935
+ torch.distributed.elastic.multiprocessing.api.SignalException: Process 542044 got signal: 15
granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/results.tsv CHANGED
@@ -1,2 +1,4 @@
1
  config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir
2
  KIMI_K2_L2 moonep matched 48 4096 8 none none 1 OK 108.60 107.70 86491 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep1
 
 
 
1
  config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir
2
  KIMI_K2_L2 moonep matched 48 4096 8 none none 1 OK 108.60 107.70 86491 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep1
3
+ KIMI_K2_L2 moonep matched 48 4096 8 none none 2 OK 108.74 107.70 86491 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep2
4
+ KIMI_K2_L2 moonep matched 48 4096 8 none none 3 OK 108.82 107.70 86491 /workspace/granularity_next_b200/tierb_models/attempts/KIMI_K2_L2/moonep/tierb/tierb_gbs8/KIMI_K2_L2__moonep__matched__seq4096__gbs8__capnone__async/rep3